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> |
b3f20785 | 44 | #include <linux/compat.h> |
0793a61d | 45 | |
76369139 FW |
46 | #include "internal.h" |
47 | ||
4e193bd4 TB |
48 | #include <asm/irq_regs.h> |
49 | ||
fadfe7be JO |
50 | static struct workqueue_struct *perf_wq; |
51 | ||
fe4b04fa | 52 | struct remote_function_call { |
e7e7ee2e IM |
53 | struct task_struct *p; |
54 | int (*func)(void *info); | |
55 | void *info; | |
56 | int ret; | |
fe4b04fa PZ |
57 | }; |
58 | ||
59 | static void remote_function(void *data) | |
60 | { | |
61 | struct remote_function_call *tfc = data; | |
62 | struct task_struct *p = tfc->p; | |
63 | ||
64 | if (p) { | |
65 | tfc->ret = -EAGAIN; | |
66 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
67 | return; | |
68 | } | |
69 | ||
70 | tfc->ret = tfc->func(tfc->info); | |
71 | } | |
72 | ||
73 | /** | |
74 | * task_function_call - call a function on the cpu on which a task runs | |
75 | * @p: the task to evaluate | |
76 | * @func: the function to be called | |
77 | * @info: the function call argument | |
78 | * | |
79 | * Calls the function @func when the task is currently running. This might | |
80 | * be on the current CPU, which just calls the function directly | |
81 | * | |
82 | * returns: @func return value, or | |
83 | * -ESRCH - when the process isn't running | |
84 | * -EAGAIN - when the process moved away | |
85 | */ | |
86 | static int | |
87 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
88 | { | |
89 | struct remote_function_call data = { | |
e7e7ee2e IM |
90 | .p = p, |
91 | .func = func, | |
92 | .info = info, | |
93 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
94 | }; |
95 | ||
96 | if (task_curr(p)) | |
97 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
98 | ||
99 | return data.ret; | |
100 | } | |
101 | ||
102 | /** | |
103 | * cpu_function_call - call a function on the cpu | |
104 | * @func: the function to be called | |
105 | * @info: the function call argument | |
106 | * | |
107 | * Calls the function @func on the remote cpu. | |
108 | * | |
109 | * returns: @func return value or -ENXIO when the cpu is offline | |
110 | */ | |
111 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
112 | { | |
113 | struct remote_function_call data = { | |
e7e7ee2e IM |
114 | .p = NULL, |
115 | .func = func, | |
116 | .info = info, | |
117 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
118 | }; |
119 | ||
120 | smp_call_function_single(cpu, remote_function, &data, 1); | |
121 | ||
122 | return data.ret; | |
123 | } | |
124 | ||
f8697762 JO |
125 | #define EVENT_OWNER_KERNEL ((void *) -1) |
126 | ||
127 | static bool is_kernel_event(struct perf_event *event) | |
128 | { | |
129 | return event->owner == EVENT_OWNER_KERNEL; | |
130 | } | |
131 | ||
e5d1367f SE |
132 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
133 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
134 | PERF_FLAG_PID_CGROUP |\ |
135 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 136 | |
bce38cd5 SE |
137 | /* |
138 | * branch priv levels that need permission checks | |
139 | */ | |
140 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
141 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
142 | PERF_SAMPLE_BRANCH_HV) | |
143 | ||
0b3fcf17 SE |
144 | enum event_type_t { |
145 | EVENT_FLEXIBLE = 0x1, | |
146 | EVENT_PINNED = 0x2, | |
147 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
148 | }; | |
149 | ||
e5d1367f SE |
150 | /* |
151 | * perf_sched_events : >0 events exist | |
152 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
153 | */ | |
c5905afb | 154 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 155 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 156 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 157 | |
cdd6c482 IM |
158 | static atomic_t nr_mmap_events __read_mostly; |
159 | static atomic_t nr_comm_events __read_mostly; | |
160 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 161 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 162 | |
108b02cf PZ |
163 | static LIST_HEAD(pmus); |
164 | static DEFINE_MUTEX(pmus_lock); | |
165 | static struct srcu_struct pmus_srcu; | |
166 | ||
0764771d | 167 | /* |
cdd6c482 | 168 | * perf event paranoia level: |
0fbdea19 IM |
169 | * -1 - not paranoid at all |
170 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 171 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 172 | * 2 - disallow kernel profiling for unpriv |
0764771d | 173 | */ |
cdd6c482 | 174 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 175 | |
20443384 FW |
176 | /* Minimum for 512 kiB + 1 user control page */ |
177 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
178 | |
179 | /* | |
cdd6c482 | 180 | * max perf event sample rate |
df58ab24 | 181 | */ |
14c63f17 DH |
182 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
183 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
184 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
185 | ||
186 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
187 | ||
188 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
189 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
190 | ||
d9494cb4 PZ |
191 | static int perf_sample_allowed_ns __read_mostly = |
192 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
193 | |
194 | void update_perf_cpu_limits(void) | |
195 | { | |
196 | u64 tmp = perf_sample_period_ns; | |
197 | ||
198 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 199 | do_div(tmp, 100); |
d9494cb4 | 200 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 201 | } |
163ec435 | 202 | |
9e630205 SE |
203 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
204 | ||
163ec435 PZ |
205 | int perf_proc_update_handler(struct ctl_table *table, int write, |
206 | void __user *buffer, size_t *lenp, | |
207 | loff_t *ppos) | |
208 | { | |
723478c8 | 209 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
210 | |
211 | if (ret || !write) | |
212 | return ret; | |
213 | ||
214 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
215 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
216 | update_perf_cpu_limits(); | |
217 | ||
218 | return 0; | |
219 | } | |
220 | ||
221 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
222 | ||
223 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
224 | void __user *buffer, size_t *lenp, | |
225 | loff_t *ppos) | |
226 | { | |
227 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
228 | ||
229 | if (ret || !write) | |
230 | return ret; | |
231 | ||
232 | update_perf_cpu_limits(); | |
163ec435 PZ |
233 | |
234 | return 0; | |
235 | } | |
1ccd1549 | 236 | |
14c63f17 DH |
237 | /* |
238 | * perf samples are done in some very critical code paths (NMIs). | |
239 | * If they take too much CPU time, the system can lock up and not | |
240 | * get any real work done. This will drop the sample rate when | |
241 | * we detect that events are taking too long. | |
242 | */ | |
243 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 244 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 245 | |
6a02ad66 | 246 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 247 | { |
6a02ad66 | 248 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 249 | u64 avg_local_sample_len; |
e5302920 | 250 | u64 local_samples_len; |
6a02ad66 PZ |
251 | |
252 | local_samples_len = __get_cpu_var(running_sample_length); | |
253 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
254 | ||
255 | printk_ratelimited(KERN_WARNING | |
256 | "perf interrupt took too long (%lld > %lld), lowering " | |
257 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 258 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
259 | sysctl_perf_event_sample_rate); |
260 | } | |
261 | ||
262 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
263 | ||
264 | void perf_sample_event_took(u64 sample_len_ns) | |
265 | { | |
d9494cb4 | 266 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
267 | u64 avg_local_sample_len; |
268 | u64 local_samples_len; | |
14c63f17 | 269 | |
d9494cb4 | 270 | if (allowed_ns == 0) |
14c63f17 DH |
271 | return; |
272 | ||
273 | /* decay the counter by 1 average sample */ | |
274 | local_samples_len = __get_cpu_var(running_sample_length); | |
275 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
276 | local_samples_len += sample_len_ns; | |
277 | __get_cpu_var(running_sample_length) = local_samples_len; | |
278 | ||
279 | /* | |
280 | * note: this will be biased artifically low until we have | |
281 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
282 | * from having to maintain a count. | |
283 | */ | |
284 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
285 | ||
d9494cb4 | 286 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
287 | return; |
288 | ||
289 | if (max_samples_per_tick <= 1) | |
290 | return; | |
291 | ||
292 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
293 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
294 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
295 | ||
14c63f17 | 296 | update_perf_cpu_limits(); |
6a02ad66 | 297 | |
cd578abb PZ |
298 | if (!irq_work_queue(&perf_duration_work)) { |
299 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
300 | "kernel.perf_event_max_sample_rate to %d\n", | |
301 | avg_local_sample_len, allowed_ns >> 1, | |
302 | sysctl_perf_event_sample_rate); | |
303 | } | |
14c63f17 DH |
304 | } |
305 | ||
cdd6c482 | 306 | static atomic64_t perf_event_id; |
a96bbc16 | 307 | |
0b3fcf17 SE |
308 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
309 | enum event_type_t event_type); | |
310 | ||
311 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
312 | enum event_type_t event_type, |
313 | struct task_struct *task); | |
314 | ||
315 | static void update_context_time(struct perf_event_context *ctx); | |
316 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 317 | |
cdd6c482 | 318 | void __weak perf_event_print_debug(void) { } |
0793a61d | 319 | |
84c79910 | 320 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 321 | { |
84c79910 | 322 | return "pmu"; |
0793a61d TG |
323 | } |
324 | ||
0b3fcf17 SE |
325 | static inline u64 perf_clock(void) |
326 | { | |
327 | return local_clock(); | |
328 | } | |
329 | ||
e5d1367f SE |
330 | static inline struct perf_cpu_context * |
331 | __get_cpu_context(struct perf_event_context *ctx) | |
332 | { | |
333 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
334 | } | |
335 | ||
facc4307 PZ |
336 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
337 | struct perf_event_context *ctx) | |
338 | { | |
339 | raw_spin_lock(&cpuctx->ctx.lock); | |
340 | if (ctx) | |
341 | raw_spin_lock(&ctx->lock); | |
342 | } | |
343 | ||
344 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
345 | struct perf_event_context *ctx) | |
346 | { | |
347 | if (ctx) | |
348 | raw_spin_unlock(&ctx->lock); | |
349 | raw_spin_unlock(&cpuctx->ctx.lock); | |
350 | } | |
351 | ||
e5d1367f SE |
352 | #ifdef CONFIG_CGROUP_PERF |
353 | ||
877c6856 LZ |
354 | /* |
355 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
356 | * This is a per-cpu dynamically allocated data structure. | |
357 | */ | |
358 | struct perf_cgroup_info { | |
359 | u64 time; | |
360 | u64 timestamp; | |
361 | }; | |
362 | ||
363 | struct perf_cgroup { | |
364 | struct cgroup_subsys_state css; | |
86e213e1 | 365 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
366 | }; |
367 | ||
3f7cce3c SE |
368 | /* |
369 | * Must ensure cgroup is pinned (css_get) before calling | |
370 | * this function. In other words, we cannot call this function | |
371 | * if there is no cgroup event for the current CPU context. | |
372 | */ | |
e5d1367f SE |
373 | static inline struct perf_cgroup * |
374 | perf_cgroup_from_task(struct task_struct *task) | |
375 | { | |
073219e9 | 376 | return container_of(task_css(task, perf_event_cgrp_id), |
8af01f56 | 377 | struct perf_cgroup, css); |
e5d1367f SE |
378 | } |
379 | ||
380 | static inline bool | |
381 | perf_cgroup_match(struct perf_event *event) | |
382 | { | |
383 | struct perf_event_context *ctx = event->ctx; | |
384 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
385 | ||
ef824fa1 TH |
386 | /* @event doesn't care about cgroup */ |
387 | if (!event->cgrp) | |
388 | return true; | |
389 | ||
390 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
391 | if (!cpuctx->cgrp) | |
392 | return false; | |
393 | ||
394 | /* | |
395 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
396 | * also enabled for all its descendant cgroups. If @cpuctx's | |
397 | * cgroup is a descendant of @event's (the test covers identity | |
398 | * case), it's a match. | |
399 | */ | |
400 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
401 | event->cgrp->css.cgroup); | |
e5d1367f SE |
402 | } |
403 | ||
e5d1367f SE |
404 | static inline void perf_put_cgroup(struct perf_event *event) |
405 | { | |
406 | css_put(&event->cgrp->css); | |
407 | } | |
408 | ||
409 | static inline void perf_detach_cgroup(struct perf_event *event) | |
410 | { | |
411 | perf_put_cgroup(event); | |
412 | event->cgrp = NULL; | |
413 | } | |
414 | ||
415 | static inline int is_cgroup_event(struct perf_event *event) | |
416 | { | |
417 | return event->cgrp != NULL; | |
418 | } | |
419 | ||
420 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
421 | { | |
422 | struct perf_cgroup_info *t; | |
423 | ||
424 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
425 | return t->time; | |
426 | } | |
427 | ||
428 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
429 | { | |
430 | struct perf_cgroup_info *info; | |
431 | u64 now; | |
432 | ||
433 | now = perf_clock(); | |
434 | ||
435 | info = this_cpu_ptr(cgrp->info); | |
436 | ||
437 | info->time += now - info->timestamp; | |
438 | info->timestamp = now; | |
439 | } | |
440 | ||
441 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
442 | { | |
443 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
444 | if (cgrp_out) | |
445 | __update_cgrp_time(cgrp_out); | |
446 | } | |
447 | ||
448 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
449 | { | |
3f7cce3c SE |
450 | struct perf_cgroup *cgrp; |
451 | ||
e5d1367f | 452 | /* |
3f7cce3c SE |
453 | * ensure we access cgroup data only when needed and |
454 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 455 | */ |
3f7cce3c | 456 | if (!is_cgroup_event(event)) |
e5d1367f SE |
457 | return; |
458 | ||
3f7cce3c SE |
459 | cgrp = perf_cgroup_from_task(current); |
460 | /* | |
461 | * Do not update time when cgroup is not active | |
462 | */ | |
463 | if (cgrp == event->cgrp) | |
464 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
465 | } |
466 | ||
467 | static inline void | |
3f7cce3c SE |
468 | perf_cgroup_set_timestamp(struct task_struct *task, |
469 | struct perf_event_context *ctx) | |
e5d1367f SE |
470 | { |
471 | struct perf_cgroup *cgrp; | |
472 | struct perf_cgroup_info *info; | |
473 | ||
3f7cce3c SE |
474 | /* |
475 | * ctx->lock held by caller | |
476 | * ensure we do not access cgroup data | |
477 | * unless we have the cgroup pinned (css_get) | |
478 | */ | |
479 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
480 | return; |
481 | ||
482 | cgrp = perf_cgroup_from_task(task); | |
483 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 484 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
485 | } |
486 | ||
487 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
488 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
489 | ||
490 | /* | |
491 | * reschedule events based on the cgroup constraint of task. | |
492 | * | |
493 | * mode SWOUT : schedule out everything | |
494 | * mode SWIN : schedule in based on cgroup for next | |
495 | */ | |
496 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
497 | { | |
498 | struct perf_cpu_context *cpuctx; | |
499 | struct pmu *pmu; | |
500 | unsigned long flags; | |
501 | ||
502 | /* | |
503 | * disable interrupts to avoid geting nr_cgroup | |
504 | * changes via __perf_event_disable(). Also | |
505 | * avoids preemption. | |
506 | */ | |
507 | local_irq_save(flags); | |
508 | ||
509 | /* | |
510 | * we reschedule only in the presence of cgroup | |
511 | * constrained events. | |
512 | */ | |
513 | rcu_read_lock(); | |
514 | ||
515 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 516 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
517 | if (cpuctx->unique_pmu != pmu) |
518 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 519 | |
e5d1367f SE |
520 | /* |
521 | * perf_cgroup_events says at least one | |
522 | * context on this CPU has cgroup events. | |
523 | * | |
524 | * ctx->nr_cgroups reports the number of cgroup | |
525 | * events for a context. | |
526 | */ | |
527 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
528 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
529 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
530 | |
531 | if (mode & PERF_CGROUP_SWOUT) { | |
532 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
533 | /* | |
534 | * must not be done before ctxswout due | |
535 | * to event_filter_match() in event_sched_out() | |
536 | */ | |
537 | cpuctx->cgrp = NULL; | |
538 | } | |
539 | ||
540 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 541 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
542 | /* |
543 | * set cgrp before ctxsw in to allow | |
544 | * event_filter_match() to not have to pass | |
545 | * task around | |
e5d1367f SE |
546 | */ |
547 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
548 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
549 | } | |
facc4307 PZ |
550 | perf_pmu_enable(cpuctx->ctx.pmu); |
551 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 552 | } |
e5d1367f SE |
553 | } |
554 | ||
555 | rcu_read_unlock(); | |
556 | ||
557 | local_irq_restore(flags); | |
558 | } | |
559 | ||
a8d757ef SE |
560 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
561 | struct task_struct *next) | |
e5d1367f | 562 | { |
a8d757ef SE |
563 | struct perf_cgroup *cgrp1; |
564 | struct perf_cgroup *cgrp2 = NULL; | |
565 | ||
566 | /* | |
567 | * we come here when we know perf_cgroup_events > 0 | |
568 | */ | |
569 | cgrp1 = perf_cgroup_from_task(task); | |
570 | ||
571 | /* | |
572 | * next is NULL when called from perf_event_enable_on_exec() | |
573 | * that will systematically cause a cgroup_switch() | |
574 | */ | |
575 | if (next) | |
576 | cgrp2 = perf_cgroup_from_task(next); | |
577 | ||
578 | /* | |
579 | * only schedule out current cgroup events if we know | |
580 | * that we are switching to a different cgroup. Otherwise, | |
581 | * do no touch the cgroup events. | |
582 | */ | |
583 | if (cgrp1 != cgrp2) | |
584 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
585 | } |
586 | ||
a8d757ef SE |
587 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
588 | struct task_struct *task) | |
e5d1367f | 589 | { |
a8d757ef SE |
590 | struct perf_cgroup *cgrp1; |
591 | struct perf_cgroup *cgrp2 = NULL; | |
592 | ||
593 | /* | |
594 | * we come here when we know perf_cgroup_events > 0 | |
595 | */ | |
596 | cgrp1 = perf_cgroup_from_task(task); | |
597 | ||
598 | /* prev can never be NULL */ | |
599 | cgrp2 = perf_cgroup_from_task(prev); | |
600 | ||
601 | /* | |
602 | * only need to schedule in cgroup events if we are changing | |
603 | * cgroup during ctxsw. Cgroup events were not scheduled | |
604 | * out of ctxsw out if that was not the case. | |
605 | */ | |
606 | if (cgrp1 != cgrp2) | |
607 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
608 | } |
609 | ||
610 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
611 | struct perf_event_attr *attr, | |
612 | struct perf_event *group_leader) | |
613 | { | |
614 | struct perf_cgroup *cgrp; | |
615 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
616 | struct fd f = fdget(fd); |
617 | int ret = 0; | |
e5d1367f | 618 | |
2903ff01 | 619 | if (!f.file) |
e5d1367f SE |
620 | return -EBADF; |
621 | ||
ec903c0c TH |
622 | css = css_tryget_online_from_dir(f.file->f_dentry, |
623 | &perf_event_cgrp_subsys); | |
3db272c0 LZ |
624 | if (IS_ERR(css)) { |
625 | ret = PTR_ERR(css); | |
626 | goto out; | |
627 | } | |
e5d1367f SE |
628 | |
629 | cgrp = container_of(css, struct perf_cgroup, css); | |
630 | event->cgrp = cgrp; | |
631 | ||
632 | /* | |
633 | * all events in a group must monitor | |
634 | * the same cgroup because a task belongs | |
635 | * to only one perf cgroup at a time | |
636 | */ | |
637 | if (group_leader && group_leader->cgrp != cgrp) { | |
638 | perf_detach_cgroup(event); | |
639 | ret = -EINVAL; | |
e5d1367f | 640 | } |
3db272c0 | 641 | out: |
2903ff01 | 642 | fdput(f); |
e5d1367f SE |
643 | return ret; |
644 | } | |
645 | ||
646 | static inline void | |
647 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
648 | { | |
649 | struct perf_cgroup_info *t; | |
650 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
651 | event->shadow_ctx_time = now - t->timestamp; | |
652 | } | |
653 | ||
654 | static inline void | |
655 | perf_cgroup_defer_enabled(struct perf_event *event) | |
656 | { | |
657 | /* | |
658 | * when the current task's perf cgroup does not match | |
659 | * the event's, we need to remember to call the | |
660 | * perf_mark_enable() function the first time a task with | |
661 | * a matching perf cgroup is scheduled in. | |
662 | */ | |
663 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
664 | event->cgrp_defer_enabled = 1; | |
665 | } | |
666 | ||
667 | static inline void | |
668 | perf_cgroup_mark_enabled(struct perf_event *event, | |
669 | struct perf_event_context *ctx) | |
670 | { | |
671 | struct perf_event *sub; | |
672 | u64 tstamp = perf_event_time(event); | |
673 | ||
674 | if (!event->cgrp_defer_enabled) | |
675 | return; | |
676 | ||
677 | event->cgrp_defer_enabled = 0; | |
678 | ||
679 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
680 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
681 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
682 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
683 | sub->cgrp_defer_enabled = 0; | |
684 | } | |
685 | } | |
686 | } | |
687 | #else /* !CONFIG_CGROUP_PERF */ | |
688 | ||
689 | static inline bool | |
690 | perf_cgroup_match(struct perf_event *event) | |
691 | { | |
692 | return true; | |
693 | } | |
694 | ||
695 | static inline void perf_detach_cgroup(struct perf_event *event) | |
696 | {} | |
697 | ||
698 | static inline int is_cgroup_event(struct perf_event *event) | |
699 | { | |
700 | return 0; | |
701 | } | |
702 | ||
703 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
704 | { | |
705 | return 0; | |
706 | } | |
707 | ||
708 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
709 | { | |
710 | } | |
711 | ||
712 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
713 | { | |
714 | } | |
715 | ||
a8d757ef SE |
716 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
717 | struct task_struct *next) | |
e5d1367f SE |
718 | { |
719 | } | |
720 | ||
a8d757ef SE |
721 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
722 | struct task_struct *task) | |
e5d1367f SE |
723 | { |
724 | } | |
725 | ||
726 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
727 | struct perf_event_attr *attr, | |
728 | struct perf_event *group_leader) | |
729 | { | |
730 | return -EINVAL; | |
731 | } | |
732 | ||
733 | static inline void | |
3f7cce3c SE |
734 | perf_cgroup_set_timestamp(struct task_struct *task, |
735 | struct perf_event_context *ctx) | |
e5d1367f SE |
736 | { |
737 | } | |
738 | ||
739 | void | |
740 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
741 | { | |
742 | } | |
743 | ||
744 | static inline void | |
745 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
746 | { | |
747 | } | |
748 | ||
749 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
750 | { | |
751 | return 0; | |
752 | } | |
753 | ||
754 | static inline void | |
755 | perf_cgroup_defer_enabled(struct perf_event *event) | |
756 | { | |
757 | } | |
758 | ||
759 | static inline void | |
760 | perf_cgroup_mark_enabled(struct perf_event *event, | |
761 | struct perf_event_context *ctx) | |
762 | { | |
763 | } | |
764 | #endif | |
765 | ||
9e630205 SE |
766 | /* |
767 | * set default to be dependent on timer tick just | |
768 | * like original code | |
769 | */ | |
770 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
771 | /* | |
772 | * function must be called with interrupts disbled | |
773 | */ | |
774 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
775 | { | |
776 | struct perf_cpu_context *cpuctx; | |
777 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
778 | int rotations = 0; | |
779 | ||
780 | WARN_ON(!irqs_disabled()); | |
781 | ||
782 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
783 | ||
784 | rotations = perf_rotate_context(cpuctx); | |
785 | ||
786 | /* | |
787 | * arm timer if needed | |
788 | */ | |
789 | if (rotations) { | |
790 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
791 | ret = HRTIMER_RESTART; | |
792 | } | |
793 | ||
794 | return ret; | |
795 | } | |
796 | ||
797 | /* CPU is going down */ | |
798 | void perf_cpu_hrtimer_cancel(int cpu) | |
799 | { | |
800 | struct perf_cpu_context *cpuctx; | |
801 | struct pmu *pmu; | |
802 | unsigned long flags; | |
803 | ||
804 | if (WARN_ON(cpu != smp_processor_id())) | |
805 | return; | |
806 | ||
807 | local_irq_save(flags); | |
808 | ||
809 | rcu_read_lock(); | |
810 | ||
811 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
812 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
813 | ||
814 | if (pmu->task_ctx_nr == perf_sw_context) | |
815 | continue; | |
816 | ||
817 | hrtimer_cancel(&cpuctx->hrtimer); | |
818 | } | |
819 | ||
820 | rcu_read_unlock(); | |
821 | ||
822 | local_irq_restore(flags); | |
823 | } | |
824 | ||
825 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
826 | { | |
827 | struct hrtimer *hr = &cpuctx->hrtimer; | |
828 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 829 | int timer; |
9e630205 SE |
830 | |
831 | /* no multiplexing needed for SW PMU */ | |
832 | if (pmu->task_ctx_nr == perf_sw_context) | |
833 | return; | |
834 | ||
62b85639 SE |
835 | /* |
836 | * check default is sane, if not set then force to | |
837 | * default interval (1/tick) | |
838 | */ | |
839 | timer = pmu->hrtimer_interval_ms; | |
840 | if (timer < 1) | |
841 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
842 | ||
843 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
844 | |
845 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
846 | hr->function = perf_cpu_hrtimer_handler; | |
847 | } | |
848 | ||
849 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
850 | { | |
851 | struct hrtimer *hr = &cpuctx->hrtimer; | |
852 | struct pmu *pmu = cpuctx->ctx.pmu; | |
853 | ||
854 | /* not for SW PMU */ | |
855 | if (pmu->task_ctx_nr == perf_sw_context) | |
856 | return; | |
857 | ||
858 | if (hrtimer_active(hr)) | |
859 | return; | |
860 | ||
861 | if (!hrtimer_callback_running(hr)) | |
862 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
863 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
864 | } | |
865 | ||
33696fc0 | 866 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 867 | { |
33696fc0 PZ |
868 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
869 | if (!(*count)++) | |
870 | pmu->pmu_disable(pmu); | |
9e35ad38 | 871 | } |
9e35ad38 | 872 | |
33696fc0 | 873 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 874 | { |
33696fc0 PZ |
875 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
876 | if (!--(*count)) | |
877 | pmu->pmu_enable(pmu); | |
9e35ad38 | 878 | } |
9e35ad38 | 879 | |
e9d2b064 PZ |
880 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
881 | ||
882 | /* | |
883 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
884 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
885 | * disabled, while rotate_context is called from IRQ context. | |
886 | */ | |
108b02cf | 887 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 888 | { |
108b02cf | 889 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 890 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 891 | |
e9d2b064 | 892 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 893 | |
d84153d6 | 894 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 895 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 896 | } |
9e35ad38 | 897 | |
cdd6c482 | 898 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 899 | { |
e5289d4a | 900 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
901 | } |
902 | ||
cdd6c482 | 903 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 904 | { |
564c2b21 PM |
905 | if (atomic_dec_and_test(&ctx->refcount)) { |
906 | if (ctx->parent_ctx) | |
907 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
908 | if (ctx->task) |
909 | put_task_struct(ctx->task); | |
cb796ff3 | 910 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 911 | } |
a63eaf34 PM |
912 | } |
913 | ||
cdd6c482 | 914 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
915 | { |
916 | if (ctx->parent_ctx) { | |
917 | put_ctx(ctx->parent_ctx); | |
918 | ctx->parent_ctx = NULL; | |
919 | } | |
5a3126d4 | 920 | ctx->generation++; |
71a851b4 PZ |
921 | } |
922 | ||
6844c09d ACM |
923 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
924 | { | |
925 | /* | |
926 | * only top level events have the pid namespace they were created in | |
927 | */ | |
928 | if (event->parent) | |
929 | event = event->parent; | |
930 | ||
931 | return task_tgid_nr_ns(p, event->ns); | |
932 | } | |
933 | ||
934 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
935 | { | |
936 | /* | |
937 | * only top level events have the pid namespace they were created in | |
938 | */ | |
939 | if (event->parent) | |
940 | event = event->parent; | |
941 | ||
942 | return task_pid_nr_ns(p, event->ns); | |
943 | } | |
944 | ||
7f453c24 | 945 | /* |
cdd6c482 | 946 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
947 | * to userspace. |
948 | */ | |
cdd6c482 | 949 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 950 | { |
cdd6c482 | 951 | u64 id = event->id; |
7f453c24 | 952 | |
cdd6c482 IM |
953 | if (event->parent) |
954 | id = event->parent->id; | |
7f453c24 PZ |
955 | |
956 | return id; | |
957 | } | |
958 | ||
25346b93 | 959 | /* |
cdd6c482 | 960 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
961 | * This has to cope with with the fact that until it is locked, |
962 | * the context could get moved to another task. | |
963 | */ | |
cdd6c482 | 964 | static struct perf_event_context * |
8dc85d54 | 965 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 966 | { |
cdd6c482 | 967 | struct perf_event_context *ctx; |
25346b93 | 968 | |
9ed6060d | 969 | retry: |
058ebd0e PZ |
970 | /* |
971 | * One of the few rules of preemptible RCU is that one cannot do | |
972 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
973 | * part of the read side critical section was preemptible -- see | |
974 | * rcu_read_unlock_special(). | |
975 | * | |
976 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
977 | * side critical section is non-preemptible. | |
978 | */ | |
979 | preempt_disable(); | |
980 | rcu_read_lock(); | |
8dc85d54 | 981 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
982 | if (ctx) { |
983 | /* | |
984 | * If this context is a clone of another, it might | |
985 | * get swapped for another underneath us by | |
cdd6c482 | 986 | * perf_event_task_sched_out, though the |
25346b93 PM |
987 | * rcu_read_lock() protects us from any context |
988 | * getting freed. Lock the context and check if it | |
989 | * got swapped before we could get the lock, and retry | |
990 | * if so. If we locked the right context, then it | |
991 | * can't get swapped on us any more. | |
992 | */ | |
e625cce1 | 993 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 994 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 995 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
996 | rcu_read_unlock(); |
997 | preempt_enable(); | |
25346b93 PM |
998 | goto retry; |
999 | } | |
b49a9e7e PZ |
1000 | |
1001 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 1002 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1003 | ctx = NULL; |
1004 | } | |
25346b93 PM |
1005 | } |
1006 | rcu_read_unlock(); | |
058ebd0e | 1007 | preempt_enable(); |
25346b93 PM |
1008 | return ctx; |
1009 | } | |
1010 | ||
1011 | /* | |
1012 | * Get the context for a task and increment its pin_count so it | |
1013 | * can't get swapped to another task. This also increments its | |
1014 | * reference count so that the context can't get freed. | |
1015 | */ | |
8dc85d54 PZ |
1016 | static struct perf_event_context * |
1017 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1018 | { |
cdd6c482 | 1019 | struct perf_event_context *ctx; |
25346b93 PM |
1020 | unsigned long flags; |
1021 | ||
8dc85d54 | 1022 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1023 | if (ctx) { |
1024 | ++ctx->pin_count; | |
e625cce1 | 1025 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1026 | } |
1027 | return ctx; | |
1028 | } | |
1029 | ||
cdd6c482 | 1030 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1031 | { |
1032 | unsigned long flags; | |
1033 | ||
e625cce1 | 1034 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1035 | --ctx->pin_count; |
e625cce1 | 1036 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1037 | } |
1038 | ||
f67218c3 PZ |
1039 | /* |
1040 | * Update the record of the current time in a context. | |
1041 | */ | |
1042 | static void update_context_time(struct perf_event_context *ctx) | |
1043 | { | |
1044 | u64 now = perf_clock(); | |
1045 | ||
1046 | ctx->time += now - ctx->timestamp; | |
1047 | ctx->timestamp = now; | |
1048 | } | |
1049 | ||
4158755d SE |
1050 | static u64 perf_event_time(struct perf_event *event) |
1051 | { | |
1052 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1053 | |
1054 | if (is_cgroup_event(event)) | |
1055 | return perf_cgroup_event_time(event); | |
1056 | ||
4158755d SE |
1057 | return ctx ? ctx->time : 0; |
1058 | } | |
1059 | ||
f67218c3 PZ |
1060 | /* |
1061 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1062 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1063 | */ |
1064 | static void update_event_times(struct perf_event *event) | |
1065 | { | |
1066 | struct perf_event_context *ctx = event->ctx; | |
1067 | u64 run_end; | |
1068 | ||
1069 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1070 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1071 | return; | |
e5d1367f SE |
1072 | /* |
1073 | * in cgroup mode, time_enabled represents | |
1074 | * the time the event was enabled AND active | |
1075 | * tasks were in the monitored cgroup. This is | |
1076 | * independent of the activity of the context as | |
1077 | * there may be a mix of cgroup and non-cgroup events. | |
1078 | * | |
1079 | * That is why we treat cgroup events differently | |
1080 | * here. | |
1081 | */ | |
1082 | if (is_cgroup_event(event)) | |
46cd6a7f | 1083 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1084 | else if (ctx->is_active) |
1085 | run_end = ctx->time; | |
acd1d7c1 PZ |
1086 | else |
1087 | run_end = event->tstamp_stopped; | |
1088 | ||
1089 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1090 | |
1091 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1092 | run_end = event->tstamp_stopped; | |
1093 | else | |
4158755d | 1094 | run_end = perf_event_time(event); |
f67218c3 PZ |
1095 | |
1096 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1097 | |
f67218c3 PZ |
1098 | } |
1099 | ||
96c21a46 PZ |
1100 | /* |
1101 | * Update total_time_enabled and total_time_running for all events in a group. | |
1102 | */ | |
1103 | static void update_group_times(struct perf_event *leader) | |
1104 | { | |
1105 | struct perf_event *event; | |
1106 | ||
1107 | update_event_times(leader); | |
1108 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1109 | update_event_times(event); | |
1110 | } | |
1111 | ||
889ff015 FW |
1112 | static struct list_head * |
1113 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1114 | { | |
1115 | if (event->attr.pinned) | |
1116 | return &ctx->pinned_groups; | |
1117 | else | |
1118 | return &ctx->flexible_groups; | |
1119 | } | |
1120 | ||
fccc714b | 1121 | /* |
cdd6c482 | 1122 | * Add a event from the lists for its context. |
fccc714b PZ |
1123 | * Must be called with ctx->mutex and ctx->lock held. |
1124 | */ | |
04289bb9 | 1125 | static void |
cdd6c482 | 1126 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1127 | { |
8a49542c PZ |
1128 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1129 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1130 | |
1131 | /* | |
8a49542c PZ |
1132 | * If we're a stand alone event or group leader, we go to the context |
1133 | * list, group events are kept attached to the group so that | |
1134 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1135 | */ |
8a49542c | 1136 | if (event->group_leader == event) { |
889ff015 FW |
1137 | struct list_head *list; |
1138 | ||
d6f962b5 FW |
1139 | if (is_software_event(event)) |
1140 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1141 | ||
889ff015 FW |
1142 | list = ctx_group_list(event, ctx); |
1143 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1144 | } |
592903cd | 1145 | |
08309379 | 1146 | if (is_cgroup_event(event)) |
e5d1367f | 1147 | ctx->nr_cgroups++; |
e5d1367f | 1148 | |
d010b332 SE |
1149 | if (has_branch_stack(event)) |
1150 | ctx->nr_branch_stack++; | |
1151 | ||
cdd6c482 | 1152 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1153 | if (!ctx->nr_events) |
108b02cf | 1154 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1155 | ctx->nr_events++; |
1156 | if (event->attr.inherit_stat) | |
bfbd3381 | 1157 | ctx->nr_stat++; |
5a3126d4 PZ |
1158 | |
1159 | ctx->generation++; | |
04289bb9 IM |
1160 | } |
1161 | ||
0231bb53 JO |
1162 | /* |
1163 | * Initialize event state based on the perf_event_attr::disabled. | |
1164 | */ | |
1165 | static inline void perf_event__state_init(struct perf_event *event) | |
1166 | { | |
1167 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1168 | PERF_EVENT_STATE_INACTIVE; | |
1169 | } | |
1170 | ||
c320c7b7 ACM |
1171 | /* |
1172 | * Called at perf_event creation and when events are attached/detached from a | |
1173 | * group. | |
1174 | */ | |
1175 | static void perf_event__read_size(struct perf_event *event) | |
1176 | { | |
1177 | int entry = sizeof(u64); /* value */ | |
1178 | int size = 0; | |
1179 | int nr = 1; | |
1180 | ||
1181 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1182 | size += sizeof(u64); | |
1183 | ||
1184 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1185 | size += sizeof(u64); | |
1186 | ||
1187 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1188 | entry += sizeof(u64); | |
1189 | ||
1190 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1191 | nr += event->group_leader->nr_siblings; | |
1192 | size += sizeof(u64); | |
1193 | } | |
1194 | ||
1195 | size += entry * nr; | |
1196 | event->read_size = size; | |
1197 | } | |
1198 | ||
1199 | static void perf_event__header_size(struct perf_event *event) | |
1200 | { | |
1201 | struct perf_sample_data *data; | |
1202 | u64 sample_type = event->attr.sample_type; | |
1203 | u16 size = 0; | |
1204 | ||
1205 | perf_event__read_size(event); | |
1206 | ||
1207 | if (sample_type & PERF_SAMPLE_IP) | |
1208 | size += sizeof(data->ip); | |
1209 | ||
6844c09d ACM |
1210 | if (sample_type & PERF_SAMPLE_ADDR) |
1211 | size += sizeof(data->addr); | |
1212 | ||
1213 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1214 | size += sizeof(data->period); | |
1215 | ||
c3feedf2 AK |
1216 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1217 | size += sizeof(data->weight); | |
1218 | ||
6844c09d ACM |
1219 | if (sample_type & PERF_SAMPLE_READ) |
1220 | size += event->read_size; | |
1221 | ||
d6be9ad6 SE |
1222 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1223 | size += sizeof(data->data_src.val); | |
1224 | ||
fdfbbd07 AK |
1225 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1226 | size += sizeof(data->txn); | |
1227 | ||
6844c09d ACM |
1228 | event->header_size = size; |
1229 | } | |
1230 | ||
1231 | static void perf_event__id_header_size(struct perf_event *event) | |
1232 | { | |
1233 | struct perf_sample_data *data; | |
1234 | u64 sample_type = event->attr.sample_type; | |
1235 | u16 size = 0; | |
1236 | ||
c320c7b7 ACM |
1237 | if (sample_type & PERF_SAMPLE_TID) |
1238 | size += sizeof(data->tid_entry); | |
1239 | ||
1240 | if (sample_type & PERF_SAMPLE_TIME) | |
1241 | size += sizeof(data->time); | |
1242 | ||
ff3d527c AH |
1243 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1244 | size += sizeof(data->id); | |
1245 | ||
c320c7b7 ACM |
1246 | if (sample_type & PERF_SAMPLE_ID) |
1247 | size += sizeof(data->id); | |
1248 | ||
1249 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1250 | size += sizeof(data->stream_id); | |
1251 | ||
1252 | if (sample_type & PERF_SAMPLE_CPU) | |
1253 | size += sizeof(data->cpu_entry); | |
1254 | ||
6844c09d | 1255 | event->id_header_size = size; |
c320c7b7 ACM |
1256 | } |
1257 | ||
8a49542c PZ |
1258 | static void perf_group_attach(struct perf_event *event) |
1259 | { | |
c320c7b7 | 1260 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1261 | |
74c3337c PZ |
1262 | /* |
1263 | * We can have double attach due to group movement in perf_event_open. | |
1264 | */ | |
1265 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1266 | return; | |
1267 | ||
8a49542c PZ |
1268 | event->attach_state |= PERF_ATTACH_GROUP; |
1269 | ||
1270 | if (group_leader == event) | |
1271 | return; | |
1272 | ||
1273 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1274 | !is_software_event(event)) | |
1275 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1276 | ||
1277 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1278 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1279 | |
1280 | perf_event__header_size(group_leader); | |
1281 | ||
1282 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1283 | perf_event__header_size(pos); | |
8a49542c PZ |
1284 | } |
1285 | ||
a63eaf34 | 1286 | /* |
cdd6c482 | 1287 | * Remove a event from the lists for its context. |
fccc714b | 1288 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1289 | */ |
04289bb9 | 1290 | static void |
cdd6c482 | 1291 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1292 | { |
68cacd29 | 1293 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1294 | /* |
1295 | * We can have double detach due to exit/hot-unplug + close. | |
1296 | */ | |
1297 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1298 | return; |
8a49542c PZ |
1299 | |
1300 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1301 | ||
68cacd29 | 1302 | if (is_cgroup_event(event)) { |
e5d1367f | 1303 | ctx->nr_cgroups--; |
68cacd29 SE |
1304 | cpuctx = __get_cpu_context(ctx); |
1305 | /* | |
1306 | * if there are no more cgroup events | |
1307 | * then cler cgrp to avoid stale pointer | |
1308 | * in update_cgrp_time_from_cpuctx() | |
1309 | */ | |
1310 | if (!ctx->nr_cgroups) | |
1311 | cpuctx->cgrp = NULL; | |
1312 | } | |
e5d1367f | 1313 | |
d010b332 SE |
1314 | if (has_branch_stack(event)) |
1315 | ctx->nr_branch_stack--; | |
1316 | ||
cdd6c482 IM |
1317 | ctx->nr_events--; |
1318 | if (event->attr.inherit_stat) | |
bfbd3381 | 1319 | ctx->nr_stat--; |
8bc20959 | 1320 | |
cdd6c482 | 1321 | list_del_rcu(&event->event_entry); |
04289bb9 | 1322 | |
8a49542c PZ |
1323 | if (event->group_leader == event) |
1324 | list_del_init(&event->group_entry); | |
5c148194 | 1325 | |
96c21a46 | 1326 | update_group_times(event); |
b2e74a26 SE |
1327 | |
1328 | /* | |
1329 | * If event was in error state, then keep it | |
1330 | * that way, otherwise bogus counts will be | |
1331 | * returned on read(). The only way to get out | |
1332 | * of error state is by explicit re-enabling | |
1333 | * of the event | |
1334 | */ | |
1335 | if (event->state > PERF_EVENT_STATE_OFF) | |
1336 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1337 | |
1338 | ctx->generation++; | |
050735b0 PZ |
1339 | } |
1340 | ||
8a49542c | 1341 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1342 | { |
1343 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1344 | struct list_head *list = NULL; |
1345 | ||
1346 | /* | |
1347 | * We can have double detach due to exit/hot-unplug + close. | |
1348 | */ | |
1349 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1350 | return; | |
1351 | ||
1352 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1353 | ||
1354 | /* | |
1355 | * If this is a sibling, remove it from its group. | |
1356 | */ | |
1357 | if (event->group_leader != event) { | |
1358 | list_del_init(&event->group_entry); | |
1359 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1360 | goto out; |
8a49542c PZ |
1361 | } |
1362 | ||
1363 | if (!list_empty(&event->group_entry)) | |
1364 | list = &event->group_entry; | |
2e2af50b | 1365 | |
04289bb9 | 1366 | /* |
cdd6c482 IM |
1367 | * If this was a group event with sibling events then |
1368 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1369 | * to whatever list we are on. |
04289bb9 | 1370 | */ |
cdd6c482 | 1371 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1372 | if (list) |
1373 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1374 | sibling->group_leader = sibling; |
d6f962b5 FW |
1375 | |
1376 | /* Inherit group flags from the previous leader */ | |
1377 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1378 | } |
c320c7b7 ACM |
1379 | |
1380 | out: | |
1381 | perf_event__header_size(event->group_leader); | |
1382 | ||
1383 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1384 | perf_event__header_size(tmp); | |
04289bb9 IM |
1385 | } |
1386 | ||
fadfe7be JO |
1387 | /* |
1388 | * User event without the task. | |
1389 | */ | |
1390 | static bool is_orphaned_event(struct perf_event *event) | |
1391 | { | |
1392 | return event && !is_kernel_event(event) && !event->owner; | |
1393 | } | |
1394 | ||
1395 | /* | |
1396 | * Event has a parent but parent's task finished and it's | |
1397 | * alive only because of children holding refference. | |
1398 | */ | |
1399 | static bool is_orphaned_child(struct perf_event *event) | |
1400 | { | |
1401 | return is_orphaned_event(event->parent); | |
1402 | } | |
1403 | ||
1404 | static void orphans_remove_work(struct work_struct *work); | |
1405 | ||
1406 | static void schedule_orphans_remove(struct perf_event_context *ctx) | |
1407 | { | |
1408 | if (!ctx->task || ctx->orphans_remove_sched || !perf_wq) | |
1409 | return; | |
1410 | ||
1411 | if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) { | |
1412 | get_ctx(ctx); | |
1413 | ctx->orphans_remove_sched = true; | |
1414 | } | |
1415 | } | |
1416 | ||
1417 | static int __init perf_workqueue_init(void) | |
1418 | { | |
1419 | perf_wq = create_singlethread_workqueue("perf"); | |
1420 | WARN(!perf_wq, "failed to create perf workqueue\n"); | |
1421 | return perf_wq ? 0 : -1; | |
1422 | } | |
1423 | ||
1424 | core_initcall(perf_workqueue_init); | |
1425 | ||
fa66f07a SE |
1426 | static inline int |
1427 | event_filter_match(struct perf_event *event) | |
1428 | { | |
e5d1367f SE |
1429 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1430 | && perf_cgroup_match(event); | |
fa66f07a SE |
1431 | } |
1432 | ||
9ffcfa6f SE |
1433 | static void |
1434 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1435 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1436 | struct perf_event_context *ctx) |
3b6f9e5c | 1437 | { |
4158755d | 1438 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1439 | u64 delta; |
1440 | /* | |
1441 | * An event which could not be activated because of | |
1442 | * filter mismatch still needs to have its timings | |
1443 | * maintained, otherwise bogus information is return | |
1444 | * via read() for time_enabled, time_running: | |
1445 | */ | |
1446 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1447 | && !event_filter_match(event)) { | |
e5d1367f | 1448 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1449 | event->tstamp_running += delta; |
4158755d | 1450 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1451 | } |
1452 | ||
cdd6c482 | 1453 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1454 | return; |
3b6f9e5c | 1455 | |
44377277 AS |
1456 | perf_pmu_disable(event->pmu); |
1457 | ||
cdd6c482 IM |
1458 | event->state = PERF_EVENT_STATE_INACTIVE; |
1459 | if (event->pending_disable) { | |
1460 | event->pending_disable = 0; | |
1461 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1462 | } |
4158755d | 1463 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1464 | event->pmu->del(event, 0); |
cdd6c482 | 1465 | event->oncpu = -1; |
3b6f9e5c | 1466 | |
cdd6c482 | 1467 | if (!is_software_event(event)) |
3b6f9e5c PM |
1468 | cpuctx->active_oncpu--; |
1469 | ctx->nr_active--; | |
0f5a2601 PZ |
1470 | if (event->attr.freq && event->attr.sample_freq) |
1471 | ctx->nr_freq--; | |
cdd6c482 | 1472 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1473 | cpuctx->exclusive = 0; |
44377277 | 1474 | |
fadfe7be JO |
1475 | if (is_orphaned_child(event)) |
1476 | schedule_orphans_remove(ctx); | |
1477 | ||
44377277 | 1478 | perf_pmu_enable(event->pmu); |
3b6f9e5c PM |
1479 | } |
1480 | ||
d859e29f | 1481 | static void |
cdd6c482 | 1482 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1483 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1484 | struct perf_event_context *ctx) |
d859e29f | 1485 | { |
cdd6c482 | 1486 | struct perf_event *event; |
fa66f07a | 1487 | int state = group_event->state; |
d859e29f | 1488 | |
cdd6c482 | 1489 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1490 | |
1491 | /* | |
1492 | * Schedule out siblings (if any): | |
1493 | */ | |
cdd6c482 IM |
1494 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1495 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1496 | |
fa66f07a | 1497 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1498 | cpuctx->exclusive = 0; |
1499 | } | |
1500 | ||
46ce0fe9 PZ |
1501 | struct remove_event { |
1502 | struct perf_event *event; | |
1503 | bool detach_group; | |
1504 | }; | |
1505 | ||
0793a61d | 1506 | /* |
cdd6c482 | 1507 | * Cross CPU call to remove a performance event |
0793a61d | 1508 | * |
cdd6c482 | 1509 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1510 | * remove it from the context list. |
1511 | */ | |
fe4b04fa | 1512 | static int __perf_remove_from_context(void *info) |
0793a61d | 1513 | { |
46ce0fe9 PZ |
1514 | struct remove_event *re = info; |
1515 | struct perf_event *event = re->event; | |
cdd6c482 | 1516 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1517 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1518 | |
e625cce1 | 1519 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1520 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1521 | if (re->detach_group) |
1522 | perf_group_detach(event); | |
cdd6c482 | 1523 | list_del_event(event, ctx); |
64ce3126 PZ |
1524 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1525 | ctx->is_active = 0; | |
1526 | cpuctx->task_ctx = NULL; | |
1527 | } | |
e625cce1 | 1528 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1529 | |
1530 | return 0; | |
0793a61d TG |
1531 | } |
1532 | ||
1533 | ||
1534 | /* | |
cdd6c482 | 1535 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1536 | * |
cdd6c482 | 1537 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1538 | * call when the task is on a CPU. |
c93f7669 | 1539 | * |
cdd6c482 IM |
1540 | * If event->ctx is a cloned context, callers must make sure that |
1541 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1542 | * remains valid. This is OK when called from perf_release since |
1543 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1544 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1545 | * context has been detached from its task. |
0793a61d | 1546 | */ |
46ce0fe9 | 1547 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1548 | { |
cdd6c482 | 1549 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1550 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1551 | struct remove_event re = { |
1552 | .event = event, | |
1553 | .detach_group = detach_group, | |
1554 | }; | |
0793a61d | 1555 | |
fe4b04fa PZ |
1556 | lockdep_assert_held(&ctx->mutex); |
1557 | ||
0793a61d TG |
1558 | if (!task) { |
1559 | /* | |
cdd6c482 | 1560 | * Per cpu events are removed via an smp call and |
af901ca1 | 1561 | * the removal is always successful. |
0793a61d | 1562 | */ |
46ce0fe9 | 1563 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1564 | return; |
1565 | } | |
1566 | ||
1567 | retry: | |
46ce0fe9 | 1568 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1569 | return; |
0793a61d | 1570 | |
e625cce1 | 1571 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1572 | /* |
fe4b04fa PZ |
1573 | * If we failed to find a running task, but find the context active now |
1574 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1575 | */ |
fe4b04fa | 1576 | if (ctx->is_active) { |
e625cce1 | 1577 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1578 | goto retry; |
1579 | } | |
1580 | ||
1581 | /* | |
fe4b04fa PZ |
1582 | * Since the task isn't running, its safe to remove the event, us |
1583 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1584 | */ |
46ce0fe9 PZ |
1585 | if (detach_group) |
1586 | perf_group_detach(event); | |
fe4b04fa | 1587 | list_del_event(event, ctx); |
e625cce1 | 1588 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1589 | } |
1590 | ||
d859e29f | 1591 | /* |
cdd6c482 | 1592 | * Cross CPU call to disable a performance event |
d859e29f | 1593 | */ |
500ad2d8 | 1594 | int __perf_event_disable(void *info) |
d859e29f | 1595 | { |
cdd6c482 | 1596 | struct perf_event *event = info; |
cdd6c482 | 1597 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1598 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1599 | |
1600 | /* | |
cdd6c482 IM |
1601 | * If this is a per-task event, need to check whether this |
1602 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1603 | * |
1604 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1605 | * flipping contexts around. | |
d859e29f | 1606 | */ |
665c2142 | 1607 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1608 | return -EINVAL; |
d859e29f | 1609 | |
e625cce1 | 1610 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1611 | |
1612 | /* | |
cdd6c482 | 1613 | * If the event is on, turn it off. |
d859e29f PM |
1614 | * If it is in error state, leave it in error state. |
1615 | */ | |
cdd6c482 | 1616 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1617 | update_context_time(ctx); |
e5d1367f | 1618 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1619 | update_group_times(event); |
1620 | if (event == event->group_leader) | |
1621 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1622 | else |
cdd6c482 IM |
1623 | event_sched_out(event, cpuctx, ctx); |
1624 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1625 | } |
1626 | ||
e625cce1 | 1627 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1628 | |
1629 | return 0; | |
d859e29f PM |
1630 | } |
1631 | ||
1632 | /* | |
cdd6c482 | 1633 | * Disable a event. |
c93f7669 | 1634 | * |
cdd6c482 IM |
1635 | * If event->ctx is a cloned context, callers must make sure that |
1636 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1637 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1638 | * perf_event_for_each_child or perf_event_for_each because they |
1639 | * hold the top-level event's child_mutex, so any descendant that | |
1640 | * goes to exit will block in sync_child_event. | |
1641 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1642 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1643 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1644 | */ |
44234adc | 1645 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1646 | { |
cdd6c482 | 1647 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1648 | struct task_struct *task = ctx->task; |
1649 | ||
1650 | if (!task) { | |
1651 | /* | |
cdd6c482 | 1652 | * Disable the event on the cpu that it's on |
d859e29f | 1653 | */ |
fe4b04fa | 1654 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1655 | return; |
1656 | } | |
1657 | ||
9ed6060d | 1658 | retry: |
fe4b04fa PZ |
1659 | if (!task_function_call(task, __perf_event_disable, event)) |
1660 | return; | |
d859e29f | 1661 | |
e625cce1 | 1662 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1663 | /* |
cdd6c482 | 1664 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1665 | */ |
cdd6c482 | 1666 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1667 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1668 | /* |
1669 | * Reload the task pointer, it might have been changed by | |
1670 | * a concurrent perf_event_context_sched_out(). | |
1671 | */ | |
1672 | task = ctx->task; | |
d859e29f PM |
1673 | goto retry; |
1674 | } | |
1675 | ||
1676 | /* | |
1677 | * Since we have the lock this context can't be scheduled | |
1678 | * in, so we can change the state safely. | |
1679 | */ | |
cdd6c482 IM |
1680 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1681 | update_group_times(event); | |
1682 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1683 | } |
e625cce1 | 1684 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1685 | } |
dcfce4a0 | 1686 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1687 | |
e5d1367f SE |
1688 | static void perf_set_shadow_time(struct perf_event *event, |
1689 | struct perf_event_context *ctx, | |
1690 | u64 tstamp) | |
1691 | { | |
1692 | /* | |
1693 | * use the correct time source for the time snapshot | |
1694 | * | |
1695 | * We could get by without this by leveraging the | |
1696 | * fact that to get to this function, the caller | |
1697 | * has most likely already called update_context_time() | |
1698 | * and update_cgrp_time_xx() and thus both timestamp | |
1699 | * are identical (or very close). Given that tstamp is, | |
1700 | * already adjusted for cgroup, we could say that: | |
1701 | * tstamp - ctx->timestamp | |
1702 | * is equivalent to | |
1703 | * tstamp - cgrp->timestamp. | |
1704 | * | |
1705 | * Then, in perf_output_read(), the calculation would | |
1706 | * work with no changes because: | |
1707 | * - event is guaranteed scheduled in | |
1708 | * - no scheduled out in between | |
1709 | * - thus the timestamp would be the same | |
1710 | * | |
1711 | * But this is a bit hairy. | |
1712 | * | |
1713 | * So instead, we have an explicit cgroup call to remain | |
1714 | * within the time time source all along. We believe it | |
1715 | * is cleaner and simpler to understand. | |
1716 | */ | |
1717 | if (is_cgroup_event(event)) | |
1718 | perf_cgroup_set_shadow_time(event, tstamp); | |
1719 | else | |
1720 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1721 | } | |
1722 | ||
4fe757dd PZ |
1723 | #define MAX_INTERRUPTS (~0ULL) |
1724 | ||
1725 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1726 | ||
235c7fc7 | 1727 | static int |
9ffcfa6f | 1728 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1729 | struct perf_cpu_context *cpuctx, |
6e37738a | 1730 | struct perf_event_context *ctx) |
235c7fc7 | 1731 | { |
4158755d | 1732 | u64 tstamp = perf_event_time(event); |
44377277 | 1733 | int ret = 0; |
4158755d | 1734 | |
63342411 PZ |
1735 | lockdep_assert_held(&ctx->lock); |
1736 | ||
cdd6c482 | 1737 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1738 | return 0; |
1739 | ||
cdd6c482 | 1740 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1741 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1742 | |
1743 | /* | |
1744 | * Unthrottle events, since we scheduled we might have missed several | |
1745 | * ticks already, also for a heavily scheduling task there is little | |
1746 | * guarantee it'll get a tick in a timely manner. | |
1747 | */ | |
1748 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1749 | perf_log_throttle(event, 1); | |
1750 | event->hw.interrupts = 0; | |
1751 | } | |
1752 | ||
235c7fc7 IM |
1753 | /* |
1754 | * The new state must be visible before we turn it on in the hardware: | |
1755 | */ | |
1756 | smp_wmb(); | |
1757 | ||
44377277 AS |
1758 | perf_pmu_disable(event->pmu); |
1759 | ||
a4eaf7f1 | 1760 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1761 | event->state = PERF_EVENT_STATE_INACTIVE; |
1762 | event->oncpu = -1; | |
44377277 AS |
1763 | ret = -EAGAIN; |
1764 | goto out; | |
235c7fc7 IM |
1765 | } |
1766 | ||
4158755d | 1767 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1768 | |
e5d1367f | 1769 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1770 | |
cdd6c482 | 1771 | if (!is_software_event(event)) |
3b6f9e5c | 1772 | cpuctx->active_oncpu++; |
235c7fc7 | 1773 | ctx->nr_active++; |
0f5a2601 PZ |
1774 | if (event->attr.freq && event->attr.sample_freq) |
1775 | ctx->nr_freq++; | |
235c7fc7 | 1776 | |
cdd6c482 | 1777 | if (event->attr.exclusive) |
3b6f9e5c PM |
1778 | cpuctx->exclusive = 1; |
1779 | ||
fadfe7be JO |
1780 | if (is_orphaned_child(event)) |
1781 | schedule_orphans_remove(ctx); | |
1782 | ||
44377277 AS |
1783 | out: |
1784 | perf_pmu_enable(event->pmu); | |
1785 | ||
1786 | return ret; | |
235c7fc7 IM |
1787 | } |
1788 | ||
6751b71e | 1789 | static int |
cdd6c482 | 1790 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1791 | struct perf_cpu_context *cpuctx, |
6e37738a | 1792 | struct perf_event_context *ctx) |
6751b71e | 1793 | { |
6bde9b6c | 1794 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1795 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1796 | u64 now = ctx->time; |
1797 | bool simulate = false; | |
6751b71e | 1798 | |
cdd6c482 | 1799 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1800 | return 0; |
1801 | ||
ad5133b7 | 1802 | pmu->start_txn(pmu); |
6bde9b6c | 1803 | |
9ffcfa6f | 1804 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1805 | pmu->cancel_txn(pmu); |
9e630205 | 1806 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1807 | return -EAGAIN; |
90151c35 | 1808 | } |
6751b71e PM |
1809 | |
1810 | /* | |
1811 | * Schedule in siblings as one group (if any): | |
1812 | */ | |
cdd6c482 | 1813 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1814 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1815 | partial_group = event; |
6751b71e PM |
1816 | goto group_error; |
1817 | } | |
1818 | } | |
1819 | ||
9ffcfa6f | 1820 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1821 | return 0; |
9ffcfa6f | 1822 | |
6751b71e PM |
1823 | group_error: |
1824 | /* | |
1825 | * Groups can be scheduled in as one unit only, so undo any | |
1826 | * partial group before returning: | |
d7842da4 SE |
1827 | * The events up to the failed event are scheduled out normally, |
1828 | * tstamp_stopped will be updated. | |
1829 | * | |
1830 | * The failed events and the remaining siblings need to have | |
1831 | * their timings updated as if they had gone thru event_sched_in() | |
1832 | * and event_sched_out(). This is required to get consistent timings | |
1833 | * across the group. This also takes care of the case where the group | |
1834 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1835 | * the time the event was actually stopped, such that time delta | |
1836 | * calculation in update_event_times() is correct. | |
6751b71e | 1837 | */ |
cdd6c482 IM |
1838 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1839 | if (event == partial_group) | |
d7842da4 SE |
1840 | simulate = true; |
1841 | ||
1842 | if (simulate) { | |
1843 | event->tstamp_running += now - event->tstamp_stopped; | |
1844 | event->tstamp_stopped = now; | |
1845 | } else { | |
1846 | event_sched_out(event, cpuctx, ctx); | |
1847 | } | |
6751b71e | 1848 | } |
9ffcfa6f | 1849 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1850 | |
ad5133b7 | 1851 | pmu->cancel_txn(pmu); |
90151c35 | 1852 | |
9e630205 SE |
1853 | perf_cpu_hrtimer_restart(cpuctx); |
1854 | ||
6751b71e PM |
1855 | return -EAGAIN; |
1856 | } | |
1857 | ||
3b6f9e5c | 1858 | /* |
cdd6c482 | 1859 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1860 | */ |
cdd6c482 | 1861 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1862 | struct perf_cpu_context *cpuctx, |
1863 | int can_add_hw) | |
1864 | { | |
1865 | /* | |
cdd6c482 | 1866 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1867 | */ |
d6f962b5 | 1868 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1869 | return 1; |
1870 | /* | |
1871 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1872 | * events can go on. |
3b6f9e5c PM |
1873 | */ |
1874 | if (cpuctx->exclusive) | |
1875 | return 0; | |
1876 | /* | |
1877 | * If this group is exclusive and there are already | |
cdd6c482 | 1878 | * events on the CPU, it can't go on. |
3b6f9e5c | 1879 | */ |
cdd6c482 | 1880 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1881 | return 0; |
1882 | /* | |
1883 | * Otherwise, try to add it if all previous groups were able | |
1884 | * to go on. | |
1885 | */ | |
1886 | return can_add_hw; | |
1887 | } | |
1888 | ||
cdd6c482 IM |
1889 | static void add_event_to_ctx(struct perf_event *event, |
1890 | struct perf_event_context *ctx) | |
53cfbf59 | 1891 | { |
4158755d SE |
1892 | u64 tstamp = perf_event_time(event); |
1893 | ||
cdd6c482 | 1894 | list_add_event(event, ctx); |
8a49542c | 1895 | perf_group_attach(event); |
4158755d SE |
1896 | event->tstamp_enabled = tstamp; |
1897 | event->tstamp_running = tstamp; | |
1898 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1899 | } |
1900 | ||
2c29ef0f PZ |
1901 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1902 | static void | |
1903 | ctx_sched_in(struct perf_event_context *ctx, | |
1904 | struct perf_cpu_context *cpuctx, | |
1905 | enum event_type_t event_type, | |
1906 | struct task_struct *task); | |
fe4b04fa | 1907 | |
dce5855b PZ |
1908 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1909 | struct perf_event_context *ctx, | |
1910 | struct task_struct *task) | |
1911 | { | |
1912 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1913 | if (ctx) | |
1914 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1915 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1916 | if (ctx) | |
1917 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1918 | } | |
1919 | ||
0793a61d | 1920 | /* |
cdd6c482 | 1921 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1922 | * |
1923 | * Must be called with ctx->mutex held | |
0793a61d | 1924 | */ |
fe4b04fa | 1925 | static int __perf_install_in_context(void *info) |
0793a61d | 1926 | { |
cdd6c482 IM |
1927 | struct perf_event *event = info; |
1928 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1929 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1930 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1931 | struct task_struct *task = current; | |
1932 | ||
b58f6b0d | 1933 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1934 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1935 | |
1936 | /* | |
2c29ef0f | 1937 | * If there was an active task_ctx schedule it out. |
0793a61d | 1938 | */ |
b58f6b0d | 1939 | if (task_ctx) |
2c29ef0f | 1940 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1941 | |
1942 | /* | |
1943 | * If the context we're installing events in is not the | |
1944 | * active task_ctx, flip them. | |
1945 | */ | |
1946 | if (ctx->task && task_ctx != ctx) { | |
1947 | if (task_ctx) | |
1948 | raw_spin_unlock(&task_ctx->lock); | |
1949 | raw_spin_lock(&ctx->lock); | |
1950 | task_ctx = ctx; | |
1951 | } | |
1952 | ||
1953 | if (task_ctx) { | |
1954 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1955 | task = task_ctx->task; |
1956 | } | |
b58f6b0d | 1957 | |
2c29ef0f | 1958 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1959 | |
4af4998b | 1960 | update_context_time(ctx); |
e5d1367f SE |
1961 | /* |
1962 | * update cgrp time only if current cgrp | |
1963 | * matches event->cgrp. Must be done before | |
1964 | * calling add_event_to_ctx() | |
1965 | */ | |
1966 | update_cgrp_time_from_event(event); | |
0793a61d | 1967 | |
cdd6c482 | 1968 | add_event_to_ctx(event, ctx); |
0793a61d | 1969 | |
d859e29f | 1970 | /* |
2c29ef0f | 1971 | * Schedule everything back in |
d859e29f | 1972 | */ |
dce5855b | 1973 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1974 | |
1975 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1976 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1977 | |
1978 | return 0; | |
0793a61d TG |
1979 | } |
1980 | ||
1981 | /* | |
cdd6c482 | 1982 | * Attach a performance event to a context |
0793a61d | 1983 | * |
cdd6c482 IM |
1984 | * First we add the event to the list with the hardware enable bit |
1985 | * in event->hw_config cleared. | |
0793a61d | 1986 | * |
cdd6c482 | 1987 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1988 | * call to enable it in the task context. The task might have been |
1989 | * scheduled away, but we check this in the smp call again. | |
1990 | */ | |
1991 | static void | |
cdd6c482 IM |
1992 | perf_install_in_context(struct perf_event_context *ctx, |
1993 | struct perf_event *event, | |
0793a61d TG |
1994 | int cpu) |
1995 | { | |
1996 | struct task_struct *task = ctx->task; | |
1997 | ||
fe4b04fa PZ |
1998 | lockdep_assert_held(&ctx->mutex); |
1999 | ||
c3f00c70 | 2000 | event->ctx = ctx; |
0cda4c02 YZ |
2001 | if (event->cpu != -1) |
2002 | event->cpu = cpu; | |
c3f00c70 | 2003 | |
0793a61d TG |
2004 | if (!task) { |
2005 | /* | |
cdd6c482 | 2006 | * Per cpu events are installed via an smp call and |
af901ca1 | 2007 | * the install is always successful. |
0793a61d | 2008 | */ |
fe4b04fa | 2009 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
2010 | return; |
2011 | } | |
2012 | ||
0793a61d | 2013 | retry: |
fe4b04fa PZ |
2014 | if (!task_function_call(task, __perf_install_in_context, event)) |
2015 | return; | |
0793a61d | 2016 | |
e625cce1 | 2017 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 2018 | /* |
fe4b04fa PZ |
2019 | * If we failed to find a running task, but find the context active now |
2020 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 2021 | */ |
fe4b04fa | 2022 | if (ctx->is_active) { |
e625cce1 | 2023 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
2024 | goto retry; |
2025 | } | |
2026 | ||
2027 | /* | |
fe4b04fa PZ |
2028 | * Since the task isn't running, its safe to add the event, us holding |
2029 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 2030 | */ |
fe4b04fa | 2031 | add_event_to_ctx(event, ctx); |
e625cce1 | 2032 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
2033 | } |
2034 | ||
fa289bec | 2035 | /* |
cdd6c482 | 2036 | * Put a event into inactive state and update time fields. |
fa289bec PM |
2037 | * Enabling the leader of a group effectively enables all |
2038 | * the group members that aren't explicitly disabled, so we | |
2039 | * have to update their ->tstamp_enabled also. | |
2040 | * Note: this works for group members as well as group leaders | |
2041 | * since the non-leader members' sibling_lists will be empty. | |
2042 | */ | |
1d9b482e | 2043 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 2044 | { |
cdd6c482 | 2045 | struct perf_event *sub; |
4158755d | 2046 | u64 tstamp = perf_event_time(event); |
fa289bec | 2047 | |
cdd6c482 | 2048 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 2049 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 2050 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
2051 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
2052 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 2053 | } |
fa289bec PM |
2054 | } |
2055 | ||
d859e29f | 2056 | /* |
cdd6c482 | 2057 | * Cross CPU call to enable a performance event |
d859e29f | 2058 | */ |
fe4b04fa | 2059 | static int __perf_event_enable(void *info) |
04289bb9 | 2060 | { |
cdd6c482 | 2061 | struct perf_event *event = info; |
cdd6c482 IM |
2062 | struct perf_event_context *ctx = event->ctx; |
2063 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2064 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2065 | int err; |
04289bb9 | 2066 | |
06f41796 JO |
2067 | /* |
2068 | * There's a time window between 'ctx->is_active' check | |
2069 | * in perf_event_enable function and this place having: | |
2070 | * - IRQs on | |
2071 | * - ctx->lock unlocked | |
2072 | * | |
2073 | * where the task could be killed and 'ctx' deactivated | |
2074 | * by perf_event_exit_task. | |
2075 | */ | |
2076 | if (!ctx->is_active) | |
fe4b04fa | 2077 | return -EINVAL; |
3cbed429 | 2078 | |
e625cce1 | 2079 | raw_spin_lock(&ctx->lock); |
4af4998b | 2080 | update_context_time(ctx); |
d859e29f | 2081 | |
cdd6c482 | 2082 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2083 | goto unlock; |
e5d1367f SE |
2084 | |
2085 | /* | |
2086 | * set current task's cgroup time reference point | |
2087 | */ | |
3f7cce3c | 2088 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2089 | |
1d9b482e | 2090 | __perf_event_mark_enabled(event); |
04289bb9 | 2091 | |
e5d1367f SE |
2092 | if (!event_filter_match(event)) { |
2093 | if (is_cgroup_event(event)) | |
2094 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2095 | goto unlock; |
e5d1367f | 2096 | } |
f4c4176f | 2097 | |
04289bb9 | 2098 | /* |
cdd6c482 | 2099 | * If the event is in a group and isn't the group leader, |
d859e29f | 2100 | * then don't put it on unless the group is on. |
04289bb9 | 2101 | */ |
cdd6c482 | 2102 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2103 | goto unlock; |
3b6f9e5c | 2104 | |
cdd6c482 | 2105 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2106 | err = -EEXIST; |
e758a33d | 2107 | } else { |
cdd6c482 | 2108 | if (event == leader) |
6e37738a | 2109 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2110 | else |
6e37738a | 2111 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2112 | } |
d859e29f PM |
2113 | |
2114 | if (err) { | |
2115 | /* | |
cdd6c482 | 2116 | * If this event can't go on and it's part of a |
d859e29f PM |
2117 | * group, then the whole group has to come off. |
2118 | */ | |
9e630205 | 2119 | if (leader != event) { |
d859e29f | 2120 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2121 | perf_cpu_hrtimer_restart(cpuctx); |
2122 | } | |
0d48696f | 2123 | if (leader->attr.pinned) { |
53cfbf59 | 2124 | update_group_times(leader); |
cdd6c482 | 2125 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2126 | } |
d859e29f PM |
2127 | } |
2128 | ||
9ed6060d | 2129 | unlock: |
e625cce1 | 2130 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2131 | |
2132 | return 0; | |
d859e29f PM |
2133 | } |
2134 | ||
2135 | /* | |
cdd6c482 | 2136 | * Enable a event. |
c93f7669 | 2137 | * |
cdd6c482 IM |
2138 | * If event->ctx is a cloned context, callers must make sure that |
2139 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2140 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2141 | * perf_event_for_each_child or perf_event_for_each as described |
2142 | * for perf_event_disable. | |
d859e29f | 2143 | */ |
44234adc | 2144 | void perf_event_enable(struct perf_event *event) |
d859e29f | 2145 | { |
cdd6c482 | 2146 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2147 | struct task_struct *task = ctx->task; |
2148 | ||
2149 | if (!task) { | |
2150 | /* | |
cdd6c482 | 2151 | * Enable the event on the cpu that it's on |
d859e29f | 2152 | */ |
fe4b04fa | 2153 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2154 | return; |
2155 | } | |
2156 | ||
e625cce1 | 2157 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2158 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2159 | goto out; |
2160 | ||
2161 | /* | |
cdd6c482 IM |
2162 | * If the event is in error state, clear that first. |
2163 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2164 | * know that it has gone back into error state, as distinct |
2165 | * from the task having been scheduled away before the | |
2166 | * cross-call arrived. | |
2167 | */ | |
cdd6c482 IM |
2168 | if (event->state == PERF_EVENT_STATE_ERROR) |
2169 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2170 | |
9ed6060d | 2171 | retry: |
fe4b04fa | 2172 | if (!ctx->is_active) { |
1d9b482e | 2173 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2174 | goto out; |
2175 | } | |
2176 | ||
e625cce1 | 2177 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2178 | |
2179 | if (!task_function_call(task, __perf_event_enable, event)) | |
2180 | return; | |
d859e29f | 2181 | |
e625cce1 | 2182 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2183 | |
2184 | /* | |
cdd6c482 | 2185 | * If the context is active and the event is still off, |
d859e29f PM |
2186 | * we need to retry the cross-call. |
2187 | */ | |
fe4b04fa PZ |
2188 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2189 | /* | |
2190 | * task could have been flipped by a concurrent | |
2191 | * perf_event_context_sched_out() | |
2192 | */ | |
2193 | task = ctx->task; | |
d859e29f | 2194 | goto retry; |
fe4b04fa | 2195 | } |
fa289bec | 2196 | |
9ed6060d | 2197 | out: |
e625cce1 | 2198 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2199 | } |
dcfce4a0 | 2200 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2201 | |
26ca5c11 | 2202 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2203 | { |
2023b359 | 2204 | /* |
cdd6c482 | 2205 | * not supported on inherited events |
2023b359 | 2206 | */ |
2e939d1d | 2207 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2208 | return -EINVAL; |
2209 | ||
cdd6c482 IM |
2210 | atomic_add(refresh, &event->event_limit); |
2211 | perf_event_enable(event); | |
2023b359 PZ |
2212 | |
2213 | return 0; | |
79f14641 | 2214 | } |
26ca5c11 | 2215 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2216 | |
5b0311e1 FW |
2217 | static void ctx_sched_out(struct perf_event_context *ctx, |
2218 | struct perf_cpu_context *cpuctx, | |
2219 | enum event_type_t event_type) | |
235c7fc7 | 2220 | { |
cdd6c482 | 2221 | struct perf_event *event; |
db24d33e | 2222 | int is_active = ctx->is_active; |
235c7fc7 | 2223 | |
db24d33e | 2224 | ctx->is_active &= ~event_type; |
cdd6c482 | 2225 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2226 | return; |
2227 | ||
4af4998b | 2228 | update_context_time(ctx); |
e5d1367f | 2229 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2230 | if (!ctx->nr_active) |
facc4307 | 2231 | return; |
5b0311e1 | 2232 | |
075e0b00 | 2233 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2234 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2235 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2236 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2237 | } |
889ff015 | 2238 | |
db24d33e | 2239 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2240 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2241 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2242 | } |
1b9a644f | 2243 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2244 | } |
2245 | ||
564c2b21 | 2246 | /* |
5a3126d4 PZ |
2247 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2248 | * cloned from the same version of the same context. | |
2249 | * | |
2250 | * Equivalence is measured using a generation number in the context that is | |
2251 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2252 | * and list_del_event(). | |
564c2b21 | 2253 | */ |
cdd6c482 IM |
2254 | static int context_equiv(struct perf_event_context *ctx1, |
2255 | struct perf_event_context *ctx2) | |
564c2b21 | 2256 | { |
5a3126d4 PZ |
2257 | /* Pinning disables the swap optimization */ |
2258 | if (ctx1->pin_count || ctx2->pin_count) | |
2259 | return 0; | |
2260 | ||
2261 | /* If ctx1 is the parent of ctx2 */ | |
2262 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2263 | return 1; | |
2264 | ||
2265 | /* If ctx2 is the parent of ctx1 */ | |
2266 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2267 | return 1; | |
2268 | ||
2269 | /* | |
2270 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2271 | * hierarchy, see perf_event_init_context(). | |
2272 | */ | |
2273 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2274 | ctx1->parent_gen == ctx2->parent_gen) | |
2275 | return 1; | |
2276 | ||
2277 | /* Unmatched */ | |
2278 | return 0; | |
564c2b21 PM |
2279 | } |
2280 | ||
cdd6c482 IM |
2281 | static void __perf_event_sync_stat(struct perf_event *event, |
2282 | struct perf_event *next_event) | |
bfbd3381 PZ |
2283 | { |
2284 | u64 value; | |
2285 | ||
cdd6c482 | 2286 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2287 | return; |
2288 | ||
2289 | /* | |
cdd6c482 | 2290 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2291 | * because we're in the middle of a context switch and have IRQs |
2292 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2293 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2294 | * don't need to use it. |
2295 | */ | |
cdd6c482 IM |
2296 | switch (event->state) { |
2297 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2298 | event->pmu->read(event); |
2299 | /* fall-through */ | |
bfbd3381 | 2300 | |
cdd6c482 IM |
2301 | case PERF_EVENT_STATE_INACTIVE: |
2302 | update_event_times(event); | |
bfbd3381 PZ |
2303 | break; |
2304 | ||
2305 | default: | |
2306 | break; | |
2307 | } | |
2308 | ||
2309 | /* | |
cdd6c482 | 2310 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2311 | * values when we flip the contexts. |
2312 | */ | |
e7850595 PZ |
2313 | value = local64_read(&next_event->count); |
2314 | value = local64_xchg(&event->count, value); | |
2315 | local64_set(&next_event->count, value); | |
bfbd3381 | 2316 | |
cdd6c482 IM |
2317 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2318 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2319 | |
bfbd3381 | 2320 | /* |
19d2e755 | 2321 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2322 | */ |
cdd6c482 IM |
2323 | perf_event_update_userpage(event); |
2324 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2325 | } |
2326 | ||
cdd6c482 IM |
2327 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2328 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2329 | { |
cdd6c482 | 2330 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2331 | |
2332 | if (!ctx->nr_stat) | |
2333 | return; | |
2334 | ||
02ffdbc8 PZ |
2335 | update_context_time(ctx); |
2336 | ||
cdd6c482 IM |
2337 | event = list_first_entry(&ctx->event_list, |
2338 | struct perf_event, event_entry); | |
bfbd3381 | 2339 | |
cdd6c482 IM |
2340 | next_event = list_first_entry(&next_ctx->event_list, |
2341 | struct perf_event, event_entry); | |
bfbd3381 | 2342 | |
cdd6c482 IM |
2343 | while (&event->event_entry != &ctx->event_list && |
2344 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2345 | |
cdd6c482 | 2346 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2347 | |
cdd6c482 IM |
2348 | event = list_next_entry(event, event_entry); |
2349 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2350 | } |
2351 | } | |
2352 | ||
fe4b04fa PZ |
2353 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2354 | struct task_struct *next) | |
0793a61d | 2355 | { |
8dc85d54 | 2356 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2357 | struct perf_event_context *next_ctx; |
5a3126d4 | 2358 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2359 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2360 | int do_switch = 1; |
0793a61d | 2361 | |
108b02cf PZ |
2362 | if (likely(!ctx)) |
2363 | return; | |
10989fb2 | 2364 | |
108b02cf PZ |
2365 | cpuctx = __get_cpu_context(ctx); |
2366 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2367 | return; |
2368 | ||
c93f7669 | 2369 | rcu_read_lock(); |
8dc85d54 | 2370 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2371 | if (!next_ctx) |
2372 | goto unlock; | |
2373 | ||
2374 | parent = rcu_dereference(ctx->parent_ctx); | |
2375 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2376 | ||
2377 | /* If neither context have a parent context; they cannot be clones. */ | |
1f9a7268 | 2378 | if (!parent || !next_parent) |
5a3126d4 PZ |
2379 | goto unlock; |
2380 | ||
2381 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2382 | /* |
2383 | * Looks like the two contexts are clones, so we might be | |
2384 | * able to optimize the context switch. We lock both | |
2385 | * contexts and check that they are clones under the | |
2386 | * lock (including re-checking that neither has been | |
2387 | * uncloned in the meantime). It doesn't matter which | |
2388 | * order we take the locks because no other cpu could | |
2389 | * be trying to lock both of these tasks. | |
2390 | */ | |
e625cce1 TG |
2391 | raw_spin_lock(&ctx->lock); |
2392 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2393 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2394 | /* |
2395 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2396 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2397 | */ |
8dc85d54 PZ |
2398 | task->perf_event_ctxp[ctxn] = next_ctx; |
2399 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2400 | ctx->task = next; |
2401 | next_ctx->task = task; | |
2402 | do_switch = 0; | |
bfbd3381 | 2403 | |
cdd6c482 | 2404 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2405 | } |
e625cce1 TG |
2406 | raw_spin_unlock(&next_ctx->lock); |
2407 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2408 | } |
5a3126d4 | 2409 | unlock: |
c93f7669 | 2410 | rcu_read_unlock(); |
564c2b21 | 2411 | |
c93f7669 | 2412 | if (do_switch) { |
facc4307 | 2413 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2414 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2415 | cpuctx->task_ctx = NULL; |
facc4307 | 2416 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2417 | } |
0793a61d TG |
2418 | } |
2419 | ||
8dc85d54 PZ |
2420 | #define for_each_task_context_nr(ctxn) \ |
2421 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2422 | ||
2423 | /* | |
2424 | * Called from scheduler to remove the events of the current task, | |
2425 | * with interrupts disabled. | |
2426 | * | |
2427 | * We stop each event and update the event value in event->count. | |
2428 | * | |
2429 | * This does not protect us against NMI, but disable() | |
2430 | * sets the disabled bit in the control field of event _before_ | |
2431 | * accessing the event control register. If a NMI hits, then it will | |
2432 | * not restart the event. | |
2433 | */ | |
ab0cce56 JO |
2434 | void __perf_event_task_sched_out(struct task_struct *task, |
2435 | struct task_struct *next) | |
8dc85d54 PZ |
2436 | { |
2437 | int ctxn; | |
2438 | ||
8dc85d54 PZ |
2439 | for_each_task_context_nr(ctxn) |
2440 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2441 | |
2442 | /* | |
2443 | * if cgroup events exist on this CPU, then we need | |
2444 | * to check if we have to switch out PMU state. | |
2445 | * cgroup event are system-wide mode only | |
2446 | */ | |
2447 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2448 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2449 | } |
2450 | ||
04dc2dbb | 2451 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2452 | { |
108b02cf | 2453 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2454 | |
a63eaf34 PM |
2455 | if (!cpuctx->task_ctx) |
2456 | return; | |
012b84da IM |
2457 | |
2458 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2459 | return; | |
2460 | ||
04dc2dbb | 2461 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2462 | cpuctx->task_ctx = NULL; |
2463 | } | |
2464 | ||
5b0311e1 FW |
2465 | /* |
2466 | * Called with IRQs disabled | |
2467 | */ | |
2468 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2469 | enum event_type_t event_type) | |
2470 | { | |
2471 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2472 | } |
2473 | ||
235c7fc7 | 2474 | static void |
5b0311e1 | 2475 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2476 | struct perf_cpu_context *cpuctx) |
0793a61d | 2477 | { |
cdd6c482 | 2478 | struct perf_event *event; |
0793a61d | 2479 | |
889ff015 FW |
2480 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2481 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2482 | continue; |
5632ab12 | 2483 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2484 | continue; |
2485 | ||
e5d1367f SE |
2486 | /* may need to reset tstamp_enabled */ |
2487 | if (is_cgroup_event(event)) | |
2488 | perf_cgroup_mark_enabled(event, ctx); | |
2489 | ||
8c9ed8e1 | 2490 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2491 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2492 | |
2493 | /* | |
2494 | * If this pinned group hasn't been scheduled, | |
2495 | * put it in error state. | |
2496 | */ | |
cdd6c482 IM |
2497 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2498 | update_group_times(event); | |
2499 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2500 | } |
3b6f9e5c | 2501 | } |
5b0311e1 FW |
2502 | } |
2503 | ||
2504 | static void | |
2505 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2506 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2507 | { |
2508 | struct perf_event *event; | |
2509 | int can_add_hw = 1; | |
3b6f9e5c | 2510 | |
889ff015 FW |
2511 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2512 | /* Ignore events in OFF or ERROR state */ | |
2513 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2514 | continue; |
04289bb9 IM |
2515 | /* |
2516 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2517 | * of events: |
04289bb9 | 2518 | */ |
5632ab12 | 2519 | if (!event_filter_match(event)) |
0793a61d TG |
2520 | continue; |
2521 | ||
e5d1367f SE |
2522 | /* may need to reset tstamp_enabled */ |
2523 | if (is_cgroup_event(event)) | |
2524 | perf_cgroup_mark_enabled(event, ctx); | |
2525 | ||
9ed6060d | 2526 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2527 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2528 | can_add_hw = 0; |
9ed6060d | 2529 | } |
0793a61d | 2530 | } |
5b0311e1 FW |
2531 | } |
2532 | ||
2533 | static void | |
2534 | ctx_sched_in(struct perf_event_context *ctx, | |
2535 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2536 | enum event_type_t event_type, |
2537 | struct task_struct *task) | |
5b0311e1 | 2538 | { |
e5d1367f | 2539 | u64 now; |
db24d33e | 2540 | int is_active = ctx->is_active; |
e5d1367f | 2541 | |
db24d33e | 2542 | ctx->is_active |= event_type; |
5b0311e1 | 2543 | if (likely(!ctx->nr_events)) |
facc4307 | 2544 | return; |
5b0311e1 | 2545 | |
e5d1367f SE |
2546 | now = perf_clock(); |
2547 | ctx->timestamp = now; | |
3f7cce3c | 2548 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2549 | /* |
2550 | * First go through the list and put on any pinned groups | |
2551 | * in order to give them the best chance of going on. | |
2552 | */ | |
db24d33e | 2553 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2554 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2555 | |
2556 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2557 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2558 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2559 | } |
2560 | ||
329c0e01 | 2561 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2562 | enum event_type_t event_type, |
2563 | struct task_struct *task) | |
329c0e01 FW |
2564 | { |
2565 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2566 | ||
e5d1367f | 2567 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2568 | } |
2569 | ||
e5d1367f SE |
2570 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2571 | struct task_struct *task) | |
235c7fc7 | 2572 | { |
108b02cf | 2573 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2574 | |
108b02cf | 2575 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2576 | if (cpuctx->task_ctx == ctx) |
2577 | return; | |
2578 | ||
facc4307 | 2579 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2580 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2581 | /* |
2582 | * We want to keep the following priority order: | |
2583 | * cpu pinned (that don't need to move), task pinned, | |
2584 | * cpu flexible, task flexible. | |
2585 | */ | |
2586 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2587 | ||
1d5f003f GN |
2588 | if (ctx->nr_events) |
2589 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2590 | |
86b47c25 GN |
2591 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2592 | ||
facc4307 PZ |
2593 | perf_pmu_enable(ctx->pmu); |
2594 | perf_ctx_unlock(cpuctx, ctx); | |
2595 | ||
b5ab4cd5 PZ |
2596 | /* |
2597 | * Since these rotations are per-cpu, we need to ensure the | |
2598 | * cpu-context we got scheduled on is actually rotating. | |
2599 | */ | |
108b02cf | 2600 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2601 | } |
2602 | ||
d010b332 SE |
2603 | /* |
2604 | * When sampling the branck stack in system-wide, it may be necessary | |
2605 | * to flush the stack on context switch. This happens when the branch | |
2606 | * stack does not tag its entries with the pid of the current task. | |
2607 | * Otherwise it becomes impossible to associate a branch entry with a | |
2608 | * task. This ambiguity is more likely to appear when the branch stack | |
2609 | * supports priv level filtering and the user sets it to monitor only | |
2610 | * at the user level (which could be a useful measurement in system-wide | |
2611 | * mode). In that case, the risk is high of having a branch stack with | |
2612 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2613 | * entries or stashing them somewhere in the PMU specific code layer. | |
2614 | * | |
2615 | * This function provides the context switch callback to the lower code | |
2616 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2617 | * with at least one active event using taken branch sampling. | |
2618 | */ | |
2619 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2620 | struct task_struct *task) | |
2621 | { | |
2622 | struct perf_cpu_context *cpuctx; | |
2623 | struct pmu *pmu; | |
2624 | unsigned long flags; | |
2625 | ||
2626 | /* no need to flush branch stack if not changing task */ | |
2627 | if (prev == task) | |
2628 | return; | |
2629 | ||
2630 | local_irq_save(flags); | |
2631 | ||
2632 | rcu_read_lock(); | |
2633 | ||
2634 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2635 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2636 | ||
2637 | /* | |
2638 | * check if the context has at least one | |
2639 | * event using PERF_SAMPLE_BRANCH_STACK | |
2640 | */ | |
2641 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2642 | && pmu->flush_branch_stack) { | |
2643 | ||
d010b332 SE |
2644 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
2645 | ||
2646 | perf_pmu_disable(pmu); | |
2647 | ||
2648 | pmu->flush_branch_stack(); | |
2649 | ||
2650 | perf_pmu_enable(pmu); | |
2651 | ||
2652 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2653 | } | |
2654 | } | |
2655 | ||
2656 | rcu_read_unlock(); | |
2657 | ||
2658 | local_irq_restore(flags); | |
2659 | } | |
2660 | ||
8dc85d54 PZ |
2661 | /* |
2662 | * Called from scheduler to add the events of the current task | |
2663 | * with interrupts disabled. | |
2664 | * | |
2665 | * We restore the event value and then enable it. | |
2666 | * | |
2667 | * This does not protect us against NMI, but enable() | |
2668 | * sets the enabled bit in the control field of event _before_ | |
2669 | * accessing the event control register. If a NMI hits, then it will | |
2670 | * keep the event running. | |
2671 | */ | |
ab0cce56 JO |
2672 | void __perf_event_task_sched_in(struct task_struct *prev, |
2673 | struct task_struct *task) | |
8dc85d54 PZ |
2674 | { |
2675 | struct perf_event_context *ctx; | |
2676 | int ctxn; | |
2677 | ||
2678 | for_each_task_context_nr(ctxn) { | |
2679 | ctx = task->perf_event_ctxp[ctxn]; | |
2680 | if (likely(!ctx)) | |
2681 | continue; | |
2682 | ||
e5d1367f | 2683 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2684 | } |
e5d1367f SE |
2685 | /* |
2686 | * if cgroup events exist on this CPU, then we need | |
2687 | * to check if we have to switch in PMU state. | |
2688 | * cgroup event are system-wide mode only | |
2689 | */ | |
2690 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2691 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2692 | |
2693 | /* check for system-wide branch_stack events */ | |
2694 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2695 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2696 | } |
2697 | ||
abd50713 PZ |
2698 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2699 | { | |
2700 | u64 frequency = event->attr.sample_freq; | |
2701 | u64 sec = NSEC_PER_SEC; | |
2702 | u64 divisor, dividend; | |
2703 | ||
2704 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2705 | ||
2706 | count_fls = fls64(count); | |
2707 | nsec_fls = fls64(nsec); | |
2708 | frequency_fls = fls64(frequency); | |
2709 | sec_fls = 30; | |
2710 | ||
2711 | /* | |
2712 | * We got @count in @nsec, with a target of sample_freq HZ | |
2713 | * the target period becomes: | |
2714 | * | |
2715 | * @count * 10^9 | |
2716 | * period = ------------------- | |
2717 | * @nsec * sample_freq | |
2718 | * | |
2719 | */ | |
2720 | ||
2721 | /* | |
2722 | * Reduce accuracy by one bit such that @a and @b converge | |
2723 | * to a similar magnitude. | |
2724 | */ | |
fe4b04fa | 2725 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2726 | do { \ |
2727 | if (a##_fls > b##_fls) { \ | |
2728 | a >>= 1; \ | |
2729 | a##_fls--; \ | |
2730 | } else { \ | |
2731 | b >>= 1; \ | |
2732 | b##_fls--; \ | |
2733 | } \ | |
2734 | } while (0) | |
2735 | ||
2736 | /* | |
2737 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2738 | * the other, so that finally we can do a u64/u64 division. | |
2739 | */ | |
2740 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2741 | REDUCE_FLS(nsec, frequency); | |
2742 | REDUCE_FLS(sec, count); | |
2743 | } | |
2744 | ||
2745 | if (count_fls + sec_fls > 64) { | |
2746 | divisor = nsec * frequency; | |
2747 | ||
2748 | while (count_fls + sec_fls > 64) { | |
2749 | REDUCE_FLS(count, sec); | |
2750 | divisor >>= 1; | |
2751 | } | |
2752 | ||
2753 | dividend = count * sec; | |
2754 | } else { | |
2755 | dividend = count * sec; | |
2756 | ||
2757 | while (nsec_fls + frequency_fls > 64) { | |
2758 | REDUCE_FLS(nsec, frequency); | |
2759 | dividend >>= 1; | |
2760 | } | |
2761 | ||
2762 | divisor = nsec * frequency; | |
2763 | } | |
2764 | ||
f6ab91ad PZ |
2765 | if (!divisor) |
2766 | return dividend; | |
2767 | ||
abd50713 PZ |
2768 | return div64_u64(dividend, divisor); |
2769 | } | |
2770 | ||
e050e3f0 SE |
2771 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2772 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2773 | ||
f39d47ff | 2774 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2775 | { |
cdd6c482 | 2776 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2777 | s64 period, sample_period; |
bd2b5b12 PZ |
2778 | s64 delta; |
2779 | ||
abd50713 | 2780 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2781 | |
2782 | delta = (s64)(period - hwc->sample_period); | |
2783 | delta = (delta + 7) / 8; /* low pass filter */ | |
2784 | ||
2785 | sample_period = hwc->sample_period + delta; | |
2786 | ||
2787 | if (!sample_period) | |
2788 | sample_period = 1; | |
2789 | ||
bd2b5b12 | 2790 | hwc->sample_period = sample_period; |
abd50713 | 2791 | |
e7850595 | 2792 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2793 | if (disable) |
2794 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2795 | ||
e7850595 | 2796 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2797 | |
2798 | if (disable) | |
2799 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2800 | } |
bd2b5b12 PZ |
2801 | } |
2802 | ||
e050e3f0 SE |
2803 | /* |
2804 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2805 | * events. At the same time, make sure, having freq events does not change | |
2806 | * the rate of unthrottling as that would introduce bias. | |
2807 | */ | |
2808 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2809 | int needs_unthr) | |
60db5e09 | 2810 | { |
cdd6c482 IM |
2811 | struct perf_event *event; |
2812 | struct hw_perf_event *hwc; | |
e050e3f0 | 2813 | u64 now, period = TICK_NSEC; |
abd50713 | 2814 | s64 delta; |
60db5e09 | 2815 | |
e050e3f0 SE |
2816 | /* |
2817 | * only need to iterate over all events iff: | |
2818 | * - context have events in frequency mode (needs freq adjust) | |
2819 | * - there are events to unthrottle on this cpu | |
2820 | */ | |
2821 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2822 | return; |
2823 | ||
e050e3f0 | 2824 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2825 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2826 | |
03541f8b | 2827 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2828 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2829 | continue; |
2830 | ||
5632ab12 | 2831 | if (!event_filter_match(event)) |
5d27c23d PZ |
2832 | continue; |
2833 | ||
44377277 AS |
2834 | perf_pmu_disable(event->pmu); |
2835 | ||
cdd6c482 | 2836 | hwc = &event->hw; |
6a24ed6c | 2837 | |
ae23bff1 | 2838 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2839 | hwc->interrupts = 0; |
cdd6c482 | 2840 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2841 | event->pmu->start(event, 0); |
a78ac325 PZ |
2842 | } |
2843 | ||
cdd6c482 | 2844 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2845 | goto next; |
60db5e09 | 2846 | |
e050e3f0 SE |
2847 | /* |
2848 | * stop the event and update event->count | |
2849 | */ | |
2850 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2851 | ||
e7850595 | 2852 | now = local64_read(&event->count); |
abd50713 PZ |
2853 | delta = now - hwc->freq_count_stamp; |
2854 | hwc->freq_count_stamp = now; | |
60db5e09 | 2855 | |
e050e3f0 SE |
2856 | /* |
2857 | * restart the event | |
2858 | * reload only if value has changed | |
f39d47ff SE |
2859 | * we have stopped the event so tell that |
2860 | * to perf_adjust_period() to avoid stopping it | |
2861 | * twice. | |
e050e3f0 | 2862 | */ |
abd50713 | 2863 | if (delta > 0) |
f39d47ff | 2864 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2865 | |
2866 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
2867 | next: |
2868 | perf_pmu_enable(event->pmu); | |
60db5e09 | 2869 | } |
e050e3f0 | 2870 | |
f39d47ff | 2871 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2872 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2873 | } |
2874 | ||
235c7fc7 | 2875 | /* |
cdd6c482 | 2876 | * Round-robin a context's events: |
235c7fc7 | 2877 | */ |
cdd6c482 | 2878 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2879 | { |
dddd3379 TG |
2880 | /* |
2881 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2882 | * disabled by the inheritance code. | |
2883 | */ | |
2884 | if (!ctx->rotate_disable) | |
2885 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2886 | } |
2887 | ||
b5ab4cd5 | 2888 | /* |
e9d2b064 PZ |
2889 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2890 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2891 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2892 | */ |
9e630205 | 2893 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2894 | { |
8dc85d54 | 2895 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2896 | int rotate = 0, remove = 1; |
7fc23a53 | 2897 | |
b5ab4cd5 | 2898 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2899 | remove = 0; |
b5ab4cd5 PZ |
2900 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2901 | rotate = 1; | |
2902 | } | |
235c7fc7 | 2903 | |
8dc85d54 | 2904 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2905 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2906 | remove = 0; |
b5ab4cd5 PZ |
2907 | if (ctx->nr_events != ctx->nr_active) |
2908 | rotate = 1; | |
2909 | } | |
9717e6cd | 2910 | |
e050e3f0 | 2911 | if (!rotate) |
0f5a2601 PZ |
2912 | goto done; |
2913 | ||
facc4307 | 2914 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2915 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2916 | |
e050e3f0 SE |
2917 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2918 | if (ctx) | |
2919 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2920 | |
e050e3f0 SE |
2921 | rotate_ctx(&cpuctx->ctx); |
2922 | if (ctx) | |
2923 | rotate_ctx(ctx); | |
235c7fc7 | 2924 | |
e050e3f0 | 2925 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2926 | |
0f5a2601 PZ |
2927 | perf_pmu_enable(cpuctx->ctx.pmu); |
2928 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2929 | done: |
e9d2b064 PZ |
2930 | if (remove) |
2931 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2932 | |
2933 | return rotate; | |
e9d2b064 PZ |
2934 | } |
2935 | ||
026249ef FW |
2936 | #ifdef CONFIG_NO_HZ_FULL |
2937 | bool perf_event_can_stop_tick(void) | |
2938 | { | |
948b26b6 | 2939 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 2940 | __this_cpu_read(perf_throttled_count)) |
026249ef | 2941 | return false; |
d84153d6 FW |
2942 | else |
2943 | return true; | |
026249ef FW |
2944 | } |
2945 | #endif | |
2946 | ||
e9d2b064 PZ |
2947 | void perf_event_task_tick(void) |
2948 | { | |
2949 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2950 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2951 | struct perf_event_context *ctx; |
2952 | int throttled; | |
b5ab4cd5 | 2953 | |
e9d2b064 PZ |
2954 | WARN_ON(!irqs_disabled()); |
2955 | ||
e050e3f0 SE |
2956 | __this_cpu_inc(perf_throttled_seq); |
2957 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2958 | ||
e9d2b064 | 2959 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2960 | ctx = &cpuctx->ctx; |
2961 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2962 | ||
2963 | ctx = cpuctx->task_ctx; | |
2964 | if (ctx) | |
2965 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2966 | } |
0793a61d TG |
2967 | } |
2968 | ||
889ff015 FW |
2969 | static int event_enable_on_exec(struct perf_event *event, |
2970 | struct perf_event_context *ctx) | |
2971 | { | |
2972 | if (!event->attr.enable_on_exec) | |
2973 | return 0; | |
2974 | ||
2975 | event->attr.enable_on_exec = 0; | |
2976 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2977 | return 0; | |
2978 | ||
1d9b482e | 2979 | __perf_event_mark_enabled(event); |
889ff015 FW |
2980 | |
2981 | return 1; | |
2982 | } | |
2983 | ||
57e7986e | 2984 | /* |
cdd6c482 | 2985 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2986 | * This expects task == current. |
2987 | */ | |
8dc85d54 | 2988 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2989 | { |
cdd6c482 | 2990 | struct perf_event *event; |
57e7986e PM |
2991 | unsigned long flags; |
2992 | int enabled = 0; | |
889ff015 | 2993 | int ret; |
57e7986e PM |
2994 | |
2995 | local_irq_save(flags); | |
cdd6c482 | 2996 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2997 | goto out; |
2998 | ||
e566b76e SE |
2999 | /* |
3000 | * We must ctxsw out cgroup events to avoid conflict | |
3001 | * when invoking perf_task_event_sched_in() later on | |
3002 | * in this function. Otherwise we end up trying to | |
3003 | * ctxswin cgroup events which are already scheduled | |
3004 | * in. | |
3005 | */ | |
a8d757ef | 3006 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 3007 | |
e625cce1 | 3008 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 3009 | task_ctx_sched_out(ctx); |
57e7986e | 3010 | |
b79387ef | 3011 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
3012 | ret = event_enable_on_exec(event, ctx); |
3013 | if (ret) | |
3014 | enabled = 1; | |
57e7986e PM |
3015 | } |
3016 | ||
3017 | /* | |
cdd6c482 | 3018 | * Unclone this context if we enabled any event. |
57e7986e | 3019 | */ |
71a851b4 PZ |
3020 | if (enabled) |
3021 | unclone_ctx(ctx); | |
57e7986e | 3022 | |
e625cce1 | 3023 | raw_spin_unlock(&ctx->lock); |
57e7986e | 3024 | |
e566b76e SE |
3025 | /* |
3026 | * Also calls ctxswin for cgroup events, if any: | |
3027 | */ | |
e5d1367f | 3028 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 3029 | out: |
57e7986e PM |
3030 | local_irq_restore(flags); |
3031 | } | |
3032 | ||
e041e328 PZ |
3033 | void perf_event_exec(void) |
3034 | { | |
3035 | struct perf_event_context *ctx; | |
3036 | int ctxn; | |
3037 | ||
3038 | rcu_read_lock(); | |
3039 | for_each_task_context_nr(ctxn) { | |
3040 | ctx = current->perf_event_ctxp[ctxn]; | |
3041 | if (!ctx) | |
3042 | continue; | |
3043 | ||
3044 | perf_event_enable_on_exec(ctx); | |
3045 | } | |
3046 | rcu_read_unlock(); | |
3047 | } | |
3048 | ||
0793a61d | 3049 | /* |
cdd6c482 | 3050 | * Cross CPU call to read the hardware event |
0793a61d | 3051 | */ |
cdd6c482 | 3052 | static void __perf_event_read(void *info) |
0793a61d | 3053 | { |
cdd6c482 IM |
3054 | struct perf_event *event = info; |
3055 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 3056 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3057 | |
e1ac3614 PM |
3058 | /* |
3059 | * If this is a task context, we need to check whether it is | |
3060 | * the current task context of this cpu. If not it has been | |
3061 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3062 | * event->count would have been updated to a recent sample |
3063 | * when the event was scheduled out. | |
e1ac3614 PM |
3064 | */ |
3065 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3066 | return; | |
3067 | ||
e625cce1 | 3068 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3069 | if (ctx->is_active) { |
542e72fc | 3070 | update_context_time(ctx); |
e5d1367f SE |
3071 | update_cgrp_time_from_event(event); |
3072 | } | |
cdd6c482 | 3073 | update_event_times(event); |
542e72fc PZ |
3074 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3075 | event->pmu->read(event); | |
e625cce1 | 3076 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3077 | } |
3078 | ||
b5e58793 PZ |
3079 | static inline u64 perf_event_count(struct perf_event *event) |
3080 | { | |
e7850595 | 3081 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
3082 | } |
3083 | ||
cdd6c482 | 3084 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3085 | { |
3086 | /* | |
cdd6c482 IM |
3087 | * If event is enabled and currently active on a CPU, update the |
3088 | * value in the event structure: | |
0793a61d | 3089 | */ |
cdd6c482 IM |
3090 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3091 | smp_call_function_single(event->oncpu, | |
3092 | __perf_event_read, event, 1); | |
3093 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3094 | struct perf_event_context *ctx = event->ctx; |
3095 | unsigned long flags; | |
3096 | ||
e625cce1 | 3097 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3098 | /* |
3099 | * may read while context is not active | |
3100 | * (e.g., thread is blocked), in that case | |
3101 | * we cannot update context time | |
3102 | */ | |
e5d1367f | 3103 | if (ctx->is_active) { |
c530ccd9 | 3104 | update_context_time(ctx); |
e5d1367f SE |
3105 | update_cgrp_time_from_event(event); |
3106 | } | |
cdd6c482 | 3107 | update_event_times(event); |
e625cce1 | 3108 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3109 | } |
3110 | ||
b5e58793 | 3111 | return perf_event_count(event); |
0793a61d TG |
3112 | } |
3113 | ||
a63eaf34 | 3114 | /* |
cdd6c482 | 3115 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3116 | */ |
eb184479 | 3117 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3118 | { |
e625cce1 | 3119 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3120 | mutex_init(&ctx->mutex); |
889ff015 FW |
3121 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3122 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3123 | INIT_LIST_HEAD(&ctx->event_list); |
3124 | atomic_set(&ctx->refcount, 1); | |
fadfe7be | 3125 | INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work); |
eb184479 PZ |
3126 | } |
3127 | ||
3128 | static struct perf_event_context * | |
3129 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3130 | { | |
3131 | struct perf_event_context *ctx; | |
3132 | ||
3133 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3134 | if (!ctx) | |
3135 | return NULL; | |
3136 | ||
3137 | __perf_event_init_context(ctx); | |
3138 | if (task) { | |
3139 | ctx->task = task; | |
3140 | get_task_struct(task); | |
0793a61d | 3141 | } |
eb184479 PZ |
3142 | ctx->pmu = pmu; |
3143 | ||
3144 | return ctx; | |
a63eaf34 PM |
3145 | } |
3146 | ||
2ebd4ffb MH |
3147 | static struct task_struct * |
3148 | find_lively_task_by_vpid(pid_t vpid) | |
3149 | { | |
3150 | struct task_struct *task; | |
3151 | int err; | |
0793a61d TG |
3152 | |
3153 | rcu_read_lock(); | |
2ebd4ffb | 3154 | if (!vpid) |
0793a61d TG |
3155 | task = current; |
3156 | else | |
2ebd4ffb | 3157 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3158 | if (task) |
3159 | get_task_struct(task); | |
3160 | rcu_read_unlock(); | |
3161 | ||
3162 | if (!task) | |
3163 | return ERR_PTR(-ESRCH); | |
3164 | ||
0793a61d | 3165 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3166 | err = -EACCES; |
3167 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3168 | goto errout; | |
3169 | ||
2ebd4ffb MH |
3170 | return task; |
3171 | errout: | |
3172 | put_task_struct(task); | |
3173 | return ERR_PTR(err); | |
3174 | ||
3175 | } | |
3176 | ||
fe4b04fa PZ |
3177 | /* |
3178 | * Returns a matching context with refcount and pincount. | |
3179 | */ | |
108b02cf | 3180 | static struct perf_event_context * |
38a81da2 | 3181 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3182 | { |
cdd6c482 | 3183 | struct perf_event_context *ctx; |
22a4f650 | 3184 | struct perf_cpu_context *cpuctx; |
25346b93 | 3185 | unsigned long flags; |
8dc85d54 | 3186 | int ctxn, err; |
0793a61d | 3187 | |
22a4ec72 | 3188 | if (!task) { |
cdd6c482 | 3189 | /* Must be root to operate on a CPU event: */ |
0764771d | 3190 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3191 | return ERR_PTR(-EACCES); |
3192 | ||
0793a61d | 3193 | /* |
cdd6c482 | 3194 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3195 | * offline CPU and activate it when the CPU comes up, but |
3196 | * that's for later. | |
3197 | */ | |
f6325e30 | 3198 | if (!cpu_online(cpu)) |
0793a61d TG |
3199 | return ERR_PTR(-ENODEV); |
3200 | ||
108b02cf | 3201 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3202 | ctx = &cpuctx->ctx; |
c93f7669 | 3203 | get_ctx(ctx); |
fe4b04fa | 3204 | ++ctx->pin_count; |
0793a61d | 3205 | |
0793a61d TG |
3206 | return ctx; |
3207 | } | |
3208 | ||
8dc85d54 PZ |
3209 | err = -EINVAL; |
3210 | ctxn = pmu->task_ctx_nr; | |
3211 | if (ctxn < 0) | |
3212 | goto errout; | |
3213 | ||
9ed6060d | 3214 | retry: |
8dc85d54 | 3215 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3216 | if (ctx) { |
71a851b4 | 3217 | unclone_ctx(ctx); |
fe4b04fa | 3218 | ++ctx->pin_count; |
e625cce1 | 3219 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3220 | } else { |
eb184479 | 3221 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3222 | err = -ENOMEM; |
3223 | if (!ctx) | |
3224 | goto errout; | |
eb184479 | 3225 | |
dbe08d82 ON |
3226 | err = 0; |
3227 | mutex_lock(&task->perf_event_mutex); | |
3228 | /* | |
3229 | * If it has already passed perf_event_exit_task(). | |
3230 | * we must see PF_EXITING, it takes this mutex too. | |
3231 | */ | |
3232 | if (task->flags & PF_EXITING) | |
3233 | err = -ESRCH; | |
3234 | else if (task->perf_event_ctxp[ctxn]) | |
3235 | err = -EAGAIN; | |
fe4b04fa | 3236 | else { |
9137fb28 | 3237 | get_ctx(ctx); |
fe4b04fa | 3238 | ++ctx->pin_count; |
dbe08d82 | 3239 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3240 | } |
dbe08d82 ON |
3241 | mutex_unlock(&task->perf_event_mutex); |
3242 | ||
3243 | if (unlikely(err)) { | |
9137fb28 | 3244 | put_ctx(ctx); |
dbe08d82 ON |
3245 | |
3246 | if (err == -EAGAIN) | |
3247 | goto retry; | |
3248 | goto errout; | |
a63eaf34 PM |
3249 | } |
3250 | } | |
3251 | ||
0793a61d | 3252 | return ctx; |
c93f7669 | 3253 | |
9ed6060d | 3254 | errout: |
c93f7669 | 3255 | return ERR_PTR(err); |
0793a61d TG |
3256 | } |
3257 | ||
6fb2915d LZ |
3258 | static void perf_event_free_filter(struct perf_event *event); |
3259 | ||
cdd6c482 | 3260 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3261 | { |
cdd6c482 | 3262 | struct perf_event *event; |
592903cd | 3263 | |
cdd6c482 IM |
3264 | event = container_of(head, struct perf_event, rcu_head); |
3265 | if (event->ns) | |
3266 | put_pid_ns(event->ns); | |
6fb2915d | 3267 | perf_event_free_filter(event); |
cdd6c482 | 3268 | kfree(event); |
592903cd PZ |
3269 | } |
3270 | ||
76369139 | 3271 | static void ring_buffer_put(struct ring_buffer *rb); |
b69cf536 PZ |
3272 | static void ring_buffer_attach(struct perf_event *event, |
3273 | struct ring_buffer *rb); | |
925d519a | 3274 | |
4beb31f3 | 3275 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3276 | { |
4beb31f3 FW |
3277 | if (event->parent) |
3278 | return; | |
3279 | ||
3280 | if (has_branch_stack(event)) { | |
3281 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3282 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3283 | } | |
3284 | if (is_cgroup_event(event)) | |
3285 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3286 | } | |
925d519a | 3287 | |
4beb31f3 FW |
3288 | static void unaccount_event(struct perf_event *event) |
3289 | { | |
3290 | if (event->parent) | |
3291 | return; | |
3292 | ||
3293 | if (event->attach_state & PERF_ATTACH_TASK) | |
3294 | static_key_slow_dec_deferred(&perf_sched_events); | |
3295 | if (event->attr.mmap || event->attr.mmap_data) | |
3296 | atomic_dec(&nr_mmap_events); | |
3297 | if (event->attr.comm) | |
3298 | atomic_dec(&nr_comm_events); | |
3299 | if (event->attr.task) | |
3300 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3301 | if (event->attr.freq) |
3302 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3303 | if (is_cgroup_event(event)) |
3304 | static_key_slow_dec_deferred(&perf_sched_events); | |
3305 | if (has_branch_stack(event)) | |
3306 | static_key_slow_dec_deferred(&perf_sched_events); | |
3307 | ||
3308 | unaccount_event_cpu(event, event->cpu); | |
3309 | } | |
925d519a | 3310 | |
766d6c07 FW |
3311 | static void __free_event(struct perf_event *event) |
3312 | { | |
cdd6c482 | 3313 | if (!event->parent) { |
927c7a9e FW |
3314 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3315 | put_callchain_buffers(); | |
f344011c | 3316 | } |
9ee318a7 | 3317 | |
766d6c07 FW |
3318 | if (event->destroy) |
3319 | event->destroy(event); | |
3320 | ||
3321 | if (event->ctx) | |
3322 | put_ctx(event->ctx); | |
3323 | ||
c464c76e YZ |
3324 | if (event->pmu) |
3325 | module_put(event->pmu->module); | |
3326 | ||
766d6c07 FW |
3327 | call_rcu(&event->rcu_head, free_event_rcu); |
3328 | } | |
683ede43 PZ |
3329 | |
3330 | static void _free_event(struct perf_event *event) | |
f1600952 | 3331 | { |
e360adbe | 3332 | irq_work_sync(&event->pending); |
925d519a | 3333 | |
4beb31f3 | 3334 | unaccount_event(event); |
9ee318a7 | 3335 | |
76369139 | 3336 | if (event->rb) { |
9bb5d40c PZ |
3337 | /* |
3338 | * Can happen when we close an event with re-directed output. | |
3339 | * | |
3340 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3341 | * over us; possibly making our ring_buffer_put() the last. | |
3342 | */ | |
3343 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3344 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3345 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3346 | } |
3347 | ||
e5d1367f SE |
3348 | if (is_cgroup_event(event)) |
3349 | perf_detach_cgroup(event); | |
3350 | ||
766d6c07 | 3351 | __free_event(event); |
f1600952 PZ |
3352 | } |
3353 | ||
683ede43 PZ |
3354 | /* |
3355 | * Used to free events which have a known refcount of 1, such as in error paths | |
3356 | * where the event isn't exposed yet and inherited events. | |
3357 | */ | |
3358 | static void free_event(struct perf_event *event) | |
0793a61d | 3359 | { |
683ede43 PZ |
3360 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3361 | "unexpected event refcount: %ld; ptr=%p\n", | |
3362 | atomic_long_read(&event->refcount), event)) { | |
3363 | /* leak to avoid use-after-free */ | |
3364 | return; | |
3365 | } | |
0793a61d | 3366 | |
683ede43 | 3367 | _free_event(event); |
0793a61d TG |
3368 | } |
3369 | ||
a66a3052 | 3370 | /* |
f8697762 | 3371 | * Remove user event from the owner task. |
a66a3052 | 3372 | */ |
f8697762 | 3373 | static void perf_remove_from_owner(struct perf_event *event) |
fb0459d7 | 3374 | { |
8882135b | 3375 | struct task_struct *owner; |
fb0459d7 | 3376 | |
8882135b PZ |
3377 | rcu_read_lock(); |
3378 | owner = ACCESS_ONCE(event->owner); | |
3379 | /* | |
3380 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3381 | * !owner it means the list deletion is complete and we can indeed | |
3382 | * free this event, otherwise we need to serialize on | |
3383 | * owner->perf_event_mutex. | |
3384 | */ | |
3385 | smp_read_barrier_depends(); | |
3386 | if (owner) { | |
3387 | /* | |
3388 | * Since delayed_put_task_struct() also drops the last | |
3389 | * task reference we can safely take a new reference | |
3390 | * while holding the rcu_read_lock(). | |
3391 | */ | |
3392 | get_task_struct(owner); | |
3393 | } | |
3394 | rcu_read_unlock(); | |
3395 | ||
3396 | if (owner) { | |
3397 | mutex_lock(&owner->perf_event_mutex); | |
3398 | /* | |
3399 | * We have to re-check the event->owner field, if it is cleared | |
3400 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3401 | * ensured they're done, and we can proceed with freeing the | |
3402 | * event. | |
3403 | */ | |
3404 | if (event->owner) | |
3405 | list_del_init(&event->owner_entry); | |
3406 | mutex_unlock(&owner->perf_event_mutex); | |
3407 | put_task_struct(owner); | |
3408 | } | |
f8697762 JO |
3409 | } |
3410 | ||
3411 | /* | |
3412 | * Called when the last reference to the file is gone. | |
3413 | */ | |
3414 | static void put_event(struct perf_event *event) | |
3415 | { | |
3416 | struct perf_event_context *ctx = event->ctx; | |
3417 | ||
3418 | if (!atomic_long_dec_and_test(&event->refcount)) | |
3419 | return; | |
3420 | ||
3421 | if (!is_kernel_event(event)) | |
3422 | perf_remove_from_owner(event); | |
8882135b | 3423 | |
683ede43 PZ |
3424 | WARN_ON_ONCE(ctx->parent_ctx); |
3425 | /* | |
3426 | * There are two ways this annotation is useful: | |
3427 | * | |
3428 | * 1) there is a lock recursion from perf_event_exit_task | |
3429 | * see the comment there. | |
3430 | * | |
3431 | * 2) there is a lock-inversion with mmap_sem through | |
3432 | * perf_event_read_group(), which takes faults while | |
3433 | * holding ctx->mutex, however this is called after | |
3434 | * the last filedesc died, so there is no possibility | |
3435 | * to trigger the AB-BA case. | |
3436 | */ | |
3437 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
3438 | perf_remove_from_context(event, true); | |
3439 | mutex_unlock(&ctx->mutex); | |
3440 | ||
3441 | _free_event(event); | |
a6fa941d AV |
3442 | } |
3443 | ||
683ede43 PZ |
3444 | int perf_event_release_kernel(struct perf_event *event) |
3445 | { | |
3446 | put_event(event); | |
3447 | return 0; | |
3448 | } | |
3449 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3450 | ||
a6fa941d AV |
3451 | static int perf_release(struct inode *inode, struct file *file) |
3452 | { | |
3453 | put_event(file->private_data); | |
3454 | return 0; | |
fb0459d7 | 3455 | } |
fb0459d7 | 3456 | |
fadfe7be JO |
3457 | /* |
3458 | * Remove all orphanes events from the context. | |
3459 | */ | |
3460 | static void orphans_remove_work(struct work_struct *work) | |
3461 | { | |
3462 | struct perf_event_context *ctx; | |
3463 | struct perf_event *event, *tmp; | |
3464 | ||
3465 | ctx = container_of(work, struct perf_event_context, | |
3466 | orphans_remove.work); | |
3467 | ||
3468 | mutex_lock(&ctx->mutex); | |
3469 | list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) { | |
3470 | struct perf_event *parent_event = event->parent; | |
3471 | ||
3472 | if (!is_orphaned_child(event)) | |
3473 | continue; | |
3474 | ||
3475 | perf_remove_from_context(event, true); | |
3476 | ||
3477 | mutex_lock(&parent_event->child_mutex); | |
3478 | list_del_init(&event->child_list); | |
3479 | mutex_unlock(&parent_event->child_mutex); | |
3480 | ||
3481 | free_event(event); | |
3482 | put_event(parent_event); | |
3483 | } | |
3484 | ||
3485 | raw_spin_lock_irq(&ctx->lock); | |
3486 | ctx->orphans_remove_sched = false; | |
3487 | raw_spin_unlock_irq(&ctx->lock); | |
3488 | mutex_unlock(&ctx->mutex); | |
3489 | ||
3490 | put_ctx(ctx); | |
3491 | } | |
3492 | ||
59ed446f | 3493 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3494 | { |
cdd6c482 | 3495 | struct perf_event *child; |
e53c0994 PZ |
3496 | u64 total = 0; |
3497 | ||
59ed446f PZ |
3498 | *enabled = 0; |
3499 | *running = 0; | |
3500 | ||
6f10581a | 3501 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3502 | total += perf_event_read(event); |
59ed446f PZ |
3503 | *enabled += event->total_time_enabled + |
3504 | atomic64_read(&event->child_total_time_enabled); | |
3505 | *running += event->total_time_running + | |
3506 | atomic64_read(&event->child_total_time_running); | |
3507 | ||
3508 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3509 | total += perf_event_read(child); |
59ed446f PZ |
3510 | *enabled += child->total_time_enabled; |
3511 | *running += child->total_time_running; | |
3512 | } | |
6f10581a | 3513 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3514 | |
3515 | return total; | |
3516 | } | |
fb0459d7 | 3517 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3518 | |
cdd6c482 | 3519 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3520 | u64 read_format, char __user *buf) |
3521 | { | |
cdd6c482 | 3522 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3523 | int n = 0, size = 0, ret = -EFAULT; |
3524 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3525 | u64 values[5]; |
59ed446f | 3526 | u64 count, enabled, running; |
abf4868b | 3527 | |
6f10581a | 3528 | mutex_lock(&ctx->mutex); |
59ed446f | 3529 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3530 | |
3531 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3532 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3533 | values[n++] = enabled; | |
3534 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3535 | values[n++] = running; | |
abf4868b PZ |
3536 | values[n++] = count; |
3537 | if (read_format & PERF_FORMAT_ID) | |
3538 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3539 | |
3540 | size = n * sizeof(u64); | |
3541 | ||
3542 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3543 | goto unlock; |
3dab77fb | 3544 | |
6f10581a | 3545 | ret = size; |
3dab77fb | 3546 | |
65abc865 | 3547 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3548 | n = 0; |
3dab77fb | 3549 | |
59ed446f | 3550 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3551 | if (read_format & PERF_FORMAT_ID) |
3552 | values[n++] = primary_event_id(sub); | |
3553 | ||
3554 | size = n * sizeof(u64); | |
3555 | ||
184d3da8 | 3556 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3557 | ret = -EFAULT; |
3558 | goto unlock; | |
3559 | } | |
abf4868b PZ |
3560 | |
3561 | ret += size; | |
3dab77fb | 3562 | } |
6f10581a PZ |
3563 | unlock: |
3564 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3565 | |
abf4868b | 3566 | return ret; |
3dab77fb PZ |
3567 | } |
3568 | ||
cdd6c482 | 3569 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3570 | u64 read_format, char __user *buf) |
3571 | { | |
59ed446f | 3572 | u64 enabled, running; |
3dab77fb PZ |
3573 | u64 values[4]; |
3574 | int n = 0; | |
3575 | ||
59ed446f PZ |
3576 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3577 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3578 | values[n++] = enabled; | |
3579 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3580 | values[n++] = running; | |
3dab77fb | 3581 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3582 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3583 | |
3584 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3585 | return -EFAULT; | |
3586 | ||
3587 | return n * sizeof(u64); | |
3588 | } | |
3589 | ||
dc633982 JO |
3590 | static bool is_event_hup(struct perf_event *event) |
3591 | { | |
3592 | bool no_children; | |
3593 | ||
3594 | if (event->state != PERF_EVENT_STATE_EXIT) | |
3595 | return false; | |
3596 | ||
3597 | mutex_lock(&event->child_mutex); | |
3598 | no_children = list_empty(&event->child_list); | |
3599 | mutex_unlock(&event->child_mutex); | |
3600 | return no_children; | |
3601 | } | |
3602 | ||
0793a61d | 3603 | /* |
cdd6c482 | 3604 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3605 | */ |
3606 | static ssize_t | |
cdd6c482 | 3607 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3608 | { |
cdd6c482 | 3609 | u64 read_format = event->attr.read_format; |
3dab77fb | 3610 | int ret; |
0793a61d | 3611 | |
3b6f9e5c | 3612 | /* |
cdd6c482 | 3613 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3614 | * error state (i.e. because it was pinned but it couldn't be |
3615 | * scheduled on to the CPU at some point). | |
3616 | */ | |
c88f2096 | 3617 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3618 | return 0; |
3619 | ||
c320c7b7 | 3620 | if (count < event->read_size) |
3dab77fb PZ |
3621 | return -ENOSPC; |
3622 | ||
cdd6c482 | 3623 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3624 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3625 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3626 | else |
cdd6c482 | 3627 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3628 | |
3dab77fb | 3629 | return ret; |
0793a61d TG |
3630 | } |
3631 | ||
0793a61d TG |
3632 | static ssize_t |
3633 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3634 | { | |
cdd6c482 | 3635 | struct perf_event *event = file->private_data; |
0793a61d | 3636 | |
cdd6c482 | 3637 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3638 | } |
3639 | ||
3640 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3641 | { | |
cdd6c482 | 3642 | struct perf_event *event = file->private_data; |
76369139 | 3643 | struct ring_buffer *rb; |
61b67684 | 3644 | unsigned int events = POLLHUP; |
c7138f37 | 3645 | |
e708d7ad | 3646 | poll_wait(file, &event->waitq, wait); |
179033b3 | 3647 | |
dc633982 | 3648 | if (is_event_hup(event)) |
179033b3 JO |
3649 | return events; |
3650 | ||
10c6db11 | 3651 | /* |
9bb5d40c PZ |
3652 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3653 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3654 | */ |
3655 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3656 | rb = event->rb; |
3657 | if (rb) | |
76369139 | 3658 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3659 | mutex_unlock(&event->mmap_mutex); |
0793a61d TG |
3660 | return events; |
3661 | } | |
3662 | ||
cdd6c482 | 3663 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3664 | { |
cdd6c482 | 3665 | (void)perf_event_read(event); |
e7850595 | 3666 | local64_set(&event->count, 0); |
cdd6c482 | 3667 | perf_event_update_userpage(event); |
3df5edad PZ |
3668 | } |
3669 | ||
c93f7669 | 3670 | /* |
cdd6c482 IM |
3671 | * Holding the top-level event's child_mutex means that any |
3672 | * descendant process that has inherited this event will block | |
3673 | * in sync_child_event if it goes to exit, thus satisfying the | |
3674 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3675 | */ |
cdd6c482 IM |
3676 | static void perf_event_for_each_child(struct perf_event *event, |
3677 | void (*func)(struct perf_event *)) | |
3df5edad | 3678 | { |
cdd6c482 | 3679 | struct perf_event *child; |
3df5edad | 3680 | |
cdd6c482 IM |
3681 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3682 | mutex_lock(&event->child_mutex); | |
3683 | func(event); | |
3684 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3685 | func(child); |
cdd6c482 | 3686 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3687 | } |
3688 | ||
cdd6c482 IM |
3689 | static void perf_event_for_each(struct perf_event *event, |
3690 | void (*func)(struct perf_event *)) | |
3df5edad | 3691 | { |
cdd6c482 IM |
3692 | struct perf_event_context *ctx = event->ctx; |
3693 | struct perf_event *sibling; | |
3df5edad | 3694 | |
75f937f2 PZ |
3695 | WARN_ON_ONCE(ctx->parent_ctx); |
3696 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3697 | event = event->group_leader; |
75f937f2 | 3698 | |
cdd6c482 | 3699 | perf_event_for_each_child(event, func); |
cdd6c482 | 3700 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3701 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3702 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3703 | } |
3704 | ||
cdd6c482 | 3705 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3706 | { |
cdd6c482 | 3707 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3708 | int ret = 0, active; |
08247e31 PZ |
3709 | u64 value; |
3710 | ||
6c7e550f | 3711 | if (!is_sampling_event(event)) |
08247e31 PZ |
3712 | return -EINVAL; |
3713 | ||
ad0cf347 | 3714 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3715 | return -EFAULT; |
3716 | ||
3717 | if (!value) | |
3718 | return -EINVAL; | |
3719 | ||
e625cce1 | 3720 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3721 | if (event->attr.freq) { |
3722 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3723 | ret = -EINVAL; |
3724 | goto unlock; | |
3725 | } | |
3726 | ||
cdd6c482 | 3727 | event->attr.sample_freq = value; |
08247e31 | 3728 | } else { |
cdd6c482 IM |
3729 | event->attr.sample_period = value; |
3730 | event->hw.sample_period = value; | |
08247e31 | 3731 | } |
bad7192b PZ |
3732 | |
3733 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3734 | if (active) { | |
3735 | perf_pmu_disable(ctx->pmu); | |
3736 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3737 | } | |
3738 | ||
3739 | local64_set(&event->hw.period_left, 0); | |
3740 | ||
3741 | if (active) { | |
3742 | event->pmu->start(event, PERF_EF_RELOAD); | |
3743 | perf_pmu_enable(ctx->pmu); | |
3744 | } | |
3745 | ||
08247e31 | 3746 | unlock: |
e625cce1 | 3747 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3748 | |
3749 | return ret; | |
3750 | } | |
3751 | ||
ac9721f3 PZ |
3752 | static const struct file_operations perf_fops; |
3753 | ||
2903ff01 | 3754 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3755 | { |
2903ff01 AV |
3756 | struct fd f = fdget(fd); |
3757 | if (!f.file) | |
3758 | return -EBADF; | |
ac9721f3 | 3759 | |
2903ff01 AV |
3760 | if (f.file->f_op != &perf_fops) { |
3761 | fdput(f); | |
3762 | return -EBADF; | |
ac9721f3 | 3763 | } |
2903ff01 AV |
3764 | *p = f; |
3765 | return 0; | |
ac9721f3 PZ |
3766 | } |
3767 | ||
3768 | static int perf_event_set_output(struct perf_event *event, | |
3769 | struct perf_event *output_event); | |
6fb2915d | 3770 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3771 | |
d859e29f PM |
3772 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3773 | { | |
cdd6c482 IM |
3774 | struct perf_event *event = file->private_data; |
3775 | void (*func)(struct perf_event *); | |
3df5edad | 3776 | u32 flags = arg; |
d859e29f PM |
3777 | |
3778 | switch (cmd) { | |
cdd6c482 IM |
3779 | case PERF_EVENT_IOC_ENABLE: |
3780 | func = perf_event_enable; | |
d859e29f | 3781 | break; |
cdd6c482 IM |
3782 | case PERF_EVENT_IOC_DISABLE: |
3783 | func = perf_event_disable; | |
79f14641 | 3784 | break; |
cdd6c482 IM |
3785 | case PERF_EVENT_IOC_RESET: |
3786 | func = perf_event_reset; | |
6de6a7b9 | 3787 | break; |
3df5edad | 3788 | |
cdd6c482 IM |
3789 | case PERF_EVENT_IOC_REFRESH: |
3790 | return perf_event_refresh(event, arg); | |
08247e31 | 3791 | |
cdd6c482 IM |
3792 | case PERF_EVENT_IOC_PERIOD: |
3793 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3794 | |
cf4957f1 JO |
3795 | case PERF_EVENT_IOC_ID: |
3796 | { | |
3797 | u64 id = primary_event_id(event); | |
3798 | ||
3799 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3800 | return -EFAULT; | |
3801 | return 0; | |
3802 | } | |
3803 | ||
cdd6c482 | 3804 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3805 | { |
ac9721f3 | 3806 | int ret; |
ac9721f3 | 3807 | if (arg != -1) { |
2903ff01 AV |
3808 | struct perf_event *output_event; |
3809 | struct fd output; | |
3810 | ret = perf_fget_light(arg, &output); | |
3811 | if (ret) | |
3812 | return ret; | |
3813 | output_event = output.file->private_data; | |
3814 | ret = perf_event_set_output(event, output_event); | |
3815 | fdput(output); | |
3816 | } else { | |
3817 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3818 | } |
ac9721f3 PZ |
3819 | return ret; |
3820 | } | |
a4be7c27 | 3821 | |
6fb2915d LZ |
3822 | case PERF_EVENT_IOC_SET_FILTER: |
3823 | return perf_event_set_filter(event, (void __user *)arg); | |
3824 | ||
d859e29f | 3825 | default: |
3df5edad | 3826 | return -ENOTTY; |
d859e29f | 3827 | } |
3df5edad PZ |
3828 | |
3829 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3830 | perf_event_for_each(event, func); |
3df5edad | 3831 | else |
cdd6c482 | 3832 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3833 | |
3834 | return 0; | |
d859e29f PM |
3835 | } |
3836 | ||
b3f20785 PM |
3837 | #ifdef CONFIG_COMPAT |
3838 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, | |
3839 | unsigned long arg) | |
3840 | { | |
3841 | switch (_IOC_NR(cmd)) { | |
3842 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): | |
3843 | case _IOC_NR(PERF_EVENT_IOC_ID): | |
3844 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ | |
3845 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { | |
3846 | cmd &= ~IOCSIZE_MASK; | |
3847 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; | |
3848 | } | |
3849 | break; | |
3850 | } | |
3851 | return perf_ioctl(file, cmd, arg); | |
3852 | } | |
3853 | #else | |
3854 | # define perf_compat_ioctl NULL | |
3855 | #endif | |
3856 | ||
cdd6c482 | 3857 | int perf_event_task_enable(void) |
771d7cde | 3858 | { |
cdd6c482 | 3859 | struct perf_event *event; |
771d7cde | 3860 | |
cdd6c482 IM |
3861 | mutex_lock(¤t->perf_event_mutex); |
3862 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3863 | perf_event_for_each_child(event, perf_event_enable); | |
3864 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3865 | |
3866 | return 0; | |
3867 | } | |
3868 | ||
cdd6c482 | 3869 | int perf_event_task_disable(void) |
771d7cde | 3870 | { |
cdd6c482 | 3871 | struct perf_event *event; |
771d7cde | 3872 | |
cdd6c482 IM |
3873 | mutex_lock(¤t->perf_event_mutex); |
3874 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3875 | perf_event_for_each_child(event, perf_event_disable); | |
3876 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3877 | |
3878 | return 0; | |
3879 | } | |
3880 | ||
cdd6c482 | 3881 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3882 | { |
a4eaf7f1 PZ |
3883 | if (event->hw.state & PERF_HES_STOPPED) |
3884 | return 0; | |
3885 | ||
cdd6c482 | 3886 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3887 | return 0; |
3888 | ||
35edc2a5 | 3889 | return event->pmu->event_idx(event); |
194002b2 PZ |
3890 | } |
3891 | ||
c4794295 | 3892 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3893 | u64 *now, |
7f310a5d EM |
3894 | u64 *enabled, |
3895 | u64 *running) | |
c4794295 | 3896 | { |
e3f3541c | 3897 | u64 ctx_time; |
c4794295 | 3898 | |
e3f3541c PZ |
3899 | *now = perf_clock(); |
3900 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3901 | *enabled = ctx_time - event->tstamp_enabled; |
3902 | *running = ctx_time - event->tstamp_running; | |
3903 | } | |
3904 | ||
fa731587 PZ |
3905 | static void perf_event_init_userpage(struct perf_event *event) |
3906 | { | |
3907 | struct perf_event_mmap_page *userpg; | |
3908 | struct ring_buffer *rb; | |
3909 | ||
3910 | rcu_read_lock(); | |
3911 | rb = rcu_dereference(event->rb); | |
3912 | if (!rb) | |
3913 | goto unlock; | |
3914 | ||
3915 | userpg = rb->user_page; | |
3916 | ||
3917 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
3918 | userpg->cap_bit0_is_deprecated = 1; | |
3919 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
3920 | ||
3921 | unlock: | |
3922 | rcu_read_unlock(); | |
3923 | } | |
3924 | ||
c7206205 | 3925 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3926 | { |
3927 | } | |
3928 | ||
38ff667b PZ |
3929 | /* |
3930 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3931 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3932 | * code calls this from NMI context. | |
3933 | */ | |
cdd6c482 | 3934 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3935 | { |
cdd6c482 | 3936 | struct perf_event_mmap_page *userpg; |
76369139 | 3937 | struct ring_buffer *rb; |
e3f3541c | 3938 | u64 enabled, running, now; |
38ff667b PZ |
3939 | |
3940 | rcu_read_lock(); | |
5ec4c599 PZ |
3941 | rb = rcu_dereference(event->rb); |
3942 | if (!rb) | |
3943 | goto unlock; | |
3944 | ||
0d641208 EM |
3945 | /* |
3946 | * compute total_time_enabled, total_time_running | |
3947 | * based on snapshot values taken when the event | |
3948 | * was last scheduled in. | |
3949 | * | |
3950 | * we cannot simply called update_context_time() | |
3951 | * because of locking issue as we can be called in | |
3952 | * NMI context | |
3953 | */ | |
e3f3541c | 3954 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 3955 | |
76369139 | 3956 | userpg = rb->user_page; |
7b732a75 PZ |
3957 | /* |
3958 | * Disable preemption so as to not let the corresponding user-space | |
3959 | * spin too long if we get preempted. | |
3960 | */ | |
3961 | preempt_disable(); | |
37d81828 | 3962 | ++userpg->lock; |
92f22a38 | 3963 | barrier(); |
cdd6c482 | 3964 | userpg->index = perf_event_index(event); |
b5e58793 | 3965 | userpg->offset = perf_event_count(event); |
365a4038 | 3966 | if (userpg->index) |
e7850595 | 3967 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3968 | |
0d641208 | 3969 | userpg->time_enabled = enabled + |
cdd6c482 | 3970 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3971 | |
0d641208 | 3972 | userpg->time_running = running + |
cdd6c482 | 3973 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3974 | |
c7206205 | 3975 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3976 | |
92f22a38 | 3977 | barrier(); |
37d81828 | 3978 | ++userpg->lock; |
7b732a75 | 3979 | preempt_enable(); |
38ff667b | 3980 | unlock: |
7b732a75 | 3981 | rcu_read_unlock(); |
37d81828 PM |
3982 | } |
3983 | ||
906010b2 PZ |
3984 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3985 | { | |
3986 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3987 | struct ring_buffer *rb; |
906010b2 PZ |
3988 | int ret = VM_FAULT_SIGBUS; |
3989 | ||
3990 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3991 | if (vmf->pgoff == 0) | |
3992 | ret = 0; | |
3993 | return ret; | |
3994 | } | |
3995 | ||
3996 | rcu_read_lock(); | |
76369139 FW |
3997 | rb = rcu_dereference(event->rb); |
3998 | if (!rb) | |
906010b2 PZ |
3999 | goto unlock; |
4000 | ||
4001 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
4002 | goto unlock; | |
4003 | ||
76369139 | 4004 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
4005 | if (!vmf->page) |
4006 | goto unlock; | |
4007 | ||
4008 | get_page(vmf->page); | |
4009 | vmf->page->mapping = vma->vm_file->f_mapping; | |
4010 | vmf->page->index = vmf->pgoff; | |
4011 | ||
4012 | ret = 0; | |
4013 | unlock: | |
4014 | rcu_read_unlock(); | |
4015 | ||
4016 | return ret; | |
4017 | } | |
4018 | ||
10c6db11 PZ |
4019 | static void ring_buffer_attach(struct perf_event *event, |
4020 | struct ring_buffer *rb) | |
4021 | { | |
b69cf536 | 4022 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
4023 | unsigned long flags; |
4024 | ||
b69cf536 PZ |
4025 | if (event->rb) { |
4026 | /* | |
4027 | * Should be impossible, we set this when removing | |
4028 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
4029 | */ | |
4030 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 4031 | |
b69cf536 PZ |
4032 | old_rb = event->rb; |
4033 | event->rcu_batches = get_state_synchronize_rcu(); | |
4034 | event->rcu_pending = 1; | |
10c6db11 | 4035 | |
b69cf536 PZ |
4036 | spin_lock_irqsave(&old_rb->event_lock, flags); |
4037 | list_del_rcu(&event->rb_entry); | |
4038 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
4039 | } | |
10c6db11 | 4040 | |
b69cf536 PZ |
4041 | if (event->rcu_pending && rb) { |
4042 | cond_synchronize_rcu(event->rcu_batches); | |
4043 | event->rcu_pending = 0; | |
4044 | } | |
10c6db11 | 4045 | |
b69cf536 PZ |
4046 | if (rb) { |
4047 | spin_lock_irqsave(&rb->event_lock, flags); | |
4048 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
4049 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
4050 | } | |
4051 | ||
4052 | rcu_assign_pointer(event->rb, rb); | |
4053 | ||
4054 | if (old_rb) { | |
4055 | ring_buffer_put(old_rb); | |
4056 | /* | |
4057 | * Since we detached before setting the new rb, so that we | |
4058 | * could attach the new rb, we could have missed a wakeup. | |
4059 | * Provide it now. | |
4060 | */ | |
4061 | wake_up_all(&event->waitq); | |
4062 | } | |
10c6db11 PZ |
4063 | } |
4064 | ||
4065 | static void ring_buffer_wakeup(struct perf_event *event) | |
4066 | { | |
4067 | struct ring_buffer *rb; | |
4068 | ||
4069 | rcu_read_lock(); | |
4070 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
4071 | if (rb) { |
4072 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
4073 | wake_up_all(&event->waitq); | |
4074 | } | |
10c6db11 PZ |
4075 | rcu_read_unlock(); |
4076 | } | |
4077 | ||
76369139 | 4078 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 4079 | { |
76369139 | 4080 | struct ring_buffer *rb; |
906010b2 | 4081 | |
76369139 FW |
4082 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
4083 | rb_free(rb); | |
7b732a75 PZ |
4084 | } |
4085 | ||
76369139 | 4086 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 4087 | { |
76369139 | 4088 | struct ring_buffer *rb; |
7b732a75 | 4089 | |
ac9721f3 | 4090 | rcu_read_lock(); |
76369139 FW |
4091 | rb = rcu_dereference(event->rb); |
4092 | if (rb) { | |
4093 | if (!atomic_inc_not_zero(&rb->refcount)) | |
4094 | rb = NULL; | |
ac9721f3 PZ |
4095 | } |
4096 | rcu_read_unlock(); | |
4097 | ||
76369139 | 4098 | return rb; |
ac9721f3 PZ |
4099 | } |
4100 | ||
76369139 | 4101 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 4102 | { |
76369139 | 4103 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 4104 | return; |
7b732a75 | 4105 | |
9bb5d40c | 4106 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 4107 | |
76369139 | 4108 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
4109 | } |
4110 | ||
4111 | static void perf_mmap_open(struct vm_area_struct *vma) | |
4112 | { | |
cdd6c482 | 4113 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4114 | |
cdd6c482 | 4115 | atomic_inc(&event->mmap_count); |
9bb5d40c | 4116 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
4117 | } |
4118 | ||
9bb5d40c PZ |
4119 | /* |
4120 | * A buffer can be mmap()ed multiple times; either directly through the same | |
4121 | * event, or through other events by use of perf_event_set_output(). | |
4122 | * | |
4123 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
4124 | * the buffer here, where we still have a VM context. This means we need | |
4125 | * to detach all events redirecting to us. | |
4126 | */ | |
7b732a75 PZ |
4127 | static void perf_mmap_close(struct vm_area_struct *vma) |
4128 | { | |
cdd6c482 | 4129 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4130 | |
b69cf536 | 4131 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
4132 | struct user_struct *mmap_user = rb->mmap_user; |
4133 | int mmap_locked = rb->mmap_locked; | |
4134 | unsigned long size = perf_data_size(rb); | |
789f90fc | 4135 | |
9bb5d40c PZ |
4136 | atomic_dec(&rb->mmap_count); |
4137 | ||
4138 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4139 | goto out_put; |
9bb5d40c | 4140 | |
b69cf536 | 4141 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4142 | mutex_unlock(&event->mmap_mutex); |
4143 | ||
4144 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4145 | if (atomic_read(&rb->mmap_count)) |
4146 | goto out_put; | |
ac9721f3 | 4147 | |
9bb5d40c PZ |
4148 | /* |
4149 | * No other mmap()s, detach from all other events that might redirect | |
4150 | * into the now unreachable buffer. Somewhat complicated by the | |
4151 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4152 | */ | |
4153 | again: | |
4154 | rcu_read_lock(); | |
4155 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4156 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4157 | /* | |
4158 | * This event is en-route to free_event() which will | |
4159 | * detach it and remove it from the list. | |
4160 | */ | |
4161 | continue; | |
4162 | } | |
4163 | rcu_read_unlock(); | |
789f90fc | 4164 | |
9bb5d40c PZ |
4165 | mutex_lock(&event->mmap_mutex); |
4166 | /* | |
4167 | * Check we didn't race with perf_event_set_output() which can | |
4168 | * swizzle the rb from under us while we were waiting to | |
4169 | * acquire mmap_mutex. | |
4170 | * | |
4171 | * If we find a different rb; ignore this event, a next | |
4172 | * iteration will no longer find it on the list. We have to | |
4173 | * still restart the iteration to make sure we're not now | |
4174 | * iterating the wrong list. | |
4175 | */ | |
b69cf536 PZ |
4176 | if (event->rb == rb) |
4177 | ring_buffer_attach(event, NULL); | |
4178 | ||
cdd6c482 | 4179 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4180 | put_event(event); |
ac9721f3 | 4181 | |
9bb5d40c PZ |
4182 | /* |
4183 | * Restart the iteration; either we're on the wrong list or | |
4184 | * destroyed its integrity by doing a deletion. | |
4185 | */ | |
4186 | goto again; | |
7b732a75 | 4187 | } |
9bb5d40c PZ |
4188 | rcu_read_unlock(); |
4189 | ||
4190 | /* | |
4191 | * It could be there's still a few 0-ref events on the list; they'll | |
4192 | * get cleaned up by free_event() -- they'll also still have their | |
4193 | * ref on the rb and will free it whenever they are done with it. | |
4194 | * | |
4195 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4196 | * undo the VM accounting. | |
4197 | */ | |
4198 | ||
4199 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4200 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4201 | free_uid(mmap_user); | |
4202 | ||
b69cf536 | 4203 | out_put: |
9bb5d40c | 4204 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4205 | } |
4206 | ||
f0f37e2f | 4207 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4208 | .open = perf_mmap_open, |
4209 | .close = perf_mmap_close, | |
4210 | .fault = perf_mmap_fault, | |
4211 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4212 | }; |
4213 | ||
4214 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4215 | { | |
cdd6c482 | 4216 | struct perf_event *event = file->private_data; |
22a4f650 | 4217 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4218 | struct user_struct *user = current_user(); |
22a4f650 | 4219 | unsigned long locked, lock_limit; |
76369139 | 4220 | struct ring_buffer *rb; |
7b732a75 PZ |
4221 | unsigned long vma_size; |
4222 | unsigned long nr_pages; | |
789f90fc | 4223 | long user_extra, extra; |
d57e34fd | 4224 | int ret = 0, flags = 0; |
37d81828 | 4225 | |
c7920614 PZ |
4226 | /* |
4227 | * Don't allow mmap() of inherited per-task counters. This would | |
4228 | * create a performance issue due to all children writing to the | |
76369139 | 4229 | * same rb. |
c7920614 PZ |
4230 | */ |
4231 | if (event->cpu == -1 && event->attr.inherit) | |
4232 | return -EINVAL; | |
4233 | ||
43a21ea8 | 4234 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4235 | return -EINVAL; |
7b732a75 PZ |
4236 | |
4237 | vma_size = vma->vm_end - vma->vm_start; | |
4238 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4239 | ||
7730d865 | 4240 | /* |
76369139 | 4241 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4242 | * can do bitmasks instead of modulo. |
4243 | */ | |
4244 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
4245 | return -EINVAL; |
4246 | ||
7b732a75 | 4247 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4248 | return -EINVAL; |
4249 | ||
7b732a75 PZ |
4250 | if (vma->vm_pgoff != 0) |
4251 | return -EINVAL; | |
37d81828 | 4252 | |
cdd6c482 | 4253 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4254 | again: |
cdd6c482 | 4255 | mutex_lock(&event->mmap_mutex); |
76369139 | 4256 | if (event->rb) { |
9bb5d40c | 4257 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4258 | ret = -EINVAL; |
9bb5d40c PZ |
4259 | goto unlock; |
4260 | } | |
4261 | ||
4262 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4263 | /* | |
4264 | * Raced against perf_mmap_close() through | |
4265 | * perf_event_set_output(). Try again, hope for better | |
4266 | * luck. | |
4267 | */ | |
4268 | mutex_unlock(&event->mmap_mutex); | |
4269 | goto again; | |
4270 | } | |
4271 | ||
ebb3c4c4 PZ |
4272 | goto unlock; |
4273 | } | |
4274 | ||
789f90fc | 4275 | user_extra = nr_pages + 1; |
cdd6c482 | 4276 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4277 | |
4278 | /* | |
4279 | * Increase the limit linearly with more CPUs: | |
4280 | */ | |
4281 | user_lock_limit *= num_online_cpus(); | |
4282 | ||
789f90fc | 4283 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4284 | |
789f90fc PZ |
4285 | extra = 0; |
4286 | if (user_locked > user_lock_limit) | |
4287 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4288 | |
78d7d407 | 4289 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4290 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4291 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4292 | |
459ec28a IM |
4293 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4294 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4295 | ret = -EPERM; |
4296 | goto unlock; | |
4297 | } | |
7b732a75 | 4298 | |
76369139 | 4299 | WARN_ON(event->rb); |
906010b2 | 4300 | |
d57e34fd | 4301 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4302 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4303 | |
4ec8363d VW |
4304 | rb = rb_alloc(nr_pages, |
4305 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4306 | event->cpu, flags); | |
4307 | ||
76369139 | 4308 | if (!rb) { |
ac9721f3 | 4309 | ret = -ENOMEM; |
ebb3c4c4 | 4310 | goto unlock; |
ac9721f3 | 4311 | } |
26cb63ad | 4312 | |
9bb5d40c | 4313 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4314 | rb->mmap_locked = extra; |
4315 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4316 | |
ac9721f3 | 4317 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4318 | vma->vm_mm->pinned_vm += extra; |
4319 | ||
9bb5d40c | 4320 | ring_buffer_attach(event, rb); |
ac9721f3 | 4321 | |
fa731587 | 4322 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4323 | perf_event_update_userpage(event); |
4324 | ||
ebb3c4c4 | 4325 | unlock: |
ac9721f3 PZ |
4326 | if (!ret) |
4327 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4328 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4329 | |
9bb5d40c PZ |
4330 | /* |
4331 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4332 | * vma. | |
4333 | */ | |
26cb63ad | 4334 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4335 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4336 | |
4337 | return ret; | |
37d81828 PM |
4338 | } |
4339 | ||
3c446b3d PZ |
4340 | static int perf_fasync(int fd, struct file *filp, int on) |
4341 | { | |
496ad9aa | 4342 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4343 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4344 | int retval; |
4345 | ||
4346 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4347 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4348 | mutex_unlock(&inode->i_mutex); |
4349 | ||
4350 | if (retval < 0) | |
4351 | return retval; | |
4352 | ||
4353 | return 0; | |
4354 | } | |
4355 | ||
0793a61d | 4356 | static const struct file_operations perf_fops = { |
3326c1ce | 4357 | .llseek = no_llseek, |
0793a61d TG |
4358 | .release = perf_release, |
4359 | .read = perf_read, | |
4360 | .poll = perf_poll, | |
d859e29f | 4361 | .unlocked_ioctl = perf_ioctl, |
b3f20785 | 4362 | .compat_ioctl = perf_compat_ioctl, |
37d81828 | 4363 | .mmap = perf_mmap, |
3c446b3d | 4364 | .fasync = perf_fasync, |
0793a61d TG |
4365 | }; |
4366 | ||
925d519a | 4367 | /* |
cdd6c482 | 4368 | * Perf event wakeup |
925d519a PZ |
4369 | * |
4370 | * If there's data, ensure we set the poll() state and publish everything | |
4371 | * to user-space before waking everybody up. | |
4372 | */ | |
4373 | ||
cdd6c482 | 4374 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4375 | { |
10c6db11 | 4376 | ring_buffer_wakeup(event); |
4c9e2542 | 4377 | |
cdd6c482 IM |
4378 | if (event->pending_kill) { |
4379 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4380 | event->pending_kill = 0; | |
4c9e2542 | 4381 | } |
925d519a PZ |
4382 | } |
4383 | ||
e360adbe | 4384 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4385 | { |
cdd6c482 IM |
4386 | struct perf_event *event = container_of(entry, |
4387 | struct perf_event, pending); | |
79f14641 | 4388 | |
cdd6c482 IM |
4389 | if (event->pending_disable) { |
4390 | event->pending_disable = 0; | |
4391 | __perf_event_disable(event); | |
79f14641 PZ |
4392 | } |
4393 | ||
cdd6c482 IM |
4394 | if (event->pending_wakeup) { |
4395 | event->pending_wakeup = 0; | |
4396 | perf_event_wakeup(event); | |
79f14641 PZ |
4397 | } |
4398 | } | |
4399 | ||
39447b38 ZY |
4400 | /* |
4401 | * We assume there is only KVM supporting the callbacks. | |
4402 | * Later on, we might change it to a list if there is | |
4403 | * another virtualization implementation supporting the callbacks. | |
4404 | */ | |
4405 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4406 | ||
4407 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4408 | { | |
4409 | perf_guest_cbs = cbs; | |
4410 | return 0; | |
4411 | } | |
4412 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4413 | ||
4414 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4415 | { | |
4416 | perf_guest_cbs = NULL; | |
4417 | return 0; | |
4418 | } | |
4419 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4420 | ||
4018994f JO |
4421 | static void |
4422 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4423 | struct pt_regs *regs, u64 mask) | |
4424 | { | |
4425 | int bit; | |
4426 | ||
4427 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4428 | sizeof(mask) * BITS_PER_BYTE) { | |
4429 | u64 val; | |
4430 | ||
4431 | val = perf_reg_value(regs, bit); | |
4432 | perf_output_put(handle, val); | |
4433 | } | |
4434 | } | |
4435 | ||
4436 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4437 | struct pt_regs *regs) | |
4438 | { | |
4439 | if (!user_mode(regs)) { | |
4440 | if (current->mm) | |
4441 | regs = task_pt_regs(current); | |
4442 | else | |
4443 | regs = NULL; | |
4444 | } | |
4445 | ||
4446 | if (regs) { | |
4447 | regs_user->regs = regs; | |
4448 | regs_user->abi = perf_reg_abi(current); | |
4449 | } | |
4450 | } | |
4451 | ||
c5ebcedb JO |
4452 | /* |
4453 | * Get remaining task size from user stack pointer. | |
4454 | * | |
4455 | * It'd be better to take stack vma map and limit this more | |
4456 | * precisly, but there's no way to get it safely under interrupt, | |
4457 | * so using TASK_SIZE as limit. | |
4458 | */ | |
4459 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4460 | { | |
4461 | unsigned long addr = perf_user_stack_pointer(regs); | |
4462 | ||
4463 | if (!addr || addr >= TASK_SIZE) | |
4464 | return 0; | |
4465 | ||
4466 | return TASK_SIZE - addr; | |
4467 | } | |
4468 | ||
4469 | static u16 | |
4470 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4471 | struct pt_regs *regs) | |
4472 | { | |
4473 | u64 task_size; | |
4474 | ||
4475 | /* No regs, no stack pointer, no dump. */ | |
4476 | if (!regs) | |
4477 | return 0; | |
4478 | ||
4479 | /* | |
4480 | * Check if we fit in with the requested stack size into the: | |
4481 | * - TASK_SIZE | |
4482 | * If we don't, we limit the size to the TASK_SIZE. | |
4483 | * | |
4484 | * - remaining sample size | |
4485 | * If we don't, we customize the stack size to | |
4486 | * fit in to the remaining sample size. | |
4487 | */ | |
4488 | ||
4489 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4490 | stack_size = min(stack_size, (u16) task_size); | |
4491 | ||
4492 | /* Current header size plus static size and dynamic size. */ | |
4493 | header_size += 2 * sizeof(u64); | |
4494 | ||
4495 | /* Do we fit in with the current stack dump size? */ | |
4496 | if ((u16) (header_size + stack_size) < header_size) { | |
4497 | /* | |
4498 | * If we overflow the maximum size for the sample, | |
4499 | * we customize the stack dump size to fit in. | |
4500 | */ | |
4501 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4502 | stack_size = round_up(stack_size, sizeof(u64)); | |
4503 | } | |
4504 | ||
4505 | return stack_size; | |
4506 | } | |
4507 | ||
4508 | static void | |
4509 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4510 | struct pt_regs *regs) | |
4511 | { | |
4512 | /* Case of a kernel thread, nothing to dump */ | |
4513 | if (!regs) { | |
4514 | u64 size = 0; | |
4515 | perf_output_put(handle, size); | |
4516 | } else { | |
4517 | unsigned long sp; | |
4518 | unsigned int rem; | |
4519 | u64 dyn_size; | |
4520 | ||
4521 | /* | |
4522 | * We dump: | |
4523 | * static size | |
4524 | * - the size requested by user or the best one we can fit | |
4525 | * in to the sample max size | |
4526 | * data | |
4527 | * - user stack dump data | |
4528 | * dynamic size | |
4529 | * - the actual dumped size | |
4530 | */ | |
4531 | ||
4532 | /* Static size. */ | |
4533 | perf_output_put(handle, dump_size); | |
4534 | ||
4535 | /* Data. */ | |
4536 | sp = perf_user_stack_pointer(regs); | |
4537 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4538 | dyn_size = dump_size - rem; | |
4539 | ||
4540 | perf_output_skip(handle, rem); | |
4541 | ||
4542 | /* Dynamic size. */ | |
4543 | perf_output_put(handle, dyn_size); | |
4544 | } | |
4545 | } | |
4546 | ||
c980d109 ACM |
4547 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4548 | struct perf_sample_data *data, | |
4549 | struct perf_event *event) | |
6844c09d ACM |
4550 | { |
4551 | u64 sample_type = event->attr.sample_type; | |
4552 | ||
4553 | data->type = sample_type; | |
4554 | header->size += event->id_header_size; | |
4555 | ||
4556 | if (sample_type & PERF_SAMPLE_TID) { | |
4557 | /* namespace issues */ | |
4558 | data->tid_entry.pid = perf_event_pid(event, current); | |
4559 | data->tid_entry.tid = perf_event_tid(event, current); | |
4560 | } | |
4561 | ||
4562 | if (sample_type & PERF_SAMPLE_TIME) | |
4563 | data->time = perf_clock(); | |
4564 | ||
ff3d527c | 4565 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4566 | data->id = primary_event_id(event); |
4567 | ||
4568 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4569 | data->stream_id = event->id; | |
4570 | ||
4571 | if (sample_type & PERF_SAMPLE_CPU) { | |
4572 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4573 | data->cpu_entry.reserved = 0; | |
4574 | } | |
4575 | } | |
4576 | ||
76369139 FW |
4577 | void perf_event_header__init_id(struct perf_event_header *header, |
4578 | struct perf_sample_data *data, | |
4579 | struct perf_event *event) | |
c980d109 ACM |
4580 | { |
4581 | if (event->attr.sample_id_all) | |
4582 | __perf_event_header__init_id(header, data, event); | |
4583 | } | |
4584 | ||
4585 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4586 | struct perf_sample_data *data) | |
4587 | { | |
4588 | u64 sample_type = data->type; | |
4589 | ||
4590 | if (sample_type & PERF_SAMPLE_TID) | |
4591 | perf_output_put(handle, data->tid_entry); | |
4592 | ||
4593 | if (sample_type & PERF_SAMPLE_TIME) | |
4594 | perf_output_put(handle, data->time); | |
4595 | ||
4596 | if (sample_type & PERF_SAMPLE_ID) | |
4597 | perf_output_put(handle, data->id); | |
4598 | ||
4599 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4600 | perf_output_put(handle, data->stream_id); | |
4601 | ||
4602 | if (sample_type & PERF_SAMPLE_CPU) | |
4603 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4604 | |
4605 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4606 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4607 | } |
4608 | ||
76369139 FW |
4609 | void perf_event__output_id_sample(struct perf_event *event, |
4610 | struct perf_output_handle *handle, | |
4611 | struct perf_sample_data *sample) | |
c980d109 ACM |
4612 | { |
4613 | if (event->attr.sample_id_all) | |
4614 | __perf_event__output_id_sample(handle, sample); | |
4615 | } | |
4616 | ||
3dab77fb | 4617 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4618 | struct perf_event *event, |
4619 | u64 enabled, u64 running) | |
3dab77fb | 4620 | { |
cdd6c482 | 4621 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4622 | u64 values[4]; |
4623 | int n = 0; | |
4624 | ||
b5e58793 | 4625 | values[n++] = perf_event_count(event); |
3dab77fb | 4626 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4627 | values[n++] = enabled + |
cdd6c482 | 4628 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4629 | } |
4630 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4631 | values[n++] = running + |
cdd6c482 | 4632 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4633 | } |
4634 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4635 | values[n++] = primary_event_id(event); |
3dab77fb | 4636 | |
76369139 | 4637 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4638 | } |
4639 | ||
4640 | /* | |
cdd6c482 | 4641 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4642 | */ |
4643 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4644 | struct perf_event *event, |
4645 | u64 enabled, u64 running) | |
3dab77fb | 4646 | { |
cdd6c482 IM |
4647 | struct perf_event *leader = event->group_leader, *sub; |
4648 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4649 | u64 values[5]; |
4650 | int n = 0; | |
4651 | ||
4652 | values[n++] = 1 + leader->nr_siblings; | |
4653 | ||
4654 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4655 | values[n++] = enabled; |
3dab77fb PZ |
4656 | |
4657 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4658 | values[n++] = running; |
3dab77fb | 4659 | |
cdd6c482 | 4660 | if (leader != event) |
3dab77fb PZ |
4661 | leader->pmu->read(leader); |
4662 | ||
b5e58793 | 4663 | values[n++] = perf_event_count(leader); |
3dab77fb | 4664 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4665 | values[n++] = primary_event_id(leader); |
3dab77fb | 4666 | |
76369139 | 4667 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4668 | |
65abc865 | 4669 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4670 | n = 0; |
4671 | ||
6f5ab001 JO |
4672 | if ((sub != event) && |
4673 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4674 | sub->pmu->read(sub); |
4675 | ||
b5e58793 | 4676 | values[n++] = perf_event_count(sub); |
3dab77fb | 4677 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4678 | values[n++] = primary_event_id(sub); |
3dab77fb | 4679 | |
76369139 | 4680 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4681 | } |
4682 | } | |
4683 | ||
eed01528 SE |
4684 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4685 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4686 | ||
3dab77fb | 4687 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4688 | struct perf_event *event) |
3dab77fb | 4689 | { |
e3f3541c | 4690 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4691 | u64 read_format = event->attr.read_format; |
4692 | ||
4693 | /* | |
4694 | * compute total_time_enabled, total_time_running | |
4695 | * based on snapshot values taken when the event | |
4696 | * was last scheduled in. | |
4697 | * | |
4698 | * we cannot simply called update_context_time() | |
4699 | * because of locking issue as we are called in | |
4700 | * NMI context | |
4701 | */ | |
c4794295 | 4702 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4703 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4704 | |
cdd6c482 | 4705 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4706 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4707 | else |
eed01528 | 4708 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4709 | } |
4710 | ||
5622f295 MM |
4711 | void perf_output_sample(struct perf_output_handle *handle, |
4712 | struct perf_event_header *header, | |
4713 | struct perf_sample_data *data, | |
cdd6c482 | 4714 | struct perf_event *event) |
5622f295 MM |
4715 | { |
4716 | u64 sample_type = data->type; | |
4717 | ||
4718 | perf_output_put(handle, *header); | |
4719 | ||
ff3d527c AH |
4720 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4721 | perf_output_put(handle, data->id); | |
4722 | ||
5622f295 MM |
4723 | if (sample_type & PERF_SAMPLE_IP) |
4724 | perf_output_put(handle, data->ip); | |
4725 | ||
4726 | if (sample_type & PERF_SAMPLE_TID) | |
4727 | perf_output_put(handle, data->tid_entry); | |
4728 | ||
4729 | if (sample_type & PERF_SAMPLE_TIME) | |
4730 | perf_output_put(handle, data->time); | |
4731 | ||
4732 | if (sample_type & PERF_SAMPLE_ADDR) | |
4733 | perf_output_put(handle, data->addr); | |
4734 | ||
4735 | if (sample_type & PERF_SAMPLE_ID) | |
4736 | perf_output_put(handle, data->id); | |
4737 | ||
4738 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4739 | perf_output_put(handle, data->stream_id); | |
4740 | ||
4741 | if (sample_type & PERF_SAMPLE_CPU) | |
4742 | perf_output_put(handle, data->cpu_entry); | |
4743 | ||
4744 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4745 | perf_output_put(handle, data->period); | |
4746 | ||
4747 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4748 | perf_output_read(handle, event); |
5622f295 MM |
4749 | |
4750 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4751 | if (data->callchain) { | |
4752 | int size = 1; | |
4753 | ||
4754 | if (data->callchain) | |
4755 | size += data->callchain->nr; | |
4756 | ||
4757 | size *= sizeof(u64); | |
4758 | ||
76369139 | 4759 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4760 | } else { |
4761 | u64 nr = 0; | |
4762 | perf_output_put(handle, nr); | |
4763 | } | |
4764 | } | |
4765 | ||
4766 | if (sample_type & PERF_SAMPLE_RAW) { | |
4767 | if (data->raw) { | |
4768 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4769 | __output_copy(handle, data->raw->data, |
4770 | data->raw->size); | |
5622f295 MM |
4771 | } else { |
4772 | struct { | |
4773 | u32 size; | |
4774 | u32 data; | |
4775 | } raw = { | |
4776 | .size = sizeof(u32), | |
4777 | .data = 0, | |
4778 | }; | |
4779 | perf_output_put(handle, raw); | |
4780 | } | |
4781 | } | |
a7ac67ea | 4782 | |
bce38cd5 SE |
4783 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4784 | if (data->br_stack) { | |
4785 | size_t size; | |
4786 | ||
4787 | size = data->br_stack->nr | |
4788 | * sizeof(struct perf_branch_entry); | |
4789 | ||
4790 | perf_output_put(handle, data->br_stack->nr); | |
4791 | perf_output_copy(handle, data->br_stack->entries, size); | |
4792 | } else { | |
4793 | /* | |
4794 | * we always store at least the value of nr | |
4795 | */ | |
4796 | u64 nr = 0; | |
4797 | perf_output_put(handle, nr); | |
4798 | } | |
4799 | } | |
4018994f JO |
4800 | |
4801 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4802 | u64 abi = data->regs_user.abi; | |
4803 | ||
4804 | /* | |
4805 | * If there are no regs to dump, notice it through | |
4806 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4807 | */ | |
4808 | perf_output_put(handle, abi); | |
4809 | ||
4810 | if (abi) { | |
4811 | u64 mask = event->attr.sample_regs_user; | |
4812 | perf_output_sample_regs(handle, | |
4813 | data->regs_user.regs, | |
4814 | mask); | |
4815 | } | |
4816 | } | |
c5ebcedb | 4817 | |
a5cdd40c | 4818 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
4819 | perf_output_sample_ustack(handle, |
4820 | data->stack_user_size, | |
4821 | data->regs_user.regs); | |
a5cdd40c | 4822 | } |
c3feedf2 AK |
4823 | |
4824 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4825 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4826 | |
4827 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4828 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 4829 | |
fdfbbd07 AK |
4830 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
4831 | perf_output_put(handle, data->txn); | |
4832 | ||
a5cdd40c PZ |
4833 | if (!event->attr.watermark) { |
4834 | int wakeup_events = event->attr.wakeup_events; | |
4835 | ||
4836 | if (wakeup_events) { | |
4837 | struct ring_buffer *rb = handle->rb; | |
4838 | int events = local_inc_return(&rb->events); | |
4839 | ||
4840 | if (events >= wakeup_events) { | |
4841 | local_sub(wakeup_events, &rb->events); | |
4842 | local_inc(&rb->wakeup); | |
4843 | } | |
4844 | } | |
4845 | } | |
5622f295 MM |
4846 | } |
4847 | ||
4848 | void perf_prepare_sample(struct perf_event_header *header, | |
4849 | struct perf_sample_data *data, | |
cdd6c482 | 4850 | struct perf_event *event, |
5622f295 | 4851 | struct pt_regs *regs) |
7b732a75 | 4852 | { |
cdd6c482 | 4853 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4854 | |
cdd6c482 | 4855 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4856 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4857 | |
4858 | header->misc = 0; | |
4859 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4860 | |
c980d109 | 4861 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4862 | |
c320c7b7 | 4863 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4864 | data->ip = perf_instruction_pointer(regs); |
4865 | ||
b23f3325 | 4866 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4867 | int size = 1; |
394ee076 | 4868 | |
e6dab5ff | 4869 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4870 | |
4871 | if (data->callchain) | |
4872 | size += data->callchain->nr; | |
4873 | ||
4874 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4875 | } |
4876 | ||
3a43ce68 | 4877 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4878 | int size = sizeof(u32); |
4879 | ||
4880 | if (data->raw) | |
4881 | size += data->raw->size; | |
4882 | else | |
4883 | size += sizeof(u32); | |
4884 | ||
4885 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4886 | header->size += size; |
7f453c24 | 4887 | } |
bce38cd5 SE |
4888 | |
4889 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4890 | int size = sizeof(u64); /* nr */ | |
4891 | if (data->br_stack) { | |
4892 | size += data->br_stack->nr | |
4893 | * sizeof(struct perf_branch_entry); | |
4894 | } | |
4895 | header->size += size; | |
4896 | } | |
4018994f JO |
4897 | |
4898 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4899 | /* regs dump ABI info */ | |
4900 | int size = sizeof(u64); | |
4901 | ||
4902 | perf_sample_regs_user(&data->regs_user, regs); | |
4903 | ||
4904 | if (data->regs_user.regs) { | |
4905 | u64 mask = event->attr.sample_regs_user; | |
4906 | size += hweight64(mask) * sizeof(u64); | |
4907 | } | |
4908 | ||
4909 | header->size += size; | |
4910 | } | |
c5ebcedb JO |
4911 | |
4912 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4913 | /* | |
4914 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4915 | * processed as the last one or have additional check added | |
4916 | * in case new sample type is added, because we could eat | |
4917 | * up the rest of the sample size. | |
4918 | */ | |
4919 | struct perf_regs_user *uregs = &data->regs_user; | |
4920 | u16 stack_size = event->attr.sample_stack_user; | |
4921 | u16 size = sizeof(u64); | |
4922 | ||
4923 | if (!uregs->abi) | |
4924 | perf_sample_regs_user(uregs, regs); | |
4925 | ||
4926 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4927 | uregs->regs); | |
4928 | ||
4929 | /* | |
4930 | * If there is something to dump, add space for the dump | |
4931 | * itself and for the field that tells the dynamic size, | |
4932 | * which is how many have been actually dumped. | |
4933 | */ | |
4934 | if (stack_size) | |
4935 | size += sizeof(u64) + stack_size; | |
4936 | ||
4937 | data->stack_user_size = stack_size; | |
4938 | header->size += size; | |
4939 | } | |
5622f295 | 4940 | } |
7f453c24 | 4941 | |
a8b0ca17 | 4942 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4943 | struct perf_sample_data *data, |
4944 | struct pt_regs *regs) | |
4945 | { | |
4946 | struct perf_output_handle handle; | |
4947 | struct perf_event_header header; | |
689802b2 | 4948 | |
927c7a9e FW |
4949 | /* protect the callchain buffers */ |
4950 | rcu_read_lock(); | |
4951 | ||
cdd6c482 | 4952 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4953 | |
a7ac67ea | 4954 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4955 | goto exit; |
0322cd6e | 4956 | |
cdd6c482 | 4957 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4958 | |
8a057d84 | 4959 | perf_output_end(&handle); |
927c7a9e FW |
4960 | |
4961 | exit: | |
4962 | rcu_read_unlock(); | |
0322cd6e PZ |
4963 | } |
4964 | ||
38b200d6 | 4965 | /* |
cdd6c482 | 4966 | * read event_id |
38b200d6 PZ |
4967 | */ |
4968 | ||
4969 | struct perf_read_event { | |
4970 | struct perf_event_header header; | |
4971 | ||
4972 | u32 pid; | |
4973 | u32 tid; | |
38b200d6 PZ |
4974 | }; |
4975 | ||
4976 | static void | |
cdd6c482 | 4977 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4978 | struct task_struct *task) |
4979 | { | |
4980 | struct perf_output_handle handle; | |
c980d109 | 4981 | struct perf_sample_data sample; |
dfc65094 | 4982 | struct perf_read_event read_event = { |
38b200d6 | 4983 | .header = { |
cdd6c482 | 4984 | .type = PERF_RECORD_READ, |
38b200d6 | 4985 | .misc = 0, |
c320c7b7 | 4986 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4987 | }, |
cdd6c482 IM |
4988 | .pid = perf_event_pid(event, task), |
4989 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4990 | }; |
3dab77fb | 4991 | int ret; |
38b200d6 | 4992 | |
c980d109 | 4993 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4994 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4995 | if (ret) |
4996 | return; | |
4997 | ||
dfc65094 | 4998 | perf_output_put(&handle, read_event); |
cdd6c482 | 4999 | perf_output_read(&handle, event); |
c980d109 | 5000 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 5001 | |
38b200d6 PZ |
5002 | perf_output_end(&handle); |
5003 | } | |
5004 | ||
52d857a8 JO |
5005 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
5006 | ||
5007 | static void | |
5008 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
5009 | perf_event_aux_output_cb output, |
5010 | void *data) | |
5011 | { | |
5012 | struct perf_event *event; | |
5013 | ||
5014 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5015 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
5016 | continue; | |
5017 | if (!event_filter_match(event)) | |
5018 | continue; | |
67516844 | 5019 | output(event, data); |
52d857a8 JO |
5020 | } |
5021 | } | |
5022 | ||
5023 | static void | |
67516844 | 5024 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
5025 | struct perf_event_context *task_ctx) |
5026 | { | |
5027 | struct perf_cpu_context *cpuctx; | |
5028 | struct perf_event_context *ctx; | |
5029 | struct pmu *pmu; | |
5030 | int ctxn; | |
5031 | ||
5032 | rcu_read_lock(); | |
5033 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
5034 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
5035 | if (cpuctx->unique_pmu != pmu) | |
5036 | goto next; | |
67516844 | 5037 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
5038 | if (task_ctx) |
5039 | goto next; | |
5040 | ctxn = pmu->task_ctx_nr; | |
5041 | if (ctxn < 0) | |
5042 | goto next; | |
5043 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
5044 | if (ctx) | |
67516844 | 5045 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
5046 | next: |
5047 | put_cpu_ptr(pmu->pmu_cpu_context); | |
5048 | } | |
5049 | ||
5050 | if (task_ctx) { | |
5051 | preempt_disable(); | |
67516844 | 5052 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
5053 | preempt_enable(); |
5054 | } | |
5055 | rcu_read_unlock(); | |
5056 | } | |
5057 | ||
60313ebe | 5058 | /* |
9f498cc5 PZ |
5059 | * task tracking -- fork/exit |
5060 | * | |
13d7a241 | 5061 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
5062 | */ |
5063 | ||
9f498cc5 | 5064 | struct perf_task_event { |
3a80b4a3 | 5065 | struct task_struct *task; |
cdd6c482 | 5066 | struct perf_event_context *task_ctx; |
60313ebe PZ |
5067 | |
5068 | struct { | |
5069 | struct perf_event_header header; | |
5070 | ||
5071 | u32 pid; | |
5072 | u32 ppid; | |
9f498cc5 PZ |
5073 | u32 tid; |
5074 | u32 ptid; | |
393b2ad8 | 5075 | u64 time; |
cdd6c482 | 5076 | } event_id; |
60313ebe PZ |
5077 | }; |
5078 | ||
67516844 JO |
5079 | static int perf_event_task_match(struct perf_event *event) |
5080 | { | |
13d7a241 SE |
5081 | return event->attr.comm || event->attr.mmap || |
5082 | event->attr.mmap2 || event->attr.mmap_data || | |
5083 | event->attr.task; | |
67516844 JO |
5084 | } |
5085 | ||
cdd6c482 | 5086 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 5087 | void *data) |
60313ebe | 5088 | { |
52d857a8 | 5089 | struct perf_task_event *task_event = data; |
60313ebe | 5090 | struct perf_output_handle handle; |
c980d109 | 5091 | struct perf_sample_data sample; |
9f498cc5 | 5092 | struct task_struct *task = task_event->task; |
c980d109 | 5093 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 5094 | |
67516844 JO |
5095 | if (!perf_event_task_match(event)) |
5096 | return; | |
5097 | ||
c980d109 | 5098 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 5099 | |
c980d109 | 5100 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 5101 | task_event->event_id.header.size); |
ef60777c | 5102 | if (ret) |
c980d109 | 5103 | goto out; |
60313ebe | 5104 | |
cdd6c482 IM |
5105 | task_event->event_id.pid = perf_event_pid(event, task); |
5106 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 5107 | |
cdd6c482 IM |
5108 | task_event->event_id.tid = perf_event_tid(event, task); |
5109 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 5110 | |
cdd6c482 | 5111 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 5112 | |
c980d109 ACM |
5113 | perf_event__output_id_sample(event, &handle, &sample); |
5114 | ||
60313ebe | 5115 | perf_output_end(&handle); |
c980d109 ACM |
5116 | out: |
5117 | task_event->event_id.header.size = size; | |
60313ebe PZ |
5118 | } |
5119 | ||
cdd6c482 IM |
5120 | static void perf_event_task(struct task_struct *task, |
5121 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 5122 | int new) |
60313ebe | 5123 | { |
9f498cc5 | 5124 | struct perf_task_event task_event; |
60313ebe | 5125 | |
cdd6c482 IM |
5126 | if (!atomic_read(&nr_comm_events) && |
5127 | !atomic_read(&nr_mmap_events) && | |
5128 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
5129 | return; |
5130 | ||
9f498cc5 | 5131 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
5132 | .task = task, |
5133 | .task_ctx = task_ctx, | |
cdd6c482 | 5134 | .event_id = { |
60313ebe | 5135 | .header = { |
cdd6c482 | 5136 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 5137 | .misc = 0, |
cdd6c482 | 5138 | .size = sizeof(task_event.event_id), |
60313ebe | 5139 | }, |
573402db PZ |
5140 | /* .pid */ |
5141 | /* .ppid */ | |
9f498cc5 PZ |
5142 | /* .tid */ |
5143 | /* .ptid */ | |
6f93d0a7 | 5144 | .time = perf_clock(), |
60313ebe PZ |
5145 | }, |
5146 | }; | |
5147 | ||
67516844 | 5148 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5149 | &task_event, |
5150 | task_ctx); | |
9f498cc5 PZ |
5151 | } |
5152 | ||
cdd6c482 | 5153 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5154 | { |
cdd6c482 | 5155 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5156 | } |
5157 | ||
8d1b2d93 PZ |
5158 | /* |
5159 | * comm tracking | |
5160 | */ | |
5161 | ||
5162 | struct perf_comm_event { | |
22a4f650 IM |
5163 | struct task_struct *task; |
5164 | char *comm; | |
8d1b2d93 PZ |
5165 | int comm_size; |
5166 | ||
5167 | struct { | |
5168 | struct perf_event_header header; | |
5169 | ||
5170 | u32 pid; | |
5171 | u32 tid; | |
cdd6c482 | 5172 | } event_id; |
8d1b2d93 PZ |
5173 | }; |
5174 | ||
67516844 JO |
5175 | static int perf_event_comm_match(struct perf_event *event) |
5176 | { | |
5177 | return event->attr.comm; | |
5178 | } | |
5179 | ||
cdd6c482 | 5180 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5181 | void *data) |
8d1b2d93 | 5182 | { |
52d857a8 | 5183 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5184 | struct perf_output_handle handle; |
c980d109 | 5185 | struct perf_sample_data sample; |
cdd6c482 | 5186 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5187 | int ret; |
5188 | ||
67516844 JO |
5189 | if (!perf_event_comm_match(event)) |
5190 | return; | |
5191 | ||
c980d109 ACM |
5192 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5193 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5194 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5195 | |
5196 | if (ret) | |
c980d109 | 5197 | goto out; |
8d1b2d93 | 5198 | |
cdd6c482 IM |
5199 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5200 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5201 | |
cdd6c482 | 5202 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5203 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5204 | comm_event->comm_size); |
c980d109 ACM |
5205 | |
5206 | perf_event__output_id_sample(event, &handle, &sample); | |
5207 | ||
8d1b2d93 | 5208 | perf_output_end(&handle); |
c980d109 ACM |
5209 | out: |
5210 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5211 | } |
5212 | ||
cdd6c482 | 5213 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5214 | { |
413ee3b4 | 5215 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5216 | unsigned int size; |
8d1b2d93 | 5217 | |
413ee3b4 | 5218 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5219 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5220 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5221 | |
5222 | comm_event->comm = comm; | |
5223 | comm_event->comm_size = size; | |
5224 | ||
cdd6c482 | 5225 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5226 | |
67516844 | 5227 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5228 | comm_event, |
5229 | NULL); | |
8d1b2d93 PZ |
5230 | } |
5231 | ||
82b89778 | 5232 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5233 | { |
9ee318a7 PZ |
5234 | struct perf_comm_event comm_event; |
5235 | ||
cdd6c482 | 5236 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5237 | return; |
a63eaf34 | 5238 | |
9ee318a7 | 5239 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5240 | .task = task, |
573402db PZ |
5241 | /* .comm */ |
5242 | /* .comm_size */ | |
cdd6c482 | 5243 | .event_id = { |
573402db | 5244 | .header = { |
cdd6c482 | 5245 | .type = PERF_RECORD_COMM, |
82b89778 | 5246 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5247 | /* .size */ |
5248 | }, | |
5249 | /* .pid */ | |
5250 | /* .tid */ | |
8d1b2d93 PZ |
5251 | }, |
5252 | }; | |
5253 | ||
cdd6c482 | 5254 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5255 | } |
5256 | ||
0a4a9391 PZ |
5257 | /* |
5258 | * mmap tracking | |
5259 | */ | |
5260 | ||
5261 | struct perf_mmap_event { | |
089dd79d PZ |
5262 | struct vm_area_struct *vma; |
5263 | ||
5264 | const char *file_name; | |
5265 | int file_size; | |
13d7a241 SE |
5266 | int maj, min; |
5267 | u64 ino; | |
5268 | u64 ino_generation; | |
f972eb63 | 5269 | u32 prot, flags; |
0a4a9391 PZ |
5270 | |
5271 | struct { | |
5272 | struct perf_event_header header; | |
5273 | ||
5274 | u32 pid; | |
5275 | u32 tid; | |
5276 | u64 start; | |
5277 | u64 len; | |
5278 | u64 pgoff; | |
cdd6c482 | 5279 | } event_id; |
0a4a9391 PZ |
5280 | }; |
5281 | ||
67516844 JO |
5282 | static int perf_event_mmap_match(struct perf_event *event, |
5283 | void *data) | |
5284 | { | |
5285 | struct perf_mmap_event *mmap_event = data; | |
5286 | struct vm_area_struct *vma = mmap_event->vma; | |
5287 | int executable = vma->vm_flags & VM_EXEC; | |
5288 | ||
5289 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5290 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5291 | } |
5292 | ||
cdd6c482 | 5293 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5294 | void *data) |
0a4a9391 | 5295 | { |
52d857a8 | 5296 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5297 | struct perf_output_handle handle; |
c980d109 | 5298 | struct perf_sample_data sample; |
cdd6c482 | 5299 | int size = mmap_event->event_id.header.size; |
c980d109 | 5300 | int ret; |
0a4a9391 | 5301 | |
67516844 JO |
5302 | if (!perf_event_mmap_match(event, data)) |
5303 | return; | |
5304 | ||
13d7a241 SE |
5305 | if (event->attr.mmap2) { |
5306 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5307 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5308 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5309 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5310 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
f972eb63 PZ |
5311 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
5312 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); | |
13d7a241 SE |
5313 | } |
5314 | ||
c980d109 ACM |
5315 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5316 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5317 | mmap_event->event_id.header.size); |
0a4a9391 | 5318 | if (ret) |
c980d109 | 5319 | goto out; |
0a4a9391 | 5320 | |
cdd6c482 IM |
5321 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5322 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5323 | |
cdd6c482 | 5324 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5325 | |
5326 | if (event->attr.mmap2) { | |
5327 | perf_output_put(&handle, mmap_event->maj); | |
5328 | perf_output_put(&handle, mmap_event->min); | |
5329 | perf_output_put(&handle, mmap_event->ino); | |
5330 | perf_output_put(&handle, mmap_event->ino_generation); | |
f972eb63 PZ |
5331 | perf_output_put(&handle, mmap_event->prot); |
5332 | perf_output_put(&handle, mmap_event->flags); | |
13d7a241 SE |
5333 | } |
5334 | ||
76369139 | 5335 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5336 | mmap_event->file_size); |
c980d109 ACM |
5337 | |
5338 | perf_event__output_id_sample(event, &handle, &sample); | |
5339 | ||
78d613eb | 5340 | perf_output_end(&handle); |
c980d109 ACM |
5341 | out: |
5342 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5343 | } |
5344 | ||
cdd6c482 | 5345 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5346 | { |
089dd79d PZ |
5347 | struct vm_area_struct *vma = mmap_event->vma; |
5348 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5349 | int maj = 0, min = 0; |
5350 | u64 ino = 0, gen = 0; | |
f972eb63 | 5351 | u32 prot = 0, flags = 0; |
0a4a9391 PZ |
5352 | unsigned int size; |
5353 | char tmp[16]; | |
5354 | char *buf = NULL; | |
2c42cfbf | 5355 | char *name; |
413ee3b4 | 5356 | |
0a4a9391 | 5357 | if (file) { |
13d7a241 SE |
5358 | struct inode *inode; |
5359 | dev_t dev; | |
3ea2f2b9 | 5360 | |
2c42cfbf | 5361 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5362 | if (!buf) { |
c7e548b4 ON |
5363 | name = "//enomem"; |
5364 | goto cpy_name; | |
0a4a9391 | 5365 | } |
413ee3b4 | 5366 | /* |
3ea2f2b9 | 5367 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5368 | * need to add enough zero bytes after the string to handle |
5369 | * the 64bit alignment we do later. | |
5370 | */ | |
3ea2f2b9 | 5371 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5372 | if (IS_ERR(name)) { |
c7e548b4 ON |
5373 | name = "//toolong"; |
5374 | goto cpy_name; | |
0a4a9391 | 5375 | } |
13d7a241 SE |
5376 | inode = file_inode(vma->vm_file); |
5377 | dev = inode->i_sb->s_dev; | |
5378 | ino = inode->i_ino; | |
5379 | gen = inode->i_generation; | |
5380 | maj = MAJOR(dev); | |
5381 | min = MINOR(dev); | |
f972eb63 PZ |
5382 | |
5383 | if (vma->vm_flags & VM_READ) | |
5384 | prot |= PROT_READ; | |
5385 | if (vma->vm_flags & VM_WRITE) | |
5386 | prot |= PROT_WRITE; | |
5387 | if (vma->vm_flags & VM_EXEC) | |
5388 | prot |= PROT_EXEC; | |
5389 | ||
5390 | if (vma->vm_flags & VM_MAYSHARE) | |
5391 | flags = MAP_SHARED; | |
5392 | else | |
5393 | flags = MAP_PRIVATE; | |
5394 | ||
5395 | if (vma->vm_flags & VM_DENYWRITE) | |
5396 | flags |= MAP_DENYWRITE; | |
5397 | if (vma->vm_flags & VM_MAYEXEC) | |
5398 | flags |= MAP_EXECUTABLE; | |
5399 | if (vma->vm_flags & VM_LOCKED) | |
5400 | flags |= MAP_LOCKED; | |
5401 | if (vma->vm_flags & VM_HUGETLB) | |
5402 | flags |= MAP_HUGETLB; | |
5403 | ||
c7e548b4 | 5404 | goto got_name; |
0a4a9391 | 5405 | } else { |
fbe26abe JO |
5406 | if (vma->vm_ops && vma->vm_ops->name) { |
5407 | name = (char *) vma->vm_ops->name(vma); | |
5408 | if (name) | |
5409 | goto cpy_name; | |
5410 | } | |
5411 | ||
2c42cfbf | 5412 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5413 | if (name) |
5414 | goto cpy_name; | |
089dd79d | 5415 | |
32c5fb7e | 5416 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5417 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5418 | name = "[heap]"; |
5419 | goto cpy_name; | |
32c5fb7e ON |
5420 | } |
5421 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5422 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5423 | name = "[stack]"; |
5424 | goto cpy_name; | |
089dd79d PZ |
5425 | } |
5426 | ||
c7e548b4 ON |
5427 | name = "//anon"; |
5428 | goto cpy_name; | |
0a4a9391 PZ |
5429 | } |
5430 | ||
c7e548b4 ON |
5431 | cpy_name: |
5432 | strlcpy(tmp, name, sizeof(tmp)); | |
5433 | name = tmp; | |
0a4a9391 | 5434 | got_name: |
2c42cfbf PZ |
5435 | /* |
5436 | * Since our buffer works in 8 byte units we need to align our string | |
5437 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5438 | * zero'd out to avoid leaking random bits to userspace. | |
5439 | */ | |
5440 | size = strlen(name)+1; | |
5441 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5442 | name[size++] = '\0'; | |
0a4a9391 PZ |
5443 | |
5444 | mmap_event->file_name = name; | |
5445 | mmap_event->file_size = size; | |
13d7a241 SE |
5446 | mmap_event->maj = maj; |
5447 | mmap_event->min = min; | |
5448 | mmap_event->ino = ino; | |
5449 | mmap_event->ino_generation = gen; | |
f972eb63 PZ |
5450 | mmap_event->prot = prot; |
5451 | mmap_event->flags = flags; | |
0a4a9391 | 5452 | |
2fe85427 SE |
5453 | if (!(vma->vm_flags & VM_EXEC)) |
5454 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5455 | ||
cdd6c482 | 5456 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5457 | |
67516844 | 5458 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5459 | mmap_event, |
5460 | NULL); | |
665c2142 | 5461 | |
0a4a9391 PZ |
5462 | kfree(buf); |
5463 | } | |
5464 | ||
3af9e859 | 5465 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5466 | { |
9ee318a7 PZ |
5467 | struct perf_mmap_event mmap_event; |
5468 | ||
cdd6c482 | 5469 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5470 | return; |
5471 | ||
5472 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5473 | .vma = vma, |
573402db PZ |
5474 | /* .file_name */ |
5475 | /* .file_size */ | |
cdd6c482 | 5476 | .event_id = { |
573402db | 5477 | .header = { |
cdd6c482 | 5478 | .type = PERF_RECORD_MMAP, |
39447b38 | 5479 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5480 | /* .size */ |
5481 | }, | |
5482 | /* .pid */ | |
5483 | /* .tid */ | |
089dd79d PZ |
5484 | .start = vma->vm_start, |
5485 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5486 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5487 | }, |
13d7a241 SE |
5488 | /* .maj (attr_mmap2 only) */ |
5489 | /* .min (attr_mmap2 only) */ | |
5490 | /* .ino (attr_mmap2 only) */ | |
5491 | /* .ino_generation (attr_mmap2 only) */ | |
f972eb63 PZ |
5492 | /* .prot (attr_mmap2 only) */ |
5493 | /* .flags (attr_mmap2 only) */ | |
0a4a9391 PZ |
5494 | }; |
5495 | ||
cdd6c482 | 5496 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5497 | } |
5498 | ||
a78ac325 PZ |
5499 | /* |
5500 | * IRQ throttle logging | |
5501 | */ | |
5502 | ||
cdd6c482 | 5503 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5504 | { |
5505 | struct perf_output_handle handle; | |
c980d109 | 5506 | struct perf_sample_data sample; |
a78ac325 PZ |
5507 | int ret; |
5508 | ||
5509 | struct { | |
5510 | struct perf_event_header header; | |
5511 | u64 time; | |
cca3f454 | 5512 | u64 id; |
7f453c24 | 5513 | u64 stream_id; |
a78ac325 PZ |
5514 | } throttle_event = { |
5515 | .header = { | |
cdd6c482 | 5516 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5517 | .misc = 0, |
5518 | .size = sizeof(throttle_event), | |
5519 | }, | |
def0a9b2 | 5520 | .time = perf_clock(), |
cdd6c482 IM |
5521 | .id = primary_event_id(event), |
5522 | .stream_id = event->id, | |
a78ac325 PZ |
5523 | }; |
5524 | ||
966ee4d6 | 5525 | if (enable) |
cdd6c482 | 5526 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5527 | |
c980d109 ACM |
5528 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5529 | ||
5530 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5531 | throttle_event.header.size); |
a78ac325 PZ |
5532 | if (ret) |
5533 | return; | |
5534 | ||
5535 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5536 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5537 | perf_output_end(&handle); |
5538 | } | |
5539 | ||
f6c7d5fe | 5540 | /* |
cdd6c482 | 5541 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5542 | */ |
5543 | ||
a8b0ca17 | 5544 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5545 | int throttle, struct perf_sample_data *data, |
5546 | struct pt_regs *regs) | |
f6c7d5fe | 5547 | { |
cdd6c482 IM |
5548 | int events = atomic_read(&event->event_limit); |
5549 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5550 | u64 seq; |
79f14641 PZ |
5551 | int ret = 0; |
5552 | ||
96398826 PZ |
5553 | /* |
5554 | * Non-sampling counters might still use the PMI to fold short | |
5555 | * hardware counters, ignore those. | |
5556 | */ | |
5557 | if (unlikely(!is_sampling_event(event))) | |
5558 | return 0; | |
5559 | ||
e050e3f0 SE |
5560 | seq = __this_cpu_read(perf_throttled_seq); |
5561 | if (seq != hwc->interrupts_seq) { | |
5562 | hwc->interrupts_seq = seq; | |
5563 | hwc->interrupts = 1; | |
5564 | } else { | |
5565 | hwc->interrupts++; | |
5566 | if (unlikely(throttle | |
5567 | && hwc->interrupts >= max_samples_per_tick)) { | |
5568 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5569 | hwc->interrupts = MAX_INTERRUPTS; |
5570 | perf_log_throttle(event, 0); | |
d84153d6 | 5571 | tick_nohz_full_kick(); |
a78ac325 PZ |
5572 | ret = 1; |
5573 | } | |
e050e3f0 | 5574 | } |
60db5e09 | 5575 | |
cdd6c482 | 5576 | if (event->attr.freq) { |
def0a9b2 | 5577 | u64 now = perf_clock(); |
abd50713 | 5578 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5579 | |
abd50713 | 5580 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5581 | |
abd50713 | 5582 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5583 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5584 | } |
5585 | ||
2023b359 PZ |
5586 | /* |
5587 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5588 | * events |
2023b359 PZ |
5589 | */ |
5590 | ||
cdd6c482 IM |
5591 | event->pending_kill = POLL_IN; |
5592 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5593 | ret = 1; |
cdd6c482 | 5594 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5595 | event->pending_disable = 1; |
5596 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5597 | } |
5598 | ||
453f19ee | 5599 | if (event->overflow_handler) |
a8b0ca17 | 5600 | event->overflow_handler(event, data, regs); |
453f19ee | 5601 | else |
a8b0ca17 | 5602 | perf_event_output(event, data, regs); |
453f19ee | 5603 | |
f506b3dc | 5604 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5605 | event->pending_wakeup = 1; |
5606 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5607 | } |
5608 | ||
79f14641 | 5609 | return ret; |
f6c7d5fe PZ |
5610 | } |
5611 | ||
a8b0ca17 | 5612 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5613 | struct perf_sample_data *data, |
5614 | struct pt_regs *regs) | |
850bc73f | 5615 | { |
a8b0ca17 | 5616 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5617 | } |
5618 | ||
15dbf27c | 5619 | /* |
cdd6c482 | 5620 | * Generic software event infrastructure |
15dbf27c PZ |
5621 | */ |
5622 | ||
b28ab83c PZ |
5623 | struct swevent_htable { |
5624 | struct swevent_hlist *swevent_hlist; | |
5625 | struct mutex hlist_mutex; | |
5626 | int hlist_refcount; | |
5627 | ||
5628 | /* Recursion avoidance in each contexts */ | |
5629 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
5630 | |
5631 | /* Keeps track of cpu being initialized/exited */ | |
5632 | bool online; | |
b28ab83c PZ |
5633 | }; |
5634 | ||
5635 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5636 | ||
7b4b6658 | 5637 | /* |
cdd6c482 IM |
5638 | * We directly increment event->count and keep a second value in |
5639 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5640 | * is kept in the range [-sample_period, 0] so that we can use the |
5641 | * sign as trigger. | |
5642 | */ | |
5643 | ||
ab573844 | 5644 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5645 | { |
cdd6c482 | 5646 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5647 | u64 period = hwc->last_period; |
5648 | u64 nr, offset; | |
5649 | s64 old, val; | |
5650 | ||
5651 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5652 | |
5653 | again: | |
e7850595 | 5654 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5655 | if (val < 0) |
5656 | return 0; | |
15dbf27c | 5657 | |
7b4b6658 PZ |
5658 | nr = div64_u64(period + val, period); |
5659 | offset = nr * period; | |
5660 | val -= offset; | |
e7850595 | 5661 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5662 | goto again; |
15dbf27c | 5663 | |
7b4b6658 | 5664 | return nr; |
15dbf27c PZ |
5665 | } |
5666 | ||
0cff784a | 5667 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5668 | struct perf_sample_data *data, |
5622f295 | 5669 | struct pt_regs *regs) |
15dbf27c | 5670 | { |
cdd6c482 | 5671 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5672 | int throttle = 0; |
15dbf27c | 5673 | |
0cff784a PZ |
5674 | if (!overflow) |
5675 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5676 | |
7b4b6658 PZ |
5677 | if (hwc->interrupts == MAX_INTERRUPTS) |
5678 | return; | |
15dbf27c | 5679 | |
7b4b6658 | 5680 | for (; overflow; overflow--) { |
a8b0ca17 | 5681 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5682 | data, regs)) { |
7b4b6658 PZ |
5683 | /* |
5684 | * We inhibit the overflow from happening when | |
5685 | * hwc->interrupts == MAX_INTERRUPTS. | |
5686 | */ | |
5687 | break; | |
5688 | } | |
cf450a73 | 5689 | throttle = 1; |
7b4b6658 | 5690 | } |
15dbf27c PZ |
5691 | } |
5692 | ||
a4eaf7f1 | 5693 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5694 | struct perf_sample_data *data, |
5622f295 | 5695 | struct pt_regs *regs) |
7b4b6658 | 5696 | { |
cdd6c482 | 5697 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5698 | |
e7850595 | 5699 | local64_add(nr, &event->count); |
d6d020e9 | 5700 | |
0cff784a PZ |
5701 | if (!regs) |
5702 | return; | |
5703 | ||
6c7e550f | 5704 | if (!is_sampling_event(event)) |
7b4b6658 | 5705 | return; |
d6d020e9 | 5706 | |
5d81e5cf AV |
5707 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5708 | data->period = nr; | |
5709 | return perf_swevent_overflow(event, 1, data, regs); | |
5710 | } else | |
5711 | data->period = event->hw.last_period; | |
5712 | ||
0cff784a | 5713 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5714 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5715 | |
e7850595 | 5716 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5717 | return; |
df1a132b | 5718 | |
a8b0ca17 | 5719 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5720 | } |
5721 | ||
f5ffe02e FW |
5722 | static int perf_exclude_event(struct perf_event *event, |
5723 | struct pt_regs *regs) | |
5724 | { | |
a4eaf7f1 | 5725 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5726 | return 1; |
a4eaf7f1 | 5727 | |
f5ffe02e FW |
5728 | if (regs) { |
5729 | if (event->attr.exclude_user && user_mode(regs)) | |
5730 | return 1; | |
5731 | ||
5732 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5733 | return 1; | |
5734 | } | |
5735 | ||
5736 | return 0; | |
5737 | } | |
5738 | ||
cdd6c482 | 5739 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5740 | enum perf_type_id type, |
6fb2915d LZ |
5741 | u32 event_id, |
5742 | struct perf_sample_data *data, | |
5743 | struct pt_regs *regs) | |
15dbf27c | 5744 | { |
cdd6c482 | 5745 | if (event->attr.type != type) |
a21ca2ca | 5746 | return 0; |
f5ffe02e | 5747 | |
cdd6c482 | 5748 | if (event->attr.config != event_id) |
15dbf27c PZ |
5749 | return 0; |
5750 | ||
f5ffe02e FW |
5751 | if (perf_exclude_event(event, regs)) |
5752 | return 0; | |
15dbf27c PZ |
5753 | |
5754 | return 1; | |
5755 | } | |
5756 | ||
76e1d904 FW |
5757 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5758 | { | |
5759 | u64 val = event_id | (type << 32); | |
5760 | ||
5761 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5762 | } | |
5763 | ||
49f135ed FW |
5764 | static inline struct hlist_head * |
5765 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5766 | { |
49f135ed FW |
5767 | u64 hash = swevent_hash(type, event_id); |
5768 | ||
5769 | return &hlist->heads[hash]; | |
5770 | } | |
76e1d904 | 5771 | |
49f135ed FW |
5772 | /* For the read side: events when they trigger */ |
5773 | static inline struct hlist_head * | |
b28ab83c | 5774 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5775 | { |
5776 | struct swevent_hlist *hlist; | |
76e1d904 | 5777 | |
b28ab83c | 5778 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5779 | if (!hlist) |
5780 | return NULL; | |
5781 | ||
49f135ed FW |
5782 | return __find_swevent_head(hlist, type, event_id); |
5783 | } | |
5784 | ||
5785 | /* For the event head insertion and removal in the hlist */ | |
5786 | static inline struct hlist_head * | |
b28ab83c | 5787 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5788 | { |
5789 | struct swevent_hlist *hlist; | |
5790 | u32 event_id = event->attr.config; | |
5791 | u64 type = event->attr.type; | |
5792 | ||
5793 | /* | |
5794 | * Event scheduling is always serialized against hlist allocation | |
5795 | * and release. Which makes the protected version suitable here. | |
5796 | * The context lock guarantees that. | |
5797 | */ | |
b28ab83c | 5798 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5799 | lockdep_is_held(&event->ctx->lock)); |
5800 | if (!hlist) | |
5801 | return NULL; | |
5802 | ||
5803 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5804 | } |
5805 | ||
5806 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5807 | u64 nr, |
76e1d904 FW |
5808 | struct perf_sample_data *data, |
5809 | struct pt_regs *regs) | |
15dbf27c | 5810 | { |
b28ab83c | 5811 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5812 | struct perf_event *event; |
76e1d904 | 5813 | struct hlist_head *head; |
15dbf27c | 5814 | |
76e1d904 | 5815 | rcu_read_lock(); |
b28ab83c | 5816 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5817 | if (!head) |
5818 | goto end; | |
5819 | ||
b67bfe0d | 5820 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5821 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5822 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5823 | } |
76e1d904 FW |
5824 | end: |
5825 | rcu_read_unlock(); | |
15dbf27c PZ |
5826 | } |
5827 | ||
4ed7c92d | 5828 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5829 | { |
b28ab83c | 5830 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5831 | |
b28ab83c | 5832 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5833 | } |
645e8cc0 | 5834 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5835 | |
fa9f90be | 5836 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5837 | { |
b28ab83c | 5838 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5839 | |
b28ab83c | 5840 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5841 | } |
15dbf27c | 5842 | |
a8b0ca17 | 5843 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5844 | { |
a4234bfc | 5845 | struct perf_sample_data data; |
4ed7c92d PZ |
5846 | int rctx; |
5847 | ||
1c024eca | 5848 | preempt_disable_notrace(); |
4ed7c92d PZ |
5849 | rctx = perf_swevent_get_recursion_context(); |
5850 | if (rctx < 0) | |
5851 | return; | |
a4234bfc | 5852 | |
fd0d000b | 5853 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5854 | |
a8b0ca17 | 5855 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5856 | |
5857 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5858 | preempt_enable_notrace(); |
b8e83514 PZ |
5859 | } |
5860 | ||
cdd6c482 | 5861 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5862 | { |
15dbf27c PZ |
5863 | } |
5864 | ||
a4eaf7f1 | 5865 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5866 | { |
b28ab83c | 5867 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5868 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5869 | struct hlist_head *head; |
5870 | ||
6c7e550f | 5871 | if (is_sampling_event(event)) { |
7b4b6658 | 5872 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5873 | perf_swevent_set_period(event); |
7b4b6658 | 5874 | } |
76e1d904 | 5875 | |
a4eaf7f1 PZ |
5876 | hwc->state = !(flags & PERF_EF_START); |
5877 | ||
b28ab83c | 5878 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
5879 | if (!head) { |
5880 | /* | |
5881 | * We can race with cpu hotplug code. Do not | |
5882 | * WARN if the cpu just got unplugged. | |
5883 | */ | |
5884 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 5885 | return -EINVAL; |
39af6b16 | 5886 | } |
76e1d904 FW |
5887 | |
5888 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5889 | ||
15dbf27c PZ |
5890 | return 0; |
5891 | } | |
5892 | ||
a4eaf7f1 | 5893 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5894 | { |
76e1d904 | 5895 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5896 | } |
5897 | ||
a4eaf7f1 | 5898 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5899 | { |
a4eaf7f1 | 5900 | event->hw.state = 0; |
d6d020e9 | 5901 | } |
aa9c4c0f | 5902 | |
a4eaf7f1 | 5903 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5904 | { |
a4eaf7f1 | 5905 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5906 | } |
5907 | ||
49f135ed FW |
5908 | /* Deref the hlist from the update side */ |
5909 | static inline struct swevent_hlist * | |
b28ab83c | 5910 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5911 | { |
b28ab83c PZ |
5912 | return rcu_dereference_protected(swhash->swevent_hlist, |
5913 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5914 | } |
5915 | ||
b28ab83c | 5916 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5917 | { |
b28ab83c | 5918 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5919 | |
49f135ed | 5920 | if (!hlist) |
76e1d904 FW |
5921 | return; |
5922 | ||
70691d4a | 5923 | RCU_INIT_POINTER(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5924 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5925 | } |
5926 | ||
5927 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5928 | { | |
b28ab83c | 5929 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5930 | |
b28ab83c | 5931 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5932 | |
b28ab83c PZ |
5933 | if (!--swhash->hlist_refcount) |
5934 | swevent_hlist_release(swhash); | |
76e1d904 | 5935 | |
b28ab83c | 5936 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5937 | } |
5938 | ||
5939 | static void swevent_hlist_put(struct perf_event *event) | |
5940 | { | |
5941 | int cpu; | |
5942 | ||
76e1d904 FW |
5943 | for_each_possible_cpu(cpu) |
5944 | swevent_hlist_put_cpu(event, cpu); | |
5945 | } | |
5946 | ||
5947 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5948 | { | |
b28ab83c | 5949 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5950 | int err = 0; |
5951 | ||
b28ab83c | 5952 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5953 | |
b28ab83c | 5954 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5955 | struct swevent_hlist *hlist; |
5956 | ||
5957 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5958 | if (!hlist) { | |
5959 | err = -ENOMEM; | |
5960 | goto exit; | |
5961 | } | |
b28ab83c | 5962 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5963 | } |
b28ab83c | 5964 | swhash->hlist_refcount++; |
9ed6060d | 5965 | exit: |
b28ab83c | 5966 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5967 | |
5968 | return err; | |
5969 | } | |
5970 | ||
5971 | static int swevent_hlist_get(struct perf_event *event) | |
5972 | { | |
5973 | int err; | |
5974 | int cpu, failed_cpu; | |
5975 | ||
76e1d904 FW |
5976 | get_online_cpus(); |
5977 | for_each_possible_cpu(cpu) { | |
5978 | err = swevent_hlist_get_cpu(event, cpu); | |
5979 | if (err) { | |
5980 | failed_cpu = cpu; | |
5981 | goto fail; | |
5982 | } | |
5983 | } | |
5984 | put_online_cpus(); | |
5985 | ||
5986 | return 0; | |
9ed6060d | 5987 | fail: |
76e1d904 FW |
5988 | for_each_possible_cpu(cpu) { |
5989 | if (cpu == failed_cpu) | |
5990 | break; | |
5991 | swevent_hlist_put_cpu(event, cpu); | |
5992 | } | |
5993 | ||
5994 | put_online_cpus(); | |
5995 | return err; | |
5996 | } | |
5997 | ||
c5905afb | 5998 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5999 | |
b0a873eb PZ |
6000 | static void sw_perf_event_destroy(struct perf_event *event) |
6001 | { | |
6002 | u64 event_id = event->attr.config; | |
95476b64 | 6003 | |
b0a873eb PZ |
6004 | WARN_ON(event->parent); |
6005 | ||
c5905afb | 6006 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6007 | swevent_hlist_put(event); |
6008 | } | |
6009 | ||
6010 | static int perf_swevent_init(struct perf_event *event) | |
6011 | { | |
8176cced | 6012 | u64 event_id = event->attr.config; |
b0a873eb PZ |
6013 | |
6014 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6015 | return -ENOENT; | |
6016 | ||
2481c5fa SE |
6017 | /* |
6018 | * no branch sampling for software events | |
6019 | */ | |
6020 | if (has_branch_stack(event)) | |
6021 | return -EOPNOTSUPP; | |
6022 | ||
b0a873eb PZ |
6023 | switch (event_id) { |
6024 | case PERF_COUNT_SW_CPU_CLOCK: | |
6025 | case PERF_COUNT_SW_TASK_CLOCK: | |
6026 | return -ENOENT; | |
6027 | ||
6028 | default: | |
6029 | break; | |
6030 | } | |
6031 | ||
ce677831 | 6032 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
6033 | return -ENOENT; |
6034 | ||
6035 | if (!event->parent) { | |
6036 | int err; | |
6037 | ||
6038 | err = swevent_hlist_get(event); | |
6039 | if (err) | |
6040 | return err; | |
6041 | ||
c5905afb | 6042 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6043 | event->destroy = sw_perf_event_destroy; |
6044 | } | |
6045 | ||
6046 | return 0; | |
6047 | } | |
6048 | ||
35edc2a5 PZ |
6049 | static int perf_swevent_event_idx(struct perf_event *event) |
6050 | { | |
6051 | return 0; | |
6052 | } | |
6053 | ||
b0a873eb | 6054 | static struct pmu perf_swevent = { |
89a1e187 | 6055 | .task_ctx_nr = perf_sw_context, |
95476b64 | 6056 | |
b0a873eb | 6057 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
6058 | .add = perf_swevent_add, |
6059 | .del = perf_swevent_del, | |
6060 | .start = perf_swevent_start, | |
6061 | .stop = perf_swevent_stop, | |
1c024eca | 6062 | .read = perf_swevent_read, |
35edc2a5 PZ |
6063 | |
6064 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
6065 | }; |
6066 | ||
b0a873eb PZ |
6067 | #ifdef CONFIG_EVENT_TRACING |
6068 | ||
1c024eca PZ |
6069 | static int perf_tp_filter_match(struct perf_event *event, |
6070 | struct perf_sample_data *data) | |
6071 | { | |
6072 | void *record = data->raw->data; | |
6073 | ||
6074 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
6075 | return 1; | |
6076 | return 0; | |
6077 | } | |
6078 | ||
6079 | static int perf_tp_event_match(struct perf_event *event, | |
6080 | struct perf_sample_data *data, | |
6081 | struct pt_regs *regs) | |
6082 | { | |
a0f7d0f7 FW |
6083 | if (event->hw.state & PERF_HES_STOPPED) |
6084 | return 0; | |
580d607c PZ |
6085 | /* |
6086 | * All tracepoints are from kernel-space. | |
6087 | */ | |
6088 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
6089 | return 0; |
6090 | ||
6091 | if (!perf_tp_filter_match(event, data)) | |
6092 | return 0; | |
6093 | ||
6094 | return 1; | |
6095 | } | |
6096 | ||
6097 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
6098 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
6099 | struct task_struct *task) | |
95476b64 FW |
6100 | { |
6101 | struct perf_sample_data data; | |
1c024eca | 6102 | struct perf_event *event; |
1c024eca | 6103 | |
95476b64 FW |
6104 | struct perf_raw_record raw = { |
6105 | .size = entry_size, | |
6106 | .data = record, | |
6107 | }; | |
6108 | ||
fd0d000b | 6109 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
6110 | data.raw = &raw; |
6111 | ||
b67bfe0d | 6112 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 6113 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 6114 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 6115 | } |
ecc55f84 | 6116 | |
e6dab5ff AV |
6117 | /* |
6118 | * If we got specified a target task, also iterate its context and | |
6119 | * deliver this event there too. | |
6120 | */ | |
6121 | if (task && task != current) { | |
6122 | struct perf_event_context *ctx; | |
6123 | struct trace_entry *entry = record; | |
6124 | ||
6125 | rcu_read_lock(); | |
6126 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
6127 | if (!ctx) | |
6128 | goto unlock; | |
6129 | ||
6130 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
6131 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6132 | continue; | |
6133 | if (event->attr.config != entry->type) | |
6134 | continue; | |
6135 | if (perf_tp_event_match(event, &data, regs)) | |
6136 | perf_swevent_event(event, count, &data, regs); | |
6137 | } | |
6138 | unlock: | |
6139 | rcu_read_unlock(); | |
6140 | } | |
6141 | ||
ecc55f84 | 6142 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
6143 | } |
6144 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
6145 | ||
cdd6c482 | 6146 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 6147 | { |
1c024eca | 6148 | perf_trace_destroy(event); |
e077df4f PZ |
6149 | } |
6150 | ||
b0a873eb | 6151 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 6152 | { |
76e1d904 FW |
6153 | int err; |
6154 | ||
b0a873eb PZ |
6155 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
6156 | return -ENOENT; | |
6157 | ||
2481c5fa SE |
6158 | /* |
6159 | * no branch sampling for tracepoint events | |
6160 | */ | |
6161 | if (has_branch_stack(event)) | |
6162 | return -EOPNOTSUPP; | |
6163 | ||
1c024eca PZ |
6164 | err = perf_trace_init(event); |
6165 | if (err) | |
b0a873eb | 6166 | return err; |
e077df4f | 6167 | |
cdd6c482 | 6168 | event->destroy = tp_perf_event_destroy; |
e077df4f | 6169 | |
b0a873eb PZ |
6170 | return 0; |
6171 | } | |
6172 | ||
6173 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
6174 | .task_ctx_nr = perf_sw_context, |
6175 | ||
b0a873eb | 6176 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6177 | .add = perf_trace_add, |
6178 | .del = perf_trace_del, | |
6179 | .start = perf_swevent_start, | |
6180 | .stop = perf_swevent_stop, | |
b0a873eb | 6181 | .read = perf_swevent_read, |
35edc2a5 PZ |
6182 | |
6183 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6184 | }; |
6185 | ||
6186 | static inline void perf_tp_register(void) | |
6187 | { | |
2e80a82a | 6188 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6189 | } |
6fb2915d LZ |
6190 | |
6191 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6192 | { | |
6193 | char *filter_str; | |
6194 | int ret; | |
6195 | ||
6196 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6197 | return -EINVAL; | |
6198 | ||
6199 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6200 | if (IS_ERR(filter_str)) | |
6201 | return PTR_ERR(filter_str); | |
6202 | ||
6203 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6204 | ||
6205 | kfree(filter_str); | |
6206 | return ret; | |
6207 | } | |
6208 | ||
6209 | static void perf_event_free_filter(struct perf_event *event) | |
6210 | { | |
6211 | ftrace_profile_free_filter(event); | |
6212 | } | |
6213 | ||
e077df4f | 6214 | #else |
6fb2915d | 6215 | |
b0a873eb | 6216 | static inline void perf_tp_register(void) |
e077df4f | 6217 | { |
e077df4f | 6218 | } |
6fb2915d LZ |
6219 | |
6220 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6221 | { | |
6222 | return -ENOENT; | |
6223 | } | |
6224 | ||
6225 | static void perf_event_free_filter(struct perf_event *event) | |
6226 | { | |
6227 | } | |
6228 | ||
07b139c8 | 6229 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6230 | |
24f1e32c | 6231 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6232 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6233 | { |
f5ffe02e FW |
6234 | struct perf_sample_data sample; |
6235 | struct pt_regs *regs = data; | |
6236 | ||
fd0d000b | 6237 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6238 | |
a4eaf7f1 | 6239 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6240 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6241 | } |
6242 | #endif | |
6243 | ||
b0a873eb PZ |
6244 | /* |
6245 | * hrtimer based swevent callback | |
6246 | */ | |
f29ac756 | 6247 | |
b0a873eb | 6248 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6249 | { |
b0a873eb PZ |
6250 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6251 | struct perf_sample_data data; | |
6252 | struct pt_regs *regs; | |
6253 | struct perf_event *event; | |
6254 | u64 period; | |
f29ac756 | 6255 | |
b0a873eb | 6256 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6257 | |
6258 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6259 | return HRTIMER_NORESTART; | |
6260 | ||
b0a873eb | 6261 | event->pmu->read(event); |
f344011c | 6262 | |
fd0d000b | 6263 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6264 | regs = get_irq_regs(); |
6265 | ||
6266 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6267 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6268 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6269 | ret = HRTIMER_NORESTART; |
6270 | } | |
24f1e32c | 6271 | |
b0a873eb PZ |
6272 | period = max_t(u64, 10000, event->hw.sample_period); |
6273 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6274 | |
b0a873eb | 6275 | return ret; |
f29ac756 PZ |
6276 | } |
6277 | ||
b0a873eb | 6278 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6279 | { |
b0a873eb | 6280 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6281 | s64 period; |
6282 | ||
6283 | if (!is_sampling_event(event)) | |
6284 | return; | |
f5ffe02e | 6285 | |
5d508e82 FBH |
6286 | period = local64_read(&hwc->period_left); |
6287 | if (period) { | |
6288 | if (period < 0) | |
6289 | period = 10000; | |
fa407f35 | 6290 | |
5d508e82 FBH |
6291 | local64_set(&hwc->period_left, 0); |
6292 | } else { | |
6293 | period = max_t(u64, 10000, hwc->sample_period); | |
6294 | } | |
6295 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6296 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6297 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6298 | } |
b0a873eb PZ |
6299 | |
6300 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6301 | { |
b0a873eb PZ |
6302 | struct hw_perf_event *hwc = &event->hw; |
6303 | ||
6c7e550f | 6304 | if (is_sampling_event(event)) { |
b0a873eb | 6305 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6306 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6307 | |
6308 | hrtimer_cancel(&hwc->hrtimer); | |
6309 | } | |
24f1e32c FW |
6310 | } |
6311 | ||
ba3dd36c PZ |
6312 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6313 | { | |
6314 | struct hw_perf_event *hwc = &event->hw; | |
6315 | ||
6316 | if (!is_sampling_event(event)) | |
6317 | return; | |
6318 | ||
6319 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6320 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6321 | ||
6322 | /* | |
6323 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6324 | * mapping and avoid the whole period adjust feedback stuff. | |
6325 | */ | |
6326 | if (event->attr.freq) { | |
6327 | long freq = event->attr.sample_freq; | |
6328 | ||
6329 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6330 | hwc->sample_period = event->attr.sample_period; | |
6331 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6332 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6333 | event->attr.freq = 0; |
6334 | } | |
6335 | } | |
6336 | ||
b0a873eb PZ |
6337 | /* |
6338 | * Software event: cpu wall time clock | |
6339 | */ | |
6340 | ||
6341 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6342 | { |
b0a873eb PZ |
6343 | s64 prev; |
6344 | u64 now; | |
6345 | ||
a4eaf7f1 | 6346 | now = local_clock(); |
b0a873eb PZ |
6347 | prev = local64_xchg(&event->hw.prev_count, now); |
6348 | local64_add(now - prev, &event->count); | |
24f1e32c | 6349 | } |
24f1e32c | 6350 | |
a4eaf7f1 | 6351 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6352 | { |
a4eaf7f1 | 6353 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6354 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6355 | } |
6356 | ||
a4eaf7f1 | 6357 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6358 | { |
b0a873eb PZ |
6359 | perf_swevent_cancel_hrtimer(event); |
6360 | cpu_clock_event_update(event); | |
6361 | } | |
f29ac756 | 6362 | |
a4eaf7f1 PZ |
6363 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6364 | { | |
6365 | if (flags & PERF_EF_START) | |
6366 | cpu_clock_event_start(event, flags); | |
6367 | ||
6368 | return 0; | |
6369 | } | |
6370 | ||
6371 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6372 | { | |
6373 | cpu_clock_event_stop(event, flags); | |
6374 | } | |
6375 | ||
b0a873eb PZ |
6376 | static void cpu_clock_event_read(struct perf_event *event) |
6377 | { | |
6378 | cpu_clock_event_update(event); | |
6379 | } | |
f344011c | 6380 | |
b0a873eb PZ |
6381 | static int cpu_clock_event_init(struct perf_event *event) |
6382 | { | |
6383 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6384 | return -ENOENT; | |
6385 | ||
6386 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6387 | return -ENOENT; | |
6388 | ||
2481c5fa SE |
6389 | /* |
6390 | * no branch sampling for software events | |
6391 | */ | |
6392 | if (has_branch_stack(event)) | |
6393 | return -EOPNOTSUPP; | |
6394 | ||
ba3dd36c PZ |
6395 | perf_swevent_init_hrtimer(event); |
6396 | ||
b0a873eb | 6397 | return 0; |
f29ac756 PZ |
6398 | } |
6399 | ||
b0a873eb | 6400 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6401 | .task_ctx_nr = perf_sw_context, |
6402 | ||
b0a873eb | 6403 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6404 | .add = cpu_clock_event_add, |
6405 | .del = cpu_clock_event_del, | |
6406 | .start = cpu_clock_event_start, | |
6407 | .stop = cpu_clock_event_stop, | |
b0a873eb | 6408 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
6409 | |
6410 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6411 | }; |
6412 | ||
6413 | /* | |
6414 | * Software event: task time clock | |
6415 | */ | |
6416 | ||
6417 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6418 | { |
b0a873eb PZ |
6419 | u64 prev; |
6420 | s64 delta; | |
5c92d124 | 6421 | |
b0a873eb PZ |
6422 | prev = local64_xchg(&event->hw.prev_count, now); |
6423 | delta = now - prev; | |
6424 | local64_add(delta, &event->count); | |
6425 | } | |
5c92d124 | 6426 | |
a4eaf7f1 | 6427 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6428 | { |
a4eaf7f1 | 6429 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6430 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6431 | } |
6432 | ||
a4eaf7f1 | 6433 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6434 | { |
6435 | perf_swevent_cancel_hrtimer(event); | |
6436 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6437 | } |
6438 | ||
6439 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6440 | { | |
6441 | if (flags & PERF_EF_START) | |
6442 | task_clock_event_start(event, flags); | |
b0a873eb | 6443 | |
a4eaf7f1 PZ |
6444 | return 0; |
6445 | } | |
6446 | ||
6447 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6448 | { | |
6449 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6450 | } |
6451 | ||
6452 | static void task_clock_event_read(struct perf_event *event) | |
6453 | { | |
768a06e2 PZ |
6454 | u64 now = perf_clock(); |
6455 | u64 delta = now - event->ctx->timestamp; | |
6456 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6457 | |
6458 | task_clock_event_update(event, time); | |
6459 | } | |
6460 | ||
6461 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6462 | { |
b0a873eb PZ |
6463 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6464 | return -ENOENT; | |
6465 | ||
6466 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6467 | return -ENOENT; | |
6468 | ||
2481c5fa SE |
6469 | /* |
6470 | * no branch sampling for software events | |
6471 | */ | |
6472 | if (has_branch_stack(event)) | |
6473 | return -EOPNOTSUPP; | |
6474 | ||
ba3dd36c PZ |
6475 | perf_swevent_init_hrtimer(event); |
6476 | ||
b0a873eb | 6477 | return 0; |
6fb2915d LZ |
6478 | } |
6479 | ||
b0a873eb | 6480 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6481 | .task_ctx_nr = perf_sw_context, |
6482 | ||
b0a873eb | 6483 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6484 | .add = task_clock_event_add, |
6485 | .del = task_clock_event_del, | |
6486 | .start = task_clock_event_start, | |
6487 | .stop = task_clock_event_stop, | |
b0a873eb | 6488 | .read = task_clock_event_read, |
35edc2a5 PZ |
6489 | |
6490 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6491 | }; |
6fb2915d | 6492 | |
ad5133b7 | 6493 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6494 | { |
e077df4f | 6495 | } |
6fb2915d | 6496 | |
ad5133b7 | 6497 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6498 | { |
ad5133b7 | 6499 | return 0; |
6fb2915d LZ |
6500 | } |
6501 | ||
ad5133b7 | 6502 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6503 | { |
ad5133b7 | 6504 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6505 | } |
6506 | ||
ad5133b7 PZ |
6507 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6508 | { | |
6509 | perf_pmu_enable(pmu); | |
6510 | return 0; | |
6511 | } | |
e077df4f | 6512 | |
ad5133b7 | 6513 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6514 | { |
ad5133b7 | 6515 | perf_pmu_enable(pmu); |
24f1e32c FW |
6516 | } |
6517 | ||
35edc2a5 PZ |
6518 | static int perf_event_idx_default(struct perf_event *event) |
6519 | { | |
6520 | return event->hw.idx + 1; | |
6521 | } | |
6522 | ||
8dc85d54 PZ |
6523 | /* |
6524 | * Ensures all contexts with the same task_ctx_nr have the same | |
6525 | * pmu_cpu_context too. | |
6526 | */ | |
9e317041 | 6527 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 6528 | { |
8dc85d54 | 6529 | struct pmu *pmu; |
b326e956 | 6530 | |
8dc85d54 PZ |
6531 | if (ctxn < 0) |
6532 | return NULL; | |
24f1e32c | 6533 | |
8dc85d54 PZ |
6534 | list_for_each_entry(pmu, &pmus, entry) { |
6535 | if (pmu->task_ctx_nr == ctxn) | |
6536 | return pmu->pmu_cpu_context; | |
6537 | } | |
24f1e32c | 6538 | |
8dc85d54 | 6539 | return NULL; |
24f1e32c FW |
6540 | } |
6541 | ||
51676957 | 6542 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6543 | { |
51676957 PZ |
6544 | int cpu; |
6545 | ||
6546 | for_each_possible_cpu(cpu) { | |
6547 | struct perf_cpu_context *cpuctx; | |
6548 | ||
6549 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6550 | ||
3f1f3320 PZ |
6551 | if (cpuctx->unique_pmu == old_pmu) |
6552 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6553 | } |
6554 | } | |
6555 | ||
6556 | static void free_pmu_context(struct pmu *pmu) | |
6557 | { | |
6558 | struct pmu *i; | |
f5ffe02e | 6559 | |
8dc85d54 | 6560 | mutex_lock(&pmus_lock); |
0475f9ea | 6561 | /* |
8dc85d54 | 6562 | * Like a real lame refcount. |
0475f9ea | 6563 | */ |
51676957 PZ |
6564 | list_for_each_entry(i, &pmus, entry) { |
6565 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6566 | update_pmu_context(i, pmu); | |
8dc85d54 | 6567 | goto out; |
51676957 | 6568 | } |
8dc85d54 | 6569 | } |
d6d020e9 | 6570 | |
51676957 | 6571 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6572 | out: |
6573 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6574 | } |
2e80a82a | 6575 | static struct idr pmu_idr; |
d6d020e9 | 6576 | |
abe43400 PZ |
6577 | static ssize_t |
6578 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6579 | { | |
6580 | struct pmu *pmu = dev_get_drvdata(dev); | |
6581 | ||
6582 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6583 | } | |
90826ca7 | 6584 | static DEVICE_ATTR_RO(type); |
abe43400 | 6585 | |
62b85639 SE |
6586 | static ssize_t |
6587 | perf_event_mux_interval_ms_show(struct device *dev, | |
6588 | struct device_attribute *attr, | |
6589 | char *page) | |
6590 | { | |
6591 | struct pmu *pmu = dev_get_drvdata(dev); | |
6592 | ||
6593 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6594 | } | |
6595 | ||
6596 | static ssize_t | |
6597 | perf_event_mux_interval_ms_store(struct device *dev, | |
6598 | struct device_attribute *attr, | |
6599 | const char *buf, size_t count) | |
6600 | { | |
6601 | struct pmu *pmu = dev_get_drvdata(dev); | |
6602 | int timer, cpu, ret; | |
6603 | ||
6604 | ret = kstrtoint(buf, 0, &timer); | |
6605 | if (ret) | |
6606 | return ret; | |
6607 | ||
6608 | if (timer < 1) | |
6609 | return -EINVAL; | |
6610 | ||
6611 | /* same value, noting to do */ | |
6612 | if (timer == pmu->hrtimer_interval_ms) | |
6613 | return count; | |
6614 | ||
6615 | pmu->hrtimer_interval_ms = timer; | |
6616 | ||
6617 | /* update all cpuctx for this PMU */ | |
6618 | for_each_possible_cpu(cpu) { | |
6619 | struct perf_cpu_context *cpuctx; | |
6620 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6621 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6622 | ||
6623 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6624 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6625 | } | |
6626 | ||
6627 | return count; | |
6628 | } | |
90826ca7 | 6629 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6630 | |
90826ca7 GKH |
6631 | static struct attribute *pmu_dev_attrs[] = { |
6632 | &dev_attr_type.attr, | |
6633 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6634 | NULL, | |
abe43400 | 6635 | }; |
90826ca7 | 6636 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6637 | |
6638 | static int pmu_bus_running; | |
6639 | static struct bus_type pmu_bus = { | |
6640 | .name = "event_source", | |
90826ca7 | 6641 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6642 | }; |
6643 | ||
6644 | static void pmu_dev_release(struct device *dev) | |
6645 | { | |
6646 | kfree(dev); | |
6647 | } | |
6648 | ||
6649 | static int pmu_dev_alloc(struct pmu *pmu) | |
6650 | { | |
6651 | int ret = -ENOMEM; | |
6652 | ||
6653 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6654 | if (!pmu->dev) | |
6655 | goto out; | |
6656 | ||
0c9d42ed | 6657 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6658 | device_initialize(pmu->dev); |
6659 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6660 | if (ret) | |
6661 | goto free_dev; | |
6662 | ||
6663 | dev_set_drvdata(pmu->dev, pmu); | |
6664 | pmu->dev->bus = &pmu_bus; | |
6665 | pmu->dev->release = pmu_dev_release; | |
6666 | ret = device_add(pmu->dev); | |
6667 | if (ret) | |
6668 | goto free_dev; | |
6669 | ||
6670 | out: | |
6671 | return ret; | |
6672 | ||
6673 | free_dev: | |
6674 | put_device(pmu->dev); | |
6675 | goto out; | |
6676 | } | |
6677 | ||
547e9fd7 | 6678 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6679 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6680 | |
03d8e80b | 6681 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6682 | { |
108b02cf | 6683 | int cpu, ret; |
24f1e32c | 6684 | |
b0a873eb | 6685 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6686 | ret = -ENOMEM; |
6687 | pmu->pmu_disable_count = alloc_percpu(int); | |
6688 | if (!pmu->pmu_disable_count) | |
6689 | goto unlock; | |
f29ac756 | 6690 | |
2e80a82a PZ |
6691 | pmu->type = -1; |
6692 | if (!name) | |
6693 | goto skip_type; | |
6694 | pmu->name = name; | |
6695 | ||
6696 | if (type < 0) { | |
0e9c3be2 TH |
6697 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6698 | if (type < 0) { | |
6699 | ret = type; | |
2e80a82a PZ |
6700 | goto free_pdc; |
6701 | } | |
6702 | } | |
6703 | pmu->type = type; | |
6704 | ||
abe43400 PZ |
6705 | if (pmu_bus_running) { |
6706 | ret = pmu_dev_alloc(pmu); | |
6707 | if (ret) | |
6708 | goto free_idr; | |
6709 | } | |
6710 | ||
2e80a82a | 6711 | skip_type: |
8dc85d54 PZ |
6712 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6713 | if (pmu->pmu_cpu_context) | |
6714 | goto got_cpu_context; | |
f29ac756 | 6715 | |
c4814202 | 6716 | ret = -ENOMEM; |
108b02cf PZ |
6717 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6718 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6719 | goto free_dev; |
f344011c | 6720 | |
108b02cf PZ |
6721 | for_each_possible_cpu(cpu) { |
6722 | struct perf_cpu_context *cpuctx; | |
6723 | ||
6724 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6725 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6726 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6727 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6728 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6729 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6730 | |
6731 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6732 | ||
e9d2b064 | 6733 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6734 | cpuctx->unique_pmu = pmu; |
108b02cf | 6735 | } |
76e1d904 | 6736 | |
8dc85d54 | 6737 | got_cpu_context: |
ad5133b7 PZ |
6738 | if (!pmu->start_txn) { |
6739 | if (pmu->pmu_enable) { | |
6740 | /* | |
6741 | * If we have pmu_enable/pmu_disable calls, install | |
6742 | * transaction stubs that use that to try and batch | |
6743 | * hardware accesses. | |
6744 | */ | |
6745 | pmu->start_txn = perf_pmu_start_txn; | |
6746 | pmu->commit_txn = perf_pmu_commit_txn; | |
6747 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6748 | } else { | |
6749 | pmu->start_txn = perf_pmu_nop_void; | |
6750 | pmu->commit_txn = perf_pmu_nop_int; | |
6751 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6752 | } |
5c92d124 | 6753 | } |
15dbf27c | 6754 | |
ad5133b7 PZ |
6755 | if (!pmu->pmu_enable) { |
6756 | pmu->pmu_enable = perf_pmu_nop_void; | |
6757 | pmu->pmu_disable = perf_pmu_nop_void; | |
6758 | } | |
6759 | ||
35edc2a5 PZ |
6760 | if (!pmu->event_idx) |
6761 | pmu->event_idx = perf_event_idx_default; | |
6762 | ||
b0a873eb | 6763 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6764 | ret = 0; |
6765 | unlock: | |
b0a873eb PZ |
6766 | mutex_unlock(&pmus_lock); |
6767 | ||
33696fc0 | 6768 | return ret; |
108b02cf | 6769 | |
abe43400 PZ |
6770 | free_dev: |
6771 | device_del(pmu->dev); | |
6772 | put_device(pmu->dev); | |
6773 | ||
2e80a82a PZ |
6774 | free_idr: |
6775 | if (pmu->type >= PERF_TYPE_MAX) | |
6776 | idr_remove(&pmu_idr, pmu->type); | |
6777 | ||
108b02cf PZ |
6778 | free_pdc: |
6779 | free_percpu(pmu->pmu_disable_count); | |
6780 | goto unlock; | |
f29ac756 | 6781 | } |
c464c76e | 6782 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 6783 | |
b0a873eb | 6784 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6785 | { |
b0a873eb PZ |
6786 | mutex_lock(&pmus_lock); |
6787 | list_del_rcu(&pmu->entry); | |
6788 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6789 | |
0475f9ea | 6790 | /* |
cde8e884 PZ |
6791 | * We dereference the pmu list under both SRCU and regular RCU, so |
6792 | * synchronize against both of those. | |
0475f9ea | 6793 | */ |
b0a873eb | 6794 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6795 | synchronize_rcu(); |
d6d020e9 | 6796 | |
33696fc0 | 6797 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6798 | if (pmu->type >= PERF_TYPE_MAX) |
6799 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6800 | device_del(pmu->dev); |
6801 | put_device(pmu->dev); | |
51676957 | 6802 | free_pmu_context(pmu); |
b0a873eb | 6803 | } |
c464c76e | 6804 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 6805 | |
b0a873eb PZ |
6806 | struct pmu *perf_init_event(struct perf_event *event) |
6807 | { | |
6808 | struct pmu *pmu = NULL; | |
6809 | int idx; | |
940c5b29 | 6810 | int ret; |
b0a873eb PZ |
6811 | |
6812 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6813 | |
6814 | rcu_read_lock(); | |
6815 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6816 | rcu_read_unlock(); | |
940c5b29 | 6817 | if (pmu) { |
c464c76e YZ |
6818 | if (!try_module_get(pmu->module)) { |
6819 | pmu = ERR_PTR(-ENODEV); | |
6820 | goto unlock; | |
6821 | } | |
7e5b2a01 | 6822 | event->pmu = pmu; |
940c5b29 LM |
6823 | ret = pmu->event_init(event); |
6824 | if (ret) | |
6825 | pmu = ERR_PTR(ret); | |
2e80a82a | 6826 | goto unlock; |
940c5b29 | 6827 | } |
2e80a82a | 6828 | |
b0a873eb | 6829 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
c464c76e YZ |
6830 | if (!try_module_get(pmu->module)) { |
6831 | pmu = ERR_PTR(-ENODEV); | |
6832 | goto unlock; | |
6833 | } | |
7e5b2a01 | 6834 | event->pmu = pmu; |
940c5b29 | 6835 | ret = pmu->event_init(event); |
b0a873eb | 6836 | if (!ret) |
e5f4d339 | 6837 | goto unlock; |
76e1d904 | 6838 | |
b0a873eb PZ |
6839 | if (ret != -ENOENT) { |
6840 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6841 | goto unlock; |
f344011c | 6842 | } |
5c92d124 | 6843 | } |
e5f4d339 PZ |
6844 | pmu = ERR_PTR(-ENOENT); |
6845 | unlock: | |
b0a873eb | 6846 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6847 | |
4aeb0b42 | 6848 | return pmu; |
5c92d124 IM |
6849 | } |
6850 | ||
4beb31f3 FW |
6851 | static void account_event_cpu(struct perf_event *event, int cpu) |
6852 | { | |
6853 | if (event->parent) | |
6854 | return; | |
6855 | ||
6856 | if (has_branch_stack(event)) { | |
6857 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6858 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
6859 | } | |
6860 | if (is_cgroup_event(event)) | |
6861 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
6862 | } | |
6863 | ||
766d6c07 FW |
6864 | static void account_event(struct perf_event *event) |
6865 | { | |
4beb31f3 FW |
6866 | if (event->parent) |
6867 | return; | |
6868 | ||
766d6c07 FW |
6869 | if (event->attach_state & PERF_ATTACH_TASK) |
6870 | static_key_slow_inc(&perf_sched_events.key); | |
6871 | if (event->attr.mmap || event->attr.mmap_data) | |
6872 | atomic_inc(&nr_mmap_events); | |
6873 | if (event->attr.comm) | |
6874 | atomic_inc(&nr_comm_events); | |
6875 | if (event->attr.task) | |
6876 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
6877 | if (event->attr.freq) { |
6878 | if (atomic_inc_return(&nr_freq_events) == 1) | |
6879 | tick_nohz_full_kick_all(); | |
6880 | } | |
4beb31f3 | 6881 | if (has_branch_stack(event)) |
766d6c07 | 6882 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 6883 | if (is_cgroup_event(event)) |
766d6c07 | 6884 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
6885 | |
6886 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
6887 | } |
6888 | ||
0793a61d | 6889 | /* |
cdd6c482 | 6890 | * Allocate and initialize a event structure |
0793a61d | 6891 | */ |
cdd6c482 | 6892 | static struct perf_event * |
c3f00c70 | 6893 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6894 | struct task_struct *task, |
6895 | struct perf_event *group_leader, | |
6896 | struct perf_event *parent_event, | |
4dc0da86 AK |
6897 | perf_overflow_handler_t overflow_handler, |
6898 | void *context) | |
0793a61d | 6899 | { |
51b0fe39 | 6900 | struct pmu *pmu; |
cdd6c482 IM |
6901 | struct perf_event *event; |
6902 | struct hw_perf_event *hwc; | |
90983b16 | 6903 | long err = -EINVAL; |
0793a61d | 6904 | |
66832eb4 ON |
6905 | if ((unsigned)cpu >= nr_cpu_ids) { |
6906 | if (!task || cpu != -1) | |
6907 | return ERR_PTR(-EINVAL); | |
6908 | } | |
6909 | ||
c3f00c70 | 6910 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6911 | if (!event) |
d5d2bc0d | 6912 | return ERR_PTR(-ENOMEM); |
0793a61d | 6913 | |
04289bb9 | 6914 | /* |
cdd6c482 | 6915 | * Single events are their own group leaders, with an |
04289bb9 IM |
6916 | * empty sibling list: |
6917 | */ | |
6918 | if (!group_leader) | |
cdd6c482 | 6919 | group_leader = event; |
04289bb9 | 6920 | |
cdd6c482 IM |
6921 | mutex_init(&event->child_mutex); |
6922 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6923 | |
cdd6c482 IM |
6924 | INIT_LIST_HEAD(&event->group_entry); |
6925 | INIT_LIST_HEAD(&event->event_entry); | |
6926 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 6927 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 6928 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
6929 | INIT_HLIST_NODE(&event->hlist_entry); |
6930 | ||
10c6db11 | 6931 | |
cdd6c482 | 6932 | init_waitqueue_head(&event->waitq); |
e360adbe | 6933 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6934 | |
cdd6c482 | 6935 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6936 | |
a6fa941d | 6937 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6938 | event->cpu = cpu; |
6939 | event->attr = *attr; | |
6940 | event->group_leader = group_leader; | |
6941 | event->pmu = NULL; | |
cdd6c482 | 6942 | event->oncpu = -1; |
a96bbc16 | 6943 | |
cdd6c482 | 6944 | event->parent = parent_event; |
b84fbc9f | 6945 | |
17cf22c3 | 6946 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6947 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6948 | |
cdd6c482 | 6949 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6950 | |
d580ff86 PZ |
6951 | if (task) { |
6952 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6953 | |
6954 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6955 | event->hw.tp_target = task; | |
d580ff86 PZ |
6956 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6957 | /* | |
6958 | * hw_breakpoint is a bit difficult here.. | |
6959 | */ | |
f22c1bb6 | 6960 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6961 | event->hw.bp_target = task; |
6962 | #endif | |
6963 | } | |
6964 | ||
4dc0da86 | 6965 | if (!overflow_handler && parent_event) { |
b326e956 | 6966 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6967 | context = parent_event->overflow_handler_context; |
6968 | } | |
66832eb4 | 6969 | |
b326e956 | 6970 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6971 | event->overflow_handler_context = context; |
97eaf530 | 6972 | |
0231bb53 | 6973 | perf_event__state_init(event); |
a86ed508 | 6974 | |
4aeb0b42 | 6975 | pmu = NULL; |
b8e83514 | 6976 | |
cdd6c482 | 6977 | hwc = &event->hw; |
bd2b5b12 | 6978 | hwc->sample_period = attr->sample_period; |
0d48696f | 6979 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6980 | hwc->sample_period = 1; |
eced1dfc | 6981 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6982 | |
e7850595 | 6983 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6984 | |
2023b359 | 6985 | /* |
cdd6c482 | 6986 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6987 | */ |
3dab77fb | 6988 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 6989 | goto err_ns; |
2023b359 | 6990 | |
b0a873eb | 6991 | pmu = perf_init_event(event); |
4aeb0b42 | 6992 | if (!pmu) |
90983b16 FW |
6993 | goto err_ns; |
6994 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 6995 | err = PTR_ERR(pmu); |
90983b16 | 6996 | goto err_ns; |
621a01ea | 6997 | } |
d5d2bc0d | 6998 | |
cdd6c482 | 6999 | if (!event->parent) { |
927c7a9e FW |
7000 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
7001 | err = get_callchain_buffers(); | |
90983b16 FW |
7002 | if (err) |
7003 | goto err_pmu; | |
d010b332 | 7004 | } |
f344011c | 7005 | } |
9ee318a7 | 7006 | |
cdd6c482 | 7007 | return event; |
90983b16 FW |
7008 | |
7009 | err_pmu: | |
7010 | if (event->destroy) | |
7011 | event->destroy(event); | |
c464c76e | 7012 | module_put(pmu->module); |
90983b16 FW |
7013 | err_ns: |
7014 | if (event->ns) | |
7015 | put_pid_ns(event->ns); | |
7016 | kfree(event); | |
7017 | ||
7018 | return ERR_PTR(err); | |
0793a61d TG |
7019 | } |
7020 | ||
cdd6c482 IM |
7021 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
7022 | struct perf_event_attr *attr) | |
974802ea | 7023 | { |
974802ea | 7024 | u32 size; |
cdf8073d | 7025 | int ret; |
974802ea PZ |
7026 | |
7027 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
7028 | return -EFAULT; | |
7029 | ||
7030 | /* | |
7031 | * zero the full structure, so that a short copy will be nice. | |
7032 | */ | |
7033 | memset(attr, 0, sizeof(*attr)); | |
7034 | ||
7035 | ret = get_user(size, &uattr->size); | |
7036 | if (ret) | |
7037 | return ret; | |
7038 | ||
7039 | if (size > PAGE_SIZE) /* silly large */ | |
7040 | goto err_size; | |
7041 | ||
7042 | if (!size) /* abi compat */ | |
7043 | size = PERF_ATTR_SIZE_VER0; | |
7044 | ||
7045 | if (size < PERF_ATTR_SIZE_VER0) | |
7046 | goto err_size; | |
7047 | ||
7048 | /* | |
7049 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
7050 | * ensure all the unknown bits are 0 - i.e. new |
7051 | * user-space does not rely on any kernel feature | |
7052 | * extensions we dont know about yet. | |
974802ea PZ |
7053 | */ |
7054 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
7055 | unsigned char __user *addr; |
7056 | unsigned char __user *end; | |
7057 | unsigned char val; | |
974802ea | 7058 | |
cdf8073d IS |
7059 | addr = (void __user *)uattr + sizeof(*attr); |
7060 | end = (void __user *)uattr + size; | |
974802ea | 7061 | |
cdf8073d | 7062 | for (; addr < end; addr++) { |
974802ea PZ |
7063 | ret = get_user(val, addr); |
7064 | if (ret) | |
7065 | return ret; | |
7066 | if (val) | |
7067 | goto err_size; | |
7068 | } | |
b3e62e35 | 7069 | size = sizeof(*attr); |
974802ea PZ |
7070 | } |
7071 | ||
7072 | ret = copy_from_user(attr, uattr, size); | |
7073 | if (ret) | |
7074 | return -EFAULT; | |
7075 | ||
cd757645 | 7076 | if (attr->__reserved_1) |
974802ea PZ |
7077 | return -EINVAL; |
7078 | ||
7079 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
7080 | return -EINVAL; | |
7081 | ||
7082 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
7083 | return -EINVAL; | |
7084 | ||
bce38cd5 SE |
7085 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
7086 | u64 mask = attr->branch_sample_type; | |
7087 | ||
7088 | /* only using defined bits */ | |
7089 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
7090 | return -EINVAL; | |
7091 | ||
7092 | /* at least one branch bit must be set */ | |
7093 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
7094 | return -EINVAL; | |
7095 | ||
bce38cd5 SE |
7096 | /* propagate priv level, when not set for branch */ |
7097 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
7098 | ||
7099 | /* exclude_kernel checked on syscall entry */ | |
7100 | if (!attr->exclude_kernel) | |
7101 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
7102 | ||
7103 | if (!attr->exclude_user) | |
7104 | mask |= PERF_SAMPLE_BRANCH_USER; | |
7105 | ||
7106 | if (!attr->exclude_hv) | |
7107 | mask |= PERF_SAMPLE_BRANCH_HV; | |
7108 | /* | |
7109 | * adjust user setting (for HW filter setup) | |
7110 | */ | |
7111 | attr->branch_sample_type = mask; | |
7112 | } | |
e712209a SE |
7113 | /* privileged levels capture (kernel, hv): check permissions */ |
7114 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
7115 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
7116 | return -EACCES; | |
bce38cd5 | 7117 | } |
4018994f | 7118 | |
c5ebcedb | 7119 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 7120 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
7121 | if (ret) |
7122 | return ret; | |
7123 | } | |
7124 | ||
7125 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
7126 | if (!arch_perf_have_user_stack_dump()) | |
7127 | return -ENOSYS; | |
7128 | ||
7129 | /* | |
7130 | * We have __u32 type for the size, but so far | |
7131 | * we can only use __u16 as maximum due to the | |
7132 | * __u16 sample size limit. | |
7133 | */ | |
7134 | if (attr->sample_stack_user >= USHRT_MAX) | |
7135 | ret = -EINVAL; | |
7136 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
7137 | ret = -EINVAL; | |
7138 | } | |
4018994f | 7139 | |
974802ea PZ |
7140 | out: |
7141 | return ret; | |
7142 | ||
7143 | err_size: | |
7144 | put_user(sizeof(*attr), &uattr->size); | |
7145 | ret = -E2BIG; | |
7146 | goto out; | |
7147 | } | |
7148 | ||
ac9721f3 PZ |
7149 | static int |
7150 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 7151 | { |
b69cf536 | 7152 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
7153 | int ret = -EINVAL; |
7154 | ||
ac9721f3 | 7155 | if (!output_event) |
a4be7c27 PZ |
7156 | goto set; |
7157 | ||
ac9721f3 PZ |
7158 | /* don't allow circular references */ |
7159 | if (event == output_event) | |
a4be7c27 PZ |
7160 | goto out; |
7161 | ||
0f139300 PZ |
7162 | /* |
7163 | * Don't allow cross-cpu buffers | |
7164 | */ | |
7165 | if (output_event->cpu != event->cpu) | |
7166 | goto out; | |
7167 | ||
7168 | /* | |
76369139 | 7169 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
7170 | */ |
7171 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
7172 | goto out; | |
7173 | ||
a4be7c27 | 7174 | set: |
cdd6c482 | 7175 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7176 | /* Can't redirect output if we've got an active mmap() */ |
7177 | if (atomic_read(&event->mmap_count)) | |
7178 | goto unlock; | |
a4be7c27 | 7179 | |
ac9721f3 | 7180 | if (output_event) { |
76369139 FW |
7181 | /* get the rb we want to redirect to */ |
7182 | rb = ring_buffer_get(output_event); | |
7183 | if (!rb) | |
ac9721f3 | 7184 | goto unlock; |
a4be7c27 PZ |
7185 | } |
7186 | ||
b69cf536 | 7187 | ring_buffer_attach(event, rb); |
9bb5d40c | 7188 | |
a4be7c27 | 7189 | ret = 0; |
ac9721f3 PZ |
7190 | unlock: |
7191 | mutex_unlock(&event->mmap_mutex); | |
7192 | ||
a4be7c27 | 7193 | out: |
a4be7c27 PZ |
7194 | return ret; |
7195 | } | |
7196 | ||
0793a61d | 7197 | /** |
cdd6c482 | 7198 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7199 | * |
cdd6c482 | 7200 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7201 | * @pid: target pid |
9f66a381 | 7202 | * @cpu: target cpu |
cdd6c482 | 7203 | * @group_fd: group leader event fd |
0793a61d | 7204 | */ |
cdd6c482 IM |
7205 | SYSCALL_DEFINE5(perf_event_open, |
7206 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7207 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7208 | { |
b04243ef PZ |
7209 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7210 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
7211 | struct perf_event_attr attr; |
7212 | struct perf_event_context *ctx; | |
7213 | struct file *event_file = NULL; | |
2903ff01 | 7214 | struct fd group = {NULL, 0}; |
38a81da2 | 7215 | struct task_struct *task = NULL; |
89a1e187 | 7216 | struct pmu *pmu; |
ea635c64 | 7217 | int event_fd; |
b04243ef | 7218 | int move_group = 0; |
dc86cabe | 7219 | int err; |
a21b0b35 | 7220 | int f_flags = O_RDWR; |
0793a61d | 7221 | |
2743a5b0 | 7222 | /* for future expandability... */ |
e5d1367f | 7223 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7224 | return -EINVAL; |
7225 | ||
dc86cabe IM |
7226 | err = perf_copy_attr(attr_uptr, &attr); |
7227 | if (err) | |
7228 | return err; | |
eab656ae | 7229 | |
0764771d PZ |
7230 | if (!attr.exclude_kernel) { |
7231 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7232 | return -EACCES; | |
7233 | } | |
7234 | ||
df58ab24 | 7235 | if (attr.freq) { |
cdd6c482 | 7236 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 7237 | return -EINVAL; |
0819b2e3 PZ |
7238 | } else { |
7239 | if (attr.sample_period & (1ULL << 63)) | |
7240 | return -EINVAL; | |
df58ab24 PZ |
7241 | } |
7242 | ||
e5d1367f SE |
7243 | /* |
7244 | * In cgroup mode, the pid argument is used to pass the fd | |
7245 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7246 | * designates the cpu on which to monitor threads from that | |
7247 | * cgroup. | |
7248 | */ | |
7249 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7250 | return -EINVAL; | |
7251 | ||
a21b0b35 YD |
7252 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7253 | f_flags |= O_CLOEXEC; | |
7254 | ||
7255 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7256 | if (event_fd < 0) |
7257 | return event_fd; | |
7258 | ||
ac9721f3 | 7259 | if (group_fd != -1) { |
2903ff01 AV |
7260 | err = perf_fget_light(group_fd, &group); |
7261 | if (err) | |
d14b12d7 | 7262 | goto err_fd; |
2903ff01 | 7263 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7264 | if (flags & PERF_FLAG_FD_OUTPUT) |
7265 | output_event = group_leader; | |
7266 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7267 | group_leader = NULL; | |
7268 | } | |
7269 | ||
e5d1367f | 7270 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7271 | task = find_lively_task_by_vpid(pid); |
7272 | if (IS_ERR(task)) { | |
7273 | err = PTR_ERR(task); | |
7274 | goto err_group_fd; | |
7275 | } | |
7276 | } | |
7277 | ||
1f4ee503 PZ |
7278 | if (task && group_leader && |
7279 | group_leader->attr.inherit != attr.inherit) { | |
7280 | err = -EINVAL; | |
7281 | goto err_task; | |
7282 | } | |
7283 | ||
fbfc623f YZ |
7284 | get_online_cpus(); |
7285 | ||
4dc0da86 AK |
7286 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
7287 | NULL, NULL); | |
d14b12d7 SE |
7288 | if (IS_ERR(event)) { |
7289 | err = PTR_ERR(event); | |
1f4ee503 | 7290 | goto err_cpus; |
d14b12d7 SE |
7291 | } |
7292 | ||
e5d1367f SE |
7293 | if (flags & PERF_FLAG_PID_CGROUP) { |
7294 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
7295 | if (err) { |
7296 | __free_event(event); | |
1f4ee503 | 7297 | goto err_cpus; |
766d6c07 | 7298 | } |
e5d1367f SE |
7299 | } |
7300 | ||
53b25335 VW |
7301 | if (is_sampling_event(event)) { |
7302 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
7303 | err = -ENOTSUPP; | |
7304 | goto err_alloc; | |
7305 | } | |
7306 | } | |
7307 | ||
766d6c07 FW |
7308 | account_event(event); |
7309 | ||
89a1e187 PZ |
7310 | /* |
7311 | * Special case software events and allow them to be part of | |
7312 | * any hardware group. | |
7313 | */ | |
7314 | pmu = event->pmu; | |
b04243ef PZ |
7315 | |
7316 | if (group_leader && | |
7317 | (is_software_event(event) != is_software_event(group_leader))) { | |
7318 | if (is_software_event(event)) { | |
7319 | /* | |
7320 | * If event and group_leader are not both a software | |
7321 | * event, and event is, then group leader is not. | |
7322 | * | |
7323 | * Allow the addition of software events to !software | |
7324 | * groups, this is safe because software events never | |
7325 | * fail to schedule. | |
7326 | */ | |
7327 | pmu = group_leader->pmu; | |
7328 | } else if (is_software_event(group_leader) && | |
7329 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7330 | /* | |
7331 | * In case the group is a pure software group, and we | |
7332 | * try to add a hardware event, move the whole group to | |
7333 | * the hardware context. | |
7334 | */ | |
7335 | move_group = 1; | |
7336 | } | |
7337 | } | |
89a1e187 PZ |
7338 | |
7339 | /* | |
7340 | * Get the target context (task or percpu): | |
7341 | */ | |
e2d37cd2 | 7342 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
7343 | if (IS_ERR(ctx)) { |
7344 | err = PTR_ERR(ctx); | |
c6be5a5c | 7345 | goto err_alloc; |
89a1e187 PZ |
7346 | } |
7347 | ||
fd1edb3a PZ |
7348 | if (task) { |
7349 | put_task_struct(task); | |
7350 | task = NULL; | |
7351 | } | |
7352 | ||
ccff286d | 7353 | /* |
cdd6c482 | 7354 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7355 | */ |
ac9721f3 | 7356 | if (group_leader) { |
dc86cabe | 7357 | err = -EINVAL; |
04289bb9 | 7358 | |
04289bb9 | 7359 | /* |
ccff286d IM |
7360 | * Do not allow a recursive hierarchy (this new sibling |
7361 | * becoming part of another group-sibling): | |
7362 | */ | |
7363 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7364 | goto err_context; |
ccff286d IM |
7365 | /* |
7366 | * Do not allow to attach to a group in a different | |
7367 | * task or CPU context: | |
04289bb9 | 7368 | */ |
b04243ef PZ |
7369 | if (move_group) { |
7370 | if (group_leader->ctx->type != ctx->type) | |
7371 | goto err_context; | |
7372 | } else { | |
7373 | if (group_leader->ctx != ctx) | |
7374 | goto err_context; | |
7375 | } | |
7376 | ||
3b6f9e5c PM |
7377 | /* |
7378 | * Only a group leader can be exclusive or pinned | |
7379 | */ | |
0d48696f | 7380 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7381 | goto err_context; |
ac9721f3 PZ |
7382 | } |
7383 | ||
7384 | if (output_event) { | |
7385 | err = perf_event_set_output(event, output_event); | |
7386 | if (err) | |
c3f00c70 | 7387 | goto err_context; |
ac9721f3 | 7388 | } |
0793a61d | 7389 | |
a21b0b35 YD |
7390 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7391 | f_flags); | |
ea635c64 AV |
7392 | if (IS_ERR(event_file)) { |
7393 | err = PTR_ERR(event_file); | |
c3f00c70 | 7394 | goto err_context; |
ea635c64 | 7395 | } |
9b51f66d | 7396 | |
b04243ef PZ |
7397 | if (move_group) { |
7398 | struct perf_event_context *gctx = group_leader->ctx; | |
7399 | ||
7400 | mutex_lock(&gctx->mutex); | |
46ce0fe9 | 7401 | perf_remove_from_context(group_leader, false); |
0231bb53 JO |
7402 | |
7403 | /* | |
7404 | * Removing from the context ends up with disabled | |
7405 | * event. What we want here is event in the initial | |
7406 | * startup state, ready to be add into new context. | |
7407 | */ | |
7408 | perf_event__state_init(group_leader); | |
b04243ef PZ |
7409 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7410 | group_entry) { | |
46ce0fe9 | 7411 | perf_remove_from_context(sibling, false); |
0231bb53 | 7412 | perf_event__state_init(sibling); |
b04243ef PZ |
7413 | put_ctx(gctx); |
7414 | } | |
7415 | mutex_unlock(&gctx->mutex); | |
7416 | put_ctx(gctx); | |
ea635c64 | 7417 | } |
9b51f66d | 7418 | |
ad3a37de | 7419 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 7420 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
7421 | |
7422 | if (move_group) { | |
0cda4c02 | 7423 | synchronize_rcu(); |
e2d37cd2 | 7424 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
7425 | get_ctx(ctx); |
7426 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
7427 | group_entry) { | |
e2d37cd2 | 7428 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
7429 | get_ctx(ctx); |
7430 | } | |
7431 | } | |
7432 | ||
e2d37cd2 | 7433 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7434 | perf_unpin_context(ctx); |
d859e29f | 7435 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7436 | |
fbfc623f YZ |
7437 | put_online_cpus(); |
7438 | ||
cdd6c482 | 7439 | event->owner = current; |
8882135b | 7440 | |
cdd6c482 IM |
7441 | mutex_lock(¤t->perf_event_mutex); |
7442 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7443 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7444 | |
c320c7b7 ACM |
7445 | /* |
7446 | * Precalculate sample_data sizes | |
7447 | */ | |
7448 | perf_event__header_size(event); | |
6844c09d | 7449 | perf_event__id_header_size(event); |
c320c7b7 | 7450 | |
8a49542c PZ |
7451 | /* |
7452 | * Drop the reference on the group_event after placing the | |
7453 | * new event on the sibling_list. This ensures destruction | |
7454 | * of the group leader will find the pointer to itself in | |
7455 | * perf_group_detach(). | |
7456 | */ | |
2903ff01 | 7457 | fdput(group); |
ea635c64 AV |
7458 | fd_install(event_fd, event_file); |
7459 | return event_fd; | |
0793a61d | 7460 | |
c3f00c70 | 7461 | err_context: |
fe4b04fa | 7462 | perf_unpin_context(ctx); |
ea635c64 | 7463 | put_ctx(ctx); |
c6be5a5c | 7464 | err_alloc: |
ea635c64 | 7465 | free_event(event); |
1f4ee503 | 7466 | err_cpus: |
fbfc623f | 7467 | put_online_cpus(); |
1f4ee503 | 7468 | err_task: |
e7d0bc04 PZ |
7469 | if (task) |
7470 | put_task_struct(task); | |
89a1e187 | 7471 | err_group_fd: |
2903ff01 | 7472 | fdput(group); |
ea635c64 AV |
7473 | err_fd: |
7474 | put_unused_fd(event_fd); | |
dc86cabe | 7475 | return err; |
0793a61d TG |
7476 | } |
7477 | ||
fb0459d7 AV |
7478 | /** |
7479 | * perf_event_create_kernel_counter | |
7480 | * | |
7481 | * @attr: attributes of the counter to create | |
7482 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7483 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7484 | */ |
7485 | struct perf_event * | |
7486 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7487 | struct task_struct *task, |
4dc0da86 AK |
7488 | perf_overflow_handler_t overflow_handler, |
7489 | void *context) | |
fb0459d7 | 7490 | { |
fb0459d7 | 7491 | struct perf_event_context *ctx; |
c3f00c70 | 7492 | struct perf_event *event; |
fb0459d7 | 7493 | int err; |
d859e29f | 7494 | |
fb0459d7 AV |
7495 | /* |
7496 | * Get the target context (task or percpu): | |
7497 | */ | |
d859e29f | 7498 | |
4dc0da86 AK |
7499 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7500 | overflow_handler, context); | |
c3f00c70 PZ |
7501 | if (IS_ERR(event)) { |
7502 | err = PTR_ERR(event); | |
7503 | goto err; | |
7504 | } | |
d859e29f | 7505 | |
f8697762 JO |
7506 | /* Mark owner so we could distinguish it from user events. */ |
7507 | event->owner = EVENT_OWNER_KERNEL; | |
7508 | ||
766d6c07 FW |
7509 | account_event(event); |
7510 | ||
38a81da2 | 7511 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7512 | if (IS_ERR(ctx)) { |
7513 | err = PTR_ERR(ctx); | |
c3f00c70 | 7514 | goto err_free; |
d859e29f | 7515 | } |
fb0459d7 | 7516 | |
fb0459d7 AV |
7517 | WARN_ON_ONCE(ctx->parent_ctx); |
7518 | mutex_lock(&ctx->mutex); | |
7519 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7520 | perf_unpin_context(ctx); |
fb0459d7 AV |
7521 | mutex_unlock(&ctx->mutex); |
7522 | ||
fb0459d7 AV |
7523 | return event; |
7524 | ||
c3f00c70 PZ |
7525 | err_free: |
7526 | free_event(event); | |
7527 | err: | |
c6567f64 | 7528 | return ERR_PTR(err); |
9b51f66d | 7529 | } |
fb0459d7 | 7530 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7531 | |
0cda4c02 YZ |
7532 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7533 | { | |
7534 | struct perf_event_context *src_ctx; | |
7535 | struct perf_event_context *dst_ctx; | |
7536 | struct perf_event *event, *tmp; | |
7537 | LIST_HEAD(events); | |
7538 | ||
7539 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7540 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7541 | ||
7542 | mutex_lock(&src_ctx->mutex); | |
7543 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7544 | event_entry) { | |
46ce0fe9 | 7545 | perf_remove_from_context(event, false); |
9a545de0 | 7546 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7547 | put_ctx(src_ctx); |
9886167d | 7548 | list_add(&event->migrate_entry, &events); |
0cda4c02 YZ |
7549 | } |
7550 | mutex_unlock(&src_ctx->mutex); | |
7551 | ||
7552 | synchronize_rcu(); | |
7553 | ||
7554 | mutex_lock(&dst_ctx->mutex); | |
9886167d PZ |
7555 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7556 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7557 | if (event->state >= PERF_EVENT_STATE_OFF) |
7558 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7559 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7560 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7561 | get_ctx(dst_ctx); | |
7562 | } | |
7563 | mutex_unlock(&dst_ctx->mutex); | |
7564 | } | |
7565 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7566 | ||
cdd6c482 | 7567 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7568 | struct task_struct *child) |
d859e29f | 7569 | { |
cdd6c482 | 7570 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7571 | u64 child_val; |
d859e29f | 7572 | |
cdd6c482 IM |
7573 | if (child_event->attr.inherit_stat) |
7574 | perf_event_read_event(child_event, child); | |
38b200d6 | 7575 | |
b5e58793 | 7576 | child_val = perf_event_count(child_event); |
d859e29f PM |
7577 | |
7578 | /* | |
7579 | * Add back the child's count to the parent's count: | |
7580 | */ | |
a6e6dea6 | 7581 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7582 | atomic64_add(child_event->total_time_enabled, |
7583 | &parent_event->child_total_time_enabled); | |
7584 | atomic64_add(child_event->total_time_running, | |
7585 | &parent_event->child_total_time_running); | |
d859e29f PM |
7586 | |
7587 | /* | |
cdd6c482 | 7588 | * Remove this event from the parent's list |
d859e29f | 7589 | */ |
cdd6c482 IM |
7590 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7591 | mutex_lock(&parent_event->child_mutex); | |
7592 | list_del_init(&child_event->child_list); | |
7593 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f | 7594 | |
dc633982 JO |
7595 | /* |
7596 | * Make sure user/parent get notified, that we just | |
7597 | * lost one event. | |
7598 | */ | |
7599 | perf_event_wakeup(parent_event); | |
7600 | ||
d859e29f | 7601 | /* |
cdd6c482 | 7602 | * Release the parent event, if this was the last |
d859e29f PM |
7603 | * reference to it. |
7604 | */ | |
a6fa941d | 7605 | put_event(parent_event); |
d859e29f PM |
7606 | } |
7607 | ||
9b51f66d | 7608 | static void |
cdd6c482 IM |
7609 | __perf_event_exit_task(struct perf_event *child_event, |
7610 | struct perf_event_context *child_ctx, | |
38b200d6 | 7611 | struct task_struct *child) |
9b51f66d | 7612 | { |
1903d50c PZ |
7613 | /* |
7614 | * Do not destroy the 'original' grouping; because of the context | |
7615 | * switch optimization the original events could've ended up in a | |
7616 | * random child task. | |
7617 | * | |
7618 | * If we were to destroy the original group, all group related | |
7619 | * operations would cease to function properly after this random | |
7620 | * child dies. | |
7621 | * | |
7622 | * Do destroy all inherited groups, we don't care about those | |
7623 | * and being thorough is better. | |
7624 | */ | |
7625 | perf_remove_from_context(child_event, !!child_event->parent); | |
0cc0c027 | 7626 | |
9b51f66d | 7627 | /* |
38b435b1 | 7628 | * It can happen that the parent exits first, and has events |
9b51f66d | 7629 | * that are still around due to the child reference. These |
38b435b1 | 7630 | * events need to be zapped. |
9b51f66d | 7631 | */ |
38b435b1 | 7632 | if (child_event->parent) { |
cdd6c482 IM |
7633 | sync_child_event(child_event, child); |
7634 | free_event(child_event); | |
179033b3 JO |
7635 | } else { |
7636 | child_event->state = PERF_EVENT_STATE_EXIT; | |
7637 | perf_event_wakeup(child_event); | |
4bcf349a | 7638 | } |
9b51f66d IM |
7639 | } |
7640 | ||
8dc85d54 | 7641 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7642 | { |
ebf905fc | 7643 | struct perf_event *child_event, *next; |
4a1c0f26 | 7644 | struct perf_event_context *child_ctx, *parent_ctx; |
a63eaf34 | 7645 | unsigned long flags; |
9b51f66d | 7646 | |
8dc85d54 | 7647 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7648 | perf_event_task(child, NULL, 0); |
9b51f66d | 7649 | return; |
9f498cc5 | 7650 | } |
9b51f66d | 7651 | |
a63eaf34 | 7652 | local_irq_save(flags); |
ad3a37de PM |
7653 | /* |
7654 | * We can't reschedule here because interrupts are disabled, | |
7655 | * and either child is current or it is a task that can't be | |
7656 | * scheduled, so we are now safe from rescheduling changing | |
7657 | * our context. | |
7658 | */ | |
806839b2 | 7659 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7660 | |
7661 | /* | |
7662 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7663 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7664 | * incremented the context's refcount before we do put_ctx below. |
7665 | */ | |
e625cce1 | 7666 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7667 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7668 | child->perf_event_ctxp[ctxn] = NULL; |
4a1c0f26 PZ |
7669 | |
7670 | /* | |
7671 | * In order to avoid freeing: child_ctx->parent_ctx->task | |
7672 | * under perf_event_context::lock, grab another reference. | |
7673 | */ | |
7674 | parent_ctx = child_ctx->parent_ctx; | |
7675 | if (parent_ctx) | |
7676 | get_ctx(parent_ctx); | |
7677 | ||
71a851b4 PZ |
7678 | /* |
7679 | * If this context is a clone; unclone it so it can't get | |
7680 | * swapped to another process while we're removing all | |
cdd6c482 | 7681 | * the events from it. |
71a851b4 PZ |
7682 | */ |
7683 | unclone_ctx(child_ctx); | |
5e942bb3 | 7684 | update_context_time(child_ctx); |
e625cce1 | 7685 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 | 7686 | |
4a1c0f26 PZ |
7687 | /* |
7688 | * Now that we no longer hold perf_event_context::lock, drop | |
7689 | * our extra child_ctx->parent_ctx reference. | |
7690 | */ | |
7691 | if (parent_ctx) | |
7692 | put_ctx(parent_ctx); | |
7693 | ||
9f498cc5 | 7694 | /* |
cdd6c482 IM |
7695 | * Report the task dead after unscheduling the events so that we |
7696 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7697 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7698 | */ |
cdd6c482 | 7699 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7700 | |
66fff224 PZ |
7701 | /* |
7702 | * We can recurse on the same lock type through: | |
7703 | * | |
cdd6c482 IM |
7704 | * __perf_event_exit_task() |
7705 | * sync_child_event() | |
a6fa941d AV |
7706 | * put_event() |
7707 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7708 | * |
7709 | * But since its the parent context it won't be the same instance. | |
7710 | */ | |
a0507c84 | 7711 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7712 | |
ebf905fc | 7713 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 7714 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 7715 | |
a63eaf34 PM |
7716 | mutex_unlock(&child_ctx->mutex); |
7717 | ||
7718 | put_ctx(child_ctx); | |
9b51f66d IM |
7719 | } |
7720 | ||
8dc85d54 PZ |
7721 | /* |
7722 | * When a child task exits, feed back event values to parent events. | |
7723 | */ | |
7724 | void perf_event_exit_task(struct task_struct *child) | |
7725 | { | |
8882135b | 7726 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7727 | int ctxn; |
7728 | ||
8882135b PZ |
7729 | mutex_lock(&child->perf_event_mutex); |
7730 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7731 | owner_entry) { | |
7732 | list_del_init(&event->owner_entry); | |
7733 | ||
7734 | /* | |
7735 | * Ensure the list deletion is visible before we clear | |
7736 | * the owner, closes a race against perf_release() where | |
7737 | * we need to serialize on the owner->perf_event_mutex. | |
7738 | */ | |
7739 | smp_wmb(); | |
7740 | event->owner = NULL; | |
7741 | } | |
7742 | mutex_unlock(&child->perf_event_mutex); | |
7743 | ||
8dc85d54 PZ |
7744 | for_each_task_context_nr(ctxn) |
7745 | perf_event_exit_task_context(child, ctxn); | |
7746 | } | |
7747 | ||
889ff015 FW |
7748 | static void perf_free_event(struct perf_event *event, |
7749 | struct perf_event_context *ctx) | |
7750 | { | |
7751 | struct perf_event *parent = event->parent; | |
7752 | ||
7753 | if (WARN_ON_ONCE(!parent)) | |
7754 | return; | |
7755 | ||
7756 | mutex_lock(&parent->child_mutex); | |
7757 | list_del_init(&event->child_list); | |
7758 | mutex_unlock(&parent->child_mutex); | |
7759 | ||
a6fa941d | 7760 | put_event(parent); |
889ff015 | 7761 | |
8a49542c | 7762 | perf_group_detach(event); |
889ff015 FW |
7763 | list_del_event(event, ctx); |
7764 | free_event(event); | |
7765 | } | |
7766 | ||
bbbee908 PZ |
7767 | /* |
7768 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7769 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7770 | */ |
cdd6c482 | 7771 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7772 | { |
8dc85d54 | 7773 | struct perf_event_context *ctx; |
cdd6c482 | 7774 | struct perf_event *event, *tmp; |
8dc85d54 | 7775 | int ctxn; |
bbbee908 | 7776 | |
8dc85d54 PZ |
7777 | for_each_task_context_nr(ctxn) { |
7778 | ctx = task->perf_event_ctxp[ctxn]; | |
7779 | if (!ctx) | |
7780 | continue; | |
bbbee908 | 7781 | |
8dc85d54 | 7782 | mutex_lock(&ctx->mutex); |
bbbee908 | 7783 | again: |
8dc85d54 PZ |
7784 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7785 | group_entry) | |
7786 | perf_free_event(event, ctx); | |
bbbee908 | 7787 | |
8dc85d54 PZ |
7788 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7789 | group_entry) | |
7790 | perf_free_event(event, ctx); | |
bbbee908 | 7791 | |
8dc85d54 PZ |
7792 | if (!list_empty(&ctx->pinned_groups) || |
7793 | !list_empty(&ctx->flexible_groups)) | |
7794 | goto again; | |
bbbee908 | 7795 | |
8dc85d54 | 7796 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7797 | |
8dc85d54 PZ |
7798 | put_ctx(ctx); |
7799 | } | |
889ff015 FW |
7800 | } |
7801 | ||
4e231c79 PZ |
7802 | void perf_event_delayed_put(struct task_struct *task) |
7803 | { | |
7804 | int ctxn; | |
7805 | ||
7806 | for_each_task_context_nr(ctxn) | |
7807 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7808 | } | |
7809 | ||
97dee4f3 PZ |
7810 | /* |
7811 | * inherit a event from parent task to child task: | |
7812 | */ | |
7813 | static struct perf_event * | |
7814 | inherit_event(struct perf_event *parent_event, | |
7815 | struct task_struct *parent, | |
7816 | struct perf_event_context *parent_ctx, | |
7817 | struct task_struct *child, | |
7818 | struct perf_event *group_leader, | |
7819 | struct perf_event_context *child_ctx) | |
7820 | { | |
7821 | struct perf_event *child_event; | |
cee010ec | 7822 | unsigned long flags; |
97dee4f3 PZ |
7823 | |
7824 | /* | |
7825 | * Instead of creating recursive hierarchies of events, | |
7826 | * we link inherited events back to the original parent, | |
7827 | * which has a filp for sure, which we use as the reference | |
7828 | * count: | |
7829 | */ | |
7830 | if (parent_event->parent) | |
7831 | parent_event = parent_event->parent; | |
7832 | ||
7833 | child_event = perf_event_alloc(&parent_event->attr, | |
7834 | parent_event->cpu, | |
d580ff86 | 7835 | child, |
97dee4f3 | 7836 | group_leader, parent_event, |
4dc0da86 | 7837 | NULL, NULL); |
97dee4f3 PZ |
7838 | if (IS_ERR(child_event)) |
7839 | return child_event; | |
a6fa941d | 7840 | |
fadfe7be JO |
7841 | if (is_orphaned_event(parent_event) || |
7842 | !atomic_long_inc_not_zero(&parent_event->refcount)) { | |
a6fa941d AV |
7843 | free_event(child_event); |
7844 | return NULL; | |
7845 | } | |
7846 | ||
97dee4f3 PZ |
7847 | get_ctx(child_ctx); |
7848 | ||
7849 | /* | |
7850 | * Make the child state follow the state of the parent event, | |
7851 | * not its attr.disabled bit. We hold the parent's mutex, | |
7852 | * so we won't race with perf_event_{en, dis}able_family. | |
7853 | */ | |
7854 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7855 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7856 | else | |
7857 | child_event->state = PERF_EVENT_STATE_OFF; | |
7858 | ||
7859 | if (parent_event->attr.freq) { | |
7860 | u64 sample_period = parent_event->hw.sample_period; | |
7861 | struct hw_perf_event *hwc = &child_event->hw; | |
7862 | ||
7863 | hwc->sample_period = sample_period; | |
7864 | hwc->last_period = sample_period; | |
7865 | ||
7866 | local64_set(&hwc->period_left, sample_period); | |
7867 | } | |
7868 | ||
7869 | child_event->ctx = child_ctx; | |
7870 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7871 | child_event->overflow_handler_context |
7872 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7873 | |
614b6780 TG |
7874 | /* |
7875 | * Precalculate sample_data sizes | |
7876 | */ | |
7877 | perf_event__header_size(child_event); | |
6844c09d | 7878 | perf_event__id_header_size(child_event); |
614b6780 | 7879 | |
97dee4f3 PZ |
7880 | /* |
7881 | * Link it up in the child's context: | |
7882 | */ | |
cee010ec | 7883 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7884 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7885 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7886 | |
97dee4f3 PZ |
7887 | /* |
7888 | * Link this into the parent event's child list | |
7889 | */ | |
7890 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7891 | mutex_lock(&parent_event->child_mutex); | |
7892 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7893 | mutex_unlock(&parent_event->child_mutex); | |
7894 | ||
7895 | return child_event; | |
7896 | } | |
7897 | ||
7898 | static int inherit_group(struct perf_event *parent_event, | |
7899 | struct task_struct *parent, | |
7900 | struct perf_event_context *parent_ctx, | |
7901 | struct task_struct *child, | |
7902 | struct perf_event_context *child_ctx) | |
7903 | { | |
7904 | struct perf_event *leader; | |
7905 | struct perf_event *sub; | |
7906 | struct perf_event *child_ctr; | |
7907 | ||
7908 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7909 | child, NULL, child_ctx); | |
7910 | if (IS_ERR(leader)) | |
7911 | return PTR_ERR(leader); | |
7912 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7913 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7914 | child, leader, child_ctx); | |
7915 | if (IS_ERR(child_ctr)) | |
7916 | return PTR_ERR(child_ctr); | |
7917 | } | |
7918 | return 0; | |
889ff015 FW |
7919 | } |
7920 | ||
7921 | static int | |
7922 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7923 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7924 | struct task_struct *child, int ctxn, |
889ff015 FW |
7925 | int *inherited_all) |
7926 | { | |
7927 | int ret; | |
8dc85d54 | 7928 | struct perf_event_context *child_ctx; |
889ff015 FW |
7929 | |
7930 | if (!event->attr.inherit) { | |
7931 | *inherited_all = 0; | |
7932 | return 0; | |
bbbee908 PZ |
7933 | } |
7934 | ||
fe4b04fa | 7935 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7936 | if (!child_ctx) { |
7937 | /* | |
7938 | * This is executed from the parent task context, so | |
7939 | * inherit events that have been marked for cloning. | |
7940 | * First allocate and initialize a context for the | |
7941 | * child. | |
7942 | */ | |
bbbee908 | 7943 | |
734df5ab | 7944 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7945 | if (!child_ctx) |
7946 | return -ENOMEM; | |
bbbee908 | 7947 | |
8dc85d54 | 7948 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7949 | } |
7950 | ||
7951 | ret = inherit_group(event, parent, parent_ctx, | |
7952 | child, child_ctx); | |
7953 | ||
7954 | if (ret) | |
7955 | *inherited_all = 0; | |
7956 | ||
7957 | return ret; | |
bbbee908 PZ |
7958 | } |
7959 | ||
9b51f66d | 7960 | /* |
cdd6c482 | 7961 | * Initialize the perf_event context in task_struct |
9b51f66d | 7962 | */ |
985c8dcb | 7963 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7964 | { |
889ff015 | 7965 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7966 | struct perf_event_context *cloned_ctx; |
7967 | struct perf_event *event; | |
9b51f66d | 7968 | struct task_struct *parent = current; |
564c2b21 | 7969 | int inherited_all = 1; |
dddd3379 | 7970 | unsigned long flags; |
6ab423e0 | 7971 | int ret = 0; |
9b51f66d | 7972 | |
8dc85d54 | 7973 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7974 | return 0; |
7975 | ||
ad3a37de | 7976 | /* |
25346b93 PM |
7977 | * If the parent's context is a clone, pin it so it won't get |
7978 | * swapped under us. | |
ad3a37de | 7979 | */ |
8dc85d54 | 7980 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
7981 | if (!parent_ctx) |
7982 | return 0; | |
25346b93 | 7983 | |
ad3a37de PM |
7984 | /* |
7985 | * No need to check if parent_ctx != NULL here; since we saw | |
7986 | * it non-NULL earlier, the only reason for it to become NULL | |
7987 | * is if we exit, and since we're currently in the middle of | |
7988 | * a fork we can't be exiting at the same time. | |
7989 | */ | |
ad3a37de | 7990 | |
9b51f66d IM |
7991 | /* |
7992 | * Lock the parent list. No need to lock the child - not PID | |
7993 | * hashed yet and not running, so nobody can access it. | |
7994 | */ | |
d859e29f | 7995 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7996 | |
7997 | /* | |
7998 | * We dont have to disable NMIs - we are only looking at | |
7999 | * the list, not manipulating it: | |
8000 | */ | |
889ff015 | 8001 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
8002 | ret = inherit_task_group(event, parent, parent_ctx, |
8003 | child, ctxn, &inherited_all); | |
889ff015 FW |
8004 | if (ret) |
8005 | break; | |
8006 | } | |
b93f7978 | 8007 | |
dddd3379 TG |
8008 | /* |
8009 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
8010 | * to allocations, but we need to prevent rotation because | |
8011 | * rotate_ctx() will change the list from interrupt context. | |
8012 | */ | |
8013 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
8014 | parent_ctx->rotate_disable = 1; | |
8015 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
8016 | ||
889ff015 | 8017 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
8018 | ret = inherit_task_group(event, parent, parent_ctx, |
8019 | child, ctxn, &inherited_all); | |
889ff015 | 8020 | if (ret) |
9b51f66d | 8021 | break; |
564c2b21 PM |
8022 | } |
8023 | ||
dddd3379 TG |
8024 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
8025 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 8026 | |
8dc85d54 | 8027 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 8028 | |
05cbaa28 | 8029 | if (child_ctx && inherited_all) { |
564c2b21 PM |
8030 | /* |
8031 | * Mark the child context as a clone of the parent | |
8032 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
8033 | * |
8034 | * Note that if the parent is a clone, the holding of | |
8035 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 8036 | */ |
c5ed5145 | 8037 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
8038 | if (cloned_ctx) { |
8039 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 8040 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
8041 | } else { |
8042 | child_ctx->parent_ctx = parent_ctx; | |
8043 | child_ctx->parent_gen = parent_ctx->generation; | |
8044 | } | |
8045 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
8046 | } |
8047 | ||
c5ed5145 | 8048 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 8049 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 8050 | |
25346b93 | 8051 | perf_unpin_context(parent_ctx); |
fe4b04fa | 8052 | put_ctx(parent_ctx); |
ad3a37de | 8053 | |
6ab423e0 | 8054 | return ret; |
9b51f66d IM |
8055 | } |
8056 | ||
8dc85d54 PZ |
8057 | /* |
8058 | * Initialize the perf_event context in task_struct | |
8059 | */ | |
8060 | int perf_event_init_task(struct task_struct *child) | |
8061 | { | |
8062 | int ctxn, ret; | |
8063 | ||
8550d7cb ON |
8064 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
8065 | mutex_init(&child->perf_event_mutex); | |
8066 | INIT_LIST_HEAD(&child->perf_event_list); | |
8067 | ||
8dc85d54 PZ |
8068 | for_each_task_context_nr(ctxn) { |
8069 | ret = perf_event_init_context(child, ctxn); | |
8070 | if (ret) | |
8071 | return ret; | |
8072 | } | |
8073 | ||
8074 | return 0; | |
8075 | } | |
8076 | ||
220b140b PM |
8077 | static void __init perf_event_init_all_cpus(void) |
8078 | { | |
b28ab83c | 8079 | struct swevent_htable *swhash; |
220b140b | 8080 | int cpu; |
220b140b PM |
8081 | |
8082 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
8083 | swhash = &per_cpu(swevent_htable, cpu); |
8084 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 8085 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
8086 | } |
8087 | } | |
8088 | ||
0db0628d | 8089 | static void perf_event_init_cpu(int cpu) |
0793a61d | 8090 | { |
108b02cf | 8091 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 8092 | |
b28ab83c | 8093 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8094 | swhash->online = true; |
4536e4d1 | 8095 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
8096 | struct swevent_hlist *hlist; |
8097 | ||
b28ab83c PZ |
8098 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
8099 | WARN_ON(!hlist); | |
8100 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 8101 | } |
b28ab83c | 8102 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
8103 | } |
8104 | ||
c277443c | 8105 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 8106 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 8107 | { |
e9d2b064 PZ |
8108 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
8109 | ||
8110 | WARN_ON(!irqs_disabled()); | |
8111 | ||
8112 | list_del_init(&cpuctx->rotation_list); | |
8113 | } | |
8114 | ||
108b02cf | 8115 | static void __perf_event_exit_context(void *__info) |
0793a61d | 8116 | { |
46ce0fe9 | 8117 | struct remove_event re = { .detach_group = false }; |
108b02cf | 8118 | struct perf_event_context *ctx = __info; |
0793a61d | 8119 | |
108b02cf | 8120 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 8121 | |
e3703f8c | 8122 | rcu_read_lock(); |
46ce0fe9 PZ |
8123 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
8124 | __perf_remove_from_context(&re); | |
e3703f8c | 8125 | rcu_read_unlock(); |
0793a61d | 8126 | } |
108b02cf PZ |
8127 | |
8128 | static void perf_event_exit_cpu_context(int cpu) | |
8129 | { | |
8130 | struct perf_event_context *ctx; | |
8131 | struct pmu *pmu; | |
8132 | int idx; | |
8133 | ||
8134 | idx = srcu_read_lock(&pmus_srcu); | |
8135 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 8136 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
8137 | |
8138 | mutex_lock(&ctx->mutex); | |
8139 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
8140 | mutex_unlock(&ctx->mutex); | |
8141 | } | |
8142 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
8143 | } |
8144 | ||
cdd6c482 | 8145 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 8146 | { |
b28ab83c | 8147 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 8148 | |
e3703f8c PZ |
8149 | perf_event_exit_cpu_context(cpu); |
8150 | ||
b28ab83c | 8151 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8152 | swhash->online = false; |
b28ab83c PZ |
8153 | swevent_hlist_release(swhash); |
8154 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
8155 | } |
8156 | #else | |
cdd6c482 | 8157 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
8158 | #endif |
8159 | ||
c277443c PZ |
8160 | static int |
8161 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
8162 | { | |
8163 | int cpu; | |
8164 | ||
8165 | for_each_online_cpu(cpu) | |
8166 | perf_event_exit_cpu(cpu); | |
8167 | ||
8168 | return NOTIFY_OK; | |
8169 | } | |
8170 | ||
8171 | /* | |
8172 | * Run the perf reboot notifier at the very last possible moment so that | |
8173 | * the generic watchdog code runs as long as possible. | |
8174 | */ | |
8175 | static struct notifier_block perf_reboot_notifier = { | |
8176 | .notifier_call = perf_reboot, | |
8177 | .priority = INT_MIN, | |
8178 | }; | |
8179 | ||
0db0628d | 8180 | static int |
0793a61d TG |
8181 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
8182 | { | |
8183 | unsigned int cpu = (long)hcpu; | |
8184 | ||
4536e4d1 | 8185 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
8186 | |
8187 | case CPU_UP_PREPARE: | |
5e11637e | 8188 | case CPU_DOWN_FAILED: |
cdd6c482 | 8189 | perf_event_init_cpu(cpu); |
0793a61d TG |
8190 | break; |
8191 | ||
5e11637e | 8192 | case CPU_UP_CANCELED: |
0793a61d | 8193 | case CPU_DOWN_PREPARE: |
cdd6c482 | 8194 | perf_event_exit_cpu(cpu); |
0793a61d | 8195 | break; |
0793a61d TG |
8196 | default: |
8197 | break; | |
8198 | } | |
8199 | ||
8200 | return NOTIFY_OK; | |
8201 | } | |
8202 | ||
cdd6c482 | 8203 | void __init perf_event_init(void) |
0793a61d | 8204 | { |
3c502e7a JW |
8205 | int ret; |
8206 | ||
2e80a82a PZ |
8207 | idr_init(&pmu_idr); |
8208 | ||
220b140b | 8209 | perf_event_init_all_cpus(); |
b0a873eb | 8210 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
8211 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
8212 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
8213 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
8214 | perf_tp_register(); |
8215 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 8216 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
8217 | |
8218 | ret = init_hw_breakpoint(); | |
8219 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
8220 | |
8221 | /* do not patch jump label more than once per second */ | |
8222 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
8223 | |
8224 | /* | |
8225 | * Build time assertion that we keep the data_head at the intended | |
8226 | * location. IOW, validation we got the __reserved[] size right. | |
8227 | */ | |
8228 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
8229 | != 1024); | |
0793a61d | 8230 | } |
abe43400 PZ |
8231 | |
8232 | static int __init perf_event_sysfs_init(void) | |
8233 | { | |
8234 | struct pmu *pmu; | |
8235 | int ret; | |
8236 | ||
8237 | mutex_lock(&pmus_lock); | |
8238 | ||
8239 | ret = bus_register(&pmu_bus); | |
8240 | if (ret) | |
8241 | goto unlock; | |
8242 | ||
8243 | list_for_each_entry(pmu, &pmus, entry) { | |
8244 | if (!pmu->name || pmu->type < 0) | |
8245 | continue; | |
8246 | ||
8247 | ret = pmu_dev_alloc(pmu); | |
8248 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8249 | } | |
8250 | pmu_bus_running = 1; | |
8251 | ret = 0; | |
8252 | ||
8253 | unlock: | |
8254 | mutex_unlock(&pmus_lock); | |
8255 | ||
8256 | return ret; | |
8257 | } | |
8258 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8259 | |
8260 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8261 | static struct cgroup_subsys_state * |
8262 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8263 | { |
8264 | struct perf_cgroup *jc; | |
e5d1367f | 8265 | |
1b15d055 | 8266 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8267 | if (!jc) |
8268 | return ERR_PTR(-ENOMEM); | |
8269 | ||
e5d1367f SE |
8270 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8271 | if (!jc->info) { | |
8272 | kfree(jc); | |
8273 | return ERR_PTR(-ENOMEM); | |
8274 | } | |
8275 | ||
e5d1367f SE |
8276 | return &jc->css; |
8277 | } | |
8278 | ||
eb95419b | 8279 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8280 | { |
eb95419b TH |
8281 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8282 | ||
e5d1367f SE |
8283 | free_percpu(jc->info); |
8284 | kfree(jc); | |
8285 | } | |
8286 | ||
8287 | static int __perf_cgroup_move(void *info) | |
8288 | { | |
8289 | struct task_struct *task = info; | |
8290 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8291 | return 0; | |
8292 | } | |
8293 | ||
eb95419b TH |
8294 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8295 | struct cgroup_taskset *tset) | |
e5d1367f | 8296 | { |
bb9d97b6 TH |
8297 | struct task_struct *task; |
8298 | ||
924f0d9a | 8299 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 8300 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8301 | } |
8302 | ||
eb95419b TH |
8303 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8304 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8305 | struct task_struct *task) |
e5d1367f SE |
8306 | { |
8307 | /* | |
8308 | * cgroup_exit() is called in the copy_process() failure path. | |
8309 | * Ignore this case since the task hasn't ran yet, this avoids | |
8310 | * trying to poke a half freed task state from generic code. | |
8311 | */ | |
8312 | if (!(task->flags & PF_EXITING)) | |
8313 | return; | |
8314 | ||
bb9d97b6 | 8315 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8316 | } |
8317 | ||
073219e9 | 8318 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
8319 | .css_alloc = perf_cgroup_css_alloc, |
8320 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8321 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8322 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8323 | }; |
8324 | #endif /* CONFIG_CGROUP_PERF */ |