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 | 251 | |
4a32fea9 | 252 | local_samples_len = __this_cpu_read(running_sample_length); |
6a02ad66 PZ |
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 */ | |
4a32fea9 | 274 | local_samples_len = __this_cpu_read(running_sample_length); |
14c63f17 DH |
275 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; |
276 | local_samples_len += sample_len_ns; | |
4a32fea9 | 277 | __this_cpu_write(running_sample_length, local_samples_len); |
14c63f17 DH |
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_detach_cgroup(struct perf_event *event) |
405 | { | |
4e2ba650 | 406 | css_put(&event->cgrp->css); |
e5d1367f SE |
407 | event->cgrp = NULL; |
408 | } | |
409 | ||
410 | static inline int is_cgroup_event(struct perf_event *event) | |
411 | { | |
412 | return event->cgrp != NULL; | |
413 | } | |
414 | ||
415 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
416 | { | |
417 | struct perf_cgroup_info *t; | |
418 | ||
419 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
420 | return t->time; | |
421 | } | |
422 | ||
423 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
424 | { | |
425 | struct perf_cgroup_info *info; | |
426 | u64 now; | |
427 | ||
428 | now = perf_clock(); | |
429 | ||
430 | info = this_cpu_ptr(cgrp->info); | |
431 | ||
432 | info->time += now - info->timestamp; | |
433 | info->timestamp = now; | |
434 | } | |
435 | ||
436 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
437 | { | |
438 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
439 | if (cgrp_out) | |
440 | __update_cgrp_time(cgrp_out); | |
441 | } | |
442 | ||
443 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
444 | { | |
3f7cce3c SE |
445 | struct perf_cgroup *cgrp; |
446 | ||
e5d1367f | 447 | /* |
3f7cce3c SE |
448 | * ensure we access cgroup data only when needed and |
449 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 450 | */ |
3f7cce3c | 451 | if (!is_cgroup_event(event)) |
e5d1367f SE |
452 | return; |
453 | ||
3f7cce3c SE |
454 | cgrp = perf_cgroup_from_task(current); |
455 | /* | |
456 | * Do not update time when cgroup is not active | |
457 | */ | |
458 | if (cgrp == event->cgrp) | |
459 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
460 | } |
461 | ||
462 | static inline void | |
3f7cce3c SE |
463 | perf_cgroup_set_timestamp(struct task_struct *task, |
464 | struct perf_event_context *ctx) | |
e5d1367f SE |
465 | { |
466 | struct perf_cgroup *cgrp; | |
467 | struct perf_cgroup_info *info; | |
468 | ||
3f7cce3c SE |
469 | /* |
470 | * ctx->lock held by caller | |
471 | * ensure we do not access cgroup data | |
472 | * unless we have the cgroup pinned (css_get) | |
473 | */ | |
474 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
475 | return; |
476 | ||
477 | cgrp = perf_cgroup_from_task(task); | |
478 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 479 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
480 | } |
481 | ||
482 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
483 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
484 | ||
485 | /* | |
486 | * reschedule events based on the cgroup constraint of task. | |
487 | * | |
488 | * mode SWOUT : schedule out everything | |
489 | * mode SWIN : schedule in based on cgroup for next | |
490 | */ | |
491 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
492 | { | |
493 | struct perf_cpu_context *cpuctx; | |
494 | struct pmu *pmu; | |
495 | unsigned long flags; | |
496 | ||
497 | /* | |
498 | * disable interrupts to avoid geting nr_cgroup | |
499 | * changes via __perf_event_disable(). Also | |
500 | * avoids preemption. | |
501 | */ | |
502 | local_irq_save(flags); | |
503 | ||
504 | /* | |
505 | * we reschedule only in the presence of cgroup | |
506 | * constrained events. | |
507 | */ | |
508 | rcu_read_lock(); | |
509 | ||
510 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 511 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
512 | if (cpuctx->unique_pmu != pmu) |
513 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 514 | |
e5d1367f SE |
515 | /* |
516 | * perf_cgroup_events says at least one | |
517 | * context on this CPU has cgroup events. | |
518 | * | |
519 | * ctx->nr_cgroups reports the number of cgroup | |
520 | * events for a context. | |
521 | */ | |
522 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
523 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
524 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
525 | |
526 | if (mode & PERF_CGROUP_SWOUT) { | |
527 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
528 | /* | |
529 | * must not be done before ctxswout due | |
530 | * to event_filter_match() in event_sched_out() | |
531 | */ | |
532 | cpuctx->cgrp = NULL; | |
533 | } | |
534 | ||
535 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 536 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
537 | /* |
538 | * set cgrp before ctxsw in to allow | |
539 | * event_filter_match() to not have to pass | |
540 | * task around | |
e5d1367f SE |
541 | */ |
542 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
543 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
544 | } | |
facc4307 PZ |
545 | perf_pmu_enable(cpuctx->ctx.pmu); |
546 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 547 | } |
e5d1367f SE |
548 | } |
549 | ||
550 | rcu_read_unlock(); | |
551 | ||
552 | local_irq_restore(flags); | |
553 | } | |
554 | ||
a8d757ef SE |
555 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
556 | struct task_struct *next) | |
e5d1367f | 557 | { |
a8d757ef SE |
558 | struct perf_cgroup *cgrp1; |
559 | struct perf_cgroup *cgrp2 = NULL; | |
560 | ||
561 | /* | |
562 | * we come here when we know perf_cgroup_events > 0 | |
563 | */ | |
564 | cgrp1 = perf_cgroup_from_task(task); | |
565 | ||
566 | /* | |
567 | * next is NULL when called from perf_event_enable_on_exec() | |
568 | * that will systematically cause a cgroup_switch() | |
569 | */ | |
570 | if (next) | |
571 | cgrp2 = perf_cgroup_from_task(next); | |
572 | ||
573 | /* | |
574 | * only schedule out current cgroup events if we know | |
575 | * that we are switching to a different cgroup. Otherwise, | |
576 | * do no touch the cgroup events. | |
577 | */ | |
578 | if (cgrp1 != cgrp2) | |
579 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
580 | } |
581 | ||
a8d757ef SE |
582 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
583 | struct task_struct *task) | |
e5d1367f | 584 | { |
a8d757ef SE |
585 | struct perf_cgroup *cgrp1; |
586 | struct perf_cgroup *cgrp2 = NULL; | |
587 | ||
588 | /* | |
589 | * we come here when we know perf_cgroup_events > 0 | |
590 | */ | |
591 | cgrp1 = perf_cgroup_from_task(task); | |
592 | ||
593 | /* prev can never be NULL */ | |
594 | cgrp2 = perf_cgroup_from_task(prev); | |
595 | ||
596 | /* | |
597 | * only need to schedule in cgroup events if we are changing | |
598 | * cgroup during ctxsw. Cgroup events were not scheduled | |
599 | * out of ctxsw out if that was not the case. | |
600 | */ | |
601 | if (cgrp1 != cgrp2) | |
602 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
603 | } |
604 | ||
605 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
606 | struct perf_event_attr *attr, | |
607 | struct perf_event *group_leader) | |
608 | { | |
609 | struct perf_cgroup *cgrp; | |
610 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
611 | struct fd f = fdget(fd); |
612 | int ret = 0; | |
e5d1367f | 613 | |
2903ff01 | 614 | if (!f.file) |
e5d1367f SE |
615 | return -EBADF; |
616 | ||
b583043e | 617 | css = css_tryget_online_from_dir(f.file->f_path.dentry, |
ec903c0c | 618 | &perf_event_cgrp_subsys); |
3db272c0 LZ |
619 | if (IS_ERR(css)) { |
620 | ret = PTR_ERR(css); | |
621 | goto out; | |
622 | } | |
e5d1367f SE |
623 | |
624 | cgrp = container_of(css, struct perf_cgroup, css); | |
625 | event->cgrp = cgrp; | |
626 | ||
627 | /* | |
628 | * all events in a group must monitor | |
629 | * the same cgroup because a task belongs | |
630 | * to only one perf cgroup at a time | |
631 | */ | |
632 | if (group_leader && group_leader->cgrp != cgrp) { | |
633 | perf_detach_cgroup(event); | |
634 | ret = -EINVAL; | |
e5d1367f | 635 | } |
3db272c0 | 636 | out: |
2903ff01 | 637 | fdput(f); |
e5d1367f SE |
638 | return ret; |
639 | } | |
640 | ||
641 | static inline void | |
642 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
643 | { | |
644 | struct perf_cgroup_info *t; | |
645 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
646 | event->shadow_ctx_time = now - t->timestamp; | |
647 | } | |
648 | ||
649 | static inline void | |
650 | perf_cgroup_defer_enabled(struct perf_event *event) | |
651 | { | |
652 | /* | |
653 | * when the current task's perf cgroup does not match | |
654 | * the event's, we need to remember to call the | |
655 | * perf_mark_enable() function the first time a task with | |
656 | * a matching perf cgroup is scheduled in. | |
657 | */ | |
658 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
659 | event->cgrp_defer_enabled = 1; | |
660 | } | |
661 | ||
662 | static inline void | |
663 | perf_cgroup_mark_enabled(struct perf_event *event, | |
664 | struct perf_event_context *ctx) | |
665 | { | |
666 | struct perf_event *sub; | |
667 | u64 tstamp = perf_event_time(event); | |
668 | ||
669 | if (!event->cgrp_defer_enabled) | |
670 | return; | |
671 | ||
672 | event->cgrp_defer_enabled = 0; | |
673 | ||
674 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
675 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
676 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
677 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
678 | sub->cgrp_defer_enabled = 0; | |
679 | } | |
680 | } | |
681 | } | |
682 | #else /* !CONFIG_CGROUP_PERF */ | |
683 | ||
684 | static inline bool | |
685 | perf_cgroup_match(struct perf_event *event) | |
686 | { | |
687 | return true; | |
688 | } | |
689 | ||
690 | static inline void perf_detach_cgroup(struct perf_event *event) | |
691 | {} | |
692 | ||
693 | static inline int is_cgroup_event(struct perf_event *event) | |
694 | { | |
695 | return 0; | |
696 | } | |
697 | ||
698 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
699 | { | |
700 | return 0; | |
701 | } | |
702 | ||
703 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
704 | { | |
705 | } | |
706 | ||
707 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
708 | { | |
709 | } | |
710 | ||
a8d757ef SE |
711 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
712 | struct task_struct *next) | |
e5d1367f SE |
713 | { |
714 | } | |
715 | ||
a8d757ef SE |
716 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
717 | struct task_struct *task) | |
e5d1367f SE |
718 | { |
719 | } | |
720 | ||
721 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
722 | struct perf_event_attr *attr, | |
723 | struct perf_event *group_leader) | |
724 | { | |
725 | return -EINVAL; | |
726 | } | |
727 | ||
728 | static inline void | |
3f7cce3c SE |
729 | perf_cgroup_set_timestamp(struct task_struct *task, |
730 | struct perf_event_context *ctx) | |
e5d1367f SE |
731 | { |
732 | } | |
733 | ||
734 | void | |
735 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
736 | { | |
737 | } | |
738 | ||
739 | static inline void | |
740 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
741 | { | |
742 | } | |
743 | ||
744 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
745 | { | |
746 | return 0; | |
747 | } | |
748 | ||
749 | static inline void | |
750 | perf_cgroup_defer_enabled(struct perf_event *event) | |
751 | { | |
752 | } | |
753 | ||
754 | static inline void | |
755 | perf_cgroup_mark_enabled(struct perf_event *event, | |
756 | struct perf_event_context *ctx) | |
757 | { | |
758 | } | |
759 | #endif | |
760 | ||
9e630205 SE |
761 | /* |
762 | * set default to be dependent on timer tick just | |
763 | * like original code | |
764 | */ | |
765 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
766 | /* | |
767 | * function must be called with interrupts disbled | |
768 | */ | |
769 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
770 | { | |
771 | struct perf_cpu_context *cpuctx; | |
772 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
773 | int rotations = 0; | |
774 | ||
775 | WARN_ON(!irqs_disabled()); | |
776 | ||
777 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
778 | ||
779 | rotations = perf_rotate_context(cpuctx); | |
780 | ||
781 | /* | |
782 | * arm timer if needed | |
783 | */ | |
784 | if (rotations) { | |
785 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
786 | ret = HRTIMER_RESTART; | |
787 | } | |
788 | ||
789 | return ret; | |
790 | } | |
791 | ||
792 | /* CPU is going down */ | |
793 | void perf_cpu_hrtimer_cancel(int cpu) | |
794 | { | |
795 | struct perf_cpu_context *cpuctx; | |
796 | struct pmu *pmu; | |
797 | unsigned long flags; | |
798 | ||
799 | if (WARN_ON(cpu != smp_processor_id())) | |
800 | return; | |
801 | ||
802 | local_irq_save(flags); | |
803 | ||
804 | rcu_read_lock(); | |
805 | ||
806 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
807 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
808 | ||
809 | if (pmu->task_ctx_nr == perf_sw_context) | |
810 | continue; | |
811 | ||
812 | hrtimer_cancel(&cpuctx->hrtimer); | |
813 | } | |
814 | ||
815 | rcu_read_unlock(); | |
816 | ||
817 | local_irq_restore(flags); | |
818 | } | |
819 | ||
820 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
821 | { | |
822 | struct hrtimer *hr = &cpuctx->hrtimer; | |
823 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 824 | int timer; |
9e630205 SE |
825 | |
826 | /* no multiplexing needed for SW PMU */ | |
827 | if (pmu->task_ctx_nr == perf_sw_context) | |
828 | return; | |
829 | ||
62b85639 SE |
830 | /* |
831 | * check default is sane, if not set then force to | |
832 | * default interval (1/tick) | |
833 | */ | |
834 | timer = pmu->hrtimer_interval_ms; | |
835 | if (timer < 1) | |
836 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
837 | ||
838 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
839 | |
840 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
841 | hr->function = perf_cpu_hrtimer_handler; | |
842 | } | |
843 | ||
844 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
845 | { | |
846 | struct hrtimer *hr = &cpuctx->hrtimer; | |
847 | struct pmu *pmu = cpuctx->ctx.pmu; | |
848 | ||
849 | /* not for SW PMU */ | |
850 | if (pmu->task_ctx_nr == perf_sw_context) | |
851 | return; | |
852 | ||
853 | if (hrtimer_active(hr)) | |
854 | return; | |
855 | ||
856 | if (!hrtimer_callback_running(hr)) | |
857 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
858 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
859 | } | |
860 | ||
33696fc0 | 861 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 862 | { |
33696fc0 PZ |
863 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
864 | if (!(*count)++) | |
865 | pmu->pmu_disable(pmu); | |
9e35ad38 | 866 | } |
9e35ad38 | 867 | |
33696fc0 | 868 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 869 | { |
33696fc0 PZ |
870 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
871 | if (!--(*count)) | |
872 | pmu->pmu_enable(pmu); | |
9e35ad38 | 873 | } |
9e35ad38 | 874 | |
e9d2b064 PZ |
875 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
876 | ||
877 | /* | |
878 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
879 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
880 | * disabled, while rotate_context is called from IRQ context. | |
881 | */ | |
108b02cf | 882 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 883 | { |
108b02cf | 884 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
4a32fea9 | 885 | struct list_head *head = this_cpu_ptr(&rotation_list); |
b5ab4cd5 | 886 | |
e9d2b064 | 887 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 888 | |
d84153d6 | 889 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 890 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 891 | } |
9e35ad38 | 892 | |
cdd6c482 | 893 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 894 | { |
e5289d4a | 895 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
896 | } |
897 | ||
cdd6c482 | 898 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 899 | { |
564c2b21 PM |
900 | if (atomic_dec_and_test(&ctx->refcount)) { |
901 | if (ctx->parent_ctx) | |
902 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
903 | if (ctx->task) |
904 | put_task_struct(ctx->task); | |
cb796ff3 | 905 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 906 | } |
a63eaf34 PM |
907 | } |
908 | ||
f63a8daa PZ |
909 | /* |
910 | * Because of perf_event::ctx migration in sys_perf_event_open::move_group and | |
911 | * perf_pmu_migrate_context() we need some magic. | |
912 | * | |
913 | * Those places that change perf_event::ctx will hold both | |
914 | * perf_event_ctx::mutex of the 'old' and 'new' ctx value. | |
915 | * | |
916 | * Lock ordering is by mutex address. There is one other site where | |
917 | * perf_event_context::mutex nests and that is put_event(). But remember that | |
918 | * that is a parent<->child context relation, and migration does not affect | |
919 | * children, therefore these two orderings should not interact. | |
920 | * | |
921 | * The change in perf_event::ctx does not affect children (as claimed above) | |
922 | * because the sys_perf_event_open() case will install a new event and break | |
923 | * the ctx parent<->child relation, and perf_pmu_migrate_context() is only | |
924 | * concerned with cpuctx and that doesn't have children. | |
925 | * | |
926 | * The places that change perf_event::ctx will issue: | |
927 | * | |
928 | * perf_remove_from_context(); | |
929 | * synchronize_rcu(); | |
930 | * perf_install_in_context(); | |
931 | * | |
932 | * to affect the change. The remove_from_context() + synchronize_rcu() should | |
933 | * quiesce the event, after which we can install it in the new location. This | |
934 | * means that only external vectors (perf_fops, prctl) can perturb the event | |
935 | * while in transit. Therefore all such accessors should also acquire | |
936 | * perf_event_context::mutex to serialize against this. | |
937 | * | |
938 | * However; because event->ctx can change while we're waiting to acquire | |
939 | * ctx->mutex we must be careful and use the below perf_event_ctx_lock() | |
940 | * function. | |
941 | * | |
942 | * Lock order: | |
943 | * task_struct::perf_event_mutex | |
944 | * perf_event_context::mutex | |
945 | * perf_event_context::lock | |
946 | * perf_event::child_mutex; | |
947 | * perf_event::mmap_mutex | |
948 | * mmap_sem | |
949 | */ | |
a83fe28e PZ |
950 | static struct perf_event_context * |
951 | perf_event_ctx_lock_nested(struct perf_event *event, int nesting) | |
f63a8daa PZ |
952 | { |
953 | struct perf_event_context *ctx; | |
954 | ||
955 | again: | |
956 | rcu_read_lock(); | |
957 | ctx = ACCESS_ONCE(event->ctx); | |
958 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
959 | rcu_read_unlock(); | |
960 | goto again; | |
961 | } | |
962 | rcu_read_unlock(); | |
963 | ||
a83fe28e | 964 | mutex_lock_nested(&ctx->mutex, nesting); |
f63a8daa PZ |
965 | if (event->ctx != ctx) { |
966 | mutex_unlock(&ctx->mutex); | |
967 | put_ctx(ctx); | |
968 | goto again; | |
969 | } | |
970 | ||
971 | return ctx; | |
972 | } | |
973 | ||
a83fe28e PZ |
974 | static inline struct perf_event_context * |
975 | perf_event_ctx_lock(struct perf_event *event) | |
976 | { | |
977 | return perf_event_ctx_lock_nested(event, 0); | |
978 | } | |
979 | ||
f63a8daa PZ |
980 | static void perf_event_ctx_unlock(struct perf_event *event, |
981 | struct perf_event_context *ctx) | |
982 | { | |
983 | mutex_unlock(&ctx->mutex); | |
984 | put_ctx(ctx); | |
985 | } | |
986 | ||
211de6eb PZ |
987 | /* |
988 | * This must be done under the ctx->lock, such as to serialize against | |
989 | * context_equiv(), therefore we cannot call put_ctx() since that might end up | |
990 | * calling scheduler related locks and ctx->lock nests inside those. | |
991 | */ | |
992 | static __must_check struct perf_event_context * | |
993 | unclone_ctx(struct perf_event_context *ctx) | |
71a851b4 | 994 | { |
211de6eb PZ |
995 | struct perf_event_context *parent_ctx = ctx->parent_ctx; |
996 | ||
997 | lockdep_assert_held(&ctx->lock); | |
998 | ||
999 | if (parent_ctx) | |
71a851b4 | 1000 | ctx->parent_ctx = NULL; |
5a3126d4 | 1001 | ctx->generation++; |
211de6eb PZ |
1002 | |
1003 | return parent_ctx; | |
71a851b4 PZ |
1004 | } |
1005 | ||
6844c09d ACM |
1006 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
1007 | { | |
1008 | /* | |
1009 | * only top level events have the pid namespace they were created in | |
1010 | */ | |
1011 | if (event->parent) | |
1012 | event = event->parent; | |
1013 | ||
1014 | return task_tgid_nr_ns(p, event->ns); | |
1015 | } | |
1016 | ||
1017 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
1018 | { | |
1019 | /* | |
1020 | * only top level events have the pid namespace they were created in | |
1021 | */ | |
1022 | if (event->parent) | |
1023 | event = event->parent; | |
1024 | ||
1025 | return task_pid_nr_ns(p, event->ns); | |
1026 | } | |
1027 | ||
7f453c24 | 1028 | /* |
cdd6c482 | 1029 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
1030 | * to userspace. |
1031 | */ | |
cdd6c482 | 1032 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 1033 | { |
cdd6c482 | 1034 | u64 id = event->id; |
7f453c24 | 1035 | |
cdd6c482 IM |
1036 | if (event->parent) |
1037 | id = event->parent->id; | |
7f453c24 PZ |
1038 | |
1039 | return id; | |
1040 | } | |
1041 | ||
25346b93 | 1042 | /* |
cdd6c482 | 1043 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
1044 | * This has to cope with with the fact that until it is locked, |
1045 | * the context could get moved to another task. | |
1046 | */ | |
cdd6c482 | 1047 | static struct perf_event_context * |
8dc85d54 | 1048 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 1049 | { |
cdd6c482 | 1050 | struct perf_event_context *ctx; |
25346b93 | 1051 | |
9ed6060d | 1052 | retry: |
058ebd0e PZ |
1053 | /* |
1054 | * One of the few rules of preemptible RCU is that one cannot do | |
1055 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
1056 | * part of the read side critical section was preemptible -- see | |
1057 | * rcu_read_unlock_special(). | |
1058 | * | |
1059 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
1060 | * side critical section is non-preemptible. | |
1061 | */ | |
1062 | preempt_disable(); | |
1063 | rcu_read_lock(); | |
8dc85d54 | 1064 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
1065 | if (ctx) { |
1066 | /* | |
1067 | * If this context is a clone of another, it might | |
1068 | * get swapped for another underneath us by | |
cdd6c482 | 1069 | * perf_event_task_sched_out, though the |
25346b93 PM |
1070 | * rcu_read_lock() protects us from any context |
1071 | * getting freed. Lock the context and check if it | |
1072 | * got swapped before we could get the lock, and retry | |
1073 | * if so. If we locked the right context, then it | |
1074 | * can't get swapped on us any more. | |
1075 | */ | |
e625cce1 | 1076 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 1077 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 1078 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
1079 | rcu_read_unlock(); |
1080 | preempt_enable(); | |
25346b93 PM |
1081 | goto retry; |
1082 | } | |
b49a9e7e PZ |
1083 | |
1084 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 1085 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1086 | ctx = NULL; |
1087 | } | |
25346b93 PM |
1088 | } |
1089 | rcu_read_unlock(); | |
058ebd0e | 1090 | preempt_enable(); |
25346b93 PM |
1091 | return ctx; |
1092 | } | |
1093 | ||
1094 | /* | |
1095 | * Get the context for a task and increment its pin_count so it | |
1096 | * can't get swapped to another task. This also increments its | |
1097 | * reference count so that the context can't get freed. | |
1098 | */ | |
8dc85d54 PZ |
1099 | static struct perf_event_context * |
1100 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1101 | { |
cdd6c482 | 1102 | struct perf_event_context *ctx; |
25346b93 PM |
1103 | unsigned long flags; |
1104 | ||
8dc85d54 | 1105 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1106 | if (ctx) { |
1107 | ++ctx->pin_count; | |
e625cce1 | 1108 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1109 | } |
1110 | return ctx; | |
1111 | } | |
1112 | ||
cdd6c482 | 1113 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1114 | { |
1115 | unsigned long flags; | |
1116 | ||
e625cce1 | 1117 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1118 | --ctx->pin_count; |
e625cce1 | 1119 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1120 | } |
1121 | ||
f67218c3 PZ |
1122 | /* |
1123 | * Update the record of the current time in a context. | |
1124 | */ | |
1125 | static void update_context_time(struct perf_event_context *ctx) | |
1126 | { | |
1127 | u64 now = perf_clock(); | |
1128 | ||
1129 | ctx->time += now - ctx->timestamp; | |
1130 | ctx->timestamp = now; | |
1131 | } | |
1132 | ||
4158755d SE |
1133 | static u64 perf_event_time(struct perf_event *event) |
1134 | { | |
1135 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1136 | |
1137 | if (is_cgroup_event(event)) | |
1138 | return perf_cgroup_event_time(event); | |
1139 | ||
4158755d SE |
1140 | return ctx ? ctx->time : 0; |
1141 | } | |
1142 | ||
f67218c3 PZ |
1143 | /* |
1144 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1145 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1146 | */ |
1147 | static void update_event_times(struct perf_event *event) | |
1148 | { | |
1149 | struct perf_event_context *ctx = event->ctx; | |
1150 | u64 run_end; | |
1151 | ||
1152 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1153 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1154 | return; | |
e5d1367f SE |
1155 | /* |
1156 | * in cgroup mode, time_enabled represents | |
1157 | * the time the event was enabled AND active | |
1158 | * tasks were in the monitored cgroup. This is | |
1159 | * independent of the activity of the context as | |
1160 | * there may be a mix of cgroup and non-cgroup events. | |
1161 | * | |
1162 | * That is why we treat cgroup events differently | |
1163 | * here. | |
1164 | */ | |
1165 | if (is_cgroup_event(event)) | |
46cd6a7f | 1166 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1167 | else if (ctx->is_active) |
1168 | run_end = ctx->time; | |
acd1d7c1 PZ |
1169 | else |
1170 | run_end = event->tstamp_stopped; | |
1171 | ||
1172 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1173 | |
1174 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1175 | run_end = event->tstamp_stopped; | |
1176 | else | |
4158755d | 1177 | run_end = perf_event_time(event); |
f67218c3 PZ |
1178 | |
1179 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1180 | |
f67218c3 PZ |
1181 | } |
1182 | ||
96c21a46 PZ |
1183 | /* |
1184 | * Update total_time_enabled and total_time_running for all events in a group. | |
1185 | */ | |
1186 | static void update_group_times(struct perf_event *leader) | |
1187 | { | |
1188 | struct perf_event *event; | |
1189 | ||
1190 | update_event_times(leader); | |
1191 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1192 | update_event_times(event); | |
1193 | } | |
1194 | ||
889ff015 FW |
1195 | static struct list_head * |
1196 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1197 | { | |
1198 | if (event->attr.pinned) | |
1199 | return &ctx->pinned_groups; | |
1200 | else | |
1201 | return &ctx->flexible_groups; | |
1202 | } | |
1203 | ||
fccc714b | 1204 | /* |
cdd6c482 | 1205 | * Add a event from the lists for its context. |
fccc714b PZ |
1206 | * Must be called with ctx->mutex and ctx->lock held. |
1207 | */ | |
04289bb9 | 1208 | static void |
cdd6c482 | 1209 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1210 | { |
8a49542c PZ |
1211 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1212 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1213 | |
1214 | /* | |
8a49542c PZ |
1215 | * If we're a stand alone event or group leader, we go to the context |
1216 | * list, group events are kept attached to the group so that | |
1217 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1218 | */ |
8a49542c | 1219 | if (event->group_leader == event) { |
889ff015 FW |
1220 | struct list_head *list; |
1221 | ||
d6f962b5 FW |
1222 | if (is_software_event(event)) |
1223 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1224 | ||
889ff015 FW |
1225 | list = ctx_group_list(event, ctx); |
1226 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1227 | } |
592903cd | 1228 | |
08309379 | 1229 | if (is_cgroup_event(event)) |
e5d1367f | 1230 | ctx->nr_cgroups++; |
e5d1367f | 1231 | |
d010b332 SE |
1232 | if (has_branch_stack(event)) |
1233 | ctx->nr_branch_stack++; | |
1234 | ||
cdd6c482 | 1235 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1236 | if (!ctx->nr_events) |
108b02cf | 1237 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1238 | ctx->nr_events++; |
1239 | if (event->attr.inherit_stat) | |
bfbd3381 | 1240 | ctx->nr_stat++; |
5a3126d4 PZ |
1241 | |
1242 | ctx->generation++; | |
04289bb9 IM |
1243 | } |
1244 | ||
0231bb53 JO |
1245 | /* |
1246 | * Initialize event state based on the perf_event_attr::disabled. | |
1247 | */ | |
1248 | static inline void perf_event__state_init(struct perf_event *event) | |
1249 | { | |
1250 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1251 | PERF_EVENT_STATE_INACTIVE; | |
1252 | } | |
1253 | ||
c320c7b7 ACM |
1254 | /* |
1255 | * Called at perf_event creation and when events are attached/detached from a | |
1256 | * group. | |
1257 | */ | |
1258 | static void perf_event__read_size(struct perf_event *event) | |
1259 | { | |
1260 | int entry = sizeof(u64); /* value */ | |
1261 | int size = 0; | |
1262 | int nr = 1; | |
1263 | ||
1264 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1265 | size += sizeof(u64); | |
1266 | ||
1267 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1268 | size += sizeof(u64); | |
1269 | ||
1270 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1271 | entry += sizeof(u64); | |
1272 | ||
1273 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1274 | nr += event->group_leader->nr_siblings; | |
1275 | size += sizeof(u64); | |
1276 | } | |
1277 | ||
1278 | size += entry * nr; | |
1279 | event->read_size = size; | |
1280 | } | |
1281 | ||
1282 | static void perf_event__header_size(struct perf_event *event) | |
1283 | { | |
1284 | struct perf_sample_data *data; | |
1285 | u64 sample_type = event->attr.sample_type; | |
1286 | u16 size = 0; | |
1287 | ||
1288 | perf_event__read_size(event); | |
1289 | ||
1290 | if (sample_type & PERF_SAMPLE_IP) | |
1291 | size += sizeof(data->ip); | |
1292 | ||
6844c09d ACM |
1293 | if (sample_type & PERF_SAMPLE_ADDR) |
1294 | size += sizeof(data->addr); | |
1295 | ||
1296 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1297 | size += sizeof(data->period); | |
1298 | ||
c3feedf2 AK |
1299 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1300 | size += sizeof(data->weight); | |
1301 | ||
6844c09d ACM |
1302 | if (sample_type & PERF_SAMPLE_READ) |
1303 | size += event->read_size; | |
1304 | ||
d6be9ad6 SE |
1305 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1306 | size += sizeof(data->data_src.val); | |
1307 | ||
fdfbbd07 AK |
1308 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1309 | size += sizeof(data->txn); | |
1310 | ||
6844c09d ACM |
1311 | event->header_size = size; |
1312 | } | |
1313 | ||
1314 | static void perf_event__id_header_size(struct perf_event *event) | |
1315 | { | |
1316 | struct perf_sample_data *data; | |
1317 | u64 sample_type = event->attr.sample_type; | |
1318 | u16 size = 0; | |
1319 | ||
c320c7b7 ACM |
1320 | if (sample_type & PERF_SAMPLE_TID) |
1321 | size += sizeof(data->tid_entry); | |
1322 | ||
1323 | if (sample_type & PERF_SAMPLE_TIME) | |
1324 | size += sizeof(data->time); | |
1325 | ||
ff3d527c AH |
1326 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1327 | size += sizeof(data->id); | |
1328 | ||
c320c7b7 ACM |
1329 | if (sample_type & PERF_SAMPLE_ID) |
1330 | size += sizeof(data->id); | |
1331 | ||
1332 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1333 | size += sizeof(data->stream_id); | |
1334 | ||
1335 | if (sample_type & PERF_SAMPLE_CPU) | |
1336 | size += sizeof(data->cpu_entry); | |
1337 | ||
6844c09d | 1338 | event->id_header_size = size; |
c320c7b7 ACM |
1339 | } |
1340 | ||
8a49542c PZ |
1341 | static void perf_group_attach(struct perf_event *event) |
1342 | { | |
c320c7b7 | 1343 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1344 | |
74c3337c PZ |
1345 | /* |
1346 | * We can have double attach due to group movement in perf_event_open. | |
1347 | */ | |
1348 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1349 | return; | |
1350 | ||
8a49542c PZ |
1351 | event->attach_state |= PERF_ATTACH_GROUP; |
1352 | ||
1353 | if (group_leader == event) | |
1354 | return; | |
1355 | ||
652884fe PZ |
1356 | WARN_ON_ONCE(group_leader->ctx != event->ctx); |
1357 | ||
8a49542c PZ |
1358 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
1359 | !is_software_event(event)) | |
1360 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1361 | ||
1362 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1363 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1364 | |
1365 | perf_event__header_size(group_leader); | |
1366 | ||
1367 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1368 | perf_event__header_size(pos); | |
8a49542c PZ |
1369 | } |
1370 | ||
a63eaf34 | 1371 | /* |
cdd6c482 | 1372 | * Remove a event from the lists for its context. |
fccc714b | 1373 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1374 | */ |
04289bb9 | 1375 | static void |
cdd6c482 | 1376 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1377 | { |
68cacd29 | 1378 | struct perf_cpu_context *cpuctx; |
652884fe PZ |
1379 | |
1380 | WARN_ON_ONCE(event->ctx != ctx); | |
1381 | lockdep_assert_held(&ctx->lock); | |
1382 | ||
8a49542c PZ |
1383 | /* |
1384 | * We can have double detach due to exit/hot-unplug + close. | |
1385 | */ | |
1386 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1387 | return; |
8a49542c PZ |
1388 | |
1389 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1390 | ||
68cacd29 | 1391 | if (is_cgroup_event(event)) { |
e5d1367f | 1392 | ctx->nr_cgroups--; |
68cacd29 SE |
1393 | cpuctx = __get_cpu_context(ctx); |
1394 | /* | |
1395 | * if there are no more cgroup events | |
1396 | * then cler cgrp to avoid stale pointer | |
1397 | * in update_cgrp_time_from_cpuctx() | |
1398 | */ | |
1399 | if (!ctx->nr_cgroups) | |
1400 | cpuctx->cgrp = NULL; | |
1401 | } | |
e5d1367f | 1402 | |
d010b332 SE |
1403 | if (has_branch_stack(event)) |
1404 | ctx->nr_branch_stack--; | |
1405 | ||
cdd6c482 IM |
1406 | ctx->nr_events--; |
1407 | if (event->attr.inherit_stat) | |
bfbd3381 | 1408 | ctx->nr_stat--; |
8bc20959 | 1409 | |
cdd6c482 | 1410 | list_del_rcu(&event->event_entry); |
04289bb9 | 1411 | |
8a49542c PZ |
1412 | if (event->group_leader == event) |
1413 | list_del_init(&event->group_entry); | |
5c148194 | 1414 | |
96c21a46 | 1415 | update_group_times(event); |
b2e74a26 SE |
1416 | |
1417 | /* | |
1418 | * If event was in error state, then keep it | |
1419 | * that way, otherwise bogus counts will be | |
1420 | * returned on read(). The only way to get out | |
1421 | * of error state is by explicit re-enabling | |
1422 | * of the event | |
1423 | */ | |
1424 | if (event->state > PERF_EVENT_STATE_OFF) | |
1425 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1426 | |
1427 | ctx->generation++; | |
050735b0 PZ |
1428 | } |
1429 | ||
8a49542c | 1430 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1431 | { |
1432 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1433 | struct list_head *list = NULL; |
1434 | ||
1435 | /* | |
1436 | * We can have double detach due to exit/hot-unplug + close. | |
1437 | */ | |
1438 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1439 | return; | |
1440 | ||
1441 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1442 | ||
1443 | /* | |
1444 | * If this is a sibling, remove it from its group. | |
1445 | */ | |
1446 | if (event->group_leader != event) { | |
1447 | list_del_init(&event->group_entry); | |
1448 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1449 | goto out; |
8a49542c PZ |
1450 | } |
1451 | ||
1452 | if (!list_empty(&event->group_entry)) | |
1453 | list = &event->group_entry; | |
2e2af50b | 1454 | |
04289bb9 | 1455 | /* |
cdd6c482 IM |
1456 | * If this was a group event with sibling events then |
1457 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1458 | * to whatever list we are on. |
04289bb9 | 1459 | */ |
cdd6c482 | 1460 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1461 | if (list) |
1462 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1463 | sibling->group_leader = sibling; |
d6f962b5 FW |
1464 | |
1465 | /* Inherit group flags from the previous leader */ | |
1466 | sibling->group_flags = event->group_flags; | |
652884fe PZ |
1467 | |
1468 | WARN_ON_ONCE(sibling->ctx != event->ctx); | |
04289bb9 | 1469 | } |
c320c7b7 ACM |
1470 | |
1471 | out: | |
1472 | perf_event__header_size(event->group_leader); | |
1473 | ||
1474 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1475 | perf_event__header_size(tmp); | |
04289bb9 IM |
1476 | } |
1477 | ||
fadfe7be JO |
1478 | /* |
1479 | * User event without the task. | |
1480 | */ | |
1481 | static bool is_orphaned_event(struct perf_event *event) | |
1482 | { | |
1483 | return event && !is_kernel_event(event) && !event->owner; | |
1484 | } | |
1485 | ||
1486 | /* | |
1487 | * Event has a parent but parent's task finished and it's | |
1488 | * alive only because of children holding refference. | |
1489 | */ | |
1490 | static bool is_orphaned_child(struct perf_event *event) | |
1491 | { | |
1492 | return is_orphaned_event(event->parent); | |
1493 | } | |
1494 | ||
1495 | static void orphans_remove_work(struct work_struct *work); | |
1496 | ||
1497 | static void schedule_orphans_remove(struct perf_event_context *ctx) | |
1498 | { | |
1499 | if (!ctx->task || ctx->orphans_remove_sched || !perf_wq) | |
1500 | return; | |
1501 | ||
1502 | if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) { | |
1503 | get_ctx(ctx); | |
1504 | ctx->orphans_remove_sched = true; | |
1505 | } | |
1506 | } | |
1507 | ||
1508 | static int __init perf_workqueue_init(void) | |
1509 | { | |
1510 | perf_wq = create_singlethread_workqueue("perf"); | |
1511 | WARN(!perf_wq, "failed to create perf workqueue\n"); | |
1512 | return perf_wq ? 0 : -1; | |
1513 | } | |
1514 | ||
1515 | core_initcall(perf_workqueue_init); | |
1516 | ||
fa66f07a SE |
1517 | static inline int |
1518 | event_filter_match(struct perf_event *event) | |
1519 | { | |
e5d1367f SE |
1520 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1521 | && perf_cgroup_match(event); | |
fa66f07a SE |
1522 | } |
1523 | ||
9ffcfa6f SE |
1524 | static void |
1525 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1526 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1527 | struct perf_event_context *ctx) |
3b6f9e5c | 1528 | { |
4158755d | 1529 | u64 tstamp = perf_event_time(event); |
fa66f07a | 1530 | u64 delta; |
652884fe PZ |
1531 | |
1532 | WARN_ON_ONCE(event->ctx != ctx); | |
1533 | lockdep_assert_held(&ctx->lock); | |
1534 | ||
fa66f07a SE |
1535 | /* |
1536 | * An event which could not be activated because of | |
1537 | * filter mismatch still needs to have its timings | |
1538 | * maintained, otherwise bogus information is return | |
1539 | * via read() for time_enabled, time_running: | |
1540 | */ | |
1541 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1542 | && !event_filter_match(event)) { | |
e5d1367f | 1543 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1544 | event->tstamp_running += delta; |
4158755d | 1545 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1546 | } |
1547 | ||
cdd6c482 | 1548 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1549 | return; |
3b6f9e5c | 1550 | |
44377277 AS |
1551 | perf_pmu_disable(event->pmu); |
1552 | ||
cdd6c482 IM |
1553 | event->state = PERF_EVENT_STATE_INACTIVE; |
1554 | if (event->pending_disable) { | |
1555 | event->pending_disable = 0; | |
1556 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1557 | } |
4158755d | 1558 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1559 | event->pmu->del(event, 0); |
cdd6c482 | 1560 | event->oncpu = -1; |
3b6f9e5c | 1561 | |
cdd6c482 | 1562 | if (!is_software_event(event)) |
3b6f9e5c PM |
1563 | cpuctx->active_oncpu--; |
1564 | ctx->nr_active--; | |
0f5a2601 PZ |
1565 | if (event->attr.freq && event->attr.sample_freq) |
1566 | ctx->nr_freq--; | |
cdd6c482 | 1567 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1568 | cpuctx->exclusive = 0; |
44377277 | 1569 | |
fadfe7be JO |
1570 | if (is_orphaned_child(event)) |
1571 | schedule_orphans_remove(ctx); | |
1572 | ||
44377277 | 1573 | perf_pmu_enable(event->pmu); |
3b6f9e5c PM |
1574 | } |
1575 | ||
d859e29f | 1576 | static void |
cdd6c482 | 1577 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1578 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1579 | struct perf_event_context *ctx) |
d859e29f | 1580 | { |
cdd6c482 | 1581 | struct perf_event *event; |
fa66f07a | 1582 | int state = group_event->state; |
d859e29f | 1583 | |
cdd6c482 | 1584 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1585 | |
1586 | /* | |
1587 | * Schedule out siblings (if any): | |
1588 | */ | |
cdd6c482 IM |
1589 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1590 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1591 | |
fa66f07a | 1592 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1593 | cpuctx->exclusive = 0; |
1594 | } | |
1595 | ||
46ce0fe9 PZ |
1596 | struct remove_event { |
1597 | struct perf_event *event; | |
1598 | bool detach_group; | |
1599 | }; | |
1600 | ||
0793a61d | 1601 | /* |
cdd6c482 | 1602 | * Cross CPU call to remove a performance event |
0793a61d | 1603 | * |
cdd6c482 | 1604 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1605 | * remove it from the context list. |
1606 | */ | |
fe4b04fa | 1607 | static int __perf_remove_from_context(void *info) |
0793a61d | 1608 | { |
46ce0fe9 PZ |
1609 | struct remove_event *re = info; |
1610 | struct perf_event *event = re->event; | |
cdd6c482 | 1611 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1612 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1613 | |
e625cce1 | 1614 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1615 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1616 | if (re->detach_group) |
1617 | perf_group_detach(event); | |
cdd6c482 | 1618 | list_del_event(event, ctx); |
64ce3126 PZ |
1619 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1620 | ctx->is_active = 0; | |
1621 | cpuctx->task_ctx = NULL; | |
1622 | } | |
e625cce1 | 1623 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1624 | |
1625 | return 0; | |
0793a61d TG |
1626 | } |
1627 | ||
1628 | ||
1629 | /* | |
cdd6c482 | 1630 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1631 | * |
cdd6c482 | 1632 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1633 | * call when the task is on a CPU. |
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 PM |
1637 | * remains valid. This is OK when called from perf_release since |
1638 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1639 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1640 | * context has been detached from its task. |
0793a61d | 1641 | */ |
46ce0fe9 | 1642 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1643 | { |
cdd6c482 | 1644 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1645 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1646 | struct remove_event re = { |
1647 | .event = event, | |
1648 | .detach_group = detach_group, | |
1649 | }; | |
0793a61d | 1650 | |
fe4b04fa PZ |
1651 | lockdep_assert_held(&ctx->mutex); |
1652 | ||
0793a61d TG |
1653 | if (!task) { |
1654 | /* | |
226424ee MR |
1655 | * Per cpu events are removed via an smp call. The removal can |
1656 | * fail if the CPU is currently offline, but in that case we | |
1657 | * already called __perf_remove_from_context from | |
1658 | * perf_event_exit_cpu. | |
0793a61d | 1659 | */ |
46ce0fe9 | 1660 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1661 | return; |
1662 | } | |
1663 | ||
1664 | retry: | |
46ce0fe9 | 1665 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1666 | return; |
0793a61d | 1667 | |
e625cce1 | 1668 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1669 | /* |
fe4b04fa PZ |
1670 | * If we failed to find a running task, but find the context active now |
1671 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1672 | */ |
fe4b04fa | 1673 | if (ctx->is_active) { |
e625cce1 | 1674 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
1675 | /* |
1676 | * Reload the task pointer, it might have been changed by | |
1677 | * a concurrent perf_event_context_sched_out(). | |
1678 | */ | |
1679 | task = ctx->task; | |
0793a61d TG |
1680 | goto retry; |
1681 | } | |
1682 | ||
1683 | /* | |
fe4b04fa PZ |
1684 | * Since the task isn't running, its safe to remove the event, us |
1685 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1686 | */ |
46ce0fe9 PZ |
1687 | if (detach_group) |
1688 | perf_group_detach(event); | |
fe4b04fa | 1689 | list_del_event(event, ctx); |
e625cce1 | 1690 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1691 | } |
1692 | ||
d859e29f | 1693 | /* |
cdd6c482 | 1694 | * Cross CPU call to disable a performance event |
d859e29f | 1695 | */ |
500ad2d8 | 1696 | int __perf_event_disable(void *info) |
d859e29f | 1697 | { |
cdd6c482 | 1698 | struct perf_event *event = info; |
cdd6c482 | 1699 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1700 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1701 | |
1702 | /* | |
cdd6c482 IM |
1703 | * If this is a per-task event, need to check whether this |
1704 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1705 | * |
1706 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1707 | * flipping contexts around. | |
d859e29f | 1708 | */ |
665c2142 | 1709 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1710 | return -EINVAL; |
d859e29f | 1711 | |
e625cce1 | 1712 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1713 | |
1714 | /* | |
cdd6c482 | 1715 | * If the event is on, turn it off. |
d859e29f PM |
1716 | * If it is in error state, leave it in error state. |
1717 | */ | |
cdd6c482 | 1718 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1719 | update_context_time(ctx); |
e5d1367f | 1720 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1721 | update_group_times(event); |
1722 | if (event == event->group_leader) | |
1723 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1724 | else |
cdd6c482 IM |
1725 | event_sched_out(event, cpuctx, ctx); |
1726 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1727 | } |
1728 | ||
e625cce1 | 1729 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1730 | |
1731 | return 0; | |
d859e29f PM |
1732 | } |
1733 | ||
1734 | /* | |
cdd6c482 | 1735 | * Disable a event. |
c93f7669 | 1736 | * |
cdd6c482 IM |
1737 | * If event->ctx is a cloned context, callers must make sure that |
1738 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1739 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1740 | * perf_event_for_each_child or perf_event_for_each because they |
1741 | * hold the top-level event's child_mutex, so any descendant that | |
1742 | * goes to exit will block in sync_child_event. | |
1743 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1744 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1745 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1746 | */ |
f63a8daa | 1747 | static void _perf_event_disable(struct perf_event *event) |
d859e29f | 1748 | { |
cdd6c482 | 1749 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1750 | struct task_struct *task = ctx->task; |
1751 | ||
1752 | if (!task) { | |
1753 | /* | |
cdd6c482 | 1754 | * Disable the event on the cpu that it's on |
d859e29f | 1755 | */ |
fe4b04fa | 1756 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1757 | return; |
1758 | } | |
1759 | ||
9ed6060d | 1760 | retry: |
fe4b04fa PZ |
1761 | if (!task_function_call(task, __perf_event_disable, event)) |
1762 | return; | |
d859e29f | 1763 | |
e625cce1 | 1764 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1765 | /* |
cdd6c482 | 1766 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1767 | */ |
cdd6c482 | 1768 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1769 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1770 | /* |
1771 | * Reload the task pointer, it might have been changed by | |
1772 | * a concurrent perf_event_context_sched_out(). | |
1773 | */ | |
1774 | task = ctx->task; | |
d859e29f PM |
1775 | goto retry; |
1776 | } | |
1777 | ||
1778 | /* | |
1779 | * Since we have the lock this context can't be scheduled | |
1780 | * in, so we can change the state safely. | |
1781 | */ | |
cdd6c482 IM |
1782 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1783 | update_group_times(event); | |
1784 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1785 | } |
e625cce1 | 1786 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1787 | } |
f63a8daa PZ |
1788 | |
1789 | /* | |
1790 | * Strictly speaking kernel users cannot create groups and therefore this | |
1791 | * interface does not need the perf_event_ctx_lock() magic. | |
1792 | */ | |
1793 | void perf_event_disable(struct perf_event *event) | |
1794 | { | |
1795 | struct perf_event_context *ctx; | |
1796 | ||
1797 | ctx = perf_event_ctx_lock(event); | |
1798 | _perf_event_disable(event); | |
1799 | perf_event_ctx_unlock(event, ctx); | |
1800 | } | |
dcfce4a0 | 1801 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1802 | |
e5d1367f SE |
1803 | static void perf_set_shadow_time(struct perf_event *event, |
1804 | struct perf_event_context *ctx, | |
1805 | u64 tstamp) | |
1806 | { | |
1807 | /* | |
1808 | * use the correct time source for the time snapshot | |
1809 | * | |
1810 | * We could get by without this by leveraging the | |
1811 | * fact that to get to this function, the caller | |
1812 | * has most likely already called update_context_time() | |
1813 | * and update_cgrp_time_xx() and thus both timestamp | |
1814 | * are identical (or very close). Given that tstamp is, | |
1815 | * already adjusted for cgroup, we could say that: | |
1816 | * tstamp - ctx->timestamp | |
1817 | * is equivalent to | |
1818 | * tstamp - cgrp->timestamp. | |
1819 | * | |
1820 | * Then, in perf_output_read(), the calculation would | |
1821 | * work with no changes because: | |
1822 | * - event is guaranteed scheduled in | |
1823 | * - no scheduled out in between | |
1824 | * - thus the timestamp would be the same | |
1825 | * | |
1826 | * But this is a bit hairy. | |
1827 | * | |
1828 | * So instead, we have an explicit cgroup call to remain | |
1829 | * within the time time source all along. We believe it | |
1830 | * is cleaner and simpler to understand. | |
1831 | */ | |
1832 | if (is_cgroup_event(event)) | |
1833 | perf_cgroup_set_shadow_time(event, tstamp); | |
1834 | else | |
1835 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1836 | } | |
1837 | ||
4fe757dd PZ |
1838 | #define MAX_INTERRUPTS (~0ULL) |
1839 | ||
1840 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1841 | ||
235c7fc7 | 1842 | static int |
9ffcfa6f | 1843 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1844 | struct perf_cpu_context *cpuctx, |
6e37738a | 1845 | struct perf_event_context *ctx) |
235c7fc7 | 1846 | { |
4158755d | 1847 | u64 tstamp = perf_event_time(event); |
44377277 | 1848 | int ret = 0; |
4158755d | 1849 | |
63342411 PZ |
1850 | lockdep_assert_held(&ctx->lock); |
1851 | ||
cdd6c482 | 1852 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1853 | return 0; |
1854 | ||
cdd6c482 | 1855 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1856 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1857 | |
1858 | /* | |
1859 | * Unthrottle events, since we scheduled we might have missed several | |
1860 | * ticks already, also for a heavily scheduling task there is little | |
1861 | * guarantee it'll get a tick in a timely manner. | |
1862 | */ | |
1863 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1864 | perf_log_throttle(event, 1); | |
1865 | event->hw.interrupts = 0; | |
1866 | } | |
1867 | ||
235c7fc7 IM |
1868 | /* |
1869 | * The new state must be visible before we turn it on in the hardware: | |
1870 | */ | |
1871 | smp_wmb(); | |
1872 | ||
44377277 AS |
1873 | perf_pmu_disable(event->pmu); |
1874 | ||
a4eaf7f1 | 1875 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1876 | event->state = PERF_EVENT_STATE_INACTIVE; |
1877 | event->oncpu = -1; | |
44377277 AS |
1878 | ret = -EAGAIN; |
1879 | goto out; | |
235c7fc7 IM |
1880 | } |
1881 | ||
4158755d | 1882 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1883 | |
e5d1367f | 1884 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1885 | |
cdd6c482 | 1886 | if (!is_software_event(event)) |
3b6f9e5c | 1887 | cpuctx->active_oncpu++; |
235c7fc7 | 1888 | ctx->nr_active++; |
0f5a2601 PZ |
1889 | if (event->attr.freq && event->attr.sample_freq) |
1890 | ctx->nr_freq++; | |
235c7fc7 | 1891 | |
cdd6c482 | 1892 | if (event->attr.exclusive) |
3b6f9e5c PM |
1893 | cpuctx->exclusive = 1; |
1894 | ||
fadfe7be JO |
1895 | if (is_orphaned_child(event)) |
1896 | schedule_orphans_remove(ctx); | |
1897 | ||
44377277 AS |
1898 | out: |
1899 | perf_pmu_enable(event->pmu); | |
1900 | ||
1901 | return ret; | |
235c7fc7 IM |
1902 | } |
1903 | ||
6751b71e | 1904 | static int |
cdd6c482 | 1905 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1906 | struct perf_cpu_context *cpuctx, |
6e37738a | 1907 | struct perf_event_context *ctx) |
6751b71e | 1908 | { |
6bde9b6c | 1909 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1910 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1911 | u64 now = ctx->time; |
1912 | bool simulate = false; | |
6751b71e | 1913 | |
cdd6c482 | 1914 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1915 | return 0; |
1916 | ||
ad5133b7 | 1917 | pmu->start_txn(pmu); |
6bde9b6c | 1918 | |
9ffcfa6f | 1919 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1920 | pmu->cancel_txn(pmu); |
9e630205 | 1921 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1922 | return -EAGAIN; |
90151c35 | 1923 | } |
6751b71e PM |
1924 | |
1925 | /* | |
1926 | * Schedule in siblings as one group (if any): | |
1927 | */ | |
cdd6c482 | 1928 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1929 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1930 | partial_group = event; |
6751b71e PM |
1931 | goto group_error; |
1932 | } | |
1933 | } | |
1934 | ||
9ffcfa6f | 1935 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1936 | return 0; |
9ffcfa6f | 1937 | |
6751b71e PM |
1938 | group_error: |
1939 | /* | |
1940 | * Groups can be scheduled in as one unit only, so undo any | |
1941 | * partial group before returning: | |
d7842da4 SE |
1942 | * The events up to the failed event are scheduled out normally, |
1943 | * tstamp_stopped will be updated. | |
1944 | * | |
1945 | * The failed events and the remaining siblings need to have | |
1946 | * their timings updated as if they had gone thru event_sched_in() | |
1947 | * and event_sched_out(). This is required to get consistent timings | |
1948 | * across the group. This also takes care of the case where the group | |
1949 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1950 | * the time the event was actually stopped, such that time delta | |
1951 | * calculation in update_event_times() is correct. | |
6751b71e | 1952 | */ |
cdd6c482 IM |
1953 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1954 | if (event == partial_group) | |
d7842da4 SE |
1955 | simulate = true; |
1956 | ||
1957 | if (simulate) { | |
1958 | event->tstamp_running += now - event->tstamp_stopped; | |
1959 | event->tstamp_stopped = now; | |
1960 | } else { | |
1961 | event_sched_out(event, cpuctx, ctx); | |
1962 | } | |
6751b71e | 1963 | } |
9ffcfa6f | 1964 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1965 | |
ad5133b7 | 1966 | pmu->cancel_txn(pmu); |
90151c35 | 1967 | |
9e630205 SE |
1968 | perf_cpu_hrtimer_restart(cpuctx); |
1969 | ||
6751b71e PM |
1970 | return -EAGAIN; |
1971 | } | |
1972 | ||
3b6f9e5c | 1973 | /* |
cdd6c482 | 1974 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1975 | */ |
cdd6c482 | 1976 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1977 | struct perf_cpu_context *cpuctx, |
1978 | int can_add_hw) | |
1979 | { | |
1980 | /* | |
cdd6c482 | 1981 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1982 | */ |
d6f962b5 | 1983 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1984 | return 1; |
1985 | /* | |
1986 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1987 | * events can go on. |
3b6f9e5c PM |
1988 | */ |
1989 | if (cpuctx->exclusive) | |
1990 | return 0; | |
1991 | /* | |
1992 | * If this group is exclusive and there are already | |
cdd6c482 | 1993 | * events on the CPU, it can't go on. |
3b6f9e5c | 1994 | */ |
cdd6c482 | 1995 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1996 | return 0; |
1997 | /* | |
1998 | * Otherwise, try to add it if all previous groups were able | |
1999 | * to go on. | |
2000 | */ | |
2001 | return can_add_hw; | |
2002 | } | |
2003 | ||
cdd6c482 IM |
2004 | static void add_event_to_ctx(struct perf_event *event, |
2005 | struct perf_event_context *ctx) | |
53cfbf59 | 2006 | { |
4158755d SE |
2007 | u64 tstamp = perf_event_time(event); |
2008 | ||
cdd6c482 | 2009 | list_add_event(event, ctx); |
8a49542c | 2010 | perf_group_attach(event); |
4158755d SE |
2011 | event->tstamp_enabled = tstamp; |
2012 | event->tstamp_running = tstamp; | |
2013 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
2014 | } |
2015 | ||
2c29ef0f PZ |
2016 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
2017 | static void | |
2018 | ctx_sched_in(struct perf_event_context *ctx, | |
2019 | struct perf_cpu_context *cpuctx, | |
2020 | enum event_type_t event_type, | |
2021 | struct task_struct *task); | |
fe4b04fa | 2022 | |
dce5855b PZ |
2023 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
2024 | struct perf_event_context *ctx, | |
2025 | struct task_struct *task) | |
2026 | { | |
2027 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
2028 | if (ctx) | |
2029 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
2030 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
2031 | if (ctx) | |
2032 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
2033 | } | |
2034 | ||
0793a61d | 2035 | /* |
cdd6c482 | 2036 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
2037 | * |
2038 | * Must be called with ctx->mutex held | |
0793a61d | 2039 | */ |
fe4b04fa | 2040 | static int __perf_install_in_context(void *info) |
0793a61d | 2041 | { |
cdd6c482 IM |
2042 | struct perf_event *event = info; |
2043 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2044 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
2045 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
2046 | struct task_struct *task = current; | |
2047 | ||
b58f6b0d | 2048 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 2049 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
2050 | |
2051 | /* | |
2c29ef0f | 2052 | * If there was an active task_ctx schedule it out. |
0793a61d | 2053 | */ |
b58f6b0d | 2054 | if (task_ctx) |
2c29ef0f | 2055 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
2056 | |
2057 | /* | |
2058 | * If the context we're installing events in is not the | |
2059 | * active task_ctx, flip them. | |
2060 | */ | |
2061 | if (ctx->task && task_ctx != ctx) { | |
2062 | if (task_ctx) | |
2063 | raw_spin_unlock(&task_ctx->lock); | |
2064 | raw_spin_lock(&ctx->lock); | |
2065 | task_ctx = ctx; | |
2066 | } | |
2067 | ||
2068 | if (task_ctx) { | |
2069 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
2070 | task = task_ctx->task; |
2071 | } | |
b58f6b0d | 2072 | |
2c29ef0f | 2073 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 2074 | |
4af4998b | 2075 | update_context_time(ctx); |
e5d1367f SE |
2076 | /* |
2077 | * update cgrp time only if current cgrp | |
2078 | * matches event->cgrp. Must be done before | |
2079 | * calling add_event_to_ctx() | |
2080 | */ | |
2081 | update_cgrp_time_from_event(event); | |
0793a61d | 2082 | |
cdd6c482 | 2083 | add_event_to_ctx(event, ctx); |
0793a61d | 2084 | |
d859e29f | 2085 | /* |
2c29ef0f | 2086 | * Schedule everything back in |
d859e29f | 2087 | */ |
dce5855b | 2088 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
2089 | |
2090 | perf_pmu_enable(cpuctx->ctx.pmu); | |
2091 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
2092 | |
2093 | return 0; | |
0793a61d TG |
2094 | } |
2095 | ||
2096 | /* | |
cdd6c482 | 2097 | * Attach a performance event to a context |
0793a61d | 2098 | * |
cdd6c482 IM |
2099 | * First we add the event to the list with the hardware enable bit |
2100 | * in event->hw_config cleared. | |
0793a61d | 2101 | * |
cdd6c482 | 2102 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
2103 | * call to enable it in the task context. The task might have been |
2104 | * scheduled away, but we check this in the smp call again. | |
2105 | */ | |
2106 | static void | |
cdd6c482 IM |
2107 | perf_install_in_context(struct perf_event_context *ctx, |
2108 | struct perf_event *event, | |
0793a61d TG |
2109 | int cpu) |
2110 | { | |
2111 | struct task_struct *task = ctx->task; | |
2112 | ||
fe4b04fa PZ |
2113 | lockdep_assert_held(&ctx->mutex); |
2114 | ||
c3f00c70 | 2115 | event->ctx = ctx; |
0cda4c02 YZ |
2116 | if (event->cpu != -1) |
2117 | event->cpu = cpu; | |
c3f00c70 | 2118 | |
0793a61d TG |
2119 | if (!task) { |
2120 | /* | |
cdd6c482 | 2121 | * Per cpu events are installed via an smp call and |
af901ca1 | 2122 | * the install is always successful. |
0793a61d | 2123 | */ |
fe4b04fa | 2124 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
2125 | return; |
2126 | } | |
2127 | ||
0793a61d | 2128 | retry: |
fe4b04fa PZ |
2129 | if (!task_function_call(task, __perf_install_in_context, event)) |
2130 | return; | |
0793a61d | 2131 | |
e625cce1 | 2132 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 2133 | /* |
fe4b04fa PZ |
2134 | * If we failed to find a running task, but find the context active now |
2135 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 2136 | */ |
fe4b04fa | 2137 | if (ctx->is_active) { |
e625cce1 | 2138 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
2139 | /* |
2140 | * Reload the task pointer, it might have been changed by | |
2141 | * a concurrent perf_event_context_sched_out(). | |
2142 | */ | |
2143 | task = ctx->task; | |
0793a61d TG |
2144 | goto retry; |
2145 | } | |
2146 | ||
2147 | /* | |
fe4b04fa PZ |
2148 | * Since the task isn't running, its safe to add the event, us holding |
2149 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 2150 | */ |
fe4b04fa | 2151 | add_event_to_ctx(event, ctx); |
e625cce1 | 2152 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
2153 | } |
2154 | ||
fa289bec | 2155 | /* |
cdd6c482 | 2156 | * Put a event into inactive state and update time fields. |
fa289bec PM |
2157 | * Enabling the leader of a group effectively enables all |
2158 | * the group members that aren't explicitly disabled, so we | |
2159 | * have to update their ->tstamp_enabled also. | |
2160 | * Note: this works for group members as well as group leaders | |
2161 | * since the non-leader members' sibling_lists will be empty. | |
2162 | */ | |
1d9b482e | 2163 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 2164 | { |
cdd6c482 | 2165 | struct perf_event *sub; |
4158755d | 2166 | u64 tstamp = perf_event_time(event); |
fa289bec | 2167 | |
cdd6c482 | 2168 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 2169 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 2170 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
2171 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
2172 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 2173 | } |
fa289bec PM |
2174 | } |
2175 | ||
d859e29f | 2176 | /* |
cdd6c482 | 2177 | * Cross CPU call to enable a performance event |
d859e29f | 2178 | */ |
fe4b04fa | 2179 | static int __perf_event_enable(void *info) |
04289bb9 | 2180 | { |
cdd6c482 | 2181 | struct perf_event *event = info; |
cdd6c482 IM |
2182 | struct perf_event_context *ctx = event->ctx; |
2183 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2184 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2185 | int err; |
04289bb9 | 2186 | |
06f41796 JO |
2187 | /* |
2188 | * There's a time window between 'ctx->is_active' check | |
2189 | * in perf_event_enable function and this place having: | |
2190 | * - IRQs on | |
2191 | * - ctx->lock unlocked | |
2192 | * | |
2193 | * where the task could be killed and 'ctx' deactivated | |
2194 | * by perf_event_exit_task. | |
2195 | */ | |
2196 | if (!ctx->is_active) | |
fe4b04fa | 2197 | return -EINVAL; |
3cbed429 | 2198 | |
e625cce1 | 2199 | raw_spin_lock(&ctx->lock); |
4af4998b | 2200 | update_context_time(ctx); |
d859e29f | 2201 | |
cdd6c482 | 2202 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2203 | goto unlock; |
e5d1367f SE |
2204 | |
2205 | /* | |
2206 | * set current task's cgroup time reference point | |
2207 | */ | |
3f7cce3c | 2208 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2209 | |
1d9b482e | 2210 | __perf_event_mark_enabled(event); |
04289bb9 | 2211 | |
e5d1367f SE |
2212 | if (!event_filter_match(event)) { |
2213 | if (is_cgroup_event(event)) | |
2214 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2215 | goto unlock; |
e5d1367f | 2216 | } |
f4c4176f | 2217 | |
04289bb9 | 2218 | /* |
cdd6c482 | 2219 | * If the event is in a group and isn't the group leader, |
d859e29f | 2220 | * then don't put it on unless the group is on. |
04289bb9 | 2221 | */ |
cdd6c482 | 2222 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2223 | goto unlock; |
3b6f9e5c | 2224 | |
cdd6c482 | 2225 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2226 | err = -EEXIST; |
e758a33d | 2227 | } else { |
cdd6c482 | 2228 | if (event == leader) |
6e37738a | 2229 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2230 | else |
6e37738a | 2231 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2232 | } |
d859e29f PM |
2233 | |
2234 | if (err) { | |
2235 | /* | |
cdd6c482 | 2236 | * If this event can't go on and it's part of a |
d859e29f PM |
2237 | * group, then the whole group has to come off. |
2238 | */ | |
9e630205 | 2239 | if (leader != event) { |
d859e29f | 2240 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2241 | perf_cpu_hrtimer_restart(cpuctx); |
2242 | } | |
0d48696f | 2243 | if (leader->attr.pinned) { |
53cfbf59 | 2244 | update_group_times(leader); |
cdd6c482 | 2245 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2246 | } |
d859e29f PM |
2247 | } |
2248 | ||
9ed6060d | 2249 | unlock: |
e625cce1 | 2250 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2251 | |
2252 | return 0; | |
d859e29f PM |
2253 | } |
2254 | ||
2255 | /* | |
cdd6c482 | 2256 | * Enable a event. |
c93f7669 | 2257 | * |
cdd6c482 IM |
2258 | * If event->ctx is a cloned context, callers must make sure that |
2259 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2260 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2261 | * perf_event_for_each_child or perf_event_for_each as described |
2262 | * for perf_event_disable. | |
d859e29f | 2263 | */ |
f63a8daa | 2264 | static void _perf_event_enable(struct perf_event *event) |
d859e29f | 2265 | { |
cdd6c482 | 2266 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2267 | struct task_struct *task = ctx->task; |
2268 | ||
2269 | if (!task) { | |
2270 | /* | |
cdd6c482 | 2271 | * Enable the event on the cpu that it's on |
d859e29f | 2272 | */ |
fe4b04fa | 2273 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2274 | return; |
2275 | } | |
2276 | ||
e625cce1 | 2277 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2278 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2279 | goto out; |
2280 | ||
2281 | /* | |
cdd6c482 IM |
2282 | * If the event is in error state, clear that first. |
2283 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2284 | * know that it has gone back into error state, as distinct |
2285 | * from the task having been scheduled away before the | |
2286 | * cross-call arrived. | |
2287 | */ | |
cdd6c482 IM |
2288 | if (event->state == PERF_EVENT_STATE_ERROR) |
2289 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2290 | |
9ed6060d | 2291 | retry: |
fe4b04fa | 2292 | if (!ctx->is_active) { |
1d9b482e | 2293 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2294 | goto out; |
2295 | } | |
2296 | ||
e625cce1 | 2297 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2298 | |
2299 | if (!task_function_call(task, __perf_event_enable, event)) | |
2300 | return; | |
d859e29f | 2301 | |
e625cce1 | 2302 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2303 | |
2304 | /* | |
cdd6c482 | 2305 | * If the context is active and the event is still off, |
d859e29f PM |
2306 | * we need to retry the cross-call. |
2307 | */ | |
fe4b04fa PZ |
2308 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2309 | /* | |
2310 | * task could have been flipped by a concurrent | |
2311 | * perf_event_context_sched_out() | |
2312 | */ | |
2313 | task = ctx->task; | |
d859e29f | 2314 | goto retry; |
fe4b04fa | 2315 | } |
fa289bec | 2316 | |
9ed6060d | 2317 | out: |
e625cce1 | 2318 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2319 | } |
f63a8daa PZ |
2320 | |
2321 | /* | |
2322 | * See perf_event_disable(); | |
2323 | */ | |
2324 | void perf_event_enable(struct perf_event *event) | |
2325 | { | |
2326 | struct perf_event_context *ctx; | |
2327 | ||
2328 | ctx = perf_event_ctx_lock(event); | |
2329 | _perf_event_enable(event); | |
2330 | perf_event_ctx_unlock(event, ctx); | |
2331 | } | |
dcfce4a0 | 2332 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2333 | |
f63a8daa | 2334 | static int _perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2335 | { |
2023b359 | 2336 | /* |
cdd6c482 | 2337 | * not supported on inherited events |
2023b359 | 2338 | */ |
2e939d1d | 2339 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2340 | return -EINVAL; |
2341 | ||
cdd6c482 | 2342 | atomic_add(refresh, &event->event_limit); |
f63a8daa | 2343 | _perf_event_enable(event); |
2023b359 PZ |
2344 | |
2345 | return 0; | |
79f14641 | 2346 | } |
f63a8daa PZ |
2347 | |
2348 | /* | |
2349 | * See perf_event_disable() | |
2350 | */ | |
2351 | int perf_event_refresh(struct perf_event *event, int refresh) | |
2352 | { | |
2353 | struct perf_event_context *ctx; | |
2354 | int ret; | |
2355 | ||
2356 | ctx = perf_event_ctx_lock(event); | |
2357 | ret = _perf_event_refresh(event, refresh); | |
2358 | perf_event_ctx_unlock(event, ctx); | |
2359 | ||
2360 | return ret; | |
2361 | } | |
26ca5c11 | 2362 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2363 | |
5b0311e1 FW |
2364 | static void ctx_sched_out(struct perf_event_context *ctx, |
2365 | struct perf_cpu_context *cpuctx, | |
2366 | enum event_type_t event_type) | |
235c7fc7 | 2367 | { |
cdd6c482 | 2368 | struct perf_event *event; |
db24d33e | 2369 | int is_active = ctx->is_active; |
235c7fc7 | 2370 | |
db24d33e | 2371 | ctx->is_active &= ~event_type; |
cdd6c482 | 2372 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2373 | return; |
2374 | ||
4af4998b | 2375 | update_context_time(ctx); |
e5d1367f | 2376 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2377 | if (!ctx->nr_active) |
facc4307 | 2378 | return; |
5b0311e1 | 2379 | |
075e0b00 | 2380 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2381 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2382 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2383 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2384 | } |
889ff015 | 2385 | |
db24d33e | 2386 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2387 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2388 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2389 | } |
1b9a644f | 2390 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2391 | } |
2392 | ||
564c2b21 | 2393 | /* |
5a3126d4 PZ |
2394 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2395 | * cloned from the same version of the same context. | |
2396 | * | |
2397 | * Equivalence is measured using a generation number in the context that is | |
2398 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2399 | * and list_del_event(). | |
564c2b21 | 2400 | */ |
cdd6c482 IM |
2401 | static int context_equiv(struct perf_event_context *ctx1, |
2402 | struct perf_event_context *ctx2) | |
564c2b21 | 2403 | { |
211de6eb PZ |
2404 | lockdep_assert_held(&ctx1->lock); |
2405 | lockdep_assert_held(&ctx2->lock); | |
2406 | ||
5a3126d4 PZ |
2407 | /* Pinning disables the swap optimization */ |
2408 | if (ctx1->pin_count || ctx2->pin_count) | |
2409 | return 0; | |
2410 | ||
2411 | /* If ctx1 is the parent of ctx2 */ | |
2412 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2413 | return 1; | |
2414 | ||
2415 | /* If ctx2 is the parent of ctx1 */ | |
2416 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2417 | return 1; | |
2418 | ||
2419 | /* | |
2420 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2421 | * hierarchy, see perf_event_init_context(). | |
2422 | */ | |
2423 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2424 | ctx1->parent_gen == ctx2->parent_gen) | |
2425 | return 1; | |
2426 | ||
2427 | /* Unmatched */ | |
2428 | return 0; | |
564c2b21 PM |
2429 | } |
2430 | ||
cdd6c482 IM |
2431 | static void __perf_event_sync_stat(struct perf_event *event, |
2432 | struct perf_event *next_event) | |
bfbd3381 PZ |
2433 | { |
2434 | u64 value; | |
2435 | ||
cdd6c482 | 2436 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2437 | return; |
2438 | ||
2439 | /* | |
cdd6c482 | 2440 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2441 | * because we're in the middle of a context switch and have IRQs |
2442 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2443 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2444 | * don't need to use it. |
2445 | */ | |
cdd6c482 IM |
2446 | switch (event->state) { |
2447 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2448 | event->pmu->read(event); |
2449 | /* fall-through */ | |
bfbd3381 | 2450 | |
cdd6c482 IM |
2451 | case PERF_EVENT_STATE_INACTIVE: |
2452 | update_event_times(event); | |
bfbd3381 PZ |
2453 | break; |
2454 | ||
2455 | default: | |
2456 | break; | |
2457 | } | |
2458 | ||
2459 | /* | |
cdd6c482 | 2460 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2461 | * values when we flip the contexts. |
2462 | */ | |
e7850595 PZ |
2463 | value = local64_read(&next_event->count); |
2464 | value = local64_xchg(&event->count, value); | |
2465 | local64_set(&next_event->count, value); | |
bfbd3381 | 2466 | |
cdd6c482 IM |
2467 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2468 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2469 | |
bfbd3381 | 2470 | /* |
19d2e755 | 2471 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2472 | */ |
cdd6c482 IM |
2473 | perf_event_update_userpage(event); |
2474 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2475 | } |
2476 | ||
cdd6c482 IM |
2477 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2478 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2479 | { |
cdd6c482 | 2480 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2481 | |
2482 | if (!ctx->nr_stat) | |
2483 | return; | |
2484 | ||
02ffdbc8 PZ |
2485 | update_context_time(ctx); |
2486 | ||
cdd6c482 IM |
2487 | event = list_first_entry(&ctx->event_list, |
2488 | struct perf_event, event_entry); | |
bfbd3381 | 2489 | |
cdd6c482 IM |
2490 | next_event = list_first_entry(&next_ctx->event_list, |
2491 | struct perf_event, event_entry); | |
bfbd3381 | 2492 | |
cdd6c482 IM |
2493 | while (&event->event_entry != &ctx->event_list && |
2494 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2495 | |
cdd6c482 | 2496 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2497 | |
cdd6c482 IM |
2498 | event = list_next_entry(event, event_entry); |
2499 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2500 | } |
2501 | } | |
2502 | ||
fe4b04fa PZ |
2503 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2504 | struct task_struct *next) | |
0793a61d | 2505 | { |
8dc85d54 | 2506 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2507 | struct perf_event_context *next_ctx; |
5a3126d4 | 2508 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2509 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2510 | int do_switch = 1; |
0793a61d | 2511 | |
108b02cf PZ |
2512 | if (likely(!ctx)) |
2513 | return; | |
10989fb2 | 2514 | |
108b02cf PZ |
2515 | cpuctx = __get_cpu_context(ctx); |
2516 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2517 | return; |
2518 | ||
c93f7669 | 2519 | rcu_read_lock(); |
8dc85d54 | 2520 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2521 | if (!next_ctx) |
2522 | goto unlock; | |
2523 | ||
2524 | parent = rcu_dereference(ctx->parent_ctx); | |
2525 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2526 | ||
2527 | /* If neither context have a parent context; they cannot be clones. */ | |
802c8a61 | 2528 | if (!parent && !next_parent) |
5a3126d4 PZ |
2529 | goto unlock; |
2530 | ||
2531 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2532 | /* |
2533 | * Looks like the two contexts are clones, so we might be | |
2534 | * able to optimize the context switch. We lock both | |
2535 | * contexts and check that they are clones under the | |
2536 | * lock (including re-checking that neither has been | |
2537 | * uncloned in the meantime). It doesn't matter which | |
2538 | * order we take the locks because no other cpu could | |
2539 | * be trying to lock both of these tasks. | |
2540 | */ | |
e625cce1 TG |
2541 | raw_spin_lock(&ctx->lock); |
2542 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2543 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2544 | /* |
2545 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2546 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2547 | */ |
8dc85d54 PZ |
2548 | task->perf_event_ctxp[ctxn] = next_ctx; |
2549 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2550 | ctx->task = next; |
2551 | next_ctx->task = task; | |
2552 | do_switch = 0; | |
bfbd3381 | 2553 | |
cdd6c482 | 2554 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2555 | } |
e625cce1 TG |
2556 | raw_spin_unlock(&next_ctx->lock); |
2557 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2558 | } |
5a3126d4 | 2559 | unlock: |
c93f7669 | 2560 | rcu_read_unlock(); |
564c2b21 | 2561 | |
c93f7669 | 2562 | if (do_switch) { |
facc4307 | 2563 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2564 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2565 | cpuctx->task_ctx = NULL; |
facc4307 | 2566 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2567 | } |
0793a61d TG |
2568 | } |
2569 | ||
8dc85d54 PZ |
2570 | #define for_each_task_context_nr(ctxn) \ |
2571 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2572 | ||
2573 | /* | |
2574 | * Called from scheduler to remove the events of the current task, | |
2575 | * with interrupts disabled. | |
2576 | * | |
2577 | * We stop each event and update the event value in event->count. | |
2578 | * | |
2579 | * This does not protect us against NMI, but disable() | |
2580 | * sets the disabled bit in the control field of event _before_ | |
2581 | * accessing the event control register. If a NMI hits, then it will | |
2582 | * not restart the event. | |
2583 | */ | |
ab0cce56 JO |
2584 | void __perf_event_task_sched_out(struct task_struct *task, |
2585 | struct task_struct *next) | |
8dc85d54 PZ |
2586 | { |
2587 | int ctxn; | |
2588 | ||
8dc85d54 PZ |
2589 | for_each_task_context_nr(ctxn) |
2590 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2591 | |
2592 | /* | |
2593 | * if cgroup events exist on this CPU, then we need | |
2594 | * to check if we have to switch out PMU state. | |
2595 | * cgroup event are system-wide mode only | |
2596 | */ | |
4a32fea9 | 2597 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2598 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2599 | } |
2600 | ||
04dc2dbb | 2601 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2602 | { |
108b02cf | 2603 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2604 | |
a63eaf34 PM |
2605 | if (!cpuctx->task_ctx) |
2606 | return; | |
012b84da IM |
2607 | |
2608 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2609 | return; | |
2610 | ||
04dc2dbb | 2611 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2612 | cpuctx->task_ctx = NULL; |
2613 | } | |
2614 | ||
5b0311e1 FW |
2615 | /* |
2616 | * Called with IRQs disabled | |
2617 | */ | |
2618 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2619 | enum event_type_t event_type) | |
2620 | { | |
2621 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2622 | } |
2623 | ||
235c7fc7 | 2624 | static void |
5b0311e1 | 2625 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2626 | struct perf_cpu_context *cpuctx) |
0793a61d | 2627 | { |
cdd6c482 | 2628 | struct perf_event *event; |
0793a61d | 2629 | |
889ff015 FW |
2630 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2631 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2632 | continue; |
5632ab12 | 2633 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2634 | continue; |
2635 | ||
e5d1367f SE |
2636 | /* may need to reset tstamp_enabled */ |
2637 | if (is_cgroup_event(event)) | |
2638 | perf_cgroup_mark_enabled(event, ctx); | |
2639 | ||
8c9ed8e1 | 2640 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2641 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2642 | |
2643 | /* | |
2644 | * If this pinned group hasn't been scheduled, | |
2645 | * put it in error state. | |
2646 | */ | |
cdd6c482 IM |
2647 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2648 | update_group_times(event); | |
2649 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2650 | } |
3b6f9e5c | 2651 | } |
5b0311e1 FW |
2652 | } |
2653 | ||
2654 | static void | |
2655 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2656 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2657 | { |
2658 | struct perf_event *event; | |
2659 | int can_add_hw = 1; | |
3b6f9e5c | 2660 | |
889ff015 FW |
2661 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2662 | /* Ignore events in OFF or ERROR state */ | |
2663 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2664 | continue; |
04289bb9 IM |
2665 | /* |
2666 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2667 | * of events: |
04289bb9 | 2668 | */ |
5632ab12 | 2669 | if (!event_filter_match(event)) |
0793a61d TG |
2670 | continue; |
2671 | ||
e5d1367f SE |
2672 | /* may need to reset tstamp_enabled */ |
2673 | if (is_cgroup_event(event)) | |
2674 | perf_cgroup_mark_enabled(event, ctx); | |
2675 | ||
9ed6060d | 2676 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2677 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2678 | can_add_hw = 0; |
9ed6060d | 2679 | } |
0793a61d | 2680 | } |
5b0311e1 FW |
2681 | } |
2682 | ||
2683 | static void | |
2684 | ctx_sched_in(struct perf_event_context *ctx, | |
2685 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2686 | enum event_type_t event_type, |
2687 | struct task_struct *task) | |
5b0311e1 | 2688 | { |
e5d1367f | 2689 | u64 now; |
db24d33e | 2690 | int is_active = ctx->is_active; |
e5d1367f | 2691 | |
db24d33e | 2692 | ctx->is_active |= event_type; |
5b0311e1 | 2693 | if (likely(!ctx->nr_events)) |
facc4307 | 2694 | return; |
5b0311e1 | 2695 | |
e5d1367f SE |
2696 | now = perf_clock(); |
2697 | ctx->timestamp = now; | |
3f7cce3c | 2698 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2699 | /* |
2700 | * First go through the list and put on any pinned groups | |
2701 | * in order to give them the best chance of going on. | |
2702 | */ | |
db24d33e | 2703 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2704 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2705 | |
2706 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2707 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2708 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2709 | } |
2710 | ||
329c0e01 | 2711 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2712 | enum event_type_t event_type, |
2713 | struct task_struct *task) | |
329c0e01 FW |
2714 | { |
2715 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2716 | ||
e5d1367f | 2717 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2718 | } |
2719 | ||
e5d1367f SE |
2720 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2721 | struct task_struct *task) | |
235c7fc7 | 2722 | { |
108b02cf | 2723 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2724 | |
108b02cf | 2725 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2726 | if (cpuctx->task_ctx == ctx) |
2727 | return; | |
2728 | ||
facc4307 | 2729 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2730 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2731 | /* |
2732 | * We want to keep the following priority order: | |
2733 | * cpu pinned (that don't need to move), task pinned, | |
2734 | * cpu flexible, task flexible. | |
2735 | */ | |
2736 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2737 | ||
1d5f003f GN |
2738 | if (ctx->nr_events) |
2739 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2740 | |
86b47c25 GN |
2741 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2742 | ||
facc4307 PZ |
2743 | perf_pmu_enable(ctx->pmu); |
2744 | perf_ctx_unlock(cpuctx, ctx); | |
2745 | ||
b5ab4cd5 PZ |
2746 | /* |
2747 | * Since these rotations are per-cpu, we need to ensure the | |
2748 | * cpu-context we got scheduled on is actually rotating. | |
2749 | */ | |
108b02cf | 2750 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2751 | } |
2752 | ||
d010b332 SE |
2753 | /* |
2754 | * When sampling the branck stack in system-wide, it may be necessary | |
2755 | * to flush the stack on context switch. This happens when the branch | |
2756 | * stack does not tag its entries with the pid of the current task. | |
2757 | * Otherwise it becomes impossible to associate a branch entry with a | |
2758 | * task. This ambiguity is more likely to appear when the branch stack | |
2759 | * supports priv level filtering and the user sets it to monitor only | |
2760 | * at the user level (which could be a useful measurement in system-wide | |
2761 | * mode). In that case, the risk is high of having a branch stack with | |
2762 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2763 | * entries or stashing them somewhere in the PMU specific code layer. | |
2764 | * | |
2765 | * This function provides the context switch callback to the lower code | |
2766 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2767 | * with at least one active event using taken branch sampling. | |
2768 | */ | |
2769 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2770 | struct task_struct *task) | |
2771 | { | |
2772 | struct perf_cpu_context *cpuctx; | |
2773 | struct pmu *pmu; | |
2774 | unsigned long flags; | |
2775 | ||
2776 | /* no need to flush branch stack if not changing task */ | |
2777 | if (prev == task) | |
2778 | return; | |
2779 | ||
2780 | local_irq_save(flags); | |
2781 | ||
2782 | rcu_read_lock(); | |
2783 | ||
2784 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2785 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2786 | ||
2787 | /* | |
2788 | * check if the context has at least one | |
2789 | * event using PERF_SAMPLE_BRANCH_STACK | |
2790 | */ | |
2791 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2792 | && pmu->flush_branch_stack) { | |
2793 | ||
d010b332 SE |
2794 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
2795 | ||
2796 | perf_pmu_disable(pmu); | |
2797 | ||
2798 | pmu->flush_branch_stack(); | |
2799 | ||
2800 | perf_pmu_enable(pmu); | |
2801 | ||
2802 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2803 | } | |
2804 | } | |
2805 | ||
2806 | rcu_read_unlock(); | |
2807 | ||
2808 | local_irq_restore(flags); | |
2809 | } | |
2810 | ||
8dc85d54 PZ |
2811 | /* |
2812 | * Called from scheduler to add the events of the current task | |
2813 | * with interrupts disabled. | |
2814 | * | |
2815 | * We restore the event value and then enable it. | |
2816 | * | |
2817 | * This does not protect us against NMI, but enable() | |
2818 | * sets the enabled bit in the control field of event _before_ | |
2819 | * accessing the event control register. If a NMI hits, then it will | |
2820 | * keep the event running. | |
2821 | */ | |
ab0cce56 JO |
2822 | void __perf_event_task_sched_in(struct task_struct *prev, |
2823 | struct task_struct *task) | |
8dc85d54 PZ |
2824 | { |
2825 | struct perf_event_context *ctx; | |
2826 | int ctxn; | |
2827 | ||
2828 | for_each_task_context_nr(ctxn) { | |
2829 | ctx = task->perf_event_ctxp[ctxn]; | |
2830 | if (likely(!ctx)) | |
2831 | continue; | |
2832 | ||
e5d1367f | 2833 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2834 | } |
e5d1367f SE |
2835 | /* |
2836 | * if cgroup events exist on this CPU, then we need | |
2837 | * to check if we have to switch in PMU state. | |
2838 | * cgroup event are system-wide mode only | |
2839 | */ | |
4a32fea9 | 2840 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2841 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2842 | |
2843 | /* check for system-wide branch_stack events */ | |
4a32fea9 | 2844 | if (atomic_read(this_cpu_ptr(&perf_branch_stack_events))) |
d010b332 | 2845 | perf_branch_stack_sched_in(prev, task); |
235c7fc7 IM |
2846 | } |
2847 | ||
abd50713 PZ |
2848 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2849 | { | |
2850 | u64 frequency = event->attr.sample_freq; | |
2851 | u64 sec = NSEC_PER_SEC; | |
2852 | u64 divisor, dividend; | |
2853 | ||
2854 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2855 | ||
2856 | count_fls = fls64(count); | |
2857 | nsec_fls = fls64(nsec); | |
2858 | frequency_fls = fls64(frequency); | |
2859 | sec_fls = 30; | |
2860 | ||
2861 | /* | |
2862 | * We got @count in @nsec, with a target of sample_freq HZ | |
2863 | * the target period becomes: | |
2864 | * | |
2865 | * @count * 10^9 | |
2866 | * period = ------------------- | |
2867 | * @nsec * sample_freq | |
2868 | * | |
2869 | */ | |
2870 | ||
2871 | /* | |
2872 | * Reduce accuracy by one bit such that @a and @b converge | |
2873 | * to a similar magnitude. | |
2874 | */ | |
fe4b04fa | 2875 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2876 | do { \ |
2877 | if (a##_fls > b##_fls) { \ | |
2878 | a >>= 1; \ | |
2879 | a##_fls--; \ | |
2880 | } else { \ | |
2881 | b >>= 1; \ | |
2882 | b##_fls--; \ | |
2883 | } \ | |
2884 | } while (0) | |
2885 | ||
2886 | /* | |
2887 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2888 | * the other, so that finally we can do a u64/u64 division. | |
2889 | */ | |
2890 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2891 | REDUCE_FLS(nsec, frequency); | |
2892 | REDUCE_FLS(sec, count); | |
2893 | } | |
2894 | ||
2895 | if (count_fls + sec_fls > 64) { | |
2896 | divisor = nsec * frequency; | |
2897 | ||
2898 | while (count_fls + sec_fls > 64) { | |
2899 | REDUCE_FLS(count, sec); | |
2900 | divisor >>= 1; | |
2901 | } | |
2902 | ||
2903 | dividend = count * sec; | |
2904 | } else { | |
2905 | dividend = count * sec; | |
2906 | ||
2907 | while (nsec_fls + frequency_fls > 64) { | |
2908 | REDUCE_FLS(nsec, frequency); | |
2909 | dividend >>= 1; | |
2910 | } | |
2911 | ||
2912 | divisor = nsec * frequency; | |
2913 | } | |
2914 | ||
f6ab91ad PZ |
2915 | if (!divisor) |
2916 | return dividend; | |
2917 | ||
abd50713 PZ |
2918 | return div64_u64(dividend, divisor); |
2919 | } | |
2920 | ||
e050e3f0 SE |
2921 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2922 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2923 | ||
f39d47ff | 2924 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2925 | { |
cdd6c482 | 2926 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2927 | s64 period, sample_period; |
bd2b5b12 PZ |
2928 | s64 delta; |
2929 | ||
abd50713 | 2930 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2931 | |
2932 | delta = (s64)(period - hwc->sample_period); | |
2933 | delta = (delta + 7) / 8; /* low pass filter */ | |
2934 | ||
2935 | sample_period = hwc->sample_period + delta; | |
2936 | ||
2937 | if (!sample_period) | |
2938 | sample_period = 1; | |
2939 | ||
bd2b5b12 | 2940 | hwc->sample_period = sample_period; |
abd50713 | 2941 | |
e7850595 | 2942 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2943 | if (disable) |
2944 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2945 | ||
e7850595 | 2946 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2947 | |
2948 | if (disable) | |
2949 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2950 | } |
bd2b5b12 PZ |
2951 | } |
2952 | ||
e050e3f0 SE |
2953 | /* |
2954 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2955 | * events. At the same time, make sure, having freq events does not change | |
2956 | * the rate of unthrottling as that would introduce bias. | |
2957 | */ | |
2958 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2959 | int needs_unthr) | |
60db5e09 | 2960 | { |
cdd6c482 IM |
2961 | struct perf_event *event; |
2962 | struct hw_perf_event *hwc; | |
e050e3f0 | 2963 | u64 now, period = TICK_NSEC; |
abd50713 | 2964 | s64 delta; |
60db5e09 | 2965 | |
e050e3f0 SE |
2966 | /* |
2967 | * only need to iterate over all events iff: | |
2968 | * - context have events in frequency mode (needs freq adjust) | |
2969 | * - there are events to unthrottle on this cpu | |
2970 | */ | |
2971 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2972 | return; |
2973 | ||
e050e3f0 | 2974 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2975 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2976 | |
03541f8b | 2977 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2978 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2979 | continue; |
2980 | ||
5632ab12 | 2981 | if (!event_filter_match(event)) |
5d27c23d PZ |
2982 | continue; |
2983 | ||
44377277 AS |
2984 | perf_pmu_disable(event->pmu); |
2985 | ||
cdd6c482 | 2986 | hwc = &event->hw; |
6a24ed6c | 2987 | |
ae23bff1 | 2988 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2989 | hwc->interrupts = 0; |
cdd6c482 | 2990 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2991 | event->pmu->start(event, 0); |
a78ac325 PZ |
2992 | } |
2993 | ||
cdd6c482 | 2994 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2995 | goto next; |
60db5e09 | 2996 | |
e050e3f0 SE |
2997 | /* |
2998 | * stop the event and update event->count | |
2999 | */ | |
3000 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3001 | ||
e7850595 | 3002 | now = local64_read(&event->count); |
abd50713 PZ |
3003 | delta = now - hwc->freq_count_stamp; |
3004 | hwc->freq_count_stamp = now; | |
60db5e09 | 3005 | |
e050e3f0 SE |
3006 | /* |
3007 | * restart the event | |
3008 | * reload only if value has changed | |
f39d47ff SE |
3009 | * we have stopped the event so tell that |
3010 | * to perf_adjust_period() to avoid stopping it | |
3011 | * twice. | |
e050e3f0 | 3012 | */ |
abd50713 | 3013 | if (delta > 0) |
f39d47ff | 3014 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
3015 | |
3016 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
3017 | next: |
3018 | perf_pmu_enable(event->pmu); | |
60db5e09 | 3019 | } |
e050e3f0 | 3020 | |
f39d47ff | 3021 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 3022 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
3023 | } |
3024 | ||
235c7fc7 | 3025 | /* |
cdd6c482 | 3026 | * Round-robin a context's events: |
235c7fc7 | 3027 | */ |
cdd6c482 | 3028 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 3029 | { |
dddd3379 TG |
3030 | /* |
3031 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
3032 | * disabled by the inheritance code. | |
3033 | */ | |
3034 | if (!ctx->rotate_disable) | |
3035 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
3036 | } |
3037 | ||
b5ab4cd5 | 3038 | /* |
e9d2b064 PZ |
3039 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
3040 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
3041 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 3042 | */ |
9e630205 | 3043 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 3044 | { |
8dc85d54 | 3045 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 3046 | int rotate = 0, remove = 1; |
7fc23a53 | 3047 | |
b5ab4cd5 | 3048 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 3049 | remove = 0; |
b5ab4cd5 PZ |
3050 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
3051 | rotate = 1; | |
3052 | } | |
235c7fc7 | 3053 | |
8dc85d54 | 3054 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 3055 | if (ctx && ctx->nr_events) { |
e9d2b064 | 3056 | remove = 0; |
b5ab4cd5 PZ |
3057 | if (ctx->nr_events != ctx->nr_active) |
3058 | rotate = 1; | |
3059 | } | |
9717e6cd | 3060 | |
e050e3f0 | 3061 | if (!rotate) |
0f5a2601 PZ |
3062 | goto done; |
3063 | ||
facc4307 | 3064 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 3065 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 3066 | |
e050e3f0 SE |
3067 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
3068 | if (ctx) | |
3069 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 3070 | |
e050e3f0 SE |
3071 | rotate_ctx(&cpuctx->ctx); |
3072 | if (ctx) | |
3073 | rotate_ctx(ctx); | |
235c7fc7 | 3074 | |
e050e3f0 | 3075 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 3076 | |
0f5a2601 PZ |
3077 | perf_pmu_enable(cpuctx->ctx.pmu); |
3078 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 3079 | done: |
e9d2b064 PZ |
3080 | if (remove) |
3081 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
3082 | |
3083 | return rotate; | |
e9d2b064 PZ |
3084 | } |
3085 | ||
026249ef FW |
3086 | #ifdef CONFIG_NO_HZ_FULL |
3087 | bool perf_event_can_stop_tick(void) | |
3088 | { | |
948b26b6 | 3089 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 3090 | __this_cpu_read(perf_throttled_count)) |
026249ef | 3091 | return false; |
d84153d6 FW |
3092 | else |
3093 | return true; | |
026249ef FW |
3094 | } |
3095 | #endif | |
3096 | ||
e9d2b064 PZ |
3097 | void perf_event_task_tick(void) |
3098 | { | |
4a32fea9 | 3099 | struct list_head *head = this_cpu_ptr(&rotation_list); |
e9d2b064 | 3100 | struct perf_cpu_context *cpuctx, *tmp; |
e050e3f0 SE |
3101 | struct perf_event_context *ctx; |
3102 | int throttled; | |
b5ab4cd5 | 3103 | |
e9d2b064 PZ |
3104 | WARN_ON(!irqs_disabled()); |
3105 | ||
e050e3f0 SE |
3106 | __this_cpu_inc(perf_throttled_seq); |
3107 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
3108 | ||
e9d2b064 | 3109 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
3110 | ctx = &cpuctx->ctx; |
3111 | perf_adjust_freq_unthr_context(ctx, throttled); | |
3112 | ||
3113 | ctx = cpuctx->task_ctx; | |
3114 | if (ctx) | |
3115 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 3116 | } |
0793a61d TG |
3117 | } |
3118 | ||
889ff015 FW |
3119 | static int event_enable_on_exec(struct perf_event *event, |
3120 | struct perf_event_context *ctx) | |
3121 | { | |
3122 | if (!event->attr.enable_on_exec) | |
3123 | return 0; | |
3124 | ||
3125 | event->attr.enable_on_exec = 0; | |
3126 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
3127 | return 0; | |
3128 | ||
1d9b482e | 3129 | __perf_event_mark_enabled(event); |
889ff015 FW |
3130 | |
3131 | return 1; | |
3132 | } | |
3133 | ||
57e7986e | 3134 | /* |
cdd6c482 | 3135 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
3136 | * This expects task == current. |
3137 | */ | |
8dc85d54 | 3138 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 3139 | { |
211de6eb | 3140 | struct perf_event_context *clone_ctx = NULL; |
cdd6c482 | 3141 | struct perf_event *event; |
57e7986e PM |
3142 | unsigned long flags; |
3143 | int enabled = 0; | |
889ff015 | 3144 | int ret; |
57e7986e PM |
3145 | |
3146 | local_irq_save(flags); | |
cdd6c482 | 3147 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
3148 | goto out; |
3149 | ||
e566b76e SE |
3150 | /* |
3151 | * We must ctxsw out cgroup events to avoid conflict | |
3152 | * when invoking perf_task_event_sched_in() later on | |
3153 | * in this function. Otherwise we end up trying to | |
3154 | * ctxswin cgroup events which are already scheduled | |
3155 | * in. | |
3156 | */ | |
a8d757ef | 3157 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 3158 | |
e625cce1 | 3159 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 3160 | task_ctx_sched_out(ctx); |
57e7986e | 3161 | |
b79387ef | 3162 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
3163 | ret = event_enable_on_exec(event, ctx); |
3164 | if (ret) | |
3165 | enabled = 1; | |
57e7986e PM |
3166 | } |
3167 | ||
3168 | /* | |
cdd6c482 | 3169 | * Unclone this context if we enabled any event. |
57e7986e | 3170 | */ |
71a851b4 | 3171 | if (enabled) |
211de6eb | 3172 | clone_ctx = unclone_ctx(ctx); |
57e7986e | 3173 | |
e625cce1 | 3174 | raw_spin_unlock(&ctx->lock); |
57e7986e | 3175 | |
e566b76e SE |
3176 | /* |
3177 | * Also calls ctxswin for cgroup events, if any: | |
3178 | */ | |
e5d1367f | 3179 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 3180 | out: |
57e7986e | 3181 | local_irq_restore(flags); |
211de6eb PZ |
3182 | |
3183 | if (clone_ctx) | |
3184 | put_ctx(clone_ctx); | |
57e7986e PM |
3185 | } |
3186 | ||
e041e328 PZ |
3187 | void perf_event_exec(void) |
3188 | { | |
3189 | struct perf_event_context *ctx; | |
3190 | int ctxn; | |
3191 | ||
3192 | rcu_read_lock(); | |
3193 | for_each_task_context_nr(ctxn) { | |
3194 | ctx = current->perf_event_ctxp[ctxn]; | |
3195 | if (!ctx) | |
3196 | continue; | |
3197 | ||
3198 | perf_event_enable_on_exec(ctx); | |
3199 | } | |
3200 | rcu_read_unlock(); | |
3201 | } | |
3202 | ||
0793a61d | 3203 | /* |
cdd6c482 | 3204 | * Cross CPU call to read the hardware event |
0793a61d | 3205 | */ |
cdd6c482 | 3206 | static void __perf_event_read(void *info) |
0793a61d | 3207 | { |
cdd6c482 IM |
3208 | struct perf_event *event = info; |
3209 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 3210 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3211 | |
e1ac3614 PM |
3212 | /* |
3213 | * If this is a task context, we need to check whether it is | |
3214 | * the current task context of this cpu. If not it has been | |
3215 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3216 | * event->count would have been updated to a recent sample |
3217 | * when the event was scheduled out. | |
e1ac3614 PM |
3218 | */ |
3219 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3220 | return; | |
3221 | ||
e625cce1 | 3222 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3223 | if (ctx->is_active) { |
542e72fc | 3224 | update_context_time(ctx); |
e5d1367f SE |
3225 | update_cgrp_time_from_event(event); |
3226 | } | |
cdd6c482 | 3227 | update_event_times(event); |
542e72fc PZ |
3228 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3229 | event->pmu->read(event); | |
e625cce1 | 3230 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3231 | } |
3232 | ||
b5e58793 PZ |
3233 | static inline u64 perf_event_count(struct perf_event *event) |
3234 | { | |
e7850595 | 3235 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
3236 | } |
3237 | ||
cdd6c482 | 3238 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3239 | { |
3240 | /* | |
cdd6c482 IM |
3241 | * If event is enabled and currently active on a CPU, update the |
3242 | * value in the event structure: | |
0793a61d | 3243 | */ |
cdd6c482 IM |
3244 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3245 | smp_call_function_single(event->oncpu, | |
3246 | __perf_event_read, event, 1); | |
3247 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3248 | struct perf_event_context *ctx = event->ctx; |
3249 | unsigned long flags; | |
3250 | ||
e625cce1 | 3251 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3252 | /* |
3253 | * may read while context is not active | |
3254 | * (e.g., thread is blocked), in that case | |
3255 | * we cannot update context time | |
3256 | */ | |
e5d1367f | 3257 | if (ctx->is_active) { |
c530ccd9 | 3258 | update_context_time(ctx); |
e5d1367f SE |
3259 | update_cgrp_time_from_event(event); |
3260 | } | |
cdd6c482 | 3261 | update_event_times(event); |
e625cce1 | 3262 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3263 | } |
3264 | ||
b5e58793 | 3265 | return perf_event_count(event); |
0793a61d TG |
3266 | } |
3267 | ||
a63eaf34 | 3268 | /* |
cdd6c482 | 3269 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3270 | */ |
eb184479 | 3271 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3272 | { |
e625cce1 | 3273 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3274 | mutex_init(&ctx->mutex); |
889ff015 FW |
3275 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3276 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3277 | INIT_LIST_HEAD(&ctx->event_list); |
3278 | atomic_set(&ctx->refcount, 1); | |
fadfe7be | 3279 | INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work); |
eb184479 PZ |
3280 | } |
3281 | ||
3282 | static struct perf_event_context * | |
3283 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3284 | { | |
3285 | struct perf_event_context *ctx; | |
3286 | ||
3287 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3288 | if (!ctx) | |
3289 | return NULL; | |
3290 | ||
3291 | __perf_event_init_context(ctx); | |
3292 | if (task) { | |
3293 | ctx->task = task; | |
3294 | get_task_struct(task); | |
0793a61d | 3295 | } |
eb184479 PZ |
3296 | ctx->pmu = pmu; |
3297 | ||
3298 | return ctx; | |
a63eaf34 PM |
3299 | } |
3300 | ||
2ebd4ffb MH |
3301 | static struct task_struct * |
3302 | find_lively_task_by_vpid(pid_t vpid) | |
3303 | { | |
3304 | struct task_struct *task; | |
3305 | int err; | |
0793a61d TG |
3306 | |
3307 | rcu_read_lock(); | |
2ebd4ffb | 3308 | if (!vpid) |
0793a61d TG |
3309 | task = current; |
3310 | else | |
2ebd4ffb | 3311 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3312 | if (task) |
3313 | get_task_struct(task); | |
3314 | rcu_read_unlock(); | |
3315 | ||
3316 | if (!task) | |
3317 | return ERR_PTR(-ESRCH); | |
3318 | ||
0793a61d | 3319 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3320 | err = -EACCES; |
3321 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3322 | goto errout; | |
3323 | ||
2ebd4ffb MH |
3324 | return task; |
3325 | errout: | |
3326 | put_task_struct(task); | |
3327 | return ERR_PTR(err); | |
3328 | ||
3329 | } | |
3330 | ||
fe4b04fa PZ |
3331 | /* |
3332 | * Returns a matching context with refcount and pincount. | |
3333 | */ | |
108b02cf | 3334 | static struct perf_event_context * |
38a81da2 | 3335 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3336 | { |
211de6eb | 3337 | struct perf_event_context *ctx, *clone_ctx = NULL; |
22a4f650 | 3338 | struct perf_cpu_context *cpuctx; |
25346b93 | 3339 | unsigned long flags; |
8dc85d54 | 3340 | int ctxn, err; |
0793a61d | 3341 | |
22a4ec72 | 3342 | if (!task) { |
cdd6c482 | 3343 | /* Must be root to operate on a CPU event: */ |
0764771d | 3344 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3345 | return ERR_PTR(-EACCES); |
3346 | ||
0793a61d | 3347 | /* |
cdd6c482 | 3348 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3349 | * offline CPU and activate it when the CPU comes up, but |
3350 | * that's for later. | |
3351 | */ | |
f6325e30 | 3352 | if (!cpu_online(cpu)) |
0793a61d TG |
3353 | return ERR_PTR(-ENODEV); |
3354 | ||
108b02cf | 3355 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3356 | ctx = &cpuctx->ctx; |
c93f7669 | 3357 | get_ctx(ctx); |
fe4b04fa | 3358 | ++ctx->pin_count; |
0793a61d | 3359 | |
0793a61d TG |
3360 | return ctx; |
3361 | } | |
3362 | ||
8dc85d54 PZ |
3363 | err = -EINVAL; |
3364 | ctxn = pmu->task_ctx_nr; | |
3365 | if (ctxn < 0) | |
3366 | goto errout; | |
3367 | ||
9ed6060d | 3368 | retry: |
8dc85d54 | 3369 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3370 | if (ctx) { |
211de6eb | 3371 | clone_ctx = unclone_ctx(ctx); |
fe4b04fa | 3372 | ++ctx->pin_count; |
e625cce1 | 3373 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
211de6eb PZ |
3374 | |
3375 | if (clone_ctx) | |
3376 | put_ctx(clone_ctx); | |
9137fb28 | 3377 | } else { |
eb184479 | 3378 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3379 | err = -ENOMEM; |
3380 | if (!ctx) | |
3381 | goto errout; | |
eb184479 | 3382 | |
dbe08d82 ON |
3383 | err = 0; |
3384 | mutex_lock(&task->perf_event_mutex); | |
3385 | /* | |
3386 | * If it has already passed perf_event_exit_task(). | |
3387 | * we must see PF_EXITING, it takes this mutex too. | |
3388 | */ | |
3389 | if (task->flags & PF_EXITING) | |
3390 | err = -ESRCH; | |
3391 | else if (task->perf_event_ctxp[ctxn]) | |
3392 | err = -EAGAIN; | |
fe4b04fa | 3393 | else { |
9137fb28 | 3394 | get_ctx(ctx); |
fe4b04fa | 3395 | ++ctx->pin_count; |
dbe08d82 | 3396 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3397 | } |
dbe08d82 ON |
3398 | mutex_unlock(&task->perf_event_mutex); |
3399 | ||
3400 | if (unlikely(err)) { | |
9137fb28 | 3401 | put_ctx(ctx); |
dbe08d82 ON |
3402 | |
3403 | if (err == -EAGAIN) | |
3404 | goto retry; | |
3405 | goto errout; | |
a63eaf34 PM |
3406 | } |
3407 | } | |
3408 | ||
0793a61d | 3409 | return ctx; |
c93f7669 | 3410 | |
9ed6060d | 3411 | errout: |
c93f7669 | 3412 | return ERR_PTR(err); |
0793a61d TG |
3413 | } |
3414 | ||
6fb2915d LZ |
3415 | static void perf_event_free_filter(struct perf_event *event); |
3416 | ||
cdd6c482 | 3417 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3418 | { |
cdd6c482 | 3419 | struct perf_event *event; |
592903cd | 3420 | |
cdd6c482 IM |
3421 | event = container_of(head, struct perf_event, rcu_head); |
3422 | if (event->ns) | |
3423 | put_pid_ns(event->ns); | |
6fb2915d | 3424 | perf_event_free_filter(event); |
cdd6c482 | 3425 | kfree(event); |
592903cd PZ |
3426 | } |
3427 | ||
76369139 | 3428 | static void ring_buffer_put(struct ring_buffer *rb); |
b69cf536 PZ |
3429 | static void ring_buffer_attach(struct perf_event *event, |
3430 | struct ring_buffer *rb); | |
925d519a | 3431 | |
4beb31f3 | 3432 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3433 | { |
4beb31f3 FW |
3434 | if (event->parent) |
3435 | return; | |
3436 | ||
3437 | if (has_branch_stack(event)) { | |
3438 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3439 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3440 | } | |
3441 | if (is_cgroup_event(event)) | |
3442 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3443 | } | |
925d519a | 3444 | |
4beb31f3 FW |
3445 | static void unaccount_event(struct perf_event *event) |
3446 | { | |
3447 | if (event->parent) | |
3448 | return; | |
3449 | ||
3450 | if (event->attach_state & PERF_ATTACH_TASK) | |
3451 | static_key_slow_dec_deferred(&perf_sched_events); | |
3452 | if (event->attr.mmap || event->attr.mmap_data) | |
3453 | atomic_dec(&nr_mmap_events); | |
3454 | if (event->attr.comm) | |
3455 | atomic_dec(&nr_comm_events); | |
3456 | if (event->attr.task) | |
3457 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3458 | if (event->attr.freq) |
3459 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3460 | if (is_cgroup_event(event)) |
3461 | static_key_slow_dec_deferred(&perf_sched_events); | |
3462 | if (has_branch_stack(event)) | |
3463 | static_key_slow_dec_deferred(&perf_sched_events); | |
3464 | ||
3465 | unaccount_event_cpu(event, event->cpu); | |
3466 | } | |
925d519a | 3467 | |
766d6c07 FW |
3468 | static void __free_event(struct perf_event *event) |
3469 | { | |
cdd6c482 | 3470 | if (!event->parent) { |
927c7a9e FW |
3471 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3472 | put_callchain_buffers(); | |
f344011c | 3473 | } |
9ee318a7 | 3474 | |
766d6c07 FW |
3475 | if (event->destroy) |
3476 | event->destroy(event); | |
3477 | ||
3478 | if (event->ctx) | |
3479 | put_ctx(event->ctx); | |
3480 | ||
c464c76e YZ |
3481 | if (event->pmu) |
3482 | module_put(event->pmu->module); | |
3483 | ||
766d6c07 FW |
3484 | call_rcu(&event->rcu_head, free_event_rcu); |
3485 | } | |
683ede43 PZ |
3486 | |
3487 | static void _free_event(struct perf_event *event) | |
f1600952 | 3488 | { |
e360adbe | 3489 | irq_work_sync(&event->pending); |
925d519a | 3490 | |
4beb31f3 | 3491 | unaccount_event(event); |
9ee318a7 | 3492 | |
76369139 | 3493 | if (event->rb) { |
9bb5d40c PZ |
3494 | /* |
3495 | * Can happen when we close an event with re-directed output. | |
3496 | * | |
3497 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3498 | * over us; possibly making our ring_buffer_put() the last. | |
3499 | */ | |
3500 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3501 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3502 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3503 | } |
3504 | ||
e5d1367f SE |
3505 | if (is_cgroup_event(event)) |
3506 | perf_detach_cgroup(event); | |
3507 | ||
766d6c07 | 3508 | __free_event(event); |
f1600952 PZ |
3509 | } |
3510 | ||
683ede43 PZ |
3511 | /* |
3512 | * Used to free events which have a known refcount of 1, such as in error paths | |
3513 | * where the event isn't exposed yet and inherited events. | |
3514 | */ | |
3515 | static void free_event(struct perf_event *event) | |
0793a61d | 3516 | { |
683ede43 PZ |
3517 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3518 | "unexpected event refcount: %ld; ptr=%p\n", | |
3519 | atomic_long_read(&event->refcount), event)) { | |
3520 | /* leak to avoid use-after-free */ | |
3521 | return; | |
3522 | } | |
0793a61d | 3523 | |
683ede43 | 3524 | _free_event(event); |
0793a61d TG |
3525 | } |
3526 | ||
a66a3052 | 3527 | /* |
f8697762 | 3528 | * Remove user event from the owner task. |
a66a3052 | 3529 | */ |
f8697762 | 3530 | static void perf_remove_from_owner(struct perf_event *event) |
fb0459d7 | 3531 | { |
8882135b | 3532 | struct task_struct *owner; |
fb0459d7 | 3533 | |
8882135b PZ |
3534 | rcu_read_lock(); |
3535 | owner = ACCESS_ONCE(event->owner); | |
3536 | /* | |
3537 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3538 | * !owner it means the list deletion is complete and we can indeed | |
3539 | * free this event, otherwise we need to serialize on | |
3540 | * owner->perf_event_mutex. | |
3541 | */ | |
3542 | smp_read_barrier_depends(); | |
3543 | if (owner) { | |
3544 | /* | |
3545 | * Since delayed_put_task_struct() also drops the last | |
3546 | * task reference we can safely take a new reference | |
3547 | * while holding the rcu_read_lock(). | |
3548 | */ | |
3549 | get_task_struct(owner); | |
3550 | } | |
3551 | rcu_read_unlock(); | |
3552 | ||
3553 | if (owner) { | |
f63a8daa PZ |
3554 | /* |
3555 | * If we're here through perf_event_exit_task() we're already | |
3556 | * holding ctx->mutex which would be an inversion wrt. the | |
3557 | * normal lock order. | |
3558 | * | |
3559 | * However we can safely take this lock because its the child | |
3560 | * ctx->mutex. | |
3561 | */ | |
3562 | mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); | |
3563 | ||
8882135b PZ |
3564 | /* |
3565 | * We have to re-check the event->owner field, if it is cleared | |
3566 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3567 | * ensured they're done, and we can proceed with freeing the | |
3568 | * event. | |
3569 | */ | |
3570 | if (event->owner) | |
3571 | list_del_init(&event->owner_entry); | |
3572 | mutex_unlock(&owner->perf_event_mutex); | |
3573 | put_task_struct(owner); | |
3574 | } | |
f8697762 JO |
3575 | } |
3576 | ||
3577 | /* | |
3578 | * Called when the last reference to the file is gone. | |
3579 | */ | |
3580 | static void put_event(struct perf_event *event) | |
3581 | { | |
a83fe28e | 3582 | struct perf_event_context *ctx; |
f8697762 JO |
3583 | |
3584 | if (!atomic_long_dec_and_test(&event->refcount)) | |
3585 | return; | |
3586 | ||
3587 | if (!is_kernel_event(event)) | |
3588 | perf_remove_from_owner(event); | |
8882135b | 3589 | |
683ede43 PZ |
3590 | /* |
3591 | * There are two ways this annotation is useful: | |
3592 | * | |
3593 | * 1) there is a lock recursion from perf_event_exit_task | |
3594 | * see the comment there. | |
3595 | * | |
3596 | * 2) there is a lock-inversion with mmap_sem through | |
3597 | * perf_event_read_group(), which takes faults while | |
3598 | * holding ctx->mutex, however this is called after | |
3599 | * the last filedesc died, so there is no possibility | |
3600 | * to trigger the AB-BA case. | |
3601 | */ | |
a83fe28e PZ |
3602 | ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING); |
3603 | WARN_ON_ONCE(ctx->parent_ctx); | |
683ede43 PZ |
3604 | perf_remove_from_context(event, true); |
3605 | mutex_unlock(&ctx->mutex); | |
3606 | ||
3607 | _free_event(event); | |
a6fa941d AV |
3608 | } |
3609 | ||
683ede43 PZ |
3610 | int perf_event_release_kernel(struct perf_event *event) |
3611 | { | |
3612 | put_event(event); | |
3613 | return 0; | |
3614 | } | |
3615 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3616 | ||
a6fa941d AV |
3617 | static int perf_release(struct inode *inode, struct file *file) |
3618 | { | |
3619 | put_event(file->private_data); | |
3620 | return 0; | |
fb0459d7 | 3621 | } |
fb0459d7 | 3622 | |
fadfe7be JO |
3623 | /* |
3624 | * Remove all orphanes events from the context. | |
3625 | */ | |
3626 | static void orphans_remove_work(struct work_struct *work) | |
3627 | { | |
3628 | struct perf_event_context *ctx; | |
3629 | struct perf_event *event, *tmp; | |
3630 | ||
3631 | ctx = container_of(work, struct perf_event_context, | |
3632 | orphans_remove.work); | |
3633 | ||
3634 | mutex_lock(&ctx->mutex); | |
3635 | list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) { | |
3636 | struct perf_event *parent_event = event->parent; | |
3637 | ||
3638 | if (!is_orphaned_child(event)) | |
3639 | continue; | |
3640 | ||
3641 | perf_remove_from_context(event, true); | |
3642 | ||
3643 | mutex_lock(&parent_event->child_mutex); | |
3644 | list_del_init(&event->child_list); | |
3645 | mutex_unlock(&parent_event->child_mutex); | |
3646 | ||
3647 | free_event(event); | |
3648 | put_event(parent_event); | |
3649 | } | |
3650 | ||
3651 | raw_spin_lock_irq(&ctx->lock); | |
3652 | ctx->orphans_remove_sched = false; | |
3653 | raw_spin_unlock_irq(&ctx->lock); | |
3654 | mutex_unlock(&ctx->mutex); | |
3655 | ||
3656 | put_ctx(ctx); | |
3657 | } | |
3658 | ||
59ed446f | 3659 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3660 | { |
cdd6c482 | 3661 | struct perf_event *child; |
e53c0994 PZ |
3662 | u64 total = 0; |
3663 | ||
59ed446f PZ |
3664 | *enabled = 0; |
3665 | *running = 0; | |
3666 | ||
6f10581a | 3667 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3668 | total += perf_event_read(event); |
59ed446f PZ |
3669 | *enabled += event->total_time_enabled + |
3670 | atomic64_read(&event->child_total_time_enabled); | |
3671 | *running += event->total_time_running + | |
3672 | atomic64_read(&event->child_total_time_running); | |
3673 | ||
3674 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3675 | total += perf_event_read(child); |
59ed446f PZ |
3676 | *enabled += child->total_time_enabled; |
3677 | *running += child->total_time_running; | |
3678 | } | |
6f10581a | 3679 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3680 | |
3681 | return total; | |
3682 | } | |
fb0459d7 | 3683 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3684 | |
cdd6c482 | 3685 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3686 | u64 read_format, char __user *buf) |
3687 | { | |
cdd6c482 | 3688 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a | 3689 | struct perf_event_context *ctx = leader->ctx; |
f63a8daa | 3690 | int n = 0, size = 0, ret; |
59ed446f | 3691 | u64 count, enabled, running; |
f63a8daa PZ |
3692 | u64 values[5]; |
3693 | ||
3694 | lockdep_assert_held(&ctx->mutex); | |
abf4868b | 3695 | |
59ed446f | 3696 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3697 | |
3698 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3699 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3700 | values[n++] = enabled; | |
3701 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3702 | values[n++] = running; | |
abf4868b PZ |
3703 | values[n++] = count; |
3704 | if (read_format & PERF_FORMAT_ID) | |
3705 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3706 | |
3707 | size = n * sizeof(u64); | |
3708 | ||
3709 | if (copy_to_user(buf, values, size)) | |
f63a8daa | 3710 | return -EFAULT; |
3dab77fb | 3711 | |
6f10581a | 3712 | ret = size; |
3dab77fb | 3713 | |
65abc865 | 3714 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3715 | n = 0; |
3dab77fb | 3716 | |
59ed446f | 3717 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3718 | if (read_format & PERF_FORMAT_ID) |
3719 | values[n++] = primary_event_id(sub); | |
3720 | ||
3721 | size = n * sizeof(u64); | |
3722 | ||
184d3da8 | 3723 | if (copy_to_user(buf + ret, values, size)) { |
f63a8daa | 3724 | return -EFAULT; |
6f10581a | 3725 | } |
abf4868b PZ |
3726 | |
3727 | ret += size; | |
3dab77fb PZ |
3728 | } |
3729 | ||
abf4868b | 3730 | return ret; |
3dab77fb PZ |
3731 | } |
3732 | ||
cdd6c482 | 3733 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3734 | u64 read_format, char __user *buf) |
3735 | { | |
59ed446f | 3736 | u64 enabled, running; |
3dab77fb PZ |
3737 | u64 values[4]; |
3738 | int n = 0; | |
3739 | ||
59ed446f PZ |
3740 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3741 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3742 | values[n++] = enabled; | |
3743 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3744 | values[n++] = running; | |
3dab77fb | 3745 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3746 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3747 | |
3748 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3749 | return -EFAULT; | |
3750 | ||
3751 | return n * sizeof(u64); | |
3752 | } | |
3753 | ||
dc633982 JO |
3754 | static bool is_event_hup(struct perf_event *event) |
3755 | { | |
3756 | bool no_children; | |
3757 | ||
3758 | if (event->state != PERF_EVENT_STATE_EXIT) | |
3759 | return false; | |
3760 | ||
3761 | mutex_lock(&event->child_mutex); | |
3762 | no_children = list_empty(&event->child_list); | |
3763 | mutex_unlock(&event->child_mutex); | |
3764 | return no_children; | |
3765 | } | |
3766 | ||
0793a61d | 3767 | /* |
cdd6c482 | 3768 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3769 | */ |
3770 | static ssize_t | |
cdd6c482 | 3771 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3772 | { |
cdd6c482 | 3773 | u64 read_format = event->attr.read_format; |
3dab77fb | 3774 | int ret; |
0793a61d | 3775 | |
3b6f9e5c | 3776 | /* |
cdd6c482 | 3777 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3778 | * error state (i.e. because it was pinned but it couldn't be |
3779 | * scheduled on to the CPU at some point). | |
3780 | */ | |
cdd6c482 | 3781 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3782 | return 0; |
3783 | ||
c320c7b7 | 3784 | if (count < event->read_size) |
3dab77fb PZ |
3785 | return -ENOSPC; |
3786 | ||
cdd6c482 | 3787 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3788 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3789 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3790 | else |
cdd6c482 | 3791 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3792 | |
3dab77fb | 3793 | return ret; |
0793a61d TG |
3794 | } |
3795 | ||
0793a61d TG |
3796 | static ssize_t |
3797 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3798 | { | |
cdd6c482 | 3799 | struct perf_event *event = file->private_data; |
f63a8daa PZ |
3800 | struct perf_event_context *ctx; |
3801 | int ret; | |
0793a61d | 3802 | |
f63a8daa PZ |
3803 | ctx = perf_event_ctx_lock(event); |
3804 | ret = perf_read_hw(event, buf, count); | |
3805 | perf_event_ctx_unlock(event, ctx); | |
3806 | ||
3807 | return ret; | |
0793a61d TG |
3808 | } |
3809 | ||
3810 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3811 | { | |
cdd6c482 | 3812 | struct perf_event *event = file->private_data; |
76369139 | 3813 | struct ring_buffer *rb; |
61b67684 | 3814 | unsigned int events = POLLHUP; |
c7138f37 | 3815 | |
e708d7ad | 3816 | poll_wait(file, &event->waitq, wait); |
179033b3 | 3817 | |
dc633982 | 3818 | if (is_event_hup(event)) |
179033b3 | 3819 | return events; |
c7138f37 | 3820 | |
10c6db11 | 3821 | /* |
9bb5d40c PZ |
3822 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3823 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3824 | */ |
3825 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3826 | rb = event->rb; |
3827 | if (rb) | |
76369139 | 3828 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3829 | mutex_unlock(&event->mmap_mutex); |
0793a61d TG |
3830 | return events; |
3831 | } | |
3832 | ||
f63a8daa | 3833 | static void _perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3834 | { |
cdd6c482 | 3835 | (void)perf_event_read(event); |
e7850595 | 3836 | local64_set(&event->count, 0); |
cdd6c482 | 3837 | perf_event_update_userpage(event); |
3df5edad PZ |
3838 | } |
3839 | ||
c93f7669 | 3840 | /* |
cdd6c482 IM |
3841 | * Holding the top-level event's child_mutex means that any |
3842 | * descendant process that has inherited this event will block | |
3843 | * in sync_child_event if it goes to exit, thus satisfying the | |
3844 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3845 | */ |
cdd6c482 IM |
3846 | static void perf_event_for_each_child(struct perf_event *event, |
3847 | void (*func)(struct perf_event *)) | |
3df5edad | 3848 | { |
cdd6c482 | 3849 | struct perf_event *child; |
3df5edad | 3850 | |
cdd6c482 | 3851 | WARN_ON_ONCE(event->ctx->parent_ctx); |
f63a8daa | 3852 | |
cdd6c482 IM |
3853 | mutex_lock(&event->child_mutex); |
3854 | func(event); | |
3855 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3856 | func(child); |
cdd6c482 | 3857 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3858 | } |
3859 | ||
cdd6c482 IM |
3860 | static void perf_event_for_each(struct perf_event *event, |
3861 | void (*func)(struct perf_event *)) | |
3df5edad | 3862 | { |
cdd6c482 IM |
3863 | struct perf_event_context *ctx = event->ctx; |
3864 | struct perf_event *sibling; | |
3df5edad | 3865 | |
f63a8daa PZ |
3866 | lockdep_assert_held(&ctx->mutex); |
3867 | ||
cdd6c482 | 3868 | event = event->group_leader; |
75f937f2 | 3869 | |
cdd6c482 | 3870 | perf_event_for_each_child(event, func); |
cdd6c482 | 3871 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3872 | perf_event_for_each_child(sibling, func); |
6de6a7b9 PZ |
3873 | } |
3874 | ||
cdd6c482 | 3875 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3876 | { |
cdd6c482 | 3877 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3878 | int ret = 0, active; |
08247e31 PZ |
3879 | u64 value; |
3880 | ||
6c7e550f | 3881 | if (!is_sampling_event(event)) |
08247e31 PZ |
3882 | return -EINVAL; |
3883 | ||
ad0cf347 | 3884 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3885 | return -EFAULT; |
3886 | ||
3887 | if (!value) | |
3888 | return -EINVAL; | |
3889 | ||
e625cce1 | 3890 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3891 | if (event->attr.freq) { |
3892 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3893 | ret = -EINVAL; |
3894 | goto unlock; | |
3895 | } | |
3896 | ||
cdd6c482 | 3897 | event->attr.sample_freq = value; |
08247e31 | 3898 | } else { |
cdd6c482 IM |
3899 | event->attr.sample_period = value; |
3900 | event->hw.sample_period = value; | |
08247e31 | 3901 | } |
bad7192b PZ |
3902 | |
3903 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3904 | if (active) { | |
3905 | perf_pmu_disable(ctx->pmu); | |
3906 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3907 | } | |
3908 | ||
3909 | local64_set(&event->hw.period_left, 0); | |
3910 | ||
3911 | if (active) { | |
3912 | event->pmu->start(event, PERF_EF_RELOAD); | |
3913 | perf_pmu_enable(ctx->pmu); | |
3914 | } | |
3915 | ||
08247e31 | 3916 | unlock: |
e625cce1 | 3917 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3918 | |
3919 | return ret; | |
3920 | } | |
3921 | ||
ac9721f3 PZ |
3922 | static const struct file_operations perf_fops; |
3923 | ||
2903ff01 | 3924 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3925 | { |
2903ff01 AV |
3926 | struct fd f = fdget(fd); |
3927 | if (!f.file) | |
3928 | return -EBADF; | |
ac9721f3 | 3929 | |
2903ff01 AV |
3930 | if (f.file->f_op != &perf_fops) { |
3931 | fdput(f); | |
3932 | return -EBADF; | |
ac9721f3 | 3933 | } |
2903ff01 AV |
3934 | *p = f; |
3935 | return 0; | |
ac9721f3 PZ |
3936 | } |
3937 | ||
3938 | static int perf_event_set_output(struct perf_event *event, | |
3939 | struct perf_event *output_event); | |
6fb2915d | 3940 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3941 | |
f63a8daa | 3942 | static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg) |
d859e29f | 3943 | { |
cdd6c482 | 3944 | void (*func)(struct perf_event *); |
3df5edad | 3945 | u32 flags = arg; |
d859e29f PM |
3946 | |
3947 | switch (cmd) { | |
cdd6c482 | 3948 | case PERF_EVENT_IOC_ENABLE: |
f63a8daa | 3949 | func = _perf_event_enable; |
d859e29f | 3950 | break; |
cdd6c482 | 3951 | case PERF_EVENT_IOC_DISABLE: |
f63a8daa | 3952 | func = _perf_event_disable; |
79f14641 | 3953 | break; |
cdd6c482 | 3954 | case PERF_EVENT_IOC_RESET: |
f63a8daa | 3955 | func = _perf_event_reset; |
6de6a7b9 | 3956 | break; |
3df5edad | 3957 | |
cdd6c482 | 3958 | case PERF_EVENT_IOC_REFRESH: |
f63a8daa | 3959 | return _perf_event_refresh(event, arg); |
08247e31 | 3960 | |
cdd6c482 IM |
3961 | case PERF_EVENT_IOC_PERIOD: |
3962 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3963 | |
cf4957f1 JO |
3964 | case PERF_EVENT_IOC_ID: |
3965 | { | |
3966 | u64 id = primary_event_id(event); | |
3967 | ||
3968 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3969 | return -EFAULT; | |
3970 | return 0; | |
3971 | } | |
3972 | ||
cdd6c482 | 3973 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3974 | { |
ac9721f3 | 3975 | int ret; |
ac9721f3 | 3976 | if (arg != -1) { |
2903ff01 AV |
3977 | struct perf_event *output_event; |
3978 | struct fd output; | |
3979 | ret = perf_fget_light(arg, &output); | |
3980 | if (ret) | |
3981 | return ret; | |
3982 | output_event = output.file->private_data; | |
3983 | ret = perf_event_set_output(event, output_event); | |
3984 | fdput(output); | |
3985 | } else { | |
3986 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3987 | } |
ac9721f3 PZ |
3988 | return ret; |
3989 | } | |
a4be7c27 | 3990 | |
6fb2915d LZ |
3991 | case PERF_EVENT_IOC_SET_FILTER: |
3992 | return perf_event_set_filter(event, (void __user *)arg); | |
3993 | ||
d859e29f | 3994 | default: |
3df5edad | 3995 | return -ENOTTY; |
d859e29f | 3996 | } |
3df5edad PZ |
3997 | |
3998 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3999 | perf_event_for_each(event, func); |
3df5edad | 4000 | else |
cdd6c482 | 4001 | perf_event_for_each_child(event, func); |
3df5edad PZ |
4002 | |
4003 | return 0; | |
d859e29f PM |
4004 | } |
4005 | ||
f63a8daa PZ |
4006 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
4007 | { | |
4008 | struct perf_event *event = file->private_data; | |
4009 | struct perf_event_context *ctx; | |
4010 | long ret; | |
4011 | ||
4012 | ctx = perf_event_ctx_lock(event); | |
4013 | ret = _perf_ioctl(event, cmd, arg); | |
4014 | perf_event_ctx_unlock(event, ctx); | |
4015 | ||
4016 | return ret; | |
4017 | } | |
4018 | ||
b3f20785 PM |
4019 | #ifdef CONFIG_COMPAT |
4020 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, | |
4021 | unsigned long arg) | |
4022 | { | |
4023 | switch (_IOC_NR(cmd)) { | |
4024 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): | |
4025 | case _IOC_NR(PERF_EVENT_IOC_ID): | |
4026 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ | |
4027 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { | |
4028 | cmd &= ~IOCSIZE_MASK; | |
4029 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; | |
4030 | } | |
4031 | break; | |
4032 | } | |
4033 | return perf_ioctl(file, cmd, arg); | |
4034 | } | |
4035 | #else | |
4036 | # define perf_compat_ioctl NULL | |
4037 | #endif | |
4038 | ||
cdd6c482 | 4039 | int perf_event_task_enable(void) |
771d7cde | 4040 | { |
f63a8daa | 4041 | struct perf_event_context *ctx; |
cdd6c482 | 4042 | struct perf_event *event; |
771d7cde | 4043 | |
cdd6c482 | 4044 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4045 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4046 | ctx = perf_event_ctx_lock(event); | |
4047 | perf_event_for_each_child(event, _perf_event_enable); | |
4048 | perf_event_ctx_unlock(event, ctx); | |
4049 | } | |
cdd6c482 | 4050 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4051 | |
4052 | return 0; | |
4053 | } | |
4054 | ||
cdd6c482 | 4055 | int perf_event_task_disable(void) |
771d7cde | 4056 | { |
f63a8daa | 4057 | struct perf_event_context *ctx; |
cdd6c482 | 4058 | struct perf_event *event; |
771d7cde | 4059 | |
cdd6c482 | 4060 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4061 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4062 | ctx = perf_event_ctx_lock(event); | |
4063 | perf_event_for_each_child(event, _perf_event_disable); | |
4064 | perf_event_ctx_unlock(event, ctx); | |
4065 | } | |
cdd6c482 | 4066 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4067 | |
4068 | return 0; | |
4069 | } | |
4070 | ||
cdd6c482 | 4071 | static int perf_event_index(struct perf_event *event) |
194002b2 | 4072 | { |
a4eaf7f1 PZ |
4073 | if (event->hw.state & PERF_HES_STOPPED) |
4074 | return 0; | |
4075 | ||
cdd6c482 | 4076 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
4077 | return 0; |
4078 | ||
35edc2a5 | 4079 | return event->pmu->event_idx(event); |
194002b2 PZ |
4080 | } |
4081 | ||
c4794295 | 4082 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 4083 | u64 *now, |
7f310a5d EM |
4084 | u64 *enabled, |
4085 | u64 *running) | |
c4794295 | 4086 | { |
e3f3541c | 4087 | u64 ctx_time; |
c4794295 | 4088 | |
e3f3541c PZ |
4089 | *now = perf_clock(); |
4090 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
4091 | *enabled = ctx_time - event->tstamp_enabled; |
4092 | *running = ctx_time - event->tstamp_running; | |
4093 | } | |
4094 | ||
fa731587 PZ |
4095 | static void perf_event_init_userpage(struct perf_event *event) |
4096 | { | |
4097 | struct perf_event_mmap_page *userpg; | |
4098 | struct ring_buffer *rb; | |
4099 | ||
4100 | rcu_read_lock(); | |
4101 | rb = rcu_dereference(event->rb); | |
4102 | if (!rb) | |
4103 | goto unlock; | |
4104 | ||
4105 | userpg = rb->user_page; | |
4106 | ||
4107 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
4108 | userpg->cap_bit0_is_deprecated = 1; | |
4109 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
4110 | ||
4111 | unlock: | |
4112 | rcu_read_unlock(); | |
4113 | } | |
4114 | ||
c7206205 | 4115 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
4116 | { |
4117 | } | |
4118 | ||
38ff667b PZ |
4119 | /* |
4120 | * Callers need to ensure there can be no nesting of this function, otherwise | |
4121 | * the seqlock logic goes bad. We can not serialize this because the arch | |
4122 | * code calls this from NMI context. | |
4123 | */ | |
cdd6c482 | 4124 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 4125 | { |
cdd6c482 | 4126 | struct perf_event_mmap_page *userpg; |
76369139 | 4127 | struct ring_buffer *rb; |
e3f3541c | 4128 | u64 enabled, running, now; |
38ff667b PZ |
4129 | |
4130 | rcu_read_lock(); | |
5ec4c599 PZ |
4131 | rb = rcu_dereference(event->rb); |
4132 | if (!rb) | |
4133 | goto unlock; | |
4134 | ||
0d641208 EM |
4135 | /* |
4136 | * compute total_time_enabled, total_time_running | |
4137 | * based on snapshot values taken when the event | |
4138 | * was last scheduled in. | |
4139 | * | |
4140 | * we cannot simply called update_context_time() | |
4141 | * because of locking issue as we can be called in | |
4142 | * NMI context | |
4143 | */ | |
e3f3541c | 4144 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 4145 | |
76369139 | 4146 | userpg = rb->user_page; |
7b732a75 PZ |
4147 | /* |
4148 | * Disable preemption so as to not let the corresponding user-space | |
4149 | * spin too long if we get preempted. | |
4150 | */ | |
4151 | preempt_disable(); | |
37d81828 | 4152 | ++userpg->lock; |
92f22a38 | 4153 | barrier(); |
cdd6c482 | 4154 | userpg->index = perf_event_index(event); |
b5e58793 | 4155 | userpg->offset = perf_event_count(event); |
365a4038 | 4156 | if (userpg->index) |
e7850595 | 4157 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 4158 | |
0d641208 | 4159 | userpg->time_enabled = enabled + |
cdd6c482 | 4160 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 4161 | |
0d641208 | 4162 | userpg->time_running = running + |
cdd6c482 | 4163 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 4164 | |
c7206205 | 4165 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 4166 | |
92f22a38 | 4167 | barrier(); |
37d81828 | 4168 | ++userpg->lock; |
7b732a75 | 4169 | preempt_enable(); |
38ff667b | 4170 | unlock: |
7b732a75 | 4171 | rcu_read_unlock(); |
37d81828 PM |
4172 | } |
4173 | ||
906010b2 PZ |
4174 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
4175 | { | |
4176 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 4177 | struct ring_buffer *rb; |
906010b2 PZ |
4178 | int ret = VM_FAULT_SIGBUS; |
4179 | ||
4180 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
4181 | if (vmf->pgoff == 0) | |
4182 | ret = 0; | |
4183 | return ret; | |
4184 | } | |
4185 | ||
4186 | rcu_read_lock(); | |
76369139 FW |
4187 | rb = rcu_dereference(event->rb); |
4188 | if (!rb) | |
906010b2 PZ |
4189 | goto unlock; |
4190 | ||
4191 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
4192 | goto unlock; | |
4193 | ||
76369139 | 4194 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
4195 | if (!vmf->page) |
4196 | goto unlock; | |
4197 | ||
4198 | get_page(vmf->page); | |
4199 | vmf->page->mapping = vma->vm_file->f_mapping; | |
4200 | vmf->page->index = vmf->pgoff; | |
4201 | ||
4202 | ret = 0; | |
4203 | unlock: | |
4204 | rcu_read_unlock(); | |
4205 | ||
4206 | return ret; | |
4207 | } | |
4208 | ||
10c6db11 PZ |
4209 | static void ring_buffer_attach(struct perf_event *event, |
4210 | struct ring_buffer *rb) | |
4211 | { | |
b69cf536 | 4212 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
4213 | unsigned long flags; |
4214 | ||
b69cf536 PZ |
4215 | if (event->rb) { |
4216 | /* | |
4217 | * Should be impossible, we set this when removing | |
4218 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
4219 | */ | |
4220 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 4221 | |
b69cf536 PZ |
4222 | old_rb = event->rb; |
4223 | event->rcu_batches = get_state_synchronize_rcu(); | |
4224 | event->rcu_pending = 1; | |
10c6db11 | 4225 | |
b69cf536 PZ |
4226 | spin_lock_irqsave(&old_rb->event_lock, flags); |
4227 | list_del_rcu(&event->rb_entry); | |
4228 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
4229 | } | |
10c6db11 | 4230 | |
b69cf536 PZ |
4231 | if (event->rcu_pending && rb) { |
4232 | cond_synchronize_rcu(event->rcu_batches); | |
4233 | event->rcu_pending = 0; | |
4234 | } | |
10c6db11 | 4235 | |
b69cf536 PZ |
4236 | if (rb) { |
4237 | spin_lock_irqsave(&rb->event_lock, flags); | |
4238 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
4239 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
4240 | } | |
4241 | ||
4242 | rcu_assign_pointer(event->rb, rb); | |
4243 | ||
4244 | if (old_rb) { | |
4245 | ring_buffer_put(old_rb); | |
4246 | /* | |
4247 | * Since we detached before setting the new rb, so that we | |
4248 | * could attach the new rb, we could have missed a wakeup. | |
4249 | * Provide it now. | |
4250 | */ | |
4251 | wake_up_all(&event->waitq); | |
4252 | } | |
10c6db11 PZ |
4253 | } |
4254 | ||
4255 | static void ring_buffer_wakeup(struct perf_event *event) | |
4256 | { | |
4257 | struct ring_buffer *rb; | |
4258 | ||
4259 | rcu_read_lock(); | |
4260 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
4261 | if (rb) { |
4262 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
4263 | wake_up_all(&event->waitq); | |
4264 | } | |
10c6db11 PZ |
4265 | rcu_read_unlock(); |
4266 | } | |
4267 | ||
76369139 | 4268 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 4269 | { |
76369139 | 4270 | struct ring_buffer *rb; |
906010b2 | 4271 | |
76369139 FW |
4272 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
4273 | rb_free(rb); | |
7b732a75 PZ |
4274 | } |
4275 | ||
76369139 | 4276 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 4277 | { |
76369139 | 4278 | struct ring_buffer *rb; |
7b732a75 | 4279 | |
ac9721f3 | 4280 | rcu_read_lock(); |
76369139 FW |
4281 | rb = rcu_dereference(event->rb); |
4282 | if (rb) { | |
4283 | if (!atomic_inc_not_zero(&rb->refcount)) | |
4284 | rb = NULL; | |
ac9721f3 PZ |
4285 | } |
4286 | rcu_read_unlock(); | |
4287 | ||
76369139 | 4288 | return rb; |
ac9721f3 PZ |
4289 | } |
4290 | ||
76369139 | 4291 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 4292 | { |
76369139 | 4293 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 4294 | return; |
7b732a75 | 4295 | |
9bb5d40c | 4296 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 4297 | |
76369139 | 4298 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
4299 | } |
4300 | ||
4301 | static void perf_mmap_open(struct vm_area_struct *vma) | |
4302 | { | |
cdd6c482 | 4303 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4304 | |
cdd6c482 | 4305 | atomic_inc(&event->mmap_count); |
9bb5d40c | 4306 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
4307 | } |
4308 | ||
9bb5d40c PZ |
4309 | /* |
4310 | * A buffer can be mmap()ed multiple times; either directly through the same | |
4311 | * event, or through other events by use of perf_event_set_output(). | |
4312 | * | |
4313 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
4314 | * the buffer here, where we still have a VM context. This means we need | |
4315 | * to detach all events redirecting to us. | |
4316 | */ | |
7b732a75 PZ |
4317 | static void perf_mmap_close(struct vm_area_struct *vma) |
4318 | { | |
cdd6c482 | 4319 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4320 | |
b69cf536 | 4321 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
4322 | struct user_struct *mmap_user = rb->mmap_user; |
4323 | int mmap_locked = rb->mmap_locked; | |
4324 | unsigned long size = perf_data_size(rb); | |
789f90fc | 4325 | |
9bb5d40c PZ |
4326 | atomic_dec(&rb->mmap_count); |
4327 | ||
4328 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4329 | goto out_put; |
9bb5d40c | 4330 | |
b69cf536 | 4331 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4332 | mutex_unlock(&event->mmap_mutex); |
4333 | ||
4334 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4335 | if (atomic_read(&rb->mmap_count)) |
4336 | goto out_put; | |
ac9721f3 | 4337 | |
9bb5d40c PZ |
4338 | /* |
4339 | * No other mmap()s, detach from all other events that might redirect | |
4340 | * into the now unreachable buffer. Somewhat complicated by the | |
4341 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4342 | */ | |
4343 | again: | |
4344 | rcu_read_lock(); | |
4345 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4346 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4347 | /* | |
4348 | * This event is en-route to free_event() which will | |
4349 | * detach it and remove it from the list. | |
4350 | */ | |
4351 | continue; | |
4352 | } | |
4353 | rcu_read_unlock(); | |
789f90fc | 4354 | |
9bb5d40c PZ |
4355 | mutex_lock(&event->mmap_mutex); |
4356 | /* | |
4357 | * Check we didn't race with perf_event_set_output() which can | |
4358 | * swizzle the rb from under us while we were waiting to | |
4359 | * acquire mmap_mutex. | |
4360 | * | |
4361 | * If we find a different rb; ignore this event, a next | |
4362 | * iteration will no longer find it on the list. We have to | |
4363 | * still restart the iteration to make sure we're not now | |
4364 | * iterating the wrong list. | |
4365 | */ | |
b69cf536 PZ |
4366 | if (event->rb == rb) |
4367 | ring_buffer_attach(event, NULL); | |
4368 | ||
cdd6c482 | 4369 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4370 | put_event(event); |
ac9721f3 | 4371 | |
9bb5d40c PZ |
4372 | /* |
4373 | * Restart the iteration; either we're on the wrong list or | |
4374 | * destroyed its integrity by doing a deletion. | |
4375 | */ | |
4376 | goto again; | |
7b732a75 | 4377 | } |
9bb5d40c PZ |
4378 | rcu_read_unlock(); |
4379 | ||
4380 | /* | |
4381 | * It could be there's still a few 0-ref events on the list; they'll | |
4382 | * get cleaned up by free_event() -- they'll also still have their | |
4383 | * ref on the rb and will free it whenever they are done with it. | |
4384 | * | |
4385 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4386 | * undo the VM accounting. | |
4387 | */ | |
4388 | ||
4389 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4390 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4391 | free_uid(mmap_user); | |
4392 | ||
b69cf536 | 4393 | out_put: |
9bb5d40c | 4394 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4395 | } |
4396 | ||
f0f37e2f | 4397 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4398 | .open = perf_mmap_open, |
4399 | .close = perf_mmap_close, | |
4400 | .fault = perf_mmap_fault, | |
4401 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4402 | }; |
4403 | ||
4404 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4405 | { | |
cdd6c482 | 4406 | struct perf_event *event = file->private_data; |
22a4f650 | 4407 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4408 | struct user_struct *user = current_user(); |
22a4f650 | 4409 | unsigned long locked, lock_limit; |
76369139 | 4410 | struct ring_buffer *rb; |
7b732a75 PZ |
4411 | unsigned long vma_size; |
4412 | unsigned long nr_pages; | |
789f90fc | 4413 | long user_extra, extra; |
d57e34fd | 4414 | int ret = 0, flags = 0; |
37d81828 | 4415 | |
c7920614 PZ |
4416 | /* |
4417 | * Don't allow mmap() of inherited per-task counters. This would | |
4418 | * create a performance issue due to all children writing to the | |
76369139 | 4419 | * same rb. |
c7920614 PZ |
4420 | */ |
4421 | if (event->cpu == -1 && event->attr.inherit) | |
4422 | return -EINVAL; | |
4423 | ||
43a21ea8 | 4424 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4425 | return -EINVAL; |
7b732a75 PZ |
4426 | |
4427 | vma_size = vma->vm_end - vma->vm_start; | |
4428 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4429 | ||
7730d865 | 4430 | /* |
76369139 | 4431 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4432 | * can do bitmasks instead of modulo. |
4433 | */ | |
4434 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
4435 | return -EINVAL; |
4436 | ||
7b732a75 | 4437 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4438 | return -EINVAL; |
4439 | ||
7b732a75 PZ |
4440 | if (vma->vm_pgoff != 0) |
4441 | return -EINVAL; | |
37d81828 | 4442 | |
cdd6c482 | 4443 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4444 | again: |
cdd6c482 | 4445 | mutex_lock(&event->mmap_mutex); |
76369139 | 4446 | if (event->rb) { |
9bb5d40c | 4447 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4448 | ret = -EINVAL; |
9bb5d40c PZ |
4449 | goto unlock; |
4450 | } | |
4451 | ||
4452 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4453 | /* | |
4454 | * Raced against perf_mmap_close() through | |
4455 | * perf_event_set_output(). Try again, hope for better | |
4456 | * luck. | |
4457 | */ | |
4458 | mutex_unlock(&event->mmap_mutex); | |
4459 | goto again; | |
4460 | } | |
4461 | ||
ebb3c4c4 PZ |
4462 | goto unlock; |
4463 | } | |
4464 | ||
789f90fc | 4465 | user_extra = nr_pages + 1; |
cdd6c482 | 4466 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4467 | |
4468 | /* | |
4469 | * Increase the limit linearly with more CPUs: | |
4470 | */ | |
4471 | user_lock_limit *= num_online_cpus(); | |
4472 | ||
789f90fc | 4473 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4474 | |
789f90fc PZ |
4475 | extra = 0; |
4476 | if (user_locked > user_lock_limit) | |
4477 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4478 | |
78d7d407 | 4479 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4480 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4481 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4482 | |
459ec28a IM |
4483 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4484 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4485 | ret = -EPERM; |
4486 | goto unlock; | |
4487 | } | |
7b732a75 | 4488 | |
76369139 | 4489 | WARN_ON(event->rb); |
906010b2 | 4490 | |
d57e34fd | 4491 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4492 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4493 | |
4ec8363d VW |
4494 | rb = rb_alloc(nr_pages, |
4495 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4496 | event->cpu, flags); | |
4497 | ||
76369139 | 4498 | if (!rb) { |
ac9721f3 | 4499 | ret = -ENOMEM; |
ebb3c4c4 | 4500 | goto unlock; |
ac9721f3 | 4501 | } |
26cb63ad | 4502 | |
9bb5d40c | 4503 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4504 | rb->mmap_locked = extra; |
4505 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4506 | |
ac9721f3 | 4507 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4508 | vma->vm_mm->pinned_vm += extra; |
4509 | ||
9bb5d40c | 4510 | ring_buffer_attach(event, rb); |
ac9721f3 | 4511 | |
fa731587 | 4512 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4513 | perf_event_update_userpage(event); |
4514 | ||
ebb3c4c4 | 4515 | unlock: |
ac9721f3 PZ |
4516 | if (!ret) |
4517 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4518 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4519 | |
9bb5d40c PZ |
4520 | /* |
4521 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4522 | * vma. | |
4523 | */ | |
26cb63ad | 4524 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4525 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4526 | |
4527 | return ret; | |
37d81828 PM |
4528 | } |
4529 | ||
3c446b3d PZ |
4530 | static int perf_fasync(int fd, struct file *filp, int on) |
4531 | { | |
496ad9aa | 4532 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4533 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4534 | int retval; |
4535 | ||
4536 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4537 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4538 | mutex_unlock(&inode->i_mutex); |
4539 | ||
4540 | if (retval < 0) | |
4541 | return retval; | |
4542 | ||
4543 | return 0; | |
4544 | } | |
4545 | ||
0793a61d | 4546 | static const struct file_operations perf_fops = { |
3326c1ce | 4547 | .llseek = no_llseek, |
0793a61d TG |
4548 | .release = perf_release, |
4549 | .read = perf_read, | |
4550 | .poll = perf_poll, | |
d859e29f | 4551 | .unlocked_ioctl = perf_ioctl, |
b3f20785 | 4552 | .compat_ioctl = perf_compat_ioctl, |
37d81828 | 4553 | .mmap = perf_mmap, |
3c446b3d | 4554 | .fasync = perf_fasync, |
0793a61d TG |
4555 | }; |
4556 | ||
925d519a | 4557 | /* |
cdd6c482 | 4558 | * Perf event wakeup |
925d519a PZ |
4559 | * |
4560 | * If there's data, ensure we set the poll() state and publish everything | |
4561 | * to user-space before waking everybody up. | |
4562 | */ | |
4563 | ||
cdd6c482 | 4564 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4565 | { |
10c6db11 | 4566 | ring_buffer_wakeup(event); |
4c9e2542 | 4567 | |
cdd6c482 IM |
4568 | if (event->pending_kill) { |
4569 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4570 | event->pending_kill = 0; | |
4c9e2542 | 4571 | } |
925d519a PZ |
4572 | } |
4573 | ||
e360adbe | 4574 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4575 | { |
cdd6c482 IM |
4576 | struct perf_event *event = container_of(entry, |
4577 | struct perf_event, pending); | |
79f14641 | 4578 | |
cdd6c482 IM |
4579 | if (event->pending_disable) { |
4580 | event->pending_disable = 0; | |
4581 | __perf_event_disable(event); | |
79f14641 PZ |
4582 | } |
4583 | ||
cdd6c482 IM |
4584 | if (event->pending_wakeup) { |
4585 | event->pending_wakeup = 0; | |
4586 | perf_event_wakeup(event); | |
79f14641 PZ |
4587 | } |
4588 | } | |
4589 | ||
39447b38 ZY |
4590 | /* |
4591 | * We assume there is only KVM supporting the callbacks. | |
4592 | * Later on, we might change it to a list if there is | |
4593 | * another virtualization implementation supporting the callbacks. | |
4594 | */ | |
4595 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4596 | ||
4597 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4598 | { | |
4599 | perf_guest_cbs = cbs; | |
4600 | return 0; | |
4601 | } | |
4602 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4603 | ||
4604 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4605 | { | |
4606 | perf_guest_cbs = NULL; | |
4607 | return 0; | |
4608 | } | |
4609 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4610 | ||
4018994f JO |
4611 | static void |
4612 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4613 | struct pt_regs *regs, u64 mask) | |
4614 | { | |
4615 | int bit; | |
4616 | ||
4617 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4618 | sizeof(mask) * BITS_PER_BYTE) { | |
4619 | u64 val; | |
4620 | ||
4621 | val = perf_reg_value(regs, bit); | |
4622 | perf_output_put(handle, val); | |
4623 | } | |
4624 | } | |
4625 | ||
60e2364e | 4626 | static void perf_sample_regs_user(struct perf_regs *regs_user, |
88a7c26a AL |
4627 | struct pt_regs *regs, |
4628 | struct pt_regs *regs_user_copy) | |
4018994f | 4629 | { |
88a7c26a AL |
4630 | if (user_mode(regs)) { |
4631 | regs_user->abi = perf_reg_abi(current); | |
2565711f | 4632 | regs_user->regs = regs; |
88a7c26a AL |
4633 | } else if (current->mm) { |
4634 | perf_get_regs_user(regs_user, regs, regs_user_copy); | |
2565711f PZ |
4635 | } else { |
4636 | regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; | |
4637 | regs_user->regs = NULL; | |
4018994f JO |
4638 | } |
4639 | } | |
4640 | ||
60e2364e SE |
4641 | static void perf_sample_regs_intr(struct perf_regs *regs_intr, |
4642 | struct pt_regs *regs) | |
4643 | { | |
4644 | regs_intr->regs = regs; | |
4645 | regs_intr->abi = perf_reg_abi(current); | |
4646 | } | |
4647 | ||
4648 | ||
c5ebcedb JO |
4649 | /* |
4650 | * Get remaining task size from user stack pointer. | |
4651 | * | |
4652 | * It'd be better to take stack vma map and limit this more | |
4653 | * precisly, but there's no way to get it safely under interrupt, | |
4654 | * so using TASK_SIZE as limit. | |
4655 | */ | |
4656 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4657 | { | |
4658 | unsigned long addr = perf_user_stack_pointer(regs); | |
4659 | ||
4660 | if (!addr || addr >= TASK_SIZE) | |
4661 | return 0; | |
4662 | ||
4663 | return TASK_SIZE - addr; | |
4664 | } | |
4665 | ||
4666 | static u16 | |
4667 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4668 | struct pt_regs *regs) | |
4669 | { | |
4670 | u64 task_size; | |
4671 | ||
4672 | /* No regs, no stack pointer, no dump. */ | |
4673 | if (!regs) | |
4674 | return 0; | |
4675 | ||
4676 | /* | |
4677 | * Check if we fit in with the requested stack size into the: | |
4678 | * - TASK_SIZE | |
4679 | * If we don't, we limit the size to the TASK_SIZE. | |
4680 | * | |
4681 | * - remaining sample size | |
4682 | * If we don't, we customize the stack size to | |
4683 | * fit in to the remaining sample size. | |
4684 | */ | |
4685 | ||
4686 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4687 | stack_size = min(stack_size, (u16) task_size); | |
4688 | ||
4689 | /* Current header size plus static size and dynamic size. */ | |
4690 | header_size += 2 * sizeof(u64); | |
4691 | ||
4692 | /* Do we fit in with the current stack dump size? */ | |
4693 | if ((u16) (header_size + stack_size) < header_size) { | |
4694 | /* | |
4695 | * If we overflow the maximum size for the sample, | |
4696 | * we customize the stack dump size to fit in. | |
4697 | */ | |
4698 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4699 | stack_size = round_up(stack_size, sizeof(u64)); | |
4700 | } | |
4701 | ||
4702 | return stack_size; | |
4703 | } | |
4704 | ||
4705 | static void | |
4706 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4707 | struct pt_regs *regs) | |
4708 | { | |
4709 | /* Case of a kernel thread, nothing to dump */ | |
4710 | if (!regs) { | |
4711 | u64 size = 0; | |
4712 | perf_output_put(handle, size); | |
4713 | } else { | |
4714 | unsigned long sp; | |
4715 | unsigned int rem; | |
4716 | u64 dyn_size; | |
4717 | ||
4718 | /* | |
4719 | * We dump: | |
4720 | * static size | |
4721 | * - the size requested by user or the best one we can fit | |
4722 | * in to the sample max size | |
4723 | * data | |
4724 | * - user stack dump data | |
4725 | * dynamic size | |
4726 | * - the actual dumped size | |
4727 | */ | |
4728 | ||
4729 | /* Static size. */ | |
4730 | perf_output_put(handle, dump_size); | |
4731 | ||
4732 | /* Data. */ | |
4733 | sp = perf_user_stack_pointer(regs); | |
4734 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4735 | dyn_size = dump_size - rem; | |
4736 | ||
4737 | perf_output_skip(handle, rem); | |
4738 | ||
4739 | /* Dynamic size. */ | |
4740 | perf_output_put(handle, dyn_size); | |
4741 | } | |
4742 | } | |
4743 | ||
c980d109 ACM |
4744 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4745 | struct perf_sample_data *data, | |
4746 | struct perf_event *event) | |
6844c09d ACM |
4747 | { |
4748 | u64 sample_type = event->attr.sample_type; | |
4749 | ||
4750 | data->type = sample_type; | |
4751 | header->size += event->id_header_size; | |
4752 | ||
4753 | if (sample_type & PERF_SAMPLE_TID) { | |
4754 | /* namespace issues */ | |
4755 | data->tid_entry.pid = perf_event_pid(event, current); | |
4756 | data->tid_entry.tid = perf_event_tid(event, current); | |
4757 | } | |
4758 | ||
4759 | if (sample_type & PERF_SAMPLE_TIME) | |
4760 | data->time = perf_clock(); | |
4761 | ||
ff3d527c | 4762 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4763 | data->id = primary_event_id(event); |
4764 | ||
4765 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4766 | data->stream_id = event->id; | |
4767 | ||
4768 | if (sample_type & PERF_SAMPLE_CPU) { | |
4769 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4770 | data->cpu_entry.reserved = 0; | |
4771 | } | |
4772 | } | |
4773 | ||
76369139 FW |
4774 | void perf_event_header__init_id(struct perf_event_header *header, |
4775 | struct perf_sample_data *data, | |
4776 | struct perf_event *event) | |
c980d109 ACM |
4777 | { |
4778 | if (event->attr.sample_id_all) | |
4779 | __perf_event_header__init_id(header, data, event); | |
4780 | } | |
4781 | ||
4782 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4783 | struct perf_sample_data *data) | |
4784 | { | |
4785 | u64 sample_type = data->type; | |
4786 | ||
4787 | if (sample_type & PERF_SAMPLE_TID) | |
4788 | perf_output_put(handle, data->tid_entry); | |
4789 | ||
4790 | if (sample_type & PERF_SAMPLE_TIME) | |
4791 | perf_output_put(handle, data->time); | |
4792 | ||
4793 | if (sample_type & PERF_SAMPLE_ID) | |
4794 | perf_output_put(handle, data->id); | |
4795 | ||
4796 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4797 | perf_output_put(handle, data->stream_id); | |
4798 | ||
4799 | if (sample_type & PERF_SAMPLE_CPU) | |
4800 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4801 | |
4802 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4803 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4804 | } |
4805 | ||
76369139 FW |
4806 | void perf_event__output_id_sample(struct perf_event *event, |
4807 | struct perf_output_handle *handle, | |
4808 | struct perf_sample_data *sample) | |
c980d109 ACM |
4809 | { |
4810 | if (event->attr.sample_id_all) | |
4811 | __perf_event__output_id_sample(handle, sample); | |
4812 | } | |
4813 | ||
3dab77fb | 4814 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4815 | struct perf_event *event, |
4816 | u64 enabled, u64 running) | |
3dab77fb | 4817 | { |
cdd6c482 | 4818 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4819 | u64 values[4]; |
4820 | int n = 0; | |
4821 | ||
b5e58793 | 4822 | values[n++] = perf_event_count(event); |
3dab77fb | 4823 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4824 | values[n++] = enabled + |
cdd6c482 | 4825 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4826 | } |
4827 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4828 | values[n++] = running + |
cdd6c482 | 4829 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4830 | } |
4831 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4832 | values[n++] = primary_event_id(event); |
3dab77fb | 4833 | |
76369139 | 4834 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4835 | } |
4836 | ||
4837 | /* | |
cdd6c482 | 4838 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4839 | */ |
4840 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4841 | struct perf_event *event, |
4842 | u64 enabled, u64 running) | |
3dab77fb | 4843 | { |
cdd6c482 IM |
4844 | struct perf_event *leader = event->group_leader, *sub; |
4845 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4846 | u64 values[5]; |
4847 | int n = 0; | |
4848 | ||
4849 | values[n++] = 1 + leader->nr_siblings; | |
4850 | ||
4851 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4852 | values[n++] = enabled; |
3dab77fb PZ |
4853 | |
4854 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4855 | values[n++] = running; |
3dab77fb | 4856 | |
cdd6c482 | 4857 | if (leader != event) |
3dab77fb PZ |
4858 | leader->pmu->read(leader); |
4859 | ||
b5e58793 | 4860 | values[n++] = perf_event_count(leader); |
3dab77fb | 4861 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4862 | values[n++] = primary_event_id(leader); |
3dab77fb | 4863 | |
76369139 | 4864 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4865 | |
65abc865 | 4866 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4867 | n = 0; |
4868 | ||
6f5ab001 JO |
4869 | if ((sub != event) && |
4870 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4871 | sub->pmu->read(sub); |
4872 | ||
b5e58793 | 4873 | values[n++] = perf_event_count(sub); |
3dab77fb | 4874 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4875 | values[n++] = primary_event_id(sub); |
3dab77fb | 4876 | |
76369139 | 4877 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4878 | } |
4879 | } | |
4880 | ||
eed01528 SE |
4881 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4882 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4883 | ||
3dab77fb | 4884 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4885 | struct perf_event *event) |
3dab77fb | 4886 | { |
e3f3541c | 4887 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4888 | u64 read_format = event->attr.read_format; |
4889 | ||
4890 | /* | |
4891 | * compute total_time_enabled, total_time_running | |
4892 | * based on snapshot values taken when the event | |
4893 | * was last scheduled in. | |
4894 | * | |
4895 | * we cannot simply called update_context_time() | |
4896 | * because of locking issue as we are called in | |
4897 | * NMI context | |
4898 | */ | |
c4794295 | 4899 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4900 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4901 | |
cdd6c482 | 4902 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4903 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4904 | else |
eed01528 | 4905 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4906 | } |
4907 | ||
5622f295 MM |
4908 | void perf_output_sample(struct perf_output_handle *handle, |
4909 | struct perf_event_header *header, | |
4910 | struct perf_sample_data *data, | |
cdd6c482 | 4911 | struct perf_event *event) |
5622f295 MM |
4912 | { |
4913 | u64 sample_type = data->type; | |
4914 | ||
4915 | perf_output_put(handle, *header); | |
4916 | ||
ff3d527c AH |
4917 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4918 | perf_output_put(handle, data->id); | |
4919 | ||
5622f295 MM |
4920 | if (sample_type & PERF_SAMPLE_IP) |
4921 | perf_output_put(handle, data->ip); | |
4922 | ||
4923 | if (sample_type & PERF_SAMPLE_TID) | |
4924 | perf_output_put(handle, data->tid_entry); | |
4925 | ||
4926 | if (sample_type & PERF_SAMPLE_TIME) | |
4927 | perf_output_put(handle, data->time); | |
4928 | ||
4929 | if (sample_type & PERF_SAMPLE_ADDR) | |
4930 | perf_output_put(handle, data->addr); | |
4931 | ||
4932 | if (sample_type & PERF_SAMPLE_ID) | |
4933 | perf_output_put(handle, data->id); | |
4934 | ||
4935 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4936 | perf_output_put(handle, data->stream_id); | |
4937 | ||
4938 | if (sample_type & PERF_SAMPLE_CPU) | |
4939 | perf_output_put(handle, data->cpu_entry); | |
4940 | ||
4941 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4942 | perf_output_put(handle, data->period); | |
4943 | ||
4944 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4945 | perf_output_read(handle, event); |
5622f295 MM |
4946 | |
4947 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4948 | if (data->callchain) { | |
4949 | int size = 1; | |
4950 | ||
4951 | if (data->callchain) | |
4952 | size += data->callchain->nr; | |
4953 | ||
4954 | size *= sizeof(u64); | |
4955 | ||
76369139 | 4956 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4957 | } else { |
4958 | u64 nr = 0; | |
4959 | perf_output_put(handle, nr); | |
4960 | } | |
4961 | } | |
4962 | ||
4963 | if (sample_type & PERF_SAMPLE_RAW) { | |
4964 | if (data->raw) { | |
4965 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4966 | __output_copy(handle, data->raw->data, |
4967 | data->raw->size); | |
5622f295 MM |
4968 | } else { |
4969 | struct { | |
4970 | u32 size; | |
4971 | u32 data; | |
4972 | } raw = { | |
4973 | .size = sizeof(u32), | |
4974 | .data = 0, | |
4975 | }; | |
4976 | perf_output_put(handle, raw); | |
4977 | } | |
4978 | } | |
a7ac67ea | 4979 | |
bce38cd5 SE |
4980 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4981 | if (data->br_stack) { | |
4982 | size_t size; | |
4983 | ||
4984 | size = data->br_stack->nr | |
4985 | * sizeof(struct perf_branch_entry); | |
4986 | ||
4987 | perf_output_put(handle, data->br_stack->nr); | |
4988 | perf_output_copy(handle, data->br_stack->entries, size); | |
4989 | } else { | |
4990 | /* | |
4991 | * we always store at least the value of nr | |
4992 | */ | |
4993 | u64 nr = 0; | |
4994 | perf_output_put(handle, nr); | |
4995 | } | |
4996 | } | |
4018994f JO |
4997 | |
4998 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4999 | u64 abi = data->regs_user.abi; | |
5000 | ||
5001 | /* | |
5002 | * If there are no regs to dump, notice it through | |
5003 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5004 | */ | |
5005 | perf_output_put(handle, abi); | |
5006 | ||
5007 | if (abi) { | |
5008 | u64 mask = event->attr.sample_regs_user; | |
5009 | perf_output_sample_regs(handle, | |
5010 | data->regs_user.regs, | |
5011 | mask); | |
5012 | } | |
5013 | } | |
c5ebcedb | 5014 | |
a5cdd40c | 5015 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
5016 | perf_output_sample_ustack(handle, |
5017 | data->stack_user_size, | |
5018 | data->regs_user.regs); | |
a5cdd40c | 5019 | } |
c3feedf2 AK |
5020 | |
5021 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
5022 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
5023 | |
5024 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
5025 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 5026 | |
fdfbbd07 AK |
5027 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
5028 | perf_output_put(handle, data->txn); | |
5029 | ||
60e2364e SE |
5030 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
5031 | u64 abi = data->regs_intr.abi; | |
5032 | /* | |
5033 | * If there are no regs to dump, notice it through | |
5034 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5035 | */ | |
5036 | perf_output_put(handle, abi); | |
5037 | ||
5038 | if (abi) { | |
5039 | u64 mask = event->attr.sample_regs_intr; | |
5040 | ||
5041 | perf_output_sample_regs(handle, | |
5042 | data->regs_intr.regs, | |
5043 | mask); | |
5044 | } | |
5045 | } | |
5046 | ||
a5cdd40c PZ |
5047 | if (!event->attr.watermark) { |
5048 | int wakeup_events = event->attr.wakeup_events; | |
5049 | ||
5050 | if (wakeup_events) { | |
5051 | struct ring_buffer *rb = handle->rb; | |
5052 | int events = local_inc_return(&rb->events); | |
5053 | ||
5054 | if (events >= wakeup_events) { | |
5055 | local_sub(wakeup_events, &rb->events); | |
5056 | local_inc(&rb->wakeup); | |
5057 | } | |
5058 | } | |
5059 | } | |
5622f295 MM |
5060 | } |
5061 | ||
5062 | void perf_prepare_sample(struct perf_event_header *header, | |
5063 | struct perf_sample_data *data, | |
cdd6c482 | 5064 | struct perf_event *event, |
5622f295 | 5065 | struct pt_regs *regs) |
7b732a75 | 5066 | { |
cdd6c482 | 5067 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 5068 | |
cdd6c482 | 5069 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 5070 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
5071 | |
5072 | header->misc = 0; | |
5073 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 5074 | |
c980d109 | 5075 | __perf_event_header__init_id(header, data, event); |
6844c09d | 5076 | |
c320c7b7 | 5077 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
5078 | data->ip = perf_instruction_pointer(regs); |
5079 | ||
b23f3325 | 5080 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 5081 | int size = 1; |
394ee076 | 5082 | |
e6dab5ff | 5083 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
5084 | |
5085 | if (data->callchain) | |
5086 | size += data->callchain->nr; | |
5087 | ||
5088 | header->size += size * sizeof(u64); | |
394ee076 PZ |
5089 | } |
5090 | ||
3a43ce68 | 5091 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
5092 | int size = sizeof(u32); |
5093 | ||
5094 | if (data->raw) | |
5095 | size += data->raw->size; | |
5096 | else | |
5097 | size += sizeof(u32); | |
5098 | ||
5099 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 5100 | header->size += size; |
7f453c24 | 5101 | } |
bce38cd5 SE |
5102 | |
5103 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
5104 | int size = sizeof(u64); /* nr */ | |
5105 | if (data->br_stack) { | |
5106 | size += data->br_stack->nr | |
5107 | * sizeof(struct perf_branch_entry); | |
5108 | } | |
5109 | header->size += size; | |
5110 | } | |
4018994f | 5111 | |
2565711f | 5112 | if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER)) |
88a7c26a AL |
5113 | perf_sample_regs_user(&data->regs_user, regs, |
5114 | &data->regs_user_copy); | |
2565711f | 5115 | |
4018994f JO |
5116 | if (sample_type & PERF_SAMPLE_REGS_USER) { |
5117 | /* regs dump ABI info */ | |
5118 | int size = sizeof(u64); | |
5119 | ||
4018994f JO |
5120 | if (data->regs_user.regs) { |
5121 | u64 mask = event->attr.sample_regs_user; | |
5122 | size += hweight64(mask) * sizeof(u64); | |
5123 | } | |
5124 | ||
5125 | header->size += size; | |
5126 | } | |
c5ebcedb JO |
5127 | |
5128 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
5129 | /* | |
5130 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
5131 | * processed as the last one or have additional check added | |
5132 | * in case new sample type is added, because we could eat | |
5133 | * up the rest of the sample size. | |
5134 | */ | |
c5ebcedb JO |
5135 | u16 stack_size = event->attr.sample_stack_user; |
5136 | u16 size = sizeof(u64); | |
5137 | ||
c5ebcedb | 5138 | stack_size = perf_sample_ustack_size(stack_size, header->size, |
2565711f | 5139 | data->regs_user.regs); |
c5ebcedb JO |
5140 | |
5141 | /* | |
5142 | * If there is something to dump, add space for the dump | |
5143 | * itself and for the field that tells the dynamic size, | |
5144 | * which is how many have been actually dumped. | |
5145 | */ | |
5146 | if (stack_size) | |
5147 | size += sizeof(u64) + stack_size; | |
5148 | ||
5149 | data->stack_user_size = stack_size; | |
5150 | header->size += size; | |
5151 | } | |
60e2364e SE |
5152 | |
5153 | if (sample_type & PERF_SAMPLE_REGS_INTR) { | |
5154 | /* regs dump ABI info */ | |
5155 | int size = sizeof(u64); | |
5156 | ||
5157 | perf_sample_regs_intr(&data->regs_intr, regs); | |
5158 | ||
5159 | if (data->regs_intr.regs) { | |
5160 | u64 mask = event->attr.sample_regs_intr; | |
5161 | ||
5162 | size += hweight64(mask) * sizeof(u64); | |
5163 | } | |
5164 | ||
5165 | header->size += size; | |
5166 | } | |
5622f295 | 5167 | } |
7f453c24 | 5168 | |
a8b0ca17 | 5169 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
5170 | struct perf_sample_data *data, |
5171 | struct pt_regs *regs) | |
5172 | { | |
5173 | struct perf_output_handle handle; | |
5174 | struct perf_event_header header; | |
689802b2 | 5175 | |
927c7a9e FW |
5176 | /* protect the callchain buffers */ |
5177 | rcu_read_lock(); | |
5178 | ||
cdd6c482 | 5179 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 5180 | |
a7ac67ea | 5181 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 5182 | goto exit; |
0322cd6e | 5183 | |
cdd6c482 | 5184 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 5185 | |
8a057d84 | 5186 | perf_output_end(&handle); |
927c7a9e FW |
5187 | |
5188 | exit: | |
5189 | rcu_read_unlock(); | |
0322cd6e PZ |
5190 | } |
5191 | ||
38b200d6 | 5192 | /* |
cdd6c482 | 5193 | * read event_id |
38b200d6 PZ |
5194 | */ |
5195 | ||
5196 | struct perf_read_event { | |
5197 | struct perf_event_header header; | |
5198 | ||
5199 | u32 pid; | |
5200 | u32 tid; | |
38b200d6 PZ |
5201 | }; |
5202 | ||
5203 | static void | |
cdd6c482 | 5204 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
5205 | struct task_struct *task) |
5206 | { | |
5207 | struct perf_output_handle handle; | |
c980d109 | 5208 | struct perf_sample_data sample; |
dfc65094 | 5209 | struct perf_read_event read_event = { |
38b200d6 | 5210 | .header = { |
cdd6c482 | 5211 | .type = PERF_RECORD_READ, |
38b200d6 | 5212 | .misc = 0, |
c320c7b7 | 5213 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 5214 | }, |
cdd6c482 IM |
5215 | .pid = perf_event_pid(event, task), |
5216 | .tid = perf_event_tid(event, task), | |
38b200d6 | 5217 | }; |
3dab77fb | 5218 | int ret; |
38b200d6 | 5219 | |
c980d109 | 5220 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 5221 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
5222 | if (ret) |
5223 | return; | |
5224 | ||
dfc65094 | 5225 | perf_output_put(&handle, read_event); |
cdd6c482 | 5226 | perf_output_read(&handle, event); |
c980d109 | 5227 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 5228 | |
38b200d6 PZ |
5229 | perf_output_end(&handle); |
5230 | } | |
5231 | ||
52d857a8 JO |
5232 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
5233 | ||
5234 | static void | |
5235 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
5236 | perf_event_aux_output_cb output, |
5237 | void *data) | |
5238 | { | |
5239 | struct perf_event *event; | |
5240 | ||
5241 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5242 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
5243 | continue; | |
5244 | if (!event_filter_match(event)) | |
5245 | continue; | |
67516844 | 5246 | output(event, data); |
52d857a8 JO |
5247 | } |
5248 | } | |
5249 | ||
5250 | static void | |
67516844 | 5251 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
5252 | struct perf_event_context *task_ctx) |
5253 | { | |
5254 | struct perf_cpu_context *cpuctx; | |
5255 | struct perf_event_context *ctx; | |
5256 | struct pmu *pmu; | |
5257 | int ctxn; | |
5258 | ||
5259 | rcu_read_lock(); | |
5260 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
5261 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
5262 | if (cpuctx->unique_pmu != pmu) | |
5263 | goto next; | |
67516844 | 5264 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
5265 | if (task_ctx) |
5266 | goto next; | |
5267 | ctxn = pmu->task_ctx_nr; | |
5268 | if (ctxn < 0) | |
5269 | goto next; | |
5270 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
5271 | if (ctx) | |
67516844 | 5272 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
5273 | next: |
5274 | put_cpu_ptr(pmu->pmu_cpu_context); | |
5275 | } | |
5276 | ||
5277 | if (task_ctx) { | |
5278 | preempt_disable(); | |
67516844 | 5279 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
5280 | preempt_enable(); |
5281 | } | |
5282 | rcu_read_unlock(); | |
5283 | } | |
5284 | ||
60313ebe | 5285 | /* |
9f498cc5 PZ |
5286 | * task tracking -- fork/exit |
5287 | * | |
13d7a241 | 5288 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
5289 | */ |
5290 | ||
9f498cc5 | 5291 | struct perf_task_event { |
3a80b4a3 | 5292 | struct task_struct *task; |
cdd6c482 | 5293 | struct perf_event_context *task_ctx; |
60313ebe PZ |
5294 | |
5295 | struct { | |
5296 | struct perf_event_header header; | |
5297 | ||
5298 | u32 pid; | |
5299 | u32 ppid; | |
9f498cc5 PZ |
5300 | u32 tid; |
5301 | u32 ptid; | |
393b2ad8 | 5302 | u64 time; |
cdd6c482 | 5303 | } event_id; |
60313ebe PZ |
5304 | }; |
5305 | ||
67516844 JO |
5306 | static int perf_event_task_match(struct perf_event *event) |
5307 | { | |
13d7a241 SE |
5308 | return event->attr.comm || event->attr.mmap || |
5309 | event->attr.mmap2 || event->attr.mmap_data || | |
5310 | event->attr.task; | |
67516844 JO |
5311 | } |
5312 | ||
cdd6c482 | 5313 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 5314 | void *data) |
60313ebe | 5315 | { |
52d857a8 | 5316 | struct perf_task_event *task_event = data; |
60313ebe | 5317 | struct perf_output_handle handle; |
c980d109 | 5318 | struct perf_sample_data sample; |
9f498cc5 | 5319 | struct task_struct *task = task_event->task; |
c980d109 | 5320 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 5321 | |
67516844 JO |
5322 | if (!perf_event_task_match(event)) |
5323 | return; | |
5324 | ||
c980d109 | 5325 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 5326 | |
c980d109 | 5327 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 5328 | task_event->event_id.header.size); |
ef60777c | 5329 | if (ret) |
c980d109 | 5330 | goto out; |
60313ebe | 5331 | |
cdd6c482 IM |
5332 | task_event->event_id.pid = perf_event_pid(event, task); |
5333 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 5334 | |
cdd6c482 IM |
5335 | task_event->event_id.tid = perf_event_tid(event, task); |
5336 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 5337 | |
cdd6c482 | 5338 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 5339 | |
c980d109 ACM |
5340 | perf_event__output_id_sample(event, &handle, &sample); |
5341 | ||
60313ebe | 5342 | perf_output_end(&handle); |
c980d109 ACM |
5343 | out: |
5344 | task_event->event_id.header.size = size; | |
60313ebe PZ |
5345 | } |
5346 | ||
cdd6c482 IM |
5347 | static void perf_event_task(struct task_struct *task, |
5348 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 5349 | int new) |
60313ebe | 5350 | { |
9f498cc5 | 5351 | struct perf_task_event task_event; |
60313ebe | 5352 | |
cdd6c482 IM |
5353 | if (!atomic_read(&nr_comm_events) && |
5354 | !atomic_read(&nr_mmap_events) && | |
5355 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
5356 | return; |
5357 | ||
9f498cc5 | 5358 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
5359 | .task = task, |
5360 | .task_ctx = task_ctx, | |
cdd6c482 | 5361 | .event_id = { |
60313ebe | 5362 | .header = { |
cdd6c482 | 5363 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 5364 | .misc = 0, |
cdd6c482 | 5365 | .size = sizeof(task_event.event_id), |
60313ebe | 5366 | }, |
573402db PZ |
5367 | /* .pid */ |
5368 | /* .ppid */ | |
9f498cc5 PZ |
5369 | /* .tid */ |
5370 | /* .ptid */ | |
6f93d0a7 | 5371 | .time = perf_clock(), |
60313ebe PZ |
5372 | }, |
5373 | }; | |
5374 | ||
67516844 | 5375 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5376 | &task_event, |
5377 | task_ctx); | |
9f498cc5 PZ |
5378 | } |
5379 | ||
cdd6c482 | 5380 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5381 | { |
cdd6c482 | 5382 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5383 | } |
5384 | ||
8d1b2d93 PZ |
5385 | /* |
5386 | * comm tracking | |
5387 | */ | |
5388 | ||
5389 | struct perf_comm_event { | |
22a4f650 IM |
5390 | struct task_struct *task; |
5391 | char *comm; | |
8d1b2d93 PZ |
5392 | int comm_size; |
5393 | ||
5394 | struct { | |
5395 | struct perf_event_header header; | |
5396 | ||
5397 | u32 pid; | |
5398 | u32 tid; | |
cdd6c482 | 5399 | } event_id; |
8d1b2d93 PZ |
5400 | }; |
5401 | ||
67516844 JO |
5402 | static int perf_event_comm_match(struct perf_event *event) |
5403 | { | |
5404 | return event->attr.comm; | |
5405 | } | |
5406 | ||
cdd6c482 | 5407 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5408 | void *data) |
8d1b2d93 | 5409 | { |
52d857a8 | 5410 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5411 | struct perf_output_handle handle; |
c980d109 | 5412 | struct perf_sample_data sample; |
cdd6c482 | 5413 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5414 | int ret; |
5415 | ||
67516844 JO |
5416 | if (!perf_event_comm_match(event)) |
5417 | return; | |
5418 | ||
c980d109 ACM |
5419 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5420 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5421 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5422 | |
5423 | if (ret) | |
c980d109 | 5424 | goto out; |
8d1b2d93 | 5425 | |
cdd6c482 IM |
5426 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5427 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5428 | |
cdd6c482 | 5429 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5430 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5431 | comm_event->comm_size); |
c980d109 ACM |
5432 | |
5433 | perf_event__output_id_sample(event, &handle, &sample); | |
5434 | ||
8d1b2d93 | 5435 | perf_output_end(&handle); |
c980d109 ACM |
5436 | out: |
5437 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5438 | } |
5439 | ||
cdd6c482 | 5440 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5441 | { |
413ee3b4 | 5442 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5443 | unsigned int size; |
8d1b2d93 | 5444 | |
413ee3b4 | 5445 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5446 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5447 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5448 | |
5449 | comm_event->comm = comm; | |
5450 | comm_event->comm_size = size; | |
5451 | ||
cdd6c482 | 5452 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5453 | |
67516844 | 5454 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5455 | comm_event, |
5456 | NULL); | |
8d1b2d93 PZ |
5457 | } |
5458 | ||
82b89778 | 5459 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5460 | { |
9ee318a7 PZ |
5461 | struct perf_comm_event comm_event; |
5462 | ||
cdd6c482 | 5463 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5464 | return; |
a63eaf34 | 5465 | |
9ee318a7 | 5466 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5467 | .task = task, |
573402db PZ |
5468 | /* .comm */ |
5469 | /* .comm_size */ | |
cdd6c482 | 5470 | .event_id = { |
573402db | 5471 | .header = { |
cdd6c482 | 5472 | .type = PERF_RECORD_COMM, |
82b89778 | 5473 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5474 | /* .size */ |
5475 | }, | |
5476 | /* .pid */ | |
5477 | /* .tid */ | |
8d1b2d93 PZ |
5478 | }, |
5479 | }; | |
5480 | ||
cdd6c482 | 5481 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5482 | } |
5483 | ||
0a4a9391 PZ |
5484 | /* |
5485 | * mmap tracking | |
5486 | */ | |
5487 | ||
5488 | struct perf_mmap_event { | |
089dd79d PZ |
5489 | struct vm_area_struct *vma; |
5490 | ||
5491 | const char *file_name; | |
5492 | int file_size; | |
13d7a241 SE |
5493 | int maj, min; |
5494 | u64 ino; | |
5495 | u64 ino_generation; | |
f972eb63 | 5496 | u32 prot, flags; |
0a4a9391 PZ |
5497 | |
5498 | struct { | |
5499 | struct perf_event_header header; | |
5500 | ||
5501 | u32 pid; | |
5502 | u32 tid; | |
5503 | u64 start; | |
5504 | u64 len; | |
5505 | u64 pgoff; | |
cdd6c482 | 5506 | } event_id; |
0a4a9391 PZ |
5507 | }; |
5508 | ||
67516844 JO |
5509 | static int perf_event_mmap_match(struct perf_event *event, |
5510 | void *data) | |
5511 | { | |
5512 | struct perf_mmap_event *mmap_event = data; | |
5513 | struct vm_area_struct *vma = mmap_event->vma; | |
5514 | int executable = vma->vm_flags & VM_EXEC; | |
5515 | ||
5516 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5517 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5518 | } |
5519 | ||
cdd6c482 | 5520 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5521 | void *data) |
0a4a9391 | 5522 | { |
52d857a8 | 5523 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5524 | struct perf_output_handle handle; |
c980d109 | 5525 | struct perf_sample_data sample; |
cdd6c482 | 5526 | int size = mmap_event->event_id.header.size; |
c980d109 | 5527 | int ret; |
0a4a9391 | 5528 | |
67516844 JO |
5529 | if (!perf_event_mmap_match(event, data)) |
5530 | return; | |
5531 | ||
13d7a241 SE |
5532 | if (event->attr.mmap2) { |
5533 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5534 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5535 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5536 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5537 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
f972eb63 PZ |
5538 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
5539 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); | |
13d7a241 SE |
5540 | } |
5541 | ||
c980d109 ACM |
5542 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5543 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5544 | mmap_event->event_id.header.size); |
0a4a9391 | 5545 | if (ret) |
c980d109 | 5546 | goto out; |
0a4a9391 | 5547 | |
cdd6c482 IM |
5548 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5549 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5550 | |
cdd6c482 | 5551 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5552 | |
5553 | if (event->attr.mmap2) { | |
5554 | perf_output_put(&handle, mmap_event->maj); | |
5555 | perf_output_put(&handle, mmap_event->min); | |
5556 | perf_output_put(&handle, mmap_event->ino); | |
5557 | perf_output_put(&handle, mmap_event->ino_generation); | |
f972eb63 PZ |
5558 | perf_output_put(&handle, mmap_event->prot); |
5559 | perf_output_put(&handle, mmap_event->flags); | |
13d7a241 SE |
5560 | } |
5561 | ||
76369139 | 5562 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5563 | mmap_event->file_size); |
c980d109 ACM |
5564 | |
5565 | perf_event__output_id_sample(event, &handle, &sample); | |
5566 | ||
78d613eb | 5567 | perf_output_end(&handle); |
c980d109 ACM |
5568 | out: |
5569 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5570 | } |
5571 | ||
cdd6c482 | 5572 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5573 | { |
089dd79d PZ |
5574 | struct vm_area_struct *vma = mmap_event->vma; |
5575 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5576 | int maj = 0, min = 0; |
5577 | u64 ino = 0, gen = 0; | |
f972eb63 | 5578 | u32 prot = 0, flags = 0; |
0a4a9391 PZ |
5579 | unsigned int size; |
5580 | char tmp[16]; | |
5581 | char *buf = NULL; | |
2c42cfbf | 5582 | char *name; |
413ee3b4 | 5583 | |
0a4a9391 | 5584 | if (file) { |
13d7a241 SE |
5585 | struct inode *inode; |
5586 | dev_t dev; | |
3ea2f2b9 | 5587 | |
2c42cfbf | 5588 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5589 | if (!buf) { |
c7e548b4 ON |
5590 | name = "//enomem"; |
5591 | goto cpy_name; | |
0a4a9391 | 5592 | } |
413ee3b4 | 5593 | /* |
3ea2f2b9 | 5594 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5595 | * need to add enough zero bytes after the string to handle |
5596 | * the 64bit alignment we do later. | |
5597 | */ | |
3ea2f2b9 | 5598 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5599 | if (IS_ERR(name)) { |
c7e548b4 ON |
5600 | name = "//toolong"; |
5601 | goto cpy_name; | |
0a4a9391 | 5602 | } |
13d7a241 SE |
5603 | inode = file_inode(vma->vm_file); |
5604 | dev = inode->i_sb->s_dev; | |
5605 | ino = inode->i_ino; | |
5606 | gen = inode->i_generation; | |
5607 | maj = MAJOR(dev); | |
5608 | min = MINOR(dev); | |
f972eb63 PZ |
5609 | |
5610 | if (vma->vm_flags & VM_READ) | |
5611 | prot |= PROT_READ; | |
5612 | if (vma->vm_flags & VM_WRITE) | |
5613 | prot |= PROT_WRITE; | |
5614 | if (vma->vm_flags & VM_EXEC) | |
5615 | prot |= PROT_EXEC; | |
5616 | ||
5617 | if (vma->vm_flags & VM_MAYSHARE) | |
5618 | flags = MAP_SHARED; | |
5619 | else | |
5620 | flags = MAP_PRIVATE; | |
5621 | ||
5622 | if (vma->vm_flags & VM_DENYWRITE) | |
5623 | flags |= MAP_DENYWRITE; | |
5624 | if (vma->vm_flags & VM_MAYEXEC) | |
5625 | flags |= MAP_EXECUTABLE; | |
5626 | if (vma->vm_flags & VM_LOCKED) | |
5627 | flags |= MAP_LOCKED; | |
5628 | if (vma->vm_flags & VM_HUGETLB) | |
5629 | flags |= MAP_HUGETLB; | |
5630 | ||
c7e548b4 | 5631 | goto got_name; |
0a4a9391 | 5632 | } else { |
fbe26abe JO |
5633 | if (vma->vm_ops && vma->vm_ops->name) { |
5634 | name = (char *) vma->vm_ops->name(vma); | |
5635 | if (name) | |
5636 | goto cpy_name; | |
5637 | } | |
5638 | ||
2c42cfbf | 5639 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5640 | if (name) |
5641 | goto cpy_name; | |
089dd79d | 5642 | |
32c5fb7e | 5643 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5644 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5645 | name = "[heap]"; |
5646 | goto cpy_name; | |
32c5fb7e ON |
5647 | } |
5648 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5649 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5650 | name = "[stack]"; |
5651 | goto cpy_name; | |
089dd79d PZ |
5652 | } |
5653 | ||
c7e548b4 ON |
5654 | name = "//anon"; |
5655 | goto cpy_name; | |
0a4a9391 PZ |
5656 | } |
5657 | ||
c7e548b4 ON |
5658 | cpy_name: |
5659 | strlcpy(tmp, name, sizeof(tmp)); | |
5660 | name = tmp; | |
0a4a9391 | 5661 | got_name: |
2c42cfbf PZ |
5662 | /* |
5663 | * Since our buffer works in 8 byte units we need to align our string | |
5664 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5665 | * zero'd out to avoid leaking random bits to userspace. | |
5666 | */ | |
5667 | size = strlen(name)+1; | |
5668 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5669 | name[size++] = '\0'; | |
0a4a9391 PZ |
5670 | |
5671 | mmap_event->file_name = name; | |
5672 | mmap_event->file_size = size; | |
13d7a241 SE |
5673 | mmap_event->maj = maj; |
5674 | mmap_event->min = min; | |
5675 | mmap_event->ino = ino; | |
5676 | mmap_event->ino_generation = gen; | |
f972eb63 PZ |
5677 | mmap_event->prot = prot; |
5678 | mmap_event->flags = flags; | |
0a4a9391 | 5679 | |
2fe85427 SE |
5680 | if (!(vma->vm_flags & VM_EXEC)) |
5681 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5682 | ||
cdd6c482 | 5683 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5684 | |
67516844 | 5685 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5686 | mmap_event, |
5687 | NULL); | |
665c2142 | 5688 | |
0a4a9391 PZ |
5689 | kfree(buf); |
5690 | } | |
5691 | ||
3af9e859 | 5692 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5693 | { |
9ee318a7 PZ |
5694 | struct perf_mmap_event mmap_event; |
5695 | ||
cdd6c482 | 5696 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5697 | return; |
5698 | ||
5699 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5700 | .vma = vma, |
573402db PZ |
5701 | /* .file_name */ |
5702 | /* .file_size */ | |
cdd6c482 | 5703 | .event_id = { |
573402db | 5704 | .header = { |
cdd6c482 | 5705 | .type = PERF_RECORD_MMAP, |
39447b38 | 5706 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5707 | /* .size */ |
5708 | }, | |
5709 | /* .pid */ | |
5710 | /* .tid */ | |
089dd79d PZ |
5711 | .start = vma->vm_start, |
5712 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5713 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5714 | }, |
13d7a241 SE |
5715 | /* .maj (attr_mmap2 only) */ |
5716 | /* .min (attr_mmap2 only) */ | |
5717 | /* .ino (attr_mmap2 only) */ | |
5718 | /* .ino_generation (attr_mmap2 only) */ | |
f972eb63 PZ |
5719 | /* .prot (attr_mmap2 only) */ |
5720 | /* .flags (attr_mmap2 only) */ | |
0a4a9391 PZ |
5721 | }; |
5722 | ||
cdd6c482 | 5723 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5724 | } |
5725 | ||
a78ac325 PZ |
5726 | /* |
5727 | * IRQ throttle logging | |
5728 | */ | |
5729 | ||
cdd6c482 | 5730 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5731 | { |
5732 | struct perf_output_handle handle; | |
c980d109 | 5733 | struct perf_sample_data sample; |
a78ac325 PZ |
5734 | int ret; |
5735 | ||
5736 | struct { | |
5737 | struct perf_event_header header; | |
5738 | u64 time; | |
cca3f454 | 5739 | u64 id; |
7f453c24 | 5740 | u64 stream_id; |
a78ac325 PZ |
5741 | } throttle_event = { |
5742 | .header = { | |
cdd6c482 | 5743 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5744 | .misc = 0, |
5745 | .size = sizeof(throttle_event), | |
5746 | }, | |
def0a9b2 | 5747 | .time = perf_clock(), |
cdd6c482 IM |
5748 | .id = primary_event_id(event), |
5749 | .stream_id = event->id, | |
a78ac325 PZ |
5750 | }; |
5751 | ||
966ee4d6 | 5752 | if (enable) |
cdd6c482 | 5753 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5754 | |
c980d109 ACM |
5755 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5756 | ||
5757 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5758 | throttle_event.header.size); |
a78ac325 PZ |
5759 | if (ret) |
5760 | return; | |
5761 | ||
5762 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5763 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5764 | perf_output_end(&handle); |
5765 | } | |
5766 | ||
f6c7d5fe | 5767 | /* |
cdd6c482 | 5768 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5769 | */ |
5770 | ||
a8b0ca17 | 5771 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5772 | int throttle, struct perf_sample_data *data, |
5773 | struct pt_regs *regs) | |
f6c7d5fe | 5774 | { |
cdd6c482 IM |
5775 | int events = atomic_read(&event->event_limit); |
5776 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5777 | u64 seq; |
79f14641 PZ |
5778 | int ret = 0; |
5779 | ||
96398826 PZ |
5780 | /* |
5781 | * Non-sampling counters might still use the PMI to fold short | |
5782 | * hardware counters, ignore those. | |
5783 | */ | |
5784 | if (unlikely(!is_sampling_event(event))) | |
5785 | return 0; | |
5786 | ||
e050e3f0 SE |
5787 | seq = __this_cpu_read(perf_throttled_seq); |
5788 | if (seq != hwc->interrupts_seq) { | |
5789 | hwc->interrupts_seq = seq; | |
5790 | hwc->interrupts = 1; | |
5791 | } else { | |
5792 | hwc->interrupts++; | |
5793 | if (unlikely(throttle | |
5794 | && hwc->interrupts >= max_samples_per_tick)) { | |
5795 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5796 | hwc->interrupts = MAX_INTERRUPTS; |
5797 | perf_log_throttle(event, 0); | |
d84153d6 | 5798 | tick_nohz_full_kick(); |
a78ac325 PZ |
5799 | ret = 1; |
5800 | } | |
e050e3f0 | 5801 | } |
60db5e09 | 5802 | |
cdd6c482 | 5803 | if (event->attr.freq) { |
def0a9b2 | 5804 | u64 now = perf_clock(); |
abd50713 | 5805 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5806 | |
abd50713 | 5807 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5808 | |
abd50713 | 5809 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5810 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5811 | } |
5812 | ||
2023b359 PZ |
5813 | /* |
5814 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5815 | * events |
2023b359 PZ |
5816 | */ |
5817 | ||
cdd6c482 IM |
5818 | event->pending_kill = POLL_IN; |
5819 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5820 | ret = 1; |
cdd6c482 | 5821 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5822 | event->pending_disable = 1; |
5823 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5824 | } |
5825 | ||
453f19ee | 5826 | if (event->overflow_handler) |
a8b0ca17 | 5827 | event->overflow_handler(event, data, regs); |
453f19ee | 5828 | else |
a8b0ca17 | 5829 | perf_event_output(event, data, regs); |
453f19ee | 5830 | |
f506b3dc | 5831 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5832 | event->pending_wakeup = 1; |
5833 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5834 | } |
5835 | ||
79f14641 | 5836 | return ret; |
f6c7d5fe PZ |
5837 | } |
5838 | ||
a8b0ca17 | 5839 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5840 | struct perf_sample_data *data, |
5841 | struct pt_regs *regs) | |
850bc73f | 5842 | { |
a8b0ca17 | 5843 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5844 | } |
5845 | ||
15dbf27c | 5846 | /* |
cdd6c482 | 5847 | * Generic software event infrastructure |
15dbf27c PZ |
5848 | */ |
5849 | ||
b28ab83c PZ |
5850 | struct swevent_htable { |
5851 | struct swevent_hlist *swevent_hlist; | |
5852 | struct mutex hlist_mutex; | |
5853 | int hlist_refcount; | |
5854 | ||
5855 | /* Recursion avoidance in each contexts */ | |
5856 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
5857 | |
5858 | /* Keeps track of cpu being initialized/exited */ | |
5859 | bool online; | |
b28ab83c PZ |
5860 | }; |
5861 | ||
5862 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5863 | ||
7b4b6658 | 5864 | /* |
cdd6c482 IM |
5865 | * We directly increment event->count and keep a second value in |
5866 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5867 | * is kept in the range [-sample_period, 0] so that we can use the |
5868 | * sign as trigger. | |
5869 | */ | |
5870 | ||
ab573844 | 5871 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5872 | { |
cdd6c482 | 5873 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5874 | u64 period = hwc->last_period; |
5875 | u64 nr, offset; | |
5876 | s64 old, val; | |
5877 | ||
5878 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5879 | |
5880 | again: | |
e7850595 | 5881 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5882 | if (val < 0) |
5883 | return 0; | |
15dbf27c | 5884 | |
7b4b6658 PZ |
5885 | nr = div64_u64(period + val, period); |
5886 | offset = nr * period; | |
5887 | val -= offset; | |
e7850595 | 5888 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5889 | goto again; |
15dbf27c | 5890 | |
7b4b6658 | 5891 | return nr; |
15dbf27c PZ |
5892 | } |
5893 | ||
0cff784a | 5894 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5895 | struct perf_sample_data *data, |
5622f295 | 5896 | struct pt_regs *regs) |
15dbf27c | 5897 | { |
cdd6c482 | 5898 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5899 | int throttle = 0; |
15dbf27c | 5900 | |
0cff784a PZ |
5901 | if (!overflow) |
5902 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5903 | |
7b4b6658 PZ |
5904 | if (hwc->interrupts == MAX_INTERRUPTS) |
5905 | return; | |
15dbf27c | 5906 | |
7b4b6658 | 5907 | for (; overflow; overflow--) { |
a8b0ca17 | 5908 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5909 | data, regs)) { |
7b4b6658 PZ |
5910 | /* |
5911 | * We inhibit the overflow from happening when | |
5912 | * hwc->interrupts == MAX_INTERRUPTS. | |
5913 | */ | |
5914 | break; | |
5915 | } | |
cf450a73 | 5916 | throttle = 1; |
7b4b6658 | 5917 | } |
15dbf27c PZ |
5918 | } |
5919 | ||
a4eaf7f1 | 5920 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5921 | struct perf_sample_data *data, |
5622f295 | 5922 | struct pt_regs *regs) |
7b4b6658 | 5923 | { |
cdd6c482 | 5924 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5925 | |
e7850595 | 5926 | local64_add(nr, &event->count); |
d6d020e9 | 5927 | |
0cff784a PZ |
5928 | if (!regs) |
5929 | return; | |
5930 | ||
6c7e550f | 5931 | if (!is_sampling_event(event)) |
7b4b6658 | 5932 | return; |
d6d020e9 | 5933 | |
5d81e5cf AV |
5934 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5935 | data->period = nr; | |
5936 | return perf_swevent_overflow(event, 1, data, regs); | |
5937 | } else | |
5938 | data->period = event->hw.last_period; | |
5939 | ||
0cff784a | 5940 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5941 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5942 | |
e7850595 | 5943 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5944 | return; |
df1a132b | 5945 | |
a8b0ca17 | 5946 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5947 | } |
5948 | ||
f5ffe02e FW |
5949 | static int perf_exclude_event(struct perf_event *event, |
5950 | struct pt_regs *regs) | |
5951 | { | |
a4eaf7f1 | 5952 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5953 | return 1; |
a4eaf7f1 | 5954 | |
f5ffe02e FW |
5955 | if (regs) { |
5956 | if (event->attr.exclude_user && user_mode(regs)) | |
5957 | return 1; | |
5958 | ||
5959 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5960 | return 1; | |
5961 | } | |
5962 | ||
5963 | return 0; | |
5964 | } | |
5965 | ||
cdd6c482 | 5966 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5967 | enum perf_type_id type, |
6fb2915d LZ |
5968 | u32 event_id, |
5969 | struct perf_sample_data *data, | |
5970 | struct pt_regs *regs) | |
15dbf27c | 5971 | { |
cdd6c482 | 5972 | if (event->attr.type != type) |
a21ca2ca | 5973 | return 0; |
f5ffe02e | 5974 | |
cdd6c482 | 5975 | if (event->attr.config != event_id) |
15dbf27c PZ |
5976 | return 0; |
5977 | ||
f5ffe02e FW |
5978 | if (perf_exclude_event(event, regs)) |
5979 | return 0; | |
15dbf27c PZ |
5980 | |
5981 | return 1; | |
5982 | } | |
5983 | ||
76e1d904 FW |
5984 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5985 | { | |
5986 | u64 val = event_id | (type << 32); | |
5987 | ||
5988 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5989 | } | |
5990 | ||
49f135ed FW |
5991 | static inline struct hlist_head * |
5992 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5993 | { |
49f135ed FW |
5994 | u64 hash = swevent_hash(type, event_id); |
5995 | ||
5996 | return &hlist->heads[hash]; | |
5997 | } | |
76e1d904 | 5998 | |
49f135ed FW |
5999 | /* For the read side: events when they trigger */ |
6000 | static inline struct hlist_head * | |
b28ab83c | 6001 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
6002 | { |
6003 | struct swevent_hlist *hlist; | |
76e1d904 | 6004 | |
b28ab83c | 6005 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
6006 | if (!hlist) |
6007 | return NULL; | |
6008 | ||
49f135ed FW |
6009 | return __find_swevent_head(hlist, type, event_id); |
6010 | } | |
6011 | ||
6012 | /* For the event head insertion and removal in the hlist */ | |
6013 | static inline struct hlist_head * | |
b28ab83c | 6014 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
6015 | { |
6016 | struct swevent_hlist *hlist; | |
6017 | u32 event_id = event->attr.config; | |
6018 | u64 type = event->attr.type; | |
6019 | ||
6020 | /* | |
6021 | * Event scheduling is always serialized against hlist allocation | |
6022 | * and release. Which makes the protected version suitable here. | |
6023 | * The context lock guarantees that. | |
6024 | */ | |
b28ab83c | 6025 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
6026 | lockdep_is_held(&event->ctx->lock)); |
6027 | if (!hlist) | |
6028 | return NULL; | |
6029 | ||
6030 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
6031 | } |
6032 | ||
6033 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 6034 | u64 nr, |
76e1d904 FW |
6035 | struct perf_sample_data *data, |
6036 | struct pt_regs *regs) | |
15dbf27c | 6037 | { |
4a32fea9 | 6038 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6039 | struct perf_event *event; |
76e1d904 | 6040 | struct hlist_head *head; |
15dbf27c | 6041 | |
76e1d904 | 6042 | rcu_read_lock(); |
b28ab83c | 6043 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
6044 | if (!head) |
6045 | goto end; | |
6046 | ||
b67bfe0d | 6047 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 6048 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 6049 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 6050 | } |
76e1d904 FW |
6051 | end: |
6052 | rcu_read_unlock(); | |
15dbf27c PZ |
6053 | } |
6054 | ||
86038c5e PZI |
6055 | DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]); |
6056 | ||
4ed7c92d | 6057 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 6058 | { |
4a32fea9 | 6059 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
96f6d444 | 6060 | |
b28ab83c | 6061 | return get_recursion_context(swhash->recursion); |
96f6d444 | 6062 | } |
645e8cc0 | 6063 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 6064 | |
fa9f90be | 6065 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 6066 | { |
4a32fea9 | 6067 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
927c7a9e | 6068 | |
b28ab83c | 6069 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 6070 | } |
15dbf27c | 6071 | |
86038c5e | 6072 | void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 6073 | { |
a4234bfc | 6074 | struct perf_sample_data data; |
4ed7c92d | 6075 | |
86038c5e | 6076 | if (WARN_ON_ONCE(!regs)) |
4ed7c92d | 6077 | return; |
a4234bfc | 6078 | |
fd0d000b | 6079 | perf_sample_data_init(&data, addr, 0); |
a8b0ca17 | 6080 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
86038c5e PZI |
6081 | } |
6082 | ||
6083 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) | |
6084 | { | |
6085 | int rctx; | |
6086 | ||
6087 | preempt_disable_notrace(); | |
6088 | rctx = perf_swevent_get_recursion_context(); | |
6089 | if (unlikely(rctx < 0)) | |
6090 | goto fail; | |
6091 | ||
6092 | ___perf_sw_event(event_id, nr, regs, addr); | |
4ed7c92d PZ |
6093 | |
6094 | perf_swevent_put_recursion_context(rctx); | |
86038c5e | 6095 | fail: |
1c024eca | 6096 | preempt_enable_notrace(); |
b8e83514 PZ |
6097 | } |
6098 | ||
cdd6c482 | 6099 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 6100 | { |
15dbf27c PZ |
6101 | } |
6102 | ||
a4eaf7f1 | 6103 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 6104 | { |
4a32fea9 | 6105 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6106 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
6107 | struct hlist_head *head; |
6108 | ||
6c7e550f | 6109 | if (is_sampling_event(event)) { |
7b4b6658 | 6110 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 6111 | perf_swevent_set_period(event); |
7b4b6658 | 6112 | } |
76e1d904 | 6113 | |
a4eaf7f1 PZ |
6114 | hwc->state = !(flags & PERF_EF_START); |
6115 | ||
b28ab83c | 6116 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
6117 | if (!head) { |
6118 | /* | |
6119 | * We can race with cpu hotplug code. Do not | |
6120 | * WARN if the cpu just got unplugged. | |
6121 | */ | |
6122 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 6123 | return -EINVAL; |
39af6b16 | 6124 | } |
76e1d904 FW |
6125 | |
6126 | hlist_add_head_rcu(&event->hlist_entry, head); | |
6127 | ||
15dbf27c PZ |
6128 | return 0; |
6129 | } | |
6130 | ||
a4eaf7f1 | 6131 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 6132 | { |
76e1d904 | 6133 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
6134 | } |
6135 | ||
a4eaf7f1 | 6136 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 6137 | { |
a4eaf7f1 | 6138 | event->hw.state = 0; |
d6d020e9 | 6139 | } |
aa9c4c0f | 6140 | |
a4eaf7f1 | 6141 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 6142 | { |
a4eaf7f1 | 6143 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
6144 | } |
6145 | ||
49f135ed FW |
6146 | /* Deref the hlist from the update side */ |
6147 | static inline struct swevent_hlist * | |
b28ab83c | 6148 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 6149 | { |
b28ab83c PZ |
6150 | return rcu_dereference_protected(swhash->swevent_hlist, |
6151 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
6152 | } |
6153 | ||
b28ab83c | 6154 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 6155 | { |
b28ab83c | 6156 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 6157 | |
49f135ed | 6158 | if (!hlist) |
76e1d904 FW |
6159 | return; |
6160 | ||
70691d4a | 6161 | RCU_INIT_POINTER(swhash->swevent_hlist, NULL); |
fa4bbc4c | 6162 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
6163 | } |
6164 | ||
6165 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
6166 | { | |
b28ab83c | 6167 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 6168 | |
b28ab83c | 6169 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6170 | |
b28ab83c PZ |
6171 | if (!--swhash->hlist_refcount) |
6172 | swevent_hlist_release(swhash); | |
76e1d904 | 6173 | |
b28ab83c | 6174 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6175 | } |
6176 | ||
6177 | static void swevent_hlist_put(struct perf_event *event) | |
6178 | { | |
6179 | int cpu; | |
6180 | ||
76e1d904 FW |
6181 | for_each_possible_cpu(cpu) |
6182 | swevent_hlist_put_cpu(event, cpu); | |
6183 | } | |
6184 | ||
6185 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
6186 | { | |
b28ab83c | 6187 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
6188 | int err = 0; |
6189 | ||
b28ab83c | 6190 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6191 | |
b28ab83c | 6192 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
6193 | struct swevent_hlist *hlist; |
6194 | ||
6195 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
6196 | if (!hlist) { | |
6197 | err = -ENOMEM; | |
6198 | goto exit; | |
6199 | } | |
b28ab83c | 6200 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 6201 | } |
b28ab83c | 6202 | swhash->hlist_refcount++; |
9ed6060d | 6203 | exit: |
b28ab83c | 6204 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6205 | |
6206 | return err; | |
6207 | } | |
6208 | ||
6209 | static int swevent_hlist_get(struct perf_event *event) | |
6210 | { | |
6211 | int err; | |
6212 | int cpu, failed_cpu; | |
6213 | ||
76e1d904 FW |
6214 | get_online_cpus(); |
6215 | for_each_possible_cpu(cpu) { | |
6216 | err = swevent_hlist_get_cpu(event, cpu); | |
6217 | if (err) { | |
6218 | failed_cpu = cpu; | |
6219 | goto fail; | |
6220 | } | |
6221 | } | |
6222 | put_online_cpus(); | |
6223 | ||
6224 | return 0; | |
9ed6060d | 6225 | fail: |
76e1d904 FW |
6226 | for_each_possible_cpu(cpu) { |
6227 | if (cpu == failed_cpu) | |
6228 | break; | |
6229 | swevent_hlist_put_cpu(event, cpu); | |
6230 | } | |
6231 | ||
6232 | put_online_cpus(); | |
6233 | return err; | |
6234 | } | |
6235 | ||
c5905afb | 6236 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 6237 | |
b0a873eb PZ |
6238 | static void sw_perf_event_destroy(struct perf_event *event) |
6239 | { | |
6240 | u64 event_id = event->attr.config; | |
95476b64 | 6241 | |
b0a873eb PZ |
6242 | WARN_ON(event->parent); |
6243 | ||
c5905afb | 6244 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6245 | swevent_hlist_put(event); |
6246 | } | |
6247 | ||
6248 | static int perf_swevent_init(struct perf_event *event) | |
6249 | { | |
8176cced | 6250 | u64 event_id = event->attr.config; |
b0a873eb PZ |
6251 | |
6252 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6253 | return -ENOENT; | |
6254 | ||
2481c5fa SE |
6255 | /* |
6256 | * no branch sampling for software events | |
6257 | */ | |
6258 | if (has_branch_stack(event)) | |
6259 | return -EOPNOTSUPP; | |
6260 | ||
b0a873eb PZ |
6261 | switch (event_id) { |
6262 | case PERF_COUNT_SW_CPU_CLOCK: | |
6263 | case PERF_COUNT_SW_TASK_CLOCK: | |
6264 | return -ENOENT; | |
6265 | ||
6266 | default: | |
6267 | break; | |
6268 | } | |
6269 | ||
ce677831 | 6270 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
6271 | return -ENOENT; |
6272 | ||
6273 | if (!event->parent) { | |
6274 | int err; | |
6275 | ||
6276 | err = swevent_hlist_get(event); | |
6277 | if (err) | |
6278 | return err; | |
6279 | ||
c5905afb | 6280 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6281 | event->destroy = sw_perf_event_destroy; |
6282 | } | |
6283 | ||
6284 | return 0; | |
6285 | } | |
6286 | ||
6287 | static struct pmu perf_swevent = { | |
89a1e187 | 6288 | .task_ctx_nr = perf_sw_context, |
95476b64 | 6289 | |
b0a873eb | 6290 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
6291 | .add = perf_swevent_add, |
6292 | .del = perf_swevent_del, | |
6293 | .start = perf_swevent_start, | |
6294 | .stop = perf_swevent_stop, | |
1c024eca | 6295 | .read = perf_swevent_read, |
1c024eca PZ |
6296 | }; |
6297 | ||
b0a873eb PZ |
6298 | #ifdef CONFIG_EVENT_TRACING |
6299 | ||
1c024eca PZ |
6300 | static int perf_tp_filter_match(struct perf_event *event, |
6301 | struct perf_sample_data *data) | |
6302 | { | |
6303 | void *record = data->raw->data; | |
6304 | ||
6305 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
6306 | return 1; | |
6307 | return 0; | |
6308 | } | |
6309 | ||
6310 | static int perf_tp_event_match(struct perf_event *event, | |
6311 | struct perf_sample_data *data, | |
6312 | struct pt_regs *regs) | |
6313 | { | |
a0f7d0f7 FW |
6314 | if (event->hw.state & PERF_HES_STOPPED) |
6315 | return 0; | |
580d607c PZ |
6316 | /* |
6317 | * All tracepoints are from kernel-space. | |
6318 | */ | |
6319 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
6320 | return 0; |
6321 | ||
6322 | if (!perf_tp_filter_match(event, data)) | |
6323 | return 0; | |
6324 | ||
6325 | return 1; | |
6326 | } | |
6327 | ||
6328 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
6329 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
6330 | struct task_struct *task) | |
95476b64 FW |
6331 | { |
6332 | struct perf_sample_data data; | |
1c024eca | 6333 | struct perf_event *event; |
1c024eca | 6334 | |
95476b64 FW |
6335 | struct perf_raw_record raw = { |
6336 | .size = entry_size, | |
6337 | .data = record, | |
6338 | }; | |
6339 | ||
fd0d000b | 6340 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
6341 | data.raw = &raw; |
6342 | ||
b67bfe0d | 6343 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 6344 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 6345 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 6346 | } |
ecc55f84 | 6347 | |
e6dab5ff AV |
6348 | /* |
6349 | * If we got specified a target task, also iterate its context and | |
6350 | * deliver this event there too. | |
6351 | */ | |
6352 | if (task && task != current) { | |
6353 | struct perf_event_context *ctx; | |
6354 | struct trace_entry *entry = record; | |
6355 | ||
6356 | rcu_read_lock(); | |
6357 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
6358 | if (!ctx) | |
6359 | goto unlock; | |
6360 | ||
6361 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
6362 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6363 | continue; | |
6364 | if (event->attr.config != entry->type) | |
6365 | continue; | |
6366 | if (perf_tp_event_match(event, &data, regs)) | |
6367 | perf_swevent_event(event, count, &data, regs); | |
6368 | } | |
6369 | unlock: | |
6370 | rcu_read_unlock(); | |
6371 | } | |
6372 | ||
ecc55f84 | 6373 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
6374 | } |
6375 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
6376 | ||
cdd6c482 | 6377 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 6378 | { |
1c024eca | 6379 | perf_trace_destroy(event); |
e077df4f PZ |
6380 | } |
6381 | ||
b0a873eb | 6382 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 6383 | { |
76e1d904 FW |
6384 | int err; |
6385 | ||
b0a873eb PZ |
6386 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
6387 | return -ENOENT; | |
6388 | ||
2481c5fa SE |
6389 | /* |
6390 | * no branch sampling for tracepoint events | |
6391 | */ | |
6392 | if (has_branch_stack(event)) | |
6393 | return -EOPNOTSUPP; | |
6394 | ||
1c024eca PZ |
6395 | err = perf_trace_init(event); |
6396 | if (err) | |
b0a873eb | 6397 | return err; |
e077df4f | 6398 | |
cdd6c482 | 6399 | event->destroy = tp_perf_event_destroy; |
e077df4f | 6400 | |
b0a873eb PZ |
6401 | return 0; |
6402 | } | |
6403 | ||
6404 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
6405 | .task_ctx_nr = perf_sw_context, |
6406 | ||
b0a873eb | 6407 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6408 | .add = perf_trace_add, |
6409 | .del = perf_trace_del, | |
6410 | .start = perf_swevent_start, | |
6411 | .stop = perf_swevent_stop, | |
b0a873eb | 6412 | .read = perf_swevent_read, |
b0a873eb PZ |
6413 | }; |
6414 | ||
6415 | static inline void perf_tp_register(void) | |
6416 | { | |
2e80a82a | 6417 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6418 | } |
6fb2915d LZ |
6419 | |
6420 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6421 | { | |
6422 | char *filter_str; | |
6423 | int ret; | |
6424 | ||
6425 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6426 | return -EINVAL; | |
6427 | ||
6428 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6429 | if (IS_ERR(filter_str)) | |
6430 | return PTR_ERR(filter_str); | |
6431 | ||
6432 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6433 | ||
6434 | kfree(filter_str); | |
6435 | return ret; | |
6436 | } | |
6437 | ||
6438 | static void perf_event_free_filter(struct perf_event *event) | |
6439 | { | |
6440 | ftrace_profile_free_filter(event); | |
6441 | } | |
6442 | ||
e077df4f | 6443 | #else |
6fb2915d | 6444 | |
b0a873eb | 6445 | static inline void perf_tp_register(void) |
e077df4f | 6446 | { |
e077df4f | 6447 | } |
6fb2915d LZ |
6448 | |
6449 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6450 | { | |
6451 | return -ENOENT; | |
6452 | } | |
6453 | ||
6454 | static void perf_event_free_filter(struct perf_event *event) | |
6455 | { | |
6456 | } | |
6457 | ||
07b139c8 | 6458 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6459 | |
24f1e32c | 6460 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6461 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6462 | { |
f5ffe02e FW |
6463 | struct perf_sample_data sample; |
6464 | struct pt_regs *regs = data; | |
6465 | ||
fd0d000b | 6466 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6467 | |
a4eaf7f1 | 6468 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6469 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6470 | } |
6471 | #endif | |
6472 | ||
b0a873eb PZ |
6473 | /* |
6474 | * hrtimer based swevent callback | |
6475 | */ | |
f29ac756 | 6476 | |
b0a873eb | 6477 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6478 | { |
b0a873eb PZ |
6479 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6480 | struct perf_sample_data data; | |
6481 | struct pt_regs *regs; | |
6482 | struct perf_event *event; | |
6483 | u64 period; | |
f29ac756 | 6484 | |
b0a873eb | 6485 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6486 | |
6487 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6488 | return HRTIMER_NORESTART; | |
6489 | ||
b0a873eb | 6490 | event->pmu->read(event); |
f344011c | 6491 | |
fd0d000b | 6492 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6493 | regs = get_irq_regs(); |
6494 | ||
6495 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6496 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6497 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6498 | ret = HRTIMER_NORESTART; |
6499 | } | |
24f1e32c | 6500 | |
b0a873eb PZ |
6501 | period = max_t(u64, 10000, event->hw.sample_period); |
6502 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6503 | |
b0a873eb | 6504 | return ret; |
f29ac756 PZ |
6505 | } |
6506 | ||
b0a873eb | 6507 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6508 | { |
b0a873eb | 6509 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6510 | s64 period; |
6511 | ||
6512 | if (!is_sampling_event(event)) | |
6513 | return; | |
f5ffe02e | 6514 | |
5d508e82 FBH |
6515 | period = local64_read(&hwc->period_left); |
6516 | if (period) { | |
6517 | if (period < 0) | |
6518 | period = 10000; | |
fa407f35 | 6519 | |
5d508e82 FBH |
6520 | local64_set(&hwc->period_left, 0); |
6521 | } else { | |
6522 | period = max_t(u64, 10000, hwc->sample_period); | |
6523 | } | |
6524 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6525 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6526 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6527 | } |
b0a873eb PZ |
6528 | |
6529 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6530 | { |
b0a873eb PZ |
6531 | struct hw_perf_event *hwc = &event->hw; |
6532 | ||
6c7e550f | 6533 | if (is_sampling_event(event)) { |
b0a873eb | 6534 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6535 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6536 | |
6537 | hrtimer_cancel(&hwc->hrtimer); | |
6538 | } | |
24f1e32c FW |
6539 | } |
6540 | ||
ba3dd36c PZ |
6541 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6542 | { | |
6543 | struct hw_perf_event *hwc = &event->hw; | |
6544 | ||
6545 | if (!is_sampling_event(event)) | |
6546 | return; | |
6547 | ||
6548 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6549 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6550 | ||
6551 | /* | |
6552 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6553 | * mapping and avoid the whole period adjust feedback stuff. | |
6554 | */ | |
6555 | if (event->attr.freq) { | |
6556 | long freq = event->attr.sample_freq; | |
6557 | ||
6558 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6559 | hwc->sample_period = event->attr.sample_period; | |
6560 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6561 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6562 | event->attr.freq = 0; |
6563 | } | |
6564 | } | |
6565 | ||
b0a873eb PZ |
6566 | /* |
6567 | * Software event: cpu wall time clock | |
6568 | */ | |
6569 | ||
6570 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6571 | { |
b0a873eb PZ |
6572 | s64 prev; |
6573 | u64 now; | |
6574 | ||
a4eaf7f1 | 6575 | now = local_clock(); |
b0a873eb PZ |
6576 | prev = local64_xchg(&event->hw.prev_count, now); |
6577 | local64_add(now - prev, &event->count); | |
24f1e32c | 6578 | } |
24f1e32c | 6579 | |
a4eaf7f1 | 6580 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6581 | { |
a4eaf7f1 | 6582 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6583 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6584 | } |
6585 | ||
a4eaf7f1 | 6586 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6587 | { |
b0a873eb PZ |
6588 | perf_swevent_cancel_hrtimer(event); |
6589 | cpu_clock_event_update(event); | |
6590 | } | |
f29ac756 | 6591 | |
a4eaf7f1 PZ |
6592 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6593 | { | |
6594 | if (flags & PERF_EF_START) | |
6595 | cpu_clock_event_start(event, flags); | |
6596 | ||
6597 | return 0; | |
6598 | } | |
6599 | ||
6600 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6601 | { | |
6602 | cpu_clock_event_stop(event, flags); | |
6603 | } | |
6604 | ||
b0a873eb PZ |
6605 | static void cpu_clock_event_read(struct perf_event *event) |
6606 | { | |
6607 | cpu_clock_event_update(event); | |
6608 | } | |
f344011c | 6609 | |
b0a873eb PZ |
6610 | static int cpu_clock_event_init(struct perf_event *event) |
6611 | { | |
6612 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6613 | return -ENOENT; | |
6614 | ||
6615 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6616 | return -ENOENT; | |
6617 | ||
2481c5fa SE |
6618 | /* |
6619 | * no branch sampling for software events | |
6620 | */ | |
6621 | if (has_branch_stack(event)) | |
6622 | return -EOPNOTSUPP; | |
6623 | ||
ba3dd36c PZ |
6624 | perf_swevent_init_hrtimer(event); |
6625 | ||
b0a873eb | 6626 | return 0; |
f29ac756 PZ |
6627 | } |
6628 | ||
b0a873eb | 6629 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6630 | .task_ctx_nr = perf_sw_context, |
6631 | ||
b0a873eb | 6632 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6633 | .add = cpu_clock_event_add, |
6634 | .del = cpu_clock_event_del, | |
6635 | .start = cpu_clock_event_start, | |
6636 | .stop = cpu_clock_event_stop, | |
b0a873eb PZ |
6637 | .read = cpu_clock_event_read, |
6638 | }; | |
6639 | ||
6640 | /* | |
6641 | * Software event: task time clock | |
6642 | */ | |
6643 | ||
6644 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6645 | { |
b0a873eb PZ |
6646 | u64 prev; |
6647 | s64 delta; | |
5c92d124 | 6648 | |
b0a873eb PZ |
6649 | prev = local64_xchg(&event->hw.prev_count, now); |
6650 | delta = now - prev; | |
6651 | local64_add(delta, &event->count); | |
6652 | } | |
5c92d124 | 6653 | |
a4eaf7f1 | 6654 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6655 | { |
a4eaf7f1 | 6656 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6657 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6658 | } |
6659 | ||
a4eaf7f1 | 6660 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6661 | { |
6662 | perf_swevent_cancel_hrtimer(event); | |
6663 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6664 | } |
6665 | ||
6666 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6667 | { | |
6668 | if (flags & PERF_EF_START) | |
6669 | task_clock_event_start(event, flags); | |
b0a873eb | 6670 | |
a4eaf7f1 PZ |
6671 | return 0; |
6672 | } | |
6673 | ||
6674 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6675 | { | |
6676 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6677 | } |
6678 | ||
6679 | static void task_clock_event_read(struct perf_event *event) | |
6680 | { | |
768a06e2 PZ |
6681 | u64 now = perf_clock(); |
6682 | u64 delta = now - event->ctx->timestamp; | |
6683 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6684 | |
6685 | task_clock_event_update(event, time); | |
6686 | } | |
6687 | ||
6688 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6689 | { |
b0a873eb PZ |
6690 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6691 | return -ENOENT; | |
6692 | ||
6693 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6694 | return -ENOENT; | |
6695 | ||
2481c5fa SE |
6696 | /* |
6697 | * no branch sampling for software events | |
6698 | */ | |
6699 | if (has_branch_stack(event)) | |
6700 | return -EOPNOTSUPP; | |
6701 | ||
ba3dd36c PZ |
6702 | perf_swevent_init_hrtimer(event); |
6703 | ||
b0a873eb | 6704 | return 0; |
6fb2915d LZ |
6705 | } |
6706 | ||
b0a873eb | 6707 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6708 | .task_ctx_nr = perf_sw_context, |
6709 | ||
b0a873eb | 6710 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6711 | .add = task_clock_event_add, |
6712 | .del = task_clock_event_del, | |
6713 | .start = task_clock_event_start, | |
6714 | .stop = task_clock_event_stop, | |
b0a873eb PZ |
6715 | .read = task_clock_event_read, |
6716 | }; | |
6fb2915d | 6717 | |
ad5133b7 | 6718 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6719 | { |
e077df4f | 6720 | } |
6fb2915d | 6721 | |
ad5133b7 | 6722 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6723 | { |
ad5133b7 | 6724 | return 0; |
6fb2915d LZ |
6725 | } |
6726 | ||
ad5133b7 | 6727 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6728 | { |
ad5133b7 | 6729 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6730 | } |
6731 | ||
ad5133b7 PZ |
6732 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6733 | { | |
6734 | perf_pmu_enable(pmu); | |
6735 | return 0; | |
6736 | } | |
e077df4f | 6737 | |
ad5133b7 | 6738 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6739 | { |
ad5133b7 | 6740 | perf_pmu_enable(pmu); |
24f1e32c FW |
6741 | } |
6742 | ||
35edc2a5 PZ |
6743 | static int perf_event_idx_default(struct perf_event *event) |
6744 | { | |
c719f560 | 6745 | return 0; |
35edc2a5 PZ |
6746 | } |
6747 | ||
8dc85d54 PZ |
6748 | /* |
6749 | * Ensures all contexts with the same task_ctx_nr have the same | |
6750 | * pmu_cpu_context too. | |
6751 | */ | |
9e317041 | 6752 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 6753 | { |
8dc85d54 | 6754 | struct pmu *pmu; |
b326e956 | 6755 | |
8dc85d54 PZ |
6756 | if (ctxn < 0) |
6757 | return NULL; | |
24f1e32c | 6758 | |
8dc85d54 PZ |
6759 | list_for_each_entry(pmu, &pmus, entry) { |
6760 | if (pmu->task_ctx_nr == ctxn) | |
6761 | return pmu->pmu_cpu_context; | |
6762 | } | |
24f1e32c | 6763 | |
8dc85d54 | 6764 | return NULL; |
24f1e32c FW |
6765 | } |
6766 | ||
51676957 | 6767 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6768 | { |
51676957 PZ |
6769 | int cpu; |
6770 | ||
6771 | for_each_possible_cpu(cpu) { | |
6772 | struct perf_cpu_context *cpuctx; | |
6773 | ||
6774 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6775 | ||
3f1f3320 PZ |
6776 | if (cpuctx->unique_pmu == old_pmu) |
6777 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6778 | } |
6779 | } | |
6780 | ||
6781 | static void free_pmu_context(struct pmu *pmu) | |
6782 | { | |
6783 | struct pmu *i; | |
f5ffe02e | 6784 | |
8dc85d54 | 6785 | mutex_lock(&pmus_lock); |
0475f9ea | 6786 | /* |
8dc85d54 | 6787 | * Like a real lame refcount. |
0475f9ea | 6788 | */ |
51676957 PZ |
6789 | list_for_each_entry(i, &pmus, entry) { |
6790 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6791 | update_pmu_context(i, pmu); | |
8dc85d54 | 6792 | goto out; |
51676957 | 6793 | } |
8dc85d54 | 6794 | } |
d6d020e9 | 6795 | |
51676957 | 6796 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6797 | out: |
6798 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6799 | } |
2e80a82a | 6800 | static struct idr pmu_idr; |
d6d020e9 | 6801 | |
abe43400 PZ |
6802 | static ssize_t |
6803 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6804 | { | |
6805 | struct pmu *pmu = dev_get_drvdata(dev); | |
6806 | ||
6807 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6808 | } | |
90826ca7 | 6809 | static DEVICE_ATTR_RO(type); |
abe43400 | 6810 | |
62b85639 SE |
6811 | static ssize_t |
6812 | perf_event_mux_interval_ms_show(struct device *dev, | |
6813 | struct device_attribute *attr, | |
6814 | char *page) | |
6815 | { | |
6816 | struct pmu *pmu = dev_get_drvdata(dev); | |
6817 | ||
6818 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6819 | } | |
6820 | ||
6821 | static ssize_t | |
6822 | perf_event_mux_interval_ms_store(struct device *dev, | |
6823 | struct device_attribute *attr, | |
6824 | const char *buf, size_t count) | |
6825 | { | |
6826 | struct pmu *pmu = dev_get_drvdata(dev); | |
6827 | int timer, cpu, ret; | |
6828 | ||
6829 | ret = kstrtoint(buf, 0, &timer); | |
6830 | if (ret) | |
6831 | return ret; | |
6832 | ||
6833 | if (timer < 1) | |
6834 | return -EINVAL; | |
6835 | ||
6836 | /* same value, noting to do */ | |
6837 | if (timer == pmu->hrtimer_interval_ms) | |
6838 | return count; | |
6839 | ||
6840 | pmu->hrtimer_interval_ms = timer; | |
6841 | ||
6842 | /* update all cpuctx for this PMU */ | |
6843 | for_each_possible_cpu(cpu) { | |
6844 | struct perf_cpu_context *cpuctx; | |
6845 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6846 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6847 | ||
6848 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6849 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6850 | } | |
6851 | ||
6852 | return count; | |
6853 | } | |
90826ca7 | 6854 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6855 | |
90826ca7 GKH |
6856 | static struct attribute *pmu_dev_attrs[] = { |
6857 | &dev_attr_type.attr, | |
6858 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6859 | NULL, | |
abe43400 | 6860 | }; |
90826ca7 | 6861 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6862 | |
6863 | static int pmu_bus_running; | |
6864 | static struct bus_type pmu_bus = { | |
6865 | .name = "event_source", | |
90826ca7 | 6866 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6867 | }; |
6868 | ||
6869 | static void pmu_dev_release(struct device *dev) | |
6870 | { | |
6871 | kfree(dev); | |
6872 | } | |
6873 | ||
6874 | static int pmu_dev_alloc(struct pmu *pmu) | |
6875 | { | |
6876 | int ret = -ENOMEM; | |
6877 | ||
6878 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6879 | if (!pmu->dev) | |
6880 | goto out; | |
6881 | ||
0c9d42ed | 6882 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6883 | device_initialize(pmu->dev); |
6884 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6885 | if (ret) | |
6886 | goto free_dev; | |
6887 | ||
6888 | dev_set_drvdata(pmu->dev, pmu); | |
6889 | pmu->dev->bus = &pmu_bus; | |
6890 | pmu->dev->release = pmu_dev_release; | |
6891 | ret = device_add(pmu->dev); | |
6892 | if (ret) | |
6893 | goto free_dev; | |
6894 | ||
6895 | out: | |
6896 | return ret; | |
6897 | ||
6898 | free_dev: | |
6899 | put_device(pmu->dev); | |
6900 | goto out; | |
6901 | } | |
6902 | ||
547e9fd7 | 6903 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6904 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6905 | |
03d8e80b | 6906 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6907 | { |
108b02cf | 6908 | int cpu, ret; |
24f1e32c | 6909 | |
b0a873eb | 6910 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6911 | ret = -ENOMEM; |
6912 | pmu->pmu_disable_count = alloc_percpu(int); | |
6913 | if (!pmu->pmu_disable_count) | |
6914 | goto unlock; | |
f29ac756 | 6915 | |
2e80a82a PZ |
6916 | pmu->type = -1; |
6917 | if (!name) | |
6918 | goto skip_type; | |
6919 | pmu->name = name; | |
6920 | ||
6921 | if (type < 0) { | |
0e9c3be2 TH |
6922 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6923 | if (type < 0) { | |
6924 | ret = type; | |
2e80a82a PZ |
6925 | goto free_pdc; |
6926 | } | |
6927 | } | |
6928 | pmu->type = type; | |
6929 | ||
abe43400 PZ |
6930 | if (pmu_bus_running) { |
6931 | ret = pmu_dev_alloc(pmu); | |
6932 | if (ret) | |
6933 | goto free_idr; | |
6934 | } | |
6935 | ||
2e80a82a | 6936 | skip_type: |
8dc85d54 PZ |
6937 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6938 | if (pmu->pmu_cpu_context) | |
6939 | goto got_cpu_context; | |
f29ac756 | 6940 | |
c4814202 | 6941 | ret = -ENOMEM; |
108b02cf PZ |
6942 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6943 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6944 | goto free_dev; |
f344011c | 6945 | |
108b02cf PZ |
6946 | for_each_possible_cpu(cpu) { |
6947 | struct perf_cpu_context *cpuctx; | |
6948 | ||
6949 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6950 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6951 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6952 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
108b02cf | 6953 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6954 | |
6955 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6956 | ||
e9d2b064 | 6957 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6958 | cpuctx->unique_pmu = pmu; |
108b02cf | 6959 | } |
76e1d904 | 6960 | |
8dc85d54 | 6961 | got_cpu_context: |
ad5133b7 PZ |
6962 | if (!pmu->start_txn) { |
6963 | if (pmu->pmu_enable) { | |
6964 | /* | |
6965 | * If we have pmu_enable/pmu_disable calls, install | |
6966 | * transaction stubs that use that to try and batch | |
6967 | * hardware accesses. | |
6968 | */ | |
6969 | pmu->start_txn = perf_pmu_start_txn; | |
6970 | pmu->commit_txn = perf_pmu_commit_txn; | |
6971 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6972 | } else { | |
6973 | pmu->start_txn = perf_pmu_nop_void; | |
6974 | pmu->commit_txn = perf_pmu_nop_int; | |
6975 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6976 | } |
5c92d124 | 6977 | } |
15dbf27c | 6978 | |
ad5133b7 PZ |
6979 | if (!pmu->pmu_enable) { |
6980 | pmu->pmu_enable = perf_pmu_nop_void; | |
6981 | pmu->pmu_disable = perf_pmu_nop_void; | |
6982 | } | |
6983 | ||
35edc2a5 PZ |
6984 | if (!pmu->event_idx) |
6985 | pmu->event_idx = perf_event_idx_default; | |
6986 | ||
b0a873eb | 6987 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6988 | ret = 0; |
6989 | unlock: | |
b0a873eb PZ |
6990 | mutex_unlock(&pmus_lock); |
6991 | ||
33696fc0 | 6992 | return ret; |
108b02cf | 6993 | |
abe43400 PZ |
6994 | free_dev: |
6995 | device_del(pmu->dev); | |
6996 | put_device(pmu->dev); | |
6997 | ||
2e80a82a PZ |
6998 | free_idr: |
6999 | if (pmu->type >= PERF_TYPE_MAX) | |
7000 | idr_remove(&pmu_idr, pmu->type); | |
7001 | ||
108b02cf PZ |
7002 | free_pdc: |
7003 | free_percpu(pmu->pmu_disable_count); | |
7004 | goto unlock; | |
f29ac756 | 7005 | } |
c464c76e | 7006 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 7007 | |
b0a873eb | 7008 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 7009 | { |
b0a873eb PZ |
7010 | mutex_lock(&pmus_lock); |
7011 | list_del_rcu(&pmu->entry); | |
7012 | mutex_unlock(&pmus_lock); | |
5c92d124 | 7013 | |
0475f9ea | 7014 | /* |
cde8e884 PZ |
7015 | * We dereference the pmu list under both SRCU and regular RCU, so |
7016 | * synchronize against both of those. | |
0475f9ea | 7017 | */ |
b0a873eb | 7018 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 7019 | synchronize_rcu(); |
d6d020e9 | 7020 | |
33696fc0 | 7021 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
7022 | if (pmu->type >= PERF_TYPE_MAX) |
7023 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
7024 | device_del(pmu->dev); |
7025 | put_device(pmu->dev); | |
51676957 | 7026 | free_pmu_context(pmu); |
b0a873eb | 7027 | } |
c464c76e | 7028 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 7029 | |
b0a873eb PZ |
7030 | struct pmu *perf_init_event(struct perf_event *event) |
7031 | { | |
7032 | struct pmu *pmu = NULL; | |
7033 | int idx; | |
940c5b29 | 7034 | int ret; |
b0a873eb PZ |
7035 | |
7036 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
7037 | |
7038 | rcu_read_lock(); | |
7039 | pmu = idr_find(&pmu_idr, event->attr.type); | |
7040 | rcu_read_unlock(); | |
940c5b29 | 7041 | if (pmu) { |
c464c76e YZ |
7042 | if (!try_module_get(pmu->module)) { |
7043 | pmu = ERR_PTR(-ENODEV); | |
7044 | goto unlock; | |
7045 | } | |
7e5b2a01 | 7046 | event->pmu = pmu; |
940c5b29 LM |
7047 | ret = pmu->event_init(event); |
7048 | if (ret) | |
7049 | pmu = ERR_PTR(ret); | |
2e80a82a | 7050 | goto unlock; |
940c5b29 | 7051 | } |
2e80a82a | 7052 | |
b0a873eb | 7053 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
c464c76e YZ |
7054 | if (!try_module_get(pmu->module)) { |
7055 | pmu = ERR_PTR(-ENODEV); | |
7056 | goto unlock; | |
7057 | } | |
7e5b2a01 | 7058 | event->pmu = pmu; |
940c5b29 | 7059 | ret = pmu->event_init(event); |
b0a873eb | 7060 | if (!ret) |
e5f4d339 | 7061 | goto unlock; |
76e1d904 | 7062 | |
b0a873eb PZ |
7063 | if (ret != -ENOENT) { |
7064 | pmu = ERR_PTR(ret); | |
e5f4d339 | 7065 | goto unlock; |
f344011c | 7066 | } |
5c92d124 | 7067 | } |
e5f4d339 PZ |
7068 | pmu = ERR_PTR(-ENOENT); |
7069 | unlock: | |
b0a873eb | 7070 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 7071 | |
4aeb0b42 | 7072 | return pmu; |
5c92d124 IM |
7073 | } |
7074 | ||
4beb31f3 FW |
7075 | static void account_event_cpu(struct perf_event *event, int cpu) |
7076 | { | |
7077 | if (event->parent) | |
7078 | return; | |
7079 | ||
7080 | if (has_branch_stack(event)) { | |
7081 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
7082 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
7083 | } | |
7084 | if (is_cgroup_event(event)) | |
7085 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
7086 | } | |
7087 | ||
766d6c07 FW |
7088 | static void account_event(struct perf_event *event) |
7089 | { | |
4beb31f3 FW |
7090 | if (event->parent) |
7091 | return; | |
7092 | ||
766d6c07 FW |
7093 | if (event->attach_state & PERF_ATTACH_TASK) |
7094 | static_key_slow_inc(&perf_sched_events.key); | |
7095 | if (event->attr.mmap || event->attr.mmap_data) | |
7096 | atomic_inc(&nr_mmap_events); | |
7097 | if (event->attr.comm) | |
7098 | atomic_inc(&nr_comm_events); | |
7099 | if (event->attr.task) | |
7100 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
7101 | if (event->attr.freq) { |
7102 | if (atomic_inc_return(&nr_freq_events) == 1) | |
7103 | tick_nohz_full_kick_all(); | |
7104 | } | |
4beb31f3 | 7105 | if (has_branch_stack(event)) |
766d6c07 | 7106 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 7107 | if (is_cgroup_event(event)) |
766d6c07 | 7108 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
7109 | |
7110 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
7111 | } |
7112 | ||
0793a61d | 7113 | /* |
cdd6c482 | 7114 | * Allocate and initialize a event structure |
0793a61d | 7115 | */ |
cdd6c482 | 7116 | static struct perf_event * |
c3f00c70 | 7117 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
7118 | struct task_struct *task, |
7119 | struct perf_event *group_leader, | |
7120 | struct perf_event *parent_event, | |
4dc0da86 AK |
7121 | perf_overflow_handler_t overflow_handler, |
7122 | void *context) | |
0793a61d | 7123 | { |
51b0fe39 | 7124 | struct pmu *pmu; |
cdd6c482 IM |
7125 | struct perf_event *event; |
7126 | struct hw_perf_event *hwc; | |
90983b16 | 7127 | long err = -EINVAL; |
0793a61d | 7128 | |
66832eb4 ON |
7129 | if ((unsigned)cpu >= nr_cpu_ids) { |
7130 | if (!task || cpu != -1) | |
7131 | return ERR_PTR(-EINVAL); | |
7132 | } | |
7133 | ||
c3f00c70 | 7134 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 7135 | if (!event) |
d5d2bc0d | 7136 | return ERR_PTR(-ENOMEM); |
0793a61d | 7137 | |
04289bb9 | 7138 | /* |
cdd6c482 | 7139 | * Single events are their own group leaders, with an |
04289bb9 IM |
7140 | * empty sibling list: |
7141 | */ | |
7142 | if (!group_leader) | |
cdd6c482 | 7143 | group_leader = event; |
04289bb9 | 7144 | |
cdd6c482 IM |
7145 | mutex_init(&event->child_mutex); |
7146 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 7147 | |
cdd6c482 IM |
7148 | INIT_LIST_HEAD(&event->group_entry); |
7149 | INIT_LIST_HEAD(&event->event_entry); | |
7150 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 7151 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 7152 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
7153 | INIT_HLIST_NODE(&event->hlist_entry); |
7154 | ||
10c6db11 | 7155 | |
cdd6c482 | 7156 | init_waitqueue_head(&event->waitq); |
e360adbe | 7157 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 7158 | |
cdd6c482 | 7159 | mutex_init(&event->mmap_mutex); |
7b732a75 | 7160 | |
a6fa941d | 7161 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
7162 | event->cpu = cpu; |
7163 | event->attr = *attr; | |
7164 | event->group_leader = group_leader; | |
7165 | event->pmu = NULL; | |
cdd6c482 | 7166 | event->oncpu = -1; |
a96bbc16 | 7167 | |
cdd6c482 | 7168 | event->parent = parent_event; |
b84fbc9f | 7169 | |
17cf22c3 | 7170 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 7171 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 7172 | |
cdd6c482 | 7173 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 7174 | |
d580ff86 PZ |
7175 | if (task) { |
7176 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
7177 | |
7178 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
7179 | event->hw.tp_target = task; | |
d580ff86 PZ |
7180 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
7181 | /* | |
7182 | * hw_breakpoint is a bit difficult here.. | |
7183 | */ | |
f22c1bb6 | 7184 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
7185 | event->hw.bp_target = task; |
7186 | #endif | |
7187 | } | |
7188 | ||
4dc0da86 | 7189 | if (!overflow_handler && parent_event) { |
b326e956 | 7190 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
7191 | context = parent_event->overflow_handler_context; |
7192 | } | |
66832eb4 | 7193 | |
b326e956 | 7194 | event->overflow_handler = overflow_handler; |
4dc0da86 | 7195 | event->overflow_handler_context = context; |
97eaf530 | 7196 | |
0231bb53 | 7197 | perf_event__state_init(event); |
a86ed508 | 7198 | |
4aeb0b42 | 7199 | pmu = NULL; |
b8e83514 | 7200 | |
cdd6c482 | 7201 | hwc = &event->hw; |
bd2b5b12 | 7202 | hwc->sample_period = attr->sample_period; |
0d48696f | 7203 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 7204 | hwc->sample_period = 1; |
eced1dfc | 7205 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 7206 | |
e7850595 | 7207 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 7208 | |
2023b359 | 7209 | /* |
cdd6c482 | 7210 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 7211 | */ |
3dab77fb | 7212 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 7213 | goto err_ns; |
2023b359 | 7214 | |
b0a873eb | 7215 | pmu = perf_init_event(event); |
4aeb0b42 | 7216 | if (!pmu) |
90983b16 FW |
7217 | goto err_ns; |
7218 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 7219 | err = PTR_ERR(pmu); |
90983b16 | 7220 | goto err_ns; |
621a01ea | 7221 | } |
d5d2bc0d | 7222 | |
cdd6c482 | 7223 | if (!event->parent) { |
927c7a9e FW |
7224 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
7225 | err = get_callchain_buffers(); | |
90983b16 FW |
7226 | if (err) |
7227 | goto err_pmu; | |
d010b332 | 7228 | } |
f344011c | 7229 | } |
9ee318a7 | 7230 | |
cdd6c482 | 7231 | return event; |
90983b16 FW |
7232 | |
7233 | err_pmu: | |
7234 | if (event->destroy) | |
7235 | event->destroy(event); | |
c464c76e | 7236 | module_put(pmu->module); |
90983b16 FW |
7237 | err_ns: |
7238 | if (event->ns) | |
7239 | put_pid_ns(event->ns); | |
7240 | kfree(event); | |
7241 | ||
7242 | return ERR_PTR(err); | |
0793a61d TG |
7243 | } |
7244 | ||
cdd6c482 IM |
7245 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
7246 | struct perf_event_attr *attr) | |
974802ea | 7247 | { |
974802ea | 7248 | u32 size; |
cdf8073d | 7249 | int ret; |
974802ea PZ |
7250 | |
7251 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
7252 | return -EFAULT; | |
7253 | ||
7254 | /* | |
7255 | * zero the full structure, so that a short copy will be nice. | |
7256 | */ | |
7257 | memset(attr, 0, sizeof(*attr)); | |
7258 | ||
7259 | ret = get_user(size, &uattr->size); | |
7260 | if (ret) | |
7261 | return ret; | |
7262 | ||
7263 | if (size > PAGE_SIZE) /* silly large */ | |
7264 | goto err_size; | |
7265 | ||
7266 | if (!size) /* abi compat */ | |
7267 | size = PERF_ATTR_SIZE_VER0; | |
7268 | ||
7269 | if (size < PERF_ATTR_SIZE_VER0) | |
7270 | goto err_size; | |
7271 | ||
7272 | /* | |
7273 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
7274 | * ensure all the unknown bits are 0 - i.e. new |
7275 | * user-space does not rely on any kernel feature | |
7276 | * extensions we dont know about yet. | |
974802ea PZ |
7277 | */ |
7278 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
7279 | unsigned char __user *addr; |
7280 | unsigned char __user *end; | |
7281 | unsigned char val; | |
974802ea | 7282 | |
cdf8073d IS |
7283 | addr = (void __user *)uattr + sizeof(*attr); |
7284 | end = (void __user *)uattr + size; | |
974802ea | 7285 | |
cdf8073d | 7286 | for (; addr < end; addr++) { |
974802ea PZ |
7287 | ret = get_user(val, addr); |
7288 | if (ret) | |
7289 | return ret; | |
7290 | if (val) | |
7291 | goto err_size; | |
7292 | } | |
b3e62e35 | 7293 | size = sizeof(*attr); |
974802ea PZ |
7294 | } |
7295 | ||
7296 | ret = copy_from_user(attr, uattr, size); | |
7297 | if (ret) | |
7298 | return -EFAULT; | |
7299 | ||
cd757645 | 7300 | if (attr->__reserved_1) |
974802ea PZ |
7301 | return -EINVAL; |
7302 | ||
7303 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
7304 | return -EINVAL; | |
7305 | ||
7306 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
7307 | return -EINVAL; | |
7308 | ||
bce38cd5 SE |
7309 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
7310 | u64 mask = attr->branch_sample_type; | |
7311 | ||
7312 | /* only using defined bits */ | |
7313 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
7314 | return -EINVAL; | |
7315 | ||
7316 | /* at least one branch bit must be set */ | |
7317 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
7318 | return -EINVAL; | |
7319 | ||
bce38cd5 SE |
7320 | /* propagate priv level, when not set for branch */ |
7321 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
7322 | ||
7323 | /* exclude_kernel checked on syscall entry */ | |
7324 | if (!attr->exclude_kernel) | |
7325 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
7326 | ||
7327 | if (!attr->exclude_user) | |
7328 | mask |= PERF_SAMPLE_BRANCH_USER; | |
7329 | ||
7330 | if (!attr->exclude_hv) | |
7331 | mask |= PERF_SAMPLE_BRANCH_HV; | |
7332 | /* | |
7333 | * adjust user setting (for HW filter setup) | |
7334 | */ | |
7335 | attr->branch_sample_type = mask; | |
7336 | } | |
e712209a SE |
7337 | /* privileged levels capture (kernel, hv): check permissions */ |
7338 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
7339 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
7340 | return -EACCES; | |
bce38cd5 | 7341 | } |
4018994f | 7342 | |
c5ebcedb | 7343 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 7344 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
7345 | if (ret) |
7346 | return ret; | |
7347 | } | |
7348 | ||
7349 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
7350 | if (!arch_perf_have_user_stack_dump()) | |
7351 | return -ENOSYS; | |
7352 | ||
7353 | /* | |
7354 | * We have __u32 type for the size, but so far | |
7355 | * we can only use __u16 as maximum due to the | |
7356 | * __u16 sample size limit. | |
7357 | */ | |
7358 | if (attr->sample_stack_user >= USHRT_MAX) | |
7359 | ret = -EINVAL; | |
7360 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
7361 | ret = -EINVAL; | |
7362 | } | |
4018994f | 7363 | |
60e2364e SE |
7364 | if (attr->sample_type & PERF_SAMPLE_REGS_INTR) |
7365 | ret = perf_reg_validate(attr->sample_regs_intr); | |
974802ea PZ |
7366 | out: |
7367 | return ret; | |
7368 | ||
7369 | err_size: | |
7370 | put_user(sizeof(*attr), &uattr->size); | |
7371 | ret = -E2BIG; | |
7372 | goto out; | |
7373 | } | |
7374 | ||
ac9721f3 PZ |
7375 | static int |
7376 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 7377 | { |
b69cf536 | 7378 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
7379 | int ret = -EINVAL; |
7380 | ||
ac9721f3 | 7381 | if (!output_event) |
a4be7c27 PZ |
7382 | goto set; |
7383 | ||
ac9721f3 PZ |
7384 | /* don't allow circular references */ |
7385 | if (event == output_event) | |
a4be7c27 PZ |
7386 | goto out; |
7387 | ||
0f139300 PZ |
7388 | /* |
7389 | * Don't allow cross-cpu buffers | |
7390 | */ | |
7391 | if (output_event->cpu != event->cpu) | |
7392 | goto out; | |
7393 | ||
7394 | /* | |
76369139 | 7395 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
7396 | */ |
7397 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
7398 | goto out; | |
7399 | ||
a4be7c27 | 7400 | set: |
cdd6c482 | 7401 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7402 | /* Can't redirect output if we've got an active mmap() */ |
7403 | if (atomic_read(&event->mmap_count)) | |
7404 | goto unlock; | |
a4be7c27 | 7405 | |
ac9721f3 | 7406 | if (output_event) { |
76369139 FW |
7407 | /* get the rb we want to redirect to */ |
7408 | rb = ring_buffer_get(output_event); | |
7409 | if (!rb) | |
ac9721f3 | 7410 | goto unlock; |
a4be7c27 PZ |
7411 | } |
7412 | ||
b69cf536 | 7413 | ring_buffer_attach(event, rb); |
9bb5d40c | 7414 | |
a4be7c27 | 7415 | ret = 0; |
ac9721f3 PZ |
7416 | unlock: |
7417 | mutex_unlock(&event->mmap_mutex); | |
7418 | ||
a4be7c27 | 7419 | out: |
a4be7c27 PZ |
7420 | return ret; |
7421 | } | |
7422 | ||
f63a8daa PZ |
7423 | static void mutex_lock_double(struct mutex *a, struct mutex *b) |
7424 | { | |
7425 | if (b < a) | |
7426 | swap(a, b); | |
7427 | ||
7428 | mutex_lock(a); | |
7429 | mutex_lock_nested(b, SINGLE_DEPTH_NESTING); | |
7430 | } | |
7431 | ||
0793a61d | 7432 | /** |
cdd6c482 | 7433 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7434 | * |
cdd6c482 | 7435 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7436 | * @pid: target pid |
9f66a381 | 7437 | * @cpu: target cpu |
cdd6c482 | 7438 | * @group_fd: group leader event fd |
0793a61d | 7439 | */ |
cdd6c482 IM |
7440 | SYSCALL_DEFINE5(perf_event_open, |
7441 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7442 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7443 | { |
b04243ef PZ |
7444 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7445 | struct perf_event *event, *sibling; | |
cdd6c482 | 7446 | struct perf_event_attr attr; |
f63a8daa | 7447 | struct perf_event_context *ctx, *uninitialized_var(gctx); |
cdd6c482 | 7448 | struct file *event_file = NULL; |
2903ff01 | 7449 | struct fd group = {NULL, 0}; |
38a81da2 | 7450 | struct task_struct *task = NULL; |
89a1e187 | 7451 | struct pmu *pmu; |
ea635c64 | 7452 | int event_fd; |
b04243ef | 7453 | int move_group = 0; |
dc86cabe | 7454 | int err; |
a21b0b35 | 7455 | int f_flags = O_RDWR; |
0793a61d | 7456 | |
2743a5b0 | 7457 | /* for future expandability... */ |
e5d1367f | 7458 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7459 | return -EINVAL; |
7460 | ||
dc86cabe IM |
7461 | err = perf_copy_attr(attr_uptr, &attr); |
7462 | if (err) | |
7463 | return err; | |
eab656ae | 7464 | |
0764771d PZ |
7465 | if (!attr.exclude_kernel) { |
7466 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7467 | return -EACCES; | |
7468 | } | |
7469 | ||
df58ab24 | 7470 | if (attr.freq) { |
cdd6c482 | 7471 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 7472 | return -EINVAL; |
0819b2e3 PZ |
7473 | } else { |
7474 | if (attr.sample_period & (1ULL << 63)) | |
7475 | return -EINVAL; | |
df58ab24 PZ |
7476 | } |
7477 | ||
e5d1367f SE |
7478 | /* |
7479 | * In cgroup mode, the pid argument is used to pass the fd | |
7480 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7481 | * designates the cpu on which to monitor threads from that | |
7482 | * cgroup. | |
7483 | */ | |
7484 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7485 | return -EINVAL; | |
7486 | ||
a21b0b35 YD |
7487 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7488 | f_flags |= O_CLOEXEC; | |
7489 | ||
7490 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7491 | if (event_fd < 0) |
7492 | return event_fd; | |
7493 | ||
ac9721f3 | 7494 | if (group_fd != -1) { |
2903ff01 AV |
7495 | err = perf_fget_light(group_fd, &group); |
7496 | if (err) | |
d14b12d7 | 7497 | goto err_fd; |
2903ff01 | 7498 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7499 | if (flags & PERF_FLAG_FD_OUTPUT) |
7500 | output_event = group_leader; | |
7501 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7502 | group_leader = NULL; | |
7503 | } | |
7504 | ||
e5d1367f | 7505 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7506 | task = find_lively_task_by_vpid(pid); |
7507 | if (IS_ERR(task)) { | |
7508 | err = PTR_ERR(task); | |
7509 | goto err_group_fd; | |
7510 | } | |
7511 | } | |
7512 | ||
1f4ee503 PZ |
7513 | if (task && group_leader && |
7514 | group_leader->attr.inherit != attr.inherit) { | |
7515 | err = -EINVAL; | |
7516 | goto err_task; | |
7517 | } | |
7518 | ||
fbfc623f YZ |
7519 | get_online_cpus(); |
7520 | ||
4dc0da86 AK |
7521 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
7522 | NULL, NULL); | |
d14b12d7 SE |
7523 | if (IS_ERR(event)) { |
7524 | err = PTR_ERR(event); | |
1f4ee503 | 7525 | goto err_cpus; |
d14b12d7 SE |
7526 | } |
7527 | ||
e5d1367f SE |
7528 | if (flags & PERF_FLAG_PID_CGROUP) { |
7529 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
7530 | if (err) { |
7531 | __free_event(event); | |
1f4ee503 | 7532 | goto err_cpus; |
766d6c07 | 7533 | } |
e5d1367f SE |
7534 | } |
7535 | ||
53b25335 VW |
7536 | if (is_sampling_event(event)) { |
7537 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
7538 | err = -ENOTSUPP; | |
7539 | goto err_alloc; | |
7540 | } | |
7541 | } | |
7542 | ||
766d6c07 FW |
7543 | account_event(event); |
7544 | ||
89a1e187 PZ |
7545 | /* |
7546 | * Special case software events and allow them to be part of | |
7547 | * any hardware group. | |
7548 | */ | |
7549 | pmu = event->pmu; | |
b04243ef PZ |
7550 | |
7551 | if (group_leader && | |
7552 | (is_software_event(event) != is_software_event(group_leader))) { | |
7553 | if (is_software_event(event)) { | |
7554 | /* | |
7555 | * If event and group_leader are not both a software | |
7556 | * event, and event is, then group leader is not. | |
7557 | * | |
7558 | * Allow the addition of software events to !software | |
7559 | * groups, this is safe because software events never | |
7560 | * fail to schedule. | |
7561 | */ | |
7562 | pmu = group_leader->pmu; | |
7563 | } else if (is_software_event(group_leader) && | |
7564 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7565 | /* | |
7566 | * In case the group is a pure software group, and we | |
7567 | * try to add a hardware event, move the whole group to | |
7568 | * the hardware context. | |
7569 | */ | |
7570 | move_group = 1; | |
7571 | } | |
7572 | } | |
89a1e187 PZ |
7573 | |
7574 | /* | |
7575 | * Get the target context (task or percpu): | |
7576 | */ | |
e2d37cd2 | 7577 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
7578 | if (IS_ERR(ctx)) { |
7579 | err = PTR_ERR(ctx); | |
c6be5a5c | 7580 | goto err_alloc; |
89a1e187 PZ |
7581 | } |
7582 | ||
fd1edb3a PZ |
7583 | if (task) { |
7584 | put_task_struct(task); | |
7585 | task = NULL; | |
7586 | } | |
7587 | ||
ccff286d | 7588 | /* |
cdd6c482 | 7589 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7590 | */ |
ac9721f3 | 7591 | if (group_leader) { |
dc86cabe | 7592 | err = -EINVAL; |
04289bb9 | 7593 | |
04289bb9 | 7594 | /* |
ccff286d IM |
7595 | * Do not allow a recursive hierarchy (this new sibling |
7596 | * becoming part of another group-sibling): | |
7597 | */ | |
7598 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7599 | goto err_context; |
ccff286d IM |
7600 | /* |
7601 | * Do not allow to attach to a group in a different | |
7602 | * task or CPU context: | |
04289bb9 | 7603 | */ |
b04243ef | 7604 | if (move_group) { |
c3c87e77 PZ |
7605 | /* |
7606 | * Make sure we're both on the same task, or both | |
7607 | * per-cpu events. | |
7608 | */ | |
7609 | if (group_leader->ctx->task != ctx->task) | |
7610 | goto err_context; | |
7611 | ||
7612 | /* | |
7613 | * Make sure we're both events for the same CPU; | |
7614 | * grouping events for different CPUs is broken; since | |
7615 | * you can never concurrently schedule them anyhow. | |
7616 | */ | |
7617 | if (group_leader->cpu != event->cpu) | |
b04243ef PZ |
7618 | goto err_context; |
7619 | } else { | |
7620 | if (group_leader->ctx != ctx) | |
7621 | goto err_context; | |
7622 | } | |
7623 | ||
3b6f9e5c PM |
7624 | /* |
7625 | * Only a group leader can be exclusive or pinned | |
7626 | */ | |
0d48696f | 7627 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7628 | goto err_context; |
ac9721f3 PZ |
7629 | } |
7630 | ||
7631 | if (output_event) { | |
7632 | err = perf_event_set_output(event, output_event); | |
7633 | if (err) | |
c3f00c70 | 7634 | goto err_context; |
ac9721f3 | 7635 | } |
0793a61d | 7636 | |
a21b0b35 YD |
7637 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7638 | f_flags); | |
ea635c64 AV |
7639 | if (IS_ERR(event_file)) { |
7640 | err = PTR_ERR(event_file); | |
c3f00c70 | 7641 | goto err_context; |
ea635c64 | 7642 | } |
9b51f66d | 7643 | |
b04243ef | 7644 | if (move_group) { |
f63a8daa PZ |
7645 | gctx = group_leader->ctx; |
7646 | ||
7647 | /* | |
7648 | * See perf_event_ctx_lock() for comments on the details | |
7649 | * of swizzling perf_event::ctx. | |
7650 | */ | |
7651 | mutex_lock_double(&gctx->mutex, &ctx->mutex); | |
b04243ef | 7652 | |
46ce0fe9 | 7653 | perf_remove_from_context(group_leader, false); |
0231bb53 | 7654 | |
b04243ef PZ |
7655 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7656 | group_entry) { | |
46ce0fe9 | 7657 | perf_remove_from_context(sibling, false); |
b04243ef PZ |
7658 | put_ctx(gctx); |
7659 | } | |
f63a8daa PZ |
7660 | } else { |
7661 | mutex_lock(&ctx->mutex); | |
ea635c64 | 7662 | } |
9b51f66d | 7663 | |
ad3a37de | 7664 | WARN_ON_ONCE(ctx->parent_ctx); |
b04243ef PZ |
7665 | |
7666 | if (move_group) { | |
f63a8daa PZ |
7667 | /* |
7668 | * Wait for everybody to stop referencing the events through | |
7669 | * the old lists, before installing it on new lists. | |
7670 | */ | |
0cda4c02 | 7671 | synchronize_rcu(); |
f63a8daa | 7672 | |
8f95b435 PZI |
7673 | /* |
7674 | * Install the group siblings before the group leader. | |
7675 | * | |
7676 | * Because a group leader will try and install the entire group | |
7677 | * (through the sibling list, which is still in-tact), we can | |
7678 | * end up with siblings installed in the wrong context. | |
7679 | * | |
7680 | * By installing siblings first we NO-OP because they're not | |
7681 | * reachable through the group lists. | |
7682 | */ | |
b04243ef PZ |
7683 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7684 | group_entry) { | |
8f95b435 | 7685 | perf_event__state_init(sibling); |
9fc81d87 | 7686 | perf_install_in_context(ctx, sibling, sibling->cpu); |
b04243ef PZ |
7687 | get_ctx(ctx); |
7688 | } | |
8f95b435 PZI |
7689 | |
7690 | /* | |
7691 | * Removing from the context ends up with disabled | |
7692 | * event. What we want here is event in the initial | |
7693 | * startup state, ready to be add into new context. | |
7694 | */ | |
7695 | perf_event__state_init(group_leader); | |
7696 | perf_install_in_context(ctx, group_leader, group_leader->cpu); | |
7697 | get_ctx(ctx); | |
b04243ef PZ |
7698 | } |
7699 | ||
e2d37cd2 | 7700 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7701 | perf_unpin_context(ctx); |
f63a8daa PZ |
7702 | |
7703 | if (move_group) { | |
7704 | mutex_unlock(&gctx->mutex); | |
7705 | put_ctx(gctx); | |
7706 | } | |
d859e29f | 7707 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7708 | |
fbfc623f YZ |
7709 | put_online_cpus(); |
7710 | ||
cdd6c482 | 7711 | event->owner = current; |
8882135b | 7712 | |
cdd6c482 IM |
7713 | mutex_lock(¤t->perf_event_mutex); |
7714 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7715 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7716 | |
c320c7b7 ACM |
7717 | /* |
7718 | * Precalculate sample_data sizes | |
7719 | */ | |
7720 | perf_event__header_size(event); | |
6844c09d | 7721 | perf_event__id_header_size(event); |
c320c7b7 | 7722 | |
8a49542c PZ |
7723 | /* |
7724 | * Drop the reference on the group_event after placing the | |
7725 | * new event on the sibling_list. This ensures destruction | |
7726 | * of the group leader will find the pointer to itself in | |
7727 | * perf_group_detach(). | |
7728 | */ | |
2903ff01 | 7729 | fdput(group); |
ea635c64 AV |
7730 | fd_install(event_fd, event_file); |
7731 | return event_fd; | |
0793a61d | 7732 | |
c3f00c70 | 7733 | err_context: |
fe4b04fa | 7734 | perf_unpin_context(ctx); |
ea635c64 | 7735 | put_ctx(ctx); |
c6be5a5c | 7736 | err_alloc: |
ea635c64 | 7737 | free_event(event); |
1f4ee503 | 7738 | err_cpus: |
fbfc623f | 7739 | put_online_cpus(); |
1f4ee503 | 7740 | err_task: |
e7d0bc04 PZ |
7741 | if (task) |
7742 | put_task_struct(task); | |
89a1e187 | 7743 | err_group_fd: |
2903ff01 | 7744 | fdput(group); |
ea635c64 AV |
7745 | err_fd: |
7746 | put_unused_fd(event_fd); | |
dc86cabe | 7747 | return err; |
0793a61d TG |
7748 | } |
7749 | ||
fb0459d7 AV |
7750 | /** |
7751 | * perf_event_create_kernel_counter | |
7752 | * | |
7753 | * @attr: attributes of the counter to create | |
7754 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7755 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7756 | */ |
7757 | struct perf_event * | |
7758 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7759 | struct task_struct *task, |
4dc0da86 AK |
7760 | perf_overflow_handler_t overflow_handler, |
7761 | void *context) | |
fb0459d7 | 7762 | { |
fb0459d7 | 7763 | struct perf_event_context *ctx; |
c3f00c70 | 7764 | struct perf_event *event; |
fb0459d7 | 7765 | int err; |
d859e29f | 7766 | |
fb0459d7 AV |
7767 | /* |
7768 | * Get the target context (task or percpu): | |
7769 | */ | |
d859e29f | 7770 | |
4dc0da86 AK |
7771 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7772 | overflow_handler, context); | |
c3f00c70 PZ |
7773 | if (IS_ERR(event)) { |
7774 | err = PTR_ERR(event); | |
7775 | goto err; | |
7776 | } | |
d859e29f | 7777 | |
f8697762 JO |
7778 | /* Mark owner so we could distinguish it from user events. */ |
7779 | event->owner = EVENT_OWNER_KERNEL; | |
7780 | ||
766d6c07 FW |
7781 | account_event(event); |
7782 | ||
38a81da2 | 7783 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7784 | if (IS_ERR(ctx)) { |
7785 | err = PTR_ERR(ctx); | |
c3f00c70 | 7786 | goto err_free; |
d859e29f | 7787 | } |
fb0459d7 | 7788 | |
fb0459d7 AV |
7789 | WARN_ON_ONCE(ctx->parent_ctx); |
7790 | mutex_lock(&ctx->mutex); | |
7791 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7792 | perf_unpin_context(ctx); |
fb0459d7 AV |
7793 | mutex_unlock(&ctx->mutex); |
7794 | ||
fb0459d7 AV |
7795 | return event; |
7796 | ||
c3f00c70 PZ |
7797 | err_free: |
7798 | free_event(event); | |
7799 | err: | |
c6567f64 | 7800 | return ERR_PTR(err); |
9b51f66d | 7801 | } |
fb0459d7 | 7802 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7803 | |
0cda4c02 YZ |
7804 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7805 | { | |
7806 | struct perf_event_context *src_ctx; | |
7807 | struct perf_event_context *dst_ctx; | |
7808 | struct perf_event *event, *tmp; | |
7809 | LIST_HEAD(events); | |
7810 | ||
7811 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7812 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7813 | ||
f63a8daa PZ |
7814 | /* |
7815 | * See perf_event_ctx_lock() for comments on the details | |
7816 | * of swizzling perf_event::ctx. | |
7817 | */ | |
7818 | mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); | |
0cda4c02 YZ |
7819 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, |
7820 | event_entry) { | |
46ce0fe9 | 7821 | perf_remove_from_context(event, false); |
9a545de0 | 7822 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7823 | put_ctx(src_ctx); |
9886167d | 7824 | list_add(&event->migrate_entry, &events); |
0cda4c02 | 7825 | } |
0cda4c02 | 7826 | |
8f95b435 PZI |
7827 | /* |
7828 | * Wait for the events to quiesce before re-instating them. | |
7829 | */ | |
0cda4c02 YZ |
7830 | synchronize_rcu(); |
7831 | ||
8f95b435 PZI |
7832 | /* |
7833 | * Re-instate events in 2 passes. | |
7834 | * | |
7835 | * Skip over group leaders and only install siblings on this first | |
7836 | * pass, siblings will not get enabled without a leader, however a | |
7837 | * leader will enable its siblings, even if those are still on the old | |
7838 | * context. | |
7839 | */ | |
7840 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { | |
7841 | if (event->group_leader == event) | |
7842 | continue; | |
7843 | ||
7844 | list_del(&event->migrate_entry); | |
7845 | if (event->state >= PERF_EVENT_STATE_OFF) | |
7846 | event->state = PERF_EVENT_STATE_INACTIVE; | |
7847 | account_event_cpu(event, dst_cpu); | |
7848 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
7849 | get_ctx(dst_ctx); | |
7850 | } | |
7851 | ||
7852 | /* | |
7853 | * Once all the siblings are setup properly, install the group leaders | |
7854 | * to make it go. | |
7855 | */ | |
9886167d PZ |
7856 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7857 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7858 | if (event->state >= PERF_EVENT_STATE_OFF) |
7859 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7860 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7861 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7862 | get_ctx(dst_ctx); | |
7863 | } | |
7864 | mutex_unlock(&dst_ctx->mutex); | |
f63a8daa | 7865 | mutex_unlock(&src_ctx->mutex); |
0cda4c02 YZ |
7866 | } |
7867 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7868 | ||
cdd6c482 | 7869 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7870 | struct task_struct *child) |
d859e29f | 7871 | { |
cdd6c482 | 7872 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7873 | u64 child_val; |
d859e29f | 7874 | |
cdd6c482 IM |
7875 | if (child_event->attr.inherit_stat) |
7876 | perf_event_read_event(child_event, child); | |
38b200d6 | 7877 | |
b5e58793 | 7878 | child_val = perf_event_count(child_event); |
d859e29f PM |
7879 | |
7880 | /* | |
7881 | * Add back the child's count to the parent's count: | |
7882 | */ | |
a6e6dea6 | 7883 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7884 | atomic64_add(child_event->total_time_enabled, |
7885 | &parent_event->child_total_time_enabled); | |
7886 | atomic64_add(child_event->total_time_running, | |
7887 | &parent_event->child_total_time_running); | |
d859e29f PM |
7888 | |
7889 | /* | |
cdd6c482 | 7890 | * Remove this event from the parent's list |
d859e29f | 7891 | */ |
cdd6c482 IM |
7892 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7893 | mutex_lock(&parent_event->child_mutex); | |
7894 | list_del_init(&child_event->child_list); | |
7895 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f | 7896 | |
dc633982 JO |
7897 | /* |
7898 | * Make sure user/parent get notified, that we just | |
7899 | * lost one event. | |
7900 | */ | |
7901 | perf_event_wakeup(parent_event); | |
7902 | ||
d859e29f | 7903 | /* |
cdd6c482 | 7904 | * Release the parent event, if this was the last |
d859e29f PM |
7905 | * reference to it. |
7906 | */ | |
a6fa941d | 7907 | put_event(parent_event); |
d859e29f PM |
7908 | } |
7909 | ||
9b51f66d | 7910 | static void |
cdd6c482 IM |
7911 | __perf_event_exit_task(struct perf_event *child_event, |
7912 | struct perf_event_context *child_ctx, | |
38b200d6 | 7913 | struct task_struct *child) |
9b51f66d | 7914 | { |
1903d50c PZ |
7915 | /* |
7916 | * Do not destroy the 'original' grouping; because of the context | |
7917 | * switch optimization the original events could've ended up in a | |
7918 | * random child task. | |
7919 | * | |
7920 | * If we were to destroy the original group, all group related | |
7921 | * operations would cease to function properly after this random | |
7922 | * child dies. | |
7923 | * | |
7924 | * Do destroy all inherited groups, we don't care about those | |
7925 | * and being thorough is better. | |
7926 | */ | |
7927 | perf_remove_from_context(child_event, !!child_event->parent); | |
0cc0c027 | 7928 | |
9b51f66d | 7929 | /* |
38b435b1 | 7930 | * It can happen that the parent exits first, and has events |
9b51f66d | 7931 | * that are still around due to the child reference. These |
38b435b1 | 7932 | * events need to be zapped. |
9b51f66d | 7933 | */ |
38b435b1 | 7934 | if (child_event->parent) { |
cdd6c482 IM |
7935 | sync_child_event(child_event, child); |
7936 | free_event(child_event); | |
179033b3 JO |
7937 | } else { |
7938 | child_event->state = PERF_EVENT_STATE_EXIT; | |
7939 | perf_event_wakeup(child_event); | |
4bcf349a | 7940 | } |
9b51f66d IM |
7941 | } |
7942 | ||
8dc85d54 | 7943 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7944 | { |
ebf905fc | 7945 | struct perf_event *child_event, *next; |
211de6eb | 7946 | struct perf_event_context *child_ctx, *clone_ctx = NULL; |
a63eaf34 | 7947 | unsigned long flags; |
9b51f66d | 7948 | |
8dc85d54 | 7949 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7950 | perf_event_task(child, NULL, 0); |
9b51f66d | 7951 | return; |
9f498cc5 | 7952 | } |
9b51f66d | 7953 | |
a63eaf34 | 7954 | local_irq_save(flags); |
ad3a37de PM |
7955 | /* |
7956 | * We can't reschedule here because interrupts are disabled, | |
7957 | * and either child is current or it is a task that can't be | |
7958 | * scheduled, so we are now safe from rescheduling changing | |
7959 | * our context. | |
7960 | */ | |
806839b2 | 7961 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7962 | |
7963 | /* | |
7964 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7965 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7966 | * incremented the context's refcount before we do put_ctx below. |
7967 | */ | |
e625cce1 | 7968 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7969 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7970 | child->perf_event_ctxp[ctxn] = NULL; |
4a1c0f26 | 7971 | |
71a851b4 PZ |
7972 | /* |
7973 | * If this context is a clone; unclone it so it can't get | |
7974 | * swapped to another process while we're removing all | |
cdd6c482 | 7975 | * the events from it. |
71a851b4 | 7976 | */ |
211de6eb | 7977 | clone_ctx = unclone_ctx(child_ctx); |
5e942bb3 | 7978 | update_context_time(child_ctx); |
e625cce1 | 7979 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 | 7980 | |
211de6eb PZ |
7981 | if (clone_ctx) |
7982 | put_ctx(clone_ctx); | |
4a1c0f26 | 7983 | |
9f498cc5 | 7984 | /* |
cdd6c482 IM |
7985 | * Report the task dead after unscheduling the events so that we |
7986 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7987 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7988 | */ |
cdd6c482 | 7989 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7990 | |
66fff224 PZ |
7991 | /* |
7992 | * We can recurse on the same lock type through: | |
7993 | * | |
cdd6c482 IM |
7994 | * __perf_event_exit_task() |
7995 | * sync_child_event() | |
a6fa941d AV |
7996 | * put_event() |
7997 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7998 | * |
7999 | * But since its the parent context it won't be the same instance. | |
8000 | */ | |
a0507c84 | 8001 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 8002 | |
ebf905fc | 8003 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 8004 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 8005 | |
a63eaf34 PM |
8006 | mutex_unlock(&child_ctx->mutex); |
8007 | ||
8008 | put_ctx(child_ctx); | |
9b51f66d IM |
8009 | } |
8010 | ||
8dc85d54 PZ |
8011 | /* |
8012 | * When a child task exits, feed back event values to parent events. | |
8013 | */ | |
8014 | void perf_event_exit_task(struct task_struct *child) | |
8015 | { | |
8882135b | 8016 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
8017 | int ctxn; |
8018 | ||
8882135b PZ |
8019 | mutex_lock(&child->perf_event_mutex); |
8020 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
8021 | owner_entry) { | |
8022 | list_del_init(&event->owner_entry); | |
8023 | ||
8024 | /* | |
8025 | * Ensure the list deletion is visible before we clear | |
8026 | * the owner, closes a race against perf_release() where | |
8027 | * we need to serialize on the owner->perf_event_mutex. | |
8028 | */ | |
8029 | smp_wmb(); | |
8030 | event->owner = NULL; | |
8031 | } | |
8032 | mutex_unlock(&child->perf_event_mutex); | |
8033 | ||
8dc85d54 PZ |
8034 | for_each_task_context_nr(ctxn) |
8035 | perf_event_exit_task_context(child, ctxn); | |
8036 | } | |
8037 | ||
889ff015 FW |
8038 | static void perf_free_event(struct perf_event *event, |
8039 | struct perf_event_context *ctx) | |
8040 | { | |
8041 | struct perf_event *parent = event->parent; | |
8042 | ||
8043 | if (WARN_ON_ONCE(!parent)) | |
8044 | return; | |
8045 | ||
8046 | mutex_lock(&parent->child_mutex); | |
8047 | list_del_init(&event->child_list); | |
8048 | mutex_unlock(&parent->child_mutex); | |
8049 | ||
a6fa941d | 8050 | put_event(parent); |
889ff015 | 8051 | |
652884fe | 8052 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 8053 | perf_group_detach(event); |
889ff015 | 8054 | list_del_event(event, ctx); |
652884fe | 8055 | raw_spin_unlock_irq(&ctx->lock); |
889ff015 FW |
8056 | free_event(event); |
8057 | } | |
8058 | ||
bbbee908 | 8059 | /* |
652884fe | 8060 | * Free an unexposed, unused context as created by inheritance by |
8dc85d54 | 8061 | * perf_event_init_task below, used by fork() in case of fail. |
652884fe PZ |
8062 | * |
8063 | * Not all locks are strictly required, but take them anyway to be nice and | |
8064 | * help out with the lockdep assertions. | |
bbbee908 | 8065 | */ |
cdd6c482 | 8066 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 8067 | { |
8dc85d54 | 8068 | struct perf_event_context *ctx; |
cdd6c482 | 8069 | struct perf_event *event, *tmp; |
8dc85d54 | 8070 | int ctxn; |
bbbee908 | 8071 | |
8dc85d54 PZ |
8072 | for_each_task_context_nr(ctxn) { |
8073 | ctx = task->perf_event_ctxp[ctxn]; | |
8074 | if (!ctx) | |
8075 | continue; | |
bbbee908 | 8076 | |
8dc85d54 | 8077 | mutex_lock(&ctx->mutex); |
bbbee908 | 8078 | again: |
8dc85d54 PZ |
8079 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
8080 | group_entry) | |
8081 | perf_free_event(event, ctx); | |
bbbee908 | 8082 | |
8dc85d54 PZ |
8083 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
8084 | group_entry) | |
8085 | perf_free_event(event, ctx); | |
bbbee908 | 8086 | |
8dc85d54 PZ |
8087 | if (!list_empty(&ctx->pinned_groups) || |
8088 | !list_empty(&ctx->flexible_groups)) | |
8089 | goto again; | |
bbbee908 | 8090 | |
8dc85d54 | 8091 | mutex_unlock(&ctx->mutex); |
bbbee908 | 8092 | |
8dc85d54 PZ |
8093 | put_ctx(ctx); |
8094 | } | |
889ff015 FW |
8095 | } |
8096 | ||
4e231c79 PZ |
8097 | void perf_event_delayed_put(struct task_struct *task) |
8098 | { | |
8099 | int ctxn; | |
8100 | ||
8101 | for_each_task_context_nr(ctxn) | |
8102 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
8103 | } | |
8104 | ||
97dee4f3 PZ |
8105 | /* |
8106 | * inherit a event from parent task to child task: | |
8107 | */ | |
8108 | static struct perf_event * | |
8109 | inherit_event(struct perf_event *parent_event, | |
8110 | struct task_struct *parent, | |
8111 | struct perf_event_context *parent_ctx, | |
8112 | struct task_struct *child, | |
8113 | struct perf_event *group_leader, | |
8114 | struct perf_event_context *child_ctx) | |
8115 | { | |
1929def9 | 8116 | enum perf_event_active_state parent_state = parent_event->state; |
97dee4f3 | 8117 | struct perf_event *child_event; |
cee010ec | 8118 | unsigned long flags; |
97dee4f3 PZ |
8119 | |
8120 | /* | |
8121 | * Instead of creating recursive hierarchies of events, | |
8122 | * we link inherited events back to the original parent, | |
8123 | * which has a filp for sure, which we use as the reference | |
8124 | * count: | |
8125 | */ | |
8126 | if (parent_event->parent) | |
8127 | parent_event = parent_event->parent; | |
8128 | ||
8129 | child_event = perf_event_alloc(&parent_event->attr, | |
8130 | parent_event->cpu, | |
d580ff86 | 8131 | child, |
97dee4f3 | 8132 | group_leader, parent_event, |
4dc0da86 | 8133 | NULL, NULL); |
97dee4f3 PZ |
8134 | if (IS_ERR(child_event)) |
8135 | return child_event; | |
a6fa941d | 8136 | |
fadfe7be JO |
8137 | if (is_orphaned_event(parent_event) || |
8138 | !atomic_long_inc_not_zero(&parent_event->refcount)) { | |
a6fa941d AV |
8139 | free_event(child_event); |
8140 | return NULL; | |
8141 | } | |
8142 | ||
97dee4f3 PZ |
8143 | get_ctx(child_ctx); |
8144 | ||
8145 | /* | |
8146 | * Make the child state follow the state of the parent event, | |
8147 | * not its attr.disabled bit. We hold the parent's mutex, | |
8148 | * so we won't race with perf_event_{en, dis}able_family. | |
8149 | */ | |
1929def9 | 8150 | if (parent_state >= PERF_EVENT_STATE_INACTIVE) |
97dee4f3 PZ |
8151 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
8152 | else | |
8153 | child_event->state = PERF_EVENT_STATE_OFF; | |
8154 | ||
8155 | if (parent_event->attr.freq) { | |
8156 | u64 sample_period = parent_event->hw.sample_period; | |
8157 | struct hw_perf_event *hwc = &child_event->hw; | |
8158 | ||
8159 | hwc->sample_period = sample_period; | |
8160 | hwc->last_period = sample_period; | |
8161 | ||
8162 | local64_set(&hwc->period_left, sample_period); | |
8163 | } | |
8164 | ||
8165 | child_event->ctx = child_ctx; | |
8166 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
8167 | child_event->overflow_handler_context |
8168 | = parent_event->overflow_handler_context; | |
97dee4f3 | 8169 | |
614b6780 TG |
8170 | /* |
8171 | * Precalculate sample_data sizes | |
8172 | */ | |
8173 | perf_event__header_size(child_event); | |
6844c09d | 8174 | perf_event__id_header_size(child_event); |
614b6780 | 8175 | |
97dee4f3 PZ |
8176 | /* |
8177 | * Link it up in the child's context: | |
8178 | */ | |
cee010ec | 8179 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 8180 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 8181 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 8182 | |
97dee4f3 PZ |
8183 | /* |
8184 | * Link this into the parent event's child list | |
8185 | */ | |
8186 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
8187 | mutex_lock(&parent_event->child_mutex); | |
8188 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
8189 | mutex_unlock(&parent_event->child_mutex); | |
8190 | ||
8191 | return child_event; | |
8192 | } | |
8193 | ||
8194 | static int inherit_group(struct perf_event *parent_event, | |
8195 | struct task_struct *parent, | |
8196 | struct perf_event_context *parent_ctx, | |
8197 | struct task_struct *child, | |
8198 | struct perf_event_context *child_ctx) | |
8199 | { | |
8200 | struct perf_event *leader; | |
8201 | struct perf_event *sub; | |
8202 | struct perf_event *child_ctr; | |
8203 | ||
8204 | leader = inherit_event(parent_event, parent, parent_ctx, | |
8205 | child, NULL, child_ctx); | |
8206 | if (IS_ERR(leader)) | |
8207 | return PTR_ERR(leader); | |
8208 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
8209 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
8210 | child, leader, child_ctx); | |
8211 | if (IS_ERR(child_ctr)) | |
8212 | return PTR_ERR(child_ctr); | |
8213 | } | |
8214 | return 0; | |
889ff015 FW |
8215 | } |
8216 | ||
8217 | static int | |
8218 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
8219 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 8220 | struct task_struct *child, int ctxn, |
889ff015 FW |
8221 | int *inherited_all) |
8222 | { | |
8223 | int ret; | |
8dc85d54 | 8224 | struct perf_event_context *child_ctx; |
889ff015 FW |
8225 | |
8226 | if (!event->attr.inherit) { | |
8227 | *inherited_all = 0; | |
8228 | return 0; | |
bbbee908 PZ |
8229 | } |
8230 | ||
fe4b04fa | 8231 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
8232 | if (!child_ctx) { |
8233 | /* | |
8234 | * This is executed from the parent task context, so | |
8235 | * inherit events that have been marked for cloning. | |
8236 | * First allocate and initialize a context for the | |
8237 | * child. | |
8238 | */ | |
bbbee908 | 8239 | |
734df5ab | 8240 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
8241 | if (!child_ctx) |
8242 | return -ENOMEM; | |
bbbee908 | 8243 | |
8dc85d54 | 8244 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
8245 | } |
8246 | ||
8247 | ret = inherit_group(event, parent, parent_ctx, | |
8248 | child, child_ctx); | |
8249 | ||
8250 | if (ret) | |
8251 | *inherited_all = 0; | |
8252 | ||
8253 | return ret; | |
bbbee908 PZ |
8254 | } |
8255 | ||
9b51f66d | 8256 | /* |
cdd6c482 | 8257 | * Initialize the perf_event context in task_struct |
9b51f66d | 8258 | */ |
985c8dcb | 8259 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 8260 | { |
889ff015 | 8261 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
8262 | struct perf_event_context *cloned_ctx; |
8263 | struct perf_event *event; | |
9b51f66d | 8264 | struct task_struct *parent = current; |
564c2b21 | 8265 | int inherited_all = 1; |
dddd3379 | 8266 | unsigned long flags; |
6ab423e0 | 8267 | int ret = 0; |
9b51f66d | 8268 | |
8dc85d54 | 8269 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
8270 | return 0; |
8271 | ||
ad3a37de | 8272 | /* |
25346b93 PM |
8273 | * If the parent's context is a clone, pin it so it won't get |
8274 | * swapped under us. | |
ad3a37de | 8275 | */ |
8dc85d54 | 8276 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
8277 | if (!parent_ctx) |
8278 | return 0; | |
25346b93 | 8279 | |
ad3a37de PM |
8280 | /* |
8281 | * No need to check if parent_ctx != NULL here; since we saw | |
8282 | * it non-NULL earlier, the only reason for it to become NULL | |
8283 | * is if we exit, and since we're currently in the middle of | |
8284 | * a fork we can't be exiting at the same time. | |
8285 | */ | |
ad3a37de | 8286 | |
9b51f66d IM |
8287 | /* |
8288 | * Lock the parent list. No need to lock the child - not PID | |
8289 | * hashed yet and not running, so nobody can access it. | |
8290 | */ | |
d859e29f | 8291 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
8292 | |
8293 | /* | |
8294 | * We dont have to disable NMIs - we are only looking at | |
8295 | * the list, not manipulating it: | |
8296 | */ | |
889ff015 | 8297 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
8298 | ret = inherit_task_group(event, parent, parent_ctx, |
8299 | child, ctxn, &inherited_all); | |
889ff015 FW |
8300 | if (ret) |
8301 | break; | |
8302 | } | |
b93f7978 | 8303 | |
dddd3379 TG |
8304 | /* |
8305 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
8306 | * to allocations, but we need to prevent rotation because | |
8307 | * rotate_ctx() will change the list from interrupt context. | |
8308 | */ | |
8309 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
8310 | parent_ctx->rotate_disable = 1; | |
8311 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
8312 | ||
889ff015 | 8313 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
8314 | ret = inherit_task_group(event, parent, parent_ctx, |
8315 | child, ctxn, &inherited_all); | |
889ff015 | 8316 | if (ret) |
9b51f66d | 8317 | break; |
564c2b21 PM |
8318 | } |
8319 | ||
dddd3379 TG |
8320 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
8321 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 8322 | |
8dc85d54 | 8323 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 8324 | |
05cbaa28 | 8325 | if (child_ctx && inherited_all) { |
564c2b21 PM |
8326 | /* |
8327 | * Mark the child context as a clone of the parent | |
8328 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
8329 | * |
8330 | * Note that if the parent is a clone, the holding of | |
8331 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 8332 | */ |
c5ed5145 | 8333 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
8334 | if (cloned_ctx) { |
8335 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 8336 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
8337 | } else { |
8338 | child_ctx->parent_ctx = parent_ctx; | |
8339 | child_ctx->parent_gen = parent_ctx->generation; | |
8340 | } | |
8341 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
8342 | } |
8343 | ||
c5ed5145 | 8344 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 8345 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 8346 | |
25346b93 | 8347 | perf_unpin_context(parent_ctx); |
fe4b04fa | 8348 | put_ctx(parent_ctx); |
ad3a37de | 8349 | |
6ab423e0 | 8350 | return ret; |
9b51f66d IM |
8351 | } |
8352 | ||
8dc85d54 PZ |
8353 | /* |
8354 | * Initialize the perf_event context in task_struct | |
8355 | */ | |
8356 | int perf_event_init_task(struct task_struct *child) | |
8357 | { | |
8358 | int ctxn, ret; | |
8359 | ||
8550d7cb ON |
8360 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
8361 | mutex_init(&child->perf_event_mutex); | |
8362 | INIT_LIST_HEAD(&child->perf_event_list); | |
8363 | ||
8dc85d54 PZ |
8364 | for_each_task_context_nr(ctxn) { |
8365 | ret = perf_event_init_context(child, ctxn); | |
6c72e350 PZ |
8366 | if (ret) { |
8367 | perf_event_free_task(child); | |
8dc85d54 | 8368 | return ret; |
6c72e350 | 8369 | } |
8dc85d54 PZ |
8370 | } |
8371 | ||
8372 | return 0; | |
8373 | } | |
8374 | ||
220b140b PM |
8375 | static void __init perf_event_init_all_cpus(void) |
8376 | { | |
b28ab83c | 8377 | struct swevent_htable *swhash; |
220b140b | 8378 | int cpu; |
220b140b PM |
8379 | |
8380 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
8381 | swhash = &per_cpu(swevent_htable, cpu); |
8382 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 8383 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
8384 | } |
8385 | } | |
8386 | ||
0db0628d | 8387 | static void perf_event_init_cpu(int cpu) |
0793a61d | 8388 | { |
108b02cf | 8389 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 8390 | |
b28ab83c | 8391 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8392 | swhash->online = true; |
4536e4d1 | 8393 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
8394 | struct swevent_hlist *hlist; |
8395 | ||
b28ab83c PZ |
8396 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
8397 | WARN_ON(!hlist); | |
8398 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 8399 | } |
b28ab83c | 8400 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
8401 | } |
8402 | ||
c277443c | 8403 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 8404 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 8405 | { |
e9d2b064 PZ |
8406 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
8407 | ||
8408 | WARN_ON(!irqs_disabled()); | |
8409 | ||
8410 | list_del_init(&cpuctx->rotation_list); | |
8411 | } | |
8412 | ||
108b02cf | 8413 | static void __perf_event_exit_context(void *__info) |
0793a61d | 8414 | { |
226424ee | 8415 | struct remove_event re = { .detach_group = true }; |
108b02cf | 8416 | struct perf_event_context *ctx = __info; |
0793a61d | 8417 | |
108b02cf | 8418 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 8419 | |
e3703f8c | 8420 | rcu_read_lock(); |
46ce0fe9 PZ |
8421 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
8422 | __perf_remove_from_context(&re); | |
e3703f8c | 8423 | rcu_read_unlock(); |
0793a61d | 8424 | } |
108b02cf PZ |
8425 | |
8426 | static void perf_event_exit_cpu_context(int cpu) | |
8427 | { | |
8428 | struct perf_event_context *ctx; | |
8429 | struct pmu *pmu; | |
8430 | int idx; | |
8431 | ||
8432 | idx = srcu_read_lock(&pmus_srcu); | |
8433 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 8434 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
8435 | |
8436 | mutex_lock(&ctx->mutex); | |
8437 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
8438 | mutex_unlock(&ctx->mutex); | |
8439 | } | |
8440 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
8441 | } |
8442 | ||
cdd6c482 | 8443 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 8444 | { |
b28ab83c | 8445 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 8446 | |
e3703f8c PZ |
8447 | perf_event_exit_cpu_context(cpu); |
8448 | ||
b28ab83c | 8449 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8450 | swhash->online = false; |
b28ab83c PZ |
8451 | swevent_hlist_release(swhash); |
8452 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
8453 | } |
8454 | #else | |
cdd6c482 | 8455 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
8456 | #endif |
8457 | ||
c277443c PZ |
8458 | static int |
8459 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
8460 | { | |
8461 | int cpu; | |
8462 | ||
8463 | for_each_online_cpu(cpu) | |
8464 | perf_event_exit_cpu(cpu); | |
8465 | ||
8466 | return NOTIFY_OK; | |
8467 | } | |
8468 | ||
8469 | /* | |
8470 | * Run the perf reboot notifier at the very last possible moment so that | |
8471 | * the generic watchdog code runs as long as possible. | |
8472 | */ | |
8473 | static struct notifier_block perf_reboot_notifier = { | |
8474 | .notifier_call = perf_reboot, | |
8475 | .priority = INT_MIN, | |
8476 | }; | |
8477 | ||
0db0628d | 8478 | static int |
0793a61d TG |
8479 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
8480 | { | |
8481 | unsigned int cpu = (long)hcpu; | |
8482 | ||
4536e4d1 | 8483 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
8484 | |
8485 | case CPU_UP_PREPARE: | |
5e11637e | 8486 | case CPU_DOWN_FAILED: |
cdd6c482 | 8487 | perf_event_init_cpu(cpu); |
0793a61d TG |
8488 | break; |
8489 | ||
5e11637e | 8490 | case CPU_UP_CANCELED: |
0793a61d | 8491 | case CPU_DOWN_PREPARE: |
cdd6c482 | 8492 | perf_event_exit_cpu(cpu); |
0793a61d | 8493 | break; |
0793a61d TG |
8494 | default: |
8495 | break; | |
8496 | } | |
8497 | ||
8498 | return NOTIFY_OK; | |
8499 | } | |
8500 | ||
cdd6c482 | 8501 | void __init perf_event_init(void) |
0793a61d | 8502 | { |
3c502e7a JW |
8503 | int ret; |
8504 | ||
2e80a82a PZ |
8505 | idr_init(&pmu_idr); |
8506 | ||
220b140b | 8507 | perf_event_init_all_cpus(); |
b0a873eb | 8508 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
8509 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
8510 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
8511 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
8512 | perf_tp_register(); |
8513 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 8514 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
8515 | |
8516 | ret = init_hw_breakpoint(); | |
8517 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
8518 | |
8519 | /* do not patch jump label more than once per second */ | |
8520 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
8521 | |
8522 | /* | |
8523 | * Build time assertion that we keep the data_head at the intended | |
8524 | * location. IOW, validation we got the __reserved[] size right. | |
8525 | */ | |
8526 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
8527 | != 1024); | |
0793a61d | 8528 | } |
abe43400 PZ |
8529 | |
8530 | static int __init perf_event_sysfs_init(void) | |
8531 | { | |
8532 | struct pmu *pmu; | |
8533 | int ret; | |
8534 | ||
8535 | mutex_lock(&pmus_lock); | |
8536 | ||
8537 | ret = bus_register(&pmu_bus); | |
8538 | if (ret) | |
8539 | goto unlock; | |
8540 | ||
8541 | list_for_each_entry(pmu, &pmus, entry) { | |
8542 | if (!pmu->name || pmu->type < 0) | |
8543 | continue; | |
8544 | ||
8545 | ret = pmu_dev_alloc(pmu); | |
8546 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8547 | } | |
8548 | pmu_bus_running = 1; | |
8549 | ret = 0; | |
8550 | ||
8551 | unlock: | |
8552 | mutex_unlock(&pmus_lock); | |
8553 | ||
8554 | return ret; | |
8555 | } | |
8556 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8557 | |
8558 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8559 | static struct cgroup_subsys_state * |
8560 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8561 | { |
8562 | struct perf_cgroup *jc; | |
e5d1367f | 8563 | |
1b15d055 | 8564 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8565 | if (!jc) |
8566 | return ERR_PTR(-ENOMEM); | |
8567 | ||
e5d1367f SE |
8568 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8569 | if (!jc->info) { | |
8570 | kfree(jc); | |
8571 | return ERR_PTR(-ENOMEM); | |
8572 | } | |
8573 | ||
e5d1367f SE |
8574 | return &jc->css; |
8575 | } | |
8576 | ||
eb95419b | 8577 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8578 | { |
eb95419b TH |
8579 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8580 | ||
e5d1367f SE |
8581 | free_percpu(jc->info); |
8582 | kfree(jc); | |
8583 | } | |
8584 | ||
8585 | static int __perf_cgroup_move(void *info) | |
8586 | { | |
8587 | struct task_struct *task = info; | |
8588 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8589 | return 0; | |
8590 | } | |
8591 | ||
eb95419b TH |
8592 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8593 | struct cgroup_taskset *tset) | |
e5d1367f | 8594 | { |
bb9d97b6 TH |
8595 | struct task_struct *task; |
8596 | ||
924f0d9a | 8597 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 8598 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8599 | } |
8600 | ||
eb95419b TH |
8601 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8602 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8603 | struct task_struct *task) |
e5d1367f SE |
8604 | { |
8605 | /* | |
8606 | * cgroup_exit() is called in the copy_process() failure path. | |
8607 | * Ignore this case since the task hasn't ran yet, this avoids | |
8608 | * trying to poke a half freed task state from generic code. | |
8609 | */ | |
8610 | if (!(task->flags & PF_EXITING)) | |
8611 | return; | |
8612 | ||
bb9d97b6 | 8613 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8614 | } |
8615 | ||
073219e9 | 8616 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
8617 | .css_alloc = perf_cgroup_css_alloc, |
8618 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8619 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8620 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8621 | }; |
8622 | #endif /* CONFIG_CGROUP_PERF */ |