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> |
0793a61d | 43 | |
76369139 FW |
44 | #include "internal.h" |
45 | ||
4e193bd4 TB |
46 | #include <asm/irq_regs.h> |
47 | ||
fe4b04fa | 48 | struct remote_function_call { |
e7e7ee2e IM |
49 | struct task_struct *p; |
50 | int (*func)(void *info); | |
51 | void *info; | |
52 | int ret; | |
fe4b04fa PZ |
53 | }; |
54 | ||
55 | static void remote_function(void *data) | |
56 | { | |
57 | struct remote_function_call *tfc = data; | |
58 | struct task_struct *p = tfc->p; | |
59 | ||
60 | if (p) { | |
61 | tfc->ret = -EAGAIN; | |
62 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
63 | return; | |
64 | } | |
65 | ||
66 | tfc->ret = tfc->func(tfc->info); | |
67 | } | |
68 | ||
69 | /** | |
70 | * task_function_call - call a function on the cpu on which a task runs | |
71 | * @p: the task to evaluate | |
72 | * @func: the function to be called | |
73 | * @info: the function call argument | |
74 | * | |
75 | * Calls the function @func when the task is currently running. This might | |
76 | * be on the current CPU, which just calls the function directly | |
77 | * | |
78 | * returns: @func return value, or | |
79 | * -ESRCH - when the process isn't running | |
80 | * -EAGAIN - when the process moved away | |
81 | */ | |
82 | static int | |
83 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
84 | { | |
85 | struct remote_function_call data = { | |
e7e7ee2e IM |
86 | .p = p, |
87 | .func = func, | |
88 | .info = info, | |
89 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
90 | }; |
91 | ||
92 | if (task_curr(p)) | |
93 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
94 | ||
95 | return data.ret; | |
96 | } | |
97 | ||
98 | /** | |
99 | * cpu_function_call - call a function on the cpu | |
100 | * @func: the function to be called | |
101 | * @info: the function call argument | |
102 | * | |
103 | * Calls the function @func on the remote cpu. | |
104 | * | |
105 | * returns: @func return value or -ENXIO when the cpu is offline | |
106 | */ | |
107 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
108 | { | |
109 | struct remote_function_call data = { | |
e7e7ee2e IM |
110 | .p = NULL, |
111 | .func = func, | |
112 | .info = info, | |
113 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
114 | }; |
115 | ||
116 | smp_call_function_single(cpu, remote_function, &data, 1); | |
117 | ||
118 | return data.ret; | |
119 | } | |
120 | ||
e5d1367f SE |
121 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
122 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
123 | PERF_FLAG_PID_CGROUP |\ |
124 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 125 | |
bce38cd5 SE |
126 | /* |
127 | * branch priv levels that need permission checks | |
128 | */ | |
129 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
130 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
131 | PERF_SAMPLE_BRANCH_HV) | |
132 | ||
0b3fcf17 SE |
133 | enum event_type_t { |
134 | EVENT_FLEXIBLE = 0x1, | |
135 | EVENT_PINNED = 0x2, | |
136 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
137 | }; | |
138 | ||
e5d1367f SE |
139 | /* |
140 | * perf_sched_events : >0 events exist | |
141 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
142 | */ | |
c5905afb | 143 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 144 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 145 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 146 | |
cdd6c482 IM |
147 | static atomic_t nr_mmap_events __read_mostly; |
148 | static atomic_t nr_comm_events __read_mostly; | |
149 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 150 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 151 | |
108b02cf PZ |
152 | static LIST_HEAD(pmus); |
153 | static DEFINE_MUTEX(pmus_lock); | |
154 | static struct srcu_struct pmus_srcu; | |
155 | ||
0764771d | 156 | /* |
cdd6c482 | 157 | * perf event paranoia level: |
0fbdea19 IM |
158 | * -1 - not paranoid at all |
159 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 160 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 161 | * 2 - disallow kernel profiling for unpriv |
0764771d | 162 | */ |
cdd6c482 | 163 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 164 | |
20443384 FW |
165 | /* Minimum for 512 kiB + 1 user control page */ |
166 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
167 | |
168 | /* | |
cdd6c482 | 169 | * max perf event sample rate |
df58ab24 | 170 | */ |
14c63f17 DH |
171 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
172 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
173 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
174 | ||
175 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
176 | ||
177 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
178 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
179 | ||
d9494cb4 PZ |
180 | static int perf_sample_allowed_ns __read_mostly = |
181 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
182 | |
183 | void update_perf_cpu_limits(void) | |
184 | { | |
185 | u64 tmp = perf_sample_period_ns; | |
186 | ||
187 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 188 | do_div(tmp, 100); |
d9494cb4 | 189 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 190 | } |
163ec435 | 191 | |
9e630205 SE |
192 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
193 | ||
163ec435 PZ |
194 | int perf_proc_update_handler(struct ctl_table *table, int write, |
195 | void __user *buffer, size_t *lenp, | |
196 | loff_t *ppos) | |
197 | { | |
723478c8 | 198 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
199 | |
200 | if (ret || !write) | |
201 | return ret; | |
202 | ||
203 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
204 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
205 | update_perf_cpu_limits(); | |
206 | ||
207 | return 0; | |
208 | } | |
209 | ||
210 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
211 | ||
212 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
213 | void __user *buffer, size_t *lenp, | |
214 | loff_t *ppos) | |
215 | { | |
216 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
217 | ||
218 | if (ret || !write) | |
219 | return ret; | |
220 | ||
221 | update_perf_cpu_limits(); | |
163ec435 PZ |
222 | |
223 | return 0; | |
224 | } | |
1ccd1549 | 225 | |
14c63f17 DH |
226 | /* |
227 | * perf samples are done in some very critical code paths (NMIs). | |
228 | * If they take too much CPU time, the system can lock up and not | |
229 | * get any real work done. This will drop the sample rate when | |
230 | * we detect that events are taking too long. | |
231 | */ | |
232 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 233 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 234 | |
6a02ad66 | 235 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 236 | { |
6a02ad66 | 237 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 238 | u64 avg_local_sample_len; |
e5302920 | 239 | u64 local_samples_len; |
6a02ad66 PZ |
240 | |
241 | local_samples_len = __get_cpu_var(running_sample_length); | |
242 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
243 | ||
244 | printk_ratelimited(KERN_WARNING | |
245 | "perf interrupt took too long (%lld > %lld), lowering " | |
246 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 247 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
248 | sysctl_perf_event_sample_rate); |
249 | } | |
250 | ||
251 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
252 | ||
253 | void perf_sample_event_took(u64 sample_len_ns) | |
254 | { | |
d9494cb4 | 255 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
256 | u64 avg_local_sample_len; |
257 | u64 local_samples_len; | |
14c63f17 | 258 | |
d9494cb4 | 259 | if (allowed_ns == 0) |
14c63f17 DH |
260 | return; |
261 | ||
262 | /* decay the counter by 1 average sample */ | |
263 | local_samples_len = __get_cpu_var(running_sample_length); | |
264 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
265 | local_samples_len += sample_len_ns; | |
266 | __get_cpu_var(running_sample_length) = local_samples_len; | |
267 | ||
268 | /* | |
269 | * note: this will be biased artifically low until we have | |
270 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
271 | * from having to maintain a count. | |
272 | */ | |
273 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
274 | ||
d9494cb4 | 275 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
276 | return; |
277 | ||
278 | if (max_samples_per_tick <= 1) | |
279 | return; | |
280 | ||
281 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
282 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
283 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
284 | ||
14c63f17 | 285 | update_perf_cpu_limits(); |
6a02ad66 | 286 | |
cd578abb PZ |
287 | if (!irq_work_queue(&perf_duration_work)) { |
288 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
289 | "kernel.perf_event_max_sample_rate to %d\n", | |
290 | avg_local_sample_len, allowed_ns >> 1, | |
291 | sysctl_perf_event_sample_rate); | |
292 | } | |
14c63f17 DH |
293 | } |
294 | ||
cdd6c482 | 295 | static atomic64_t perf_event_id; |
a96bbc16 | 296 | |
0b3fcf17 SE |
297 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
298 | enum event_type_t event_type); | |
299 | ||
300 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
301 | enum event_type_t event_type, |
302 | struct task_struct *task); | |
303 | ||
304 | static void update_context_time(struct perf_event_context *ctx); | |
305 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 306 | |
cdd6c482 | 307 | void __weak perf_event_print_debug(void) { } |
0793a61d | 308 | |
84c79910 | 309 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 310 | { |
84c79910 | 311 | return "pmu"; |
0793a61d TG |
312 | } |
313 | ||
0b3fcf17 SE |
314 | static inline u64 perf_clock(void) |
315 | { | |
316 | return local_clock(); | |
317 | } | |
318 | ||
e5d1367f SE |
319 | static inline struct perf_cpu_context * |
320 | __get_cpu_context(struct perf_event_context *ctx) | |
321 | { | |
322 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
323 | } | |
324 | ||
facc4307 PZ |
325 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
326 | struct perf_event_context *ctx) | |
327 | { | |
328 | raw_spin_lock(&cpuctx->ctx.lock); | |
329 | if (ctx) | |
330 | raw_spin_lock(&ctx->lock); | |
331 | } | |
332 | ||
333 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
334 | struct perf_event_context *ctx) | |
335 | { | |
336 | if (ctx) | |
337 | raw_spin_unlock(&ctx->lock); | |
338 | raw_spin_unlock(&cpuctx->ctx.lock); | |
339 | } | |
340 | ||
e5d1367f SE |
341 | #ifdef CONFIG_CGROUP_PERF |
342 | ||
877c6856 LZ |
343 | /* |
344 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
345 | * This is a per-cpu dynamically allocated data structure. | |
346 | */ | |
347 | struct perf_cgroup_info { | |
348 | u64 time; | |
349 | u64 timestamp; | |
350 | }; | |
351 | ||
352 | struct perf_cgroup { | |
353 | struct cgroup_subsys_state css; | |
86e213e1 | 354 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
355 | }; |
356 | ||
3f7cce3c SE |
357 | /* |
358 | * Must ensure cgroup is pinned (css_get) before calling | |
359 | * this function. In other words, we cannot call this function | |
360 | * if there is no cgroup event for the current CPU context. | |
361 | */ | |
e5d1367f SE |
362 | static inline struct perf_cgroup * |
363 | perf_cgroup_from_task(struct task_struct *task) | |
364 | { | |
073219e9 | 365 | return container_of(task_css(task, perf_event_cgrp_id), |
8af01f56 | 366 | struct perf_cgroup, css); |
e5d1367f SE |
367 | } |
368 | ||
369 | static inline bool | |
370 | perf_cgroup_match(struct perf_event *event) | |
371 | { | |
372 | struct perf_event_context *ctx = event->ctx; | |
373 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
374 | ||
ef824fa1 TH |
375 | /* @event doesn't care about cgroup */ |
376 | if (!event->cgrp) | |
377 | return true; | |
378 | ||
379 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
380 | if (!cpuctx->cgrp) | |
381 | return false; | |
382 | ||
383 | /* | |
384 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
385 | * also enabled for all its descendant cgroups. If @cpuctx's | |
386 | * cgroup is a descendant of @event's (the test covers identity | |
387 | * case), it's a match. | |
388 | */ | |
389 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
390 | event->cgrp->css.cgroup); | |
e5d1367f SE |
391 | } |
392 | ||
e5d1367f SE |
393 | static inline void perf_put_cgroup(struct perf_event *event) |
394 | { | |
395 | css_put(&event->cgrp->css); | |
396 | } | |
397 | ||
398 | static inline void perf_detach_cgroup(struct perf_event *event) | |
399 | { | |
400 | perf_put_cgroup(event); | |
401 | event->cgrp = NULL; | |
402 | } | |
403 | ||
404 | static inline int is_cgroup_event(struct perf_event *event) | |
405 | { | |
406 | return event->cgrp != NULL; | |
407 | } | |
408 | ||
409 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
410 | { | |
411 | struct perf_cgroup_info *t; | |
412 | ||
413 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
414 | return t->time; | |
415 | } | |
416 | ||
417 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
418 | { | |
419 | struct perf_cgroup_info *info; | |
420 | u64 now; | |
421 | ||
422 | now = perf_clock(); | |
423 | ||
424 | info = this_cpu_ptr(cgrp->info); | |
425 | ||
426 | info->time += now - info->timestamp; | |
427 | info->timestamp = now; | |
428 | } | |
429 | ||
430 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
431 | { | |
432 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
433 | if (cgrp_out) | |
434 | __update_cgrp_time(cgrp_out); | |
435 | } | |
436 | ||
437 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
438 | { | |
3f7cce3c SE |
439 | struct perf_cgroup *cgrp; |
440 | ||
e5d1367f | 441 | /* |
3f7cce3c SE |
442 | * ensure we access cgroup data only when needed and |
443 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 444 | */ |
3f7cce3c | 445 | if (!is_cgroup_event(event)) |
e5d1367f SE |
446 | return; |
447 | ||
3f7cce3c SE |
448 | cgrp = perf_cgroup_from_task(current); |
449 | /* | |
450 | * Do not update time when cgroup is not active | |
451 | */ | |
452 | if (cgrp == event->cgrp) | |
453 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
454 | } |
455 | ||
456 | static inline void | |
3f7cce3c SE |
457 | perf_cgroup_set_timestamp(struct task_struct *task, |
458 | struct perf_event_context *ctx) | |
e5d1367f SE |
459 | { |
460 | struct perf_cgroup *cgrp; | |
461 | struct perf_cgroup_info *info; | |
462 | ||
3f7cce3c SE |
463 | /* |
464 | * ctx->lock held by caller | |
465 | * ensure we do not access cgroup data | |
466 | * unless we have the cgroup pinned (css_get) | |
467 | */ | |
468 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
469 | return; |
470 | ||
471 | cgrp = perf_cgroup_from_task(task); | |
472 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 473 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
474 | } |
475 | ||
476 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
477 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
478 | ||
479 | /* | |
480 | * reschedule events based on the cgroup constraint of task. | |
481 | * | |
482 | * mode SWOUT : schedule out everything | |
483 | * mode SWIN : schedule in based on cgroup for next | |
484 | */ | |
485 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
486 | { | |
487 | struct perf_cpu_context *cpuctx; | |
488 | struct pmu *pmu; | |
489 | unsigned long flags; | |
490 | ||
491 | /* | |
492 | * disable interrupts to avoid geting nr_cgroup | |
493 | * changes via __perf_event_disable(). Also | |
494 | * avoids preemption. | |
495 | */ | |
496 | local_irq_save(flags); | |
497 | ||
498 | /* | |
499 | * we reschedule only in the presence of cgroup | |
500 | * constrained events. | |
501 | */ | |
502 | rcu_read_lock(); | |
503 | ||
504 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 505 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
506 | if (cpuctx->unique_pmu != pmu) |
507 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 508 | |
e5d1367f SE |
509 | /* |
510 | * perf_cgroup_events says at least one | |
511 | * context on this CPU has cgroup events. | |
512 | * | |
513 | * ctx->nr_cgroups reports the number of cgroup | |
514 | * events for a context. | |
515 | */ | |
516 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
517 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
518 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
519 | |
520 | if (mode & PERF_CGROUP_SWOUT) { | |
521 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
522 | /* | |
523 | * must not be done before ctxswout due | |
524 | * to event_filter_match() in event_sched_out() | |
525 | */ | |
526 | cpuctx->cgrp = NULL; | |
527 | } | |
528 | ||
529 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 530 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
531 | /* |
532 | * set cgrp before ctxsw in to allow | |
533 | * event_filter_match() to not have to pass | |
534 | * task around | |
e5d1367f SE |
535 | */ |
536 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
537 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
538 | } | |
facc4307 PZ |
539 | perf_pmu_enable(cpuctx->ctx.pmu); |
540 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 541 | } |
e5d1367f SE |
542 | } |
543 | ||
544 | rcu_read_unlock(); | |
545 | ||
546 | local_irq_restore(flags); | |
547 | } | |
548 | ||
a8d757ef SE |
549 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
550 | struct task_struct *next) | |
e5d1367f | 551 | { |
a8d757ef SE |
552 | struct perf_cgroup *cgrp1; |
553 | struct perf_cgroup *cgrp2 = NULL; | |
554 | ||
555 | /* | |
556 | * we come here when we know perf_cgroup_events > 0 | |
557 | */ | |
558 | cgrp1 = perf_cgroup_from_task(task); | |
559 | ||
560 | /* | |
561 | * next is NULL when called from perf_event_enable_on_exec() | |
562 | * that will systematically cause a cgroup_switch() | |
563 | */ | |
564 | if (next) | |
565 | cgrp2 = perf_cgroup_from_task(next); | |
566 | ||
567 | /* | |
568 | * only schedule out current cgroup events if we know | |
569 | * that we are switching to a different cgroup. Otherwise, | |
570 | * do no touch the cgroup events. | |
571 | */ | |
572 | if (cgrp1 != cgrp2) | |
573 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
574 | } |
575 | ||
a8d757ef SE |
576 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
577 | struct task_struct *task) | |
e5d1367f | 578 | { |
a8d757ef SE |
579 | struct perf_cgroup *cgrp1; |
580 | struct perf_cgroup *cgrp2 = NULL; | |
581 | ||
582 | /* | |
583 | * we come here when we know perf_cgroup_events > 0 | |
584 | */ | |
585 | cgrp1 = perf_cgroup_from_task(task); | |
586 | ||
587 | /* prev can never be NULL */ | |
588 | cgrp2 = perf_cgroup_from_task(prev); | |
589 | ||
590 | /* | |
591 | * only need to schedule in cgroup events if we are changing | |
592 | * cgroup during ctxsw. Cgroup events were not scheduled | |
593 | * out of ctxsw out if that was not the case. | |
594 | */ | |
595 | if (cgrp1 != cgrp2) | |
596 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
597 | } |
598 | ||
599 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
600 | struct perf_event_attr *attr, | |
601 | struct perf_event *group_leader) | |
602 | { | |
603 | struct perf_cgroup *cgrp; | |
604 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
605 | struct fd f = fdget(fd); |
606 | int ret = 0; | |
e5d1367f | 607 | |
2903ff01 | 608 | if (!f.file) |
e5d1367f SE |
609 | return -EBADF; |
610 | ||
5a17f543 | 611 | css = css_tryget_from_dir(f.file->f_dentry, &perf_event_cgrp_subsys); |
3db272c0 LZ |
612 | if (IS_ERR(css)) { |
613 | ret = PTR_ERR(css); | |
614 | goto out; | |
615 | } | |
e5d1367f SE |
616 | |
617 | cgrp = container_of(css, struct perf_cgroup, css); | |
618 | event->cgrp = cgrp; | |
619 | ||
620 | /* | |
621 | * all events in a group must monitor | |
622 | * the same cgroup because a task belongs | |
623 | * to only one perf cgroup at a time | |
624 | */ | |
625 | if (group_leader && group_leader->cgrp != cgrp) { | |
626 | perf_detach_cgroup(event); | |
627 | ret = -EINVAL; | |
e5d1367f | 628 | } |
3db272c0 | 629 | out: |
2903ff01 | 630 | fdput(f); |
e5d1367f SE |
631 | return ret; |
632 | } | |
633 | ||
634 | static inline void | |
635 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
636 | { | |
637 | struct perf_cgroup_info *t; | |
638 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
639 | event->shadow_ctx_time = now - t->timestamp; | |
640 | } | |
641 | ||
642 | static inline void | |
643 | perf_cgroup_defer_enabled(struct perf_event *event) | |
644 | { | |
645 | /* | |
646 | * when the current task's perf cgroup does not match | |
647 | * the event's, we need to remember to call the | |
648 | * perf_mark_enable() function the first time a task with | |
649 | * a matching perf cgroup is scheduled in. | |
650 | */ | |
651 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
652 | event->cgrp_defer_enabled = 1; | |
653 | } | |
654 | ||
655 | static inline void | |
656 | perf_cgroup_mark_enabled(struct perf_event *event, | |
657 | struct perf_event_context *ctx) | |
658 | { | |
659 | struct perf_event *sub; | |
660 | u64 tstamp = perf_event_time(event); | |
661 | ||
662 | if (!event->cgrp_defer_enabled) | |
663 | return; | |
664 | ||
665 | event->cgrp_defer_enabled = 0; | |
666 | ||
667 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
668 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
669 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
670 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
671 | sub->cgrp_defer_enabled = 0; | |
672 | } | |
673 | } | |
674 | } | |
675 | #else /* !CONFIG_CGROUP_PERF */ | |
676 | ||
677 | static inline bool | |
678 | perf_cgroup_match(struct perf_event *event) | |
679 | { | |
680 | return true; | |
681 | } | |
682 | ||
683 | static inline void perf_detach_cgroup(struct perf_event *event) | |
684 | {} | |
685 | ||
686 | static inline int is_cgroup_event(struct perf_event *event) | |
687 | { | |
688 | return 0; | |
689 | } | |
690 | ||
691 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
692 | { | |
693 | return 0; | |
694 | } | |
695 | ||
696 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
697 | { | |
698 | } | |
699 | ||
700 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
701 | { | |
702 | } | |
703 | ||
a8d757ef SE |
704 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
705 | struct task_struct *next) | |
e5d1367f SE |
706 | { |
707 | } | |
708 | ||
a8d757ef SE |
709 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
710 | struct task_struct *task) | |
e5d1367f SE |
711 | { |
712 | } | |
713 | ||
714 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
715 | struct perf_event_attr *attr, | |
716 | struct perf_event *group_leader) | |
717 | { | |
718 | return -EINVAL; | |
719 | } | |
720 | ||
721 | static inline void | |
3f7cce3c SE |
722 | perf_cgroup_set_timestamp(struct task_struct *task, |
723 | struct perf_event_context *ctx) | |
e5d1367f SE |
724 | { |
725 | } | |
726 | ||
727 | void | |
728 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
729 | { | |
730 | } | |
731 | ||
732 | static inline void | |
733 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
734 | { | |
735 | } | |
736 | ||
737 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
738 | { | |
739 | return 0; | |
740 | } | |
741 | ||
742 | static inline void | |
743 | perf_cgroup_defer_enabled(struct perf_event *event) | |
744 | { | |
745 | } | |
746 | ||
747 | static inline void | |
748 | perf_cgroup_mark_enabled(struct perf_event *event, | |
749 | struct perf_event_context *ctx) | |
750 | { | |
751 | } | |
752 | #endif | |
753 | ||
9e630205 SE |
754 | /* |
755 | * set default to be dependent on timer tick just | |
756 | * like original code | |
757 | */ | |
758 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
759 | /* | |
760 | * function must be called with interrupts disbled | |
761 | */ | |
762 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
763 | { | |
764 | struct perf_cpu_context *cpuctx; | |
765 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
766 | int rotations = 0; | |
767 | ||
768 | WARN_ON(!irqs_disabled()); | |
769 | ||
770 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
771 | ||
772 | rotations = perf_rotate_context(cpuctx); | |
773 | ||
774 | /* | |
775 | * arm timer if needed | |
776 | */ | |
777 | if (rotations) { | |
778 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
779 | ret = HRTIMER_RESTART; | |
780 | } | |
781 | ||
782 | return ret; | |
783 | } | |
784 | ||
785 | /* CPU is going down */ | |
786 | void perf_cpu_hrtimer_cancel(int cpu) | |
787 | { | |
788 | struct perf_cpu_context *cpuctx; | |
789 | struct pmu *pmu; | |
790 | unsigned long flags; | |
791 | ||
792 | if (WARN_ON(cpu != smp_processor_id())) | |
793 | return; | |
794 | ||
795 | local_irq_save(flags); | |
796 | ||
797 | rcu_read_lock(); | |
798 | ||
799 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
800 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
801 | ||
802 | if (pmu->task_ctx_nr == perf_sw_context) | |
803 | continue; | |
804 | ||
805 | hrtimer_cancel(&cpuctx->hrtimer); | |
806 | } | |
807 | ||
808 | rcu_read_unlock(); | |
809 | ||
810 | local_irq_restore(flags); | |
811 | } | |
812 | ||
813 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
814 | { | |
815 | struct hrtimer *hr = &cpuctx->hrtimer; | |
816 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 817 | int timer; |
9e630205 SE |
818 | |
819 | /* no multiplexing needed for SW PMU */ | |
820 | if (pmu->task_ctx_nr == perf_sw_context) | |
821 | return; | |
822 | ||
62b85639 SE |
823 | /* |
824 | * check default is sane, if not set then force to | |
825 | * default interval (1/tick) | |
826 | */ | |
827 | timer = pmu->hrtimer_interval_ms; | |
828 | if (timer < 1) | |
829 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
830 | ||
831 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
832 | |
833 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
834 | hr->function = perf_cpu_hrtimer_handler; | |
835 | } | |
836 | ||
837 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
838 | { | |
839 | struct hrtimer *hr = &cpuctx->hrtimer; | |
840 | struct pmu *pmu = cpuctx->ctx.pmu; | |
841 | ||
842 | /* not for SW PMU */ | |
843 | if (pmu->task_ctx_nr == perf_sw_context) | |
844 | return; | |
845 | ||
846 | if (hrtimer_active(hr)) | |
847 | return; | |
848 | ||
849 | if (!hrtimer_callback_running(hr)) | |
850 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
851 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
852 | } | |
853 | ||
33696fc0 | 854 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 855 | { |
33696fc0 PZ |
856 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
857 | if (!(*count)++) | |
858 | pmu->pmu_disable(pmu); | |
9e35ad38 | 859 | } |
9e35ad38 | 860 | |
33696fc0 | 861 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 862 | { |
33696fc0 PZ |
863 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
864 | if (!--(*count)) | |
865 | pmu->pmu_enable(pmu); | |
9e35ad38 | 866 | } |
9e35ad38 | 867 | |
e9d2b064 PZ |
868 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
869 | ||
870 | /* | |
871 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
872 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
873 | * disabled, while rotate_context is called from IRQ context. | |
874 | */ | |
108b02cf | 875 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 876 | { |
108b02cf | 877 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 878 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 879 | |
e9d2b064 | 880 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 881 | |
d84153d6 | 882 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 883 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 884 | } |
9e35ad38 | 885 | |
cdd6c482 | 886 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 887 | { |
e5289d4a | 888 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
889 | } |
890 | ||
cdd6c482 | 891 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 892 | { |
564c2b21 PM |
893 | if (atomic_dec_and_test(&ctx->refcount)) { |
894 | if (ctx->parent_ctx) | |
895 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
896 | if (ctx->task) |
897 | put_task_struct(ctx->task); | |
cb796ff3 | 898 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 899 | } |
a63eaf34 PM |
900 | } |
901 | ||
cdd6c482 | 902 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
903 | { |
904 | if (ctx->parent_ctx) { | |
905 | put_ctx(ctx->parent_ctx); | |
906 | ctx->parent_ctx = NULL; | |
907 | } | |
5a3126d4 | 908 | ctx->generation++; |
71a851b4 PZ |
909 | } |
910 | ||
6844c09d ACM |
911 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
912 | { | |
913 | /* | |
914 | * only top level events have the pid namespace they were created in | |
915 | */ | |
916 | if (event->parent) | |
917 | event = event->parent; | |
918 | ||
919 | return task_tgid_nr_ns(p, event->ns); | |
920 | } | |
921 | ||
922 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
923 | { | |
924 | /* | |
925 | * only top level events have the pid namespace they were created in | |
926 | */ | |
927 | if (event->parent) | |
928 | event = event->parent; | |
929 | ||
930 | return task_pid_nr_ns(p, event->ns); | |
931 | } | |
932 | ||
7f453c24 | 933 | /* |
cdd6c482 | 934 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
935 | * to userspace. |
936 | */ | |
cdd6c482 | 937 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 938 | { |
cdd6c482 | 939 | u64 id = event->id; |
7f453c24 | 940 | |
cdd6c482 IM |
941 | if (event->parent) |
942 | id = event->parent->id; | |
7f453c24 PZ |
943 | |
944 | return id; | |
945 | } | |
946 | ||
25346b93 | 947 | /* |
cdd6c482 | 948 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
949 | * This has to cope with with the fact that until it is locked, |
950 | * the context could get moved to another task. | |
951 | */ | |
cdd6c482 | 952 | static struct perf_event_context * |
8dc85d54 | 953 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 954 | { |
cdd6c482 | 955 | struct perf_event_context *ctx; |
25346b93 | 956 | |
9ed6060d | 957 | retry: |
058ebd0e PZ |
958 | /* |
959 | * One of the few rules of preemptible RCU is that one cannot do | |
960 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
961 | * part of the read side critical section was preemptible -- see | |
962 | * rcu_read_unlock_special(). | |
963 | * | |
964 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
965 | * side critical section is non-preemptible. | |
966 | */ | |
967 | preempt_disable(); | |
968 | rcu_read_lock(); | |
8dc85d54 | 969 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
970 | if (ctx) { |
971 | /* | |
972 | * If this context is a clone of another, it might | |
973 | * get swapped for another underneath us by | |
cdd6c482 | 974 | * perf_event_task_sched_out, though the |
25346b93 PM |
975 | * rcu_read_lock() protects us from any context |
976 | * getting freed. Lock the context and check if it | |
977 | * got swapped before we could get the lock, and retry | |
978 | * if so. If we locked the right context, then it | |
979 | * can't get swapped on us any more. | |
980 | */ | |
e625cce1 | 981 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 982 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 983 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
984 | rcu_read_unlock(); |
985 | preempt_enable(); | |
25346b93 PM |
986 | goto retry; |
987 | } | |
b49a9e7e PZ |
988 | |
989 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 990 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
991 | ctx = NULL; |
992 | } | |
25346b93 PM |
993 | } |
994 | rcu_read_unlock(); | |
058ebd0e | 995 | preempt_enable(); |
25346b93 PM |
996 | return ctx; |
997 | } | |
998 | ||
999 | /* | |
1000 | * Get the context for a task and increment its pin_count so it | |
1001 | * can't get swapped to another task. This also increments its | |
1002 | * reference count so that the context can't get freed. | |
1003 | */ | |
8dc85d54 PZ |
1004 | static struct perf_event_context * |
1005 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1006 | { |
cdd6c482 | 1007 | struct perf_event_context *ctx; |
25346b93 PM |
1008 | unsigned long flags; |
1009 | ||
8dc85d54 | 1010 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1011 | if (ctx) { |
1012 | ++ctx->pin_count; | |
e625cce1 | 1013 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1014 | } |
1015 | return ctx; | |
1016 | } | |
1017 | ||
cdd6c482 | 1018 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1019 | { |
1020 | unsigned long flags; | |
1021 | ||
e625cce1 | 1022 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1023 | --ctx->pin_count; |
e625cce1 | 1024 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1025 | } |
1026 | ||
f67218c3 PZ |
1027 | /* |
1028 | * Update the record of the current time in a context. | |
1029 | */ | |
1030 | static void update_context_time(struct perf_event_context *ctx) | |
1031 | { | |
1032 | u64 now = perf_clock(); | |
1033 | ||
1034 | ctx->time += now - ctx->timestamp; | |
1035 | ctx->timestamp = now; | |
1036 | } | |
1037 | ||
4158755d SE |
1038 | static u64 perf_event_time(struct perf_event *event) |
1039 | { | |
1040 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1041 | |
1042 | if (is_cgroup_event(event)) | |
1043 | return perf_cgroup_event_time(event); | |
1044 | ||
4158755d SE |
1045 | return ctx ? ctx->time : 0; |
1046 | } | |
1047 | ||
f67218c3 PZ |
1048 | /* |
1049 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1050 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1051 | */ |
1052 | static void update_event_times(struct perf_event *event) | |
1053 | { | |
1054 | struct perf_event_context *ctx = event->ctx; | |
1055 | u64 run_end; | |
1056 | ||
1057 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1058 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1059 | return; | |
e5d1367f SE |
1060 | /* |
1061 | * in cgroup mode, time_enabled represents | |
1062 | * the time the event was enabled AND active | |
1063 | * tasks were in the monitored cgroup. This is | |
1064 | * independent of the activity of the context as | |
1065 | * there may be a mix of cgroup and non-cgroup events. | |
1066 | * | |
1067 | * That is why we treat cgroup events differently | |
1068 | * here. | |
1069 | */ | |
1070 | if (is_cgroup_event(event)) | |
46cd6a7f | 1071 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1072 | else if (ctx->is_active) |
1073 | run_end = ctx->time; | |
acd1d7c1 PZ |
1074 | else |
1075 | run_end = event->tstamp_stopped; | |
1076 | ||
1077 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1078 | |
1079 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1080 | run_end = event->tstamp_stopped; | |
1081 | else | |
4158755d | 1082 | run_end = perf_event_time(event); |
f67218c3 PZ |
1083 | |
1084 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1085 | |
f67218c3 PZ |
1086 | } |
1087 | ||
96c21a46 PZ |
1088 | /* |
1089 | * Update total_time_enabled and total_time_running for all events in a group. | |
1090 | */ | |
1091 | static void update_group_times(struct perf_event *leader) | |
1092 | { | |
1093 | struct perf_event *event; | |
1094 | ||
1095 | update_event_times(leader); | |
1096 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1097 | update_event_times(event); | |
1098 | } | |
1099 | ||
889ff015 FW |
1100 | static struct list_head * |
1101 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1102 | { | |
1103 | if (event->attr.pinned) | |
1104 | return &ctx->pinned_groups; | |
1105 | else | |
1106 | return &ctx->flexible_groups; | |
1107 | } | |
1108 | ||
fccc714b | 1109 | /* |
cdd6c482 | 1110 | * Add a event from the lists for its context. |
fccc714b PZ |
1111 | * Must be called with ctx->mutex and ctx->lock held. |
1112 | */ | |
04289bb9 | 1113 | static void |
cdd6c482 | 1114 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1115 | { |
8a49542c PZ |
1116 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1117 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1118 | |
1119 | /* | |
8a49542c PZ |
1120 | * If we're a stand alone event or group leader, we go to the context |
1121 | * list, group events are kept attached to the group so that | |
1122 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1123 | */ |
8a49542c | 1124 | if (event->group_leader == event) { |
889ff015 FW |
1125 | struct list_head *list; |
1126 | ||
d6f962b5 FW |
1127 | if (is_software_event(event)) |
1128 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1129 | ||
889ff015 FW |
1130 | list = ctx_group_list(event, ctx); |
1131 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1132 | } |
592903cd | 1133 | |
08309379 | 1134 | if (is_cgroup_event(event)) |
e5d1367f | 1135 | ctx->nr_cgroups++; |
e5d1367f | 1136 | |
d010b332 SE |
1137 | if (has_branch_stack(event)) |
1138 | ctx->nr_branch_stack++; | |
1139 | ||
cdd6c482 | 1140 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1141 | if (!ctx->nr_events) |
108b02cf | 1142 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1143 | ctx->nr_events++; |
1144 | if (event->attr.inherit_stat) | |
bfbd3381 | 1145 | ctx->nr_stat++; |
5a3126d4 PZ |
1146 | |
1147 | ctx->generation++; | |
04289bb9 IM |
1148 | } |
1149 | ||
0231bb53 JO |
1150 | /* |
1151 | * Initialize event state based on the perf_event_attr::disabled. | |
1152 | */ | |
1153 | static inline void perf_event__state_init(struct perf_event *event) | |
1154 | { | |
1155 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1156 | PERF_EVENT_STATE_INACTIVE; | |
1157 | } | |
1158 | ||
c320c7b7 ACM |
1159 | /* |
1160 | * Called at perf_event creation and when events are attached/detached from a | |
1161 | * group. | |
1162 | */ | |
1163 | static void perf_event__read_size(struct perf_event *event) | |
1164 | { | |
1165 | int entry = sizeof(u64); /* value */ | |
1166 | int size = 0; | |
1167 | int nr = 1; | |
1168 | ||
1169 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1170 | size += sizeof(u64); | |
1171 | ||
1172 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1173 | size += sizeof(u64); | |
1174 | ||
1175 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1176 | entry += sizeof(u64); | |
1177 | ||
1178 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1179 | nr += event->group_leader->nr_siblings; | |
1180 | size += sizeof(u64); | |
1181 | } | |
1182 | ||
1183 | size += entry * nr; | |
1184 | event->read_size = size; | |
1185 | } | |
1186 | ||
1187 | static void perf_event__header_size(struct perf_event *event) | |
1188 | { | |
1189 | struct perf_sample_data *data; | |
1190 | u64 sample_type = event->attr.sample_type; | |
1191 | u16 size = 0; | |
1192 | ||
1193 | perf_event__read_size(event); | |
1194 | ||
1195 | if (sample_type & PERF_SAMPLE_IP) | |
1196 | size += sizeof(data->ip); | |
1197 | ||
6844c09d ACM |
1198 | if (sample_type & PERF_SAMPLE_ADDR) |
1199 | size += sizeof(data->addr); | |
1200 | ||
1201 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1202 | size += sizeof(data->period); | |
1203 | ||
c3feedf2 AK |
1204 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1205 | size += sizeof(data->weight); | |
1206 | ||
6844c09d ACM |
1207 | if (sample_type & PERF_SAMPLE_READ) |
1208 | size += event->read_size; | |
1209 | ||
d6be9ad6 SE |
1210 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1211 | size += sizeof(data->data_src.val); | |
1212 | ||
fdfbbd07 AK |
1213 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1214 | size += sizeof(data->txn); | |
1215 | ||
6844c09d ACM |
1216 | event->header_size = size; |
1217 | } | |
1218 | ||
1219 | static void perf_event__id_header_size(struct perf_event *event) | |
1220 | { | |
1221 | struct perf_sample_data *data; | |
1222 | u64 sample_type = event->attr.sample_type; | |
1223 | u16 size = 0; | |
1224 | ||
c320c7b7 ACM |
1225 | if (sample_type & PERF_SAMPLE_TID) |
1226 | size += sizeof(data->tid_entry); | |
1227 | ||
1228 | if (sample_type & PERF_SAMPLE_TIME) | |
1229 | size += sizeof(data->time); | |
1230 | ||
ff3d527c AH |
1231 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1232 | size += sizeof(data->id); | |
1233 | ||
c320c7b7 ACM |
1234 | if (sample_type & PERF_SAMPLE_ID) |
1235 | size += sizeof(data->id); | |
1236 | ||
1237 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1238 | size += sizeof(data->stream_id); | |
1239 | ||
1240 | if (sample_type & PERF_SAMPLE_CPU) | |
1241 | size += sizeof(data->cpu_entry); | |
1242 | ||
6844c09d | 1243 | event->id_header_size = size; |
c320c7b7 ACM |
1244 | } |
1245 | ||
8a49542c PZ |
1246 | static void perf_group_attach(struct perf_event *event) |
1247 | { | |
c320c7b7 | 1248 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1249 | |
74c3337c PZ |
1250 | /* |
1251 | * We can have double attach due to group movement in perf_event_open. | |
1252 | */ | |
1253 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1254 | return; | |
1255 | ||
8a49542c PZ |
1256 | event->attach_state |= PERF_ATTACH_GROUP; |
1257 | ||
1258 | if (group_leader == event) | |
1259 | return; | |
1260 | ||
1261 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1262 | !is_software_event(event)) | |
1263 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1264 | ||
1265 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1266 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1267 | |
1268 | perf_event__header_size(group_leader); | |
1269 | ||
1270 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1271 | perf_event__header_size(pos); | |
8a49542c PZ |
1272 | } |
1273 | ||
a63eaf34 | 1274 | /* |
cdd6c482 | 1275 | * Remove a event from the lists for its context. |
fccc714b | 1276 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1277 | */ |
04289bb9 | 1278 | static void |
cdd6c482 | 1279 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1280 | { |
68cacd29 | 1281 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1282 | /* |
1283 | * We can have double detach due to exit/hot-unplug + close. | |
1284 | */ | |
1285 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1286 | return; |
8a49542c PZ |
1287 | |
1288 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1289 | ||
68cacd29 | 1290 | if (is_cgroup_event(event)) { |
e5d1367f | 1291 | ctx->nr_cgroups--; |
68cacd29 SE |
1292 | cpuctx = __get_cpu_context(ctx); |
1293 | /* | |
1294 | * if there are no more cgroup events | |
1295 | * then cler cgrp to avoid stale pointer | |
1296 | * in update_cgrp_time_from_cpuctx() | |
1297 | */ | |
1298 | if (!ctx->nr_cgroups) | |
1299 | cpuctx->cgrp = NULL; | |
1300 | } | |
e5d1367f | 1301 | |
d010b332 SE |
1302 | if (has_branch_stack(event)) |
1303 | ctx->nr_branch_stack--; | |
1304 | ||
cdd6c482 IM |
1305 | ctx->nr_events--; |
1306 | if (event->attr.inherit_stat) | |
bfbd3381 | 1307 | ctx->nr_stat--; |
8bc20959 | 1308 | |
cdd6c482 | 1309 | list_del_rcu(&event->event_entry); |
04289bb9 | 1310 | |
8a49542c PZ |
1311 | if (event->group_leader == event) |
1312 | list_del_init(&event->group_entry); | |
5c148194 | 1313 | |
96c21a46 | 1314 | update_group_times(event); |
b2e74a26 SE |
1315 | |
1316 | /* | |
1317 | * If event was in error state, then keep it | |
1318 | * that way, otherwise bogus counts will be | |
1319 | * returned on read(). The only way to get out | |
1320 | * of error state is by explicit re-enabling | |
1321 | * of the event | |
1322 | */ | |
1323 | if (event->state > PERF_EVENT_STATE_OFF) | |
1324 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1325 | |
1326 | ctx->generation++; | |
050735b0 PZ |
1327 | } |
1328 | ||
8a49542c | 1329 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1330 | { |
1331 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1332 | struct list_head *list = NULL; |
1333 | ||
1334 | /* | |
1335 | * We can have double detach due to exit/hot-unplug + close. | |
1336 | */ | |
1337 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1338 | return; | |
1339 | ||
1340 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1341 | ||
1342 | /* | |
1343 | * If this is a sibling, remove it from its group. | |
1344 | */ | |
1345 | if (event->group_leader != event) { | |
1346 | list_del_init(&event->group_entry); | |
1347 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1348 | goto out; |
8a49542c PZ |
1349 | } |
1350 | ||
1351 | if (!list_empty(&event->group_entry)) | |
1352 | list = &event->group_entry; | |
2e2af50b | 1353 | |
04289bb9 | 1354 | /* |
cdd6c482 IM |
1355 | * If this was a group event with sibling events then |
1356 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1357 | * to whatever list we are on. |
04289bb9 | 1358 | */ |
cdd6c482 | 1359 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1360 | if (list) |
1361 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1362 | sibling->group_leader = sibling; |
d6f962b5 FW |
1363 | |
1364 | /* Inherit group flags from the previous leader */ | |
1365 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1366 | } |
c320c7b7 ACM |
1367 | |
1368 | out: | |
1369 | perf_event__header_size(event->group_leader); | |
1370 | ||
1371 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1372 | perf_event__header_size(tmp); | |
04289bb9 IM |
1373 | } |
1374 | ||
fa66f07a SE |
1375 | static inline int |
1376 | event_filter_match(struct perf_event *event) | |
1377 | { | |
e5d1367f SE |
1378 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1379 | && perf_cgroup_match(event); | |
fa66f07a SE |
1380 | } |
1381 | ||
9ffcfa6f SE |
1382 | static void |
1383 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1384 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1385 | struct perf_event_context *ctx) |
3b6f9e5c | 1386 | { |
4158755d | 1387 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1388 | u64 delta; |
1389 | /* | |
1390 | * An event which could not be activated because of | |
1391 | * filter mismatch still needs to have its timings | |
1392 | * maintained, otherwise bogus information is return | |
1393 | * via read() for time_enabled, time_running: | |
1394 | */ | |
1395 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1396 | && !event_filter_match(event)) { | |
e5d1367f | 1397 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1398 | event->tstamp_running += delta; |
4158755d | 1399 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1400 | } |
1401 | ||
cdd6c482 | 1402 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1403 | return; |
3b6f9e5c | 1404 | |
44377277 AS |
1405 | perf_pmu_disable(event->pmu); |
1406 | ||
cdd6c482 IM |
1407 | event->state = PERF_EVENT_STATE_INACTIVE; |
1408 | if (event->pending_disable) { | |
1409 | event->pending_disable = 0; | |
1410 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1411 | } |
4158755d | 1412 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1413 | event->pmu->del(event, 0); |
cdd6c482 | 1414 | event->oncpu = -1; |
3b6f9e5c | 1415 | |
cdd6c482 | 1416 | if (!is_software_event(event)) |
3b6f9e5c PM |
1417 | cpuctx->active_oncpu--; |
1418 | ctx->nr_active--; | |
0f5a2601 PZ |
1419 | if (event->attr.freq && event->attr.sample_freq) |
1420 | ctx->nr_freq--; | |
cdd6c482 | 1421 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1422 | cpuctx->exclusive = 0; |
44377277 AS |
1423 | |
1424 | perf_pmu_enable(event->pmu); | |
3b6f9e5c PM |
1425 | } |
1426 | ||
d859e29f | 1427 | static void |
cdd6c482 | 1428 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1429 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1430 | struct perf_event_context *ctx) |
d859e29f | 1431 | { |
cdd6c482 | 1432 | struct perf_event *event; |
fa66f07a | 1433 | int state = group_event->state; |
d859e29f | 1434 | |
cdd6c482 | 1435 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1436 | |
1437 | /* | |
1438 | * Schedule out siblings (if any): | |
1439 | */ | |
cdd6c482 IM |
1440 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1441 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1442 | |
fa66f07a | 1443 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1444 | cpuctx->exclusive = 0; |
1445 | } | |
1446 | ||
46ce0fe9 PZ |
1447 | struct remove_event { |
1448 | struct perf_event *event; | |
1449 | bool detach_group; | |
1450 | }; | |
1451 | ||
0793a61d | 1452 | /* |
cdd6c482 | 1453 | * Cross CPU call to remove a performance event |
0793a61d | 1454 | * |
cdd6c482 | 1455 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1456 | * remove it from the context list. |
1457 | */ | |
fe4b04fa | 1458 | static int __perf_remove_from_context(void *info) |
0793a61d | 1459 | { |
46ce0fe9 PZ |
1460 | struct remove_event *re = info; |
1461 | struct perf_event *event = re->event; | |
cdd6c482 | 1462 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1463 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1464 | |
e625cce1 | 1465 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1466 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1467 | if (re->detach_group) |
1468 | perf_group_detach(event); | |
cdd6c482 | 1469 | list_del_event(event, ctx); |
64ce3126 PZ |
1470 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1471 | ctx->is_active = 0; | |
1472 | cpuctx->task_ctx = NULL; | |
1473 | } | |
e625cce1 | 1474 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1475 | |
1476 | return 0; | |
0793a61d TG |
1477 | } |
1478 | ||
1479 | ||
1480 | /* | |
cdd6c482 | 1481 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1482 | * |
cdd6c482 | 1483 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1484 | * call when the task is on a CPU. |
c93f7669 | 1485 | * |
cdd6c482 IM |
1486 | * If event->ctx is a cloned context, callers must make sure that |
1487 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1488 | * remains valid. This is OK when called from perf_release since |
1489 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1490 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1491 | * context has been detached from its task. |
0793a61d | 1492 | */ |
46ce0fe9 | 1493 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1494 | { |
cdd6c482 | 1495 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1496 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1497 | struct remove_event re = { |
1498 | .event = event, | |
1499 | .detach_group = detach_group, | |
1500 | }; | |
0793a61d | 1501 | |
fe4b04fa PZ |
1502 | lockdep_assert_held(&ctx->mutex); |
1503 | ||
0793a61d TG |
1504 | if (!task) { |
1505 | /* | |
cdd6c482 | 1506 | * Per cpu events are removed via an smp call and |
af901ca1 | 1507 | * the removal is always successful. |
0793a61d | 1508 | */ |
46ce0fe9 | 1509 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1510 | return; |
1511 | } | |
1512 | ||
1513 | retry: | |
46ce0fe9 | 1514 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1515 | return; |
0793a61d | 1516 | |
e625cce1 | 1517 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1518 | /* |
fe4b04fa PZ |
1519 | * If we failed to find a running task, but find the context active now |
1520 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1521 | */ |
fe4b04fa | 1522 | if (ctx->is_active) { |
e625cce1 | 1523 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1524 | goto retry; |
1525 | } | |
1526 | ||
1527 | /* | |
fe4b04fa PZ |
1528 | * Since the task isn't running, its safe to remove the event, us |
1529 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1530 | */ |
46ce0fe9 PZ |
1531 | if (detach_group) |
1532 | perf_group_detach(event); | |
fe4b04fa | 1533 | list_del_event(event, ctx); |
e625cce1 | 1534 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1535 | } |
1536 | ||
d859e29f | 1537 | /* |
cdd6c482 | 1538 | * Cross CPU call to disable a performance event |
d859e29f | 1539 | */ |
500ad2d8 | 1540 | int __perf_event_disable(void *info) |
d859e29f | 1541 | { |
cdd6c482 | 1542 | struct perf_event *event = info; |
cdd6c482 | 1543 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1544 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1545 | |
1546 | /* | |
cdd6c482 IM |
1547 | * If this is a per-task event, need to check whether this |
1548 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1549 | * |
1550 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1551 | * flipping contexts around. | |
d859e29f | 1552 | */ |
665c2142 | 1553 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1554 | return -EINVAL; |
d859e29f | 1555 | |
e625cce1 | 1556 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1557 | |
1558 | /* | |
cdd6c482 | 1559 | * If the event is on, turn it off. |
d859e29f PM |
1560 | * If it is in error state, leave it in error state. |
1561 | */ | |
cdd6c482 | 1562 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1563 | update_context_time(ctx); |
e5d1367f | 1564 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1565 | update_group_times(event); |
1566 | if (event == event->group_leader) | |
1567 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1568 | else |
cdd6c482 IM |
1569 | event_sched_out(event, cpuctx, ctx); |
1570 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1571 | } |
1572 | ||
e625cce1 | 1573 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1574 | |
1575 | return 0; | |
d859e29f PM |
1576 | } |
1577 | ||
1578 | /* | |
cdd6c482 | 1579 | * Disable a event. |
c93f7669 | 1580 | * |
cdd6c482 IM |
1581 | * If event->ctx is a cloned context, callers must make sure that |
1582 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1583 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1584 | * perf_event_for_each_child or perf_event_for_each because they |
1585 | * hold the top-level event's child_mutex, so any descendant that | |
1586 | * goes to exit will block in sync_child_event. | |
1587 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1588 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1589 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1590 | */ |
44234adc | 1591 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1592 | { |
cdd6c482 | 1593 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1594 | struct task_struct *task = ctx->task; |
1595 | ||
1596 | if (!task) { | |
1597 | /* | |
cdd6c482 | 1598 | * Disable the event on the cpu that it's on |
d859e29f | 1599 | */ |
fe4b04fa | 1600 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1601 | return; |
1602 | } | |
1603 | ||
9ed6060d | 1604 | retry: |
fe4b04fa PZ |
1605 | if (!task_function_call(task, __perf_event_disable, event)) |
1606 | return; | |
d859e29f | 1607 | |
e625cce1 | 1608 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1609 | /* |
cdd6c482 | 1610 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1611 | */ |
cdd6c482 | 1612 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1613 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1614 | /* |
1615 | * Reload the task pointer, it might have been changed by | |
1616 | * a concurrent perf_event_context_sched_out(). | |
1617 | */ | |
1618 | task = ctx->task; | |
d859e29f PM |
1619 | goto retry; |
1620 | } | |
1621 | ||
1622 | /* | |
1623 | * Since we have the lock this context can't be scheduled | |
1624 | * in, so we can change the state safely. | |
1625 | */ | |
cdd6c482 IM |
1626 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1627 | update_group_times(event); | |
1628 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1629 | } |
e625cce1 | 1630 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1631 | } |
dcfce4a0 | 1632 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1633 | |
e5d1367f SE |
1634 | static void perf_set_shadow_time(struct perf_event *event, |
1635 | struct perf_event_context *ctx, | |
1636 | u64 tstamp) | |
1637 | { | |
1638 | /* | |
1639 | * use the correct time source for the time snapshot | |
1640 | * | |
1641 | * We could get by without this by leveraging the | |
1642 | * fact that to get to this function, the caller | |
1643 | * has most likely already called update_context_time() | |
1644 | * and update_cgrp_time_xx() and thus both timestamp | |
1645 | * are identical (or very close). Given that tstamp is, | |
1646 | * already adjusted for cgroup, we could say that: | |
1647 | * tstamp - ctx->timestamp | |
1648 | * is equivalent to | |
1649 | * tstamp - cgrp->timestamp. | |
1650 | * | |
1651 | * Then, in perf_output_read(), the calculation would | |
1652 | * work with no changes because: | |
1653 | * - event is guaranteed scheduled in | |
1654 | * - no scheduled out in between | |
1655 | * - thus the timestamp would be the same | |
1656 | * | |
1657 | * But this is a bit hairy. | |
1658 | * | |
1659 | * So instead, we have an explicit cgroup call to remain | |
1660 | * within the time time source all along. We believe it | |
1661 | * is cleaner and simpler to understand. | |
1662 | */ | |
1663 | if (is_cgroup_event(event)) | |
1664 | perf_cgroup_set_shadow_time(event, tstamp); | |
1665 | else | |
1666 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1667 | } | |
1668 | ||
4fe757dd PZ |
1669 | #define MAX_INTERRUPTS (~0ULL) |
1670 | ||
1671 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1672 | ||
235c7fc7 | 1673 | static int |
9ffcfa6f | 1674 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1675 | struct perf_cpu_context *cpuctx, |
6e37738a | 1676 | struct perf_event_context *ctx) |
235c7fc7 | 1677 | { |
4158755d | 1678 | u64 tstamp = perf_event_time(event); |
44377277 | 1679 | int ret = 0; |
4158755d | 1680 | |
63342411 PZ |
1681 | lockdep_assert_held(&ctx->lock); |
1682 | ||
cdd6c482 | 1683 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1684 | return 0; |
1685 | ||
cdd6c482 | 1686 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1687 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1688 | |
1689 | /* | |
1690 | * Unthrottle events, since we scheduled we might have missed several | |
1691 | * ticks already, also for a heavily scheduling task there is little | |
1692 | * guarantee it'll get a tick in a timely manner. | |
1693 | */ | |
1694 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1695 | perf_log_throttle(event, 1); | |
1696 | event->hw.interrupts = 0; | |
1697 | } | |
1698 | ||
235c7fc7 IM |
1699 | /* |
1700 | * The new state must be visible before we turn it on in the hardware: | |
1701 | */ | |
1702 | smp_wmb(); | |
1703 | ||
44377277 AS |
1704 | perf_pmu_disable(event->pmu); |
1705 | ||
a4eaf7f1 | 1706 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1707 | event->state = PERF_EVENT_STATE_INACTIVE; |
1708 | event->oncpu = -1; | |
44377277 AS |
1709 | ret = -EAGAIN; |
1710 | goto out; | |
235c7fc7 IM |
1711 | } |
1712 | ||
4158755d | 1713 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1714 | |
e5d1367f | 1715 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1716 | |
cdd6c482 | 1717 | if (!is_software_event(event)) |
3b6f9e5c | 1718 | cpuctx->active_oncpu++; |
235c7fc7 | 1719 | ctx->nr_active++; |
0f5a2601 PZ |
1720 | if (event->attr.freq && event->attr.sample_freq) |
1721 | ctx->nr_freq++; | |
235c7fc7 | 1722 | |
cdd6c482 | 1723 | if (event->attr.exclusive) |
3b6f9e5c PM |
1724 | cpuctx->exclusive = 1; |
1725 | ||
44377277 AS |
1726 | out: |
1727 | perf_pmu_enable(event->pmu); | |
1728 | ||
1729 | return ret; | |
235c7fc7 IM |
1730 | } |
1731 | ||
6751b71e | 1732 | static int |
cdd6c482 | 1733 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1734 | struct perf_cpu_context *cpuctx, |
6e37738a | 1735 | struct perf_event_context *ctx) |
6751b71e | 1736 | { |
6bde9b6c | 1737 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1738 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1739 | u64 now = ctx->time; |
1740 | bool simulate = false; | |
6751b71e | 1741 | |
cdd6c482 | 1742 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1743 | return 0; |
1744 | ||
ad5133b7 | 1745 | pmu->start_txn(pmu); |
6bde9b6c | 1746 | |
9ffcfa6f | 1747 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1748 | pmu->cancel_txn(pmu); |
9e630205 | 1749 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1750 | return -EAGAIN; |
90151c35 | 1751 | } |
6751b71e PM |
1752 | |
1753 | /* | |
1754 | * Schedule in siblings as one group (if any): | |
1755 | */ | |
cdd6c482 | 1756 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1757 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1758 | partial_group = event; |
6751b71e PM |
1759 | goto group_error; |
1760 | } | |
1761 | } | |
1762 | ||
9ffcfa6f | 1763 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1764 | return 0; |
9ffcfa6f | 1765 | |
6751b71e PM |
1766 | group_error: |
1767 | /* | |
1768 | * Groups can be scheduled in as one unit only, so undo any | |
1769 | * partial group before returning: | |
d7842da4 SE |
1770 | * The events up to the failed event are scheduled out normally, |
1771 | * tstamp_stopped will be updated. | |
1772 | * | |
1773 | * The failed events and the remaining siblings need to have | |
1774 | * their timings updated as if they had gone thru event_sched_in() | |
1775 | * and event_sched_out(). This is required to get consistent timings | |
1776 | * across the group. This also takes care of the case where the group | |
1777 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1778 | * the time the event was actually stopped, such that time delta | |
1779 | * calculation in update_event_times() is correct. | |
6751b71e | 1780 | */ |
cdd6c482 IM |
1781 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1782 | if (event == partial_group) | |
d7842da4 SE |
1783 | simulate = true; |
1784 | ||
1785 | if (simulate) { | |
1786 | event->tstamp_running += now - event->tstamp_stopped; | |
1787 | event->tstamp_stopped = now; | |
1788 | } else { | |
1789 | event_sched_out(event, cpuctx, ctx); | |
1790 | } | |
6751b71e | 1791 | } |
9ffcfa6f | 1792 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1793 | |
ad5133b7 | 1794 | pmu->cancel_txn(pmu); |
90151c35 | 1795 | |
9e630205 SE |
1796 | perf_cpu_hrtimer_restart(cpuctx); |
1797 | ||
6751b71e PM |
1798 | return -EAGAIN; |
1799 | } | |
1800 | ||
3b6f9e5c | 1801 | /* |
cdd6c482 | 1802 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1803 | */ |
cdd6c482 | 1804 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1805 | struct perf_cpu_context *cpuctx, |
1806 | int can_add_hw) | |
1807 | { | |
1808 | /* | |
cdd6c482 | 1809 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1810 | */ |
d6f962b5 | 1811 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1812 | return 1; |
1813 | /* | |
1814 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1815 | * events can go on. |
3b6f9e5c PM |
1816 | */ |
1817 | if (cpuctx->exclusive) | |
1818 | return 0; | |
1819 | /* | |
1820 | * If this group is exclusive and there are already | |
cdd6c482 | 1821 | * events on the CPU, it can't go on. |
3b6f9e5c | 1822 | */ |
cdd6c482 | 1823 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1824 | return 0; |
1825 | /* | |
1826 | * Otherwise, try to add it if all previous groups were able | |
1827 | * to go on. | |
1828 | */ | |
1829 | return can_add_hw; | |
1830 | } | |
1831 | ||
cdd6c482 IM |
1832 | static void add_event_to_ctx(struct perf_event *event, |
1833 | struct perf_event_context *ctx) | |
53cfbf59 | 1834 | { |
4158755d SE |
1835 | u64 tstamp = perf_event_time(event); |
1836 | ||
cdd6c482 | 1837 | list_add_event(event, ctx); |
8a49542c | 1838 | perf_group_attach(event); |
4158755d SE |
1839 | event->tstamp_enabled = tstamp; |
1840 | event->tstamp_running = tstamp; | |
1841 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1842 | } |
1843 | ||
2c29ef0f PZ |
1844 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1845 | static void | |
1846 | ctx_sched_in(struct perf_event_context *ctx, | |
1847 | struct perf_cpu_context *cpuctx, | |
1848 | enum event_type_t event_type, | |
1849 | struct task_struct *task); | |
fe4b04fa | 1850 | |
dce5855b PZ |
1851 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1852 | struct perf_event_context *ctx, | |
1853 | struct task_struct *task) | |
1854 | { | |
1855 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1856 | if (ctx) | |
1857 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1858 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1859 | if (ctx) | |
1860 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1861 | } | |
1862 | ||
0793a61d | 1863 | /* |
cdd6c482 | 1864 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1865 | * |
1866 | * Must be called with ctx->mutex held | |
0793a61d | 1867 | */ |
fe4b04fa | 1868 | static int __perf_install_in_context(void *info) |
0793a61d | 1869 | { |
cdd6c482 IM |
1870 | struct perf_event *event = info; |
1871 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1872 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1873 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1874 | struct task_struct *task = current; | |
1875 | ||
b58f6b0d | 1876 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1877 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1878 | |
1879 | /* | |
2c29ef0f | 1880 | * If there was an active task_ctx schedule it out. |
0793a61d | 1881 | */ |
b58f6b0d | 1882 | if (task_ctx) |
2c29ef0f | 1883 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1884 | |
1885 | /* | |
1886 | * If the context we're installing events in is not the | |
1887 | * active task_ctx, flip them. | |
1888 | */ | |
1889 | if (ctx->task && task_ctx != ctx) { | |
1890 | if (task_ctx) | |
1891 | raw_spin_unlock(&task_ctx->lock); | |
1892 | raw_spin_lock(&ctx->lock); | |
1893 | task_ctx = ctx; | |
1894 | } | |
1895 | ||
1896 | if (task_ctx) { | |
1897 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1898 | task = task_ctx->task; |
1899 | } | |
b58f6b0d | 1900 | |
2c29ef0f | 1901 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1902 | |
4af4998b | 1903 | update_context_time(ctx); |
e5d1367f SE |
1904 | /* |
1905 | * update cgrp time only if current cgrp | |
1906 | * matches event->cgrp. Must be done before | |
1907 | * calling add_event_to_ctx() | |
1908 | */ | |
1909 | update_cgrp_time_from_event(event); | |
0793a61d | 1910 | |
cdd6c482 | 1911 | add_event_to_ctx(event, ctx); |
0793a61d | 1912 | |
d859e29f | 1913 | /* |
2c29ef0f | 1914 | * Schedule everything back in |
d859e29f | 1915 | */ |
dce5855b | 1916 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1917 | |
1918 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1919 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1920 | |
1921 | return 0; | |
0793a61d TG |
1922 | } |
1923 | ||
1924 | /* | |
cdd6c482 | 1925 | * Attach a performance event to a context |
0793a61d | 1926 | * |
cdd6c482 IM |
1927 | * First we add the event to the list with the hardware enable bit |
1928 | * in event->hw_config cleared. | |
0793a61d | 1929 | * |
cdd6c482 | 1930 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1931 | * call to enable it in the task context. The task might have been |
1932 | * scheduled away, but we check this in the smp call again. | |
1933 | */ | |
1934 | static void | |
cdd6c482 IM |
1935 | perf_install_in_context(struct perf_event_context *ctx, |
1936 | struct perf_event *event, | |
0793a61d TG |
1937 | int cpu) |
1938 | { | |
1939 | struct task_struct *task = ctx->task; | |
1940 | ||
fe4b04fa PZ |
1941 | lockdep_assert_held(&ctx->mutex); |
1942 | ||
c3f00c70 | 1943 | event->ctx = ctx; |
0cda4c02 YZ |
1944 | if (event->cpu != -1) |
1945 | event->cpu = cpu; | |
c3f00c70 | 1946 | |
0793a61d TG |
1947 | if (!task) { |
1948 | /* | |
cdd6c482 | 1949 | * Per cpu events are installed via an smp call and |
af901ca1 | 1950 | * the install is always successful. |
0793a61d | 1951 | */ |
fe4b04fa | 1952 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1953 | return; |
1954 | } | |
1955 | ||
0793a61d | 1956 | retry: |
fe4b04fa PZ |
1957 | if (!task_function_call(task, __perf_install_in_context, event)) |
1958 | return; | |
0793a61d | 1959 | |
e625cce1 | 1960 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1961 | /* |
fe4b04fa PZ |
1962 | * If we failed to find a running task, but find the context active now |
1963 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1964 | */ |
fe4b04fa | 1965 | if (ctx->is_active) { |
e625cce1 | 1966 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1967 | goto retry; |
1968 | } | |
1969 | ||
1970 | /* | |
fe4b04fa PZ |
1971 | * Since the task isn't running, its safe to add the event, us holding |
1972 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1973 | */ |
fe4b04fa | 1974 | add_event_to_ctx(event, ctx); |
e625cce1 | 1975 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1976 | } |
1977 | ||
fa289bec | 1978 | /* |
cdd6c482 | 1979 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1980 | * Enabling the leader of a group effectively enables all |
1981 | * the group members that aren't explicitly disabled, so we | |
1982 | * have to update their ->tstamp_enabled also. | |
1983 | * Note: this works for group members as well as group leaders | |
1984 | * since the non-leader members' sibling_lists will be empty. | |
1985 | */ | |
1d9b482e | 1986 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1987 | { |
cdd6c482 | 1988 | struct perf_event *sub; |
4158755d | 1989 | u64 tstamp = perf_event_time(event); |
fa289bec | 1990 | |
cdd6c482 | 1991 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1992 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1993 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1994 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1995 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1996 | } |
fa289bec PM |
1997 | } |
1998 | ||
d859e29f | 1999 | /* |
cdd6c482 | 2000 | * Cross CPU call to enable a performance event |
d859e29f | 2001 | */ |
fe4b04fa | 2002 | static int __perf_event_enable(void *info) |
04289bb9 | 2003 | { |
cdd6c482 | 2004 | struct perf_event *event = info; |
cdd6c482 IM |
2005 | struct perf_event_context *ctx = event->ctx; |
2006 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2007 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2008 | int err; |
04289bb9 | 2009 | |
06f41796 JO |
2010 | /* |
2011 | * There's a time window between 'ctx->is_active' check | |
2012 | * in perf_event_enable function and this place having: | |
2013 | * - IRQs on | |
2014 | * - ctx->lock unlocked | |
2015 | * | |
2016 | * where the task could be killed and 'ctx' deactivated | |
2017 | * by perf_event_exit_task. | |
2018 | */ | |
2019 | if (!ctx->is_active) | |
fe4b04fa | 2020 | return -EINVAL; |
3cbed429 | 2021 | |
e625cce1 | 2022 | raw_spin_lock(&ctx->lock); |
4af4998b | 2023 | update_context_time(ctx); |
d859e29f | 2024 | |
cdd6c482 | 2025 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2026 | goto unlock; |
e5d1367f SE |
2027 | |
2028 | /* | |
2029 | * set current task's cgroup time reference point | |
2030 | */ | |
3f7cce3c | 2031 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2032 | |
1d9b482e | 2033 | __perf_event_mark_enabled(event); |
04289bb9 | 2034 | |
e5d1367f SE |
2035 | if (!event_filter_match(event)) { |
2036 | if (is_cgroup_event(event)) | |
2037 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2038 | goto unlock; |
e5d1367f | 2039 | } |
f4c4176f | 2040 | |
04289bb9 | 2041 | /* |
cdd6c482 | 2042 | * If the event is in a group and isn't the group leader, |
d859e29f | 2043 | * then don't put it on unless the group is on. |
04289bb9 | 2044 | */ |
cdd6c482 | 2045 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2046 | goto unlock; |
3b6f9e5c | 2047 | |
cdd6c482 | 2048 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2049 | err = -EEXIST; |
e758a33d | 2050 | } else { |
cdd6c482 | 2051 | if (event == leader) |
6e37738a | 2052 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2053 | else |
6e37738a | 2054 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2055 | } |
d859e29f PM |
2056 | |
2057 | if (err) { | |
2058 | /* | |
cdd6c482 | 2059 | * If this event can't go on and it's part of a |
d859e29f PM |
2060 | * group, then the whole group has to come off. |
2061 | */ | |
9e630205 | 2062 | if (leader != event) { |
d859e29f | 2063 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2064 | perf_cpu_hrtimer_restart(cpuctx); |
2065 | } | |
0d48696f | 2066 | if (leader->attr.pinned) { |
53cfbf59 | 2067 | update_group_times(leader); |
cdd6c482 | 2068 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2069 | } |
d859e29f PM |
2070 | } |
2071 | ||
9ed6060d | 2072 | unlock: |
e625cce1 | 2073 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2074 | |
2075 | return 0; | |
d859e29f PM |
2076 | } |
2077 | ||
2078 | /* | |
cdd6c482 | 2079 | * Enable a event. |
c93f7669 | 2080 | * |
cdd6c482 IM |
2081 | * If event->ctx is a cloned context, callers must make sure that |
2082 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2083 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2084 | * perf_event_for_each_child or perf_event_for_each as described |
2085 | * for perf_event_disable. | |
d859e29f | 2086 | */ |
44234adc | 2087 | void perf_event_enable(struct perf_event *event) |
d859e29f | 2088 | { |
cdd6c482 | 2089 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2090 | struct task_struct *task = ctx->task; |
2091 | ||
2092 | if (!task) { | |
2093 | /* | |
cdd6c482 | 2094 | * Enable the event on the cpu that it's on |
d859e29f | 2095 | */ |
fe4b04fa | 2096 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2097 | return; |
2098 | } | |
2099 | ||
e625cce1 | 2100 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2101 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2102 | goto out; |
2103 | ||
2104 | /* | |
cdd6c482 IM |
2105 | * If the event is in error state, clear that first. |
2106 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2107 | * know that it has gone back into error state, as distinct |
2108 | * from the task having been scheduled away before the | |
2109 | * cross-call arrived. | |
2110 | */ | |
cdd6c482 IM |
2111 | if (event->state == PERF_EVENT_STATE_ERROR) |
2112 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2113 | |
9ed6060d | 2114 | retry: |
fe4b04fa | 2115 | if (!ctx->is_active) { |
1d9b482e | 2116 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2117 | goto out; |
2118 | } | |
2119 | ||
e625cce1 | 2120 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2121 | |
2122 | if (!task_function_call(task, __perf_event_enable, event)) | |
2123 | return; | |
d859e29f | 2124 | |
e625cce1 | 2125 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2126 | |
2127 | /* | |
cdd6c482 | 2128 | * If the context is active and the event is still off, |
d859e29f PM |
2129 | * we need to retry the cross-call. |
2130 | */ | |
fe4b04fa PZ |
2131 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2132 | /* | |
2133 | * task could have been flipped by a concurrent | |
2134 | * perf_event_context_sched_out() | |
2135 | */ | |
2136 | task = ctx->task; | |
d859e29f | 2137 | goto retry; |
fe4b04fa | 2138 | } |
fa289bec | 2139 | |
9ed6060d | 2140 | out: |
e625cce1 | 2141 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2142 | } |
dcfce4a0 | 2143 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2144 | |
26ca5c11 | 2145 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2146 | { |
2023b359 | 2147 | /* |
cdd6c482 | 2148 | * not supported on inherited events |
2023b359 | 2149 | */ |
2e939d1d | 2150 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2151 | return -EINVAL; |
2152 | ||
cdd6c482 IM |
2153 | atomic_add(refresh, &event->event_limit); |
2154 | perf_event_enable(event); | |
2023b359 PZ |
2155 | |
2156 | return 0; | |
79f14641 | 2157 | } |
26ca5c11 | 2158 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2159 | |
5b0311e1 FW |
2160 | static void ctx_sched_out(struct perf_event_context *ctx, |
2161 | struct perf_cpu_context *cpuctx, | |
2162 | enum event_type_t event_type) | |
235c7fc7 | 2163 | { |
cdd6c482 | 2164 | struct perf_event *event; |
db24d33e | 2165 | int is_active = ctx->is_active; |
235c7fc7 | 2166 | |
db24d33e | 2167 | ctx->is_active &= ~event_type; |
cdd6c482 | 2168 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2169 | return; |
2170 | ||
4af4998b | 2171 | update_context_time(ctx); |
e5d1367f | 2172 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2173 | if (!ctx->nr_active) |
facc4307 | 2174 | return; |
5b0311e1 | 2175 | |
075e0b00 | 2176 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2177 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2178 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2179 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2180 | } |
889ff015 | 2181 | |
db24d33e | 2182 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2183 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2184 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2185 | } |
1b9a644f | 2186 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2187 | } |
2188 | ||
564c2b21 | 2189 | /* |
5a3126d4 PZ |
2190 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2191 | * cloned from the same version of the same context. | |
2192 | * | |
2193 | * Equivalence is measured using a generation number in the context that is | |
2194 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2195 | * and list_del_event(). | |
564c2b21 | 2196 | */ |
cdd6c482 IM |
2197 | static int context_equiv(struct perf_event_context *ctx1, |
2198 | struct perf_event_context *ctx2) | |
564c2b21 | 2199 | { |
5a3126d4 PZ |
2200 | /* Pinning disables the swap optimization */ |
2201 | if (ctx1->pin_count || ctx2->pin_count) | |
2202 | return 0; | |
2203 | ||
2204 | /* If ctx1 is the parent of ctx2 */ | |
2205 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2206 | return 1; | |
2207 | ||
2208 | /* If ctx2 is the parent of ctx1 */ | |
2209 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2210 | return 1; | |
2211 | ||
2212 | /* | |
2213 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2214 | * hierarchy, see perf_event_init_context(). | |
2215 | */ | |
2216 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2217 | ctx1->parent_gen == ctx2->parent_gen) | |
2218 | return 1; | |
2219 | ||
2220 | /* Unmatched */ | |
2221 | return 0; | |
564c2b21 PM |
2222 | } |
2223 | ||
cdd6c482 IM |
2224 | static void __perf_event_sync_stat(struct perf_event *event, |
2225 | struct perf_event *next_event) | |
bfbd3381 PZ |
2226 | { |
2227 | u64 value; | |
2228 | ||
cdd6c482 | 2229 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2230 | return; |
2231 | ||
2232 | /* | |
cdd6c482 | 2233 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2234 | * because we're in the middle of a context switch and have IRQs |
2235 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2236 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2237 | * don't need to use it. |
2238 | */ | |
cdd6c482 IM |
2239 | switch (event->state) { |
2240 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2241 | event->pmu->read(event); |
2242 | /* fall-through */ | |
bfbd3381 | 2243 | |
cdd6c482 IM |
2244 | case PERF_EVENT_STATE_INACTIVE: |
2245 | update_event_times(event); | |
bfbd3381 PZ |
2246 | break; |
2247 | ||
2248 | default: | |
2249 | break; | |
2250 | } | |
2251 | ||
2252 | /* | |
cdd6c482 | 2253 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2254 | * values when we flip the contexts. |
2255 | */ | |
e7850595 PZ |
2256 | value = local64_read(&next_event->count); |
2257 | value = local64_xchg(&event->count, value); | |
2258 | local64_set(&next_event->count, value); | |
bfbd3381 | 2259 | |
cdd6c482 IM |
2260 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2261 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2262 | |
bfbd3381 | 2263 | /* |
19d2e755 | 2264 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2265 | */ |
cdd6c482 IM |
2266 | perf_event_update_userpage(event); |
2267 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2268 | } |
2269 | ||
cdd6c482 IM |
2270 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2271 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2272 | { |
cdd6c482 | 2273 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2274 | |
2275 | if (!ctx->nr_stat) | |
2276 | return; | |
2277 | ||
02ffdbc8 PZ |
2278 | update_context_time(ctx); |
2279 | ||
cdd6c482 IM |
2280 | event = list_first_entry(&ctx->event_list, |
2281 | struct perf_event, event_entry); | |
bfbd3381 | 2282 | |
cdd6c482 IM |
2283 | next_event = list_first_entry(&next_ctx->event_list, |
2284 | struct perf_event, event_entry); | |
bfbd3381 | 2285 | |
cdd6c482 IM |
2286 | while (&event->event_entry != &ctx->event_list && |
2287 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2288 | |
cdd6c482 | 2289 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2290 | |
cdd6c482 IM |
2291 | event = list_next_entry(event, event_entry); |
2292 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2293 | } |
2294 | } | |
2295 | ||
fe4b04fa PZ |
2296 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2297 | struct task_struct *next) | |
0793a61d | 2298 | { |
8dc85d54 | 2299 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2300 | struct perf_event_context *next_ctx; |
5a3126d4 | 2301 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2302 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2303 | int do_switch = 1; |
0793a61d | 2304 | |
108b02cf PZ |
2305 | if (likely(!ctx)) |
2306 | return; | |
10989fb2 | 2307 | |
108b02cf PZ |
2308 | cpuctx = __get_cpu_context(ctx); |
2309 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2310 | return; |
2311 | ||
c93f7669 | 2312 | rcu_read_lock(); |
8dc85d54 | 2313 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2314 | if (!next_ctx) |
2315 | goto unlock; | |
2316 | ||
2317 | parent = rcu_dereference(ctx->parent_ctx); | |
2318 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2319 | ||
2320 | /* If neither context have a parent context; they cannot be clones. */ | |
2321 | if (!parent && !next_parent) | |
2322 | goto unlock; | |
2323 | ||
2324 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2325 | /* |
2326 | * Looks like the two contexts are clones, so we might be | |
2327 | * able to optimize the context switch. We lock both | |
2328 | * contexts and check that they are clones under the | |
2329 | * lock (including re-checking that neither has been | |
2330 | * uncloned in the meantime). It doesn't matter which | |
2331 | * order we take the locks because no other cpu could | |
2332 | * be trying to lock both of these tasks. | |
2333 | */ | |
e625cce1 TG |
2334 | raw_spin_lock(&ctx->lock); |
2335 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2336 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2337 | /* |
2338 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2339 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2340 | */ |
8dc85d54 PZ |
2341 | task->perf_event_ctxp[ctxn] = next_ctx; |
2342 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2343 | ctx->task = next; |
2344 | next_ctx->task = task; | |
2345 | do_switch = 0; | |
bfbd3381 | 2346 | |
cdd6c482 | 2347 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2348 | } |
e625cce1 TG |
2349 | raw_spin_unlock(&next_ctx->lock); |
2350 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2351 | } |
5a3126d4 | 2352 | unlock: |
c93f7669 | 2353 | rcu_read_unlock(); |
564c2b21 | 2354 | |
c93f7669 | 2355 | if (do_switch) { |
facc4307 | 2356 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2357 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2358 | cpuctx->task_ctx = NULL; |
facc4307 | 2359 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2360 | } |
0793a61d TG |
2361 | } |
2362 | ||
8dc85d54 PZ |
2363 | #define for_each_task_context_nr(ctxn) \ |
2364 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2365 | ||
2366 | /* | |
2367 | * Called from scheduler to remove the events of the current task, | |
2368 | * with interrupts disabled. | |
2369 | * | |
2370 | * We stop each event and update the event value in event->count. | |
2371 | * | |
2372 | * This does not protect us against NMI, but disable() | |
2373 | * sets the disabled bit in the control field of event _before_ | |
2374 | * accessing the event control register. If a NMI hits, then it will | |
2375 | * not restart the event. | |
2376 | */ | |
ab0cce56 JO |
2377 | void __perf_event_task_sched_out(struct task_struct *task, |
2378 | struct task_struct *next) | |
8dc85d54 PZ |
2379 | { |
2380 | int ctxn; | |
2381 | ||
8dc85d54 PZ |
2382 | for_each_task_context_nr(ctxn) |
2383 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2384 | |
2385 | /* | |
2386 | * if cgroup events exist on this CPU, then we need | |
2387 | * to check if we have to switch out PMU state. | |
2388 | * cgroup event are system-wide mode only | |
2389 | */ | |
2390 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2391 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2392 | } |
2393 | ||
04dc2dbb | 2394 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2395 | { |
108b02cf | 2396 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2397 | |
a63eaf34 PM |
2398 | if (!cpuctx->task_ctx) |
2399 | return; | |
012b84da IM |
2400 | |
2401 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2402 | return; | |
2403 | ||
04dc2dbb | 2404 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2405 | cpuctx->task_ctx = NULL; |
2406 | } | |
2407 | ||
5b0311e1 FW |
2408 | /* |
2409 | * Called with IRQs disabled | |
2410 | */ | |
2411 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2412 | enum event_type_t event_type) | |
2413 | { | |
2414 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2415 | } |
2416 | ||
235c7fc7 | 2417 | static void |
5b0311e1 | 2418 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2419 | struct perf_cpu_context *cpuctx) |
0793a61d | 2420 | { |
cdd6c482 | 2421 | struct perf_event *event; |
0793a61d | 2422 | |
889ff015 FW |
2423 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2424 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2425 | continue; |
5632ab12 | 2426 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2427 | continue; |
2428 | ||
e5d1367f SE |
2429 | /* may need to reset tstamp_enabled */ |
2430 | if (is_cgroup_event(event)) | |
2431 | perf_cgroup_mark_enabled(event, ctx); | |
2432 | ||
8c9ed8e1 | 2433 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2434 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2435 | |
2436 | /* | |
2437 | * If this pinned group hasn't been scheduled, | |
2438 | * put it in error state. | |
2439 | */ | |
cdd6c482 IM |
2440 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2441 | update_group_times(event); | |
2442 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2443 | } |
3b6f9e5c | 2444 | } |
5b0311e1 FW |
2445 | } |
2446 | ||
2447 | static void | |
2448 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2449 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2450 | { |
2451 | struct perf_event *event; | |
2452 | int can_add_hw = 1; | |
3b6f9e5c | 2453 | |
889ff015 FW |
2454 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2455 | /* Ignore events in OFF or ERROR state */ | |
2456 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2457 | continue; |
04289bb9 IM |
2458 | /* |
2459 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2460 | * of events: |
04289bb9 | 2461 | */ |
5632ab12 | 2462 | if (!event_filter_match(event)) |
0793a61d TG |
2463 | continue; |
2464 | ||
e5d1367f SE |
2465 | /* may need to reset tstamp_enabled */ |
2466 | if (is_cgroup_event(event)) | |
2467 | perf_cgroup_mark_enabled(event, ctx); | |
2468 | ||
9ed6060d | 2469 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2470 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2471 | can_add_hw = 0; |
9ed6060d | 2472 | } |
0793a61d | 2473 | } |
5b0311e1 FW |
2474 | } |
2475 | ||
2476 | static void | |
2477 | ctx_sched_in(struct perf_event_context *ctx, | |
2478 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2479 | enum event_type_t event_type, |
2480 | struct task_struct *task) | |
5b0311e1 | 2481 | { |
e5d1367f | 2482 | u64 now; |
db24d33e | 2483 | int is_active = ctx->is_active; |
e5d1367f | 2484 | |
db24d33e | 2485 | ctx->is_active |= event_type; |
5b0311e1 | 2486 | if (likely(!ctx->nr_events)) |
facc4307 | 2487 | return; |
5b0311e1 | 2488 | |
e5d1367f SE |
2489 | now = perf_clock(); |
2490 | ctx->timestamp = now; | |
3f7cce3c | 2491 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2492 | /* |
2493 | * First go through the list and put on any pinned groups | |
2494 | * in order to give them the best chance of going on. | |
2495 | */ | |
db24d33e | 2496 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2497 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2498 | |
2499 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2500 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2501 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2502 | } |
2503 | ||
329c0e01 | 2504 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2505 | enum event_type_t event_type, |
2506 | struct task_struct *task) | |
329c0e01 FW |
2507 | { |
2508 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2509 | ||
e5d1367f | 2510 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2511 | } |
2512 | ||
e5d1367f SE |
2513 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2514 | struct task_struct *task) | |
235c7fc7 | 2515 | { |
108b02cf | 2516 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2517 | |
108b02cf | 2518 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2519 | if (cpuctx->task_ctx == ctx) |
2520 | return; | |
2521 | ||
facc4307 | 2522 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2523 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2524 | /* |
2525 | * We want to keep the following priority order: | |
2526 | * cpu pinned (that don't need to move), task pinned, | |
2527 | * cpu flexible, task flexible. | |
2528 | */ | |
2529 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2530 | ||
1d5f003f GN |
2531 | if (ctx->nr_events) |
2532 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2533 | |
86b47c25 GN |
2534 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2535 | ||
facc4307 PZ |
2536 | perf_pmu_enable(ctx->pmu); |
2537 | perf_ctx_unlock(cpuctx, ctx); | |
2538 | ||
b5ab4cd5 PZ |
2539 | /* |
2540 | * Since these rotations are per-cpu, we need to ensure the | |
2541 | * cpu-context we got scheduled on is actually rotating. | |
2542 | */ | |
108b02cf | 2543 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2544 | } |
2545 | ||
d010b332 SE |
2546 | /* |
2547 | * When sampling the branck stack in system-wide, it may be necessary | |
2548 | * to flush the stack on context switch. This happens when the branch | |
2549 | * stack does not tag its entries with the pid of the current task. | |
2550 | * Otherwise it becomes impossible to associate a branch entry with a | |
2551 | * task. This ambiguity is more likely to appear when the branch stack | |
2552 | * supports priv level filtering and the user sets it to monitor only | |
2553 | * at the user level (which could be a useful measurement in system-wide | |
2554 | * mode). In that case, the risk is high of having a branch stack with | |
2555 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2556 | * entries or stashing them somewhere in the PMU specific code layer. | |
2557 | * | |
2558 | * This function provides the context switch callback to the lower code | |
2559 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2560 | * with at least one active event using taken branch sampling. | |
2561 | */ | |
2562 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2563 | struct task_struct *task) | |
2564 | { | |
2565 | struct perf_cpu_context *cpuctx; | |
2566 | struct pmu *pmu; | |
2567 | unsigned long flags; | |
2568 | ||
2569 | /* no need to flush branch stack if not changing task */ | |
2570 | if (prev == task) | |
2571 | return; | |
2572 | ||
2573 | local_irq_save(flags); | |
2574 | ||
2575 | rcu_read_lock(); | |
2576 | ||
2577 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2578 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2579 | ||
2580 | /* | |
2581 | * check if the context has at least one | |
2582 | * event using PERF_SAMPLE_BRANCH_STACK | |
2583 | */ | |
2584 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2585 | && pmu->flush_branch_stack) { | |
2586 | ||
d010b332 SE |
2587 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
2588 | ||
2589 | perf_pmu_disable(pmu); | |
2590 | ||
2591 | pmu->flush_branch_stack(); | |
2592 | ||
2593 | perf_pmu_enable(pmu); | |
2594 | ||
2595 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2596 | } | |
2597 | } | |
2598 | ||
2599 | rcu_read_unlock(); | |
2600 | ||
2601 | local_irq_restore(flags); | |
2602 | } | |
2603 | ||
8dc85d54 PZ |
2604 | /* |
2605 | * Called from scheduler to add the events of the current task | |
2606 | * with interrupts disabled. | |
2607 | * | |
2608 | * We restore the event value and then enable it. | |
2609 | * | |
2610 | * This does not protect us against NMI, but enable() | |
2611 | * sets the enabled bit in the control field of event _before_ | |
2612 | * accessing the event control register. If a NMI hits, then it will | |
2613 | * keep the event running. | |
2614 | */ | |
ab0cce56 JO |
2615 | void __perf_event_task_sched_in(struct task_struct *prev, |
2616 | struct task_struct *task) | |
8dc85d54 PZ |
2617 | { |
2618 | struct perf_event_context *ctx; | |
2619 | int ctxn; | |
2620 | ||
2621 | for_each_task_context_nr(ctxn) { | |
2622 | ctx = task->perf_event_ctxp[ctxn]; | |
2623 | if (likely(!ctx)) | |
2624 | continue; | |
2625 | ||
e5d1367f | 2626 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2627 | } |
e5d1367f SE |
2628 | /* |
2629 | * if cgroup events exist on this CPU, then we need | |
2630 | * to check if we have to switch in PMU state. | |
2631 | * cgroup event are system-wide mode only | |
2632 | */ | |
2633 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2634 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2635 | |
2636 | /* check for system-wide branch_stack events */ | |
2637 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2638 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2639 | } |
2640 | ||
abd50713 PZ |
2641 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2642 | { | |
2643 | u64 frequency = event->attr.sample_freq; | |
2644 | u64 sec = NSEC_PER_SEC; | |
2645 | u64 divisor, dividend; | |
2646 | ||
2647 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2648 | ||
2649 | count_fls = fls64(count); | |
2650 | nsec_fls = fls64(nsec); | |
2651 | frequency_fls = fls64(frequency); | |
2652 | sec_fls = 30; | |
2653 | ||
2654 | /* | |
2655 | * We got @count in @nsec, with a target of sample_freq HZ | |
2656 | * the target period becomes: | |
2657 | * | |
2658 | * @count * 10^9 | |
2659 | * period = ------------------- | |
2660 | * @nsec * sample_freq | |
2661 | * | |
2662 | */ | |
2663 | ||
2664 | /* | |
2665 | * Reduce accuracy by one bit such that @a and @b converge | |
2666 | * to a similar magnitude. | |
2667 | */ | |
fe4b04fa | 2668 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2669 | do { \ |
2670 | if (a##_fls > b##_fls) { \ | |
2671 | a >>= 1; \ | |
2672 | a##_fls--; \ | |
2673 | } else { \ | |
2674 | b >>= 1; \ | |
2675 | b##_fls--; \ | |
2676 | } \ | |
2677 | } while (0) | |
2678 | ||
2679 | /* | |
2680 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2681 | * the other, so that finally we can do a u64/u64 division. | |
2682 | */ | |
2683 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2684 | REDUCE_FLS(nsec, frequency); | |
2685 | REDUCE_FLS(sec, count); | |
2686 | } | |
2687 | ||
2688 | if (count_fls + sec_fls > 64) { | |
2689 | divisor = nsec * frequency; | |
2690 | ||
2691 | while (count_fls + sec_fls > 64) { | |
2692 | REDUCE_FLS(count, sec); | |
2693 | divisor >>= 1; | |
2694 | } | |
2695 | ||
2696 | dividend = count * sec; | |
2697 | } else { | |
2698 | dividend = count * sec; | |
2699 | ||
2700 | while (nsec_fls + frequency_fls > 64) { | |
2701 | REDUCE_FLS(nsec, frequency); | |
2702 | dividend >>= 1; | |
2703 | } | |
2704 | ||
2705 | divisor = nsec * frequency; | |
2706 | } | |
2707 | ||
f6ab91ad PZ |
2708 | if (!divisor) |
2709 | return dividend; | |
2710 | ||
abd50713 PZ |
2711 | return div64_u64(dividend, divisor); |
2712 | } | |
2713 | ||
e050e3f0 SE |
2714 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2715 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2716 | ||
f39d47ff | 2717 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2718 | { |
cdd6c482 | 2719 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2720 | s64 period, sample_period; |
bd2b5b12 PZ |
2721 | s64 delta; |
2722 | ||
abd50713 | 2723 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2724 | |
2725 | delta = (s64)(period - hwc->sample_period); | |
2726 | delta = (delta + 7) / 8; /* low pass filter */ | |
2727 | ||
2728 | sample_period = hwc->sample_period + delta; | |
2729 | ||
2730 | if (!sample_period) | |
2731 | sample_period = 1; | |
2732 | ||
bd2b5b12 | 2733 | hwc->sample_period = sample_period; |
abd50713 | 2734 | |
e7850595 | 2735 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2736 | if (disable) |
2737 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2738 | ||
e7850595 | 2739 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2740 | |
2741 | if (disable) | |
2742 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2743 | } |
bd2b5b12 PZ |
2744 | } |
2745 | ||
e050e3f0 SE |
2746 | /* |
2747 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2748 | * events. At the same time, make sure, having freq events does not change | |
2749 | * the rate of unthrottling as that would introduce bias. | |
2750 | */ | |
2751 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2752 | int needs_unthr) | |
60db5e09 | 2753 | { |
cdd6c482 IM |
2754 | struct perf_event *event; |
2755 | struct hw_perf_event *hwc; | |
e050e3f0 | 2756 | u64 now, period = TICK_NSEC; |
abd50713 | 2757 | s64 delta; |
60db5e09 | 2758 | |
e050e3f0 SE |
2759 | /* |
2760 | * only need to iterate over all events iff: | |
2761 | * - context have events in frequency mode (needs freq adjust) | |
2762 | * - there are events to unthrottle on this cpu | |
2763 | */ | |
2764 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2765 | return; |
2766 | ||
e050e3f0 | 2767 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2768 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2769 | |
03541f8b | 2770 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2771 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2772 | continue; |
2773 | ||
5632ab12 | 2774 | if (!event_filter_match(event)) |
5d27c23d PZ |
2775 | continue; |
2776 | ||
44377277 AS |
2777 | perf_pmu_disable(event->pmu); |
2778 | ||
cdd6c482 | 2779 | hwc = &event->hw; |
6a24ed6c | 2780 | |
ae23bff1 | 2781 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2782 | hwc->interrupts = 0; |
cdd6c482 | 2783 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2784 | event->pmu->start(event, 0); |
a78ac325 PZ |
2785 | } |
2786 | ||
cdd6c482 | 2787 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2788 | goto next; |
60db5e09 | 2789 | |
e050e3f0 SE |
2790 | /* |
2791 | * stop the event and update event->count | |
2792 | */ | |
2793 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2794 | ||
e7850595 | 2795 | now = local64_read(&event->count); |
abd50713 PZ |
2796 | delta = now - hwc->freq_count_stamp; |
2797 | hwc->freq_count_stamp = now; | |
60db5e09 | 2798 | |
e050e3f0 SE |
2799 | /* |
2800 | * restart the event | |
2801 | * reload only if value has changed | |
f39d47ff SE |
2802 | * we have stopped the event so tell that |
2803 | * to perf_adjust_period() to avoid stopping it | |
2804 | * twice. | |
e050e3f0 | 2805 | */ |
abd50713 | 2806 | if (delta > 0) |
f39d47ff | 2807 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2808 | |
2809 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
2810 | next: |
2811 | perf_pmu_enable(event->pmu); | |
60db5e09 | 2812 | } |
e050e3f0 | 2813 | |
f39d47ff | 2814 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2815 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2816 | } |
2817 | ||
235c7fc7 | 2818 | /* |
cdd6c482 | 2819 | * Round-robin a context's events: |
235c7fc7 | 2820 | */ |
cdd6c482 | 2821 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2822 | { |
dddd3379 TG |
2823 | /* |
2824 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2825 | * disabled by the inheritance code. | |
2826 | */ | |
2827 | if (!ctx->rotate_disable) | |
2828 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2829 | } |
2830 | ||
b5ab4cd5 | 2831 | /* |
e9d2b064 PZ |
2832 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2833 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2834 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2835 | */ |
9e630205 | 2836 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2837 | { |
8dc85d54 | 2838 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2839 | int rotate = 0, remove = 1; |
7fc23a53 | 2840 | |
b5ab4cd5 | 2841 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2842 | remove = 0; |
b5ab4cd5 PZ |
2843 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2844 | rotate = 1; | |
2845 | } | |
235c7fc7 | 2846 | |
8dc85d54 | 2847 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2848 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2849 | remove = 0; |
b5ab4cd5 PZ |
2850 | if (ctx->nr_events != ctx->nr_active) |
2851 | rotate = 1; | |
2852 | } | |
9717e6cd | 2853 | |
e050e3f0 | 2854 | if (!rotate) |
0f5a2601 PZ |
2855 | goto done; |
2856 | ||
facc4307 | 2857 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2858 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2859 | |
e050e3f0 SE |
2860 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2861 | if (ctx) | |
2862 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2863 | |
e050e3f0 SE |
2864 | rotate_ctx(&cpuctx->ctx); |
2865 | if (ctx) | |
2866 | rotate_ctx(ctx); | |
235c7fc7 | 2867 | |
e050e3f0 | 2868 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2869 | |
0f5a2601 PZ |
2870 | perf_pmu_enable(cpuctx->ctx.pmu); |
2871 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2872 | done: |
e9d2b064 PZ |
2873 | if (remove) |
2874 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2875 | |
2876 | return rotate; | |
e9d2b064 PZ |
2877 | } |
2878 | ||
026249ef FW |
2879 | #ifdef CONFIG_NO_HZ_FULL |
2880 | bool perf_event_can_stop_tick(void) | |
2881 | { | |
948b26b6 | 2882 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 2883 | __this_cpu_read(perf_throttled_count)) |
026249ef | 2884 | return false; |
d84153d6 FW |
2885 | else |
2886 | return true; | |
026249ef FW |
2887 | } |
2888 | #endif | |
2889 | ||
e9d2b064 PZ |
2890 | void perf_event_task_tick(void) |
2891 | { | |
2892 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2893 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2894 | struct perf_event_context *ctx; |
2895 | int throttled; | |
b5ab4cd5 | 2896 | |
e9d2b064 PZ |
2897 | WARN_ON(!irqs_disabled()); |
2898 | ||
e050e3f0 SE |
2899 | __this_cpu_inc(perf_throttled_seq); |
2900 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2901 | ||
e9d2b064 | 2902 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2903 | ctx = &cpuctx->ctx; |
2904 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2905 | ||
2906 | ctx = cpuctx->task_ctx; | |
2907 | if (ctx) | |
2908 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2909 | } |
0793a61d TG |
2910 | } |
2911 | ||
889ff015 FW |
2912 | static int event_enable_on_exec(struct perf_event *event, |
2913 | struct perf_event_context *ctx) | |
2914 | { | |
2915 | if (!event->attr.enable_on_exec) | |
2916 | return 0; | |
2917 | ||
2918 | event->attr.enable_on_exec = 0; | |
2919 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2920 | return 0; | |
2921 | ||
1d9b482e | 2922 | __perf_event_mark_enabled(event); |
889ff015 FW |
2923 | |
2924 | return 1; | |
2925 | } | |
2926 | ||
57e7986e | 2927 | /* |
cdd6c482 | 2928 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2929 | * This expects task == current. |
2930 | */ | |
8dc85d54 | 2931 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2932 | { |
cdd6c482 | 2933 | struct perf_event *event; |
57e7986e PM |
2934 | unsigned long flags; |
2935 | int enabled = 0; | |
889ff015 | 2936 | int ret; |
57e7986e PM |
2937 | |
2938 | local_irq_save(flags); | |
cdd6c482 | 2939 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2940 | goto out; |
2941 | ||
e566b76e SE |
2942 | /* |
2943 | * We must ctxsw out cgroup events to avoid conflict | |
2944 | * when invoking perf_task_event_sched_in() later on | |
2945 | * in this function. Otherwise we end up trying to | |
2946 | * ctxswin cgroup events which are already scheduled | |
2947 | * in. | |
2948 | */ | |
a8d757ef | 2949 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2950 | |
e625cce1 | 2951 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2952 | task_ctx_sched_out(ctx); |
57e7986e | 2953 | |
b79387ef | 2954 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2955 | ret = event_enable_on_exec(event, ctx); |
2956 | if (ret) | |
2957 | enabled = 1; | |
57e7986e PM |
2958 | } |
2959 | ||
2960 | /* | |
cdd6c482 | 2961 | * Unclone this context if we enabled any event. |
57e7986e | 2962 | */ |
71a851b4 PZ |
2963 | if (enabled) |
2964 | unclone_ctx(ctx); | |
57e7986e | 2965 | |
e625cce1 | 2966 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2967 | |
e566b76e SE |
2968 | /* |
2969 | * Also calls ctxswin for cgroup events, if any: | |
2970 | */ | |
e5d1367f | 2971 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2972 | out: |
57e7986e PM |
2973 | local_irq_restore(flags); |
2974 | } | |
2975 | ||
0793a61d | 2976 | /* |
cdd6c482 | 2977 | * Cross CPU call to read the hardware event |
0793a61d | 2978 | */ |
cdd6c482 | 2979 | static void __perf_event_read(void *info) |
0793a61d | 2980 | { |
cdd6c482 IM |
2981 | struct perf_event *event = info; |
2982 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2983 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2984 | |
e1ac3614 PM |
2985 | /* |
2986 | * If this is a task context, we need to check whether it is | |
2987 | * the current task context of this cpu. If not it has been | |
2988 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2989 | * event->count would have been updated to a recent sample |
2990 | * when the event was scheduled out. | |
e1ac3614 PM |
2991 | */ |
2992 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2993 | return; | |
2994 | ||
e625cce1 | 2995 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2996 | if (ctx->is_active) { |
542e72fc | 2997 | update_context_time(ctx); |
e5d1367f SE |
2998 | update_cgrp_time_from_event(event); |
2999 | } | |
cdd6c482 | 3000 | update_event_times(event); |
542e72fc PZ |
3001 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3002 | event->pmu->read(event); | |
e625cce1 | 3003 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3004 | } |
3005 | ||
b5e58793 PZ |
3006 | static inline u64 perf_event_count(struct perf_event *event) |
3007 | { | |
e7850595 | 3008 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
3009 | } |
3010 | ||
cdd6c482 | 3011 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3012 | { |
3013 | /* | |
cdd6c482 IM |
3014 | * If event is enabled and currently active on a CPU, update the |
3015 | * value in the event structure: | |
0793a61d | 3016 | */ |
cdd6c482 IM |
3017 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3018 | smp_call_function_single(event->oncpu, | |
3019 | __perf_event_read, event, 1); | |
3020 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3021 | struct perf_event_context *ctx = event->ctx; |
3022 | unsigned long flags; | |
3023 | ||
e625cce1 | 3024 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3025 | /* |
3026 | * may read while context is not active | |
3027 | * (e.g., thread is blocked), in that case | |
3028 | * we cannot update context time | |
3029 | */ | |
e5d1367f | 3030 | if (ctx->is_active) { |
c530ccd9 | 3031 | update_context_time(ctx); |
e5d1367f SE |
3032 | update_cgrp_time_from_event(event); |
3033 | } | |
cdd6c482 | 3034 | update_event_times(event); |
e625cce1 | 3035 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3036 | } |
3037 | ||
b5e58793 | 3038 | return perf_event_count(event); |
0793a61d TG |
3039 | } |
3040 | ||
a63eaf34 | 3041 | /* |
cdd6c482 | 3042 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3043 | */ |
eb184479 | 3044 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3045 | { |
e625cce1 | 3046 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3047 | mutex_init(&ctx->mutex); |
889ff015 FW |
3048 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3049 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3050 | INIT_LIST_HEAD(&ctx->event_list); |
3051 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
3052 | } |
3053 | ||
3054 | static struct perf_event_context * | |
3055 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3056 | { | |
3057 | struct perf_event_context *ctx; | |
3058 | ||
3059 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3060 | if (!ctx) | |
3061 | return NULL; | |
3062 | ||
3063 | __perf_event_init_context(ctx); | |
3064 | if (task) { | |
3065 | ctx->task = task; | |
3066 | get_task_struct(task); | |
0793a61d | 3067 | } |
eb184479 PZ |
3068 | ctx->pmu = pmu; |
3069 | ||
3070 | return ctx; | |
a63eaf34 PM |
3071 | } |
3072 | ||
2ebd4ffb MH |
3073 | static struct task_struct * |
3074 | find_lively_task_by_vpid(pid_t vpid) | |
3075 | { | |
3076 | struct task_struct *task; | |
3077 | int err; | |
0793a61d TG |
3078 | |
3079 | rcu_read_lock(); | |
2ebd4ffb | 3080 | if (!vpid) |
0793a61d TG |
3081 | task = current; |
3082 | else | |
2ebd4ffb | 3083 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3084 | if (task) |
3085 | get_task_struct(task); | |
3086 | rcu_read_unlock(); | |
3087 | ||
3088 | if (!task) | |
3089 | return ERR_PTR(-ESRCH); | |
3090 | ||
0793a61d | 3091 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3092 | err = -EACCES; |
3093 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3094 | goto errout; | |
3095 | ||
2ebd4ffb MH |
3096 | return task; |
3097 | errout: | |
3098 | put_task_struct(task); | |
3099 | return ERR_PTR(err); | |
3100 | ||
3101 | } | |
3102 | ||
fe4b04fa PZ |
3103 | /* |
3104 | * Returns a matching context with refcount and pincount. | |
3105 | */ | |
108b02cf | 3106 | static struct perf_event_context * |
38a81da2 | 3107 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3108 | { |
cdd6c482 | 3109 | struct perf_event_context *ctx; |
22a4f650 | 3110 | struct perf_cpu_context *cpuctx; |
25346b93 | 3111 | unsigned long flags; |
8dc85d54 | 3112 | int ctxn, err; |
0793a61d | 3113 | |
22a4ec72 | 3114 | if (!task) { |
cdd6c482 | 3115 | /* Must be root to operate on a CPU event: */ |
0764771d | 3116 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3117 | return ERR_PTR(-EACCES); |
3118 | ||
0793a61d | 3119 | /* |
cdd6c482 | 3120 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3121 | * offline CPU and activate it when the CPU comes up, but |
3122 | * that's for later. | |
3123 | */ | |
f6325e30 | 3124 | if (!cpu_online(cpu)) |
0793a61d TG |
3125 | return ERR_PTR(-ENODEV); |
3126 | ||
108b02cf | 3127 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3128 | ctx = &cpuctx->ctx; |
c93f7669 | 3129 | get_ctx(ctx); |
fe4b04fa | 3130 | ++ctx->pin_count; |
0793a61d | 3131 | |
0793a61d TG |
3132 | return ctx; |
3133 | } | |
3134 | ||
8dc85d54 PZ |
3135 | err = -EINVAL; |
3136 | ctxn = pmu->task_ctx_nr; | |
3137 | if (ctxn < 0) | |
3138 | goto errout; | |
3139 | ||
9ed6060d | 3140 | retry: |
8dc85d54 | 3141 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3142 | if (ctx) { |
71a851b4 | 3143 | unclone_ctx(ctx); |
fe4b04fa | 3144 | ++ctx->pin_count; |
e625cce1 | 3145 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3146 | } else { |
eb184479 | 3147 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3148 | err = -ENOMEM; |
3149 | if (!ctx) | |
3150 | goto errout; | |
eb184479 | 3151 | |
dbe08d82 ON |
3152 | err = 0; |
3153 | mutex_lock(&task->perf_event_mutex); | |
3154 | /* | |
3155 | * If it has already passed perf_event_exit_task(). | |
3156 | * we must see PF_EXITING, it takes this mutex too. | |
3157 | */ | |
3158 | if (task->flags & PF_EXITING) | |
3159 | err = -ESRCH; | |
3160 | else if (task->perf_event_ctxp[ctxn]) | |
3161 | err = -EAGAIN; | |
fe4b04fa | 3162 | else { |
9137fb28 | 3163 | get_ctx(ctx); |
fe4b04fa | 3164 | ++ctx->pin_count; |
dbe08d82 | 3165 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3166 | } |
dbe08d82 ON |
3167 | mutex_unlock(&task->perf_event_mutex); |
3168 | ||
3169 | if (unlikely(err)) { | |
9137fb28 | 3170 | put_ctx(ctx); |
dbe08d82 ON |
3171 | |
3172 | if (err == -EAGAIN) | |
3173 | goto retry; | |
3174 | goto errout; | |
a63eaf34 PM |
3175 | } |
3176 | } | |
3177 | ||
0793a61d | 3178 | return ctx; |
c93f7669 | 3179 | |
9ed6060d | 3180 | errout: |
c93f7669 | 3181 | return ERR_PTR(err); |
0793a61d TG |
3182 | } |
3183 | ||
6fb2915d LZ |
3184 | static void perf_event_free_filter(struct perf_event *event); |
3185 | ||
cdd6c482 | 3186 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3187 | { |
cdd6c482 | 3188 | struct perf_event *event; |
592903cd | 3189 | |
cdd6c482 IM |
3190 | event = container_of(head, struct perf_event, rcu_head); |
3191 | if (event->ns) | |
3192 | put_pid_ns(event->ns); | |
6fb2915d | 3193 | perf_event_free_filter(event); |
cdd6c482 | 3194 | kfree(event); |
592903cd PZ |
3195 | } |
3196 | ||
76369139 | 3197 | static void ring_buffer_put(struct ring_buffer *rb); |
9bb5d40c | 3198 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb); |
925d519a | 3199 | |
4beb31f3 | 3200 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3201 | { |
4beb31f3 FW |
3202 | if (event->parent) |
3203 | return; | |
3204 | ||
3205 | if (has_branch_stack(event)) { | |
3206 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3207 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3208 | } | |
3209 | if (is_cgroup_event(event)) | |
3210 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3211 | } | |
925d519a | 3212 | |
4beb31f3 FW |
3213 | static void unaccount_event(struct perf_event *event) |
3214 | { | |
3215 | if (event->parent) | |
3216 | return; | |
3217 | ||
3218 | if (event->attach_state & PERF_ATTACH_TASK) | |
3219 | static_key_slow_dec_deferred(&perf_sched_events); | |
3220 | if (event->attr.mmap || event->attr.mmap_data) | |
3221 | atomic_dec(&nr_mmap_events); | |
3222 | if (event->attr.comm) | |
3223 | atomic_dec(&nr_comm_events); | |
3224 | if (event->attr.task) | |
3225 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3226 | if (event->attr.freq) |
3227 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3228 | if (is_cgroup_event(event)) |
3229 | static_key_slow_dec_deferred(&perf_sched_events); | |
3230 | if (has_branch_stack(event)) | |
3231 | static_key_slow_dec_deferred(&perf_sched_events); | |
3232 | ||
3233 | unaccount_event_cpu(event, event->cpu); | |
3234 | } | |
925d519a | 3235 | |
766d6c07 FW |
3236 | static void __free_event(struct perf_event *event) |
3237 | { | |
cdd6c482 | 3238 | if (!event->parent) { |
927c7a9e FW |
3239 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3240 | put_callchain_buffers(); | |
f344011c | 3241 | } |
9ee318a7 | 3242 | |
766d6c07 FW |
3243 | if (event->destroy) |
3244 | event->destroy(event); | |
3245 | ||
3246 | if (event->ctx) | |
3247 | put_ctx(event->ctx); | |
3248 | ||
c464c76e YZ |
3249 | if (event->pmu) |
3250 | module_put(event->pmu->module); | |
3251 | ||
766d6c07 FW |
3252 | call_rcu(&event->rcu_head, free_event_rcu); |
3253 | } | |
cdd6c482 | 3254 | static void free_event(struct perf_event *event) |
f1600952 | 3255 | { |
e360adbe | 3256 | irq_work_sync(&event->pending); |
925d519a | 3257 | |
4beb31f3 | 3258 | unaccount_event(event); |
9ee318a7 | 3259 | |
76369139 | 3260 | if (event->rb) { |
9bb5d40c PZ |
3261 | struct ring_buffer *rb; |
3262 | ||
3263 | /* | |
3264 | * Can happen when we close an event with re-directed output. | |
3265 | * | |
3266 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3267 | * over us; possibly making our ring_buffer_put() the last. | |
3268 | */ | |
3269 | mutex_lock(&event->mmap_mutex); | |
3270 | rb = event->rb; | |
3271 | if (rb) { | |
3272 | rcu_assign_pointer(event->rb, NULL); | |
3273 | ring_buffer_detach(event, rb); | |
3274 | ring_buffer_put(rb); /* could be last */ | |
3275 | } | |
3276 | mutex_unlock(&event->mmap_mutex); | |
a4be7c27 PZ |
3277 | } |
3278 | ||
e5d1367f SE |
3279 | if (is_cgroup_event(event)) |
3280 | perf_detach_cgroup(event); | |
3281 | ||
0c67b408 | 3282 | |
766d6c07 | 3283 | __free_event(event); |
f1600952 PZ |
3284 | } |
3285 | ||
a66a3052 | 3286 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 3287 | { |
cdd6c482 | 3288 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 3289 | |
ad3a37de | 3290 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
3291 | /* |
3292 | * There are two ways this annotation is useful: | |
3293 | * | |
3294 | * 1) there is a lock recursion from perf_event_exit_task | |
3295 | * see the comment there. | |
3296 | * | |
3297 | * 2) there is a lock-inversion with mmap_sem through | |
3298 | * perf_event_read_group(), which takes faults while | |
3299 | * holding ctx->mutex, however this is called after | |
3300 | * the last filedesc died, so there is no possibility | |
3301 | * to trigger the AB-BA case. | |
3302 | */ | |
3303 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
46ce0fe9 | 3304 | perf_remove_from_context(event, true); |
d859e29f | 3305 | mutex_unlock(&ctx->mutex); |
0793a61d | 3306 | |
cdd6c482 | 3307 | free_event(event); |
0793a61d TG |
3308 | |
3309 | return 0; | |
3310 | } | |
a66a3052 | 3311 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 3312 | |
a66a3052 PZ |
3313 | /* |
3314 | * Called when the last reference to the file is gone. | |
3315 | */ | |
a6fa941d | 3316 | static void put_event(struct perf_event *event) |
fb0459d7 | 3317 | { |
8882135b | 3318 | struct task_struct *owner; |
fb0459d7 | 3319 | |
a6fa941d AV |
3320 | if (!atomic_long_dec_and_test(&event->refcount)) |
3321 | return; | |
fb0459d7 | 3322 | |
8882135b PZ |
3323 | rcu_read_lock(); |
3324 | owner = ACCESS_ONCE(event->owner); | |
3325 | /* | |
3326 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3327 | * !owner it means the list deletion is complete and we can indeed | |
3328 | * free this event, otherwise we need to serialize on | |
3329 | * owner->perf_event_mutex. | |
3330 | */ | |
3331 | smp_read_barrier_depends(); | |
3332 | if (owner) { | |
3333 | /* | |
3334 | * Since delayed_put_task_struct() also drops the last | |
3335 | * task reference we can safely take a new reference | |
3336 | * while holding the rcu_read_lock(). | |
3337 | */ | |
3338 | get_task_struct(owner); | |
3339 | } | |
3340 | rcu_read_unlock(); | |
3341 | ||
3342 | if (owner) { | |
3343 | mutex_lock(&owner->perf_event_mutex); | |
3344 | /* | |
3345 | * We have to re-check the event->owner field, if it is cleared | |
3346 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3347 | * ensured they're done, and we can proceed with freeing the | |
3348 | * event. | |
3349 | */ | |
3350 | if (event->owner) | |
3351 | list_del_init(&event->owner_entry); | |
3352 | mutex_unlock(&owner->perf_event_mutex); | |
3353 | put_task_struct(owner); | |
3354 | } | |
3355 | ||
a6fa941d AV |
3356 | perf_event_release_kernel(event); |
3357 | } | |
3358 | ||
3359 | static int perf_release(struct inode *inode, struct file *file) | |
3360 | { | |
3361 | put_event(file->private_data); | |
3362 | return 0; | |
fb0459d7 | 3363 | } |
fb0459d7 | 3364 | |
59ed446f | 3365 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3366 | { |
cdd6c482 | 3367 | struct perf_event *child; |
e53c0994 PZ |
3368 | u64 total = 0; |
3369 | ||
59ed446f PZ |
3370 | *enabled = 0; |
3371 | *running = 0; | |
3372 | ||
6f10581a | 3373 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3374 | total += perf_event_read(event); |
59ed446f PZ |
3375 | *enabled += event->total_time_enabled + |
3376 | atomic64_read(&event->child_total_time_enabled); | |
3377 | *running += event->total_time_running + | |
3378 | atomic64_read(&event->child_total_time_running); | |
3379 | ||
3380 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3381 | total += perf_event_read(child); |
59ed446f PZ |
3382 | *enabled += child->total_time_enabled; |
3383 | *running += child->total_time_running; | |
3384 | } | |
6f10581a | 3385 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3386 | |
3387 | return total; | |
3388 | } | |
fb0459d7 | 3389 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3390 | |
cdd6c482 | 3391 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3392 | u64 read_format, char __user *buf) |
3393 | { | |
cdd6c482 | 3394 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3395 | int n = 0, size = 0, ret = -EFAULT; |
3396 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3397 | u64 values[5]; |
59ed446f | 3398 | u64 count, enabled, running; |
abf4868b | 3399 | |
6f10581a | 3400 | mutex_lock(&ctx->mutex); |
59ed446f | 3401 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3402 | |
3403 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3404 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3405 | values[n++] = enabled; | |
3406 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3407 | values[n++] = running; | |
abf4868b PZ |
3408 | values[n++] = count; |
3409 | if (read_format & PERF_FORMAT_ID) | |
3410 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3411 | |
3412 | size = n * sizeof(u64); | |
3413 | ||
3414 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3415 | goto unlock; |
3dab77fb | 3416 | |
6f10581a | 3417 | ret = size; |
3dab77fb | 3418 | |
65abc865 | 3419 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3420 | n = 0; |
3dab77fb | 3421 | |
59ed446f | 3422 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3423 | if (read_format & PERF_FORMAT_ID) |
3424 | values[n++] = primary_event_id(sub); | |
3425 | ||
3426 | size = n * sizeof(u64); | |
3427 | ||
184d3da8 | 3428 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3429 | ret = -EFAULT; |
3430 | goto unlock; | |
3431 | } | |
abf4868b PZ |
3432 | |
3433 | ret += size; | |
3dab77fb | 3434 | } |
6f10581a PZ |
3435 | unlock: |
3436 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3437 | |
abf4868b | 3438 | return ret; |
3dab77fb PZ |
3439 | } |
3440 | ||
cdd6c482 | 3441 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3442 | u64 read_format, char __user *buf) |
3443 | { | |
59ed446f | 3444 | u64 enabled, running; |
3dab77fb PZ |
3445 | u64 values[4]; |
3446 | int n = 0; | |
3447 | ||
59ed446f PZ |
3448 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3449 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3450 | values[n++] = enabled; | |
3451 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3452 | values[n++] = running; | |
3dab77fb | 3453 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3454 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3455 | |
3456 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3457 | return -EFAULT; | |
3458 | ||
3459 | return n * sizeof(u64); | |
3460 | } | |
3461 | ||
0793a61d | 3462 | /* |
cdd6c482 | 3463 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3464 | */ |
3465 | static ssize_t | |
cdd6c482 | 3466 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3467 | { |
cdd6c482 | 3468 | u64 read_format = event->attr.read_format; |
3dab77fb | 3469 | int ret; |
0793a61d | 3470 | |
3b6f9e5c | 3471 | /* |
cdd6c482 | 3472 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3473 | * error state (i.e. because it was pinned but it couldn't be |
3474 | * scheduled on to the CPU at some point). | |
3475 | */ | |
cdd6c482 | 3476 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3477 | return 0; |
3478 | ||
c320c7b7 | 3479 | if (count < event->read_size) |
3dab77fb PZ |
3480 | return -ENOSPC; |
3481 | ||
cdd6c482 | 3482 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3483 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3484 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3485 | else |
cdd6c482 | 3486 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3487 | |
3dab77fb | 3488 | return ret; |
0793a61d TG |
3489 | } |
3490 | ||
0793a61d TG |
3491 | static ssize_t |
3492 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3493 | { | |
cdd6c482 | 3494 | struct perf_event *event = file->private_data; |
0793a61d | 3495 | |
cdd6c482 | 3496 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3497 | } |
3498 | ||
3499 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3500 | { | |
cdd6c482 | 3501 | struct perf_event *event = file->private_data; |
76369139 | 3502 | struct ring_buffer *rb; |
c33a0bc4 | 3503 | unsigned int events = POLL_HUP; |
c7138f37 | 3504 | |
10c6db11 | 3505 | /* |
9bb5d40c PZ |
3506 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3507 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3508 | */ |
3509 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3510 | rb = event->rb; |
3511 | if (rb) | |
76369139 | 3512 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3513 | mutex_unlock(&event->mmap_mutex); |
3514 | ||
cdd6c482 | 3515 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3516 | |
0793a61d TG |
3517 | return events; |
3518 | } | |
3519 | ||
cdd6c482 | 3520 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3521 | { |
cdd6c482 | 3522 | (void)perf_event_read(event); |
e7850595 | 3523 | local64_set(&event->count, 0); |
cdd6c482 | 3524 | perf_event_update_userpage(event); |
3df5edad PZ |
3525 | } |
3526 | ||
c93f7669 | 3527 | /* |
cdd6c482 IM |
3528 | * Holding the top-level event's child_mutex means that any |
3529 | * descendant process that has inherited this event will block | |
3530 | * in sync_child_event if it goes to exit, thus satisfying the | |
3531 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3532 | */ |
cdd6c482 IM |
3533 | static void perf_event_for_each_child(struct perf_event *event, |
3534 | void (*func)(struct perf_event *)) | |
3df5edad | 3535 | { |
cdd6c482 | 3536 | struct perf_event *child; |
3df5edad | 3537 | |
cdd6c482 IM |
3538 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3539 | mutex_lock(&event->child_mutex); | |
3540 | func(event); | |
3541 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3542 | func(child); |
cdd6c482 | 3543 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3544 | } |
3545 | ||
cdd6c482 IM |
3546 | static void perf_event_for_each(struct perf_event *event, |
3547 | void (*func)(struct perf_event *)) | |
3df5edad | 3548 | { |
cdd6c482 IM |
3549 | struct perf_event_context *ctx = event->ctx; |
3550 | struct perf_event *sibling; | |
3df5edad | 3551 | |
75f937f2 PZ |
3552 | WARN_ON_ONCE(ctx->parent_ctx); |
3553 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3554 | event = event->group_leader; |
75f937f2 | 3555 | |
cdd6c482 | 3556 | perf_event_for_each_child(event, func); |
cdd6c482 | 3557 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3558 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3559 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3560 | } |
3561 | ||
cdd6c482 | 3562 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3563 | { |
cdd6c482 | 3564 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3565 | int ret = 0, active; |
08247e31 PZ |
3566 | u64 value; |
3567 | ||
6c7e550f | 3568 | if (!is_sampling_event(event)) |
08247e31 PZ |
3569 | return -EINVAL; |
3570 | ||
ad0cf347 | 3571 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3572 | return -EFAULT; |
3573 | ||
3574 | if (!value) | |
3575 | return -EINVAL; | |
3576 | ||
e625cce1 | 3577 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3578 | if (event->attr.freq) { |
3579 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3580 | ret = -EINVAL; |
3581 | goto unlock; | |
3582 | } | |
3583 | ||
cdd6c482 | 3584 | event->attr.sample_freq = value; |
08247e31 | 3585 | } else { |
cdd6c482 IM |
3586 | event->attr.sample_period = value; |
3587 | event->hw.sample_period = value; | |
08247e31 | 3588 | } |
bad7192b PZ |
3589 | |
3590 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3591 | if (active) { | |
3592 | perf_pmu_disable(ctx->pmu); | |
3593 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3594 | } | |
3595 | ||
3596 | local64_set(&event->hw.period_left, 0); | |
3597 | ||
3598 | if (active) { | |
3599 | event->pmu->start(event, PERF_EF_RELOAD); | |
3600 | perf_pmu_enable(ctx->pmu); | |
3601 | } | |
3602 | ||
08247e31 | 3603 | unlock: |
e625cce1 | 3604 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3605 | |
3606 | return ret; | |
3607 | } | |
3608 | ||
ac9721f3 PZ |
3609 | static const struct file_operations perf_fops; |
3610 | ||
2903ff01 | 3611 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3612 | { |
2903ff01 AV |
3613 | struct fd f = fdget(fd); |
3614 | if (!f.file) | |
3615 | return -EBADF; | |
ac9721f3 | 3616 | |
2903ff01 AV |
3617 | if (f.file->f_op != &perf_fops) { |
3618 | fdput(f); | |
3619 | return -EBADF; | |
ac9721f3 | 3620 | } |
2903ff01 AV |
3621 | *p = f; |
3622 | return 0; | |
ac9721f3 PZ |
3623 | } |
3624 | ||
3625 | static int perf_event_set_output(struct perf_event *event, | |
3626 | struct perf_event *output_event); | |
6fb2915d | 3627 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3628 | |
d859e29f PM |
3629 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3630 | { | |
cdd6c482 IM |
3631 | struct perf_event *event = file->private_data; |
3632 | void (*func)(struct perf_event *); | |
3df5edad | 3633 | u32 flags = arg; |
d859e29f PM |
3634 | |
3635 | switch (cmd) { | |
cdd6c482 IM |
3636 | case PERF_EVENT_IOC_ENABLE: |
3637 | func = perf_event_enable; | |
d859e29f | 3638 | break; |
cdd6c482 IM |
3639 | case PERF_EVENT_IOC_DISABLE: |
3640 | func = perf_event_disable; | |
79f14641 | 3641 | break; |
cdd6c482 IM |
3642 | case PERF_EVENT_IOC_RESET: |
3643 | func = perf_event_reset; | |
6de6a7b9 | 3644 | break; |
3df5edad | 3645 | |
cdd6c482 IM |
3646 | case PERF_EVENT_IOC_REFRESH: |
3647 | return perf_event_refresh(event, arg); | |
08247e31 | 3648 | |
cdd6c482 IM |
3649 | case PERF_EVENT_IOC_PERIOD: |
3650 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3651 | |
cf4957f1 JO |
3652 | case PERF_EVENT_IOC_ID: |
3653 | { | |
3654 | u64 id = primary_event_id(event); | |
3655 | ||
3656 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3657 | return -EFAULT; | |
3658 | return 0; | |
3659 | } | |
3660 | ||
cdd6c482 | 3661 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3662 | { |
ac9721f3 | 3663 | int ret; |
ac9721f3 | 3664 | if (arg != -1) { |
2903ff01 AV |
3665 | struct perf_event *output_event; |
3666 | struct fd output; | |
3667 | ret = perf_fget_light(arg, &output); | |
3668 | if (ret) | |
3669 | return ret; | |
3670 | output_event = output.file->private_data; | |
3671 | ret = perf_event_set_output(event, output_event); | |
3672 | fdput(output); | |
3673 | } else { | |
3674 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3675 | } |
ac9721f3 PZ |
3676 | return ret; |
3677 | } | |
a4be7c27 | 3678 | |
6fb2915d LZ |
3679 | case PERF_EVENT_IOC_SET_FILTER: |
3680 | return perf_event_set_filter(event, (void __user *)arg); | |
3681 | ||
d859e29f | 3682 | default: |
3df5edad | 3683 | return -ENOTTY; |
d859e29f | 3684 | } |
3df5edad PZ |
3685 | |
3686 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3687 | perf_event_for_each(event, func); |
3df5edad | 3688 | else |
cdd6c482 | 3689 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3690 | |
3691 | return 0; | |
d859e29f PM |
3692 | } |
3693 | ||
cdd6c482 | 3694 | int perf_event_task_enable(void) |
771d7cde | 3695 | { |
cdd6c482 | 3696 | struct perf_event *event; |
771d7cde | 3697 | |
cdd6c482 IM |
3698 | mutex_lock(¤t->perf_event_mutex); |
3699 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3700 | perf_event_for_each_child(event, perf_event_enable); | |
3701 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3702 | |
3703 | return 0; | |
3704 | } | |
3705 | ||
cdd6c482 | 3706 | int perf_event_task_disable(void) |
771d7cde | 3707 | { |
cdd6c482 | 3708 | struct perf_event *event; |
771d7cde | 3709 | |
cdd6c482 IM |
3710 | mutex_lock(¤t->perf_event_mutex); |
3711 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3712 | perf_event_for_each_child(event, perf_event_disable); | |
3713 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3714 | |
3715 | return 0; | |
3716 | } | |
3717 | ||
cdd6c482 | 3718 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3719 | { |
a4eaf7f1 PZ |
3720 | if (event->hw.state & PERF_HES_STOPPED) |
3721 | return 0; | |
3722 | ||
cdd6c482 | 3723 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3724 | return 0; |
3725 | ||
35edc2a5 | 3726 | return event->pmu->event_idx(event); |
194002b2 PZ |
3727 | } |
3728 | ||
c4794295 | 3729 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3730 | u64 *now, |
7f310a5d EM |
3731 | u64 *enabled, |
3732 | u64 *running) | |
c4794295 | 3733 | { |
e3f3541c | 3734 | u64 ctx_time; |
c4794295 | 3735 | |
e3f3541c PZ |
3736 | *now = perf_clock(); |
3737 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3738 | *enabled = ctx_time - event->tstamp_enabled; |
3739 | *running = ctx_time - event->tstamp_running; | |
3740 | } | |
3741 | ||
fa731587 PZ |
3742 | static void perf_event_init_userpage(struct perf_event *event) |
3743 | { | |
3744 | struct perf_event_mmap_page *userpg; | |
3745 | struct ring_buffer *rb; | |
3746 | ||
3747 | rcu_read_lock(); | |
3748 | rb = rcu_dereference(event->rb); | |
3749 | if (!rb) | |
3750 | goto unlock; | |
3751 | ||
3752 | userpg = rb->user_page; | |
3753 | ||
3754 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
3755 | userpg->cap_bit0_is_deprecated = 1; | |
3756 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
3757 | ||
3758 | unlock: | |
3759 | rcu_read_unlock(); | |
3760 | } | |
3761 | ||
c7206205 | 3762 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3763 | { |
3764 | } | |
3765 | ||
38ff667b PZ |
3766 | /* |
3767 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3768 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3769 | * code calls this from NMI context. | |
3770 | */ | |
cdd6c482 | 3771 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3772 | { |
cdd6c482 | 3773 | struct perf_event_mmap_page *userpg; |
76369139 | 3774 | struct ring_buffer *rb; |
e3f3541c | 3775 | u64 enabled, running, now; |
38ff667b PZ |
3776 | |
3777 | rcu_read_lock(); | |
5ec4c599 PZ |
3778 | rb = rcu_dereference(event->rb); |
3779 | if (!rb) | |
3780 | goto unlock; | |
3781 | ||
0d641208 EM |
3782 | /* |
3783 | * compute total_time_enabled, total_time_running | |
3784 | * based on snapshot values taken when the event | |
3785 | * was last scheduled in. | |
3786 | * | |
3787 | * we cannot simply called update_context_time() | |
3788 | * because of locking issue as we can be called in | |
3789 | * NMI context | |
3790 | */ | |
e3f3541c | 3791 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 3792 | |
76369139 | 3793 | userpg = rb->user_page; |
7b732a75 PZ |
3794 | /* |
3795 | * Disable preemption so as to not let the corresponding user-space | |
3796 | * spin too long if we get preempted. | |
3797 | */ | |
3798 | preempt_disable(); | |
37d81828 | 3799 | ++userpg->lock; |
92f22a38 | 3800 | barrier(); |
cdd6c482 | 3801 | userpg->index = perf_event_index(event); |
b5e58793 | 3802 | userpg->offset = perf_event_count(event); |
365a4038 | 3803 | if (userpg->index) |
e7850595 | 3804 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3805 | |
0d641208 | 3806 | userpg->time_enabled = enabled + |
cdd6c482 | 3807 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3808 | |
0d641208 | 3809 | userpg->time_running = running + |
cdd6c482 | 3810 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3811 | |
c7206205 | 3812 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3813 | |
92f22a38 | 3814 | barrier(); |
37d81828 | 3815 | ++userpg->lock; |
7b732a75 | 3816 | preempt_enable(); |
38ff667b | 3817 | unlock: |
7b732a75 | 3818 | rcu_read_unlock(); |
37d81828 PM |
3819 | } |
3820 | ||
906010b2 PZ |
3821 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3822 | { | |
3823 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3824 | struct ring_buffer *rb; |
906010b2 PZ |
3825 | int ret = VM_FAULT_SIGBUS; |
3826 | ||
3827 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3828 | if (vmf->pgoff == 0) | |
3829 | ret = 0; | |
3830 | return ret; | |
3831 | } | |
3832 | ||
3833 | rcu_read_lock(); | |
76369139 FW |
3834 | rb = rcu_dereference(event->rb); |
3835 | if (!rb) | |
906010b2 PZ |
3836 | goto unlock; |
3837 | ||
3838 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3839 | goto unlock; | |
3840 | ||
76369139 | 3841 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3842 | if (!vmf->page) |
3843 | goto unlock; | |
3844 | ||
3845 | get_page(vmf->page); | |
3846 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3847 | vmf->page->index = vmf->pgoff; | |
3848 | ||
3849 | ret = 0; | |
3850 | unlock: | |
3851 | rcu_read_unlock(); | |
3852 | ||
3853 | return ret; | |
3854 | } | |
3855 | ||
10c6db11 PZ |
3856 | static void ring_buffer_attach(struct perf_event *event, |
3857 | struct ring_buffer *rb) | |
3858 | { | |
3859 | unsigned long flags; | |
3860 | ||
3861 | if (!list_empty(&event->rb_entry)) | |
3862 | return; | |
3863 | ||
3864 | spin_lock_irqsave(&rb->event_lock, flags); | |
9bb5d40c PZ |
3865 | if (list_empty(&event->rb_entry)) |
3866 | list_add(&event->rb_entry, &rb->event_list); | |
10c6db11 PZ |
3867 | spin_unlock_irqrestore(&rb->event_lock, flags); |
3868 | } | |
3869 | ||
9bb5d40c | 3870 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb) |
10c6db11 PZ |
3871 | { |
3872 | unsigned long flags; | |
3873 | ||
3874 | if (list_empty(&event->rb_entry)) | |
3875 | return; | |
3876 | ||
3877 | spin_lock_irqsave(&rb->event_lock, flags); | |
3878 | list_del_init(&event->rb_entry); | |
3879 | wake_up_all(&event->waitq); | |
3880 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3881 | } | |
3882 | ||
3883 | static void ring_buffer_wakeup(struct perf_event *event) | |
3884 | { | |
3885 | struct ring_buffer *rb; | |
3886 | ||
3887 | rcu_read_lock(); | |
3888 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3889 | if (rb) { |
3890 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3891 | wake_up_all(&event->waitq); | |
3892 | } | |
10c6db11 PZ |
3893 | rcu_read_unlock(); |
3894 | } | |
3895 | ||
76369139 | 3896 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3897 | { |
76369139 | 3898 | struct ring_buffer *rb; |
906010b2 | 3899 | |
76369139 FW |
3900 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3901 | rb_free(rb); | |
7b732a75 PZ |
3902 | } |
3903 | ||
76369139 | 3904 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3905 | { |
76369139 | 3906 | struct ring_buffer *rb; |
7b732a75 | 3907 | |
ac9721f3 | 3908 | rcu_read_lock(); |
76369139 FW |
3909 | rb = rcu_dereference(event->rb); |
3910 | if (rb) { | |
3911 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3912 | rb = NULL; | |
ac9721f3 PZ |
3913 | } |
3914 | rcu_read_unlock(); | |
3915 | ||
76369139 | 3916 | return rb; |
ac9721f3 PZ |
3917 | } |
3918 | ||
76369139 | 3919 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3920 | { |
76369139 | 3921 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3922 | return; |
7b732a75 | 3923 | |
9bb5d40c | 3924 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3925 | |
76369139 | 3926 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3927 | } |
3928 | ||
3929 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3930 | { | |
cdd6c482 | 3931 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3932 | |
cdd6c482 | 3933 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3934 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3935 | } |
3936 | ||
9bb5d40c PZ |
3937 | /* |
3938 | * A buffer can be mmap()ed multiple times; either directly through the same | |
3939 | * event, or through other events by use of perf_event_set_output(). | |
3940 | * | |
3941 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
3942 | * the buffer here, where we still have a VM context. This means we need | |
3943 | * to detach all events redirecting to us. | |
3944 | */ | |
7b732a75 PZ |
3945 | static void perf_mmap_close(struct vm_area_struct *vma) |
3946 | { | |
cdd6c482 | 3947 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3948 | |
9bb5d40c PZ |
3949 | struct ring_buffer *rb = event->rb; |
3950 | struct user_struct *mmap_user = rb->mmap_user; | |
3951 | int mmap_locked = rb->mmap_locked; | |
3952 | unsigned long size = perf_data_size(rb); | |
789f90fc | 3953 | |
9bb5d40c PZ |
3954 | atomic_dec(&rb->mmap_count); |
3955 | ||
3956 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
3957 | return; | |
3958 | ||
3959 | /* Detach current event from the buffer. */ | |
3960 | rcu_assign_pointer(event->rb, NULL); | |
3961 | ring_buffer_detach(event, rb); | |
3962 | mutex_unlock(&event->mmap_mutex); | |
3963 | ||
3964 | /* If there's still other mmap()s of this buffer, we're done. */ | |
3965 | if (atomic_read(&rb->mmap_count)) { | |
3966 | ring_buffer_put(rb); /* can't be last */ | |
3967 | return; | |
3968 | } | |
ac9721f3 | 3969 | |
9bb5d40c PZ |
3970 | /* |
3971 | * No other mmap()s, detach from all other events that might redirect | |
3972 | * into the now unreachable buffer. Somewhat complicated by the | |
3973 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
3974 | */ | |
3975 | again: | |
3976 | rcu_read_lock(); | |
3977 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
3978 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
3979 | /* | |
3980 | * This event is en-route to free_event() which will | |
3981 | * detach it and remove it from the list. | |
3982 | */ | |
3983 | continue; | |
3984 | } | |
3985 | rcu_read_unlock(); | |
789f90fc | 3986 | |
9bb5d40c PZ |
3987 | mutex_lock(&event->mmap_mutex); |
3988 | /* | |
3989 | * Check we didn't race with perf_event_set_output() which can | |
3990 | * swizzle the rb from under us while we were waiting to | |
3991 | * acquire mmap_mutex. | |
3992 | * | |
3993 | * If we find a different rb; ignore this event, a next | |
3994 | * iteration will no longer find it on the list. We have to | |
3995 | * still restart the iteration to make sure we're not now | |
3996 | * iterating the wrong list. | |
3997 | */ | |
3998 | if (event->rb == rb) { | |
3999 | rcu_assign_pointer(event->rb, NULL); | |
4000 | ring_buffer_detach(event, rb); | |
4001 | ring_buffer_put(rb); /* can't be last, we still have one */ | |
26cb63ad | 4002 | } |
cdd6c482 | 4003 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4004 | put_event(event); |
ac9721f3 | 4005 | |
9bb5d40c PZ |
4006 | /* |
4007 | * Restart the iteration; either we're on the wrong list or | |
4008 | * destroyed its integrity by doing a deletion. | |
4009 | */ | |
4010 | goto again; | |
7b732a75 | 4011 | } |
9bb5d40c PZ |
4012 | rcu_read_unlock(); |
4013 | ||
4014 | /* | |
4015 | * It could be there's still a few 0-ref events on the list; they'll | |
4016 | * get cleaned up by free_event() -- they'll also still have their | |
4017 | * ref on the rb and will free it whenever they are done with it. | |
4018 | * | |
4019 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4020 | * undo the VM accounting. | |
4021 | */ | |
4022 | ||
4023 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4024 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4025 | free_uid(mmap_user); | |
4026 | ||
4027 | ring_buffer_put(rb); /* could be last */ | |
37d81828 PM |
4028 | } |
4029 | ||
f0f37e2f | 4030 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4031 | .open = perf_mmap_open, |
4032 | .close = perf_mmap_close, | |
4033 | .fault = perf_mmap_fault, | |
4034 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4035 | }; |
4036 | ||
4037 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4038 | { | |
cdd6c482 | 4039 | struct perf_event *event = file->private_data; |
22a4f650 | 4040 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4041 | struct user_struct *user = current_user(); |
22a4f650 | 4042 | unsigned long locked, lock_limit; |
76369139 | 4043 | struct ring_buffer *rb; |
7b732a75 PZ |
4044 | unsigned long vma_size; |
4045 | unsigned long nr_pages; | |
789f90fc | 4046 | long user_extra, extra; |
d57e34fd | 4047 | int ret = 0, flags = 0; |
37d81828 | 4048 | |
c7920614 PZ |
4049 | /* |
4050 | * Don't allow mmap() of inherited per-task counters. This would | |
4051 | * create a performance issue due to all children writing to the | |
76369139 | 4052 | * same rb. |
c7920614 PZ |
4053 | */ |
4054 | if (event->cpu == -1 && event->attr.inherit) | |
4055 | return -EINVAL; | |
4056 | ||
43a21ea8 | 4057 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4058 | return -EINVAL; |
7b732a75 PZ |
4059 | |
4060 | vma_size = vma->vm_end - vma->vm_start; | |
4061 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4062 | ||
7730d865 | 4063 | /* |
76369139 | 4064 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4065 | * can do bitmasks instead of modulo. |
4066 | */ | |
4067 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
4068 | return -EINVAL; |
4069 | ||
7b732a75 | 4070 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4071 | return -EINVAL; |
4072 | ||
7b732a75 PZ |
4073 | if (vma->vm_pgoff != 0) |
4074 | return -EINVAL; | |
37d81828 | 4075 | |
cdd6c482 | 4076 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4077 | again: |
cdd6c482 | 4078 | mutex_lock(&event->mmap_mutex); |
76369139 | 4079 | if (event->rb) { |
9bb5d40c | 4080 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4081 | ret = -EINVAL; |
9bb5d40c PZ |
4082 | goto unlock; |
4083 | } | |
4084 | ||
4085 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4086 | /* | |
4087 | * Raced against perf_mmap_close() through | |
4088 | * perf_event_set_output(). Try again, hope for better | |
4089 | * luck. | |
4090 | */ | |
4091 | mutex_unlock(&event->mmap_mutex); | |
4092 | goto again; | |
4093 | } | |
4094 | ||
ebb3c4c4 PZ |
4095 | goto unlock; |
4096 | } | |
4097 | ||
789f90fc | 4098 | user_extra = nr_pages + 1; |
cdd6c482 | 4099 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4100 | |
4101 | /* | |
4102 | * Increase the limit linearly with more CPUs: | |
4103 | */ | |
4104 | user_lock_limit *= num_online_cpus(); | |
4105 | ||
789f90fc | 4106 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4107 | |
789f90fc PZ |
4108 | extra = 0; |
4109 | if (user_locked > user_lock_limit) | |
4110 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4111 | |
78d7d407 | 4112 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4113 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4114 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4115 | |
459ec28a IM |
4116 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4117 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4118 | ret = -EPERM; |
4119 | goto unlock; | |
4120 | } | |
7b732a75 | 4121 | |
76369139 | 4122 | WARN_ON(event->rb); |
906010b2 | 4123 | |
d57e34fd | 4124 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4125 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4126 | |
4ec8363d VW |
4127 | rb = rb_alloc(nr_pages, |
4128 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4129 | event->cpu, flags); | |
4130 | ||
76369139 | 4131 | if (!rb) { |
ac9721f3 | 4132 | ret = -ENOMEM; |
ebb3c4c4 | 4133 | goto unlock; |
ac9721f3 | 4134 | } |
26cb63ad | 4135 | |
9bb5d40c | 4136 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4137 | rb->mmap_locked = extra; |
4138 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4139 | |
ac9721f3 | 4140 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4141 | vma->vm_mm->pinned_vm += extra; |
4142 | ||
9bb5d40c | 4143 | ring_buffer_attach(event, rb); |
26cb63ad | 4144 | rcu_assign_pointer(event->rb, rb); |
ac9721f3 | 4145 | |
fa731587 | 4146 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4147 | perf_event_update_userpage(event); |
4148 | ||
ebb3c4c4 | 4149 | unlock: |
ac9721f3 PZ |
4150 | if (!ret) |
4151 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4152 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4153 | |
9bb5d40c PZ |
4154 | /* |
4155 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4156 | * vma. | |
4157 | */ | |
26cb63ad | 4158 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4159 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4160 | |
4161 | return ret; | |
37d81828 PM |
4162 | } |
4163 | ||
3c446b3d PZ |
4164 | static int perf_fasync(int fd, struct file *filp, int on) |
4165 | { | |
496ad9aa | 4166 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4167 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4168 | int retval; |
4169 | ||
4170 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4171 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4172 | mutex_unlock(&inode->i_mutex); |
4173 | ||
4174 | if (retval < 0) | |
4175 | return retval; | |
4176 | ||
4177 | return 0; | |
4178 | } | |
4179 | ||
0793a61d | 4180 | static const struct file_operations perf_fops = { |
3326c1ce | 4181 | .llseek = no_llseek, |
0793a61d TG |
4182 | .release = perf_release, |
4183 | .read = perf_read, | |
4184 | .poll = perf_poll, | |
d859e29f PM |
4185 | .unlocked_ioctl = perf_ioctl, |
4186 | .compat_ioctl = perf_ioctl, | |
37d81828 | 4187 | .mmap = perf_mmap, |
3c446b3d | 4188 | .fasync = perf_fasync, |
0793a61d TG |
4189 | }; |
4190 | ||
925d519a | 4191 | /* |
cdd6c482 | 4192 | * Perf event wakeup |
925d519a PZ |
4193 | * |
4194 | * If there's data, ensure we set the poll() state and publish everything | |
4195 | * to user-space before waking everybody up. | |
4196 | */ | |
4197 | ||
cdd6c482 | 4198 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4199 | { |
10c6db11 | 4200 | ring_buffer_wakeup(event); |
4c9e2542 | 4201 | |
cdd6c482 IM |
4202 | if (event->pending_kill) { |
4203 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4204 | event->pending_kill = 0; | |
4c9e2542 | 4205 | } |
925d519a PZ |
4206 | } |
4207 | ||
e360adbe | 4208 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4209 | { |
cdd6c482 IM |
4210 | struct perf_event *event = container_of(entry, |
4211 | struct perf_event, pending); | |
79f14641 | 4212 | |
cdd6c482 IM |
4213 | if (event->pending_disable) { |
4214 | event->pending_disable = 0; | |
4215 | __perf_event_disable(event); | |
79f14641 PZ |
4216 | } |
4217 | ||
cdd6c482 IM |
4218 | if (event->pending_wakeup) { |
4219 | event->pending_wakeup = 0; | |
4220 | perf_event_wakeup(event); | |
79f14641 PZ |
4221 | } |
4222 | } | |
4223 | ||
39447b38 ZY |
4224 | /* |
4225 | * We assume there is only KVM supporting the callbacks. | |
4226 | * Later on, we might change it to a list if there is | |
4227 | * another virtualization implementation supporting the callbacks. | |
4228 | */ | |
4229 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4230 | ||
4231 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4232 | { | |
4233 | perf_guest_cbs = cbs; | |
4234 | return 0; | |
4235 | } | |
4236 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4237 | ||
4238 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4239 | { | |
4240 | perf_guest_cbs = NULL; | |
4241 | return 0; | |
4242 | } | |
4243 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4244 | ||
4018994f JO |
4245 | static void |
4246 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4247 | struct pt_regs *regs, u64 mask) | |
4248 | { | |
4249 | int bit; | |
4250 | ||
4251 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4252 | sizeof(mask) * BITS_PER_BYTE) { | |
4253 | u64 val; | |
4254 | ||
4255 | val = perf_reg_value(regs, bit); | |
4256 | perf_output_put(handle, val); | |
4257 | } | |
4258 | } | |
4259 | ||
4260 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4261 | struct pt_regs *regs) | |
4262 | { | |
4263 | if (!user_mode(regs)) { | |
4264 | if (current->mm) | |
4265 | regs = task_pt_regs(current); | |
4266 | else | |
4267 | regs = NULL; | |
4268 | } | |
4269 | ||
4270 | if (regs) { | |
4271 | regs_user->regs = regs; | |
4272 | regs_user->abi = perf_reg_abi(current); | |
4273 | } | |
4274 | } | |
4275 | ||
c5ebcedb JO |
4276 | /* |
4277 | * Get remaining task size from user stack pointer. | |
4278 | * | |
4279 | * It'd be better to take stack vma map and limit this more | |
4280 | * precisly, but there's no way to get it safely under interrupt, | |
4281 | * so using TASK_SIZE as limit. | |
4282 | */ | |
4283 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4284 | { | |
4285 | unsigned long addr = perf_user_stack_pointer(regs); | |
4286 | ||
4287 | if (!addr || addr >= TASK_SIZE) | |
4288 | return 0; | |
4289 | ||
4290 | return TASK_SIZE - addr; | |
4291 | } | |
4292 | ||
4293 | static u16 | |
4294 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4295 | struct pt_regs *regs) | |
4296 | { | |
4297 | u64 task_size; | |
4298 | ||
4299 | /* No regs, no stack pointer, no dump. */ | |
4300 | if (!regs) | |
4301 | return 0; | |
4302 | ||
4303 | /* | |
4304 | * Check if we fit in with the requested stack size into the: | |
4305 | * - TASK_SIZE | |
4306 | * If we don't, we limit the size to the TASK_SIZE. | |
4307 | * | |
4308 | * - remaining sample size | |
4309 | * If we don't, we customize the stack size to | |
4310 | * fit in to the remaining sample size. | |
4311 | */ | |
4312 | ||
4313 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4314 | stack_size = min(stack_size, (u16) task_size); | |
4315 | ||
4316 | /* Current header size plus static size and dynamic size. */ | |
4317 | header_size += 2 * sizeof(u64); | |
4318 | ||
4319 | /* Do we fit in with the current stack dump size? */ | |
4320 | if ((u16) (header_size + stack_size) < header_size) { | |
4321 | /* | |
4322 | * If we overflow the maximum size for the sample, | |
4323 | * we customize the stack dump size to fit in. | |
4324 | */ | |
4325 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4326 | stack_size = round_up(stack_size, sizeof(u64)); | |
4327 | } | |
4328 | ||
4329 | return stack_size; | |
4330 | } | |
4331 | ||
4332 | static void | |
4333 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4334 | struct pt_regs *regs) | |
4335 | { | |
4336 | /* Case of a kernel thread, nothing to dump */ | |
4337 | if (!regs) { | |
4338 | u64 size = 0; | |
4339 | perf_output_put(handle, size); | |
4340 | } else { | |
4341 | unsigned long sp; | |
4342 | unsigned int rem; | |
4343 | u64 dyn_size; | |
4344 | ||
4345 | /* | |
4346 | * We dump: | |
4347 | * static size | |
4348 | * - the size requested by user or the best one we can fit | |
4349 | * in to the sample max size | |
4350 | * data | |
4351 | * - user stack dump data | |
4352 | * dynamic size | |
4353 | * - the actual dumped size | |
4354 | */ | |
4355 | ||
4356 | /* Static size. */ | |
4357 | perf_output_put(handle, dump_size); | |
4358 | ||
4359 | /* Data. */ | |
4360 | sp = perf_user_stack_pointer(regs); | |
4361 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4362 | dyn_size = dump_size - rem; | |
4363 | ||
4364 | perf_output_skip(handle, rem); | |
4365 | ||
4366 | /* Dynamic size. */ | |
4367 | perf_output_put(handle, dyn_size); | |
4368 | } | |
4369 | } | |
4370 | ||
c980d109 ACM |
4371 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4372 | struct perf_sample_data *data, | |
4373 | struct perf_event *event) | |
6844c09d ACM |
4374 | { |
4375 | u64 sample_type = event->attr.sample_type; | |
4376 | ||
4377 | data->type = sample_type; | |
4378 | header->size += event->id_header_size; | |
4379 | ||
4380 | if (sample_type & PERF_SAMPLE_TID) { | |
4381 | /* namespace issues */ | |
4382 | data->tid_entry.pid = perf_event_pid(event, current); | |
4383 | data->tid_entry.tid = perf_event_tid(event, current); | |
4384 | } | |
4385 | ||
4386 | if (sample_type & PERF_SAMPLE_TIME) | |
4387 | data->time = perf_clock(); | |
4388 | ||
ff3d527c | 4389 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4390 | data->id = primary_event_id(event); |
4391 | ||
4392 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4393 | data->stream_id = event->id; | |
4394 | ||
4395 | if (sample_type & PERF_SAMPLE_CPU) { | |
4396 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4397 | data->cpu_entry.reserved = 0; | |
4398 | } | |
4399 | } | |
4400 | ||
76369139 FW |
4401 | void perf_event_header__init_id(struct perf_event_header *header, |
4402 | struct perf_sample_data *data, | |
4403 | struct perf_event *event) | |
c980d109 ACM |
4404 | { |
4405 | if (event->attr.sample_id_all) | |
4406 | __perf_event_header__init_id(header, data, event); | |
4407 | } | |
4408 | ||
4409 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4410 | struct perf_sample_data *data) | |
4411 | { | |
4412 | u64 sample_type = data->type; | |
4413 | ||
4414 | if (sample_type & PERF_SAMPLE_TID) | |
4415 | perf_output_put(handle, data->tid_entry); | |
4416 | ||
4417 | if (sample_type & PERF_SAMPLE_TIME) | |
4418 | perf_output_put(handle, data->time); | |
4419 | ||
4420 | if (sample_type & PERF_SAMPLE_ID) | |
4421 | perf_output_put(handle, data->id); | |
4422 | ||
4423 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4424 | perf_output_put(handle, data->stream_id); | |
4425 | ||
4426 | if (sample_type & PERF_SAMPLE_CPU) | |
4427 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4428 | |
4429 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4430 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4431 | } |
4432 | ||
76369139 FW |
4433 | void perf_event__output_id_sample(struct perf_event *event, |
4434 | struct perf_output_handle *handle, | |
4435 | struct perf_sample_data *sample) | |
c980d109 ACM |
4436 | { |
4437 | if (event->attr.sample_id_all) | |
4438 | __perf_event__output_id_sample(handle, sample); | |
4439 | } | |
4440 | ||
3dab77fb | 4441 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4442 | struct perf_event *event, |
4443 | u64 enabled, u64 running) | |
3dab77fb | 4444 | { |
cdd6c482 | 4445 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4446 | u64 values[4]; |
4447 | int n = 0; | |
4448 | ||
b5e58793 | 4449 | values[n++] = perf_event_count(event); |
3dab77fb | 4450 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4451 | values[n++] = enabled + |
cdd6c482 | 4452 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4453 | } |
4454 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4455 | values[n++] = running + |
cdd6c482 | 4456 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4457 | } |
4458 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4459 | values[n++] = primary_event_id(event); |
3dab77fb | 4460 | |
76369139 | 4461 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4462 | } |
4463 | ||
4464 | /* | |
cdd6c482 | 4465 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4466 | */ |
4467 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4468 | struct perf_event *event, |
4469 | u64 enabled, u64 running) | |
3dab77fb | 4470 | { |
cdd6c482 IM |
4471 | struct perf_event *leader = event->group_leader, *sub; |
4472 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4473 | u64 values[5]; |
4474 | int n = 0; | |
4475 | ||
4476 | values[n++] = 1 + leader->nr_siblings; | |
4477 | ||
4478 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4479 | values[n++] = enabled; |
3dab77fb PZ |
4480 | |
4481 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4482 | values[n++] = running; |
3dab77fb | 4483 | |
cdd6c482 | 4484 | if (leader != event) |
3dab77fb PZ |
4485 | leader->pmu->read(leader); |
4486 | ||
b5e58793 | 4487 | values[n++] = perf_event_count(leader); |
3dab77fb | 4488 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4489 | values[n++] = primary_event_id(leader); |
3dab77fb | 4490 | |
76369139 | 4491 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4492 | |
65abc865 | 4493 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4494 | n = 0; |
4495 | ||
6f5ab001 JO |
4496 | if ((sub != event) && |
4497 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4498 | sub->pmu->read(sub); |
4499 | ||
b5e58793 | 4500 | values[n++] = perf_event_count(sub); |
3dab77fb | 4501 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4502 | values[n++] = primary_event_id(sub); |
3dab77fb | 4503 | |
76369139 | 4504 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4505 | } |
4506 | } | |
4507 | ||
eed01528 SE |
4508 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4509 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4510 | ||
3dab77fb | 4511 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4512 | struct perf_event *event) |
3dab77fb | 4513 | { |
e3f3541c | 4514 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4515 | u64 read_format = event->attr.read_format; |
4516 | ||
4517 | /* | |
4518 | * compute total_time_enabled, total_time_running | |
4519 | * based on snapshot values taken when the event | |
4520 | * was last scheduled in. | |
4521 | * | |
4522 | * we cannot simply called update_context_time() | |
4523 | * because of locking issue as we are called in | |
4524 | * NMI context | |
4525 | */ | |
c4794295 | 4526 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4527 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4528 | |
cdd6c482 | 4529 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4530 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4531 | else |
eed01528 | 4532 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4533 | } |
4534 | ||
5622f295 MM |
4535 | void perf_output_sample(struct perf_output_handle *handle, |
4536 | struct perf_event_header *header, | |
4537 | struct perf_sample_data *data, | |
cdd6c482 | 4538 | struct perf_event *event) |
5622f295 MM |
4539 | { |
4540 | u64 sample_type = data->type; | |
4541 | ||
4542 | perf_output_put(handle, *header); | |
4543 | ||
ff3d527c AH |
4544 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4545 | perf_output_put(handle, data->id); | |
4546 | ||
5622f295 MM |
4547 | if (sample_type & PERF_SAMPLE_IP) |
4548 | perf_output_put(handle, data->ip); | |
4549 | ||
4550 | if (sample_type & PERF_SAMPLE_TID) | |
4551 | perf_output_put(handle, data->tid_entry); | |
4552 | ||
4553 | if (sample_type & PERF_SAMPLE_TIME) | |
4554 | perf_output_put(handle, data->time); | |
4555 | ||
4556 | if (sample_type & PERF_SAMPLE_ADDR) | |
4557 | perf_output_put(handle, data->addr); | |
4558 | ||
4559 | if (sample_type & PERF_SAMPLE_ID) | |
4560 | perf_output_put(handle, data->id); | |
4561 | ||
4562 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4563 | perf_output_put(handle, data->stream_id); | |
4564 | ||
4565 | if (sample_type & PERF_SAMPLE_CPU) | |
4566 | perf_output_put(handle, data->cpu_entry); | |
4567 | ||
4568 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4569 | perf_output_put(handle, data->period); | |
4570 | ||
4571 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4572 | perf_output_read(handle, event); |
5622f295 MM |
4573 | |
4574 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4575 | if (data->callchain) { | |
4576 | int size = 1; | |
4577 | ||
4578 | if (data->callchain) | |
4579 | size += data->callchain->nr; | |
4580 | ||
4581 | size *= sizeof(u64); | |
4582 | ||
76369139 | 4583 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4584 | } else { |
4585 | u64 nr = 0; | |
4586 | perf_output_put(handle, nr); | |
4587 | } | |
4588 | } | |
4589 | ||
4590 | if (sample_type & PERF_SAMPLE_RAW) { | |
4591 | if (data->raw) { | |
4592 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4593 | __output_copy(handle, data->raw->data, |
4594 | data->raw->size); | |
5622f295 MM |
4595 | } else { |
4596 | struct { | |
4597 | u32 size; | |
4598 | u32 data; | |
4599 | } raw = { | |
4600 | .size = sizeof(u32), | |
4601 | .data = 0, | |
4602 | }; | |
4603 | perf_output_put(handle, raw); | |
4604 | } | |
4605 | } | |
a7ac67ea | 4606 | |
bce38cd5 SE |
4607 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4608 | if (data->br_stack) { | |
4609 | size_t size; | |
4610 | ||
4611 | size = data->br_stack->nr | |
4612 | * sizeof(struct perf_branch_entry); | |
4613 | ||
4614 | perf_output_put(handle, data->br_stack->nr); | |
4615 | perf_output_copy(handle, data->br_stack->entries, size); | |
4616 | } else { | |
4617 | /* | |
4618 | * we always store at least the value of nr | |
4619 | */ | |
4620 | u64 nr = 0; | |
4621 | perf_output_put(handle, nr); | |
4622 | } | |
4623 | } | |
4018994f JO |
4624 | |
4625 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4626 | u64 abi = data->regs_user.abi; | |
4627 | ||
4628 | /* | |
4629 | * If there are no regs to dump, notice it through | |
4630 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4631 | */ | |
4632 | perf_output_put(handle, abi); | |
4633 | ||
4634 | if (abi) { | |
4635 | u64 mask = event->attr.sample_regs_user; | |
4636 | perf_output_sample_regs(handle, | |
4637 | data->regs_user.regs, | |
4638 | mask); | |
4639 | } | |
4640 | } | |
c5ebcedb | 4641 | |
a5cdd40c | 4642 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
4643 | perf_output_sample_ustack(handle, |
4644 | data->stack_user_size, | |
4645 | data->regs_user.regs); | |
a5cdd40c | 4646 | } |
c3feedf2 AK |
4647 | |
4648 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4649 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4650 | |
4651 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4652 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 4653 | |
fdfbbd07 AK |
4654 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
4655 | perf_output_put(handle, data->txn); | |
4656 | ||
a5cdd40c PZ |
4657 | if (!event->attr.watermark) { |
4658 | int wakeup_events = event->attr.wakeup_events; | |
4659 | ||
4660 | if (wakeup_events) { | |
4661 | struct ring_buffer *rb = handle->rb; | |
4662 | int events = local_inc_return(&rb->events); | |
4663 | ||
4664 | if (events >= wakeup_events) { | |
4665 | local_sub(wakeup_events, &rb->events); | |
4666 | local_inc(&rb->wakeup); | |
4667 | } | |
4668 | } | |
4669 | } | |
5622f295 MM |
4670 | } |
4671 | ||
4672 | void perf_prepare_sample(struct perf_event_header *header, | |
4673 | struct perf_sample_data *data, | |
cdd6c482 | 4674 | struct perf_event *event, |
5622f295 | 4675 | struct pt_regs *regs) |
7b732a75 | 4676 | { |
cdd6c482 | 4677 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4678 | |
cdd6c482 | 4679 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4680 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4681 | |
4682 | header->misc = 0; | |
4683 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4684 | |
c980d109 | 4685 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4686 | |
c320c7b7 | 4687 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4688 | data->ip = perf_instruction_pointer(regs); |
4689 | ||
b23f3325 | 4690 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4691 | int size = 1; |
394ee076 | 4692 | |
e6dab5ff | 4693 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4694 | |
4695 | if (data->callchain) | |
4696 | size += data->callchain->nr; | |
4697 | ||
4698 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4699 | } |
4700 | ||
3a43ce68 | 4701 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4702 | int size = sizeof(u32); |
4703 | ||
4704 | if (data->raw) | |
4705 | size += data->raw->size; | |
4706 | else | |
4707 | size += sizeof(u32); | |
4708 | ||
4709 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4710 | header->size += size; |
7f453c24 | 4711 | } |
bce38cd5 SE |
4712 | |
4713 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4714 | int size = sizeof(u64); /* nr */ | |
4715 | if (data->br_stack) { | |
4716 | size += data->br_stack->nr | |
4717 | * sizeof(struct perf_branch_entry); | |
4718 | } | |
4719 | header->size += size; | |
4720 | } | |
4018994f JO |
4721 | |
4722 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4723 | /* regs dump ABI info */ | |
4724 | int size = sizeof(u64); | |
4725 | ||
4726 | perf_sample_regs_user(&data->regs_user, regs); | |
4727 | ||
4728 | if (data->regs_user.regs) { | |
4729 | u64 mask = event->attr.sample_regs_user; | |
4730 | size += hweight64(mask) * sizeof(u64); | |
4731 | } | |
4732 | ||
4733 | header->size += size; | |
4734 | } | |
c5ebcedb JO |
4735 | |
4736 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4737 | /* | |
4738 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4739 | * processed as the last one or have additional check added | |
4740 | * in case new sample type is added, because we could eat | |
4741 | * up the rest of the sample size. | |
4742 | */ | |
4743 | struct perf_regs_user *uregs = &data->regs_user; | |
4744 | u16 stack_size = event->attr.sample_stack_user; | |
4745 | u16 size = sizeof(u64); | |
4746 | ||
4747 | if (!uregs->abi) | |
4748 | perf_sample_regs_user(uregs, regs); | |
4749 | ||
4750 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4751 | uregs->regs); | |
4752 | ||
4753 | /* | |
4754 | * If there is something to dump, add space for the dump | |
4755 | * itself and for the field that tells the dynamic size, | |
4756 | * which is how many have been actually dumped. | |
4757 | */ | |
4758 | if (stack_size) | |
4759 | size += sizeof(u64) + stack_size; | |
4760 | ||
4761 | data->stack_user_size = stack_size; | |
4762 | header->size += size; | |
4763 | } | |
5622f295 | 4764 | } |
7f453c24 | 4765 | |
a8b0ca17 | 4766 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4767 | struct perf_sample_data *data, |
4768 | struct pt_regs *regs) | |
4769 | { | |
4770 | struct perf_output_handle handle; | |
4771 | struct perf_event_header header; | |
689802b2 | 4772 | |
927c7a9e FW |
4773 | /* protect the callchain buffers */ |
4774 | rcu_read_lock(); | |
4775 | ||
cdd6c482 | 4776 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4777 | |
a7ac67ea | 4778 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4779 | goto exit; |
0322cd6e | 4780 | |
cdd6c482 | 4781 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4782 | |
8a057d84 | 4783 | perf_output_end(&handle); |
927c7a9e FW |
4784 | |
4785 | exit: | |
4786 | rcu_read_unlock(); | |
0322cd6e PZ |
4787 | } |
4788 | ||
38b200d6 | 4789 | /* |
cdd6c482 | 4790 | * read event_id |
38b200d6 PZ |
4791 | */ |
4792 | ||
4793 | struct perf_read_event { | |
4794 | struct perf_event_header header; | |
4795 | ||
4796 | u32 pid; | |
4797 | u32 tid; | |
38b200d6 PZ |
4798 | }; |
4799 | ||
4800 | static void | |
cdd6c482 | 4801 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4802 | struct task_struct *task) |
4803 | { | |
4804 | struct perf_output_handle handle; | |
c980d109 | 4805 | struct perf_sample_data sample; |
dfc65094 | 4806 | struct perf_read_event read_event = { |
38b200d6 | 4807 | .header = { |
cdd6c482 | 4808 | .type = PERF_RECORD_READ, |
38b200d6 | 4809 | .misc = 0, |
c320c7b7 | 4810 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4811 | }, |
cdd6c482 IM |
4812 | .pid = perf_event_pid(event, task), |
4813 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4814 | }; |
3dab77fb | 4815 | int ret; |
38b200d6 | 4816 | |
c980d109 | 4817 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4818 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4819 | if (ret) |
4820 | return; | |
4821 | ||
dfc65094 | 4822 | perf_output_put(&handle, read_event); |
cdd6c482 | 4823 | perf_output_read(&handle, event); |
c980d109 | 4824 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4825 | |
38b200d6 PZ |
4826 | perf_output_end(&handle); |
4827 | } | |
4828 | ||
52d857a8 JO |
4829 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
4830 | ||
4831 | static void | |
4832 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
4833 | perf_event_aux_output_cb output, |
4834 | void *data) | |
4835 | { | |
4836 | struct perf_event *event; | |
4837 | ||
4838 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4839 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4840 | continue; | |
4841 | if (!event_filter_match(event)) | |
4842 | continue; | |
67516844 | 4843 | output(event, data); |
52d857a8 JO |
4844 | } |
4845 | } | |
4846 | ||
4847 | static void | |
67516844 | 4848 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
4849 | struct perf_event_context *task_ctx) |
4850 | { | |
4851 | struct perf_cpu_context *cpuctx; | |
4852 | struct perf_event_context *ctx; | |
4853 | struct pmu *pmu; | |
4854 | int ctxn; | |
4855 | ||
4856 | rcu_read_lock(); | |
4857 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4858 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4859 | if (cpuctx->unique_pmu != pmu) | |
4860 | goto next; | |
67516844 | 4861 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
4862 | if (task_ctx) |
4863 | goto next; | |
4864 | ctxn = pmu->task_ctx_nr; | |
4865 | if (ctxn < 0) | |
4866 | goto next; | |
4867 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4868 | if (ctx) | |
67516844 | 4869 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
4870 | next: |
4871 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4872 | } | |
4873 | ||
4874 | if (task_ctx) { | |
4875 | preempt_disable(); | |
67516844 | 4876 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
4877 | preempt_enable(); |
4878 | } | |
4879 | rcu_read_unlock(); | |
4880 | } | |
4881 | ||
60313ebe | 4882 | /* |
9f498cc5 PZ |
4883 | * task tracking -- fork/exit |
4884 | * | |
13d7a241 | 4885 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
4886 | */ |
4887 | ||
9f498cc5 | 4888 | struct perf_task_event { |
3a80b4a3 | 4889 | struct task_struct *task; |
cdd6c482 | 4890 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4891 | |
4892 | struct { | |
4893 | struct perf_event_header header; | |
4894 | ||
4895 | u32 pid; | |
4896 | u32 ppid; | |
9f498cc5 PZ |
4897 | u32 tid; |
4898 | u32 ptid; | |
393b2ad8 | 4899 | u64 time; |
cdd6c482 | 4900 | } event_id; |
60313ebe PZ |
4901 | }; |
4902 | ||
67516844 JO |
4903 | static int perf_event_task_match(struct perf_event *event) |
4904 | { | |
13d7a241 SE |
4905 | return event->attr.comm || event->attr.mmap || |
4906 | event->attr.mmap2 || event->attr.mmap_data || | |
4907 | event->attr.task; | |
67516844 JO |
4908 | } |
4909 | ||
cdd6c482 | 4910 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4911 | void *data) |
60313ebe | 4912 | { |
52d857a8 | 4913 | struct perf_task_event *task_event = data; |
60313ebe | 4914 | struct perf_output_handle handle; |
c980d109 | 4915 | struct perf_sample_data sample; |
9f498cc5 | 4916 | struct task_struct *task = task_event->task; |
c980d109 | 4917 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4918 | |
67516844 JO |
4919 | if (!perf_event_task_match(event)) |
4920 | return; | |
4921 | ||
c980d109 | 4922 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4923 | |
c980d109 | 4924 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4925 | task_event->event_id.header.size); |
ef60777c | 4926 | if (ret) |
c980d109 | 4927 | goto out; |
60313ebe | 4928 | |
cdd6c482 IM |
4929 | task_event->event_id.pid = perf_event_pid(event, task); |
4930 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4931 | |
cdd6c482 IM |
4932 | task_event->event_id.tid = perf_event_tid(event, task); |
4933 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4934 | |
cdd6c482 | 4935 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4936 | |
c980d109 ACM |
4937 | perf_event__output_id_sample(event, &handle, &sample); |
4938 | ||
60313ebe | 4939 | perf_output_end(&handle); |
c980d109 ACM |
4940 | out: |
4941 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4942 | } |
4943 | ||
cdd6c482 IM |
4944 | static void perf_event_task(struct task_struct *task, |
4945 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4946 | int new) |
60313ebe | 4947 | { |
9f498cc5 | 4948 | struct perf_task_event task_event; |
60313ebe | 4949 | |
cdd6c482 IM |
4950 | if (!atomic_read(&nr_comm_events) && |
4951 | !atomic_read(&nr_mmap_events) && | |
4952 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4953 | return; |
4954 | ||
9f498cc5 | 4955 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4956 | .task = task, |
4957 | .task_ctx = task_ctx, | |
cdd6c482 | 4958 | .event_id = { |
60313ebe | 4959 | .header = { |
cdd6c482 | 4960 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4961 | .misc = 0, |
cdd6c482 | 4962 | .size = sizeof(task_event.event_id), |
60313ebe | 4963 | }, |
573402db PZ |
4964 | /* .pid */ |
4965 | /* .ppid */ | |
9f498cc5 PZ |
4966 | /* .tid */ |
4967 | /* .ptid */ | |
6f93d0a7 | 4968 | .time = perf_clock(), |
60313ebe PZ |
4969 | }, |
4970 | }; | |
4971 | ||
67516844 | 4972 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
4973 | &task_event, |
4974 | task_ctx); | |
9f498cc5 PZ |
4975 | } |
4976 | ||
cdd6c482 | 4977 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4978 | { |
cdd6c482 | 4979 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4980 | } |
4981 | ||
8d1b2d93 PZ |
4982 | /* |
4983 | * comm tracking | |
4984 | */ | |
4985 | ||
4986 | struct perf_comm_event { | |
22a4f650 IM |
4987 | struct task_struct *task; |
4988 | char *comm; | |
8d1b2d93 PZ |
4989 | int comm_size; |
4990 | ||
4991 | struct { | |
4992 | struct perf_event_header header; | |
4993 | ||
4994 | u32 pid; | |
4995 | u32 tid; | |
cdd6c482 | 4996 | } event_id; |
8d1b2d93 PZ |
4997 | }; |
4998 | ||
67516844 JO |
4999 | static int perf_event_comm_match(struct perf_event *event) |
5000 | { | |
5001 | return event->attr.comm; | |
5002 | } | |
5003 | ||
cdd6c482 | 5004 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5005 | void *data) |
8d1b2d93 | 5006 | { |
52d857a8 | 5007 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5008 | struct perf_output_handle handle; |
c980d109 | 5009 | struct perf_sample_data sample; |
cdd6c482 | 5010 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5011 | int ret; |
5012 | ||
67516844 JO |
5013 | if (!perf_event_comm_match(event)) |
5014 | return; | |
5015 | ||
c980d109 ACM |
5016 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5017 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5018 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5019 | |
5020 | if (ret) | |
c980d109 | 5021 | goto out; |
8d1b2d93 | 5022 | |
cdd6c482 IM |
5023 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5024 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5025 | |
cdd6c482 | 5026 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5027 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5028 | comm_event->comm_size); |
c980d109 ACM |
5029 | |
5030 | perf_event__output_id_sample(event, &handle, &sample); | |
5031 | ||
8d1b2d93 | 5032 | perf_output_end(&handle); |
c980d109 ACM |
5033 | out: |
5034 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5035 | } |
5036 | ||
cdd6c482 | 5037 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5038 | { |
413ee3b4 | 5039 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5040 | unsigned int size; |
8d1b2d93 | 5041 | |
413ee3b4 | 5042 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5043 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5044 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5045 | |
5046 | comm_event->comm = comm; | |
5047 | comm_event->comm_size = size; | |
5048 | ||
cdd6c482 | 5049 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5050 | |
67516844 | 5051 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5052 | comm_event, |
5053 | NULL); | |
8d1b2d93 PZ |
5054 | } |
5055 | ||
cdd6c482 | 5056 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 5057 | { |
9ee318a7 | 5058 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
5059 | struct perf_event_context *ctx; |
5060 | int ctxn; | |
9ee318a7 | 5061 | |
c79aa0d9 | 5062 | rcu_read_lock(); |
8dc85d54 PZ |
5063 | for_each_task_context_nr(ctxn) { |
5064 | ctx = task->perf_event_ctxp[ctxn]; | |
5065 | if (!ctx) | |
5066 | continue; | |
9ee318a7 | 5067 | |
8dc85d54 PZ |
5068 | perf_event_enable_on_exec(ctx); |
5069 | } | |
c79aa0d9 | 5070 | rcu_read_unlock(); |
9ee318a7 | 5071 | |
cdd6c482 | 5072 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5073 | return; |
a63eaf34 | 5074 | |
9ee318a7 | 5075 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5076 | .task = task, |
573402db PZ |
5077 | /* .comm */ |
5078 | /* .comm_size */ | |
cdd6c482 | 5079 | .event_id = { |
573402db | 5080 | .header = { |
cdd6c482 | 5081 | .type = PERF_RECORD_COMM, |
573402db PZ |
5082 | .misc = 0, |
5083 | /* .size */ | |
5084 | }, | |
5085 | /* .pid */ | |
5086 | /* .tid */ | |
8d1b2d93 PZ |
5087 | }, |
5088 | }; | |
5089 | ||
cdd6c482 | 5090 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5091 | } |
5092 | ||
0a4a9391 PZ |
5093 | /* |
5094 | * mmap tracking | |
5095 | */ | |
5096 | ||
5097 | struct perf_mmap_event { | |
089dd79d PZ |
5098 | struct vm_area_struct *vma; |
5099 | ||
5100 | const char *file_name; | |
5101 | int file_size; | |
13d7a241 SE |
5102 | int maj, min; |
5103 | u64 ino; | |
5104 | u64 ino_generation; | |
0a4a9391 PZ |
5105 | |
5106 | struct { | |
5107 | struct perf_event_header header; | |
5108 | ||
5109 | u32 pid; | |
5110 | u32 tid; | |
5111 | u64 start; | |
5112 | u64 len; | |
5113 | u64 pgoff; | |
cdd6c482 | 5114 | } event_id; |
0a4a9391 PZ |
5115 | }; |
5116 | ||
67516844 JO |
5117 | static int perf_event_mmap_match(struct perf_event *event, |
5118 | void *data) | |
5119 | { | |
5120 | struct perf_mmap_event *mmap_event = data; | |
5121 | struct vm_area_struct *vma = mmap_event->vma; | |
5122 | int executable = vma->vm_flags & VM_EXEC; | |
5123 | ||
5124 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5125 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5126 | } |
5127 | ||
cdd6c482 | 5128 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5129 | void *data) |
0a4a9391 | 5130 | { |
52d857a8 | 5131 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5132 | struct perf_output_handle handle; |
c980d109 | 5133 | struct perf_sample_data sample; |
cdd6c482 | 5134 | int size = mmap_event->event_id.header.size; |
c980d109 | 5135 | int ret; |
0a4a9391 | 5136 | |
67516844 JO |
5137 | if (!perf_event_mmap_match(event, data)) |
5138 | return; | |
5139 | ||
13d7a241 SE |
5140 | if (event->attr.mmap2) { |
5141 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5142 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5143 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5144 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5145 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
13d7a241 SE |
5146 | } |
5147 | ||
c980d109 ACM |
5148 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5149 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5150 | mmap_event->event_id.header.size); |
0a4a9391 | 5151 | if (ret) |
c980d109 | 5152 | goto out; |
0a4a9391 | 5153 | |
cdd6c482 IM |
5154 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5155 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5156 | |
cdd6c482 | 5157 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5158 | |
5159 | if (event->attr.mmap2) { | |
5160 | perf_output_put(&handle, mmap_event->maj); | |
5161 | perf_output_put(&handle, mmap_event->min); | |
5162 | perf_output_put(&handle, mmap_event->ino); | |
5163 | perf_output_put(&handle, mmap_event->ino_generation); | |
5164 | } | |
5165 | ||
76369139 | 5166 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5167 | mmap_event->file_size); |
c980d109 ACM |
5168 | |
5169 | perf_event__output_id_sample(event, &handle, &sample); | |
5170 | ||
78d613eb | 5171 | perf_output_end(&handle); |
c980d109 ACM |
5172 | out: |
5173 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5174 | } |
5175 | ||
cdd6c482 | 5176 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5177 | { |
089dd79d PZ |
5178 | struct vm_area_struct *vma = mmap_event->vma; |
5179 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5180 | int maj = 0, min = 0; |
5181 | u64 ino = 0, gen = 0; | |
0a4a9391 PZ |
5182 | unsigned int size; |
5183 | char tmp[16]; | |
5184 | char *buf = NULL; | |
2c42cfbf | 5185 | char *name; |
413ee3b4 | 5186 | |
0a4a9391 | 5187 | if (file) { |
13d7a241 SE |
5188 | struct inode *inode; |
5189 | dev_t dev; | |
3ea2f2b9 | 5190 | |
2c42cfbf | 5191 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5192 | if (!buf) { |
c7e548b4 ON |
5193 | name = "//enomem"; |
5194 | goto cpy_name; | |
0a4a9391 | 5195 | } |
413ee3b4 | 5196 | /* |
3ea2f2b9 | 5197 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5198 | * need to add enough zero bytes after the string to handle |
5199 | * the 64bit alignment we do later. | |
5200 | */ | |
3ea2f2b9 | 5201 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5202 | if (IS_ERR(name)) { |
c7e548b4 ON |
5203 | name = "//toolong"; |
5204 | goto cpy_name; | |
0a4a9391 | 5205 | } |
13d7a241 SE |
5206 | inode = file_inode(vma->vm_file); |
5207 | dev = inode->i_sb->s_dev; | |
5208 | ino = inode->i_ino; | |
5209 | gen = inode->i_generation; | |
5210 | maj = MAJOR(dev); | |
5211 | min = MINOR(dev); | |
c7e548b4 | 5212 | goto got_name; |
0a4a9391 | 5213 | } else { |
2c42cfbf | 5214 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5215 | if (name) |
5216 | goto cpy_name; | |
089dd79d | 5217 | |
32c5fb7e | 5218 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5219 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5220 | name = "[heap]"; |
5221 | goto cpy_name; | |
32c5fb7e ON |
5222 | } |
5223 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5224 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5225 | name = "[stack]"; |
5226 | goto cpy_name; | |
089dd79d PZ |
5227 | } |
5228 | ||
c7e548b4 ON |
5229 | name = "//anon"; |
5230 | goto cpy_name; | |
0a4a9391 PZ |
5231 | } |
5232 | ||
c7e548b4 ON |
5233 | cpy_name: |
5234 | strlcpy(tmp, name, sizeof(tmp)); | |
5235 | name = tmp; | |
0a4a9391 | 5236 | got_name: |
2c42cfbf PZ |
5237 | /* |
5238 | * Since our buffer works in 8 byte units we need to align our string | |
5239 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5240 | * zero'd out to avoid leaking random bits to userspace. | |
5241 | */ | |
5242 | size = strlen(name)+1; | |
5243 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5244 | name[size++] = '\0'; | |
0a4a9391 PZ |
5245 | |
5246 | mmap_event->file_name = name; | |
5247 | mmap_event->file_size = size; | |
13d7a241 SE |
5248 | mmap_event->maj = maj; |
5249 | mmap_event->min = min; | |
5250 | mmap_event->ino = ino; | |
5251 | mmap_event->ino_generation = gen; | |
0a4a9391 | 5252 | |
2fe85427 SE |
5253 | if (!(vma->vm_flags & VM_EXEC)) |
5254 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5255 | ||
cdd6c482 | 5256 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5257 | |
67516844 | 5258 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5259 | mmap_event, |
5260 | NULL); | |
665c2142 | 5261 | |
0a4a9391 PZ |
5262 | kfree(buf); |
5263 | } | |
5264 | ||
3af9e859 | 5265 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5266 | { |
9ee318a7 PZ |
5267 | struct perf_mmap_event mmap_event; |
5268 | ||
cdd6c482 | 5269 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5270 | return; |
5271 | ||
5272 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5273 | .vma = vma, |
573402db PZ |
5274 | /* .file_name */ |
5275 | /* .file_size */ | |
cdd6c482 | 5276 | .event_id = { |
573402db | 5277 | .header = { |
cdd6c482 | 5278 | .type = PERF_RECORD_MMAP, |
39447b38 | 5279 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5280 | /* .size */ |
5281 | }, | |
5282 | /* .pid */ | |
5283 | /* .tid */ | |
089dd79d PZ |
5284 | .start = vma->vm_start, |
5285 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5286 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5287 | }, |
13d7a241 SE |
5288 | /* .maj (attr_mmap2 only) */ |
5289 | /* .min (attr_mmap2 only) */ | |
5290 | /* .ino (attr_mmap2 only) */ | |
5291 | /* .ino_generation (attr_mmap2 only) */ | |
0a4a9391 PZ |
5292 | }; |
5293 | ||
cdd6c482 | 5294 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5295 | } |
5296 | ||
a78ac325 PZ |
5297 | /* |
5298 | * IRQ throttle logging | |
5299 | */ | |
5300 | ||
cdd6c482 | 5301 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5302 | { |
5303 | struct perf_output_handle handle; | |
c980d109 | 5304 | struct perf_sample_data sample; |
a78ac325 PZ |
5305 | int ret; |
5306 | ||
5307 | struct { | |
5308 | struct perf_event_header header; | |
5309 | u64 time; | |
cca3f454 | 5310 | u64 id; |
7f453c24 | 5311 | u64 stream_id; |
a78ac325 PZ |
5312 | } throttle_event = { |
5313 | .header = { | |
cdd6c482 | 5314 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5315 | .misc = 0, |
5316 | .size = sizeof(throttle_event), | |
5317 | }, | |
def0a9b2 | 5318 | .time = perf_clock(), |
cdd6c482 IM |
5319 | .id = primary_event_id(event), |
5320 | .stream_id = event->id, | |
a78ac325 PZ |
5321 | }; |
5322 | ||
966ee4d6 | 5323 | if (enable) |
cdd6c482 | 5324 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5325 | |
c980d109 ACM |
5326 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5327 | ||
5328 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5329 | throttle_event.header.size); |
a78ac325 PZ |
5330 | if (ret) |
5331 | return; | |
5332 | ||
5333 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5334 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5335 | perf_output_end(&handle); |
5336 | } | |
5337 | ||
f6c7d5fe | 5338 | /* |
cdd6c482 | 5339 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5340 | */ |
5341 | ||
a8b0ca17 | 5342 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5343 | int throttle, struct perf_sample_data *data, |
5344 | struct pt_regs *regs) | |
f6c7d5fe | 5345 | { |
cdd6c482 IM |
5346 | int events = atomic_read(&event->event_limit); |
5347 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5348 | u64 seq; |
79f14641 PZ |
5349 | int ret = 0; |
5350 | ||
96398826 PZ |
5351 | /* |
5352 | * Non-sampling counters might still use the PMI to fold short | |
5353 | * hardware counters, ignore those. | |
5354 | */ | |
5355 | if (unlikely(!is_sampling_event(event))) | |
5356 | return 0; | |
5357 | ||
e050e3f0 SE |
5358 | seq = __this_cpu_read(perf_throttled_seq); |
5359 | if (seq != hwc->interrupts_seq) { | |
5360 | hwc->interrupts_seq = seq; | |
5361 | hwc->interrupts = 1; | |
5362 | } else { | |
5363 | hwc->interrupts++; | |
5364 | if (unlikely(throttle | |
5365 | && hwc->interrupts >= max_samples_per_tick)) { | |
5366 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5367 | hwc->interrupts = MAX_INTERRUPTS; |
5368 | perf_log_throttle(event, 0); | |
d84153d6 | 5369 | tick_nohz_full_kick(); |
a78ac325 PZ |
5370 | ret = 1; |
5371 | } | |
e050e3f0 | 5372 | } |
60db5e09 | 5373 | |
cdd6c482 | 5374 | if (event->attr.freq) { |
def0a9b2 | 5375 | u64 now = perf_clock(); |
abd50713 | 5376 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5377 | |
abd50713 | 5378 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5379 | |
abd50713 | 5380 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5381 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5382 | } |
5383 | ||
2023b359 PZ |
5384 | /* |
5385 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5386 | * events |
2023b359 PZ |
5387 | */ |
5388 | ||
cdd6c482 IM |
5389 | event->pending_kill = POLL_IN; |
5390 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5391 | ret = 1; |
cdd6c482 | 5392 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5393 | event->pending_disable = 1; |
5394 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5395 | } |
5396 | ||
453f19ee | 5397 | if (event->overflow_handler) |
a8b0ca17 | 5398 | event->overflow_handler(event, data, regs); |
453f19ee | 5399 | else |
a8b0ca17 | 5400 | perf_event_output(event, data, regs); |
453f19ee | 5401 | |
f506b3dc | 5402 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5403 | event->pending_wakeup = 1; |
5404 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5405 | } |
5406 | ||
79f14641 | 5407 | return ret; |
f6c7d5fe PZ |
5408 | } |
5409 | ||
a8b0ca17 | 5410 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5411 | struct perf_sample_data *data, |
5412 | struct pt_regs *regs) | |
850bc73f | 5413 | { |
a8b0ca17 | 5414 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5415 | } |
5416 | ||
15dbf27c | 5417 | /* |
cdd6c482 | 5418 | * Generic software event infrastructure |
15dbf27c PZ |
5419 | */ |
5420 | ||
b28ab83c PZ |
5421 | struct swevent_htable { |
5422 | struct swevent_hlist *swevent_hlist; | |
5423 | struct mutex hlist_mutex; | |
5424 | int hlist_refcount; | |
5425 | ||
5426 | /* Recursion avoidance in each contexts */ | |
5427 | int recursion[PERF_NR_CONTEXTS]; | |
5428 | }; | |
5429 | ||
5430 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5431 | ||
7b4b6658 | 5432 | /* |
cdd6c482 IM |
5433 | * We directly increment event->count and keep a second value in |
5434 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5435 | * is kept in the range [-sample_period, 0] so that we can use the |
5436 | * sign as trigger. | |
5437 | */ | |
5438 | ||
ab573844 | 5439 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5440 | { |
cdd6c482 | 5441 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5442 | u64 period = hwc->last_period; |
5443 | u64 nr, offset; | |
5444 | s64 old, val; | |
5445 | ||
5446 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5447 | |
5448 | again: | |
e7850595 | 5449 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5450 | if (val < 0) |
5451 | return 0; | |
15dbf27c | 5452 | |
7b4b6658 PZ |
5453 | nr = div64_u64(period + val, period); |
5454 | offset = nr * period; | |
5455 | val -= offset; | |
e7850595 | 5456 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5457 | goto again; |
15dbf27c | 5458 | |
7b4b6658 | 5459 | return nr; |
15dbf27c PZ |
5460 | } |
5461 | ||
0cff784a | 5462 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5463 | struct perf_sample_data *data, |
5622f295 | 5464 | struct pt_regs *regs) |
15dbf27c | 5465 | { |
cdd6c482 | 5466 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5467 | int throttle = 0; |
15dbf27c | 5468 | |
0cff784a PZ |
5469 | if (!overflow) |
5470 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5471 | |
7b4b6658 PZ |
5472 | if (hwc->interrupts == MAX_INTERRUPTS) |
5473 | return; | |
15dbf27c | 5474 | |
7b4b6658 | 5475 | for (; overflow; overflow--) { |
a8b0ca17 | 5476 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5477 | data, regs)) { |
7b4b6658 PZ |
5478 | /* |
5479 | * We inhibit the overflow from happening when | |
5480 | * hwc->interrupts == MAX_INTERRUPTS. | |
5481 | */ | |
5482 | break; | |
5483 | } | |
cf450a73 | 5484 | throttle = 1; |
7b4b6658 | 5485 | } |
15dbf27c PZ |
5486 | } |
5487 | ||
a4eaf7f1 | 5488 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5489 | struct perf_sample_data *data, |
5622f295 | 5490 | struct pt_regs *regs) |
7b4b6658 | 5491 | { |
cdd6c482 | 5492 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5493 | |
e7850595 | 5494 | local64_add(nr, &event->count); |
d6d020e9 | 5495 | |
0cff784a PZ |
5496 | if (!regs) |
5497 | return; | |
5498 | ||
6c7e550f | 5499 | if (!is_sampling_event(event)) |
7b4b6658 | 5500 | return; |
d6d020e9 | 5501 | |
5d81e5cf AV |
5502 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5503 | data->period = nr; | |
5504 | return perf_swevent_overflow(event, 1, data, regs); | |
5505 | } else | |
5506 | data->period = event->hw.last_period; | |
5507 | ||
0cff784a | 5508 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5509 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5510 | |
e7850595 | 5511 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5512 | return; |
df1a132b | 5513 | |
a8b0ca17 | 5514 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5515 | } |
5516 | ||
f5ffe02e FW |
5517 | static int perf_exclude_event(struct perf_event *event, |
5518 | struct pt_regs *regs) | |
5519 | { | |
a4eaf7f1 | 5520 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5521 | return 1; |
a4eaf7f1 | 5522 | |
f5ffe02e FW |
5523 | if (regs) { |
5524 | if (event->attr.exclude_user && user_mode(regs)) | |
5525 | return 1; | |
5526 | ||
5527 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5528 | return 1; | |
5529 | } | |
5530 | ||
5531 | return 0; | |
5532 | } | |
5533 | ||
cdd6c482 | 5534 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5535 | enum perf_type_id type, |
6fb2915d LZ |
5536 | u32 event_id, |
5537 | struct perf_sample_data *data, | |
5538 | struct pt_regs *regs) | |
15dbf27c | 5539 | { |
cdd6c482 | 5540 | if (event->attr.type != type) |
a21ca2ca | 5541 | return 0; |
f5ffe02e | 5542 | |
cdd6c482 | 5543 | if (event->attr.config != event_id) |
15dbf27c PZ |
5544 | return 0; |
5545 | ||
f5ffe02e FW |
5546 | if (perf_exclude_event(event, regs)) |
5547 | return 0; | |
15dbf27c PZ |
5548 | |
5549 | return 1; | |
5550 | } | |
5551 | ||
76e1d904 FW |
5552 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5553 | { | |
5554 | u64 val = event_id | (type << 32); | |
5555 | ||
5556 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5557 | } | |
5558 | ||
49f135ed FW |
5559 | static inline struct hlist_head * |
5560 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5561 | { |
49f135ed FW |
5562 | u64 hash = swevent_hash(type, event_id); |
5563 | ||
5564 | return &hlist->heads[hash]; | |
5565 | } | |
76e1d904 | 5566 | |
49f135ed FW |
5567 | /* For the read side: events when they trigger */ |
5568 | static inline struct hlist_head * | |
b28ab83c | 5569 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5570 | { |
5571 | struct swevent_hlist *hlist; | |
76e1d904 | 5572 | |
b28ab83c | 5573 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5574 | if (!hlist) |
5575 | return NULL; | |
5576 | ||
49f135ed FW |
5577 | return __find_swevent_head(hlist, type, event_id); |
5578 | } | |
5579 | ||
5580 | /* For the event head insertion and removal in the hlist */ | |
5581 | static inline struct hlist_head * | |
b28ab83c | 5582 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5583 | { |
5584 | struct swevent_hlist *hlist; | |
5585 | u32 event_id = event->attr.config; | |
5586 | u64 type = event->attr.type; | |
5587 | ||
5588 | /* | |
5589 | * Event scheduling is always serialized against hlist allocation | |
5590 | * and release. Which makes the protected version suitable here. | |
5591 | * The context lock guarantees that. | |
5592 | */ | |
b28ab83c | 5593 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5594 | lockdep_is_held(&event->ctx->lock)); |
5595 | if (!hlist) | |
5596 | return NULL; | |
5597 | ||
5598 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5599 | } |
5600 | ||
5601 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5602 | u64 nr, |
76e1d904 FW |
5603 | struct perf_sample_data *data, |
5604 | struct pt_regs *regs) | |
15dbf27c | 5605 | { |
b28ab83c | 5606 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5607 | struct perf_event *event; |
76e1d904 | 5608 | struct hlist_head *head; |
15dbf27c | 5609 | |
76e1d904 | 5610 | rcu_read_lock(); |
b28ab83c | 5611 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5612 | if (!head) |
5613 | goto end; | |
5614 | ||
b67bfe0d | 5615 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5616 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5617 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5618 | } |
76e1d904 FW |
5619 | end: |
5620 | rcu_read_unlock(); | |
15dbf27c PZ |
5621 | } |
5622 | ||
4ed7c92d | 5623 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5624 | { |
b28ab83c | 5625 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5626 | |
b28ab83c | 5627 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5628 | } |
645e8cc0 | 5629 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5630 | |
fa9f90be | 5631 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5632 | { |
b28ab83c | 5633 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5634 | |
b28ab83c | 5635 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5636 | } |
15dbf27c | 5637 | |
a8b0ca17 | 5638 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5639 | { |
a4234bfc | 5640 | struct perf_sample_data data; |
4ed7c92d PZ |
5641 | int rctx; |
5642 | ||
1c024eca | 5643 | preempt_disable_notrace(); |
4ed7c92d PZ |
5644 | rctx = perf_swevent_get_recursion_context(); |
5645 | if (rctx < 0) | |
5646 | return; | |
a4234bfc | 5647 | |
fd0d000b | 5648 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5649 | |
a8b0ca17 | 5650 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5651 | |
5652 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5653 | preempt_enable_notrace(); |
b8e83514 PZ |
5654 | } |
5655 | ||
cdd6c482 | 5656 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5657 | { |
15dbf27c PZ |
5658 | } |
5659 | ||
a4eaf7f1 | 5660 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5661 | { |
b28ab83c | 5662 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5663 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5664 | struct hlist_head *head; |
5665 | ||
6c7e550f | 5666 | if (is_sampling_event(event)) { |
7b4b6658 | 5667 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5668 | perf_swevent_set_period(event); |
7b4b6658 | 5669 | } |
76e1d904 | 5670 | |
a4eaf7f1 PZ |
5671 | hwc->state = !(flags & PERF_EF_START); |
5672 | ||
b28ab83c | 5673 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5674 | if (WARN_ON_ONCE(!head)) |
5675 | return -EINVAL; | |
5676 | ||
5677 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5678 | ||
15dbf27c PZ |
5679 | return 0; |
5680 | } | |
5681 | ||
a4eaf7f1 | 5682 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5683 | { |
76e1d904 | 5684 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5685 | } |
5686 | ||
a4eaf7f1 | 5687 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5688 | { |
a4eaf7f1 | 5689 | event->hw.state = 0; |
d6d020e9 | 5690 | } |
aa9c4c0f | 5691 | |
a4eaf7f1 | 5692 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5693 | { |
a4eaf7f1 | 5694 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5695 | } |
5696 | ||
49f135ed FW |
5697 | /* Deref the hlist from the update side */ |
5698 | static inline struct swevent_hlist * | |
b28ab83c | 5699 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5700 | { |
b28ab83c PZ |
5701 | return rcu_dereference_protected(swhash->swevent_hlist, |
5702 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5703 | } |
5704 | ||
b28ab83c | 5705 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5706 | { |
b28ab83c | 5707 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5708 | |
49f135ed | 5709 | if (!hlist) |
76e1d904 FW |
5710 | return; |
5711 | ||
b28ab83c | 5712 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5713 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5714 | } |
5715 | ||
5716 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5717 | { | |
b28ab83c | 5718 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5719 | |
b28ab83c | 5720 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5721 | |
b28ab83c PZ |
5722 | if (!--swhash->hlist_refcount) |
5723 | swevent_hlist_release(swhash); | |
76e1d904 | 5724 | |
b28ab83c | 5725 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5726 | } |
5727 | ||
5728 | static void swevent_hlist_put(struct perf_event *event) | |
5729 | { | |
5730 | int cpu; | |
5731 | ||
76e1d904 FW |
5732 | for_each_possible_cpu(cpu) |
5733 | swevent_hlist_put_cpu(event, cpu); | |
5734 | } | |
5735 | ||
5736 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5737 | { | |
b28ab83c | 5738 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5739 | int err = 0; |
5740 | ||
b28ab83c | 5741 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5742 | |
b28ab83c | 5743 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5744 | struct swevent_hlist *hlist; |
5745 | ||
5746 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5747 | if (!hlist) { | |
5748 | err = -ENOMEM; | |
5749 | goto exit; | |
5750 | } | |
b28ab83c | 5751 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5752 | } |
b28ab83c | 5753 | swhash->hlist_refcount++; |
9ed6060d | 5754 | exit: |
b28ab83c | 5755 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5756 | |
5757 | return err; | |
5758 | } | |
5759 | ||
5760 | static int swevent_hlist_get(struct perf_event *event) | |
5761 | { | |
5762 | int err; | |
5763 | int cpu, failed_cpu; | |
5764 | ||
76e1d904 FW |
5765 | get_online_cpus(); |
5766 | for_each_possible_cpu(cpu) { | |
5767 | err = swevent_hlist_get_cpu(event, cpu); | |
5768 | if (err) { | |
5769 | failed_cpu = cpu; | |
5770 | goto fail; | |
5771 | } | |
5772 | } | |
5773 | put_online_cpus(); | |
5774 | ||
5775 | return 0; | |
9ed6060d | 5776 | fail: |
76e1d904 FW |
5777 | for_each_possible_cpu(cpu) { |
5778 | if (cpu == failed_cpu) | |
5779 | break; | |
5780 | swevent_hlist_put_cpu(event, cpu); | |
5781 | } | |
5782 | ||
5783 | put_online_cpus(); | |
5784 | return err; | |
5785 | } | |
5786 | ||
c5905afb | 5787 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5788 | |
b0a873eb PZ |
5789 | static void sw_perf_event_destroy(struct perf_event *event) |
5790 | { | |
5791 | u64 event_id = event->attr.config; | |
95476b64 | 5792 | |
b0a873eb PZ |
5793 | WARN_ON(event->parent); |
5794 | ||
c5905afb | 5795 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5796 | swevent_hlist_put(event); |
5797 | } | |
5798 | ||
5799 | static int perf_swevent_init(struct perf_event *event) | |
5800 | { | |
8176cced | 5801 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5802 | |
5803 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5804 | return -ENOENT; | |
5805 | ||
2481c5fa SE |
5806 | /* |
5807 | * no branch sampling for software events | |
5808 | */ | |
5809 | if (has_branch_stack(event)) | |
5810 | return -EOPNOTSUPP; | |
5811 | ||
b0a873eb PZ |
5812 | switch (event_id) { |
5813 | case PERF_COUNT_SW_CPU_CLOCK: | |
5814 | case PERF_COUNT_SW_TASK_CLOCK: | |
5815 | return -ENOENT; | |
5816 | ||
5817 | default: | |
5818 | break; | |
5819 | } | |
5820 | ||
ce677831 | 5821 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5822 | return -ENOENT; |
5823 | ||
5824 | if (!event->parent) { | |
5825 | int err; | |
5826 | ||
5827 | err = swevent_hlist_get(event); | |
5828 | if (err) | |
5829 | return err; | |
5830 | ||
c5905afb | 5831 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5832 | event->destroy = sw_perf_event_destroy; |
5833 | } | |
5834 | ||
5835 | return 0; | |
5836 | } | |
5837 | ||
35edc2a5 PZ |
5838 | static int perf_swevent_event_idx(struct perf_event *event) |
5839 | { | |
5840 | return 0; | |
5841 | } | |
5842 | ||
b0a873eb | 5843 | static struct pmu perf_swevent = { |
89a1e187 | 5844 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5845 | |
b0a873eb | 5846 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5847 | .add = perf_swevent_add, |
5848 | .del = perf_swevent_del, | |
5849 | .start = perf_swevent_start, | |
5850 | .stop = perf_swevent_stop, | |
1c024eca | 5851 | .read = perf_swevent_read, |
35edc2a5 PZ |
5852 | |
5853 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5854 | }; |
5855 | ||
b0a873eb PZ |
5856 | #ifdef CONFIG_EVENT_TRACING |
5857 | ||
1c024eca PZ |
5858 | static int perf_tp_filter_match(struct perf_event *event, |
5859 | struct perf_sample_data *data) | |
5860 | { | |
5861 | void *record = data->raw->data; | |
5862 | ||
5863 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5864 | return 1; | |
5865 | return 0; | |
5866 | } | |
5867 | ||
5868 | static int perf_tp_event_match(struct perf_event *event, | |
5869 | struct perf_sample_data *data, | |
5870 | struct pt_regs *regs) | |
5871 | { | |
a0f7d0f7 FW |
5872 | if (event->hw.state & PERF_HES_STOPPED) |
5873 | return 0; | |
580d607c PZ |
5874 | /* |
5875 | * All tracepoints are from kernel-space. | |
5876 | */ | |
5877 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5878 | return 0; |
5879 | ||
5880 | if (!perf_tp_filter_match(event, data)) | |
5881 | return 0; | |
5882 | ||
5883 | return 1; | |
5884 | } | |
5885 | ||
5886 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5887 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5888 | struct task_struct *task) | |
95476b64 FW |
5889 | { |
5890 | struct perf_sample_data data; | |
1c024eca | 5891 | struct perf_event *event; |
1c024eca | 5892 | |
95476b64 FW |
5893 | struct perf_raw_record raw = { |
5894 | .size = entry_size, | |
5895 | .data = record, | |
5896 | }; | |
5897 | ||
fd0d000b | 5898 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5899 | data.raw = &raw; |
5900 | ||
b67bfe0d | 5901 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5902 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5903 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5904 | } |
ecc55f84 | 5905 | |
e6dab5ff AV |
5906 | /* |
5907 | * If we got specified a target task, also iterate its context and | |
5908 | * deliver this event there too. | |
5909 | */ | |
5910 | if (task && task != current) { | |
5911 | struct perf_event_context *ctx; | |
5912 | struct trace_entry *entry = record; | |
5913 | ||
5914 | rcu_read_lock(); | |
5915 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5916 | if (!ctx) | |
5917 | goto unlock; | |
5918 | ||
5919 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5920 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5921 | continue; | |
5922 | if (event->attr.config != entry->type) | |
5923 | continue; | |
5924 | if (perf_tp_event_match(event, &data, regs)) | |
5925 | perf_swevent_event(event, count, &data, regs); | |
5926 | } | |
5927 | unlock: | |
5928 | rcu_read_unlock(); | |
5929 | } | |
5930 | ||
ecc55f84 | 5931 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5932 | } |
5933 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5934 | ||
cdd6c482 | 5935 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5936 | { |
1c024eca | 5937 | perf_trace_destroy(event); |
e077df4f PZ |
5938 | } |
5939 | ||
b0a873eb | 5940 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5941 | { |
76e1d904 FW |
5942 | int err; |
5943 | ||
b0a873eb PZ |
5944 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5945 | return -ENOENT; | |
5946 | ||
2481c5fa SE |
5947 | /* |
5948 | * no branch sampling for tracepoint events | |
5949 | */ | |
5950 | if (has_branch_stack(event)) | |
5951 | return -EOPNOTSUPP; | |
5952 | ||
1c024eca PZ |
5953 | err = perf_trace_init(event); |
5954 | if (err) | |
b0a873eb | 5955 | return err; |
e077df4f | 5956 | |
cdd6c482 | 5957 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5958 | |
b0a873eb PZ |
5959 | return 0; |
5960 | } | |
5961 | ||
5962 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5963 | .task_ctx_nr = perf_sw_context, |
5964 | ||
b0a873eb | 5965 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5966 | .add = perf_trace_add, |
5967 | .del = perf_trace_del, | |
5968 | .start = perf_swevent_start, | |
5969 | .stop = perf_swevent_stop, | |
b0a873eb | 5970 | .read = perf_swevent_read, |
35edc2a5 PZ |
5971 | |
5972 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5973 | }; |
5974 | ||
5975 | static inline void perf_tp_register(void) | |
5976 | { | |
2e80a82a | 5977 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5978 | } |
6fb2915d LZ |
5979 | |
5980 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5981 | { | |
5982 | char *filter_str; | |
5983 | int ret; | |
5984 | ||
5985 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5986 | return -EINVAL; | |
5987 | ||
5988 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5989 | if (IS_ERR(filter_str)) | |
5990 | return PTR_ERR(filter_str); | |
5991 | ||
5992 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5993 | ||
5994 | kfree(filter_str); | |
5995 | return ret; | |
5996 | } | |
5997 | ||
5998 | static void perf_event_free_filter(struct perf_event *event) | |
5999 | { | |
6000 | ftrace_profile_free_filter(event); | |
6001 | } | |
6002 | ||
e077df4f | 6003 | #else |
6fb2915d | 6004 | |
b0a873eb | 6005 | static inline void perf_tp_register(void) |
e077df4f | 6006 | { |
e077df4f | 6007 | } |
6fb2915d LZ |
6008 | |
6009 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6010 | { | |
6011 | return -ENOENT; | |
6012 | } | |
6013 | ||
6014 | static void perf_event_free_filter(struct perf_event *event) | |
6015 | { | |
6016 | } | |
6017 | ||
07b139c8 | 6018 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6019 | |
24f1e32c | 6020 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6021 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6022 | { |
f5ffe02e FW |
6023 | struct perf_sample_data sample; |
6024 | struct pt_regs *regs = data; | |
6025 | ||
fd0d000b | 6026 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6027 | |
a4eaf7f1 | 6028 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6029 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6030 | } |
6031 | #endif | |
6032 | ||
b0a873eb PZ |
6033 | /* |
6034 | * hrtimer based swevent callback | |
6035 | */ | |
f29ac756 | 6036 | |
b0a873eb | 6037 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6038 | { |
b0a873eb PZ |
6039 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6040 | struct perf_sample_data data; | |
6041 | struct pt_regs *regs; | |
6042 | struct perf_event *event; | |
6043 | u64 period; | |
f29ac756 | 6044 | |
b0a873eb | 6045 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6046 | |
6047 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6048 | return HRTIMER_NORESTART; | |
6049 | ||
b0a873eb | 6050 | event->pmu->read(event); |
f344011c | 6051 | |
fd0d000b | 6052 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6053 | regs = get_irq_regs(); |
6054 | ||
6055 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6056 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6057 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6058 | ret = HRTIMER_NORESTART; |
6059 | } | |
24f1e32c | 6060 | |
b0a873eb PZ |
6061 | period = max_t(u64, 10000, event->hw.sample_period); |
6062 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6063 | |
b0a873eb | 6064 | return ret; |
f29ac756 PZ |
6065 | } |
6066 | ||
b0a873eb | 6067 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6068 | { |
b0a873eb | 6069 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6070 | s64 period; |
6071 | ||
6072 | if (!is_sampling_event(event)) | |
6073 | return; | |
f5ffe02e | 6074 | |
5d508e82 FBH |
6075 | period = local64_read(&hwc->period_left); |
6076 | if (period) { | |
6077 | if (period < 0) | |
6078 | period = 10000; | |
fa407f35 | 6079 | |
5d508e82 FBH |
6080 | local64_set(&hwc->period_left, 0); |
6081 | } else { | |
6082 | period = max_t(u64, 10000, hwc->sample_period); | |
6083 | } | |
6084 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6085 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6086 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6087 | } |
b0a873eb PZ |
6088 | |
6089 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6090 | { |
b0a873eb PZ |
6091 | struct hw_perf_event *hwc = &event->hw; |
6092 | ||
6c7e550f | 6093 | if (is_sampling_event(event)) { |
b0a873eb | 6094 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6095 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6096 | |
6097 | hrtimer_cancel(&hwc->hrtimer); | |
6098 | } | |
24f1e32c FW |
6099 | } |
6100 | ||
ba3dd36c PZ |
6101 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6102 | { | |
6103 | struct hw_perf_event *hwc = &event->hw; | |
6104 | ||
6105 | if (!is_sampling_event(event)) | |
6106 | return; | |
6107 | ||
6108 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6109 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6110 | ||
6111 | /* | |
6112 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6113 | * mapping and avoid the whole period adjust feedback stuff. | |
6114 | */ | |
6115 | if (event->attr.freq) { | |
6116 | long freq = event->attr.sample_freq; | |
6117 | ||
6118 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6119 | hwc->sample_period = event->attr.sample_period; | |
6120 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6121 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6122 | event->attr.freq = 0; |
6123 | } | |
6124 | } | |
6125 | ||
b0a873eb PZ |
6126 | /* |
6127 | * Software event: cpu wall time clock | |
6128 | */ | |
6129 | ||
6130 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6131 | { |
b0a873eb PZ |
6132 | s64 prev; |
6133 | u64 now; | |
6134 | ||
a4eaf7f1 | 6135 | now = local_clock(); |
b0a873eb PZ |
6136 | prev = local64_xchg(&event->hw.prev_count, now); |
6137 | local64_add(now - prev, &event->count); | |
24f1e32c | 6138 | } |
24f1e32c | 6139 | |
a4eaf7f1 | 6140 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6141 | { |
a4eaf7f1 | 6142 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6143 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6144 | } |
6145 | ||
a4eaf7f1 | 6146 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6147 | { |
b0a873eb PZ |
6148 | perf_swevent_cancel_hrtimer(event); |
6149 | cpu_clock_event_update(event); | |
6150 | } | |
f29ac756 | 6151 | |
a4eaf7f1 PZ |
6152 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6153 | { | |
6154 | if (flags & PERF_EF_START) | |
6155 | cpu_clock_event_start(event, flags); | |
6156 | ||
6157 | return 0; | |
6158 | } | |
6159 | ||
6160 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6161 | { | |
6162 | cpu_clock_event_stop(event, flags); | |
6163 | } | |
6164 | ||
b0a873eb PZ |
6165 | static void cpu_clock_event_read(struct perf_event *event) |
6166 | { | |
6167 | cpu_clock_event_update(event); | |
6168 | } | |
f344011c | 6169 | |
b0a873eb PZ |
6170 | static int cpu_clock_event_init(struct perf_event *event) |
6171 | { | |
6172 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6173 | return -ENOENT; | |
6174 | ||
6175 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6176 | return -ENOENT; | |
6177 | ||
2481c5fa SE |
6178 | /* |
6179 | * no branch sampling for software events | |
6180 | */ | |
6181 | if (has_branch_stack(event)) | |
6182 | return -EOPNOTSUPP; | |
6183 | ||
ba3dd36c PZ |
6184 | perf_swevent_init_hrtimer(event); |
6185 | ||
b0a873eb | 6186 | return 0; |
f29ac756 PZ |
6187 | } |
6188 | ||
b0a873eb | 6189 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6190 | .task_ctx_nr = perf_sw_context, |
6191 | ||
b0a873eb | 6192 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6193 | .add = cpu_clock_event_add, |
6194 | .del = cpu_clock_event_del, | |
6195 | .start = cpu_clock_event_start, | |
6196 | .stop = cpu_clock_event_stop, | |
b0a873eb | 6197 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
6198 | |
6199 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6200 | }; |
6201 | ||
6202 | /* | |
6203 | * Software event: task time clock | |
6204 | */ | |
6205 | ||
6206 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6207 | { |
b0a873eb PZ |
6208 | u64 prev; |
6209 | s64 delta; | |
5c92d124 | 6210 | |
b0a873eb PZ |
6211 | prev = local64_xchg(&event->hw.prev_count, now); |
6212 | delta = now - prev; | |
6213 | local64_add(delta, &event->count); | |
6214 | } | |
5c92d124 | 6215 | |
a4eaf7f1 | 6216 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6217 | { |
a4eaf7f1 | 6218 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6219 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6220 | } |
6221 | ||
a4eaf7f1 | 6222 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6223 | { |
6224 | perf_swevent_cancel_hrtimer(event); | |
6225 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6226 | } |
6227 | ||
6228 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6229 | { | |
6230 | if (flags & PERF_EF_START) | |
6231 | task_clock_event_start(event, flags); | |
b0a873eb | 6232 | |
a4eaf7f1 PZ |
6233 | return 0; |
6234 | } | |
6235 | ||
6236 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6237 | { | |
6238 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6239 | } |
6240 | ||
6241 | static void task_clock_event_read(struct perf_event *event) | |
6242 | { | |
768a06e2 PZ |
6243 | u64 now = perf_clock(); |
6244 | u64 delta = now - event->ctx->timestamp; | |
6245 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6246 | |
6247 | task_clock_event_update(event, time); | |
6248 | } | |
6249 | ||
6250 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6251 | { |
b0a873eb PZ |
6252 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6253 | return -ENOENT; | |
6254 | ||
6255 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6256 | return -ENOENT; | |
6257 | ||
2481c5fa SE |
6258 | /* |
6259 | * no branch sampling for software events | |
6260 | */ | |
6261 | if (has_branch_stack(event)) | |
6262 | return -EOPNOTSUPP; | |
6263 | ||
ba3dd36c PZ |
6264 | perf_swevent_init_hrtimer(event); |
6265 | ||
b0a873eb | 6266 | return 0; |
6fb2915d LZ |
6267 | } |
6268 | ||
b0a873eb | 6269 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6270 | .task_ctx_nr = perf_sw_context, |
6271 | ||
b0a873eb | 6272 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6273 | .add = task_clock_event_add, |
6274 | .del = task_clock_event_del, | |
6275 | .start = task_clock_event_start, | |
6276 | .stop = task_clock_event_stop, | |
b0a873eb | 6277 | .read = task_clock_event_read, |
35edc2a5 PZ |
6278 | |
6279 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6280 | }; |
6fb2915d | 6281 | |
ad5133b7 | 6282 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6283 | { |
e077df4f | 6284 | } |
6fb2915d | 6285 | |
ad5133b7 | 6286 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6287 | { |
ad5133b7 | 6288 | return 0; |
6fb2915d LZ |
6289 | } |
6290 | ||
ad5133b7 | 6291 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6292 | { |
ad5133b7 | 6293 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6294 | } |
6295 | ||
ad5133b7 PZ |
6296 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6297 | { | |
6298 | perf_pmu_enable(pmu); | |
6299 | return 0; | |
6300 | } | |
e077df4f | 6301 | |
ad5133b7 | 6302 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6303 | { |
ad5133b7 | 6304 | perf_pmu_enable(pmu); |
24f1e32c FW |
6305 | } |
6306 | ||
35edc2a5 PZ |
6307 | static int perf_event_idx_default(struct perf_event *event) |
6308 | { | |
6309 | return event->hw.idx + 1; | |
6310 | } | |
6311 | ||
8dc85d54 PZ |
6312 | /* |
6313 | * Ensures all contexts with the same task_ctx_nr have the same | |
6314 | * pmu_cpu_context too. | |
6315 | */ | |
9e317041 | 6316 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 6317 | { |
8dc85d54 | 6318 | struct pmu *pmu; |
b326e956 | 6319 | |
8dc85d54 PZ |
6320 | if (ctxn < 0) |
6321 | return NULL; | |
24f1e32c | 6322 | |
8dc85d54 PZ |
6323 | list_for_each_entry(pmu, &pmus, entry) { |
6324 | if (pmu->task_ctx_nr == ctxn) | |
6325 | return pmu->pmu_cpu_context; | |
6326 | } | |
24f1e32c | 6327 | |
8dc85d54 | 6328 | return NULL; |
24f1e32c FW |
6329 | } |
6330 | ||
51676957 | 6331 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6332 | { |
51676957 PZ |
6333 | int cpu; |
6334 | ||
6335 | for_each_possible_cpu(cpu) { | |
6336 | struct perf_cpu_context *cpuctx; | |
6337 | ||
6338 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6339 | ||
3f1f3320 PZ |
6340 | if (cpuctx->unique_pmu == old_pmu) |
6341 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6342 | } |
6343 | } | |
6344 | ||
6345 | static void free_pmu_context(struct pmu *pmu) | |
6346 | { | |
6347 | struct pmu *i; | |
f5ffe02e | 6348 | |
8dc85d54 | 6349 | mutex_lock(&pmus_lock); |
0475f9ea | 6350 | /* |
8dc85d54 | 6351 | * Like a real lame refcount. |
0475f9ea | 6352 | */ |
51676957 PZ |
6353 | list_for_each_entry(i, &pmus, entry) { |
6354 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6355 | update_pmu_context(i, pmu); | |
8dc85d54 | 6356 | goto out; |
51676957 | 6357 | } |
8dc85d54 | 6358 | } |
d6d020e9 | 6359 | |
51676957 | 6360 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6361 | out: |
6362 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6363 | } |
2e80a82a | 6364 | static struct idr pmu_idr; |
d6d020e9 | 6365 | |
abe43400 PZ |
6366 | static ssize_t |
6367 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6368 | { | |
6369 | struct pmu *pmu = dev_get_drvdata(dev); | |
6370 | ||
6371 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6372 | } | |
90826ca7 | 6373 | static DEVICE_ATTR_RO(type); |
abe43400 | 6374 | |
62b85639 SE |
6375 | static ssize_t |
6376 | perf_event_mux_interval_ms_show(struct device *dev, | |
6377 | struct device_attribute *attr, | |
6378 | char *page) | |
6379 | { | |
6380 | struct pmu *pmu = dev_get_drvdata(dev); | |
6381 | ||
6382 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6383 | } | |
6384 | ||
6385 | static ssize_t | |
6386 | perf_event_mux_interval_ms_store(struct device *dev, | |
6387 | struct device_attribute *attr, | |
6388 | const char *buf, size_t count) | |
6389 | { | |
6390 | struct pmu *pmu = dev_get_drvdata(dev); | |
6391 | int timer, cpu, ret; | |
6392 | ||
6393 | ret = kstrtoint(buf, 0, &timer); | |
6394 | if (ret) | |
6395 | return ret; | |
6396 | ||
6397 | if (timer < 1) | |
6398 | return -EINVAL; | |
6399 | ||
6400 | /* same value, noting to do */ | |
6401 | if (timer == pmu->hrtimer_interval_ms) | |
6402 | return count; | |
6403 | ||
6404 | pmu->hrtimer_interval_ms = timer; | |
6405 | ||
6406 | /* update all cpuctx for this PMU */ | |
6407 | for_each_possible_cpu(cpu) { | |
6408 | struct perf_cpu_context *cpuctx; | |
6409 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6410 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6411 | ||
6412 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6413 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6414 | } | |
6415 | ||
6416 | return count; | |
6417 | } | |
90826ca7 | 6418 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6419 | |
90826ca7 GKH |
6420 | static struct attribute *pmu_dev_attrs[] = { |
6421 | &dev_attr_type.attr, | |
6422 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6423 | NULL, | |
abe43400 | 6424 | }; |
90826ca7 | 6425 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6426 | |
6427 | static int pmu_bus_running; | |
6428 | static struct bus_type pmu_bus = { | |
6429 | .name = "event_source", | |
90826ca7 | 6430 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6431 | }; |
6432 | ||
6433 | static void pmu_dev_release(struct device *dev) | |
6434 | { | |
6435 | kfree(dev); | |
6436 | } | |
6437 | ||
6438 | static int pmu_dev_alloc(struct pmu *pmu) | |
6439 | { | |
6440 | int ret = -ENOMEM; | |
6441 | ||
6442 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6443 | if (!pmu->dev) | |
6444 | goto out; | |
6445 | ||
0c9d42ed | 6446 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6447 | device_initialize(pmu->dev); |
6448 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6449 | if (ret) | |
6450 | goto free_dev; | |
6451 | ||
6452 | dev_set_drvdata(pmu->dev, pmu); | |
6453 | pmu->dev->bus = &pmu_bus; | |
6454 | pmu->dev->release = pmu_dev_release; | |
6455 | ret = device_add(pmu->dev); | |
6456 | if (ret) | |
6457 | goto free_dev; | |
6458 | ||
6459 | out: | |
6460 | return ret; | |
6461 | ||
6462 | free_dev: | |
6463 | put_device(pmu->dev); | |
6464 | goto out; | |
6465 | } | |
6466 | ||
547e9fd7 | 6467 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6468 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6469 | |
03d8e80b | 6470 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6471 | { |
108b02cf | 6472 | int cpu, ret; |
24f1e32c | 6473 | |
b0a873eb | 6474 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6475 | ret = -ENOMEM; |
6476 | pmu->pmu_disable_count = alloc_percpu(int); | |
6477 | if (!pmu->pmu_disable_count) | |
6478 | goto unlock; | |
f29ac756 | 6479 | |
2e80a82a PZ |
6480 | pmu->type = -1; |
6481 | if (!name) | |
6482 | goto skip_type; | |
6483 | pmu->name = name; | |
6484 | ||
6485 | if (type < 0) { | |
0e9c3be2 TH |
6486 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6487 | if (type < 0) { | |
6488 | ret = type; | |
2e80a82a PZ |
6489 | goto free_pdc; |
6490 | } | |
6491 | } | |
6492 | pmu->type = type; | |
6493 | ||
abe43400 PZ |
6494 | if (pmu_bus_running) { |
6495 | ret = pmu_dev_alloc(pmu); | |
6496 | if (ret) | |
6497 | goto free_idr; | |
6498 | } | |
6499 | ||
2e80a82a | 6500 | skip_type: |
8dc85d54 PZ |
6501 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6502 | if (pmu->pmu_cpu_context) | |
6503 | goto got_cpu_context; | |
f29ac756 | 6504 | |
c4814202 | 6505 | ret = -ENOMEM; |
108b02cf PZ |
6506 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6507 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6508 | goto free_dev; |
f344011c | 6509 | |
108b02cf PZ |
6510 | for_each_possible_cpu(cpu) { |
6511 | struct perf_cpu_context *cpuctx; | |
6512 | ||
6513 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6514 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6515 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6516 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6517 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6518 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6519 | |
6520 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6521 | ||
e9d2b064 | 6522 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6523 | cpuctx->unique_pmu = pmu; |
108b02cf | 6524 | } |
76e1d904 | 6525 | |
8dc85d54 | 6526 | got_cpu_context: |
ad5133b7 PZ |
6527 | if (!pmu->start_txn) { |
6528 | if (pmu->pmu_enable) { | |
6529 | /* | |
6530 | * If we have pmu_enable/pmu_disable calls, install | |
6531 | * transaction stubs that use that to try and batch | |
6532 | * hardware accesses. | |
6533 | */ | |
6534 | pmu->start_txn = perf_pmu_start_txn; | |
6535 | pmu->commit_txn = perf_pmu_commit_txn; | |
6536 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6537 | } else { | |
6538 | pmu->start_txn = perf_pmu_nop_void; | |
6539 | pmu->commit_txn = perf_pmu_nop_int; | |
6540 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6541 | } |
5c92d124 | 6542 | } |
15dbf27c | 6543 | |
ad5133b7 PZ |
6544 | if (!pmu->pmu_enable) { |
6545 | pmu->pmu_enable = perf_pmu_nop_void; | |
6546 | pmu->pmu_disable = perf_pmu_nop_void; | |
6547 | } | |
6548 | ||
35edc2a5 PZ |
6549 | if (!pmu->event_idx) |
6550 | pmu->event_idx = perf_event_idx_default; | |
6551 | ||
b0a873eb | 6552 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6553 | ret = 0; |
6554 | unlock: | |
b0a873eb PZ |
6555 | mutex_unlock(&pmus_lock); |
6556 | ||
33696fc0 | 6557 | return ret; |
108b02cf | 6558 | |
abe43400 PZ |
6559 | free_dev: |
6560 | device_del(pmu->dev); | |
6561 | put_device(pmu->dev); | |
6562 | ||
2e80a82a PZ |
6563 | free_idr: |
6564 | if (pmu->type >= PERF_TYPE_MAX) | |
6565 | idr_remove(&pmu_idr, pmu->type); | |
6566 | ||
108b02cf PZ |
6567 | free_pdc: |
6568 | free_percpu(pmu->pmu_disable_count); | |
6569 | goto unlock; | |
f29ac756 | 6570 | } |
c464c76e | 6571 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 6572 | |
b0a873eb | 6573 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6574 | { |
b0a873eb PZ |
6575 | mutex_lock(&pmus_lock); |
6576 | list_del_rcu(&pmu->entry); | |
6577 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6578 | |
0475f9ea | 6579 | /* |
cde8e884 PZ |
6580 | * We dereference the pmu list under both SRCU and regular RCU, so |
6581 | * synchronize against both of those. | |
0475f9ea | 6582 | */ |
b0a873eb | 6583 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6584 | synchronize_rcu(); |
d6d020e9 | 6585 | |
33696fc0 | 6586 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6587 | if (pmu->type >= PERF_TYPE_MAX) |
6588 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6589 | device_del(pmu->dev); |
6590 | put_device(pmu->dev); | |
51676957 | 6591 | free_pmu_context(pmu); |
b0a873eb | 6592 | } |
c464c76e | 6593 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 6594 | |
b0a873eb PZ |
6595 | struct pmu *perf_init_event(struct perf_event *event) |
6596 | { | |
6597 | struct pmu *pmu = NULL; | |
6598 | int idx; | |
940c5b29 | 6599 | int ret; |
b0a873eb PZ |
6600 | |
6601 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6602 | |
6603 | rcu_read_lock(); | |
6604 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6605 | rcu_read_unlock(); | |
940c5b29 | 6606 | if (pmu) { |
c464c76e YZ |
6607 | if (!try_module_get(pmu->module)) { |
6608 | pmu = ERR_PTR(-ENODEV); | |
6609 | goto unlock; | |
6610 | } | |
7e5b2a01 | 6611 | event->pmu = pmu; |
940c5b29 LM |
6612 | ret = pmu->event_init(event); |
6613 | if (ret) | |
6614 | pmu = ERR_PTR(ret); | |
2e80a82a | 6615 | goto unlock; |
940c5b29 | 6616 | } |
2e80a82a | 6617 | |
b0a873eb | 6618 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
c464c76e YZ |
6619 | if (!try_module_get(pmu->module)) { |
6620 | pmu = ERR_PTR(-ENODEV); | |
6621 | goto unlock; | |
6622 | } | |
7e5b2a01 | 6623 | event->pmu = pmu; |
940c5b29 | 6624 | ret = pmu->event_init(event); |
b0a873eb | 6625 | if (!ret) |
e5f4d339 | 6626 | goto unlock; |
76e1d904 | 6627 | |
b0a873eb PZ |
6628 | if (ret != -ENOENT) { |
6629 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6630 | goto unlock; |
f344011c | 6631 | } |
5c92d124 | 6632 | } |
e5f4d339 PZ |
6633 | pmu = ERR_PTR(-ENOENT); |
6634 | unlock: | |
b0a873eb | 6635 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6636 | |
4aeb0b42 | 6637 | return pmu; |
5c92d124 IM |
6638 | } |
6639 | ||
4beb31f3 FW |
6640 | static void account_event_cpu(struct perf_event *event, int cpu) |
6641 | { | |
6642 | if (event->parent) | |
6643 | return; | |
6644 | ||
6645 | if (has_branch_stack(event)) { | |
6646 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6647 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
6648 | } | |
6649 | if (is_cgroup_event(event)) | |
6650 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
6651 | } | |
6652 | ||
766d6c07 FW |
6653 | static void account_event(struct perf_event *event) |
6654 | { | |
4beb31f3 FW |
6655 | if (event->parent) |
6656 | return; | |
6657 | ||
766d6c07 FW |
6658 | if (event->attach_state & PERF_ATTACH_TASK) |
6659 | static_key_slow_inc(&perf_sched_events.key); | |
6660 | if (event->attr.mmap || event->attr.mmap_data) | |
6661 | atomic_inc(&nr_mmap_events); | |
6662 | if (event->attr.comm) | |
6663 | atomic_inc(&nr_comm_events); | |
6664 | if (event->attr.task) | |
6665 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
6666 | if (event->attr.freq) { |
6667 | if (atomic_inc_return(&nr_freq_events) == 1) | |
6668 | tick_nohz_full_kick_all(); | |
6669 | } | |
4beb31f3 | 6670 | if (has_branch_stack(event)) |
766d6c07 | 6671 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 6672 | if (is_cgroup_event(event)) |
766d6c07 | 6673 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
6674 | |
6675 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
6676 | } |
6677 | ||
0793a61d | 6678 | /* |
cdd6c482 | 6679 | * Allocate and initialize a event structure |
0793a61d | 6680 | */ |
cdd6c482 | 6681 | static struct perf_event * |
c3f00c70 | 6682 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6683 | struct task_struct *task, |
6684 | struct perf_event *group_leader, | |
6685 | struct perf_event *parent_event, | |
4dc0da86 AK |
6686 | perf_overflow_handler_t overflow_handler, |
6687 | void *context) | |
0793a61d | 6688 | { |
51b0fe39 | 6689 | struct pmu *pmu; |
cdd6c482 IM |
6690 | struct perf_event *event; |
6691 | struct hw_perf_event *hwc; | |
90983b16 | 6692 | long err = -EINVAL; |
0793a61d | 6693 | |
66832eb4 ON |
6694 | if ((unsigned)cpu >= nr_cpu_ids) { |
6695 | if (!task || cpu != -1) | |
6696 | return ERR_PTR(-EINVAL); | |
6697 | } | |
6698 | ||
c3f00c70 | 6699 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6700 | if (!event) |
d5d2bc0d | 6701 | return ERR_PTR(-ENOMEM); |
0793a61d | 6702 | |
04289bb9 | 6703 | /* |
cdd6c482 | 6704 | * Single events are their own group leaders, with an |
04289bb9 IM |
6705 | * empty sibling list: |
6706 | */ | |
6707 | if (!group_leader) | |
cdd6c482 | 6708 | group_leader = event; |
04289bb9 | 6709 | |
cdd6c482 IM |
6710 | mutex_init(&event->child_mutex); |
6711 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6712 | |
cdd6c482 IM |
6713 | INIT_LIST_HEAD(&event->group_entry); |
6714 | INIT_LIST_HEAD(&event->event_entry); | |
6715 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 6716 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 6717 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
6718 | INIT_HLIST_NODE(&event->hlist_entry); |
6719 | ||
10c6db11 | 6720 | |
cdd6c482 | 6721 | init_waitqueue_head(&event->waitq); |
e360adbe | 6722 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6723 | |
cdd6c482 | 6724 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6725 | |
a6fa941d | 6726 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6727 | event->cpu = cpu; |
6728 | event->attr = *attr; | |
6729 | event->group_leader = group_leader; | |
6730 | event->pmu = NULL; | |
cdd6c482 | 6731 | event->oncpu = -1; |
a96bbc16 | 6732 | |
cdd6c482 | 6733 | event->parent = parent_event; |
b84fbc9f | 6734 | |
17cf22c3 | 6735 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6736 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6737 | |
cdd6c482 | 6738 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6739 | |
d580ff86 PZ |
6740 | if (task) { |
6741 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6742 | |
6743 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6744 | event->hw.tp_target = task; | |
d580ff86 PZ |
6745 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6746 | /* | |
6747 | * hw_breakpoint is a bit difficult here.. | |
6748 | */ | |
f22c1bb6 | 6749 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6750 | event->hw.bp_target = task; |
6751 | #endif | |
6752 | } | |
6753 | ||
4dc0da86 | 6754 | if (!overflow_handler && parent_event) { |
b326e956 | 6755 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6756 | context = parent_event->overflow_handler_context; |
6757 | } | |
66832eb4 | 6758 | |
b326e956 | 6759 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6760 | event->overflow_handler_context = context; |
97eaf530 | 6761 | |
0231bb53 | 6762 | perf_event__state_init(event); |
a86ed508 | 6763 | |
4aeb0b42 | 6764 | pmu = NULL; |
b8e83514 | 6765 | |
cdd6c482 | 6766 | hwc = &event->hw; |
bd2b5b12 | 6767 | hwc->sample_period = attr->sample_period; |
0d48696f | 6768 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6769 | hwc->sample_period = 1; |
eced1dfc | 6770 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6771 | |
e7850595 | 6772 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6773 | |
2023b359 | 6774 | /* |
cdd6c482 | 6775 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6776 | */ |
3dab77fb | 6777 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 6778 | goto err_ns; |
2023b359 | 6779 | |
b0a873eb | 6780 | pmu = perf_init_event(event); |
4aeb0b42 | 6781 | if (!pmu) |
90983b16 FW |
6782 | goto err_ns; |
6783 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 6784 | err = PTR_ERR(pmu); |
90983b16 | 6785 | goto err_ns; |
621a01ea | 6786 | } |
d5d2bc0d | 6787 | |
cdd6c482 | 6788 | if (!event->parent) { |
927c7a9e FW |
6789 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6790 | err = get_callchain_buffers(); | |
90983b16 FW |
6791 | if (err) |
6792 | goto err_pmu; | |
d010b332 | 6793 | } |
f344011c | 6794 | } |
9ee318a7 | 6795 | |
cdd6c482 | 6796 | return event; |
90983b16 FW |
6797 | |
6798 | err_pmu: | |
6799 | if (event->destroy) | |
6800 | event->destroy(event); | |
c464c76e | 6801 | module_put(pmu->module); |
90983b16 FW |
6802 | err_ns: |
6803 | if (event->ns) | |
6804 | put_pid_ns(event->ns); | |
6805 | kfree(event); | |
6806 | ||
6807 | return ERR_PTR(err); | |
0793a61d TG |
6808 | } |
6809 | ||
cdd6c482 IM |
6810 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6811 | struct perf_event_attr *attr) | |
974802ea | 6812 | { |
974802ea | 6813 | u32 size; |
cdf8073d | 6814 | int ret; |
974802ea PZ |
6815 | |
6816 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6817 | return -EFAULT; | |
6818 | ||
6819 | /* | |
6820 | * zero the full structure, so that a short copy will be nice. | |
6821 | */ | |
6822 | memset(attr, 0, sizeof(*attr)); | |
6823 | ||
6824 | ret = get_user(size, &uattr->size); | |
6825 | if (ret) | |
6826 | return ret; | |
6827 | ||
6828 | if (size > PAGE_SIZE) /* silly large */ | |
6829 | goto err_size; | |
6830 | ||
6831 | if (!size) /* abi compat */ | |
6832 | size = PERF_ATTR_SIZE_VER0; | |
6833 | ||
6834 | if (size < PERF_ATTR_SIZE_VER0) | |
6835 | goto err_size; | |
6836 | ||
6837 | /* | |
6838 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6839 | * ensure all the unknown bits are 0 - i.e. new |
6840 | * user-space does not rely on any kernel feature | |
6841 | * extensions we dont know about yet. | |
974802ea PZ |
6842 | */ |
6843 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6844 | unsigned char __user *addr; |
6845 | unsigned char __user *end; | |
6846 | unsigned char val; | |
974802ea | 6847 | |
cdf8073d IS |
6848 | addr = (void __user *)uattr + sizeof(*attr); |
6849 | end = (void __user *)uattr + size; | |
974802ea | 6850 | |
cdf8073d | 6851 | for (; addr < end; addr++) { |
974802ea PZ |
6852 | ret = get_user(val, addr); |
6853 | if (ret) | |
6854 | return ret; | |
6855 | if (val) | |
6856 | goto err_size; | |
6857 | } | |
b3e62e35 | 6858 | size = sizeof(*attr); |
974802ea PZ |
6859 | } |
6860 | ||
6861 | ret = copy_from_user(attr, uattr, size); | |
6862 | if (ret) | |
6863 | return -EFAULT; | |
6864 | ||
3090ffb5 SE |
6865 | /* disabled for now */ |
6866 | if (attr->mmap2) | |
6867 | return -EINVAL; | |
6868 | ||
cd757645 | 6869 | if (attr->__reserved_1) |
974802ea PZ |
6870 | return -EINVAL; |
6871 | ||
6872 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6873 | return -EINVAL; | |
6874 | ||
6875 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6876 | return -EINVAL; | |
6877 | ||
bce38cd5 SE |
6878 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6879 | u64 mask = attr->branch_sample_type; | |
6880 | ||
6881 | /* only using defined bits */ | |
6882 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6883 | return -EINVAL; | |
6884 | ||
6885 | /* at least one branch bit must be set */ | |
6886 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6887 | return -EINVAL; | |
6888 | ||
bce38cd5 SE |
6889 | /* propagate priv level, when not set for branch */ |
6890 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6891 | ||
6892 | /* exclude_kernel checked on syscall entry */ | |
6893 | if (!attr->exclude_kernel) | |
6894 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6895 | ||
6896 | if (!attr->exclude_user) | |
6897 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6898 | ||
6899 | if (!attr->exclude_hv) | |
6900 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6901 | /* | |
6902 | * adjust user setting (for HW filter setup) | |
6903 | */ | |
6904 | attr->branch_sample_type = mask; | |
6905 | } | |
e712209a SE |
6906 | /* privileged levels capture (kernel, hv): check permissions */ |
6907 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
6908 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6909 | return -EACCES; | |
bce38cd5 | 6910 | } |
4018994f | 6911 | |
c5ebcedb | 6912 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6913 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6914 | if (ret) |
6915 | return ret; | |
6916 | } | |
6917 | ||
6918 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6919 | if (!arch_perf_have_user_stack_dump()) | |
6920 | return -ENOSYS; | |
6921 | ||
6922 | /* | |
6923 | * We have __u32 type for the size, but so far | |
6924 | * we can only use __u16 as maximum due to the | |
6925 | * __u16 sample size limit. | |
6926 | */ | |
6927 | if (attr->sample_stack_user >= USHRT_MAX) | |
6928 | ret = -EINVAL; | |
6929 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6930 | ret = -EINVAL; | |
6931 | } | |
4018994f | 6932 | |
974802ea PZ |
6933 | out: |
6934 | return ret; | |
6935 | ||
6936 | err_size: | |
6937 | put_user(sizeof(*attr), &uattr->size); | |
6938 | ret = -E2BIG; | |
6939 | goto out; | |
6940 | } | |
6941 | ||
ac9721f3 PZ |
6942 | static int |
6943 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6944 | { |
76369139 | 6945 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6946 | int ret = -EINVAL; |
6947 | ||
ac9721f3 | 6948 | if (!output_event) |
a4be7c27 PZ |
6949 | goto set; |
6950 | ||
ac9721f3 PZ |
6951 | /* don't allow circular references */ |
6952 | if (event == output_event) | |
a4be7c27 PZ |
6953 | goto out; |
6954 | ||
0f139300 PZ |
6955 | /* |
6956 | * Don't allow cross-cpu buffers | |
6957 | */ | |
6958 | if (output_event->cpu != event->cpu) | |
6959 | goto out; | |
6960 | ||
6961 | /* | |
76369139 | 6962 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6963 | */ |
6964 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6965 | goto out; | |
6966 | ||
a4be7c27 | 6967 | set: |
cdd6c482 | 6968 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6969 | /* Can't redirect output if we've got an active mmap() */ |
6970 | if (atomic_read(&event->mmap_count)) | |
6971 | goto unlock; | |
a4be7c27 | 6972 | |
9bb5d40c PZ |
6973 | old_rb = event->rb; |
6974 | ||
ac9721f3 | 6975 | if (output_event) { |
76369139 FW |
6976 | /* get the rb we want to redirect to */ |
6977 | rb = ring_buffer_get(output_event); | |
6978 | if (!rb) | |
ac9721f3 | 6979 | goto unlock; |
a4be7c27 PZ |
6980 | } |
6981 | ||
10c6db11 PZ |
6982 | if (old_rb) |
6983 | ring_buffer_detach(event, old_rb); | |
9bb5d40c PZ |
6984 | |
6985 | if (rb) | |
6986 | ring_buffer_attach(event, rb); | |
6987 | ||
6988 | rcu_assign_pointer(event->rb, rb); | |
6989 | ||
6990 | if (old_rb) { | |
6991 | ring_buffer_put(old_rb); | |
6992 | /* | |
6993 | * Since we detached before setting the new rb, so that we | |
6994 | * could attach the new rb, we could have missed a wakeup. | |
6995 | * Provide it now. | |
6996 | */ | |
6997 | wake_up_all(&event->waitq); | |
6998 | } | |
6999 | ||
a4be7c27 | 7000 | ret = 0; |
ac9721f3 PZ |
7001 | unlock: |
7002 | mutex_unlock(&event->mmap_mutex); | |
7003 | ||
a4be7c27 | 7004 | out: |
a4be7c27 PZ |
7005 | return ret; |
7006 | } | |
7007 | ||
0793a61d | 7008 | /** |
cdd6c482 | 7009 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7010 | * |
cdd6c482 | 7011 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7012 | * @pid: target pid |
9f66a381 | 7013 | * @cpu: target cpu |
cdd6c482 | 7014 | * @group_fd: group leader event fd |
0793a61d | 7015 | */ |
cdd6c482 IM |
7016 | SYSCALL_DEFINE5(perf_event_open, |
7017 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7018 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7019 | { |
b04243ef PZ |
7020 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7021 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
7022 | struct perf_event_attr attr; |
7023 | struct perf_event_context *ctx; | |
7024 | struct file *event_file = NULL; | |
2903ff01 | 7025 | struct fd group = {NULL, 0}; |
38a81da2 | 7026 | struct task_struct *task = NULL; |
89a1e187 | 7027 | struct pmu *pmu; |
ea635c64 | 7028 | int event_fd; |
b04243ef | 7029 | int move_group = 0; |
dc86cabe | 7030 | int err; |
a21b0b35 | 7031 | int f_flags = O_RDWR; |
0793a61d | 7032 | |
2743a5b0 | 7033 | /* for future expandability... */ |
e5d1367f | 7034 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7035 | return -EINVAL; |
7036 | ||
dc86cabe IM |
7037 | err = perf_copy_attr(attr_uptr, &attr); |
7038 | if (err) | |
7039 | return err; | |
eab656ae | 7040 | |
0764771d PZ |
7041 | if (!attr.exclude_kernel) { |
7042 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7043 | return -EACCES; | |
7044 | } | |
7045 | ||
df58ab24 | 7046 | if (attr.freq) { |
cdd6c482 | 7047 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
7048 | return -EINVAL; |
7049 | } | |
7050 | ||
e5d1367f SE |
7051 | /* |
7052 | * In cgroup mode, the pid argument is used to pass the fd | |
7053 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7054 | * designates the cpu on which to monitor threads from that | |
7055 | * cgroup. | |
7056 | */ | |
7057 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7058 | return -EINVAL; | |
7059 | ||
a21b0b35 YD |
7060 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7061 | f_flags |= O_CLOEXEC; | |
7062 | ||
7063 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7064 | if (event_fd < 0) |
7065 | return event_fd; | |
7066 | ||
ac9721f3 | 7067 | if (group_fd != -1) { |
2903ff01 AV |
7068 | err = perf_fget_light(group_fd, &group); |
7069 | if (err) | |
d14b12d7 | 7070 | goto err_fd; |
2903ff01 | 7071 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7072 | if (flags & PERF_FLAG_FD_OUTPUT) |
7073 | output_event = group_leader; | |
7074 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7075 | group_leader = NULL; | |
7076 | } | |
7077 | ||
e5d1367f | 7078 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7079 | task = find_lively_task_by_vpid(pid); |
7080 | if (IS_ERR(task)) { | |
7081 | err = PTR_ERR(task); | |
7082 | goto err_group_fd; | |
7083 | } | |
7084 | } | |
7085 | ||
1f4ee503 PZ |
7086 | if (task && group_leader && |
7087 | group_leader->attr.inherit != attr.inherit) { | |
7088 | err = -EINVAL; | |
7089 | goto err_task; | |
7090 | } | |
7091 | ||
fbfc623f YZ |
7092 | get_online_cpus(); |
7093 | ||
4dc0da86 AK |
7094 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
7095 | NULL, NULL); | |
d14b12d7 SE |
7096 | if (IS_ERR(event)) { |
7097 | err = PTR_ERR(event); | |
1f4ee503 | 7098 | goto err_cpus; |
d14b12d7 SE |
7099 | } |
7100 | ||
e5d1367f SE |
7101 | if (flags & PERF_FLAG_PID_CGROUP) { |
7102 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
7103 | if (err) { |
7104 | __free_event(event); | |
1f4ee503 | 7105 | goto err_cpus; |
766d6c07 | 7106 | } |
e5d1367f SE |
7107 | } |
7108 | ||
766d6c07 FW |
7109 | account_event(event); |
7110 | ||
89a1e187 PZ |
7111 | /* |
7112 | * Special case software events and allow them to be part of | |
7113 | * any hardware group. | |
7114 | */ | |
7115 | pmu = event->pmu; | |
b04243ef PZ |
7116 | |
7117 | if (group_leader && | |
7118 | (is_software_event(event) != is_software_event(group_leader))) { | |
7119 | if (is_software_event(event)) { | |
7120 | /* | |
7121 | * If event and group_leader are not both a software | |
7122 | * event, and event is, then group leader is not. | |
7123 | * | |
7124 | * Allow the addition of software events to !software | |
7125 | * groups, this is safe because software events never | |
7126 | * fail to schedule. | |
7127 | */ | |
7128 | pmu = group_leader->pmu; | |
7129 | } else if (is_software_event(group_leader) && | |
7130 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7131 | /* | |
7132 | * In case the group is a pure software group, and we | |
7133 | * try to add a hardware event, move the whole group to | |
7134 | * the hardware context. | |
7135 | */ | |
7136 | move_group = 1; | |
7137 | } | |
7138 | } | |
89a1e187 PZ |
7139 | |
7140 | /* | |
7141 | * Get the target context (task or percpu): | |
7142 | */ | |
e2d37cd2 | 7143 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
7144 | if (IS_ERR(ctx)) { |
7145 | err = PTR_ERR(ctx); | |
c6be5a5c | 7146 | goto err_alloc; |
89a1e187 PZ |
7147 | } |
7148 | ||
fd1edb3a PZ |
7149 | if (task) { |
7150 | put_task_struct(task); | |
7151 | task = NULL; | |
7152 | } | |
7153 | ||
ccff286d | 7154 | /* |
cdd6c482 | 7155 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7156 | */ |
ac9721f3 | 7157 | if (group_leader) { |
dc86cabe | 7158 | err = -EINVAL; |
04289bb9 | 7159 | |
04289bb9 | 7160 | /* |
ccff286d IM |
7161 | * Do not allow a recursive hierarchy (this new sibling |
7162 | * becoming part of another group-sibling): | |
7163 | */ | |
7164 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7165 | goto err_context; |
ccff286d IM |
7166 | /* |
7167 | * Do not allow to attach to a group in a different | |
7168 | * task or CPU context: | |
04289bb9 | 7169 | */ |
b04243ef PZ |
7170 | if (move_group) { |
7171 | if (group_leader->ctx->type != ctx->type) | |
7172 | goto err_context; | |
7173 | } else { | |
7174 | if (group_leader->ctx != ctx) | |
7175 | goto err_context; | |
7176 | } | |
7177 | ||
3b6f9e5c PM |
7178 | /* |
7179 | * Only a group leader can be exclusive or pinned | |
7180 | */ | |
0d48696f | 7181 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7182 | goto err_context; |
ac9721f3 PZ |
7183 | } |
7184 | ||
7185 | if (output_event) { | |
7186 | err = perf_event_set_output(event, output_event); | |
7187 | if (err) | |
c3f00c70 | 7188 | goto err_context; |
ac9721f3 | 7189 | } |
0793a61d | 7190 | |
a21b0b35 YD |
7191 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7192 | f_flags); | |
ea635c64 AV |
7193 | if (IS_ERR(event_file)) { |
7194 | err = PTR_ERR(event_file); | |
c3f00c70 | 7195 | goto err_context; |
ea635c64 | 7196 | } |
9b51f66d | 7197 | |
b04243ef PZ |
7198 | if (move_group) { |
7199 | struct perf_event_context *gctx = group_leader->ctx; | |
7200 | ||
7201 | mutex_lock(&gctx->mutex); | |
46ce0fe9 | 7202 | perf_remove_from_context(group_leader, false); |
0231bb53 JO |
7203 | |
7204 | /* | |
7205 | * Removing from the context ends up with disabled | |
7206 | * event. What we want here is event in the initial | |
7207 | * startup state, ready to be add into new context. | |
7208 | */ | |
7209 | perf_event__state_init(group_leader); | |
b04243ef PZ |
7210 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7211 | group_entry) { | |
46ce0fe9 | 7212 | perf_remove_from_context(sibling, false); |
0231bb53 | 7213 | perf_event__state_init(sibling); |
b04243ef PZ |
7214 | put_ctx(gctx); |
7215 | } | |
7216 | mutex_unlock(&gctx->mutex); | |
7217 | put_ctx(gctx); | |
ea635c64 | 7218 | } |
9b51f66d | 7219 | |
ad3a37de | 7220 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 7221 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
7222 | |
7223 | if (move_group) { | |
0cda4c02 | 7224 | synchronize_rcu(); |
e2d37cd2 | 7225 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
7226 | get_ctx(ctx); |
7227 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
7228 | group_entry) { | |
e2d37cd2 | 7229 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
7230 | get_ctx(ctx); |
7231 | } | |
7232 | } | |
7233 | ||
e2d37cd2 | 7234 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7235 | perf_unpin_context(ctx); |
d859e29f | 7236 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7237 | |
fbfc623f YZ |
7238 | put_online_cpus(); |
7239 | ||
cdd6c482 | 7240 | event->owner = current; |
8882135b | 7241 | |
cdd6c482 IM |
7242 | mutex_lock(¤t->perf_event_mutex); |
7243 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7244 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7245 | |
c320c7b7 ACM |
7246 | /* |
7247 | * Precalculate sample_data sizes | |
7248 | */ | |
7249 | perf_event__header_size(event); | |
6844c09d | 7250 | perf_event__id_header_size(event); |
c320c7b7 | 7251 | |
8a49542c PZ |
7252 | /* |
7253 | * Drop the reference on the group_event after placing the | |
7254 | * new event on the sibling_list. This ensures destruction | |
7255 | * of the group leader will find the pointer to itself in | |
7256 | * perf_group_detach(). | |
7257 | */ | |
2903ff01 | 7258 | fdput(group); |
ea635c64 AV |
7259 | fd_install(event_fd, event_file); |
7260 | return event_fd; | |
0793a61d | 7261 | |
c3f00c70 | 7262 | err_context: |
fe4b04fa | 7263 | perf_unpin_context(ctx); |
ea635c64 | 7264 | put_ctx(ctx); |
c6be5a5c | 7265 | err_alloc: |
ea635c64 | 7266 | free_event(event); |
1f4ee503 | 7267 | err_cpus: |
fbfc623f | 7268 | put_online_cpus(); |
1f4ee503 | 7269 | err_task: |
e7d0bc04 PZ |
7270 | if (task) |
7271 | put_task_struct(task); | |
89a1e187 | 7272 | err_group_fd: |
2903ff01 | 7273 | fdput(group); |
ea635c64 AV |
7274 | err_fd: |
7275 | put_unused_fd(event_fd); | |
dc86cabe | 7276 | return err; |
0793a61d TG |
7277 | } |
7278 | ||
fb0459d7 AV |
7279 | /** |
7280 | * perf_event_create_kernel_counter | |
7281 | * | |
7282 | * @attr: attributes of the counter to create | |
7283 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7284 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7285 | */ |
7286 | struct perf_event * | |
7287 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7288 | struct task_struct *task, |
4dc0da86 AK |
7289 | perf_overflow_handler_t overflow_handler, |
7290 | void *context) | |
fb0459d7 | 7291 | { |
fb0459d7 | 7292 | struct perf_event_context *ctx; |
c3f00c70 | 7293 | struct perf_event *event; |
fb0459d7 | 7294 | int err; |
d859e29f | 7295 | |
fb0459d7 AV |
7296 | /* |
7297 | * Get the target context (task or percpu): | |
7298 | */ | |
d859e29f | 7299 | |
4dc0da86 AK |
7300 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7301 | overflow_handler, context); | |
c3f00c70 PZ |
7302 | if (IS_ERR(event)) { |
7303 | err = PTR_ERR(event); | |
7304 | goto err; | |
7305 | } | |
d859e29f | 7306 | |
766d6c07 FW |
7307 | account_event(event); |
7308 | ||
38a81da2 | 7309 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7310 | if (IS_ERR(ctx)) { |
7311 | err = PTR_ERR(ctx); | |
c3f00c70 | 7312 | goto err_free; |
d859e29f | 7313 | } |
fb0459d7 | 7314 | |
fb0459d7 AV |
7315 | WARN_ON_ONCE(ctx->parent_ctx); |
7316 | mutex_lock(&ctx->mutex); | |
7317 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7318 | perf_unpin_context(ctx); |
fb0459d7 AV |
7319 | mutex_unlock(&ctx->mutex); |
7320 | ||
fb0459d7 AV |
7321 | return event; |
7322 | ||
c3f00c70 PZ |
7323 | err_free: |
7324 | free_event(event); | |
7325 | err: | |
c6567f64 | 7326 | return ERR_PTR(err); |
9b51f66d | 7327 | } |
fb0459d7 | 7328 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7329 | |
0cda4c02 YZ |
7330 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7331 | { | |
7332 | struct perf_event_context *src_ctx; | |
7333 | struct perf_event_context *dst_ctx; | |
7334 | struct perf_event *event, *tmp; | |
7335 | LIST_HEAD(events); | |
7336 | ||
7337 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7338 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7339 | ||
7340 | mutex_lock(&src_ctx->mutex); | |
7341 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7342 | event_entry) { | |
46ce0fe9 | 7343 | perf_remove_from_context(event, false); |
9a545de0 | 7344 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7345 | put_ctx(src_ctx); |
9886167d | 7346 | list_add(&event->migrate_entry, &events); |
0cda4c02 YZ |
7347 | } |
7348 | mutex_unlock(&src_ctx->mutex); | |
7349 | ||
7350 | synchronize_rcu(); | |
7351 | ||
7352 | mutex_lock(&dst_ctx->mutex); | |
9886167d PZ |
7353 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7354 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7355 | if (event->state >= PERF_EVENT_STATE_OFF) |
7356 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7357 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7358 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7359 | get_ctx(dst_ctx); | |
7360 | } | |
7361 | mutex_unlock(&dst_ctx->mutex); | |
7362 | } | |
7363 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7364 | ||
cdd6c482 | 7365 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7366 | struct task_struct *child) |
d859e29f | 7367 | { |
cdd6c482 | 7368 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7369 | u64 child_val; |
d859e29f | 7370 | |
cdd6c482 IM |
7371 | if (child_event->attr.inherit_stat) |
7372 | perf_event_read_event(child_event, child); | |
38b200d6 | 7373 | |
b5e58793 | 7374 | child_val = perf_event_count(child_event); |
d859e29f PM |
7375 | |
7376 | /* | |
7377 | * Add back the child's count to the parent's count: | |
7378 | */ | |
a6e6dea6 | 7379 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7380 | atomic64_add(child_event->total_time_enabled, |
7381 | &parent_event->child_total_time_enabled); | |
7382 | atomic64_add(child_event->total_time_running, | |
7383 | &parent_event->child_total_time_running); | |
d859e29f PM |
7384 | |
7385 | /* | |
cdd6c482 | 7386 | * Remove this event from the parent's list |
d859e29f | 7387 | */ |
cdd6c482 IM |
7388 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7389 | mutex_lock(&parent_event->child_mutex); | |
7390 | list_del_init(&child_event->child_list); | |
7391 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7392 | |
7393 | /* | |
cdd6c482 | 7394 | * Release the parent event, if this was the last |
d859e29f PM |
7395 | * reference to it. |
7396 | */ | |
a6fa941d | 7397 | put_event(parent_event); |
d859e29f PM |
7398 | } |
7399 | ||
9b51f66d | 7400 | static void |
cdd6c482 IM |
7401 | __perf_event_exit_task(struct perf_event *child_event, |
7402 | struct perf_event_context *child_ctx, | |
38b200d6 | 7403 | struct task_struct *child) |
9b51f66d | 7404 | { |
15a2d4de | 7405 | perf_remove_from_context(child_event, true); |
0cc0c027 | 7406 | |
9b51f66d | 7407 | /* |
38b435b1 | 7408 | * It can happen that the parent exits first, and has events |
9b51f66d | 7409 | * that are still around due to the child reference. These |
38b435b1 | 7410 | * events need to be zapped. |
9b51f66d | 7411 | */ |
38b435b1 | 7412 | if (child_event->parent) { |
cdd6c482 IM |
7413 | sync_child_event(child_event, child); |
7414 | free_event(child_event); | |
4bcf349a | 7415 | } |
9b51f66d IM |
7416 | } |
7417 | ||
8dc85d54 | 7418 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7419 | { |
cdd6c482 IM |
7420 | struct perf_event *child_event, *tmp; |
7421 | struct perf_event_context *child_ctx; | |
a63eaf34 | 7422 | unsigned long flags; |
9b51f66d | 7423 | |
8dc85d54 | 7424 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7425 | perf_event_task(child, NULL, 0); |
9b51f66d | 7426 | return; |
9f498cc5 | 7427 | } |
9b51f66d | 7428 | |
a63eaf34 | 7429 | local_irq_save(flags); |
ad3a37de PM |
7430 | /* |
7431 | * We can't reschedule here because interrupts are disabled, | |
7432 | * and either child is current or it is a task that can't be | |
7433 | * scheduled, so we are now safe from rescheduling changing | |
7434 | * our context. | |
7435 | */ | |
806839b2 | 7436 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7437 | |
7438 | /* | |
7439 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7440 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7441 | * incremented the context's refcount before we do put_ctx below. |
7442 | */ | |
e625cce1 | 7443 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7444 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7445 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
7446 | /* |
7447 | * If this context is a clone; unclone it so it can't get | |
7448 | * swapped to another process while we're removing all | |
cdd6c482 | 7449 | * the events from it. |
71a851b4 PZ |
7450 | */ |
7451 | unclone_ctx(child_ctx); | |
5e942bb3 | 7452 | update_context_time(child_ctx); |
e625cce1 | 7453 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
7454 | |
7455 | /* | |
cdd6c482 IM |
7456 | * Report the task dead after unscheduling the events so that we |
7457 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7458 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7459 | */ |
cdd6c482 | 7460 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7461 | |
66fff224 PZ |
7462 | /* |
7463 | * We can recurse on the same lock type through: | |
7464 | * | |
cdd6c482 IM |
7465 | * __perf_event_exit_task() |
7466 | * sync_child_event() | |
a6fa941d AV |
7467 | * put_event() |
7468 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7469 | * |
7470 | * But since its the parent context it won't be the same instance. | |
7471 | */ | |
a0507c84 | 7472 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7473 | |
8bc20959 | 7474 | again: |
889ff015 FW |
7475 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
7476 | group_entry) | |
7477 | __perf_event_exit_task(child_event, child_ctx, child); | |
7478 | ||
7479 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 7480 | group_entry) |
cdd6c482 | 7481 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
7482 | |
7483 | /* | |
cdd6c482 | 7484 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
7485 | * its siblings to the list, but we obtained 'tmp' before that which |
7486 | * will still point to the list head terminating the iteration. | |
7487 | */ | |
889ff015 FW |
7488 | if (!list_empty(&child_ctx->pinned_groups) || |
7489 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 7490 | goto again; |
a63eaf34 PM |
7491 | |
7492 | mutex_unlock(&child_ctx->mutex); | |
7493 | ||
7494 | put_ctx(child_ctx); | |
9b51f66d IM |
7495 | } |
7496 | ||
8dc85d54 PZ |
7497 | /* |
7498 | * When a child task exits, feed back event values to parent events. | |
7499 | */ | |
7500 | void perf_event_exit_task(struct task_struct *child) | |
7501 | { | |
8882135b | 7502 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7503 | int ctxn; |
7504 | ||
8882135b PZ |
7505 | mutex_lock(&child->perf_event_mutex); |
7506 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7507 | owner_entry) { | |
7508 | list_del_init(&event->owner_entry); | |
7509 | ||
7510 | /* | |
7511 | * Ensure the list deletion is visible before we clear | |
7512 | * the owner, closes a race against perf_release() where | |
7513 | * we need to serialize on the owner->perf_event_mutex. | |
7514 | */ | |
7515 | smp_wmb(); | |
7516 | event->owner = NULL; | |
7517 | } | |
7518 | mutex_unlock(&child->perf_event_mutex); | |
7519 | ||
8dc85d54 PZ |
7520 | for_each_task_context_nr(ctxn) |
7521 | perf_event_exit_task_context(child, ctxn); | |
7522 | } | |
7523 | ||
889ff015 FW |
7524 | static void perf_free_event(struct perf_event *event, |
7525 | struct perf_event_context *ctx) | |
7526 | { | |
7527 | struct perf_event *parent = event->parent; | |
7528 | ||
7529 | if (WARN_ON_ONCE(!parent)) | |
7530 | return; | |
7531 | ||
7532 | mutex_lock(&parent->child_mutex); | |
7533 | list_del_init(&event->child_list); | |
7534 | mutex_unlock(&parent->child_mutex); | |
7535 | ||
a6fa941d | 7536 | put_event(parent); |
889ff015 | 7537 | |
8a49542c | 7538 | perf_group_detach(event); |
889ff015 FW |
7539 | list_del_event(event, ctx); |
7540 | free_event(event); | |
7541 | } | |
7542 | ||
bbbee908 PZ |
7543 | /* |
7544 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7545 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7546 | */ |
cdd6c482 | 7547 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7548 | { |
8dc85d54 | 7549 | struct perf_event_context *ctx; |
cdd6c482 | 7550 | struct perf_event *event, *tmp; |
8dc85d54 | 7551 | int ctxn; |
bbbee908 | 7552 | |
8dc85d54 PZ |
7553 | for_each_task_context_nr(ctxn) { |
7554 | ctx = task->perf_event_ctxp[ctxn]; | |
7555 | if (!ctx) | |
7556 | continue; | |
bbbee908 | 7557 | |
8dc85d54 | 7558 | mutex_lock(&ctx->mutex); |
bbbee908 | 7559 | again: |
8dc85d54 PZ |
7560 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7561 | group_entry) | |
7562 | perf_free_event(event, ctx); | |
bbbee908 | 7563 | |
8dc85d54 PZ |
7564 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7565 | group_entry) | |
7566 | perf_free_event(event, ctx); | |
bbbee908 | 7567 | |
8dc85d54 PZ |
7568 | if (!list_empty(&ctx->pinned_groups) || |
7569 | !list_empty(&ctx->flexible_groups)) | |
7570 | goto again; | |
bbbee908 | 7571 | |
8dc85d54 | 7572 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7573 | |
8dc85d54 PZ |
7574 | put_ctx(ctx); |
7575 | } | |
889ff015 FW |
7576 | } |
7577 | ||
4e231c79 PZ |
7578 | void perf_event_delayed_put(struct task_struct *task) |
7579 | { | |
7580 | int ctxn; | |
7581 | ||
7582 | for_each_task_context_nr(ctxn) | |
7583 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7584 | } | |
7585 | ||
97dee4f3 PZ |
7586 | /* |
7587 | * inherit a event from parent task to child task: | |
7588 | */ | |
7589 | static struct perf_event * | |
7590 | inherit_event(struct perf_event *parent_event, | |
7591 | struct task_struct *parent, | |
7592 | struct perf_event_context *parent_ctx, | |
7593 | struct task_struct *child, | |
7594 | struct perf_event *group_leader, | |
7595 | struct perf_event_context *child_ctx) | |
7596 | { | |
7597 | struct perf_event *child_event; | |
cee010ec | 7598 | unsigned long flags; |
97dee4f3 PZ |
7599 | |
7600 | /* | |
7601 | * Instead of creating recursive hierarchies of events, | |
7602 | * we link inherited events back to the original parent, | |
7603 | * which has a filp for sure, which we use as the reference | |
7604 | * count: | |
7605 | */ | |
7606 | if (parent_event->parent) | |
7607 | parent_event = parent_event->parent; | |
7608 | ||
7609 | child_event = perf_event_alloc(&parent_event->attr, | |
7610 | parent_event->cpu, | |
d580ff86 | 7611 | child, |
97dee4f3 | 7612 | group_leader, parent_event, |
4dc0da86 | 7613 | NULL, NULL); |
97dee4f3 PZ |
7614 | if (IS_ERR(child_event)) |
7615 | return child_event; | |
a6fa941d AV |
7616 | |
7617 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7618 | free_event(child_event); | |
7619 | return NULL; | |
7620 | } | |
7621 | ||
97dee4f3 PZ |
7622 | get_ctx(child_ctx); |
7623 | ||
7624 | /* | |
7625 | * Make the child state follow the state of the parent event, | |
7626 | * not its attr.disabled bit. We hold the parent's mutex, | |
7627 | * so we won't race with perf_event_{en, dis}able_family. | |
7628 | */ | |
7629 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7630 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7631 | else | |
7632 | child_event->state = PERF_EVENT_STATE_OFF; | |
7633 | ||
7634 | if (parent_event->attr.freq) { | |
7635 | u64 sample_period = parent_event->hw.sample_period; | |
7636 | struct hw_perf_event *hwc = &child_event->hw; | |
7637 | ||
7638 | hwc->sample_period = sample_period; | |
7639 | hwc->last_period = sample_period; | |
7640 | ||
7641 | local64_set(&hwc->period_left, sample_period); | |
7642 | } | |
7643 | ||
7644 | child_event->ctx = child_ctx; | |
7645 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7646 | child_event->overflow_handler_context |
7647 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7648 | |
614b6780 TG |
7649 | /* |
7650 | * Precalculate sample_data sizes | |
7651 | */ | |
7652 | perf_event__header_size(child_event); | |
6844c09d | 7653 | perf_event__id_header_size(child_event); |
614b6780 | 7654 | |
97dee4f3 PZ |
7655 | /* |
7656 | * Link it up in the child's context: | |
7657 | */ | |
cee010ec | 7658 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7659 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7660 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7661 | |
97dee4f3 PZ |
7662 | /* |
7663 | * Link this into the parent event's child list | |
7664 | */ | |
7665 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7666 | mutex_lock(&parent_event->child_mutex); | |
7667 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7668 | mutex_unlock(&parent_event->child_mutex); | |
7669 | ||
7670 | return child_event; | |
7671 | } | |
7672 | ||
7673 | static int inherit_group(struct perf_event *parent_event, | |
7674 | struct task_struct *parent, | |
7675 | struct perf_event_context *parent_ctx, | |
7676 | struct task_struct *child, | |
7677 | struct perf_event_context *child_ctx) | |
7678 | { | |
7679 | struct perf_event *leader; | |
7680 | struct perf_event *sub; | |
7681 | struct perf_event *child_ctr; | |
7682 | ||
7683 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7684 | child, NULL, child_ctx); | |
7685 | if (IS_ERR(leader)) | |
7686 | return PTR_ERR(leader); | |
7687 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7688 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7689 | child, leader, child_ctx); | |
7690 | if (IS_ERR(child_ctr)) | |
7691 | return PTR_ERR(child_ctr); | |
7692 | } | |
7693 | return 0; | |
889ff015 FW |
7694 | } |
7695 | ||
7696 | static int | |
7697 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7698 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7699 | struct task_struct *child, int ctxn, |
889ff015 FW |
7700 | int *inherited_all) |
7701 | { | |
7702 | int ret; | |
8dc85d54 | 7703 | struct perf_event_context *child_ctx; |
889ff015 FW |
7704 | |
7705 | if (!event->attr.inherit) { | |
7706 | *inherited_all = 0; | |
7707 | return 0; | |
bbbee908 PZ |
7708 | } |
7709 | ||
fe4b04fa | 7710 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7711 | if (!child_ctx) { |
7712 | /* | |
7713 | * This is executed from the parent task context, so | |
7714 | * inherit events that have been marked for cloning. | |
7715 | * First allocate and initialize a context for the | |
7716 | * child. | |
7717 | */ | |
bbbee908 | 7718 | |
734df5ab | 7719 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7720 | if (!child_ctx) |
7721 | return -ENOMEM; | |
bbbee908 | 7722 | |
8dc85d54 | 7723 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7724 | } |
7725 | ||
7726 | ret = inherit_group(event, parent, parent_ctx, | |
7727 | child, child_ctx); | |
7728 | ||
7729 | if (ret) | |
7730 | *inherited_all = 0; | |
7731 | ||
7732 | return ret; | |
bbbee908 PZ |
7733 | } |
7734 | ||
9b51f66d | 7735 | /* |
cdd6c482 | 7736 | * Initialize the perf_event context in task_struct |
9b51f66d | 7737 | */ |
8dc85d54 | 7738 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7739 | { |
889ff015 | 7740 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7741 | struct perf_event_context *cloned_ctx; |
7742 | struct perf_event *event; | |
9b51f66d | 7743 | struct task_struct *parent = current; |
564c2b21 | 7744 | int inherited_all = 1; |
dddd3379 | 7745 | unsigned long flags; |
6ab423e0 | 7746 | int ret = 0; |
9b51f66d | 7747 | |
8dc85d54 | 7748 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7749 | return 0; |
7750 | ||
ad3a37de | 7751 | /* |
25346b93 PM |
7752 | * If the parent's context is a clone, pin it so it won't get |
7753 | * swapped under us. | |
ad3a37de | 7754 | */ |
8dc85d54 | 7755 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
7756 | if (!parent_ctx) |
7757 | return 0; | |
25346b93 | 7758 | |
ad3a37de PM |
7759 | /* |
7760 | * No need to check if parent_ctx != NULL here; since we saw | |
7761 | * it non-NULL earlier, the only reason for it to become NULL | |
7762 | * is if we exit, and since we're currently in the middle of | |
7763 | * a fork we can't be exiting at the same time. | |
7764 | */ | |
ad3a37de | 7765 | |
9b51f66d IM |
7766 | /* |
7767 | * Lock the parent list. No need to lock the child - not PID | |
7768 | * hashed yet and not running, so nobody can access it. | |
7769 | */ | |
d859e29f | 7770 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7771 | |
7772 | /* | |
7773 | * We dont have to disable NMIs - we are only looking at | |
7774 | * the list, not manipulating it: | |
7775 | */ | |
889ff015 | 7776 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7777 | ret = inherit_task_group(event, parent, parent_ctx, |
7778 | child, ctxn, &inherited_all); | |
889ff015 FW |
7779 | if (ret) |
7780 | break; | |
7781 | } | |
b93f7978 | 7782 | |
dddd3379 TG |
7783 | /* |
7784 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7785 | * to allocations, but we need to prevent rotation because | |
7786 | * rotate_ctx() will change the list from interrupt context. | |
7787 | */ | |
7788 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7789 | parent_ctx->rotate_disable = 1; | |
7790 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7791 | ||
889ff015 | 7792 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7793 | ret = inherit_task_group(event, parent, parent_ctx, |
7794 | child, ctxn, &inherited_all); | |
889ff015 | 7795 | if (ret) |
9b51f66d | 7796 | break; |
564c2b21 PM |
7797 | } |
7798 | ||
dddd3379 TG |
7799 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7800 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7801 | |
8dc85d54 | 7802 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7803 | |
05cbaa28 | 7804 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7805 | /* |
7806 | * Mark the child context as a clone of the parent | |
7807 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7808 | * |
7809 | * Note that if the parent is a clone, the holding of | |
7810 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7811 | */ |
c5ed5145 | 7812 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7813 | if (cloned_ctx) { |
7814 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7815 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7816 | } else { |
7817 | child_ctx->parent_ctx = parent_ctx; | |
7818 | child_ctx->parent_gen = parent_ctx->generation; | |
7819 | } | |
7820 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7821 | } |
7822 | ||
c5ed5145 | 7823 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7824 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7825 | |
25346b93 | 7826 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7827 | put_ctx(parent_ctx); |
ad3a37de | 7828 | |
6ab423e0 | 7829 | return ret; |
9b51f66d IM |
7830 | } |
7831 | ||
8dc85d54 PZ |
7832 | /* |
7833 | * Initialize the perf_event context in task_struct | |
7834 | */ | |
7835 | int perf_event_init_task(struct task_struct *child) | |
7836 | { | |
7837 | int ctxn, ret; | |
7838 | ||
8550d7cb ON |
7839 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7840 | mutex_init(&child->perf_event_mutex); | |
7841 | INIT_LIST_HEAD(&child->perf_event_list); | |
7842 | ||
8dc85d54 PZ |
7843 | for_each_task_context_nr(ctxn) { |
7844 | ret = perf_event_init_context(child, ctxn); | |
7845 | if (ret) | |
7846 | return ret; | |
7847 | } | |
7848 | ||
7849 | return 0; | |
7850 | } | |
7851 | ||
220b140b PM |
7852 | static void __init perf_event_init_all_cpus(void) |
7853 | { | |
b28ab83c | 7854 | struct swevent_htable *swhash; |
220b140b | 7855 | int cpu; |
220b140b PM |
7856 | |
7857 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7858 | swhash = &per_cpu(swevent_htable, cpu); |
7859 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7860 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7861 | } |
7862 | } | |
7863 | ||
0db0628d | 7864 | static void perf_event_init_cpu(int cpu) |
0793a61d | 7865 | { |
108b02cf | 7866 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7867 | |
b28ab83c | 7868 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7869 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7870 | struct swevent_hlist *hlist; |
7871 | ||
b28ab83c PZ |
7872 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7873 | WARN_ON(!hlist); | |
7874 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7875 | } |
b28ab83c | 7876 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7877 | } |
7878 | ||
c277443c | 7879 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7880 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7881 | { |
e9d2b064 PZ |
7882 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7883 | ||
7884 | WARN_ON(!irqs_disabled()); | |
7885 | ||
7886 | list_del_init(&cpuctx->rotation_list); | |
7887 | } | |
7888 | ||
108b02cf | 7889 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7890 | { |
46ce0fe9 | 7891 | struct remove_event re = { .detach_group = false }; |
108b02cf | 7892 | struct perf_event_context *ctx = __info; |
0793a61d | 7893 | |
108b02cf | 7894 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7895 | |
e3703f8c | 7896 | rcu_read_lock(); |
46ce0fe9 PZ |
7897 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
7898 | __perf_remove_from_context(&re); | |
e3703f8c | 7899 | rcu_read_unlock(); |
0793a61d | 7900 | } |
108b02cf PZ |
7901 | |
7902 | static void perf_event_exit_cpu_context(int cpu) | |
7903 | { | |
7904 | struct perf_event_context *ctx; | |
7905 | struct pmu *pmu; | |
7906 | int idx; | |
7907 | ||
7908 | idx = srcu_read_lock(&pmus_srcu); | |
7909 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7910 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7911 | |
7912 | mutex_lock(&ctx->mutex); | |
7913 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7914 | mutex_unlock(&ctx->mutex); | |
7915 | } | |
7916 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7917 | } |
7918 | ||
cdd6c482 | 7919 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7920 | { |
b28ab83c | 7921 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7922 | |
e3703f8c PZ |
7923 | perf_event_exit_cpu_context(cpu); |
7924 | ||
b28ab83c PZ |
7925 | mutex_lock(&swhash->hlist_mutex); |
7926 | swevent_hlist_release(swhash); | |
7927 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
7928 | } |
7929 | #else | |
cdd6c482 | 7930 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7931 | #endif |
7932 | ||
c277443c PZ |
7933 | static int |
7934 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7935 | { | |
7936 | int cpu; | |
7937 | ||
7938 | for_each_online_cpu(cpu) | |
7939 | perf_event_exit_cpu(cpu); | |
7940 | ||
7941 | return NOTIFY_OK; | |
7942 | } | |
7943 | ||
7944 | /* | |
7945 | * Run the perf reboot notifier at the very last possible moment so that | |
7946 | * the generic watchdog code runs as long as possible. | |
7947 | */ | |
7948 | static struct notifier_block perf_reboot_notifier = { | |
7949 | .notifier_call = perf_reboot, | |
7950 | .priority = INT_MIN, | |
7951 | }; | |
7952 | ||
0db0628d | 7953 | static int |
0793a61d TG |
7954 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
7955 | { | |
7956 | unsigned int cpu = (long)hcpu; | |
7957 | ||
4536e4d1 | 7958 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7959 | |
7960 | case CPU_UP_PREPARE: | |
5e11637e | 7961 | case CPU_DOWN_FAILED: |
cdd6c482 | 7962 | perf_event_init_cpu(cpu); |
0793a61d TG |
7963 | break; |
7964 | ||
5e11637e | 7965 | case CPU_UP_CANCELED: |
0793a61d | 7966 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7967 | perf_event_exit_cpu(cpu); |
0793a61d | 7968 | break; |
0793a61d TG |
7969 | default: |
7970 | break; | |
7971 | } | |
7972 | ||
7973 | return NOTIFY_OK; | |
7974 | } | |
7975 | ||
cdd6c482 | 7976 | void __init perf_event_init(void) |
0793a61d | 7977 | { |
3c502e7a JW |
7978 | int ret; |
7979 | ||
2e80a82a PZ |
7980 | idr_init(&pmu_idr); |
7981 | ||
220b140b | 7982 | perf_event_init_all_cpus(); |
b0a873eb | 7983 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7984 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7985 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7986 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7987 | perf_tp_register(); |
7988 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7989 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7990 | |
7991 | ret = init_hw_breakpoint(); | |
7992 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7993 | |
7994 | /* do not patch jump label more than once per second */ | |
7995 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7996 | |
7997 | /* | |
7998 | * Build time assertion that we keep the data_head at the intended | |
7999 | * location. IOW, validation we got the __reserved[] size right. | |
8000 | */ | |
8001 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
8002 | != 1024); | |
0793a61d | 8003 | } |
abe43400 PZ |
8004 | |
8005 | static int __init perf_event_sysfs_init(void) | |
8006 | { | |
8007 | struct pmu *pmu; | |
8008 | int ret; | |
8009 | ||
8010 | mutex_lock(&pmus_lock); | |
8011 | ||
8012 | ret = bus_register(&pmu_bus); | |
8013 | if (ret) | |
8014 | goto unlock; | |
8015 | ||
8016 | list_for_each_entry(pmu, &pmus, entry) { | |
8017 | if (!pmu->name || pmu->type < 0) | |
8018 | continue; | |
8019 | ||
8020 | ret = pmu_dev_alloc(pmu); | |
8021 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8022 | } | |
8023 | pmu_bus_running = 1; | |
8024 | ret = 0; | |
8025 | ||
8026 | unlock: | |
8027 | mutex_unlock(&pmus_lock); | |
8028 | ||
8029 | return ret; | |
8030 | } | |
8031 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8032 | |
8033 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8034 | static struct cgroup_subsys_state * |
8035 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8036 | { |
8037 | struct perf_cgroup *jc; | |
e5d1367f | 8038 | |
1b15d055 | 8039 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8040 | if (!jc) |
8041 | return ERR_PTR(-ENOMEM); | |
8042 | ||
e5d1367f SE |
8043 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8044 | if (!jc->info) { | |
8045 | kfree(jc); | |
8046 | return ERR_PTR(-ENOMEM); | |
8047 | } | |
8048 | ||
e5d1367f SE |
8049 | return &jc->css; |
8050 | } | |
8051 | ||
eb95419b | 8052 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8053 | { |
eb95419b TH |
8054 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8055 | ||
e5d1367f SE |
8056 | free_percpu(jc->info); |
8057 | kfree(jc); | |
8058 | } | |
8059 | ||
8060 | static int __perf_cgroup_move(void *info) | |
8061 | { | |
8062 | struct task_struct *task = info; | |
8063 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8064 | return 0; | |
8065 | } | |
8066 | ||
eb95419b TH |
8067 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8068 | struct cgroup_taskset *tset) | |
e5d1367f | 8069 | { |
bb9d97b6 TH |
8070 | struct task_struct *task; |
8071 | ||
924f0d9a | 8072 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 8073 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8074 | } |
8075 | ||
eb95419b TH |
8076 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8077 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8078 | struct task_struct *task) |
e5d1367f SE |
8079 | { |
8080 | /* | |
8081 | * cgroup_exit() is called in the copy_process() failure path. | |
8082 | * Ignore this case since the task hasn't ran yet, this avoids | |
8083 | * trying to poke a half freed task state from generic code. | |
8084 | */ | |
8085 | if (!(task->flags & PF_EXITING)) | |
8086 | return; | |
8087 | ||
bb9d97b6 | 8088 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8089 | } |
8090 | ||
073219e9 | 8091 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
8092 | .css_alloc = perf_cgroup_css_alloc, |
8093 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8094 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8095 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8096 | }; |
8097 | #endif /* CONFIG_CGROUP_PERF */ |