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> |
39bed6cb | 37 | #include <linux/cgroup.h> |
cdd6c482 | 38 | #include <linux/perf_event.h> |
af658dca | 39 | #include <linux/trace_events.h> |
3c502e7a | 40 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 41 | #include <linux/mm_types.h> |
c464c76e | 42 | #include <linux/module.h> |
f972eb63 | 43 | #include <linux/mman.h> |
b3f20785 | 44 | #include <linux/compat.h> |
2541517c AS |
45 | #include <linux/bpf.h> |
46 | #include <linux/filter.h> | |
0793a61d | 47 | |
76369139 FW |
48 | #include "internal.h" |
49 | ||
4e193bd4 TB |
50 | #include <asm/irq_regs.h> |
51 | ||
fadfe7be JO |
52 | static struct workqueue_struct *perf_wq; |
53 | ||
272325c4 PZ |
54 | typedef int (*remote_function_f)(void *); |
55 | ||
fe4b04fa | 56 | struct remote_function_call { |
e7e7ee2e | 57 | struct task_struct *p; |
272325c4 | 58 | remote_function_f func; |
e7e7ee2e IM |
59 | void *info; |
60 | int ret; | |
fe4b04fa PZ |
61 | }; |
62 | ||
63 | static void remote_function(void *data) | |
64 | { | |
65 | struct remote_function_call *tfc = data; | |
66 | struct task_struct *p = tfc->p; | |
67 | ||
68 | if (p) { | |
69 | tfc->ret = -EAGAIN; | |
70 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
71 | return; | |
72 | } | |
73 | ||
74 | tfc->ret = tfc->func(tfc->info); | |
75 | } | |
76 | ||
77 | /** | |
78 | * task_function_call - call a function on the cpu on which a task runs | |
79 | * @p: the task to evaluate | |
80 | * @func: the function to be called | |
81 | * @info: the function call argument | |
82 | * | |
83 | * Calls the function @func when the task is currently running. This might | |
84 | * be on the current CPU, which just calls the function directly | |
85 | * | |
86 | * returns: @func return value, or | |
87 | * -ESRCH - when the process isn't running | |
88 | * -EAGAIN - when the process moved away | |
89 | */ | |
90 | static int | |
272325c4 | 91 | task_function_call(struct task_struct *p, remote_function_f func, void *info) |
fe4b04fa PZ |
92 | { |
93 | struct remote_function_call data = { | |
e7e7ee2e IM |
94 | .p = p, |
95 | .func = func, | |
96 | .info = info, | |
97 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
98 | }; |
99 | ||
100 | if (task_curr(p)) | |
101 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
102 | ||
103 | return data.ret; | |
104 | } | |
105 | ||
106 | /** | |
107 | * cpu_function_call - call a function on the cpu | |
108 | * @func: the function to be called | |
109 | * @info: the function call argument | |
110 | * | |
111 | * Calls the function @func on the remote cpu. | |
112 | * | |
113 | * returns: @func return value or -ENXIO when the cpu is offline | |
114 | */ | |
272325c4 | 115 | static int cpu_function_call(int cpu, remote_function_f func, void *info) |
fe4b04fa PZ |
116 | { |
117 | struct remote_function_call data = { | |
e7e7ee2e IM |
118 | .p = NULL, |
119 | .func = func, | |
120 | .info = info, | |
121 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
122 | }; |
123 | ||
124 | smp_call_function_single(cpu, remote_function, &data, 1); | |
125 | ||
126 | return data.ret; | |
127 | } | |
128 | ||
f8697762 JO |
129 | #define EVENT_OWNER_KERNEL ((void *) -1) |
130 | ||
131 | static bool is_kernel_event(struct perf_event *event) | |
132 | { | |
133 | return event->owner == EVENT_OWNER_KERNEL; | |
134 | } | |
135 | ||
e5d1367f SE |
136 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
137 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
138 | PERF_FLAG_PID_CGROUP |\ |
139 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 140 | |
bce38cd5 SE |
141 | /* |
142 | * branch priv levels that need permission checks | |
143 | */ | |
144 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
145 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
146 | PERF_SAMPLE_BRANCH_HV) | |
147 | ||
0b3fcf17 SE |
148 | enum event_type_t { |
149 | EVENT_FLEXIBLE = 0x1, | |
150 | EVENT_PINNED = 0x2, | |
151 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
152 | }; | |
153 | ||
e5d1367f SE |
154 | /* |
155 | * perf_sched_events : >0 events exist | |
156 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
157 | */ | |
c5905afb | 158 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 159 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
ba532500 | 160 | static DEFINE_PER_CPU(int, perf_sched_cb_usages); |
e5d1367f | 161 | |
cdd6c482 IM |
162 | static atomic_t nr_mmap_events __read_mostly; |
163 | static atomic_t nr_comm_events __read_mostly; | |
164 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 165 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 166 | |
108b02cf PZ |
167 | static LIST_HEAD(pmus); |
168 | static DEFINE_MUTEX(pmus_lock); | |
169 | static struct srcu_struct pmus_srcu; | |
170 | ||
0764771d | 171 | /* |
cdd6c482 | 172 | * perf event paranoia level: |
0fbdea19 IM |
173 | * -1 - not paranoid at all |
174 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 175 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 176 | * 2 - disallow kernel profiling for unpriv |
0764771d | 177 | */ |
cdd6c482 | 178 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 179 | |
20443384 FW |
180 | /* Minimum for 512 kiB + 1 user control page */ |
181 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
182 | |
183 | /* | |
cdd6c482 | 184 | * max perf event sample rate |
df58ab24 | 185 | */ |
14c63f17 DH |
186 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
187 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
188 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
189 | ||
190 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
191 | ||
192 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
193 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
194 | ||
d9494cb4 PZ |
195 | static int perf_sample_allowed_ns __read_mostly = |
196 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
197 | |
198 | void update_perf_cpu_limits(void) | |
199 | { | |
200 | u64 tmp = perf_sample_period_ns; | |
201 | ||
202 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 203 | do_div(tmp, 100); |
d9494cb4 | 204 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 205 | } |
163ec435 | 206 | |
9e630205 SE |
207 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
208 | ||
163ec435 PZ |
209 | int perf_proc_update_handler(struct ctl_table *table, int write, |
210 | void __user *buffer, size_t *lenp, | |
211 | loff_t *ppos) | |
212 | { | |
723478c8 | 213 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
214 | |
215 | if (ret || !write) | |
216 | return ret; | |
217 | ||
218 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
219 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
220 | update_perf_cpu_limits(); | |
221 | ||
222 | return 0; | |
223 | } | |
224 | ||
225 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
226 | ||
227 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
228 | void __user *buffer, size_t *lenp, | |
229 | loff_t *ppos) | |
230 | { | |
231 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
232 | ||
233 | if (ret || !write) | |
234 | return ret; | |
235 | ||
236 | update_perf_cpu_limits(); | |
163ec435 PZ |
237 | |
238 | return 0; | |
239 | } | |
1ccd1549 | 240 | |
14c63f17 DH |
241 | /* |
242 | * perf samples are done in some very critical code paths (NMIs). | |
243 | * If they take too much CPU time, the system can lock up and not | |
244 | * get any real work done. This will drop the sample rate when | |
245 | * we detect that events are taking too long. | |
246 | */ | |
247 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 248 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 249 | |
6a02ad66 | 250 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 251 | { |
6a02ad66 | 252 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 253 | u64 avg_local_sample_len; |
e5302920 | 254 | u64 local_samples_len; |
6a02ad66 | 255 | |
4a32fea9 | 256 | local_samples_len = __this_cpu_read(running_sample_length); |
6a02ad66 PZ |
257 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; |
258 | ||
259 | printk_ratelimited(KERN_WARNING | |
260 | "perf interrupt took too long (%lld > %lld), lowering " | |
261 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 262 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
263 | sysctl_perf_event_sample_rate); |
264 | } | |
265 | ||
266 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
267 | ||
268 | void perf_sample_event_took(u64 sample_len_ns) | |
269 | { | |
d9494cb4 | 270 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
271 | u64 avg_local_sample_len; |
272 | u64 local_samples_len; | |
14c63f17 | 273 | |
d9494cb4 | 274 | if (allowed_ns == 0) |
14c63f17 DH |
275 | return; |
276 | ||
277 | /* decay the counter by 1 average sample */ | |
4a32fea9 | 278 | local_samples_len = __this_cpu_read(running_sample_length); |
14c63f17 DH |
279 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; |
280 | local_samples_len += sample_len_ns; | |
4a32fea9 | 281 | __this_cpu_write(running_sample_length, local_samples_len); |
14c63f17 DH |
282 | |
283 | /* | |
284 | * note: this will be biased artifically low until we have | |
285 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
286 | * from having to maintain a count. | |
287 | */ | |
288 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
289 | ||
d9494cb4 | 290 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
291 | return; |
292 | ||
293 | if (max_samples_per_tick <= 1) | |
294 | return; | |
295 | ||
296 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
297 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
298 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
299 | ||
14c63f17 | 300 | update_perf_cpu_limits(); |
6a02ad66 | 301 | |
cd578abb PZ |
302 | if (!irq_work_queue(&perf_duration_work)) { |
303 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
304 | "kernel.perf_event_max_sample_rate to %d\n", | |
305 | avg_local_sample_len, allowed_ns >> 1, | |
306 | sysctl_perf_event_sample_rate); | |
307 | } | |
14c63f17 DH |
308 | } |
309 | ||
cdd6c482 | 310 | static atomic64_t perf_event_id; |
a96bbc16 | 311 | |
0b3fcf17 SE |
312 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
313 | enum event_type_t event_type); | |
314 | ||
315 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
316 | enum event_type_t event_type, |
317 | struct task_struct *task); | |
318 | ||
319 | static void update_context_time(struct perf_event_context *ctx); | |
320 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 321 | |
cdd6c482 | 322 | void __weak perf_event_print_debug(void) { } |
0793a61d | 323 | |
84c79910 | 324 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 325 | { |
84c79910 | 326 | return "pmu"; |
0793a61d TG |
327 | } |
328 | ||
0b3fcf17 SE |
329 | static inline u64 perf_clock(void) |
330 | { | |
331 | return local_clock(); | |
332 | } | |
333 | ||
34f43927 PZ |
334 | static inline u64 perf_event_clock(struct perf_event *event) |
335 | { | |
336 | return event->clock(); | |
337 | } | |
338 | ||
e5d1367f SE |
339 | static inline struct perf_cpu_context * |
340 | __get_cpu_context(struct perf_event_context *ctx) | |
341 | { | |
342 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
343 | } | |
344 | ||
facc4307 PZ |
345 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
346 | struct perf_event_context *ctx) | |
347 | { | |
348 | raw_spin_lock(&cpuctx->ctx.lock); | |
349 | if (ctx) | |
350 | raw_spin_lock(&ctx->lock); | |
351 | } | |
352 | ||
353 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
354 | struct perf_event_context *ctx) | |
355 | { | |
356 | if (ctx) | |
357 | raw_spin_unlock(&ctx->lock); | |
358 | raw_spin_unlock(&cpuctx->ctx.lock); | |
359 | } | |
360 | ||
e5d1367f SE |
361 | #ifdef CONFIG_CGROUP_PERF |
362 | ||
e5d1367f SE |
363 | static inline bool |
364 | perf_cgroup_match(struct perf_event *event) | |
365 | { | |
366 | struct perf_event_context *ctx = event->ctx; | |
367 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
368 | ||
ef824fa1 TH |
369 | /* @event doesn't care about cgroup */ |
370 | if (!event->cgrp) | |
371 | return true; | |
372 | ||
373 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
374 | if (!cpuctx->cgrp) | |
375 | return false; | |
376 | ||
377 | /* | |
378 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
379 | * also enabled for all its descendant cgroups. If @cpuctx's | |
380 | * cgroup is a descendant of @event's (the test covers identity | |
381 | * case), it's a match. | |
382 | */ | |
383 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
384 | event->cgrp->css.cgroup); | |
e5d1367f SE |
385 | } |
386 | ||
e5d1367f SE |
387 | static inline void perf_detach_cgroup(struct perf_event *event) |
388 | { | |
4e2ba650 | 389 | css_put(&event->cgrp->css); |
e5d1367f SE |
390 | event->cgrp = NULL; |
391 | } | |
392 | ||
393 | static inline int is_cgroup_event(struct perf_event *event) | |
394 | { | |
395 | return event->cgrp != NULL; | |
396 | } | |
397 | ||
398 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
399 | { | |
400 | struct perf_cgroup_info *t; | |
401 | ||
402 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
403 | return t->time; | |
404 | } | |
405 | ||
406 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
407 | { | |
408 | struct perf_cgroup_info *info; | |
409 | u64 now; | |
410 | ||
411 | now = perf_clock(); | |
412 | ||
413 | info = this_cpu_ptr(cgrp->info); | |
414 | ||
415 | info->time += now - info->timestamp; | |
416 | info->timestamp = now; | |
417 | } | |
418 | ||
419 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
420 | { | |
421 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
422 | if (cgrp_out) | |
423 | __update_cgrp_time(cgrp_out); | |
424 | } | |
425 | ||
426 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
427 | { | |
3f7cce3c SE |
428 | struct perf_cgroup *cgrp; |
429 | ||
e5d1367f | 430 | /* |
3f7cce3c SE |
431 | * ensure we access cgroup data only when needed and |
432 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 433 | */ |
3f7cce3c | 434 | if (!is_cgroup_event(event)) |
e5d1367f SE |
435 | return; |
436 | ||
3f7cce3c SE |
437 | cgrp = perf_cgroup_from_task(current); |
438 | /* | |
439 | * Do not update time when cgroup is not active | |
440 | */ | |
441 | if (cgrp == event->cgrp) | |
442 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
443 | } |
444 | ||
445 | static inline void | |
3f7cce3c SE |
446 | perf_cgroup_set_timestamp(struct task_struct *task, |
447 | struct perf_event_context *ctx) | |
e5d1367f SE |
448 | { |
449 | struct perf_cgroup *cgrp; | |
450 | struct perf_cgroup_info *info; | |
451 | ||
3f7cce3c SE |
452 | /* |
453 | * ctx->lock held by caller | |
454 | * ensure we do not access cgroup data | |
455 | * unless we have the cgroup pinned (css_get) | |
456 | */ | |
457 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
458 | return; |
459 | ||
460 | cgrp = perf_cgroup_from_task(task); | |
461 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 462 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
463 | } |
464 | ||
465 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
466 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
467 | ||
468 | /* | |
469 | * reschedule events based on the cgroup constraint of task. | |
470 | * | |
471 | * mode SWOUT : schedule out everything | |
472 | * mode SWIN : schedule in based on cgroup for next | |
473 | */ | |
474 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
475 | { | |
476 | struct perf_cpu_context *cpuctx; | |
477 | struct pmu *pmu; | |
478 | unsigned long flags; | |
479 | ||
480 | /* | |
481 | * disable interrupts to avoid geting nr_cgroup | |
482 | * changes via __perf_event_disable(). Also | |
483 | * avoids preemption. | |
484 | */ | |
485 | local_irq_save(flags); | |
486 | ||
487 | /* | |
488 | * we reschedule only in the presence of cgroup | |
489 | * constrained events. | |
490 | */ | |
491 | rcu_read_lock(); | |
492 | ||
493 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 494 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
495 | if (cpuctx->unique_pmu != pmu) |
496 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 497 | |
e5d1367f SE |
498 | /* |
499 | * perf_cgroup_events says at least one | |
500 | * context on this CPU has cgroup events. | |
501 | * | |
502 | * ctx->nr_cgroups reports the number of cgroup | |
503 | * events for a context. | |
504 | */ | |
505 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
506 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
507 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
508 | |
509 | if (mode & PERF_CGROUP_SWOUT) { | |
510 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
511 | /* | |
512 | * must not be done before ctxswout due | |
513 | * to event_filter_match() in event_sched_out() | |
514 | */ | |
515 | cpuctx->cgrp = NULL; | |
516 | } | |
517 | ||
518 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 519 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
520 | /* |
521 | * set cgrp before ctxsw in to allow | |
522 | * event_filter_match() to not have to pass | |
523 | * task around | |
e5d1367f SE |
524 | */ |
525 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
526 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
527 | } | |
facc4307 PZ |
528 | perf_pmu_enable(cpuctx->ctx.pmu); |
529 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 530 | } |
e5d1367f SE |
531 | } |
532 | ||
533 | rcu_read_unlock(); | |
534 | ||
535 | local_irq_restore(flags); | |
536 | } | |
537 | ||
a8d757ef SE |
538 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
539 | struct task_struct *next) | |
e5d1367f | 540 | { |
a8d757ef SE |
541 | struct perf_cgroup *cgrp1; |
542 | struct perf_cgroup *cgrp2 = NULL; | |
543 | ||
544 | /* | |
545 | * we come here when we know perf_cgroup_events > 0 | |
546 | */ | |
547 | cgrp1 = perf_cgroup_from_task(task); | |
548 | ||
549 | /* | |
550 | * next is NULL when called from perf_event_enable_on_exec() | |
551 | * that will systematically cause a cgroup_switch() | |
552 | */ | |
553 | if (next) | |
554 | cgrp2 = perf_cgroup_from_task(next); | |
555 | ||
556 | /* | |
557 | * only schedule out current cgroup events if we know | |
558 | * that we are switching to a different cgroup. Otherwise, | |
559 | * do no touch the cgroup events. | |
560 | */ | |
561 | if (cgrp1 != cgrp2) | |
562 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
563 | } |
564 | ||
a8d757ef SE |
565 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
566 | struct task_struct *task) | |
e5d1367f | 567 | { |
a8d757ef SE |
568 | struct perf_cgroup *cgrp1; |
569 | struct perf_cgroup *cgrp2 = NULL; | |
570 | ||
571 | /* | |
572 | * we come here when we know perf_cgroup_events > 0 | |
573 | */ | |
574 | cgrp1 = perf_cgroup_from_task(task); | |
575 | ||
576 | /* prev can never be NULL */ | |
577 | cgrp2 = perf_cgroup_from_task(prev); | |
578 | ||
579 | /* | |
580 | * only need to schedule in cgroup events if we are changing | |
581 | * cgroup during ctxsw. Cgroup events were not scheduled | |
582 | * out of ctxsw out if that was not the case. | |
583 | */ | |
584 | if (cgrp1 != cgrp2) | |
585 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
586 | } |
587 | ||
588 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
589 | struct perf_event_attr *attr, | |
590 | struct perf_event *group_leader) | |
591 | { | |
592 | struct perf_cgroup *cgrp; | |
593 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
594 | struct fd f = fdget(fd); |
595 | int ret = 0; | |
e5d1367f | 596 | |
2903ff01 | 597 | if (!f.file) |
e5d1367f SE |
598 | return -EBADF; |
599 | ||
b583043e | 600 | css = css_tryget_online_from_dir(f.file->f_path.dentry, |
ec903c0c | 601 | &perf_event_cgrp_subsys); |
3db272c0 LZ |
602 | if (IS_ERR(css)) { |
603 | ret = PTR_ERR(css); | |
604 | goto out; | |
605 | } | |
e5d1367f SE |
606 | |
607 | cgrp = container_of(css, struct perf_cgroup, css); | |
608 | event->cgrp = cgrp; | |
609 | ||
610 | /* | |
611 | * all events in a group must monitor | |
612 | * the same cgroup because a task belongs | |
613 | * to only one perf cgroup at a time | |
614 | */ | |
615 | if (group_leader && group_leader->cgrp != cgrp) { | |
616 | perf_detach_cgroup(event); | |
617 | ret = -EINVAL; | |
e5d1367f | 618 | } |
3db272c0 | 619 | out: |
2903ff01 | 620 | fdput(f); |
e5d1367f SE |
621 | return ret; |
622 | } | |
623 | ||
624 | static inline void | |
625 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
626 | { | |
627 | struct perf_cgroup_info *t; | |
628 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
629 | event->shadow_ctx_time = now - t->timestamp; | |
630 | } | |
631 | ||
632 | static inline void | |
633 | perf_cgroup_defer_enabled(struct perf_event *event) | |
634 | { | |
635 | /* | |
636 | * when the current task's perf cgroup does not match | |
637 | * the event's, we need to remember to call the | |
638 | * perf_mark_enable() function the first time a task with | |
639 | * a matching perf cgroup is scheduled in. | |
640 | */ | |
641 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
642 | event->cgrp_defer_enabled = 1; | |
643 | } | |
644 | ||
645 | static inline void | |
646 | perf_cgroup_mark_enabled(struct perf_event *event, | |
647 | struct perf_event_context *ctx) | |
648 | { | |
649 | struct perf_event *sub; | |
650 | u64 tstamp = perf_event_time(event); | |
651 | ||
652 | if (!event->cgrp_defer_enabled) | |
653 | return; | |
654 | ||
655 | event->cgrp_defer_enabled = 0; | |
656 | ||
657 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
658 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
659 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
660 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
661 | sub->cgrp_defer_enabled = 0; | |
662 | } | |
663 | } | |
664 | } | |
665 | #else /* !CONFIG_CGROUP_PERF */ | |
666 | ||
667 | static inline bool | |
668 | perf_cgroup_match(struct perf_event *event) | |
669 | { | |
670 | return true; | |
671 | } | |
672 | ||
673 | static inline void perf_detach_cgroup(struct perf_event *event) | |
674 | {} | |
675 | ||
676 | static inline int is_cgroup_event(struct perf_event *event) | |
677 | { | |
678 | return 0; | |
679 | } | |
680 | ||
681 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
682 | { | |
683 | return 0; | |
684 | } | |
685 | ||
686 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
687 | { | |
688 | } | |
689 | ||
690 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
691 | { | |
692 | } | |
693 | ||
a8d757ef SE |
694 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
695 | struct task_struct *next) | |
e5d1367f SE |
696 | { |
697 | } | |
698 | ||
a8d757ef SE |
699 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
700 | struct task_struct *task) | |
e5d1367f SE |
701 | { |
702 | } | |
703 | ||
704 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
705 | struct perf_event_attr *attr, | |
706 | struct perf_event *group_leader) | |
707 | { | |
708 | return -EINVAL; | |
709 | } | |
710 | ||
711 | static inline void | |
3f7cce3c SE |
712 | perf_cgroup_set_timestamp(struct task_struct *task, |
713 | struct perf_event_context *ctx) | |
e5d1367f SE |
714 | { |
715 | } | |
716 | ||
717 | void | |
718 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
719 | { | |
720 | } | |
721 | ||
722 | static inline void | |
723 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
724 | { | |
725 | } | |
726 | ||
727 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
728 | { | |
729 | return 0; | |
730 | } | |
731 | ||
732 | static inline void | |
733 | perf_cgroup_defer_enabled(struct perf_event *event) | |
734 | { | |
735 | } | |
736 | ||
737 | static inline void | |
738 | perf_cgroup_mark_enabled(struct perf_event *event, | |
739 | struct perf_event_context *ctx) | |
740 | { | |
741 | } | |
742 | #endif | |
743 | ||
9e630205 SE |
744 | /* |
745 | * set default to be dependent on timer tick just | |
746 | * like original code | |
747 | */ | |
748 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
749 | /* | |
750 | * function must be called with interrupts disbled | |
751 | */ | |
272325c4 | 752 | static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr) |
9e630205 SE |
753 | { |
754 | struct perf_cpu_context *cpuctx; | |
9e630205 SE |
755 | int rotations = 0; |
756 | ||
757 | WARN_ON(!irqs_disabled()); | |
758 | ||
759 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
9e630205 SE |
760 | rotations = perf_rotate_context(cpuctx); |
761 | ||
4cfafd30 PZ |
762 | raw_spin_lock(&cpuctx->hrtimer_lock); |
763 | if (rotations) | |
9e630205 | 764 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); |
4cfafd30 PZ |
765 | else |
766 | cpuctx->hrtimer_active = 0; | |
767 | raw_spin_unlock(&cpuctx->hrtimer_lock); | |
9e630205 | 768 | |
4cfafd30 | 769 | return rotations ? HRTIMER_RESTART : HRTIMER_NORESTART; |
9e630205 SE |
770 | } |
771 | ||
272325c4 | 772 | static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) |
9e630205 | 773 | { |
272325c4 | 774 | struct hrtimer *timer = &cpuctx->hrtimer; |
9e630205 | 775 | struct pmu *pmu = cpuctx->ctx.pmu; |
272325c4 | 776 | u64 interval; |
9e630205 SE |
777 | |
778 | /* no multiplexing needed for SW PMU */ | |
779 | if (pmu->task_ctx_nr == perf_sw_context) | |
780 | return; | |
781 | ||
62b85639 SE |
782 | /* |
783 | * check default is sane, if not set then force to | |
784 | * default interval (1/tick) | |
785 | */ | |
272325c4 PZ |
786 | interval = pmu->hrtimer_interval_ms; |
787 | if (interval < 1) | |
788 | interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
62b85639 | 789 | |
272325c4 | 790 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval); |
9e630205 | 791 | |
4cfafd30 PZ |
792 | raw_spin_lock_init(&cpuctx->hrtimer_lock); |
793 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); | |
272325c4 | 794 | timer->function = perf_mux_hrtimer_handler; |
9e630205 SE |
795 | } |
796 | ||
272325c4 | 797 | static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx) |
9e630205 | 798 | { |
272325c4 | 799 | struct hrtimer *timer = &cpuctx->hrtimer; |
9e630205 | 800 | struct pmu *pmu = cpuctx->ctx.pmu; |
4cfafd30 | 801 | unsigned long flags; |
9e630205 SE |
802 | |
803 | /* not for SW PMU */ | |
804 | if (pmu->task_ctx_nr == perf_sw_context) | |
272325c4 | 805 | return 0; |
9e630205 | 806 | |
4cfafd30 PZ |
807 | raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags); |
808 | if (!cpuctx->hrtimer_active) { | |
809 | cpuctx->hrtimer_active = 1; | |
810 | hrtimer_forward_now(timer, cpuctx->hrtimer_interval); | |
811 | hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED); | |
812 | } | |
813 | raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags); | |
9e630205 | 814 | |
272325c4 | 815 | return 0; |
9e630205 SE |
816 | } |
817 | ||
33696fc0 | 818 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 819 | { |
33696fc0 PZ |
820 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
821 | if (!(*count)++) | |
822 | pmu->pmu_disable(pmu); | |
9e35ad38 | 823 | } |
9e35ad38 | 824 | |
33696fc0 | 825 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 826 | { |
33696fc0 PZ |
827 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
828 | if (!--(*count)) | |
829 | pmu->pmu_enable(pmu); | |
9e35ad38 | 830 | } |
9e35ad38 | 831 | |
2fde4f94 | 832 | static DEFINE_PER_CPU(struct list_head, active_ctx_list); |
e9d2b064 PZ |
833 | |
834 | /* | |
2fde4f94 MR |
835 | * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and |
836 | * perf_event_task_tick() are fully serialized because they're strictly cpu | |
837 | * affine and perf_event_ctx{activate,deactivate} are called with IRQs | |
838 | * disabled, while perf_event_task_tick is called from IRQ context. | |
e9d2b064 | 839 | */ |
2fde4f94 | 840 | static void perf_event_ctx_activate(struct perf_event_context *ctx) |
9e35ad38 | 841 | { |
2fde4f94 | 842 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
b5ab4cd5 | 843 | |
e9d2b064 | 844 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 845 | |
2fde4f94 MR |
846 | WARN_ON(!list_empty(&ctx->active_ctx_list)); |
847 | ||
848 | list_add(&ctx->active_ctx_list, head); | |
849 | } | |
850 | ||
851 | static void perf_event_ctx_deactivate(struct perf_event_context *ctx) | |
852 | { | |
853 | WARN_ON(!irqs_disabled()); | |
854 | ||
855 | WARN_ON(list_empty(&ctx->active_ctx_list)); | |
856 | ||
857 | list_del_init(&ctx->active_ctx_list); | |
9e35ad38 | 858 | } |
9e35ad38 | 859 | |
cdd6c482 | 860 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 861 | { |
e5289d4a | 862 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
863 | } |
864 | ||
4af57ef2 YZ |
865 | static void free_ctx(struct rcu_head *head) |
866 | { | |
867 | struct perf_event_context *ctx; | |
868 | ||
869 | ctx = container_of(head, struct perf_event_context, rcu_head); | |
870 | kfree(ctx->task_ctx_data); | |
871 | kfree(ctx); | |
872 | } | |
873 | ||
cdd6c482 | 874 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 875 | { |
564c2b21 PM |
876 | if (atomic_dec_and_test(&ctx->refcount)) { |
877 | if (ctx->parent_ctx) | |
878 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
879 | if (ctx->task) |
880 | put_task_struct(ctx->task); | |
4af57ef2 | 881 | call_rcu(&ctx->rcu_head, free_ctx); |
564c2b21 | 882 | } |
a63eaf34 PM |
883 | } |
884 | ||
f63a8daa PZ |
885 | /* |
886 | * Because of perf_event::ctx migration in sys_perf_event_open::move_group and | |
887 | * perf_pmu_migrate_context() we need some magic. | |
888 | * | |
889 | * Those places that change perf_event::ctx will hold both | |
890 | * perf_event_ctx::mutex of the 'old' and 'new' ctx value. | |
891 | * | |
8b10c5e2 PZ |
892 | * Lock ordering is by mutex address. There are two other sites where |
893 | * perf_event_context::mutex nests and those are: | |
894 | * | |
895 | * - perf_event_exit_task_context() [ child , 0 ] | |
896 | * __perf_event_exit_task() | |
897 | * sync_child_event() | |
898 | * put_event() [ parent, 1 ] | |
899 | * | |
900 | * - perf_event_init_context() [ parent, 0 ] | |
901 | * inherit_task_group() | |
902 | * inherit_group() | |
903 | * inherit_event() | |
904 | * perf_event_alloc() | |
905 | * perf_init_event() | |
906 | * perf_try_init_event() [ child , 1 ] | |
907 | * | |
908 | * While it appears there is an obvious deadlock here -- the parent and child | |
909 | * nesting levels are inverted between the two. This is in fact safe because | |
910 | * life-time rules separate them. That is an exiting task cannot fork, and a | |
911 | * spawning task cannot (yet) exit. | |
912 | * | |
913 | * But remember that that these are parent<->child context relations, and | |
914 | * migration does not affect children, therefore these two orderings should not | |
915 | * interact. | |
f63a8daa PZ |
916 | * |
917 | * The change in perf_event::ctx does not affect children (as claimed above) | |
918 | * because the sys_perf_event_open() case will install a new event and break | |
919 | * the ctx parent<->child relation, and perf_pmu_migrate_context() is only | |
920 | * concerned with cpuctx and that doesn't have children. | |
921 | * | |
922 | * The places that change perf_event::ctx will issue: | |
923 | * | |
924 | * perf_remove_from_context(); | |
925 | * synchronize_rcu(); | |
926 | * perf_install_in_context(); | |
927 | * | |
928 | * to affect the change. The remove_from_context() + synchronize_rcu() should | |
929 | * quiesce the event, after which we can install it in the new location. This | |
930 | * means that only external vectors (perf_fops, prctl) can perturb the event | |
931 | * while in transit. Therefore all such accessors should also acquire | |
932 | * perf_event_context::mutex to serialize against this. | |
933 | * | |
934 | * However; because event->ctx can change while we're waiting to acquire | |
935 | * ctx->mutex we must be careful and use the below perf_event_ctx_lock() | |
936 | * function. | |
937 | * | |
938 | * Lock order: | |
939 | * task_struct::perf_event_mutex | |
940 | * perf_event_context::mutex | |
941 | * perf_event_context::lock | |
942 | * perf_event::child_mutex; | |
943 | * perf_event::mmap_mutex | |
944 | * mmap_sem | |
945 | */ | |
a83fe28e PZ |
946 | static struct perf_event_context * |
947 | perf_event_ctx_lock_nested(struct perf_event *event, int nesting) | |
f63a8daa PZ |
948 | { |
949 | struct perf_event_context *ctx; | |
950 | ||
951 | again: | |
952 | rcu_read_lock(); | |
953 | ctx = ACCESS_ONCE(event->ctx); | |
954 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
955 | rcu_read_unlock(); | |
956 | goto again; | |
957 | } | |
958 | rcu_read_unlock(); | |
959 | ||
a83fe28e | 960 | mutex_lock_nested(&ctx->mutex, nesting); |
f63a8daa PZ |
961 | if (event->ctx != ctx) { |
962 | mutex_unlock(&ctx->mutex); | |
963 | put_ctx(ctx); | |
964 | goto again; | |
965 | } | |
966 | ||
967 | return ctx; | |
968 | } | |
969 | ||
a83fe28e PZ |
970 | static inline struct perf_event_context * |
971 | perf_event_ctx_lock(struct perf_event *event) | |
972 | { | |
973 | return perf_event_ctx_lock_nested(event, 0); | |
974 | } | |
975 | ||
f63a8daa PZ |
976 | static void perf_event_ctx_unlock(struct perf_event *event, |
977 | struct perf_event_context *ctx) | |
978 | { | |
979 | mutex_unlock(&ctx->mutex); | |
980 | put_ctx(ctx); | |
981 | } | |
982 | ||
211de6eb PZ |
983 | /* |
984 | * This must be done under the ctx->lock, such as to serialize against | |
985 | * context_equiv(), therefore we cannot call put_ctx() since that might end up | |
986 | * calling scheduler related locks and ctx->lock nests inside those. | |
987 | */ | |
988 | static __must_check struct perf_event_context * | |
989 | unclone_ctx(struct perf_event_context *ctx) | |
71a851b4 | 990 | { |
211de6eb PZ |
991 | struct perf_event_context *parent_ctx = ctx->parent_ctx; |
992 | ||
993 | lockdep_assert_held(&ctx->lock); | |
994 | ||
995 | if (parent_ctx) | |
71a851b4 | 996 | ctx->parent_ctx = NULL; |
5a3126d4 | 997 | ctx->generation++; |
211de6eb PZ |
998 | |
999 | return parent_ctx; | |
71a851b4 PZ |
1000 | } |
1001 | ||
6844c09d ACM |
1002 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
1003 | { | |
1004 | /* | |
1005 | * only top level events have the pid namespace they were created in | |
1006 | */ | |
1007 | if (event->parent) | |
1008 | event = event->parent; | |
1009 | ||
1010 | return task_tgid_nr_ns(p, event->ns); | |
1011 | } | |
1012 | ||
1013 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
1014 | { | |
1015 | /* | |
1016 | * only top level events have the pid namespace they were created in | |
1017 | */ | |
1018 | if (event->parent) | |
1019 | event = event->parent; | |
1020 | ||
1021 | return task_pid_nr_ns(p, event->ns); | |
1022 | } | |
1023 | ||
7f453c24 | 1024 | /* |
cdd6c482 | 1025 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
1026 | * to userspace. |
1027 | */ | |
cdd6c482 | 1028 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 1029 | { |
cdd6c482 | 1030 | u64 id = event->id; |
7f453c24 | 1031 | |
cdd6c482 IM |
1032 | if (event->parent) |
1033 | id = event->parent->id; | |
7f453c24 PZ |
1034 | |
1035 | return id; | |
1036 | } | |
1037 | ||
25346b93 | 1038 | /* |
cdd6c482 | 1039 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
1040 | * This has to cope with with the fact that until it is locked, |
1041 | * the context could get moved to another task. | |
1042 | */ | |
cdd6c482 | 1043 | static struct perf_event_context * |
8dc85d54 | 1044 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 1045 | { |
cdd6c482 | 1046 | struct perf_event_context *ctx; |
25346b93 | 1047 | |
9ed6060d | 1048 | retry: |
058ebd0e PZ |
1049 | /* |
1050 | * One of the few rules of preemptible RCU is that one cannot do | |
1051 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
1052 | * part of the read side critical section was preemptible -- see | |
1053 | * rcu_read_unlock_special(). | |
1054 | * | |
1055 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
1056 | * side critical section is non-preemptible. | |
1057 | */ | |
1058 | preempt_disable(); | |
1059 | rcu_read_lock(); | |
8dc85d54 | 1060 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
1061 | if (ctx) { |
1062 | /* | |
1063 | * If this context is a clone of another, it might | |
1064 | * get swapped for another underneath us by | |
cdd6c482 | 1065 | * perf_event_task_sched_out, though the |
25346b93 PM |
1066 | * rcu_read_lock() protects us from any context |
1067 | * getting freed. Lock the context and check if it | |
1068 | * got swapped before we could get the lock, and retry | |
1069 | * if so. If we locked the right context, then it | |
1070 | * can't get swapped on us any more. | |
1071 | */ | |
e625cce1 | 1072 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 1073 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 1074 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
1075 | rcu_read_unlock(); |
1076 | preempt_enable(); | |
25346b93 PM |
1077 | goto retry; |
1078 | } | |
b49a9e7e PZ |
1079 | |
1080 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 1081 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1082 | ctx = NULL; |
1083 | } | |
25346b93 PM |
1084 | } |
1085 | rcu_read_unlock(); | |
058ebd0e | 1086 | preempt_enable(); |
25346b93 PM |
1087 | return ctx; |
1088 | } | |
1089 | ||
1090 | /* | |
1091 | * Get the context for a task and increment its pin_count so it | |
1092 | * can't get swapped to another task. This also increments its | |
1093 | * reference count so that the context can't get freed. | |
1094 | */ | |
8dc85d54 PZ |
1095 | static struct perf_event_context * |
1096 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1097 | { |
cdd6c482 | 1098 | struct perf_event_context *ctx; |
25346b93 PM |
1099 | unsigned long flags; |
1100 | ||
8dc85d54 | 1101 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1102 | if (ctx) { |
1103 | ++ctx->pin_count; | |
e625cce1 | 1104 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1105 | } |
1106 | return ctx; | |
1107 | } | |
1108 | ||
cdd6c482 | 1109 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1110 | { |
1111 | unsigned long flags; | |
1112 | ||
e625cce1 | 1113 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1114 | --ctx->pin_count; |
e625cce1 | 1115 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1116 | } |
1117 | ||
f67218c3 PZ |
1118 | /* |
1119 | * Update the record of the current time in a context. | |
1120 | */ | |
1121 | static void update_context_time(struct perf_event_context *ctx) | |
1122 | { | |
1123 | u64 now = perf_clock(); | |
1124 | ||
1125 | ctx->time += now - ctx->timestamp; | |
1126 | ctx->timestamp = now; | |
1127 | } | |
1128 | ||
4158755d SE |
1129 | static u64 perf_event_time(struct perf_event *event) |
1130 | { | |
1131 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1132 | |
1133 | if (is_cgroup_event(event)) | |
1134 | return perf_cgroup_event_time(event); | |
1135 | ||
4158755d SE |
1136 | return ctx ? ctx->time : 0; |
1137 | } | |
1138 | ||
f67218c3 PZ |
1139 | /* |
1140 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1141 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1142 | */ |
1143 | static void update_event_times(struct perf_event *event) | |
1144 | { | |
1145 | struct perf_event_context *ctx = event->ctx; | |
1146 | u64 run_end; | |
1147 | ||
1148 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1149 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1150 | return; | |
e5d1367f SE |
1151 | /* |
1152 | * in cgroup mode, time_enabled represents | |
1153 | * the time the event was enabled AND active | |
1154 | * tasks were in the monitored cgroup. This is | |
1155 | * independent of the activity of the context as | |
1156 | * there may be a mix of cgroup and non-cgroup events. | |
1157 | * | |
1158 | * That is why we treat cgroup events differently | |
1159 | * here. | |
1160 | */ | |
1161 | if (is_cgroup_event(event)) | |
46cd6a7f | 1162 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1163 | else if (ctx->is_active) |
1164 | run_end = ctx->time; | |
acd1d7c1 PZ |
1165 | else |
1166 | run_end = event->tstamp_stopped; | |
1167 | ||
1168 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1169 | |
1170 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1171 | run_end = event->tstamp_stopped; | |
1172 | else | |
4158755d | 1173 | run_end = perf_event_time(event); |
f67218c3 PZ |
1174 | |
1175 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1176 | |
f67218c3 PZ |
1177 | } |
1178 | ||
96c21a46 PZ |
1179 | /* |
1180 | * Update total_time_enabled and total_time_running for all events in a group. | |
1181 | */ | |
1182 | static void update_group_times(struct perf_event *leader) | |
1183 | { | |
1184 | struct perf_event *event; | |
1185 | ||
1186 | update_event_times(leader); | |
1187 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1188 | update_event_times(event); | |
1189 | } | |
1190 | ||
889ff015 FW |
1191 | static struct list_head * |
1192 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1193 | { | |
1194 | if (event->attr.pinned) | |
1195 | return &ctx->pinned_groups; | |
1196 | else | |
1197 | return &ctx->flexible_groups; | |
1198 | } | |
1199 | ||
fccc714b | 1200 | /* |
cdd6c482 | 1201 | * Add a event from the lists for its context. |
fccc714b PZ |
1202 | * Must be called with ctx->mutex and ctx->lock held. |
1203 | */ | |
04289bb9 | 1204 | static void |
cdd6c482 | 1205 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1206 | { |
8a49542c PZ |
1207 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1208 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1209 | |
1210 | /* | |
8a49542c PZ |
1211 | * If we're a stand alone event or group leader, we go to the context |
1212 | * list, group events are kept attached to the group so that | |
1213 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1214 | */ |
8a49542c | 1215 | if (event->group_leader == event) { |
889ff015 FW |
1216 | struct list_head *list; |
1217 | ||
d6f962b5 FW |
1218 | if (is_software_event(event)) |
1219 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1220 | ||
889ff015 FW |
1221 | list = ctx_group_list(event, ctx); |
1222 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1223 | } |
592903cd | 1224 | |
08309379 | 1225 | if (is_cgroup_event(event)) |
e5d1367f | 1226 | ctx->nr_cgroups++; |
e5d1367f | 1227 | |
cdd6c482 IM |
1228 | list_add_rcu(&event->event_entry, &ctx->event_list); |
1229 | ctx->nr_events++; | |
1230 | if (event->attr.inherit_stat) | |
bfbd3381 | 1231 | ctx->nr_stat++; |
5a3126d4 PZ |
1232 | |
1233 | ctx->generation++; | |
04289bb9 IM |
1234 | } |
1235 | ||
0231bb53 JO |
1236 | /* |
1237 | * Initialize event state based on the perf_event_attr::disabled. | |
1238 | */ | |
1239 | static inline void perf_event__state_init(struct perf_event *event) | |
1240 | { | |
1241 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1242 | PERF_EVENT_STATE_INACTIVE; | |
1243 | } | |
1244 | ||
c320c7b7 ACM |
1245 | /* |
1246 | * Called at perf_event creation and when events are attached/detached from a | |
1247 | * group. | |
1248 | */ | |
1249 | static void perf_event__read_size(struct perf_event *event) | |
1250 | { | |
1251 | int entry = sizeof(u64); /* value */ | |
1252 | int size = 0; | |
1253 | int nr = 1; | |
1254 | ||
1255 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1256 | size += sizeof(u64); | |
1257 | ||
1258 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1259 | size += sizeof(u64); | |
1260 | ||
1261 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1262 | entry += sizeof(u64); | |
1263 | ||
1264 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1265 | nr += event->group_leader->nr_siblings; | |
1266 | size += sizeof(u64); | |
1267 | } | |
1268 | ||
1269 | size += entry * nr; | |
1270 | event->read_size = size; | |
1271 | } | |
1272 | ||
1273 | static void perf_event__header_size(struct perf_event *event) | |
1274 | { | |
1275 | struct perf_sample_data *data; | |
1276 | u64 sample_type = event->attr.sample_type; | |
1277 | u16 size = 0; | |
1278 | ||
1279 | perf_event__read_size(event); | |
1280 | ||
1281 | if (sample_type & PERF_SAMPLE_IP) | |
1282 | size += sizeof(data->ip); | |
1283 | ||
6844c09d ACM |
1284 | if (sample_type & PERF_SAMPLE_ADDR) |
1285 | size += sizeof(data->addr); | |
1286 | ||
1287 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1288 | size += sizeof(data->period); | |
1289 | ||
c3feedf2 AK |
1290 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1291 | size += sizeof(data->weight); | |
1292 | ||
6844c09d ACM |
1293 | if (sample_type & PERF_SAMPLE_READ) |
1294 | size += event->read_size; | |
1295 | ||
d6be9ad6 SE |
1296 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1297 | size += sizeof(data->data_src.val); | |
1298 | ||
fdfbbd07 AK |
1299 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1300 | size += sizeof(data->txn); | |
1301 | ||
6844c09d ACM |
1302 | event->header_size = size; |
1303 | } | |
1304 | ||
1305 | static void perf_event__id_header_size(struct perf_event *event) | |
1306 | { | |
1307 | struct perf_sample_data *data; | |
1308 | u64 sample_type = event->attr.sample_type; | |
1309 | u16 size = 0; | |
1310 | ||
c320c7b7 ACM |
1311 | if (sample_type & PERF_SAMPLE_TID) |
1312 | size += sizeof(data->tid_entry); | |
1313 | ||
1314 | if (sample_type & PERF_SAMPLE_TIME) | |
1315 | size += sizeof(data->time); | |
1316 | ||
ff3d527c AH |
1317 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1318 | size += sizeof(data->id); | |
1319 | ||
c320c7b7 ACM |
1320 | if (sample_type & PERF_SAMPLE_ID) |
1321 | size += sizeof(data->id); | |
1322 | ||
1323 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1324 | size += sizeof(data->stream_id); | |
1325 | ||
1326 | if (sample_type & PERF_SAMPLE_CPU) | |
1327 | size += sizeof(data->cpu_entry); | |
1328 | ||
6844c09d | 1329 | event->id_header_size = size; |
c320c7b7 ACM |
1330 | } |
1331 | ||
8a49542c PZ |
1332 | static void perf_group_attach(struct perf_event *event) |
1333 | { | |
c320c7b7 | 1334 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1335 | |
74c3337c PZ |
1336 | /* |
1337 | * We can have double attach due to group movement in perf_event_open. | |
1338 | */ | |
1339 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1340 | return; | |
1341 | ||
8a49542c PZ |
1342 | event->attach_state |= PERF_ATTACH_GROUP; |
1343 | ||
1344 | if (group_leader == event) | |
1345 | return; | |
1346 | ||
652884fe PZ |
1347 | WARN_ON_ONCE(group_leader->ctx != event->ctx); |
1348 | ||
8a49542c PZ |
1349 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
1350 | !is_software_event(event)) | |
1351 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1352 | ||
1353 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1354 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1355 | |
1356 | perf_event__header_size(group_leader); | |
1357 | ||
1358 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1359 | perf_event__header_size(pos); | |
8a49542c PZ |
1360 | } |
1361 | ||
a63eaf34 | 1362 | /* |
cdd6c482 | 1363 | * Remove a event from the lists for its context. |
fccc714b | 1364 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1365 | */ |
04289bb9 | 1366 | static void |
cdd6c482 | 1367 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1368 | { |
68cacd29 | 1369 | struct perf_cpu_context *cpuctx; |
652884fe PZ |
1370 | |
1371 | WARN_ON_ONCE(event->ctx != ctx); | |
1372 | lockdep_assert_held(&ctx->lock); | |
1373 | ||
8a49542c PZ |
1374 | /* |
1375 | * We can have double detach due to exit/hot-unplug + close. | |
1376 | */ | |
1377 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1378 | return; |
8a49542c PZ |
1379 | |
1380 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1381 | ||
68cacd29 | 1382 | if (is_cgroup_event(event)) { |
e5d1367f | 1383 | ctx->nr_cgroups--; |
68cacd29 SE |
1384 | cpuctx = __get_cpu_context(ctx); |
1385 | /* | |
1386 | * if there are no more cgroup events | |
1387 | * then cler cgrp to avoid stale pointer | |
1388 | * in update_cgrp_time_from_cpuctx() | |
1389 | */ | |
1390 | if (!ctx->nr_cgroups) | |
1391 | cpuctx->cgrp = NULL; | |
1392 | } | |
e5d1367f | 1393 | |
cdd6c482 IM |
1394 | ctx->nr_events--; |
1395 | if (event->attr.inherit_stat) | |
bfbd3381 | 1396 | ctx->nr_stat--; |
8bc20959 | 1397 | |
cdd6c482 | 1398 | list_del_rcu(&event->event_entry); |
04289bb9 | 1399 | |
8a49542c PZ |
1400 | if (event->group_leader == event) |
1401 | list_del_init(&event->group_entry); | |
5c148194 | 1402 | |
96c21a46 | 1403 | update_group_times(event); |
b2e74a26 SE |
1404 | |
1405 | /* | |
1406 | * If event was in error state, then keep it | |
1407 | * that way, otherwise bogus counts will be | |
1408 | * returned on read(). The only way to get out | |
1409 | * of error state is by explicit re-enabling | |
1410 | * of the event | |
1411 | */ | |
1412 | if (event->state > PERF_EVENT_STATE_OFF) | |
1413 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1414 | |
1415 | ctx->generation++; | |
050735b0 PZ |
1416 | } |
1417 | ||
8a49542c | 1418 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1419 | { |
1420 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1421 | struct list_head *list = NULL; |
1422 | ||
1423 | /* | |
1424 | * We can have double detach due to exit/hot-unplug + close. | |
1425 | */ | |
1426 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1427 | return; | |
1428 | ||
1429 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1430 | ||
1431 | /* | |
1432 | * If this is a sibling, remove it from its group. | |
1433 | */ | |
1434 | if (event->group_leader != event) { | |
1435 | list_del_init(&event->group_entry); | |
1436 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1437 | goto out; |
8a49542c PZ |
1438 | } |
1439 | ||
1440 | if (!list_empty(&event->group_entry)) | |
1441 | list = &event->group_entry; | |
2e2af50b | 1442 | |
04289bb9 | 1443 | /* |
cdd6c482 IM |
1444 | * If this was a group event with sibling events then |
1445 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1446 | * to whatever list we are on. |
04289bb9 | 1447 | */ |
cdd6c482 | 1448 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1449 | if (list) |
1450 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1451 | sibling->group_leader = sibling; |
d6f962b5 FW |
1452 | |
1453 | /* Inherit group flags from the previous leader */ | |
1454 | sibling->group_flags = event->group_flags; | |
652884fe PZ |
1455 | |
1456 | WARN_ON_ONCE(sibling->ctx != event->ctx); | |
04289bb9 | 1457 | } |
c320c7b7 ACM |
1458 | |
1459 | out: | |
1460 | perf_event__header_size(event->group_leader); | |
1461 | ||
1462 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1463 | perf_event__header_size(tmp); | |
04289bb9 IM |
1464 | } |
1465 | ||
fadfe7be JO |
1466 | /* |
1467 | * User event without the task. | |
1468 | */ | |
1469 | static bool is_orphaned_event(struct perf_event *event) | |
1470 | { | |
1471 | return event && !is_kernel_event(event) && !event->owner; | |
1472 | } | |
1473 | ||
1474 | /* | |
1475 | * Event has a parent but parent's task finished and it's | |
1476 | * alive only because of children holding refference. | |
1477 | */ | |
1478 | static bool is_orphaned_child(struct perf_event *event) | |
1479 | { | |
1480 | return is_orphaned_event(event->parent); | |
1481 | } | |
1482 | ||
1483 | static void orphans_remove_work(struct work_struct *work); | |
1484 | ||
1485 | static void schedule_orphans_remove(struct perf_event_context *ctx) | |
1486 | { | |
1487 | if (!ctx->task || ctx->orphans_remove_sched || !perf_wq) | |
1488 | return; | |
1489 | ||
1490 | if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) { | |
1491 | get_ctx(ctx); | |
1492 | ctx->orphans_remove_sched = true; | |
1493 | } | |
1494 | } | |
1495 | ||
1496 | static int __init perf_workqueue_init(void) | |
1497 | { | |
1498 | perf_wq = create_singlethread_workqueue("perf"); | |
1499 | WARN(!perf_wq, "failed to create perf workqueue\n"); | |
1500 | return perf_wq ? 0 : -1; | |
1501 | } | |
1502 | ||
1503 | core_initcall(perf_workqueue_init); | |
1504 | ||
66eb579e MR |
1505 | static inline int pmu_filter_match(struct perf_event *event) |
1506 | { | |
1507 | struct pmu *pmu = event->pmu; | |
1508 | return pmu->filter_match ? pmu->filter_match(event) : 1; | |
1509 | } | |
1510 | ||
fa66f07a SE |
1511 | static inline int |
1512 | event_filter_match(struct perf_event *event) | |
1513 | { | |
e5d1367f | 1514 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
66eb579e | 1515 | && perf_cgroup_match(event) && pmu_filter_match(event); |
fa66f07a SE |
1516 | } |
1517 | ||
9ffcfa6f SE |
1518 | static void |
1519 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1520 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1521 | struct perf_event_context *ctx) |
3b6f9e5c | 1522 | { |
4158755d | 1523 | u64 tstamp = perf_event_time(event); |
fa66f07a | 1524 | u64 delta; |
652884fe PZ |
1525 | |
1526 | WARN_ON_ONCE(event->ctx != ctx); | |
1527 | lockdep_assert_held(&ctx->lock); | |
1528 | ||
fa66f07a SE |
1529 | /* |
1530 | * An event which could not be activated because of | |
1531 | * filter mismatch still needs to have its timings | |
1532 | * maintained, otherwise bogus information is return | |
1533 | * via read() for time_enabled, time_running: | |
1534 | */ | |
1535 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1536 | && !event_filter_match(event)) { | |
e5d1367f | 1537 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1538 | event->tstamp_running += delta; |
4158755d | 1539 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1540 | } |
1541 | ||
cdd6c482 | 1542 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1543 | return; |
3b6f9e5c | 1544 | |
44377277 AS |
1545 | perf_pmu_disable(event->pmu); |
1546 | ||
cdd6c482 IM |
1547 | event->state = PERF_EVENT_STATE_INACTIVE; |
1548 | if (event->pending_disable) { | |
1549 | event->pending_disable = 0; | |
1550 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1551 | } |
4158755d | 1552 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1553 | event->pmu->del(event, 0); |
cdd6c482 | 1554 | event->oncpu = -1; |
3b6f9e5c | 1555 | |
cdd6c482 | 1556 | if (!is_software_event(event)) |
3b6f9e5c | 1557 | cpuctx->active_oncpu--; |
2fde4f94 MR |
1558 | if (!--ctx->nr_active) |
1559 | perf_event_ctx_deactivate(ctx); | |
0f5a2601 PZ |
1560 | if (event->attr.freq && event->attr.sample_freq) |
1561 | ctx->nr_freq--; | |
cdd6c482 | 1562 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1563 | cpuctx->exclusive = 0; |
44377277 | 1564 | |
fadfe7be JO |
1565 | if (is_orphaned_child(event)) |
1566 | schedule_orphans_remove(ctx); | |
1567 | ||
44377277 | 1568 | perf_pmu_enable(event->pmu); |
3b6f9e5c PM |
1569 | } |
1570 | ||
d859e29f | 1571 | static void |
cdd6c482 | 1572 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1573 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1574 | struct perf_event_context *ctx) |
d859e29f | 1575 | { |
cdd6c482 | 1576 | struct perf_event *event; |
fa66f07a | 1577 | int state = group_event->state; |
d859e29f | 1578 | |
cdd6c482 | 1579 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1580 | |
1581 | /* | |
1582 | * Schedule out siblings (if any): | |
1583 | */ | |
cdd6c482 IM |
1584 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1585 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1586 | |
fa66f07a | 1587 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1588 | cpuctx->exclusive = 0; |
1589 | } | |
1590 | ||
46ce0fe9 PZ |
1591 | struct remove_event { |
1592 | struct perf_event *event; | |
1593 | bool detach_group; | |
1594 | }; | |
1595 | ||
0793a61d | 1596 | /* |
cdd6c482 | 1597 | * Cross CPU call to remove a performance event |
0793a61d | 1598 | * |
cdd6c482 | 1599 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1600 | * remove it from the context list. |
1601 | */ | |
fe4b04fa | 1602 | static int __perf_remove_from_context(void *info) |
0793a61d | 1603 | { |
46ce0fe9 PZ |
1604 | struct remove_event *re = info; |
1605 | struct perf_event *event = re->event; | |
cdd6c482 | 1606 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1607 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1608 | |
e625cce1 | 1609 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1610 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1611 | if (re->detach_group) |
1612 | perf_group_detach(event); | |
cdd6c482 | 1613 | list_del_event(event, ctx); |
64ce3126 PZ |
1614 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1615 | ctx->is_active = 0; | |
1616 | cpuctx->task_ctx = NULL; | |
1617 | } | |
e625cce1 | 1618 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1619 | |
1620 | return 0; | |
0793a61d TG |
1621 | } |
1622 | ||
1623 | ||
1624 | /* | |
cdd6c482 | 1625 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1626 | * |
cdd6c482 | 1627 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1628 | * call when the task is on a CPU. |
c93f7669 | 1629 | * |
cdd6c482 IM |
1630 | * If event->ctx is a cloned context, callers must make sure that |
1631 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1632 | * remains valid. This is OK when called from perf_release since |
1633 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1634 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1635 | * context has been detached from its task. |
0793a61d | 1636 | */ |
46ce0fe9 | 1637 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1638 | { |
cdd6c482 | 1639 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1640 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1641 | struct remove_event re = { |
1642 | .event = event, | |
1643 | .detach_group = detach_group, | |
1644 | }; | |
0793a61d | 1645 | |
fe4b04fa PZ |
1646 | lockdep_assert_held(&ctx->mutex); |
1647 | ||
0793a61d TG |
1648 | if (!task) { |
1649 | /* | |
226424ee MR |
1650 | * Per cpu events are removed via an smp call. The removal can |
1651 | * fail if the CPU is currently offline, but in that case we | |
1652 | * already called __perf_remove_from_context from | |
1653 | * perf_event_exit_cpu. | |
0793a61d | 1654 | */ |
46ce0fe9 | 1655 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1656 | return; |
1657 | } | |
1658 | ||
1659 | retry: | |
46ce0fe9 | 1660 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1661 | return; |
0793a61d | 1662 | |
e625cce1 | 1663 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1664 | /* |
fe4b04fa PZ |
1665 | * If we failed to find a running task, but find the context active now |
1666 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1667 | */ |
fe4b04fa | 1668 | if (ctx->is_active) { |
e625cce1 | 1669 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
1670 | /* |
1671 | * Reload the task pointer, it might have been changed by | |
1672 | * a concurrent perf_event_context_sched_out(). | |
1673 | */ | |
1674 | task = ctx->task; | |
0793a61d TG |
1675 | goto retry; |
1676 | } | |
1677 | ||
1678 | /* | |
fe4b04fa PZ |
1679 | * Since the task isn't running, its safe to remove the event, us |
1680 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1681 | */ |
46ce0fe9 PZ |
1682 | if (detach_group) |
1683 | perf_group_detach(event); | |
fe4b04fa | 1684 | list_del_event(event, ctx); |
e625cce1 | 1685 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1686 | } |
1687 | ||
d859e29f | 1688 | /* |
cdd6c482 | 1689 | * Cross CPU call to disable a performance event |
d859e29f | 1690 | */ |
500ad2d8 | 1691 | int __perf_event_disable(void *info) |
d859e29f | 1692 | { |
cdd6c482 | 1693 | struct perf_event *event = info; |
cdd6c482 | 1694 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1695 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1696 | |
1697 | /* | |
cdd6c482 IM |
1698 | * If this is a per-task event, need to check whether this |
1699 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1700 | * |
1701 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1702 | * flipping contexts around. | |
d859e29f | 1703 | */ |
665c2142 | 1704 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1705 | return -EINVAL; |
d859e29f | 1706 | |
e625cce1 | 1707 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1708 | |
1709 | /* | |
cdd6c482 | 1710 | * If the event is on, turn it off. |
d859e29f PM |
1711 | * If it is in error state, leave it in error state. |
1712 | */ | |
cdd6c482 | 1713 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1714 | update_context_time(ctx); |
e5d1367f | 1715 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1716 | update_group_times(event); |
1717 | if (event == event->group_leader) | |
1718 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1719 | else |
cdd6c482 IM |
1720 | event_sched_out(event, cpuctx, ctx); |
1721 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1722 | } |
1723 | ||
e625cce1 | 1724 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1725 | |
1726 | return 0; | |
d859e29f PM |
1727 | } |
1728 | ||
1729 | /* | |
cdd6c482 | 1730 | * Disable a event. |
c93f7669 | 1731 | * |
cdd6c482 IM |
1732 | * If event->ctx is a cloned context, callers must make sure that |
1733 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1734 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1735 | * perf_event_for_each_child or perf_event_for_each because they |
1736 | * hold the top-level event's child_mutex, so any descendant that | |
1737 | * goes to exit will block in sync_child_event. | |
1738 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1739 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1740 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1741 | */ |
f63a8daa | 1742 | static void _perf_event_disable(struct perf_event *event) |
d859e29f | 1743 | { |
cdd6c482 | 1744 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1745 | struct task_struct *task = ctx->task; |
1746 | ||
1747 | if (!task) { | |
1748 | /* | |
cdd6c482 | 1749 | * Disable the event on the cpu that it's on |
d859e29f | 1750 | */ |
fe4b04fa | 1751 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1752 | return; |
1753 | } | |
1754 | ||
9ed6060d | 1755 | retry: |
fe4b04fa PZ |
1756 | if (!task_function_call(task, __perf_event_disable, event)) |
1757 | return; | |
d859e29f | 1758 | |
e625cce1 | 1759 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1760 | /* |
cdd6c482 | 1761 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1762 | */ |
cdd6c482 | 1763 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1764 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1765 | /* |
1766 | * Reload the task pointer, it might have been changed by | |
1767 | * a concurrent perf_event_context_sched_out(). | |
1768 | */ | |
1769 | task = ctx->task; | |
d859e29f PM |
1770 | goto retry; |
1771 | } | |
1772 | ||
1773 | /* | |
1774 | * Since we have the lock this context can't be scheduled | |
1775 | * in, so we can change the state safely. | |
1776 | */ | |
cdd6c482 IM |
1777 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1778 | update_group_times(event); | |
1779 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1780 | } |
e625cce1 | 1781 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1782 | } |
f63a8daa PZ |
1783 | |
1784 | /* | |
1785 | * Strictly speaking kernel users cannot create groups and therefore this | |
1786 | * interface does not need the perf_event_ctx_lock() magic. | |
1787 | */ | |
1788 | void perf_event_disable(struct perf_event *event) | |
1789 | { | |
1790 | struct perf_event_context *ctx; | |
1791 | ||
1792 | ctx = perf_event_ctx_lock(event); | |
1793 | _perf_event_disable(event); | |
1794 | perf_event_ctx_unlock(event, ctx); | |
1795 | } | |
dcfce4a0 | 1796 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1797 | |
e5d1367f SE |
1798 | static void perf_set_shadow_time(struct perf_event *event, |
1799 | struct perf_event_context *ctx, | |
1800 | u64 tstamp) | |
1801 | { | |
1802 | /* | |
1803 | * use the correct time source for the time snapshot | |
1804 | * | |
1805 | * We could get by without this by leveraging the | |
1806 | * fact that to get to this function, the caller | |
1807 | * has most likely already called update_context_time() | |
1808 | * and update_cgrp_time_xx() and thus both timestamp | |
1809 | * are identical (or very close). Given that tstamp is, | |
1810 | * already adjusted for cgroup, we could say that: | |
1811 | * tstamp - ctx->timestamp | |
1812 | * is equivalent to | |
1813 | * tstamp - cgrp->timestamp. | |
1814 | * | |
1815 | * Then, in perf_output_read(), the calculation would | |
1816 | * work with no changes because: | |
1817 | * - event is guaranteed scheduled in | |
1818 | * - no scheduled out in between | |
1819 | * - thus the timestamp would be the same | |
1820 | * | |
1821 | * But this is a bit hairy. | |
1822 | * | |
1823 | * So instead, we have an explicit cgroup call to remain | |
1824 | * within the time time source all along. We believe it | |
1825 | * is cleaner and simpler to understand. | |
1826 | */ | |
1827 | if (is_cgroup_event(event)) | |
1828 | perf_cgroup_set_shadow_time(event, tstamp); | |
1829 | else | |
1830 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1831 | } | |
1832 | ||
4fe757dd PZ |
1833 | #define MAX_INTERRUPTS (~0ULL) |
1834 | ||
1835 | static void perf_log_throttle(struct perf_event *event, int enable); | |
ec0d7729 | 1836 | static void perf_log_itrace_start(struct perf_event *event); |
4fe757dd | 1837 | |
235c7fc7 | 1838 | static int |
9ffcfa6f | 1839 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1840 | struct perf_cpu_context *cpuctx, |
6e37738a | 1841 | struct perf_event_context *ctx) |
235c7fc7 | 1842 | { |
4158755d | 1843 | u64 tstamp = perf_event_time(event); |
44377277 | 1844 | int ret = 0; |
4158755d | 1845 | |
63342411 PZ |
1846 | lockdep_assert_held(&ctx->lock); |
1847 | ||
cdd6c482 | 1848 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1849 | return 0; |
1850 | ||
cdd6c482 | 1851 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1852 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1853 | |
1854 | /* | |
1855 | * Unthrottle events, since we scheduled we might have missed several | |
1856 | * ticks already, also for a heavily scheduling task there is little | |
1857 | * guarantee it'll get a tick in a timely manner. | |
1858 | */ | |
1859 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1860 | perf_log_throttle(event, 1); | |
1861 | event->hw.interrupts = 0; | |
1862 | } | |
1863 | ||
235c7fc7 IM |
1864 | /* |
1865 | * The new state must be visible before we turn it on in the hardware: | |
1866 | */ | |
1867 | smp_wmb(); | |
1868 | ||
44377277 AS |
1869 | perf_pmu_disable(event->pmu); |
1870 | ||
72f669c0 SL |
1871 | perf_set_shadow_time(event, ctx, tstamp); |
1872 | ||
ec0d7729 AS |
1873 | perf_log_itrace_start(event); |
1874 | ||
a4eaf7f1 | 1875 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1876 | event->state = PERF_EVENT_STATE_INACTIVE; |
1877 | event->oncpu = -1; | |
44377277 AS |
1878 | ret = -EAGAIN; |
1879 | goto out; | |
235c7fc7 IM |
1880 | } |
1881 | ||
00a2916f PZ |
1882 | event->tstamp_running += tstamp - event->tstamp_stopped; |
1883 | ||
cdd6c482 | 1884 | if (!is_software_event(event)) |
3b6f9e5c | 1885 | cpuctx->active_oncpu++; |
2fde4f94 MR |
1886 | if (!ctx->nr_active++) |
1887 | perf_event_ctx_activate(ctx); | |
0f5a2601 PZ |
1888 | if (event->attr.freq && event->attr.sample_freq) |
1889 | ctx->nr_freq++; | |
235c7fc7 | 1890 | |
cdd6c482 | 1891 | if (event->attr.exclusive) |
3b6f9e5c PM |
1892 | cpuctx->exclusive = 1; |
1893 | ||
fadfe7be JO |
1894 | if (is_orphaned_child(event)) |
1895 | schedule_orphans_remove(ctx); | |
1896 | ||
44377277 AS |
1897 | out: |
1898 | perf_pmu_enable(event->pmu); | |
1899 | ||
1900 | return ret; | |
235c7fc7 IM |
1901 | } |
1902 | ||
6751b71e | 1903 | static int |
cdd6c482 | 1904 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1905 | struct perf_cpu_context *cpuctx, |
6e37738a | 1906 | struct perf_event_context *ctx) |
6751b71e | 1907 | { |
6bde9b6c | 1908 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1909 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1910 | u64 now = ctx->time; |
1911 | bool simulate = false; | |
6751b71e | 1912 | |
cdd6c482 | 1913 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1914 | return 0; |
1915 | ||
ad5133b7 | 1916 | pmu->start_txn(pmu); |
6bde9b6c | 1917 | |
9ffcfa6f | 1918 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1919 | pmu->cancel_txn(pmu); |
272325c4 | 1920 | perf_mux_hrtimer_restart(cpuctx); |
6751b71e | 1921 | return -EAGAIN; |
90151c35 | 1922 | } |
6751b71e PM |
1923 | |
1924 | /* | |
1925 | * Schedule in siblings as one group (if any): | |
1926 | */ | |
cdd6c482 | 1927 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1928 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1929 | partial_group = event; |
6751b71e PM |
1930 | goto group_error; |
1931 | } | |
1932 | } | |
1933 | ||
9ffcfa6f | 1934 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1935 | return 0; |
9ffcfa6f | 1936 | |
6751b71e PM |
1937 | group_error: |
1938 | /* | |
1939 | * Groups can be scheduled in as one unit only, so undo any | |
1940 | * partial group before returning: | |
d7842da4 SE |
1941 | * The events up to the failed event are scheduled out normally, |
1942 | * tstamp_stopped will be updated. | |
1943 | * | |
1944 | * The failed events and the remaining siblings need to have | |
1945 | * their timings updated as if they had gone thru event_sched_in() | |
1946 | * and event_sched_out(). This is required to get consistent timings | |
1947 | * across the group. This also takes care of the case where the group | |
1948 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1949 | * the time the event was actually stopped, such that time delta | |
1950 | * calculation in update_event_times() is correct. | |
6751b71e | 1951 | */ |
cdd6c482 IM |
1952 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1953 | if (event == partial_group) | |
d7842da4 SE |
1954 | simulate = true; |
1955 | ||
1956 | if (simulate) { | |
1957 | event->tstamp_running += now - event->tstamp_stopped; | |
1958 | event->tstamp_stopped = now; | |
1959 | } else { | |
1960 | event_sched_out(event, cpuctx, ctx); | |
1961 | } | |
6751b71e | 1962 | } |
9ffcfa6f | 1963 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1964 | |
ad5133b7 | 1965 | pmu->cancel_txn(pmu); |
90151c35 | 1966 | |
272325c4 | 1967 | perf_mux_hrtimer_restart(cpuctx); |
9e630205 | 1968 | |
6751b71e PM |
1969 | return -EAGAIN; |
1970 | } | |
1971 | ||
3b6f9e5c | 1972 | /* |
cdd6c482 | 1973 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1974 | */ |
cdd6c482 | 1975 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1976 | struct perf_cpu_context *cpuctx, |
1977 | int can_add_hw) | |
1978 | { | |
1979 | /* | |
cdd6c482 | 1980 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1981 | */ |
d6f962b5 | 1982 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1983 | return 1; |
1984 | /* | |
1985 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1986 | * events can go on. |
3b6f9e5c PM |
1987 | */ |
1988 | if (cpuctx->exclusive) | |
1989 | return 0; | |
1990 | /* | |
1991 | * If this group is exclusive and there are already | |
cdd6c482 | 1992 | * events on the CPU, it can't go on. |
3b6f9e5c | 1993 | */ |
cdd6c482 | 1994 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1995 | return 0; |
1996 | /* | |
1997 | * Otherwise, try to add it if all previous groups were able | |
1998 | * to go on. | |
1999 | */ | |
2000 | return can_add_hw; | |
2001 | } | |
2002 | ||
cdd6c482 IM |
2003 | static void add_event_to_ctx(struct perf_event *event, |
2004 | struct perf_event_context *ctx) | |
53cfbf59 | 2005 | { |
4158755d SE |
2006 | u64 tstamp = perf_event_time(event); |
2007 | ||
cdd6c482 | 2008 | list_add_event(event, ctx); |
8a49542c | 2009 | perf_group_attach(event); |
4158755d SE |
2010 | event->tstamp_enabled = tstamp; |
2011 | event->tstamp_running = tstamp; | |
2012 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
2013 | } |
2014 | ||
2c29ef0f PZ |
2015 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
2016 | static void | |
2017 | ctx_sched_in(struct perf_event_context *ctx, | |
2018 | struct perf_cpu_context *cpuctx, | |
2019 | enum event_type_t event_type, | |
2020 | struct task_struct *task); | |
fe4b04fa | 2021 | |
dce5855b PZ |
2022 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
2023 | struct perf_event_context *ctx, | |
2024 | struct task_struct *task) | |
2025 | { | |
2026 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
2027 | if (ctx) | |
2028 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
2029 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
2030 | if (ctx) | |
2031 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
2032 | } | |
2033 | ||
0793a61d | 2034 | /* |
cdd6c482 | 2035 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
2036 | * |
2037 | * Must be called with ctx->mutex held | |
0793a61d | 2038 | */ |
fe4b04fa | 2039 | static int __perf_install_in_context(void *info) |
0793a61d | 2040 | { |
cdd6c482 IM |
2041 | struct perf_event *event = info; |
2042 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2043 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
2044 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
2045 | struct task_struct *task = current; | |
2046 | ||
b58f6b0d | 2047 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 2048 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
2049 | |
2050 | /* | |
2c29ef0f | 2051 | * If there was an active task_ctx schedule it out. |
0793a61d | 2052 | */ |
b58f6b0d | 2053 | if (task_ctx) |
2c29ef0f | 2054 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
2055 | |
2056 | /* | |
2057 | * If the context we're installing events in is not the | |
2058 | * active task_ctx, flip them. | |
2059 | */ | |
2060 | if (ctx->task && task_ctx != ctx) { | |
2061 | if (task_ctx) | |
2062 | raw_spin_unlock(&task_ctx->lock); | |
2063 | raw_spin_lock(&ctx->lock); | |
2064 | task_ctx = ctx; | |
2065 | } | |
2066 | ||
2067 | if (task_ctx) { | |
2068 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
2069 | task = task_ctx->task; |
2070 | } | |
b58f6b0d | 2071 | |
2c29ef0f | 2072 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 2073 | |
4af4998b | 2074 | update_context_time(ctx); |
e5d1367f SE |
2075 | /* |
2076 | * update cgrp time only if current cgrp | |
2077 | * matches event->cgrp. Must be done before | |
2078 | * calling add_event_to_ctx() | |
2079 | */ | |
2080 | update_cgrp_time_from_event(event); | |
0793a61d | 2081 | |
cdd6c482 | 2082 | add_event_to_ctx(event, ctx); |
0793a61d | 2083 | |
d859e29f | 2084 | /* |
2c29ef0f | 2085 | * Schedule everything back in |
d859e29f | 2086 | */ |
dce5855b | 2087 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
2088 | |
2089 | perf_pmu_enable(cpuctx->ctx.pmu); | |
2090 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
2091 | |
2092 | return 0; | |
0793a61d TG |
2093 | } |
2094 | ||
2095 | /* | |
cdd6c482 | 2096 | * Attach a performance event to a context |
0793a61d | 2097 | * |
cdd6c482 IM |
2098 | * First we add the event to the list with the hardware enable bit |
2099 | * in event->hw_config cleared. | |
0793a61d | 2100 | * |
cdd6c482 | 2101 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
2102 | * call to enable it in the task context. The task might have been |
2103 | * scheduled away, but we check this in the smp call again. | |
2104 | */ | |
2105 | static void | |
cdd6c482 IM |
2106 | perf_install_in_context(struct perf_event_context *ctx, |
2107 | struct perf_event *event, | |
0793a61d TG |
2108 | int cpu) |
2109 | { | |
2110 | struct task_struct *task = ctx->task; | |
2111 | ||
fe4b04fa PZ |
2112 | lockdep_assert_held(&ctx->mutex); |
2113 | ||
c3f00c70 | 2114 | event->ctx = ctx; |
0cda4c02 YZ |
2115 | if (event->cpu != -1) |
2116 | event->cpu = cpu; | |
c3f00c70 | 2117 | |
0793a61d TG |
2118 | if (!task) { |
2119 | /* | |
cdd6c482 | 2120 | * Per cpu events are installed via an smp call and |
af901ca1 | 2121 | * the install is always successful. |
0793a61d | 2122 | */ |
fe4b04fa | 2123 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
2124 | return; |
2125 | } | |
2126 | ||
0793a61d | 2127 | retry: |
fe4b04fa PZ |
2128 | if (!task_function_call(task, __perf_install_in_context, event)) |
2129 | return; | |
0793a61d | 2130 | |
e625cce1 | 2131 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 2132 | /* |
fe4b04fa PZ |
2133 | * If we failed to find a running task, but find the context active now |
2134 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 2135 | */ |
fe4b04fa | 2136 | if (ctx->is_active) { |
e625cce1 | 2137 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
2138 | /* |
2139 | * Reload the task pointer, it might have been changed by | |
2140 | * a concurrent perf_event_context_sched_out(). | |
2141 | */ | |
2142 | task = ctx->task; | |
0793a61d TG |
2143 | goto retry; |
2144 | } | |
2145 | ||
2146 | /* | |
fe4b04fa PZ |
2147 | * Since the task isn't running, its safe to add the event, us holding |
2148 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 2149 | */ |
fe4b04fa | 2150 | add_event_to_ctx(event, ctx); |
e625cce1 | 2151 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
2152 | } |
2153 | ||
fa289bec | 2154 | /* |
cdd6c482 | 2155 | * Put a event into inactive state and update time fields. |
fa289bec PM |
2156 | * Enabling the leader of a group effectively enables all |
2157 | * the group members that aren't explicitly disabled, so we | |
2158 | * have to update their ->tstamp_enabled also. | |
2159 | * Note: this works for group members as well as group leaders | |
2160 | * since the non-leader members' sibling_lists will be empty. | |
2161 | */ | |
1d9b482e | 2162 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 2163 | { |
cdd6c482 | 2164 | struct perf_event *sub; |
4158755d | 2165 | u64 tstamp = perf_event_time(event); |
fa289bec | 2166 | |
cdd6c482 | 2167 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 2168 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 2169 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
2170 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
2171 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 2172 | } |
fa289bec PM |
2173 | } |
2174 | ||
d859e29f | 2175 | /* |
cdd6c482 | 2176 | * Cross CPU call to enable a performance event |
d859e29f | 2177 | */ |
fe4b04fa | 2178 | static int __perf_event_enable(void *info) |
04289bb9 | 2179 | { |
cdd6c482 | 2180 | struct perf_event *event = info; |
cdd6c482 IM |
2181 | struct perf_event_context *ctx = event->ctx; |
2182 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2183 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2184 | int err; |
04289bb9 | 2185 | |
06f41796 JO |
2186 | /* |
2187 | * There's a time window between 'ctx->is_active' check | |
2188 | * in perf_event_enable function and this place having: | |
2189 | * - IRQs on | |
2190 | * - ctx->lock unlocked | |
2191 | * | |
2192 | * where the task could be killed and 'ctx' deactivated | |
2193 | * by perf_event_exit_task. | |
2194 | */ | |
2195 | if (!ctx->is_active) | |
fe4b04fa | 2196 | return -EINVAL; |
3cbed429 | 2197 | |
e625cce1 | 2198 | raw_spin_lock(&ctx->lock); |
4af4998b | 2199 | update_context_time(ctx); |
d859e29f | 2200 | |
cdd6c482 | 2201 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2202 | goto unlock; |
e5d1367f SE |
2203 | |
2204 | /* | |
2205 | * set current task's cgroup time reference point | |
2206 | */ | |
3f7cce3c | 2207 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2208 | |
1d9b482e | 2209 | __perf_event_mark_enabled(event); |
04289bb9 | 2210 | |
e5d1367f SE |
2211 | if (!event_filter_match(event)) { |
2212 | if (is_cgroup_event(event)) | |
2213 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2214 | goto unlock; |
e5d1367f | 2215 | } |
f4c4176f | 2216 | |
04289bb9 | 2217 | /* |
cdd6c482 | 2218 | * If the event is in a group and isn't the group leader, |
d859e29f | 2219 | * then don't put it on unless the group is on. |
04289bb9 | 2220 | */ |
cdd6c482 | 2221 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2222 | goto unlock; |
3b6f9e5c | 2223 | |
cdd6c482 | 2224 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2225 | err = -EEXIST; |
e758a33d | 2226 | } else { |
cdd6c482 | 2227 | if (event == leader) |
6e37738a | 2228 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2229 | else |
6e37738a | 2230 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2231 | } |
d859e29f PM |
2232 | |
2233 | if (err) { | |
2234 | /* | |
cdd6c482 | 2235 | * If this event can't go on and it's part of a |
d859e29f PM |
2236 | * group, then the whole group has to come off. |
2237 | */ | |
9e630205 | 2238 | if (leader != event) { |
d859e29f | 2239 | group_sched_out(leader, cpuctx, ctx); |
272325c4 | 2240 | perf_mux_hrtimer_restart(cpuctx); |
9e630205 | 2241 | } |
0d48696f | 2242 | if (leader->attr.pinned) { |
53cfbf59 | 2243 | update_group_times(leader); |
cdd6c482 | 2244 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2245 | } |
d859e29f PM |
2246 | } |
2247 | ||
9ed6060d | 2248 | unlock: |
e625cce1 | 2249 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2250 | |
2251 | return 0; | |
d859e29f PM |
2252 | } |
2253 | ||
2254 | /* | |
cdd6c482 | 2255 | * Enable a event. |
c93f7669 | 2256 | * |
cdd6c482 IM |
2257 | * If event->ctx is a cloned context, callers must make sure that |
2258 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2259 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2260 | * perf_event_for_each_child or perf_event_for_each as described |
2261 | * for perf_event_disable. | |
d859e29f | 2262 | */ |
f63a8daa | 2263 | static void _perf_event_enable(struct perf_event *event) |
d859e29f | 2264 | { |
cdd6c482 | 2265 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2266 | struct task_struct *task = ctx->task; |
2267 | ||
2268 | if (!task) { | |
2269 | /* | |
cdd6c482 | 2270 | * Enable the event on the cpu that it's on |
d859e29f | 2271 | */ |
fe4b04fa | 2272 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2273 | return; |
2274 | } | |
2275 | ||
e625cce1 | 2276 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2277 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2278 | goto out; |
2279 | ||
2280 | /* | |
cdd6c482 IM |
2281 | * If the event is in error state, clear that first. |
2282 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2283 | * know that it has gone back into error state, as distinct |
2284 | * from the task having been scheduled away before the | |
2285 | * cross-call arrived. | |
2286 | */ | |
cdd6c482 IM |
2287 | if (event->state == PERF_EVENT_STATE_ERROR) |
2288 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2289 | |
9ed6060d | 2290 | retry: |
fe4b04fa | 2291 | if (!ctx->is_active) { |
1d9b482e | 2292 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2293 | goto out; |
2294 | } | |
2295 | ||
e625cce1 | 2296 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2297 | |
2298 | if (!task_function_call(task, __perf_event_enable, event)) | |
2299 | return; | |
d859e29f | 2300 | |
e625cce1 | 2301 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2302 | |
2303 | /* | |
cdd6c482 | 2304 | * If the context is active and the event is still off, |
d859e29f PM |
2305 | * we need to retry the cross-call. |
2306 | */ | |
fe4b04fa PZ |
2307 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2308 | /* | |
2309 | * task could have been flipped by a concurrent | |
2310 | * perf_event_context_sched_out() | |
2311 | */ | |
2312 | task = ctx->task; | |
d859e29f | 2313 | goto retry; |
fe4b04fa | 2314 | } |
fa289bec | 2315 | |
9ed6060d | 2316 | out: |
e625cce1 | 2317 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2318 | } |
f63a8daa PZ |
2319 | |
2320 | /* | |
2321 | * See perf_event_disable(); | |
2322 | */ | |
2323 | void perf_event_enable(struct perf_event *event) | |
2324 | { | |
2325 | struct perf_event_context *ctx; | |
2326 | ||
2327 | ctx = perf_event_ctx_lock(event); | |
2328 | _perf_event_enable(event); | |
2329 | perf_event_ctx_unlock(event, ctx); | |
2330 | } | |
dcfce4a0 | 2331 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2332 | |
f63a8daa | 2333 | static int _perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2334 | { |
2023b359 | 2335 | /* |
cdd6c482 | 2336 | * not supported on inherited events |
2023b359 | 2337 | */ |
2e939d1d | 2338 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2339 | return -EINVAL; |
2340 | ||
cdd6c482 | 2341 | atomic_add(refresh, &event->event_limit); |
f63a8daa | 2342 | _perf_event_enable(event); |
2023b359 PZ |
2343 | |
2344 | return 0; | |
79f14641 | 2345 | } |
f63a8daa PZ |
2346 | |
2347 | /* | |
2348 | * See perf_event_disable() | |
2349 | */ | |
2350 | int perf_event_refresh(struct perf_event *event, int refresh) | |
2351 | { | |
2352 | struct perf_event_context *ctx; | |
2353 | int ret; | |
2354 | ||
2355 | ctx = perf_event_ctx_lock(event); | |
2356 | ret = _perf_event_refresh(event, refresh); | |
2357 | perf_event_ctx_unlock(event, ctx); | |
2358 | ||
2359 | return ret; | |
2360 | } | |
26ca5c11 | 2361 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2362 | |
5b0311e1 FW |
2363 | static void ctx_sched_out(struct perf_event_context *ctx, |
2364 | struct perf_cpu_context *cpuctx, | |
2365 | enum event_type_t event_type) | |
235c7fc7 | 2366 | { |
cdd6c482 | 2367 | struct perf_event *event; |
db24d33e | 2368 | int is_active = ctx->is_active; |
235c7fc7 | 2369 | |
db24d33e | 2370 | ctx->is_active &= ~event_type; |
cdd6c482 | 2371 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2372 | return; |
2373 | ||
4af4998b | 2374 | update_context_time(ctx); |
e5d1367f | 2375 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2376 | if (!ctx->nr_active) |
facc4307 | 2377 | return; |
5b0311e1 | 2378 | |
075e0b00 | 2379 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2380 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2381 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2382 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2383 | } |
889ff015 | 2384 | |
db24d33e | 2385 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2386 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2387 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2388 | } |
1b9a644f | 2389 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2390 | } |
2391 | ||
564c2b21 | 2392 | /* |
5a3126d4 PZ |
2393 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2394 | * cloned from the same version of the same context. | |
2395 | * | |
2396 | * Equivalence is measured using a generation number in the context that is | |
2397 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2398 | * and list_del_event(). | |
564c2b21 | 2399 | */ |
cdd6c482 IM |
2400 | static int context_equiv(struct perf_event_context *ctx1, |
2401 | struct perf_event_context *ctx2) | |
564c2b21 | 2402 | { |
211de6eb PZ |
2403 | lockdep_assert_held(&ctx1->lock); |
2404 | lockdep_assert_held(&ctx2->lock); | |
2405 | ||
5a3126d4 PZ |
2406 | /* Pinning disables the swap optimization */ |
2407 | if (ctx1->pin_count || ctx2->pin_count) | |
2408 | return 0; | |
2409 | ||
2410 | /* If ctx1 is the parent of ctx2 */ | |
2411 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2412 | return 1; | |
2413 | ||
2414 | /* If ctx2 is the parent of ctx1 */ | |
2415 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2416 | return 1; | |
2417 | ||
2418 | /* | |
2419 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2420 | * hierarchy, see perf_event_init_context(). | |
2421 | */ | |
2422 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2423 | ctx1->parent_gen == ctx2->parent_gen) | |
2424 | return 1; | |
2425 | ||
2426 | /* Unmatched */ | |
2427 | return 0; | |
564c2b21 PM |
2428 | } |
2429 | ||
cdd6c482 IM |
2430 | static void __perf_event_sync_stat(struct perf_event *event, |
2431 | struct perf_event *next_event) | |
bfbd3381 PZ |
2432 | { |
2433 | u64 value; | |
2434 | ||
cdd6c482 | 2435 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2436 | return; |
2437 | ||
2438 | /* | |
cdd6c482 | 2439 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2440 | * because we're in the middle of a context switch and have IRQs |
2441 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2442 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2443 | * don't need to use it. |
2444 | */ | |
cdd6c482 IM |
2445 | switch (event->state) { |
2446 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2447 | event->pmu->read(event); |
2448 | /* fall-through */ | |
bfbd3381 | 2449 | |
cdd6c482 IM |
2450 | case PERF_EVENT_STATE_INACTIVE: |
2451 | update_event_times(event); | |
bfbd3381 PZ |
2452 | break; |
2453 | ||
2454 | default: | |
2455 | break; | |
2456 | } | |
2457 | ||
2458 | /* | |
cdd6c482 | 2459 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2460 | * values when we flip the contexts. |
2461 | */ | |
e7850595 PZ |
2462 | value = local64_read(&next_event->count); |
2463 | value = local64_xchg(&event->count, value); | |
2464 | local64_set(&next_event->count, value); | |
bfbd3381 | 2465 | |
cdd6c482 IM |
2466 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2467 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2468 | |
bfbd3381 | 2469 | /* |
19d2e755 | 2470 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2471 | */ |
cdd6c482 IM |
2472 | perf_event_update_userpage(event); |
2473 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2474 | } |
2475 | ||
cdd6c482 IM |
2476 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2477 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2478 | { |
cdd6c482 | 2479 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2480 | |
2481 | if (!ctx->nr_stat) | |
2482 | return; | |
2483 | ||
02ffdbc8 PZ |
2484 | update_context_time(ctx); |
2485 | ||
cdd6c482 IM |
2486 | event = list_first_entry(&ctx->event_list, |
2487 | struct perf_event, event_entry); | |
bfbd3381 | 2488 | |
cdd6c482 IM |
2489 | next_event = list_first_entry(&next_ctx->event_list, |
2490 | struct perf_event, event_entry); | |
bfbd3381 | 2491 | |
cdd6c482 IM |
2492 | while (&event->event_entry != &ctx->event_list && |
2493 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2494 | |
cdd6c482 | 2495 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2496 | |
cdd6c482 IM |
2497 | event = list_next_entry(event, event_entry); |
2498 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2499 | } |
2500 | } | |
2501 | ||
fe4b04fa PZ |
2502 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2503 | struct task_struct *next) | |
0793a61d | 2504 | { |
8dc85d54 | 2505 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2506 | struct perf_event_context *next_ctx; |
5a3126d4 | 2507 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2508 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2509 | int do_switch = 1; |
0793a61d | 2510 | |
108b02cf PZ |
2511 | if (likely(!ctx)) |
2512 | return; | |
10989fb2 | 2513 | |
108b02cf PZ |
2514 | cpuctx = __get_cpu_context(ctx); |
2515 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2516 | return; |
2517 | ||
c93f7669 | 2518 | rcu_read_lock(); |
8dc85d54 | 2519 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2520 | if (!next_ctx) |
2521 | goto unlock; | |
2522 | ||
2523 | parent = rcu_dereference(ctx->parent_ctx); | |
2524 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2525 | ||
2526 | /* If neither context have a parent context; they cannot be clones. */ | |
802c8a61 | 2527 | if (!parent && !next_parent) |
5a3126d4 PZ |
2528 | goto unlock; |
2529 | ||
2530 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2531 | /* |
2532 | * Looks like the two contexts are clones, so we might be | |
2533 | * able to optimize the context switch. We lock both | |
2534 | * contexts and check that they are clones under the | |
2535 | * lock (including re-checking that neither has been | |
2536 | * uncloned in the meantime). It doesn't matter which | |
2537 | * order we take the locks because no other cpu could | |
2538 | * be trying to lock both of these tasks. | |
2539 | */ | |
e625cce1 TG |
2540 | raw_spin_lock(&ctx->lock); |
2541 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2542 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2543 | /* |
2544 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2545 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2546 | */ |
8dc85d54 PZ |
2547 | task->perf_event_ctxp[ctxn] = next_ctx; |
2548 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2549 | ctx->task = next; |
2550 | next_ctx->task = task; | |
5a158c3c YZ |
2551 | |
2552 | swap(ctx->task_ctx_data, next_ctx->task_ctx_data); | |
2553 | ||
c93f7669 | 2554 | do_switch = 0; |
bfbd3381 | 2555 | |
cdd6c482 | 2556 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2557 | } |
e625cce1 TG |
2558 | raw_spin_unlock(&next_ctx->lock); |
2559 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2560 | } |
5a3126d4 | 2561 | unlock: |
c93f7669 | 2562 | rcu_read_unlock(); |
564c2b21 | 2563 | |
c93f7669 | 2564 | if (do_switch) { |
facc4307 | 2565 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2566 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2567 | cpuctx->task_ctx = NULL; |
facc4307 | 2568 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2569 | } |
0793a61d TG |
2570 | } |
2571 | ||
ba532500 YZ |
2572 | void perf_sched_cb_dec(struct pmu *pmu) |
2573 | { | |
2574 | this_cpu_dec(perf_sched_cb_usages); | |
2575 | } | |
2576 | ||
2577 | void perf_sched_cb_inc(struct pmu *pmu) | |
2578 | { | |
2579 | this_cpu_inc(perf_sched_cb_usages); | |
2580 | } | |
2581 | ||
2582 | /* | |
2583 | * This function provides the context switch callback to the lower code | |
2584 | * layer. It is invoked ONLY when the context switch callback is enabled. | |
2585 | */ | |
2586 | static void perf_pmu_sched_task(struct task_struct *prev, | |
2587 | struct task_struct *next, | |
2588 | bool sched_in) | |
2589 | { | |
2590 | struct perf_cpu_context *cpuctx; | |
2591 | struct pmu *pmu; | |
2592 | unsigned long flags; | |
2593 | ||
2594 | if (prev == next) | |
2595 | return; | |
2596 | ||
2597 | local_irq_save(flags); | |
2598 | ||
2599 | rcu_read_lock(); | |
2600 | ||
2601 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2602 | if (pmu->sched_task) { | |
2603 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2604 | ||
2605 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2606 | ||
2607 | perf_pmu_disable(pmu); | |
2608 | ||
2609 | pmu->sched_task(cpuctx->task_ctx, sched_in); | |
2610 | ||
2611 | perf_pmu_enable(pmu); | |
2612 | ||
2613 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2614 | } | |
2615 | } | |
2616 | ||
2617 | rcu_read_unlock(); | |
2618 | ||
2619 | local_irq_restore(flags); | |
2620 | } | |
2621 | ||
8dc85d54 PZ |
2622 | #define for_each_task_context_nr(ctxn) \ |
2623 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2624 | ||
2625 | /* | |
2626 | * Called from scheduler to remove the events of the current task, | |
2627 | * with interrupts disabled. | |
2628 | * | |
2629 | * We stop each event and update the event value in event->count. | |
2630 | * | |
2631 | * This does not protect us against NMI, but disable() | |
2632 | * sets the disabled bit in the control field of event _before_ | |
2633 | * accessing the event control register. If a NMI hits, then it will | |
2634 | * not restart the event. | |
2635 | */ | |
ab0cce56 JO |
2636 | void __perf_event_task_sched_out(struct task_struct *task, |
2637 | struct task_struct *next) | |
8dc85d54 PZ |
2638 | { |
2639 | int ctxn; | |
2640 | ||
ba532500 YZ |
2641 | if (__this_cpu_read(perf_sched_cb_usages)) |
2642 | perf_pmu_sched_task(task, next, false); | |
2643 | ||
8dc85d54 PZ |
2644 | for_each_task_context_nr(ctxn) |
2645 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2646 | |
2647 | /* | |
2648 | * if cgroup events exist on this CPU, then we need | |
2649 | * to check if we have to switch out PMU state. | |
2650 | * cgroup event are system-wide mode only | |
2651 | */ | |
4a32fea9 | 2652 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2653 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2654 | } |
2655 | ||
04dc2dbb | 2656 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2657 | { |
108b02cf | 2658 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2659 | |
a63eaf34 PM |
2660 | if (!cpuctx->task_ctx) |
2661 | return; | |
012b84da IM |
2662 | |
2663 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2664 | return; | |
2665 | ||
04dc2dbb | 2666 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2667 | cpuctx->task_ctx = NULL; |
2668 | } | |
2669 | ||
5b0311e1 FW |
2670 | /* |
2671 | * Called with IRQs disabled | |
2672 | */ | |
2673 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2674 | enum event_type_t event_type) | |
2675 | { | |
2676 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2677 | } |
2678 | ||
235c7fc7 | 2679 | static void |
5b0311e1 | 2680 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2681 | struct perf_cpu_context *cpuctx) |
0793a61d | 2682 | { |
cdd6c482 | 2683 | struct perf_event *event; |
0793a61d | 2684 | |
889ff015 FW |
2685 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2686 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2687 | continue; |
5632ab12 | 2688 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2689 | continue; |
2690 | ||
e5d1367f SE |
2691 | /* may need to reset tstamp_enabled */ |
2692 | if (is_cgroup_event(event)) | |
2693 | perf_cgroup_mark_enabled(event, ctx); | |
2694 | ||
8c9ed8e1 | 2695 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2696 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2697 | |
2698 | /* | |
2699 | * If this pinned group hasn't been scheduled, | |
2700 | * put it in error state. | |
2701 | */ | |
cdd6c482 IM |
2702 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2703 | update_group_times(event); | |
2704 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2705 | } |
3b6f9e5c | 2706 | } |
5b0311e1 FW |
2707 | } |
2708 | ||
2709 | static void | |
2710 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2711 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2712 | { |
2713 | struct perf_event *event; | |
2714 | int can_add_hw = 1; | |
3b6f9e5c | 2715 | |
889ff015 FW |
2716 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2717 | /* Ignore events in OFF or ERROR state */ | |
2718 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2719 | continue; |
04289bb9 IM |
2720 | /* |
2721 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2722 | * of events: |
04289bb9 | 2723 | */ |
5632ab12 | 2724 | if (!event_filter_match(event)) |
0793a61d TG |
2725 | continue; |
2726 | ||
e5d1367f SE |
2727 | /* may need to reset tstamp_enabled */ |
2728 | if (is_cgroup_event(event)) | |
2729 | perf_cgroup_mark_enabled(event, ctx); | |
2730 | ||
9ed6060d | 2731 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2732 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2733 | can_add_hw = 0; |
9ed6060d | 2734 | } |
0793a61d | 2735 | } |
5b0311e1 FW |
2736 | } |
2737 | ||
2738 | static void | |
2739 | ctx_sched_in(struct perf_event_context *ctx, | |
2740 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2741 | enum event_type_t event_type, |
2742 | struct task_struct *task) | |
5b0311e1 | 2743 | { |
e5d1367f | 2744 | u64 now; |
db24d33e | 2745 | int is_active = ctx->is_active; |
e5d1367f | 2746 | |
db24d33e | 2747 | ctx->is_active |= event_type; |
5b0311e1 | 2748 | if (likely(!ctx->nr_events)) |
facc4307 | 2749 | return; |
5b0311e1 | 2750 | |
e5d1367f SE |
2751 | now = perf_clock(); |
2752 | ctx->timestamp = now; | |
3f7cce3c | 2753 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2754 | /* |
2755 | * First go through the list and put on any pinned groups | |
2756 | * in order to give them the best chance of going on. | |
2757 | */ | |
db24d33e | 2758 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2759 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2760 | |
2761 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2762 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2763 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2764 | } |
2765 | ||
329c0e01 | 2766 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2767 | enum event_type_t event_type, |
2768 | struct task_struct *task) | |
329c0e01 FW |
2769 | { |
2770 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2771 | ||
e5d1367f | 2772 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2773 | } |
2774 | ||
e5d1367f SE |
2775 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2776 | struct task_struct *task) | |
235c7fc7 | 2777 | { |
108b02cf | 2778 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2779 | |
108b02cf | 2780 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2781 | if (cpuctx->task_ctx == ctx) |
2782 | return; | |
2783 | ||
facc4307 | 2784 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2785 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2786 | /* |
2787 | * We want to keep the following priority order: | |
2788 | * cpu pinned (that don't need to move), task pinned, | |
2789 | * cpu flexible, task flexible. | |
2790 | */ | |
2791 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2792 | ||
1d5f003f GN |
2793 | if (ctx->nr_events) |
2794 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2795 | |
86b47c25 GN |
2796 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2797 | ||
facc4307 PZ |
2798 | perf_pmu_enable(ctx->pmu); |
2799 | perf_ctx_unlock(cpuctx, ctx); | |
235c7fc7 IM |
2800 | } |
2801 | ||
8dc85d54 PZ |
2802 | /* |
2803 | * Called from scheduler to add the events of the current task | |
2804 | * with interrupts disabled. | |
2805 | * | |
2806 | * We restore the event value and then enable it. | |
2807 | * | |
2808 | * This does not protect us against NMI, but enable() | |
2809 | * sets the enabled bit in the control field of event _before_ | |
2810 | * accessing the event control register. If a NMI hits, then it will | |
2811 | * keep the event running. | |
2812 | */ | |
ab0cce56 JO |
2813 | void __perf_event_task_sched_in(struct task_struct *prev, |
2814 | struct task_struct *task) | |
8dc85d54 PZ |
2815 | { |
2816 | struct perf_event_context *ctx; | |
2817 | int ctxn; | |
2818 | ||
2819 | for_each_task_context_nr(ctxn) { | |
2820 | ctx = task->perf_event_ctxp[ctxn]; | |
2821 | if (likely(!ctx)) | |
2822 | continue; | |
2823 | ||
e5d1367f | 2824 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2825 | } |
e5d1367f SE |
2826 | /* |
2827 | * if cgroup events exist on this CPU, then we need | |
2828 | * to check if we have to switch in PMU state. | |
2829 | * cgroup event are system-wide mode only | |
2830 | */ | |
4a32fea9 | 2831 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2832 | perf_cgroup_sched_in(prev, task); |
d010b332 | 2833 | |
ba532500 YZ |
2834 | if (__this_cpu_read(perf_sched_cb_usages)) |
2835 | perf_pmu_sched_task(prev, task, true); | |
235c7fc7 IM |
2836 | } |
2837 | ||
abd50713 PZ |
2838 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2839 | { | |
2840 | u64 frequency = event->attr.sample_freq; | |
2841 | u64 sec = NSEC_PER_SEC; | |
2842 | u64 divisor, dividend; | |
2843 | ||
2844 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2845 | ||
2846 | count_fls = fls64(count); | |
2847 | nsec_fls = fls64(nsec); | |
2848 | frequency_fls = fls64(frequency); | |
2849 | sec_fls = 30; | |
2850 | ||
2851 | /* | |
2852 | * We got @count in @nsec, with a target of sample_freq HZ | |
2853 | * the target period becomes: | |
2854 | * | |
2855 | * @count * 10^9 | |
2856 | * period = ------------------- | |
2857 | * @nsec * sample_freq | |
2858 | * | |
2859 | */ | |
2860 | ||
2861 | /* | |
2862 | * Reduce accuracy by one bit such that @a and @b converge | |
2863 | * to a similar magnitude. | |
2864 | */ | |
fe4b04fa | 2865 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2866 | do { \ |
2867 | if (a##_fls > b##_fls) { \ | |
2868 | a >>= 1; \ | |
2869 | a##_fls--; \ | |
2870 | } else { \ | |
2871 | b >>= 1; \ | |
2872 | b##_fls--; \ | |
2873 | } \ | |
2874 | } while (0) | |
2875 | ||
2876 | /* | |
2877 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2878 | * the other, so that finally we can do a u64/u64 division. | |
2879 | */ | |
2880 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2881 | REDUCE_FLS(nsec, frequency); | |
2882 | REDUCE_FLS(sec, count); | |
2883 | } | |
2884 | ||
2885 | if (count_fls + sec_fls > 64) { | |
2886 | divisor = nsec * frequency; | |
2887 | ||
2888 | while (count_fls + sec_fls > 64) { | |
2889 | REDUCE_FLS(count, sec); | |
2890 | divisor >>= 1; | |
2891 | } | |
2892 | ||
2893 | dividend = count * sec; | |
2894 | } else { | |
2895 | dividend = count * sec; | |
2896 | ||
2897 | while (nsec_fls + frequency_fls > 64) { | |
2898 | REDUCE_FLS(nsec, frequency); | |
2899 | dividend >>= 1; | |
2900 | } | |
2901 | ||
2902 | divisor = nsec * frequency; | |
2903 | } | |
2904 | ||
f6ab91ad PZ |
2905 | if (!divisor) |
2906 | return dividend; | |
2907 | ||
abd50713 PZ |
2908 | return div64_u64(dividend, divisor); |
2909 | } | |
2910 | ||
e050e3f0 SE |
2911 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2912 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2913 | ||
f39d47ff | 2914 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2915 | { |
cdd6c482 | 2916 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2917 | s64 period, sample_period; |
bd2b5b12 PZ |
2918 | s64 delta; |
2919 | ||
abd50713 | 2920 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2921 | |
2922 | delta = (s64)(period - hwc->sample_period); | |
2923 | delta = (delta + 7) / 8; /* low pass filter */ | |
2924 | ||
2925 | sample_period = hwc->sample_period + delta; | |
2926 | ||
2927 | if (!sample_period) | |
2928 | sample_period = 1; | |
2929 | ||
bd2b5b12 | 2930 | hwc->sample_period = sample_period; |
abd50713 | 2931 | |
e7850595 | 2932 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2933 | if (disable) |
2934 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2935 | ||
e7850595 | 2936 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2937 | |
2938 | if (disable) | |
2939 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2940 | } |
bd2b5b12 PZ |
2941 | } |
2942 | ||
e050e3f0 SE |
2943 | /* |
2944 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2945 | * events. At the same time, make sure, having freq events does not change | |
2946 | * the rate of unthrottling as that would introduce bias. | |
2947 | */ | |
2948 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2949 | int needs_unthr) | |
60db5e09 | 2950 | { |
cdd6c482 IM |
2951 | struct perf_event *event; |
2952 | struct hw_perf_event *hwc; | |
e050e3f0 | 2953 | u64 now, period = TICK_NSEC; |
abd50713 | 2954 | s64 delta; |
60db5e09 | 2955 | |
e050e3f0 SE |
2956 | /* |
2957 | * only need to iterate over all events iff: | |
2958 | * - context have events in frequency mode (needs freq adjust) | |
2959 | * - there are events to unthrottle on this cpu | |
2960 | */ | |
2961 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2962 | return; |
2963 | ||
e050e3f0 | 2964 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2965 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2966 | |
03541f8b | 2967 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2968 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2969 | continue; |
2970 | ||
5632ab12 | 2971 | if (!event_filter_match(event)) |
5d27c23d PZ |
2972 | continue; |
2973 | ||
44377277 AS |
2974 | perf_pmu_disable(event->pmu); |
2975 | ||
cdd6c482 | 2976 | hwc = &event->hw; |
6a24ed6c | 2977 | |
ae23bff1 | 2978 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2979 | hwc->interrupts = 0; |
cdd6c482 | 2980 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2981 | event->pmu->start(event, 0); |
a78ac325 PZ |
2982 | } |
2983 | ||
cdd6c482 | 2984 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2985 | goto next; |
60db5e09 | 2986 | |
e050e3f0 SE |
2987 | /* |
2988 | * stop the event and update event->count | |
2989 | */ | |
2990 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2991 | ||
e7850595 | 2992 | now = local64_read(&event->count); |
abd50713 PZ |
2993 | delta = now - hwc->freq_count_stamp; |
2994 | hwc->freq_count_stamp = now; | |
60db5e09 | 2995 | |
e050e3f0 SE |
2996 | /* |
2997 | * restart the event | |
2998 | * reload only if value has changed | |
f39d47ff SE |
2999 | * we have stopped the event so tell that |
3000 | * to perf_adjust_period() to avoid stopping it | |
3001 | * twice. | |
e050e3f0 | 3002 | */ |
abd50713 | 3003 | if (delta > 0) |
f39d47ff | 3004 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
3005 | |
3006 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
3007 | next: |
3008 | perf_pmu_enable(event->pmu); | |
60db5e09 | 3009 | } |
e050e3f0 | 3010 | |
f39d47ff | 3011 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 3012 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
3013 | } |
3014 | ||
235c7fc7 | 3015 | /* |
cdd6c482 | 3016 | * Round-robin a context's events: |
235c7fc7 | 3017 | */ |
cdd6c482 | 3018 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 3019 | { |
dddd3379 TG |
3020 | /* |
3021 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
3022 | * disabled by the inheritance code. | |
3023 | */ | |
3024 | if (!ctx->rotate_disable) | |
3025 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
3026 | } |
3027 | ||
9e630205 | 3028 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 3029 | { |
8dc85d54 | 3030 | struct perf_event_context *ctx = NULL; |
2fde4f94 | 3031 | int rotate = 0; |
7fc23a53 | 3032 | |
b5ab4cd5 | 3033 | if (cpuctx->ctx.nr_events) { |
b5ab4cd5 PZ |
3034 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
3035 | rotate = 1; | |
3036 | } | |
235c7fc7 | 3037 | |
8dc85d54 | 3038 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 3039 | if (ctx && ctx->nr_events) { |
b5ab4cd5 PZ |
3040 | if (ctx->nr_events != ctx->nr_active) |
3041 | rotate = 1; | |
3042 | } | |
9717e6cd | 3043 | |
e050e3f0 | 3044 | if (!rotate) |
0f5a2601 PZ |
3045 | goto done; |
3046 | ||
facc4307 | 3047 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 3048 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 3049 | |
e050e3f0 SE |
3050 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
3051 | if (ctx) | |
3052 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 3053 | |
e050e3f0 SE |
3054 | rotate_ctx(&cpuctx->ctx); |
3055 | if (ctx) | |
3056 | rotate_ctx(ctx); | |
235c7fc7 | 3057 | |
e050e3f0 | 3058 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 3059 | |
0f5a2601 PZ |
3060 | perf_pmu_enable(cpuctx->ctx.pmu); |
3061 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 3062 | done: |
9e630205 SE |
3063 | |
3064 | return rotate; | |
e9d2b064 PZ |
3065 | } |
3066 | ||
026249ef FW |
3067 | #ifdef CONFIG_NO_HZ_FULL |
3068 | bool perf_event_can_stop_tick(void) | |
3069 | { | |
948b26b6 | 3070 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 3071 | __this_cpu_read(perf_throttled_count)) |
026249ef | 3072 | return false; |
d84153d6 FW |
3073 | else |
3074 | return true; | |
026249ef FW |
3075 | } |
3076 | #endif | |
3077 | ||
e9d2b064 PZ |
3078 | void perf_event_task_tick(void) |
3079 | { | |
2fde4f94 MR |
3080 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
3081 | struct perf_event_context *ctx, *tmp; | |
e050e3f0 | 3082 | int throttled; |
b5ab4cd5 | 3083 | |
e9d2b064 PZ |
3084 | WARN_ON(!irqs_disabled()); |
3085 | ||
e050e3f0 SE |
3086 | __this_cpu_inc(perf_throttled_seq); |
3087 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
3088 | ||
2fde4f94 | 3089 | list_for_each_entry_safe(ctx, tmp, head, active_ctx_list) |
e050e3f0 | 3090 | perf_adjust_freq_unthr_context(ctx, throttled); |
0793a61d TG |
3091 | } |
3092 | ||
889ff015 FW |
3093 | static int event_enable_on_exec(struct perf_event *event, |
3094 | struct perf_event_context *ctx) | |
3095 | { | |
3096 | if (!event->attr.enable_on_exec) | |
3097 | return 0; | |
3098 | ||
3099 | event->attr.enable_on_exec = 0; | |
3100 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
3101 | return 0; | |
3102 | ||
1d9b482e | 3103 | __perf_event_mark_enabled(event); |
889ff015 FW |
3104 | |
3105 | return 1; | |
3106 | } | |
3107 | ||
57e7986e | 3108 | /* |
cdd6c482 | 3109 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
3110 | * This expects task == current. |
3111 | */ | |
8dc85d54 | 3112 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 3113 | { |
211de6eb | 3114 | struct perf_event_context *clone_ctx = NULL; |
cdd6c482 | 3115 | struct perf_event *event; |
57e7986e PM |
3116 | unsigned long flags; |
3117 | int enabled = 0; | |
889ff015 | 3118 | int ret; |
57e7986e PM |
3119 | |
3120 | local_irq_save(flags); | |
cdd6c482 | 3121 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
3122 | goto out; |
3123 | ||
e566b76e SE |
3124 | /* |
3125 | * We must ctxsw out cgroup events to avoid conflict | |
3126 | * when invoking perf_task_event_sched_in() later on | |
3127 | * in this function. Otherwise we end up trying to | |
3128 | * ctxswin cgroup events which are already scheduled | |
3129 | * in. | |
3130 | */ | |
a8d757ef | 3131 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 3132 | |
e625cce1 | 3133 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 3134 | task_ctx_sched_out(ctx); |
57e7986e | 3135 | |
b79387ef | 3136 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
3137 | ret = event_enable_on_exec(event, ctx); |
3138 | if (ret) | |
3139 | enabled = 1; | |
57e7986e PM |
3140 | } |
3141 | ||
3142 | /* | |
cdd6c482 | 3143 | * Unclone this context if we enabled any event. |
57e7986e | 3144 | */ |
71a851b4 | 3145 | if (enabled) |
211de6eb | 3146 | clone_ctx = unclone_ctx(ctx); |
57e7986e | 3147 | |
e625cce1 | 3148 | raw_spin_unlock(&ctx->lock); |
57e7986e | 3149 | |
e566b76e SE |
3150 | /* |
3151 | * Also calls ctxswin for cgroup events, if any: | |
3152 | */ | |
e5d1367f | 3153 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 3154 | out: |
57e7986e | 3155 | local_irq_restore(flags); |
211de6eb PZ |
3156 | |
3157 | if (clone_ctx) | |
3158 | put_ctx(clone_ctx); | |
57e7986e PM |
3159 | } |
3160 | ||
e041e328 PZ |
3161 | void perf_event_exec(void) |
3162 | { | |
3163 | struct perf_event_context *ctx; | |
3164 | int ctxn; | |
3165 | ||
3166 | rcu_read_lock(); | |
3167 | for_each_task_context_nr(ctxn) { | |
3168 | ctx = current->perf_event_ctxp[ctxn]; | |
3169 | if (!ctx) | |
3170 | continue; | |
3171 | ||
3172 | perf_event_enable_on_exec(ctx); | |
3173 | } | |
3174 | rcu_read_unlock(); | |
3175 | } | |
3176 | ||
0793a61d | 3177 | /* |
cdd6c482 | 3178 | * Cross CPU call to read the hardware event |
0793a61d | 3179 | */ |
cdd6c482 | 3180 | static void __perf_event_read(void *info) |
0793a61d | 3181 | { |
cdd6c482 IM |
3182 | struct perf_event *event = info; |
3183 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 3184 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3185 | |
e1ac3614 PM |
3186 | /* |
3187 | * If this is a task context, we need to check whether it is | |
3188 | * the current task context of this cpu. If not it has been | |
3189 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3190 | * event->count would have been updated to a recent sample |
3191 | * when the event was scheduled out. | |
e1ac3614 PM |
3192 | */ |
3193 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3194 | return; | |
3195 | ||
e625cce1 | 3196 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3197 | if (ctx->is_active) { |
542e72fc | 3198 | update_context_time(ctx); |
e5d1367f SE |
3199 | update_cgrp_time_from_event(event); |
3200 | } | |
cdd6c482 | 3201 | update_event_times(event); |
542e72fc PZ |
3202 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3203 | event->pmu->read(event); | |
e625cce1 | 3204 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3205 | } |
3206 | ||
b5e58793 PZ |
3207 | static inline u64 perf_event_count(struct perf_event *event) |
3208 | { | |
eacd3ecc MF |
3209 | if (event->pmu->count) |
3210 | return event->pmu->count(event); | |
3211 | ||
3212 | return __perf_event_count(event); | |
b5e58793 PZ |
3213 | } |
3214 | ||
cdd6c482 | 3215 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3216 | { |
3217 | /* | |
cdd6c482 IM |
3218 | * If event is enabled and currently active on a CPU, update the |
3219 | * value in the event structure: | |
0793a61d | 3220 | */ |
cdd6c482 IM |
3221 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3222 | smp_call_function_single(event->oncpu, | |
3223 | __perf_event_read, event, 1); | |
3224 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3225 | struct perf_event_context *ctx = event->ctx; |
3226 | unsigned long flags; | |
3227 | ||
e625cce1 | 3228 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3229 | /* |
3230 | * may read while context is not active | |
3231 | * (e.g., thread is blocked), in that case | |
3232 | * we cannot update context time | |
3233 | */ | |
e5d1367f | 3234 | if (ctx->is_active) { |
c530ccd9 | 3235 | update_context_time(ctx); |
e5d1367f SE |
3236 | update_cgrp_time_from_event(event); |
3237 | } | |
cdd6c482 | 3238 | update_event_times(event); |
e625cce1 | 3239 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3240 | } |
3241 | ||
b5e58793 | 3242 | return perf_event_count(event); |
0793a61d TG |
3243 | } |
3244 | ||
a63eaf34 | 3245 | /* |
cdd6c482 | 3246 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3247 | */ |
eb184479 | 3248 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3249 | { |
e625cce1 | 3250 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3251 | mutex_init(&ctx->mutex); |
2fde4f94 | 3252 | INIT_LIST_HEAD(&ctx->active_ctx_list); |
889ff015 FW |
3253 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3254 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3255 | INIT_LIST_HEAD(&ctx->event_list); |
3256 | atomic_set(&ctx->refcount, 1); | |
fadfe7be | 3257 | INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work); |
eb184479 PZ |
3258 | } |
3259 | ||
3260 | static struct perf_event_context * | |
3261 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3262 | { | |
3263 | struct perf_event_context *ctx; | |
3264 | ||
3265 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3266 | if (!ctx) | |
3267 | return NULL; | |
3268 | ||
3269 | __perf_event_init_context(ctx); | |
3270 | if (task) { | |
3271 | ctx->task = task; | |
3272 | get_task_struct(task); | |
0793a61d | 3273 | } |
eb184479 PZ |
3274 | ctx->pmu = pmu; |
3275 | ||
3276 | return ctx; | |
a63eaf34 PM |
3277 | } |
3278 | ||
2ebd4ffb MH |
3279 | static struct task_struct * |
3280 | find_lively_task_by_vpid(pid_t vpid) | |
3281 | { | |
3282 | struct task_struct *task; | |
3283 | int err; | |
0793a61d TG |
3284 | |
3285 | rcu_read_lock(); | |
2ebd4ffb | 3286 | if (!vpid) |
0793a61d TG |
3287 | task = current; |
3288 | else | |
2ebd4ffb | 3289 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3290 | if (task) |
3291 | get_task_struct(task); | |
3292 | rcu_read_unlock(); | |
3293 | ||
3294 | if (!task) | |
3295 | return ERR_PTR(-ESRCH); | |
3296 | ||
0793a61d | 3297 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3298 | err = -EACCES; |
3299 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3300 | goto errout; | |
3301 | ||
2ebd4ffb MH |
3302 | return task; |
3303 | errout: | |
3304 | put_task_struct(task); | |
3305 | return ERR_PTR(err); | |
3306 | ||
3307 | } | |
3308 | ||
fe4b04fa PZ |
3309 | /* |
3310 | * Returns a matching context with refcount and pincount. | |
3311 | */ | |
108b02cf | 3312 | static struct perf_event_context * |
4af57ef2 YZ |
3313 | find_get_context(struct pmu *pmu, struct task_struct *task, |
3314 | struct perf_event *event) | |
0793a61d | 3315 | { |
211de6eb | 3316 | struct perf_event_context *ctx, *clone_ctx = NULL; |
22a4f650 | 3317 | struct perf_cpu_context *cpuctx; |
4af57ef2 | 3318 | void *task_ctx_data = NULL; |
25346b93 | 3319 | unsigned long flags; |
8dc85d54 | 3320 | int ctxn, err; |
4af57ef2 | 3321 | int cpu = event->cpu; |
0793a61d | 3322 | |
22a4ec72 | 3323 | if (!task) { |
cdd6c482 | 3324 | /* Must be root to operate on a CPU event: */ |
0764771d | 3325 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3326 | return ERR_PTR(-EACCES); |
3327 | ||
0793a61d | 3328 | /* |
cdd6c482 | 3329 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3330 | * offline CPU and activate it when the CPU comes up, but |
3331 | * that's for later. | |
3332 | */ | |
f6325e30 | 3333 | if (!cpu_online(cpu)) |
0793a61d TG |
3334 | return ERR_PTR(-ENODEV); |
3335 | ||
108b02cf | 3336 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3337 | ctx = &cpuctx->ctx; |
c93f7669 | 3338 | get_ctx(ctx); |
fe4b04fa | 3339 | ++ctx->pin_count; |
0793a61d | 3340 | |
0793a61d TG |
3341 | return ctx; |
3342 | } | |
3343 | ||
8dc85d54 PZ |
3344 | err = -EINVAL; |
3345 | ctxn = pmu->task_ctx_nr; | |
3346 | if (ctxn < 0) | |
3347 | goto errout; | |
3348 | ||
4af57ef2 YZ |
3349 | if (event->attach_state & PERF_ATTACH_TASK_DATA) { |
3350 | task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL); | |
3351 | if (!task_ctx_data) { | |
3352 | err = -ENOMEM; | |
3353 | goto errout; | |
3354 | } | |
3355 | } | |
3356 | ||
9ed6060d | 3357 | retry: |
8dc85d54 | 3358 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3359 | if (ctx) { |
211de6eb | 3360 | clone_ctx = unclone_ctx(ctx); |
fe4b04fa | 3361 | ++ctx->pin_count; |
4af57ef2 YZ |
3362 | |
3363 | if (task_ctx_data && !ctx->task_ctx_data) { | |
3364 | ctx->task_ctx_data = task_ctx_data; | |
3365 | task_ctx_data = NULL; | |
3366 | } | |
e625cce1 | 3367 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
211de6eb PZ |
3368 | |
3369 | if (clone_ctx) | |
3370 | put_ctx(clone_ctx); | |
9137fb28 | 3371 | } else { |
eb184479 | 3372 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3373 | err = -ENOMEM; |
3374 | if (!ctx) | |
3375 | goto errout; | |
eb184479 | 3376 | |
4af57ef2 YZ |
3377 | if (task_ctx_data) { |
3378 | ctx->task_ctx_data = task_ctx_data; | |
3379 | task_ctx_data = NULL; | |
3380 | } | |
3381 | ||
dbe08d82 ON |
3382 | err = 0; |
3383 | mutex_lock(&task->perf_event_mutex); | |
3384 | /* | |
3385 | * If it has already passed perf_event_exit_task(). | |
3386 | * we must see PF_EXITING, it takes this mutex too. | |
3387 | */ | |
3388 | if (task->flags & PF_EXITING) | |
3389 | err = -ESRCH; | |
3390 | else if (task->perf_event_ctxp[ctxn]) | |
3391 | err = -EAGAIN; | |
fe4b04fa | 3392 | else { |
9137fb28 | 3393 | get_ctx(ctx); |
fe4b04fa | 3394 | ++ctx->pin_count; |
dbe08d82 | 3395 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3396 | } |
dbe08d82 ON |
3397 | mutex_unlock(&task->perf_event_mutex); |
3398 | ||
3399 | if (unlikely(err)) { | |
9137fb28 | 3400 | put_ctx(ctx); |
dbe08d82 ON |
3401 | |
3402 | if (err == -EAGAIN) | |
3403 | goto retry; | |
3404 | goto errout; | |
a63eaf34 PM |
3405 | } |
3406 | } | |
3407 | ||
4af57ef2 | 3408 | kfree(task_ctx_data); |
0793a61d | 3409 | return ctx; |
c93f7669 | 3410 | |
9ed6060d | 3411 | errout: |
4af57ef2 | 3412 | kfree(task_ctx_data); |
c93f7669 | 3413 | return ERR_PTR(err); |
0793a61d TG |
3414 | } |
3415 | ||
6fb2915d | 3416 | static void perf_event_free_filter(struct perf_event *event); |
2541517c | 3417 | static void perf_event_free_bpf_prog(struct perf_event *event); |
6fb2915d | 3418 | |
cdd6c482 | 3419 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3420 | { |
cdd6c482 | 3421 | struct perf_event *event; |
592903cd | 3422 | |
cdd6c482 IM |
3423 | event = container_of(head, struct perf_event, rcu_head); |
3424 | if (event->ns) | |
3425 | put_pid_ns(event->ns); | |
6fb2915d | 3426 | perf_event_free_filter(event); |
cdd6c482 | 3427 | kfree(event); |
592903cd PZ |
3428 | } |
3429 | ||
b69cf536 PZ |
3430 | static void ring_buffer_attach(struct perf_event *event, |
3431 | struct ring_buffer *rb); | |
925d519a | 3432 | |
4beb31f3 | 3433 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3434 | { |
4beb31f3 FW |
3435 | if (event->parent) |
3436 | return; | |
3437 | ||
4beb31f3 FW |
3438 | if (is_cgroup_event(event)) |
3439 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3440 | } | |
925d519a | 3441 | |
4beb31f3 FW |
3442 | static void unaccount_event(struct perf_event *event) |
3443 | { | |
3444 | if (event->parent) | |
3445 | return; | |
3446 | ||
3447 | if (event->attach_state & PERF_ATTACH_TASK) | |
3448 | static_key_slow_dec_deferred(&perf_sched_events); | |
3449 | if (event->attr.mmap || event->attr.mmap_data) | |
3450 | atomic_dec(&nr_mmap_events); | |
3451 | if (event->attr.comm) | |
3452 | atomic_dec(&nr_comm_events); | |
3453 | if (event->attr.task) | |
3454 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3455 | if (event->attr.freq) |
3456 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3457 | if (is_cgroup_event(event)) |
3458 | static_key_slow_dec_deferred(&perf_sched_events); | |
3459 | if (has_branch_stack(event)) | |
3460 | static_key_slow_dec_deferred(&perf_sched_events); | |
3461 | ||
3462 | unaccount_event_cpu(event, event->cpu); | |
3463 | } | |
925d519a | 3464 | |
bed5b25a AS |
3465 | /* |
3466 | * The following implement mutual exclusion of events on "exclusive" pmus | |
3467 | * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled | |
3468 | * at a time, so we disallow creating events that might conflict, namely: | |
3469 | * | |
3470 | * 1) cpu-wide events in the presence of per-task events, | |
3471 | * 2) per-task events in the presence of cpu-wide events, | |
3472 | * 3) two matching events on the same context. | |
3473 | * | |
3474 | * The former two cases are handled in the allocation path (perf_event_alloc(), | |
3475 | * __free_event()), the latter -- before the first perf_install_in_context(). | |
3476 | */ | |
3477 | static int exclusive_event_init(struct perf_event *event) | |
3478 | { | |
3479 | struct pmu *pmu = event->pmu; | |
3480 | ||
3481 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) | |
3482 | return 0; | |
3483 | ||
3484 | /* | |
3485 | * Prevent co-existence of per-task and cpu-wide events on the | |
3486 | * same exclusive pmu. | |
3487 | * | |
3488 | * Negative pmu::exclusive_cnt means there are cpu-wide | |
3489 | * events on this "exclusive" pmu, positive means there are | |
3490 | * per-task events. | |
3491 | * | |
3492 | * Since this is called in perf_event_alloc() path, event::ctx | |
3493 | * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK | |
3494 | * to mean "per-task event", because unlike other attach states it | |
3495 | * never gets cleared. | |
3496 | */ | |
3497 | if (event->attach_state & PERF_ATTACH_TASK) { | |
3498 | if (!atomic_inc_unless_negative(&pmu->exclusive_cnt)) | |
3499 | return -EBUSY; | |
3500 | } else { | |
3501 | if (!atomic_dec_unless_positive(&pmu->exclusive_cnt)) | |
3502 | return -EBUSY; | |
3503 | } | |
3504 | ||
3505 | return 0; | |
3506 | } | |
3507 | ||
3508 | static void exclusive_event_destroy(struct perf_event *event) | |
3509 | { | |
3510 | struct pmu *pmu = event->pmu; | |
3511 | ||
3512 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) | |
3513 | return; | |
3514 | ||
3515 | /* see comment in exclusive_event_init() */ | |
3516 | if (event->attach_state & PERF_ATTACH_TASK) | |
3517 | atomic_dec(&pmu->exclusive_cnt); | |
3518 | else | |
3519 | atomic_inc(&pmu->exclusive_cnt); | |
3520 | } | |
3521 | ||
3522 | static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2) | |
3523 | { | |
3524 | if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && | |
3525 | (e1->cpu == e2->cpu || | |
3526 | e1->cpu == -1 || | |
3527 | e2->cpu == -1)) | |
3528 | return true; | |
3529 | return false; | |
3530 | } | |
3531 | ||
3532 | /* Called under the same ctx::mutex as perf_install_in_context() */ | |
3533 | static bool exclusive_event_installable(struct perf_event *event, | |
3534 | struct perf_event_context *ctx) | |
3535 | { | |
3536 | struct perf_event *iter_event; | |
3537 | struct pmu *pmu = event->pmu; | |
3538 | ||
3539 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) | |
3540 | return true; | |
3541 | ||
3542 | list_for_each_entry(iter_event, &ctx->event_list, event_entry) { | |
3543 | if (exclusive_event_match(iter_event, event)) | |
3544 | return false; | |
3545 | } | |
3546 | ||
3547 | return true; | |
3548 | } | |
3549 | ||
766d6c07 FW |
3550 | static void __free_event(struct perf_event *event) |
3551 | { | |
cdd6c482 | 3552 | if (!event->parent) { |
927c7a9e FW |
3553 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3554 | put_callchain_buffers(); | |
f344011c | 3555 | } |
9ee318a7 | 3556 | |
dead9f29 AS |
3557 | perf_event_free_bpf_prog(event); |
3558 | ||
766d6c07 FW |
3559 | if (event->destroy) |
3560 | event->destroy(event); | |
3561 | ||
3562 | if (event->ctx) | |
3563 | put_ctx(event->ctx); | |
3564 | ||
bed5b25a AS |
3565 | if (event->pmu) { |
3566 | exclusive_event_destroy(event); | |
c464c76e | 3567 | module_put(event->pmu->module); |
bed5b25a | 3568 | } |
c464c76e | 3569 | |
766d6c07 FW |
3570 | call_rcu(&event->rcu_head, free_event_rcu); |
3571 | } | |
683ede43 PZ |
3572 | |
3573 | static void _free_event(struct perf_event *event) | |
f1600952 | 3574 | { |
e360adbe | 3575 | irq_work_sync(&event->pending); |
925d519a | 3576 | |
4beb31f3 | 3577 | unaccount_event(event); |
9ee318a7 | 3578 | |
76369139 | 3579 | if (event->rb) { |
9bb5d40c PZ |
3580 | /* |
3581 | * Can happen when we close an event with re-directed output. | |
3582 | * | |
3583 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3584 | * over us; possibly making our ring_buffer_put() the last. | |
3585 | */ | |
3586 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3587 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3588 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3589 | } |
3590 | ||
e5d1367f SE |
3591 | if (is_cgroup_event(event)) |
3592 | perf_detach_cgroup(event); | |
3593 | ||
766d6c07 | 3594 | __free_event(event); |
f1600952 PZ |
3595 | } |
3596 | ||
683ede43 PZ |
3597 | /* |
3598 | * Used to free events which have a known refcount of 1, such as in error paths | |
3599 | * where the event isn't exposed yet and inherited events. | |
3600 | */ | |
3601 | static void free_event(struct perf_event *event) | |
0793a61d | 3602 | { |
683ede43 PZ |
3603 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3604 | "unexpected event refcount: %ld; ptr=%p\n", | |
3605 | atomic_long_read(&event->refcount), event)) { | |
3606 | /* leak to avoid use-after-free */ | |
3607 | return; | |
3608 | } | |
0793a61d | 3609 | |
683ede43 | 3610 | _free_event(event); |
0793a61d TG |
3611 | } |
3612 | ||
a66a3052 | 3613 | /* |
f8697762 | 3614 | * Remove user event from the owner task. |
a66a3052 | 3615 | */ |
f8697762 | 3616 | static void perf_remove_from_owner(struct perf_event *event) |
fb0459d7 | 3617 | { |
8882135b | 3618 | struct task_struct *owner; |
fb0459d7 | 3619 | |
8882135b PZ |
3620 | rcu_read_lock(); |
3621 | owner = ACCESS_ONCE(event->owner); | |
3622 | /* | |
3623 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3624 | * !owner it means the list deletion is complete and we can indeed | |
3625 | * free this event, otherwise we need to serialize on | |
3626 | * owner->perf_event_mutex. | |
3627 | */ | |
3628 | smp_read_barrier_depends(); | |
3629 | if (owner) { | |
3630 | /* | |
3631 | * Since delayed_put_task_struct() also drops the last | |
3632 | * task reference we can safely take a new reference | |
3633 | * while holding the rcu_read_lock(). | |
3634 | */ | |
3635 | get_task_struct(owner); | |
3636 | } | |
3637 | rcu_read_unlock(); | |
3638 | ||
3639 | if (owner) { | |
f63a8daa PZ |
3640 | /* |
3641 | * If we're here through perf_event_exit_task() we're already | |
3642 | * holding ctx->mutex which would be an inversion wrt. the | |
3643 | * normal lock order. | |
3644 | * | |
3645 | * However we can safely take this lock because its the child | |
3646 | * ctx->mutex. | |
3647 | */ | |
3648 | mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); | |
3649 | ||
8882135b PZ |
3650 | /* |
3651 | * We have to re-check the event->owner field, if it is cleared | |
3652 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3653 | * ensured they're done, and we can proceed with freeing the | |
3654 | * event. | |
3655 | */ | |
3656 | if (event->owner) | |
3657 | list_del_init(&event->owner_entry); | |
3658 | mutex_unlock(&owner->perf_event_mutex); | |
3659 | put_task_struct(owner); | |
3660 | } | |
f8697762 JO |
3661 | } |
3662 | ||
f8697762 JO |
3663 | static void put_event(struct perf_event *event) |
3664 | { | |
a83fe28e | 3665 | struct perf_event_context *ctx; |
f8697762 JO |
3666 | |
3667 | if (!atomic_long_dec_and_test(&event->refcount)) | |
3668 | return; | |
3669 | ||
3670 | if (!is_kernel_event(event)) | |
3671 | perf_remove_from_owner(event); | |
8882135b | 3672 | |
683ede43 PZ |
3673 | /* |
3674 | * There are two ways this annotation is useful: | |
3675 | * | |
3676 | * 1) there is a lock recursion from perf_event_exit_task | |
3677 | * see the comment there. | |
3678 | * | |
3679 | * 2) there is a lock-inversion with mmap_sem through | |
3680 | * perf_event_read_group(), which takes faults while | |
3681 | * holding ctx->mutex, however this is called after | |
3682 | * the last filedesc died, so there is no possibility | |
3683 | * to trigger the AB-BA case. | |
3684 | */ | |
a83fe28e PZ |
3685 | ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING); |
3686 | WARN_ON_ONCE(ctx->parent_ctx); | |
683ede43 | 3687 | perf_remove_from_context(event, true); |
d415a7f1 | 3688 | perf_event_ctx_unlock(event, ctx); |
683ede43 PZ |
3689 | |
3690 | _free_event(event); | |
a6fa941d AV |
3691 | } |
3692 | ||
683ede43 PZ |
3693 | int perf_event_release_kernel(struct perf_event *event) |
3694 | { | |
3695 | put_event(event); | |
3696 | return 0; | |
3697 | } | |
3698 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3699 | ||
8b10c5e2 PZ |
3700 | /* |
3701 | * Called when the last reference to the file is gone. | |
3702 | */ | |
a6fa941d AV |
3703 | static int perf_release(struct inode *inode, struct file *file) |
3704 | { | |
3705 | put_event(file->private_data); | |
3706 | return 0; | |
fb0459d7 | 3707 | } |
fb0459d7 | 3708 | |
fadfe7be JO |
3709 | /* |
3710 | * Remove all orphanes events from the context. | |
3711 | */ | |
3712 | static void orphans_remove_work(struct work_struct *work) | |
3713 | { | |
3714 | struct perf_event_context *ctx; | |
3715 | struct perf_event *event, *tmp; | |
3716 | ||
3717 | ctx = container_of(work, struct perf_event_context, | |
3718 | orphans_remove.work); | |
3719 | ||
3720 | mutex_lock(&ctx->mutex); | |
3721 | list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) { | |
3722 | struct perf_event *parent_event = event->parent; | |
3723 | ||
3724 | if (!is_orphaned_child(event)) | |
3725 | continue; | |
3726 | ||
3727 | perf_remove_from_context(event, true); | |
3728 | ||
3729 | mutex_lock(&parent_event->child_mutex); | |
3730 | list_del_init(&event->child_list); | |
3731 | mutex_unlock(&parent_event->child_mutex); | |
3732 | ||
3733 | free_event(event); | |
3734 | put_event(parent_event); | |
3735 | } | |
3736 | ||
3737 | raw_spin_lock_irq(&ctx->lock); | |
3738 | ctx->orphans_remove_sched = false; | |
3739 | raw_spin_unlock_irq(&ctx->lock); | |
3740 | mutex_unlock(&ctx->mutex); | |
3741 | ||
3742 | put_ctx(ctx); | |
3743 | } | |
3744 | ||
59ed446f | 3745 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3746 | { |
cdd6c482 | 3747 | struct perf_event *child; |
e53c0994 PZ |
3748 | u64 total = 0; |
3749 | ||
59ed446f PZ |
3750 | *enabled = 0; |
3751 | *running = 0; | |
3752 | ||
6f10581a | 3753 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3754 | total += perf_event_read(event); |
59ed446f PZ |
3755 | *enabled += event->total_time_enabled + |
3756 | atomic64_read(&event->child_total_time_enabled); | |
3757 | *running += event->total_time_running + | |
3758 | atomic64_read(&event->child_total_time_running); | |
3759 | ||
3760 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3761 | total += perf_event_read(child); |
59ed446f PZ |
3762 | *enabled += child->total_time_enabled; |
3763 | *running += child->total_time_running; | |
3764 | } | |
6f10581a | 3765 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3766 | |
3767 | return total; | |
3768 | } | |
fb0459d7 | 3769 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3770 | |
cdd6c482 | 3771 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3772 | u64 read_format, char __user *buf) |
3773 | { | |
cdd6c482 | 3774 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a | 3775 | struct perf_event_context *ctx = leader->ctx; |
f63a8daa | 3776 | int n = 0, size = 0, ret; |
59ed446f | 3777 | u64 count, enabled, running; |
f63a8daa PZ |
3778 | u64 values[5]; |
3779 | ||
3780 | lockdep_assert_held(&ctx->mutex); | |
abf4868b | 3781 | |
59ed446f | 3782 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3783 | |
3784 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3785 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3786 | values[n++] = enabled; | |
3787 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3788 | values[n++] = running; | |
abf4868b PZ |
3789 | values[n++] = count; |
3790 | if (read_format & PERF_FORMAT_ID) | |
3791 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3792 | |
3793 | size = n * sizeof(u64); | |
3794 | ||
3795 | if (copy_to_user(buf, values, size)) | |
f63a8daa | 3796 | return -EFAULT; |
3dab77fb | 3797 | |
6f10581a | 3798 | ret = size; |
3dab77fb | 3799 | |
65abc865 | 3800 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3801 | n = 0; |
3dab77fb | 3802 | |
59ed446f | 3803 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3804 | if (read_format & PERF_FORMAT_ID) |
3805 | values[n++] = primary_event_id(sub); | |
3806 | ||
3807 | size = n * sizeof(u64); | |
3808 | ||
184d3da8 | 3809 | if (copy_to_user(buf + ret, values, size)) { |
f63a8daa | 3810 | return -EFAULT; |
6f10581a | 3811 | } |
abf4868b PZ |
3812 | |
3813 | ret += size; | |
3dab77fb PZ |
3814 | } |
3815 | ||
abf4868b | 3816 | return ret; |
3dab77fb PZ |
3817 | } |
3818 | ||
cdd6c482 | 3819 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3820 | u64 read_format, char __user *buf) |
3821 | { | |
59ed446f | 3822 | u64 enabled, running; |
3dab77fb PZ |
3823 | u64 values[4]; |
3824 | int n = 0; | |
3825 | ||
59ed446f PZ |
3826 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3827 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3828 | values[n++] = enabled; | |
3829 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3830 | values[n++] = running; | |
3dab77fb | 3831 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3832 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3833 | |
3834 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3835 | return -EFAULT; | |
3836 | ||
3837 | return n * sizeof(u64); | |
3838 | } | |
3839 | ||
dc633982 JO |
3840 | static bool is_event_hup(struct perf_event *event) |
3841 | { | |
3842 | bool no_children; | |
3843 | ||
3844 | if (event->state != PERF_EVENT_STATE_EXIT) | |
3845 | return false; | |
3846 | ||
3847 | mutex_lock(&event->child_mutex); | |
3848 | no_children = list_empty(&event->child_list); | |
3849 | mutex_unlock(&event->child_mutex); | |
3850 | return no_children; | |
3851 | } | |
3852 | ||
0793a61d | 3853 | /* |
cdd6c482 | 3854 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3855 | */ |
3856 | static ssize_t | |
cdd6c482 | 3857 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3858 | { |
cdd6c482 | 3859 | u64 read_format = event->attr.read_format; |
3dab77fb | 3860 | int ret; |
0793a61d | 3861 | |
3b6f9e5c | 3862 | /* |
cdd6c482 | 3863 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3864 | * error state (i.e. because it was pinned but it couldn't be |
3865 | * scheduled on to the CPU at some point). | |
3866 | */ | |
cdd6c482 | 3867 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3868 | return 0; |
3869 | ||
c320c7b7 | 3870 | if (count < event->read_size) |
3dab77fb PZ |
3871 | return -ENOSPC; |
3872 | ||
cdd6c482 | 3873 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3874 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3875 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3876 | else |
cdd6c482 | 3877 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3878 | |
3dab77fb | 3879 | return ret; |
0793a61d TG |
3880 | } |
3881 | ||
0793a61d TG |
3882 | static ssize_t |
3883 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3884 | { | |
cdd6c482 | 3885 | struct perf_event *event = file->private_data; |
f63a8daa PZ |
3886 | struct perf_event_context *ctx; |
3887 | int ret; | |
0793a61d | 3888 | |
f63a8daa PZ |
3889 | ctx = perf_event_ctx_lock(event); |
3890 | ret = perf_read_hw(event, buf, count); | |
3891 | perf_event_ctx_unlock(event, ctx); | |
3892 | ||
3893 | return ret; | |
0793a61d TG |
3894 | } |
3895 | ||
3896 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3897 | { | |
cdd6c482 | 3898 | struct perf_event *event = file->private_data; |
76369139 | 3899 | struct ring_buffer *rb; |
61b67684 | 3900 | unsigned int events = POLLHUP; |
c7138f37 | 3901 | |
e708d7ad | 3902 | poll_wait(file, &event->waitq, wait); |
179033b3 | 3903 | |
dc633982 | 3904 | if (is_event_hup(event)) |
179033b3 | 3905 | return events; |
c7138f37 | 3906 | |
10c6db11 | 3907 | /* |
9bb5d40c PZ |
3908 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3909 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3910 | */ |
3911 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3912 | rb = event->rb; |
3913 | if (rb) | |
76369139 | 3914 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3915 | mutex_unlock(&event->mmap_mutex); |
0793a61d TG |
3916 | return events; |
3917 | } | |
3918 | ||
f63a8daa | 3919 | static void _perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3920 | { |
cdd6c482 | 3921 | (void)perf_event_read(event); |
e7850595 | 3922 | local64_set(&event->count, 0); |
cdd6c482 | 3923 | perf_event_update_userpage(event); |
3df5edad PZ |
3924 | } |
3925 | ||
c93f7669 | 3926 | /* |
cdd6c482 IM |
3927 | * Holding the top-level event's child_mutex means that any |
3928 | * descendant process that has inherited this event will block | |
3929 | * in sync_child_event if it goes to exit, thus satisfying the | |
3930 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3931 | */ |
cdd6c482 IM |
3932 | static void perf_event_for_each_child(struct perf_event *event, |
3933 | void (*func)(struct perf_event *)) | |
3df5edad | 3934 | { |
cdd6c482 | 3935 | struct perf_event *child; |
3df5edad | 3936 | |
cdd6c482 | 3937 | WARN_ON_ONCE(event->ctx->parent_ctx); |
f63a8daa | 3938 | |
cdd6c482 IM |
3939 | mutex_lock(&event->child_mutex); |
3940 | func(event); | |
3941 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3942 | func(child); |
cdd6c482 | 3943 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3944 | } |
3945 | ||
cdd6c482 IM |
3946 | static void perf_event_for_each(struct perf_event *event, |
3947 | void (*func)(struct perf_event *)) | |
3df5edad | 3948 | { |
cdd6c482 IM |
3949 | struct perf_event_context *ctx = event->ctx; |
3950 | struct perf_event *sibling; | |
3df5edad | 3951 | |
f63a8daa PZ |
3952 | lockdep_assert_held(&ctx->mutex); |
3953 | ||
cdd6c482 | 3954 | event = event->group_leader; |
75f937f2 | 3955 | |
cdd6c482 | 3956 | perf_event_for_each_child(event, func); |
cdd6c482 | 3957 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3958 | perf_event_for_each_child(sibling, func); |
6de6a7b9 PZ |
3959 | } |
3960 | ||
c7999c6f PZ |
3961 | struct period_event { |
3962 | struct perf_event *event; | |
08247e31 | 3963 | u64 value; |
c7999c6f | 3964 | }; |
08247e31 | 3965 | |
c7999c6f PZ |
3966 | static int __perf_event_period(void *info) |
3967 | { | |
3968 | struct period_event *pe = info; | |
3969 | struct perf_event *event = pe->event; | |
3970 | struct perf_event_context *ctx = event->ctx; | |
3971 | u64 value = pe->value; | |
3972 | bool active; | |
08247e31 | 3973 | |
c7999c6f | 3974 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 3975 | if (event->attr.freq) { |
cdd6c482 | 3976 | event->attr.sample_freq = value; |
08247e31 | 3977 | } else { |
cdd6c482 IM |
3978 | event->attr.sample_period = value; |
3979 | event->hw.sample_period = value; | |
08247e31 | 3980 | } |
bad7192b PZ |
3981 | |
3982 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3983 | if (active) { | |
3984 | perf_pmu_disable(ctx->pmu); | |
3985 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3986 | } | |
3987 | ||
3988 | local64_set(&event->hw.period_left, 0); | |
3989 | ||
3990 | if (active) { | |
3991 | event->pmu->start(event, PERF_EF_RELOAD); | |
3992 | perf_pmu_enable(ctx->pmu); | |
3993 | } | |
c7999c6f | 3994 | raw_spin_unlock(&ctx->lock); |
bad7192b | 3995 | |
c7999c6f PZ |
3996 | return 0; |
3997 | } | |
3998 | ||
3999 | static int perf_event_period(struct perf_event *event, u64 __user *arg) | |
4000 | { | |
4001 | struct period_event pe = { .event = event, }; | |
4002 | struct perf_event_context *ctx = event->ctx; | |
4003 | struct task_struct *task; | |
4004 | u64 value; | |
4005 | ||
4006 | if (!is_sampling_event(event)) | |
4007 | return -EINVAL; | |
4008 | ||
4009 | if (copy_from_user(&value, arg, sizeof(value))) | |
4010 | return -EFAULT; | |
4011 | ||
4012 | if (!value) | |
4013 | return -EINVAL; | |
4014 | ||
4015 | if (event->attr.freq && value > sysctl_perf_event_sample_rate) | |
4016 | return -EINVAL; | |
4017 | ||
4018 | task = ctx->task; | |
4019 | pe.value = value; | |
4020 | ||
4021 | if (!task) { | |
4022 | cpu_function_call(event->cpu, __perf_event_period, &pe); | |
4023 | return 0; | |
4024 | } | |
4025 | ||
4026 | retry: | |
4027 | if (!task_function_call(task, __perf_event_period, &pe)) | |
4028 | return 0; | |
4029 | ||
4030 | raw_spin_lock_irq(&ctx->lock); | |
4031 | if (ctx->is_active) { | |
4032 | raw_spin_unlock_irq(&ctx->lock); | |
4033 | task = ctx->task; | |
4034 | goto retry; | |
4035 | } | |
4036 | ||
4037 | __perf_event_period(&pe); | |
e625cce1 | 4038 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 | 4039 | |
c7999c6f | 4040 | return 0; |
08247e31 PZ |
4041 | } |
4042 | ||
ac9721f3 PZ |
4043 | static const struct file_operations perf_fops; |
4044 | ||
2903ff01 | 4045 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 4046 | { |
2903ff01 AV |
4047 | struct fd f = fdget(fd); |
4048 | if (!f.file) | |
4049 | return -EBADF; | |
ac9721f3 | 4050 | |
2903ff01 AV |
4051 | if (f.file->f_op != &perf_fops) { |
4052 | fdput(f); | |
4053 | return -EBADF; | |
ac9721f3 | 4054 | } |
2903ff01 AV |
4055 | *p = f; |
4056 | return 0; | |
ac9721f3 PZ |
4057 | } |
4058 | ||
4059 | static int perf_event_set_output(struct perf_event *event, | |
4060 | struct perf_event *output_event); | |
6fb2915d | 4061 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
2541517c | 4062 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd); |
a4be7c27 | 4063 | |
f63a8daa | 4064 | static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg) |
d859e29f | 4065 | { |
cdd6c482 | 4066 | void (*func)(struct perf_event *); |
3df5edad | 4067 | u32 flags = arg; |
d859e29f PM |
4068 | |
4069 | switch (cmd) { | |
cdd6c482 | 4070 | case PERF_EVENT_IOC_ENABLE: |
f63a8daa | 4071 | func = _perf_event_enable; |
d859e29f | 4072 | break; |
cdd6c482 | 4073 | case PERF_EVENT_IOC_DISABLE: |
f63a8daa | 4074 | func = _perf_event_disable; |
79f14641 | 4075 | break; |
cdd6c482 | 4076 | case PERF_EVENT_IOC_RESET: |
f63a8daa | 4077 | func = _perf_event_reset; |
6de6a7b9 | 4078 | break; |
3df5edad | 4079 | |
cdd6c482 | 4080 | case PERF_EVENT_IOC_REFRESH: |
f63a8daa | 4081 | return _perf_event_refresh(event, arg); |
08247e31 | 4082 | |
cdd6c482 IM |
4083 | case PERF_EVENT_IOC_PERIOD: |
4084 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 4085 | |
cf4957f1 JO |
4086 | case PERF_EVENT_IOC_ID: |
4087 | { | |
4088 | u64 id = primary_event_id(event); | |
4089 | ||
4090 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
4091 | return -EFAULT; | |
4092 | return 0; | |
4093 | } | |
4094 | ||
cdd6c482 | 4095 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 4096 | { |
ac9721f3 | 4097 | int ret; |
ac9721f3 | 4098 | if (arg != -1) { |
2903ff01 AV |
4099 | struct perf_event *output_event; |
4100 | struct fd output; | |
4101 | ret = perf_fget_light(arg, &output); | |
4102 | if (ret) | |
4103 | return ret; | |
4104 | output_event = output.file->private_data; | |
4105 | ret = perf_event_set_output(event, output_event); | |
4106 | fdput(output); | |
4107 | } else { | |
4108 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 4109 | } |
ac9721f3 PZ |
4110 | return ret; |
4111 | } | |
a4be7c27 | 4112 | |
6fb2915d LZ |
4113 | case PERF_EVENT_IOC_SET_FILTER: |
4114 | return perf_event_set_filter(event, (void __user *)arg); | |
4115 | ||
2541517c AS |
4116 | case PERF_EVENT_IOC_SET_BPF: |
4117 | return perf_event_set_bpf_prog(event, arg); | |
4118 | ||
d859e29f | 4119 | default: |
3df5edad | 4120 | return -ENOTTY; |
d859e29f | 4121 | } |
3df5edad PZ |
4122 | |
4123 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 4124 | perf_event_for_each(event, func); |
3df5edad | 4125 | else |
cdd6c482 | 4126 | perf_event_for_each_child(event, func); |
3df5edad PZ |
4127 | |
4128 | return 0; | |
d859e29f PM |
4129 | } |
4130 | ||
f63a8daa PZ |
4131 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
4132 | { | |
4133 | struct perf_event *event = file->private_data; | |
4134 | struct perf_event_context *ctx; | |
4135 | long ret; | |
4136 | ||
4137 | ctx = perf_event_ctx_lock(event); | |
4138 | ret = _perf_ioctl(event, cmd, arg); | |
4139 | perf_event_ctx_unlock(event, ctx); | |
4140 | ||
4141 | return ret; | |
4142 | } | |
4143 | ||
b3f20785 PM |
4144 | #ifdef CONFIG_COMPAT |
4145 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, | |
4146 | unsigned long arg) | |
4147 | { | |
4148 | switch (_IOC_NR(cmd)) { | |
4149 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): | |
4150 | case _IOC_NR(PERF_EVENT_IOC_ID): | |
4151 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ | |
4152 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { | |
4153 | cmd &= ~IOCSIZE_MASK; | |
4154 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; | |
4155 | } | |
4156 | break; | |
4157 | } | |
4158 | return perf_ioctl(file, cmd, arg); | |
4159 | } | |
4160 | #else | |
4161 | # define perf_compat_ioctl NULL | |
4162 | #endif | |
4163 | ||
cdd6c482 | 4164 | int perf_event_task_enable(void) |
771d7cde | 4165 | { |
f63a8daa | 4166 | struct perf_event_context *ctx; |
cdd6c482 | 4167 | struct perf_event *event; |
771d7cde | 4168 | |
cdd6c482 | 4169 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4170 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4171 | ctx = perf_event_ctx_lock(event); | |
4172 | perf_event_for_each_child(event, _perf_event_enable); | |
4173 | perf_event_ctx_unlock(event, ctx); | |
4174 | } | |
cdd6c482 | 4175 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4176 | |
4177 | return 0; | |
4178 | } | |
4179 | ||
cdd6c482 | 4180 | int perf_event_task_disable(void) |
771d7cde | 4181 | { |
f63a8daa | 4182 | struct perf_event_context *ctx; |
cdd6c482 | 4183 | struct perf_event *event; |
771d7cde | 4184 | |
cdd6c482 | 4185 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4186 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4187 | ctx = perf_event_ctx_lock(event); | |
4188 | perf_event_for_each_child(event, _perf_event_disable); | |
4189 | perf_event_ctx_unlock(event, ctx); | |
4190 | } | |
cdd6c482 | 4191 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4192 | |
4193 | return 0; | |
4194 | } | |
4195 | ||
cdd6c482 | 4196 | static int perf_event_index(struct perf_event *event) |
194002b2 | 4197 | { |
a4eaf7f1 PZ |
4198 | if (event->hw.state & PERF_HES_STOPPED) |
4199 | return 0; | |
4200 | ||
cdd6c482 | 4201 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
4202 | return 0; |
4203 | ||
35edc2a5 | 4204 | return event->pmu->event_idx(event); |
194002b2 PZ |
4205 | } |
4206 | ||
c4794295 | 4207 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 4208 | u64 *now, |
7f310a5d EM |
4209 | u64 *enabled, |
4210 | u64 *running) | |
c4794295 | 4211 | { |
e3f3541c | 4212 | u64 ctx_time; |
c4794295 | 4213 | |
e3f3541c PZ |
4214 | *now = perf_clock(); |
4215 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
4216 | *enabled = ctx_time - event->tstamp_enabled; |
4217 | *running = ctx_time - event->tstamp_running; | |
4218 | } | |
4219 | ||
fa731587 PZ |
4220 | static void perf_event_init_userpage(struct perf_event *event) |
4221 | { | |
4222 | struct perf_event_mmap_page *userpg; | |
4223 | struct ring_buffer *rb; | |
4224 | ||
4225 | rcu_read_lock(); | |
4226 | rb = rcu_dereference(event->rb); | |
4227 | if (!rb) | |
4228 | goto unlock; | |
4229 | ||
4230 | userpg = rb->user_page; | |
4231 | ||
4232 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
4233 | userpg->cap_bit0_is_deprecated = 1; | |
4234 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
e8c6deac AS |
4235 | userpg->data_offset = PAGE_SIZE; |
4236 | userpg->data_size = perf_data_size(rb); | |
fa731587 PZ |
4237 | |
4238 | unlock: | |
4239 | rcu_read_unlock(); | |
4240 | } | |
4241 | ||
c1317ec2 AL |
4242 | void __weak arch_perf_update_userpage( |
4243 | struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now) | |
e3f3541c PZ |
4244 | { |
4245 | } | |
4246 | ||
38ff667b PZ |
4247 | /* |
4248 | * Callers need to ensure there can be no nesting of this function, otherwise | |
4249 | * the seqlock logic goes bad. We can not serialize this because the arch | |
4250 | * code calls this from NMI context. | |
4251 | */ | |
cdd6c482 | 4252 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 4253 | { |
cdd6c482 | 4254 | struct perf_event_mmap_page *userpg; |
76369139 | 4255 | struct ring_buffer *rb; |
e3f3541c | 4256 | u64 enabled, running, now; |
38ff667b PZ |
4257 | |
4258 | rcu_read_lock(); | |
5ec4c599 PZ |
4259 | rb = rcu_dereference(event->rb); |
4260 | if (!rb) | |
4261 | goto unlock; | |
4262 | ||
0d641208 EM |
4263 | /* |
4264 | * compute total_time_enabled, total_time_running | |
4265 | * based on snapshot values taken when the event | |
4266 | * was last scheduled in. | |
4267 | * | |
4268 | * we cannot simply called update_context_time() | |
4269 | * because of locking issue as we can be called in | |
4270 | * NMI context | |
4271 | */ | |
e3f3541c | 4272 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 4273 | |
76369139 | 4274 | userpg = rb->user_page; |
7b732a75 PZ |
4275 | /* |
4276 | * Disable preemption so as to not let the corresponding user-space | |
4277 | * spin too long if we get preempted. | |
4278 | */ | |
4279 | preempt_disable(); | |
37d81828 | 4280 | ++userpg->lock; |
92f22a38 | 4281 | barrier(); |
cdd6c482 | 4282 | userpg->index = perf_event_index(event); |
b5e58793 | 4283 | userpg->offset = perf_event_count(event); |
365a4038 | 4284 | if (userpg->index) |
e7850595 | 4285 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 4286 | |
0d641208 | 4287 | userpg->time_enabled = enabled + |
cdd6c482 | 4288 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 4289 | |
0d641208 | 4290 | userpg->time_running = running + |
cdd6c482 | 4291 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 4292 | |
c1317ec2 | 4293 | arch_perf_update_userpage(event, userpg, now); |
e3f3541c | 4294 | |
92f22a38 | 4295 | barrier(); |
37d81828 | 4296 | ++userpg->lock; |
7b732a75 | 4297 | preempt_enable(); |
38ff667b | 4298 | unlock: |
7b732a75 | 4299 | rcu_read_unlock(); |
37d81828 PM |
4300 | } |
4301 | ||
906010b2 PZ |
4302 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
4303 | { | |
4304 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 4305 | struct ring_buffer *rb; |
906010b2 PZ |
4306 | int ret = VM_FAULT_SIGBUS; |
4307 | ||
4308 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
4309 | if (vmf->pgoff == 0) | |
4310 | ret = 0; | |
4311 | return ret; | |
4312 | } | |
4313 | ||
4314 | rcu_read_lock(); | |
76369139 FW |
4315 | rb = rcu_dereference(event->rb); |
4316 | if (!rb) | |
906010b2 PZ |
4317 | goto unlock; |
4318 | ||
4319 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
4320 | goto unlock; | |
4321 | ||
76369139 | 4322 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
4323 | if (!vmf->page) |
4324 | goto unlock; | |
4325 | ||
4326 | get_page(vmf->page); | |
4327 | vmf->page->mapping = vma->vm_file->f_mapping; | |
4328 | vmf->page->index = vmf->pgoff; | |
4329 | ||
4330 | ret = 0; | |
4331 | unlock: | |
4332 | rcu_read_unlock(); | |
4333 | ||
4334 | return ret; | |
4335 | } | |
4336 | ||
10c6db11 PZ |
4337 | static void ring_buffer_attach(struct perf_event *event, |
4338 | struct ring_buffer *rb) | |
4339 | { | |
b69cf536 | 4340 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
4341 | unsigned long flags; |
4342 | ||
b69cf536 PZ |
4343 | if (event->rb) { |
4344 | /* | |
4345 | * Should be impossible, we set this when removing | |
4346 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
4347 | */ | |
4348 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 4349 | |
b69cf536 | 4350 | old_rb = event->rb; |
b69cf536 PZ |
4351 | spin_lock_irqsave(&old_rb->event_lock, flags); |
4352 | list_del_rcu(&event->rb_entry); | |
4353 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
10c6db11 | 4354 | |
2f993cf0 ON |
4355 | event->rcu_batches = get_state_synchronize_rcu(); |
4356 | event->rcu_pending = 1; | |
b69cf536 | 4357 | } |
10c6db11 | 4358 | |
b69cf536 | 4359 | if (rb) { |
2f993cf0 ON |
4360 | if (event->rcu_pending) { |
4361 | cond_synchronize_rcu(event->rcu_batches); | |
4362 | event->rcu_pending = 0; | |
4363 | } | |
4364 | ||
b69cf536 PZ |
4365 | spin_lock_irqsave(&rb->event_lock, flags); |
4366 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
4367 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
4368 | } | |
4369 | ||
4370 | rcu_assign_pointer(event->rb, rb); | |
4371 | ||
4372 | if (old_rb) { | |
4373 | ring_buffer_put(old_rb); | |
4374 | /* | |
4375 | * Since we detached before setting the new rb, so that we | |
4376 | * could attach the new rb, we could have missed a wakeup. | |
4377 | * Provide it now. | |
4378 | */ | |
4379 | wake_up_all(&event->waitq); | |
4380 | } | |
10c6db11 PZ |
4381 | } |
4382 | ||
4383 | static void ring_buffer_wakeup(struct perf_event *event) | |
4384 | { | |
4385 | struct ring_buffer *rb; | |
4386 | ||
4387 | rcu_read_lock(); | |
4388 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
4389 | if (rb) { |
4390 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
4391 | wake_up_all(&event->waitq); | |
4392 | } | |
10c6db11 PZ |
4393 | rcu_read_unlock(); |
4394 | } | |
4395 | ||
fdc26706 | 4396 | struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 4397 | { |
76369139 | 4398 | struct ring_buffer *rb; |
7b732a75 | 4399 | |
ac9721f3 | 4400 | rcu_read_lock(); |
76369139 FW |
4401 | rb = rcu_dereference(event->rb); |
4402 | if (rb) { | |
4403 | if (!atomic_inc_not_zero(&rb->refcount)) | |
4404 | rb = NULL; | |
ac9721f3 PZ |
4405 | } |
4406 | rcu_read_unlock(); | |
4407 | ||
76369139 | 4408 | return rb; |
ac9721f3 PZ |
4409 | } |
4410 | ||
fdc26706 | 4411 | void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 4412 | { |
76369139 | 4413 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 4414 | return; |
7b732a75 | 4415 | |
9bb5d40c | 4416 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 4417 | |
76369139 | 4418 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
4419 | } |
4420 | ||
4421 | static void perf_mmap_open(struct vm_area_struct *vma) | |
4422 | { | |
cdd6c482 | 4423 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4424 | |
cdd6c482 | 4425 | atomic_inc(&event->mmap_count); |
9bb5d40c | 4426 | atomic_inc(&event->rb->mmap_count); |
1e0fb9ec | 4427 | |
45bfb2e5 PZ |
4428 | if (vma->vm_pgoff) |
4429 | atomic_inc(&event->rb->aux_mmap_count); | |
4430 | ||
1e0fb9ec AL |
4431 | if (event->pmu->event_mapped) |
4432 | event->pmu->event_mapped(event); | |
7b732a75 PZ |
4433 | } |
4434 | ||
9bb5d40c PZ |
4435 | /* |
4436 | * A buffer can be mmap()ed multiple times; either directly through the same | |
4437 | * event, or through other events by use of perf_event_set_output(). | |
4438 | * | |
4439 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
4440 | * the buffer here, where we still have a VM context. This means we need | |
4441 | * to detach all events redirecting to us. | |
4442 | */ | |
7b732a75 PZ |
4443 | static void perf_mmap_close(struct vm_area_struct *vma) |
4444 | { | |
cdd6c482 | 4445 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4446 | |
b69cf536 | 4447 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
4448 | struct user_struct *mmap_user = rb->mmap_user; |
4449 | int mmap_locked = rb->mmap_locked; | |
4450 | unsigned long size = perf_data_size(rb); | |
789f90fc | 4451 | |
1e0fb9ec AL |
4452 | if (event->pmu->event_unmapped) |
4453 | event->pmu->event_unmapped(event); | |
4454 | ||
45bfb2e5 PZ |
4455 | /* |
4456 | * rb->aux_mmap_count will always drop before rb->mmap_count and | |
4457 | * event->mmap_count, so it is ok to use event->mmap_mutex to | |
4458 | * serialize with perf_mmap here. | |
4459 | */ | |
4460 | if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff && | |
4461 | atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) { | |
4462 | atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm); | |
4463 | vma->vm_mm->pinned_vm -= rb->aux_mmap_locked; | |
4464 | ||
4465 | rb_free_aux(rb); | |
4466 | mutex_unlock(&event->mmap_mutex); | |
4467 | } | |
4468 | ||
9bb5d40c PZ |
4469 | atomic_dec(&rb->mmap_count); |
4470 | ||
4471 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4472 | goto out_put; |
9bb5d40c | 4473 | |
b69cf536 | 4474 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4475 | mutex_unlock(&event->mmap_mutex); |
4476 | ||
4477 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4478 | if (atomic_read(&rb->mmap_count)) |
4479 | goto out_put; | |
ac9721f3 | 4480 | |
9bb5d40c PZ |
4481 | /* |
4482 | * No other mmap()s, detach from all other events that might redirect | |
4483 | * into the now unreachable buffer. Somewhat complicated by the | |
4484 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4485 | */ | |
4486 | again: | |
4487 | rcu_read_lock(); | |
4488 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4489 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4490 | /* | |
4491 | * This event is en-route to free_event() which will | |
4492 | * detach it and remove it from the list. | |
4493 | */ | |
4494 | continue; | |
4495 | } | |
4496 | rcu_read_unlock(); | |
789f90fc | 4497 | |
9bb5d40c PZ |
4498 | mutex_lock(&event->mmap_mutex); |
4499 | /* | |
4500 | * Check we didn't race with perf_event_set_output() which can | |
4501 | * swizzle the rb from under us while we were waiting to | |
4502 | * acquire mmap_mutex. | |
4503 | * | |
4504 | * If we find a different rb; ignore this event, a next | |
4505 | * iteration will no longer find it on the list. We have to | |
4506 | * still restart the iteration to make sure we're not now | |
4507 | * iterating the wrong list. | |
4508 | */ | |
b69cf536 PZ |
4509 | if (event->rb == rb) |
4510 | ring_buffer_attach(event, NULL); | |
4511 | ||
cdd6c482 | 4512 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4513 | put_event(event); |
ac9721f3 | 4514 | |
9bb5d40c PZ |
4515 | /* |
4516 | * Restart the iteration; either we're on the wrong list or | |
4517 | * destroyed its integrity by doing a deletion. | |
4518 | */ | |
4519 | goto again; | |
7b732a75 | 4520 | } |
9bb5d40c PZ |
4521 | rcu_read_unlock(); |
4522 | ||
4523 | /* | |
4524 | * It could be there's still a few 0-ref events on the list; they'll | |
4525 | * get cleaned up by free_event() -- they'll also still have their | |
4526 | * ref on the rb and will free it whenever they are done with it. | |
4527 | * | |
4528 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4529 | * undo the VM accounting. | |
4530 | */ | |
4531 | ||
4532 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4533 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4534 | free_uid(mmap_user); | |
4535 | ||
b69cf536 | 4536 | out_put: |
9bb5d40c | 4537 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4538 | } |
4539 | ||
f0f37e2f | 4540 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 | 4541 | .open = perf_mmap_open, |
45bfb2e5 | 4542 | .close = perf_mmap_close, /* non mergable */ |
43a21ea8 PZ |
4543 | .fault = perf_mmap_fault, |
4544 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4545 | }; |
4546 | ||
4547 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4548 | { | |
cdd6c482 | 4549 | struct perf_event *event = file->private_data; |
22a4f650 | 4550 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4551 | struct user_struct *user = current_user(); |
22a4f650 | 4552 | unsigned long locked, lock_limit; |
45bfb2e5 | 4553 | struct ring_buffer *rb = NULL; |
7b732a75 PZ |
4554 | unsigned long vma_size; |
4555 | unsigned long nr_pages; | |
45bfb2e5 | 4556 | long user_extra = 0, extra = 0; |
d57e34fd | 4557 | int ret = 0, flags = 0; |
37d81828 | 4558 | |
c7920614 PZ |
4559 | /* |
4560 | * Don't allow mmap() of inherited per-task counters. This would | |
4561 | * create a performance issue due to all children writing to the | |
76369139 | 4562 | * same rb. |
c7920614 PZ |
4563 | */ |
4564 | if (event->cpu == -1 && event->attr.inherit) | |
4565 | return -EINVAL; | |
4566 | ||
43a21ea8 | 4567 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4568 | return -EINVAL; |
7b732a75 PZ |
4569 | |
4570 | vma_size = vma->vm_end - vma->vm_start; | |
45bfb2e5 PZ |
4571 | |
4572 | if (vma->vm_pgoff == 0) { | |
4573 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4574 | } else { | |
4575 | /* | |
4576 | * AUX area mapping: if rb->aux_nr_pages != 0, it's already | |
4577 | * mapped, all subsequent mappings should have the same size | |
4578 | * and offset. Must be above the normal perf buffer. | |
4579 | */ | |
4580 | u64 aux_offset, aux_size; | |
4581 | ||
4582 | if (!event->rb) | |
4583 | return -EINVAL; | |
4584 | ||
4585 | nr_pages = vma_size / PAGE_SIZE; | |
4586 | ||
4587 | mutex_lock(&event->mmap_mutex); | |
4588 | ret = -EINVAL; | |
4589 | ||
4590 | rb = event->rb; | |
4591 | if (!rb) | |
4592 | goto aux_unlock; | |
4593 | ||
4594 | aux_offset = ACCESS_ONCE(rb->user_page->aux_offset); | |
4595 | aux_size = ACCESS_ONCE(rb->user_page->aux_size); | |
4596 | ||
4597 | if (aux_offset < perf_data_size(rb) + PAGE_SIZE) | |
4598 | goto aux_unlock; | |
4599 | ||
4600 | if (aux_offset != vma->vm_pgoff << PAGE_SHIFT) | |
4601 | goto aux_unlock; | |
4602 | ||
4603 | /* already mapped with a different offset */ | |
4604 | if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff) | |
4605 | goto aux_unlock; | |
4606 | ||
4607 | if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE) | |
4608 | goto aux_unlock; | |
4609 | ||
4610 | /* already mapped with a different size */ | |
4611 | if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages) | |
4612 | goto aux_unlock; | |
4613 | ||
4614 | if (!is_power_of_2(nr_pages)) | |
4615 | goto aux_unlock; | |
4616 | ||
4617 | if (!atomic_inc_not_zero(&rb->mmap_count)) | |
4618 | goto aux_unlock; | |
4619 | ||
4620 | if (rb_has_aux(rb)) { | |
4621 | atomic_inc(&rb->aux_mmap_count); | |
4622 | ret = 0; | |
4623 | goto unlock; | |
4624 | } | |
4625 | ||
4626 | atomic_set(&rb->aux_mmap_count, 1); | |
4627 | user_extra = nr_pages; | |
4628 | ||
4629 | goto accounting; | |
4630 | } | |
7b732a75 | 4631 | |
7730d865 | 4632 | /* |
76369139 | 4633 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4634 | * can do bitmasks instead of modulo. |
4635 | */ | |
2ed11312 | 4636 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) |
37d81828 PM |
4637 | return -EINVAL; |
4638 | ||
7b732a75 | 4639 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4640 | return -EINVAL; |
4641 | ||
cdd6c482 | 4642 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4643 | again: |
cdd6c482 | 4644 | mutex_lock(&event->mmap_mutex); |
76369139 | 4645 | if (event->rb) { |
9bb5d40c | 4646 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4647 | ret = -EINVAL; |
9bb5d40c PZ |
4648 | goto unlock; |
4649 | } | |
4650 | ||
4651 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4652 | /* | |
4653 | * Raced against perf_mmap_close() through | |
4654 | * perf_event_set_output(). Try again, hope for better | |
4655 | * luck. | |
4656 | */ | |
4657 | mutex_unlock(&event->mmap_mutex); | |
4658 | goto again; | |
4659 | } | |
4660 | ||
ebb3c4c4 PZ |
4661 | goto unlock; |
4662 | } | |
4663 | ||
789f90fc | 4664 | user_extra = nr_pages + 1; |
45bfb2e5 PZ |
4665 | |
4666 | accounting: | |
cdd6c482 | 4667 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4668 | |
4669 | /* | |
4670 | * Increase the limit linearly with more CPUs: | |
4671 | */ | |
4672 | user_lock_limit *= num_online_cpus(); | |
4673 | ||
789f90fc | 4674 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4675 | |
789f90fc PZ |
4676 | if (user_locked > user_lock_limit) |
4677 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4678 | |
78d7d407 | 4679 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4680 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4681 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4682 | |
459ec28a IM |
4683 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4684 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4685 | ret = -EPERM; |
4686 | goto unlock; | |
4687 | } | |
7b732a75 | 4688 | |
45bfb2e5 | 4689 | WARN_ON(!rb && event->rb); |
906010b2 | 4690 | |
d57e34fd | 4691 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4692 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4693 | |
76369139 | 4694 | if (!rb) { |
45bfb2e5 PZ |
4695 | rb = rb_alloc(nr_pages, |
4696 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4697 | event->cpu, flags); | |
26cb63ad | 4698 | |
45bfb2e5 PZ |
4699 | if (!rb) { |
4700 | ret = -ENOMEM; | |
4701 | goto unlock; | |
4702 | } | |
43a21ea8 | 4703 | |
45bfb2e5 PZ |
4704 | atomic_set(&rb->mmap_count, 1); |
4705 | rb->mmap_user = get_current_user(); | |
4706 | rb->mmap_locked = extra; | |
26cb63ad | 4707 | |
45bfb2e5 | 4708 | ring_buffer_attach(event, rb); |
ac9721f3 | 4709 | |
45bfb2e5 PZ |
4710 | perf_event_init_userpage(event); |
4711 | perf_event_update_userpage(event); | |
4712 | } else { | |
1a594131 AS |
4713 | ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages, |
4714 | event->attr.aux_watermark, flags); | |
45bfb2e5 PZ |
4715 | if (!ret) |
4716 | rb->aux_mmap_locked = extra; | |
4717 | } | |
9a0f05cb | 4718 | |
ebb3c4c4 | 4719 | unlock: |
45bfb2e5 PZ |
4720 | if (!ret) { |
4721 | atomic_long_add(user_extra, &user->locked_vm); | |
4722 | vma->vm_mm->pinned_vm += extra; | |
4723 | ||
ac9721f3 | 4724 | atomic_inc(&event->mmap_count); |
45bfb2e5 PZ |
4725 | } else if (rb) { |
4726 | atomic_dec(&rb->mmap_count); | |
4727 | } | |
4728 | aux_unlock: | |
cdd6c482 | 4729 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4730 | |
9bb5d40c PZ |
4731 | /* |
4732 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4733 | * vma. | |
4734 | */ | |
26cb63ad | 4735 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4736 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 | 4737 | |
1e0fb9ec AL |
4738 | if (event->pmu->event_mapped) |
4739 | event->pmu->event_mapped(event); | |
4740 | ||
7b732a75 | 4741 | return ret; |
37d81828 PM |
4742 | } |
4743 | ||
3c446b3d PZ |
4744 | static int perf_fasync(int fd, struct file *filp, int on) |
4745 | { | |
496ad9aa | 4746 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4747 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4748 | int retval; |
4749 | ||
4750 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4751 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4752 | mutex_unlock(&inode->i_mutex); |
4753 | ||
4754 | if (retval < 0) | |
4755 | return retval; | |
4756 | ||
4757 | return 0; | |
4758 | } | |
4759 | ||
0793a61d | 4760 | static const struct file_operations perf_fops = { |
3326c1ce | 4761 | .llseek = no_llseek, |
0793a61d TG |
4762 | .release = perf_release, |
4763 | .read = perf_read, | |
4764 | .poll = perf_poll, | |
d859e29f | 4765 | .unlocked_ioctl = perf_ioctl, |
b3f20785 | 4766 | .compat_ioctl = perf_compat_ioctl, |
37d81828 | 4767 | .mmap = perf_mmap, |
3c446b3d | 4768 | .fasync = perf_fasync, |
0793a61d TG |
4769 | }; |
4770 | ||
925d519a | 4771 | /* |
cdd6c482 | 4772 | * Perf event wakeup |
925d519a PZ |
4773 | * |
4774 | * If there's data, ensure we set the poll() state and publish everything | |
4775 | * to user-space before waking everybody up. | |
4776 | */ | |
4777 | ||
fed66e2c PZ |
4778 | static inline struct fasync_struct **perf_event_fasync(struct perf_event *event) |
4779 | { | |
4780 | /* only the parent has fasync state */ | |
4781 | if (event->parent) | |
4782 | event = event->parent; | |
4783 | return &event->fasync; | |
4784 | } | |
4785 | ||
cdd6c482 | 4786 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4787 | { |
10c6db11 | 4788 | ring_buffer_wakeup(event); |
4c9e2542 | 4789 | |
cdd6c482 | 4790 | if (event->pending_kill) { |
fed66e2c | 4791 | kill_fasync(perf_event_fasync(event), SIGIO, event->pending_kill); |
cdd6c482 | 4792 | event->pending_kill = 0; |
4c9e2542 | 4793 | } |
925d519a PZ |
4794 | } |
4795 | ||
e360adbe | 4796 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4797 | { |
cdd6c482 IM |
4798 | struct perf_event *event = container_of(entry, |
4799 | struct perf_event, pending); | |
d525211f PZ |
4800 | int rctx; |
4801 | ||
4802 | rctx = perf_swevent_get_recursion_context(); | |
4803 | /* | |
4804 | * If we 'fail' here, that's OK, it means recursion is already disabled | |
4805 | * and we won't recurse 'further'. | |
4806 | */ | |
79f14641 | 4807 | |
cdd6c482 IM |
4808 | if (event->pending_disable) { |
4809 | event->pending_disable = 0; | |
4810 | __perf_event_disable(event); | |
79f14641 PZ |
4811 | } |
4812 | ||
cdd6c482 IM |
4813 | if (event->pending_wakeup) { |
4814 | event->pending_wakeup = 0; | |
4815 | perf_event_wakeup(event); | |
79f14641 | 4816 | } |
d525211f PZ |
4817 | |
4818 | if (rctx >= 0) | |
4819 | perf_swevent_put_recursion_context(rctx); | |
79f14641 PZ |
4820 | } |
4821 | ||
39447b38 ZY |
4822 | /* |
4823 | * We assume there is only KVM supporting the callbacks. | |
4824 | * Later on, we might change it to a list if there is | |
4825 | * another virtualization implementation supporting the callbacks. | |
4826 | */ | |
4827 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4828 | ||
4829 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4830 | { | |
4831 | perf_guest_cbs = cbs; | |
4832 | return 0; | |
4833 | } | |
4834 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4835 | ||
4836 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4837 | { | |
4838 | perf_guest_cbs = NULL; | |
4839 | return 0; | |
4840 | } | |
4841 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4842 | ||
4018994f JO |
4843 | static void |
4844 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4845 | struct pt_regs *regs, u64 mask) | |
4846 | { | |
4847 | int bit; | |
4848 | ||
4849 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4850 | sizeof(mask) * BITS_PER_BYTE) { | |
4851 | u64 val; | |
4852 | ||
4853 | val = perf_reg_value(regs, bit); | |
4854 | perf_output_put(handle, val); | |
4855 | } | |
4856 | } | |
4857 | ||
60e2364e | 4858 | static void perf_sample_regs_user(struct perf_regs *regs_user, |
88a7c26a AL |
4859 | struct pt_regs *regs, |
4860 | struct pt_regs *regs_user_copy) | |
4018994f | 4861 | { |
88a7c26a AL |
4862 | if (user_mode(regs)) { |
4863 | regs_user->abi = perf_reg_abi(current); | |
2565711f | 4864 | regs_user->regs = regs; |
88a7c26a AL |
4865 | } else if (current->mm) { |
4866 | perf_get_regs_user(regs_user, regs, regs_user_copy); | |
2565711f PZ |
4867 | } else { |
4868 | regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; | |
4869 | regs_user->regs = NULL; | |
4018994f JO |
4870 | } |
4871 | } | |
4872 | ||
60e2364e SE |
4873 | static void perf_sample_regs_intr(struct perf_regs *regs_intr, |
4874 | struct pt_regs *regs) | |
4875 | { | |
4876 | regs_intr->regs = regs; | |
4877 | regs_intr->abi = perf_reg_abi(current); | |
4878 | } | |
4879 | ||
4880 | ||
c5ebcedb JO |
4881 | /* |
4882 | * Get remaining task size from user stack pointer. | |
4883 | * | |
4884 | * It'd be better to take stack vma map and limit this more | |
4885 | * precisly, but there's no way to get it safely under interrupt, | |
4886 | * so using TASK_SIZE as limit. | |
4887 | */ | |
4888 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4889 | { | |
4890 | unsigned long addr = perf_user_stack_pointer(regs); | |
4891 | ||
4892 | if (!addr || addr >= TASK_SIZE) | |
4893 | return 0; | |
4894 | ||
4895 | return TASK_SIZE - addr; | |
4896 | } | |
4897 | ||
4898 | static u16 | |
4899 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4900 | struct pt_regs *regs) | |
4901 | { | |
4902 | u64 task_size; | |
4903 | ||
4904 | /* No regs, no stack pointer, no dump. */ | |
4905 | if (!regs) | |
4906 | return 0; | |
4907 | ||
4908 | /* | |
4909 | * Check if we fit in with the requested stack size into the: | |
4910 | * - TASK_SIZE | |
4911 | * If we don't, we limit the size to the TASK_SIZE. | |
4912 | * | |
4913 | * - remaining sample size | |
4914 | * If we don't, we customize the stack size to | |
4915 | * fit in to the remaining sample size. | |
4916 | */ | |
4917 | ||
4918 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4919 | stack_size = min(stack_size, (u16) task_size); | |
4920 | ||
4921 | /* Current header size plus static size and dynamic size. */ | |
4922 | header_size += 2 * sizeof(u64); | |
4923 | ||
4924 | /* Do we fit in with the current stack dump size? */ | |
4925 | if ((u16) (header_size + stack_size) < header_size) { | |
4926 | /* | |
4927 | * If we overflow the maximum size for the sample, | |
4928 | * we customize the stack dump size to fit in. | |
4929 | */ | |
4930 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4931 | stack_size = round_up(stack_size, sizeof(u64)); | |
4932 | } | |
4933 | ||
4934 | return stack_size; | |
4935 | } | |
4936 | ||
4937 | static void | |
4938 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4939 | struct pt_regs *regs) | |
4940 | { | |
4941 | /* Case of a kernel thread, nothing to dump */ | |
4942 | if (!regs) { | |
4943 | u64 size = 0; | |
4944 | perf_output_put(handle, size); | |
4945 | } else { | |
4946 | unsigned long sp; | |
4947 | unsigned int rem; | |
4948 | u64 dyn_size; | |
4949 | ||
4950 | /* | |
4951 | * We dump: | |
4952 | * static size | |
4953 | * - the size requested by user or the best one we can fit | |
4954 | * in to the sample max size | |
4955 | * data | |
4956 | * - user stack dump data | |
4957 | * dynamic size | |
4958 | * - the actual dumped size | |
4959 | */ | |
4960 | ||
4961 | /* Static size. */ | |
4962 | perf_output_put(handle, dump_size); | |
4963 | ||
4964 | /* Data. */ | |
4965 | sp = perf_user_stack_pointer(regs); | |
4966 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4967 | dyn_size = dump_size - rem; | |
4968 | ||
4969 | perf_output_skip(handle, rem); | |
4970 | ||
4971 | /* Dynamic size. */ | |
4972 | perf_output_put(handle, dyn_size); | |
4973 | } | |
4974 | } | |
4975 | ||
c980d109 ACM |
4976 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4977 | struct perf_sample_data *data, | |
4978 | struct perf_event *event) | |
6844c09d ACM |
4979 | { |
4980 | u64 sample_type = event->attr.sample_type; | |
4981 | ||
4982 | data->type = sample_type; | |
4983 | header->size += event->id_header_size; | |
4984 | ||
4985 | if (sample_type & PERF_SAMPLE_TID) { | |
4986 | /* namespace issues */ | |
4987 | data->tid_entry.pid = perf_event_pid(event, current); | |
4988 | data->tid_entry.tid = perf_event_tid(event, current); | |
4989 | } | |
4990 | ||
4991 | if (sample_type & PERF_SAMPLE_TIME) | |
34f43927 | 4992 | data->time = perf_event_clock(event); |
6844c09d | 4993 | |
ff3d527c | 4994 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4995 | data->id = primary_event_id(event); |
4996 | ||
4997 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4998 | data->stream_id = event->id; | |
4999 | ||
5000 | if (sample_type & PERF_SAMPLE_CPU) { | |
5001 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
5002 | data->cpu_entry.reserved = 0; | |
5003 | } | |
5004 | } | |
5005 | ||
76369139 FW |
5006 | void perf_event_header__init_id(struct perf_event_header *header, |
5007 | struct perf_sample_data *data, | |
5008 | struct perf_event *event) | |
c980d109 ACM |
5009 | { |
5010 | if (event->attr.sample_id_all) | |
5011 | __perf_event_header__init_id(header, data, event); | |
5012 | } | |
5013 | ||
5014 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
5015 | struct perf_sample_data *data) | |
5016 | { | |
5017 | u64 sample_type = data->type; | |
5018 | ||
5019 | if (sample_type & PERF_SAMPLE_TID) | |
5020 | perf_output_put(handle, data->tid_entry); | |
5021 | ||
5022 | if (sample_type & PERF_SAMPLE_TIME) | |
5023 | perf_output_put(handle, data->time); | |
5024 | ||
5025 | if (sample_type & PERF_SAMPLE_ID) | |
5026 | perf_output_put(handle, data->id); | |
5027 | ||
5028 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
5029 | perf_output_put(handle, data->stream_id); | |
5030 | ||
5031 | if (sample_type & PERF_SAMPLE_CPU) | |
5032 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
5033 | |
5034 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
5035 | perf_output_put(handle, data->id); | |
c980d109 ACM |
5036 | } |
5037 | ||
76369139 FW |
5038 | void perf_event__output_id_sample(struct perf_event *event, |
5039 | struct perf_output_handle *handle, | |
5040 | struct perf_sample_data *sample) | |
c980d109 ACM |
5041 | { |
5042 | if (event->attr.sample_id_all) | |
5043 | __perf_event__output_id_sample(handle, sample); | |
5044 | } | |
5045 | ||
3dab77fb | 5046 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
5047 | struct perf_event *event, |
5048 | u64 enabled, u64 running) | |
3dab77fb | 5049 | { |
cdd6c482 | 5050 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
5051 | u64 values[4]; |
5052 | int n = 0; | |
5053 | ||
b5e58793 | 5054 | values[n++] = perf_event_count(event); |
3dab77fb | 5055 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 5056 | values[n++] = enabled + |
cdd6c482 | 5057 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
5058 | } |
5059 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 5060 | values[n++] = running + |
cdd6c482 | 5061 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
5062 | } |
5063 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 5064 | values[n++] = primary_event_id(event); |
3dab77fb | 5065 | |
76369139 | 5066 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
5067 | } |
5068 | ||
5069 | /* | |
cdd6c482 | 5070 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
5071 | */ |
5072 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
5073 | struct perf_event *event, |
5074 | u64 enabled, u64 running) | |
3dab77fb | 5075 | { |
cdd6c482 IM |
5076 | struct perf_event *leader = event->group_leader, *sub; |
5077 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
5078 | u64 values[5]; |
5079 | int n = 0; | |
5080 | ||
5081 | values[n++] = 1 + leader->nr_siblings; | |
5082 | ||
5083 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 5084 | values[n++] = enabled; |
3dab77fb PZ |
5085 | |
5086 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 5087 | values[n++] = running; |
3dab77fb | 5088 | |
cdd6c482 | 5089 | if (leader != event) |
3dab77fb PZ |
5090 | leader->pmu->read(leader); |
5091 | ||
b5e58793 | 5092 | values[n++] = perf_event_count(leader); |
3dab77fb | 5093 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 5094 | values[n++] = primary_event_id(leader); |
3dab77fb | 5095 | |
76369139 | 5096 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 5097 | |
65abc865 | 5098 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
5099 | n = 0; |
5100 | ||
6f5ab001 JO |
5101 | if ((sub != event) && |
5102 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
5103 | sub->pmu->read(sub); |
5104 | ||
b5e58793 | 5105 | values[n++] = perf_event_count(sub); |
3dab77fb | 5106 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 5107 | values[n++] = primary_event_id(sub); |
3dab77fb | 5108 | |
76369139 | 5109 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
5110 | } |
5111 | } | |
5112 | ||
eed01528 SE |
5113 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
5114 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
5115 | ||
3dab77fb | 5116 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 5117 | struct perf_event *event) |
3dab77fb | 5118 | { |
e3f3541c | 5119 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
5120 | u64 read_format = event->attr.read_format; |
5121 | ||
5122 | /* | |
5123 | * compute total_time_enabled, total_time_running | |
5124 | * based on snapshot values taken when the event | |
5125 | * was last scheduled in. | |
5126 | * | |
5127 | * we cannot simply called update_context_time() | |
5128 | * because of locking issue as we are called in | |
5129 | * NMI context | |
5130 | */ | |
c4794295 | 5131 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 5132 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 5133 | |
cdd6c482 | 5134 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 5135 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 5136 | else |
eed01528 | 5137 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
5138 | } |
5139 | ||
5622f295 MM |
5140 | void perf_output_sample(struct perf_output_handle *handle, |
5141 | struct perf_event_header *header, | |
5142 | struct perf_sample_data *data, | |
cdd6c482 | 5143 | struct perf_event *event) |
5622f295 MM |
5144 | { |
5145 | u64 sample_type = data->type; | |
5146 | ||
5147 | perf_output_put(handle, *header); | |
5148 | ||
ff3d527c AH |
5149 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
5150 | perf_output_put(handle, data->id); | |
5151 | ||
5622f295 MM |
5152 | if (sample_type & PERF_SAMPLE_IP) |
5153 | perf_output_put(handle, data->ip); | |
5154 | ||
5155 | if (sample_type & PERF_SAMPLE_TID) | |
5156 | perf_output_put(handle, data->tid_entry); | |
5157 | ||
5158 | if (sample_type & PERF_SAMPLE_TIME) | |
5159 | perf_output_put(handle, data->time); | |
5160 | ||
5161 | if (sample_type & PERF_SAMPLE_ADDR) | |
5162 | perf_output_put(handle, data->addr); | |
5163 | ||
5164 | if (sample_type & PERF_SAMPLE_ID) | |
5165 | perf_output_put(handle, data->id); | |
5166 | ||
5167 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
5168 | perf_output_put(handle, data->stream_id); | |
5169 | ||
5170 | if (sample_type & PERF_SAMPLE_CPU) | |
5171 | perf_output_put(handle, data->cpu_entry); | |
5172 | ||
5173 | if (sample_type & PERF_SAMPLE_PERIOD) | |
5174 | perf_output_put(handle, data->period); | |
5175 | ||
5176 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 5177 | perf_output_read(handle, event); |
5622f295 MM |
5178 | |
5179 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
5180 | if (data->callchain) { | |
5181 | int size = 1; | |
5182 | ||
5183 | if (data->callchain) | |
5184 | size += data->callchain->nr; | |
5185 | ||
5186 | size *= sizeof(u64); | |
5187 | ||
76369139 | 5188 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
5189 | } else { |
5190 | u64 nr = 0; | |
5191 | perf_output_put(handle, nr); | |
5192 | } | |
5193 | } | |
5194 | ||
5195 | if (sample_type & PERF_SAMPLE_RAW) { | |
5196 | if (data->raw) { | |
5197 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
5198 | __output_copy(handle, data->raw->data, |
5199 | data->raw->size); | |
5622f295 MM |
5200 | } else { |
5201 | struct { | |
5202 | u32 size; | |
5203 | u32 data; | |
5204 | } raw = { | |
5205 | .size = sizeof(u32), | |
5206 | .data = 0, | |
5207 | }; | |
5208 | perf_output_put(handle, raw); | |
5209 | } | |
5210 | } | |
a7ac67ea | 5211 | |
bce38cd5 SE |
5212 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
5213 | if (data->br_stack) { | |
5214 | size_t size; | |
5215 | ||
5216 | size = data->br_stack->nr | |
5217 | * sizeof(struct perf_branch_entry); | |
5218 | ||
5219 | perf_output_put(handle, data->br_stack->nr); | |
5220 | perf_output_copy(handle, data->br_stack->entries, size); | |
5221 | } else { | |
5222 | /* | |
5223 | * we always store at least the value of nr | |
5224 | */ | |
5225 | u64 nr = 0; | |
5226 | perf_output_put(handle, nr); | |
5227 | } | |
5228 | } | |
4018994f JO |
5229 | |
5230 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
5231 | u64 abi = data->regs_user.abi; | |
5232 | ||
5233 | /* | |
5234 | * If there are no regs to dump, notice it through | |
5235 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5236 | */ | |
5237 | perf_output_put(handle, abi); | |
5238 | ||
5239 | if (abi) { | |
5240 | u64 mask = event->attr.sample_regs_user; | |
5241 | perf_output_sample_regs(handle, | |
5242 | data->regs_user.regs, | |
5243 | mask); | |
5244 | } | |
5245 | } | |
c5ebcedb | 5246 | |
a5cdd40c | 5247 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
5248 | perf_output_sample_ustack(handle, |
5249 | data->stack_user_size, | |
5250 | data->regs_user.regs); | |
a5cdd40c | 5251 | } |
c3feedf2 AK |
5252 | |
5253 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
5254 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
5255 | |
5256 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
5257 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 5258 | |
fdfbbd07 AK |
5259 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
5260 | perf_output_put(handle, data->txn); | |
5261 | ||
60e2364e SE |
5262 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
5263 | u64 abi = data->regs_intr.abi; | |
5264 | /* | |
5265 | * If there are no regs to dump, notice it through | |
5266 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5267 | */ | |
5268 | perf_output_put(handle, abi); | |
5269 | ||
5270 | if (abi) { | |
5271 | u64 mask = event->attr.sample_regs_intr; | |
5272 | ||
5273 | perf_output_sample_regs(handle, | |
5274 | data->regs_intr.regs, | |
5275 | mask); | |
5276 | } | |
5277 | } | |
5278 | ||
a5cdd40c PZ |
5279 | if (!event->attr.watermark) { |
5280 | int wakeup_events = event->attr.wakeup_events; | |
5281 | ||
5282 | if (wakeup_events) { | |
5283 | struct ring_buffer *rb = handle->rb; | |
5284 | int events = local_inc_return(&rb->events); | |
5285 | ||
5286 | if (events >= wakeup_events) { | |
5287 | local_sub(wakeup_events, &rb->events); | |
5288 | local_inc(&rb->wakeup); | |
5289 | } | |
5290 | } | |
5291 | } | |
5622f295 MM |
5292 | } |
5293 | ||
5294 | void perf_prepare_sample(struct perf_event_header *header, | |
5295 | struct perf_sample_data *data, | |
cdd6c482 | 5296 | struct perf_event *event, |
5622f295 | 5297 | struct pt_regs *regs) |
7b732a75 | 5298 | { |
cdd6c482 | 5299 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 5300 | |
cdd6c482 | 5301 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 5302 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
5303 | |
5304 | header->misc = 0; | |
5305 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 5306 | |
c980d109 | 5307 | __perf_event_header__init_id(header, data, event); |
6844c09d | 5308 | |
c320c7b7 | 5309 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
5310 | data->ip = perf_instruction_pointer(regs); |
5311 | ||
b23f3325 | 5312 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 5313 | int size = 1; |
394ee076 | 5314 | |
e6dab5ff | 5315 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
5316 | |
5317 | if (data->callchain) | |
5318 | size += data->callchain->nr; | |
5319 | ||
5320 | header->size += size * sizeof(u64); | |
394ee076 PZ |
5321 | } |
5322 | ||
3a43ce68 | 5323 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
5324 | int size = sizeof(u32); |
5325 | ||
5326 | if (data->raw) | |
5327 | size += data->raw->size; | |
5328 | else | |
5329 | size += sizeof(u32); | |
5330 | ||
5331 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 5332 | header->size += size; |
7f453c24 | 5333 | } |
bce38cd5 SE |
5334 | |
5335 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
5336 | int size = sizeof(u64); /* nr */ | |
5337 | if (data->br_stack) { | |
5338 | size += data->br_stack->nr | |
5339 | * sizeof(struct perf_branch_entry); | |
5340 | } | |
5341 | header->size += size; | |
5342 | } | |
4018994f | 5343 | |
2565711f | 5344 | if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER)) |
88a7c26a AL |
5345 | perf_sample_regs_user(&data->regs_user, regs, |
5346 | &data->regs_user_copy); | |
2565711f | 5347 | |
4018994f JO |
5348 | if (sample_type & PERF_SAMPLE_REGS_USER) { |
5349 | /* regs dump ABI info */ | |
5350 | int size = sizeof(u64); | |
5351 | ||
4018994f JO |
5352 | if (data->regs_user.regs) { |
5353 | u64 mask = event->attr.sample_regs_user; | |
5354 | size += hweight64(mask) * sizeof(u64); | |
5355 | } | |
5356 | ||
5357 | header->size += size; | |
5358 | } | |
c5ebcedb JO |
5359 | |
5360 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
5361 | /* | |
5362 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
5363 | * processed as the last one or have additional check added | |
5364 | * in case new sample type is added, because we could eat | |
5365 | * up the rest of the sample size. | |
5366 | */ | |
c5ebcedb JO |
5367 | u16 stack_size = event->attr.sample_stack_user; |
5368 | u16 size = sizeof(u64); | |
5369 | ||
c5ebcedb | 5370 | stack_size = perf_sample_ustack_size(stack_size, header->size, |
2565711f | 5371 | data->regs_user.regs); |
c5ebcedb JO |
5372 | |
5373 | /* | |
5374 | * If there is something to dump, add space for the dump | |
5375 | * itself and for the field that tells the dynamic size, | |
5376 | * which is how many have been actually dumped. | |
5377 | */ | |
5378 | if (stack_size) | |
5379 | size += sizeof(u64) + stack_size; | |
5380 | ||
5381 | data->stack_user_size = stack_size; | |
5382 | header->size += size; | |
5383 | } | |
60e2364e SE |
5384 | |
5385 | if (sample_type & PERF_SAMPLE_REGS_INTR) { | |
5386 | /* regs dump ABI info */ | |
5387 | int size = sizeof(u64); | |
5388 | ||
5389 | perf_sample_regs_intr(&data->regs_intr, regs); | |
5390 | ||
5391 | if (data->regs_intr.regs) { | |
5392 | u64 mask = event->attr.sample_regs_intr; | |
5393 | ||
5394 | size += hweight64(mask) * sizeof(u64); | |
5395 | } | |
5396 | ||
5397 | header->size += size; | |
5398 | } | |
5622f295 | 5399 | } |
7f453c24 | 5400 | |
21509084 YZ |
5401 | void perf_event_output(struct perf_event *event, |
5402 | struct perf_sample_data *data, | |
5403 | struct pt_regs *regs) | |
5622f295 MM |
5404 | { |
5405 | struct perf_output_handle handle; | |
5406 | struct perf_event_header header; | |
689802b2 | 5407 | |
927c7a9e FW |
5408 | /* protect the callchain buffers */ |
5409 | rcu_read_lock(); | |
5410 | ||
cdd6c482 | 5411 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 5412 | |
a7ac67ea | 5413 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 5414 | goto exit; |
0322cd6e | 5415 | |
cdd6c482 | 5416 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 5417 | |
8a057d84 | 5418 | perf_output_end(&handle); |
927c7a9e FW |
5419 | |
5420 | exit: | |
5421 | rcu_read_unlock(); | |
0322cd6e PZ |
5422 | } |
5423 | ||
38b200d6 | 5424 | /* |
cdd6c482 | 5425 | * read event_id |
38b200d6 PZ |
5426 | */ |
5427 | ||
5428 | struct perf_read_event { | |
5429 | struct perf_event_header header; | |
5430 | ||
5431 | u32 pid; | |
5432 | u32 tid; | |
38b200d6 PZ |
5433 | }; |
5434 | ||
5435 | static void | |
cdd6c482 | 5436 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
5437 | struct task_struct *task) |
5438 | { | |
5439 | struct perf_output_handle handle; | |
c980d109 | 5440 | struct perf_sample_data sample; |
dfc65094 | 5441 | struct perf_read_event read_event = { |
38b200d6 | 5442 | .header = { |
cdd6c482 | 5443 | .type = PERF_RECORD_READ, |
38b200d6 | 5444 | .misc = 0, |
c320c7b7 | 5445 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 5446 | }, |
cdd6c482 IM |
5447 | .pid = perf_event_pid(event, task), |
5448 | .tid = perf_event_tid(event, task), | |
38b200d6 | 5449 | }; |
3dab77fb | 5450 | int ret; |
38b200d6 | 5451 | |
c980d109 | 5452 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 5453 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
5454 | if (ret) |
5455 | return; | |
5456 | ||
dfc65094 | 5457 | perf_output_put(&handle, read_event); |
cdd6c482 | 5458 | perf_output_read(&handle, event); |
c980d109 | 5459 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 5460 | |
38b200d6 PZ |
5461 | perf_output_end(&handle); |
5462 | } | |
5463 | ||
52d857a8 JO |
5464 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
5465 | ||
5466 | static void | |
5467 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
5468 | perf_event_aux_output_cb output, |
5469 | void *data) | |
5470 | { | |
5471 | struct perf_event *event; | |
5472 | ||
5473 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5474 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
5475 | continue; | |
5476 | if (!event_filter_match(event)) | |
5477 | continue; | |
67516844 | 5478 | output(event, data); |
52d857a8 JO |
5479 | } |
5480 | } | |
5481 | ||
5482 | static void | |
67516844 | 5483 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
5484 | struct perf_event_context *task_ctx) |
5485 | { | |
5486 | struct perf_cpu_context *cpuctx; | |
5487 | struct perf_event_context *ctx; | |
5488 | struct pmu *pmu; | |
5489 | int ctxn; | |
5490 | ||
5491 | rcu_read_lock(); | |
5492 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
5493 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
5494 | if (cpuctx->unique_pmu != pmu) | |
5495 | goto next; | |
67516844 | 5496 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
5497 | if (task_ctx) |
5498 | goto next; | |
5499 | ctxn = pmu->task_ctx_nr; | |
5500 | if (ctxn < 0) | |
5501 | goto next; | |
5502 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
5503 | if (ctx) | |
67516844 | 5504 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
5505 | next: |
5506 | put_cpu_ptr(pmu->pmu_cpu_context); | |
5507 | } | |
5508 | ||
5509 | if (task_ctx) { | |
5510 | preempt_disable(); | |
67516844 | 5511 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
5512 | preempt_enable(); |
5513 | } | |
5514 | rcu_read_unlock(); | |
5515 | } | |
5516 | ||
60313ebe | 5517 | /* |
9f498cc5 PZ |
5518 | * task tracking -- fork/exit |
5519 | * | |
13d7a241 | 5520 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
5521 | */ |
5522 | ||
9f498cc5 | 5523 | struct perf_task_event { |
3a80b4a3 | 5524 | struct task_struct *task; |
cdd6c482 | 5525 | struct perf_event_context *task_ctx; |
60313ebe PZ |
5526 | |
5527 | struct { | |
5528 | struct perf_event_header header; | |
5529 | ||
5530 | u32 pid; | |
5531 | u32 ppid; | |
9f498cc5 PZ |
5532 | u32 tid; |
5533 | u32 ptid; | |
393b2ad8 | 5534 | u64 time; |
cdd6c482 | 5535 | } event_id; |
60313ebe PZ |
5536 | }; |
5537 | ||
67516844 JO |
5538 | static int perf_event_task_match(struct perf_event *event) |
5539 | { | |
13d7a241 SE |
5540 | return event->attr.comm || event->attr.mmap || |
5541 | event->attr.mmap2 || event->attr.mmap_data || | |
5542 | event->attr.task; | |
67516844 JO |
5543 | } |
5544 | ||
cdd6c482 | 5545 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 5546 | void *data) |
60313ebe | 5547 | { |
52d857a8 | 5548 | struct perf_task_event *task_event = data; |
60313ebe | 5549 | struct perf_output_handle handle; |
c980d109 | 5550 | struct perf_sample_data sample; |
9f498cc5 | 5551 | struct task_struct *task = task_event->task; |
c980d109 | 5552 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 5553 | |
67516844 JO |
5554 | if (!perf_event_task_match(event)) |
5555 | return; | |
5556 | ||
c980d109 | 5557 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 5558 | |
c980d109 | 5559 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 5560 | task_event->event_id.header.size); |
ef60777c | 5561 | if (ret) |
c980d109 | 5562 | goto out; |
60313ebe | 5563 | |
cdd6c482 IM |
5564 | task_event->event_id.pid = perf_event_pid(event, task); |
5565 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 5566 | |
cdd6c482 IM |
5567 | task_event->event_id.tid = perf_event_tid(event, task); |
5568 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 5569 | |
34f43927 PZ |
5570 | task_event->event_id.time = perf_event_clock(event); |
5571 | ||
cdd6c482 | 5572 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 5573 | |
c980d109 ACM |
5574 | perf_event__output_id_sample(event, &handle, &sample); |
5575 | ||
60313ebe | 5576 | perf_output_end(&handle); |
c980d109 ACM |
5577 | out: |
5578 | task_event->event_id.header.size = size; | |
60313ebe PZ |
5579 | } |
5580 | ||
cdd6c482 IM |
5581 | static void perf_event_task(struct task_struct *task, |
5582 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 5583 | int new) |
60313ebe | 5584 | { |
9f498cc5 | 5585 | struct perf_task_event task_event; |
60313ebe | 5586 | |
cdd6c482 IM |
5587 | if (!atomic_read(&nr_comm_events) && |
5588 | !atomic_read(&nr_mmap_events) && | |
5589 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
5590 | return; |
5591 | ||
9f498cc5 | 5592 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
5593 | .task = task, |
5594 | .task_ctx = task_ctx, | |
cdd6c482 | 5595 | .event_id = { |
60313ebe | 5596 | .header = { |
cdd6c482 | 5597 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 5598 | .misc = 0, |
cdd6c482 | 5599 | .size = sizeof(task_event.event_id), |
60313ebe | 5600 | }, |
573402db PZ |
5601 | /* .pid */ |
5602 | /* .ppid */ | |
9f498cc5 PZ |
5603 | /* .tid */ |
5604 | /* .ptid */ | |
34f43927 | 5605 | /* .time */ |
60313ebe PZ |
5606 | }, |
5607 | }; | |
5608 | ||
67516844 | 5609 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5610 | &task_event, |
5611 | task_ctx); | |
9f498cc5 PZ |
5612 | } |
5613 | ||
cdd6c482 | 5614 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5615 | { |
cdd6c482 | 5616 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5617 | } |
5618 | ||
8d1b2d93 PZ |
5619 | /* |
5620 | * comm tracking | |
5621 | */ | |
5622 | ||
5623 | struct perf_comm_event { | |
22a4f650 IM |
5624 | struct task_struct *task; |
5625 | char *comm; | |
8d1b2d93 PZ |
5626 | int comm_size; |
5627 | ||
5628 | struct { | |
5629 | struct perf_event_header header; | |
5630 | ||
5631 | u32 pid; | |
5632 | u32 tid; | |
cdd6c482 | 5633 | } event_id; |
8d1b2d93 PZ |
5634 | }; |
5635 | ||
67516844 JO |
5636 | static int perf_event_comm_match(struct perf_event *event) |
5637 | { | |
5638 | return event->attr.comm; | |
5639 | } | |
5640 | ||
cdd6c482 | 5641 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5642 | void *data) |
8d1b2d93 | 5643 | { |
52d857a8 | 5644 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5645 | struct perf_output_handle handle; |
c980d109 | 5646 | struct perf_sample_data sample; |
cdd6c482 | 5647 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5648 | int ret; |
5649 | ||
67516844 JO |
5650 | if (!perf_event_comm_match(event)) |
5651 | return; | |
5652 | ||
c980d109 ACM |
5653 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5654 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5655 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5656 | |
5657 | if (ret) | |
c980d109 | 5658 | goto out; |
8d1b2d93 | 5659 | |
cdd6c482 IM |
5660 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5661 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5662 | |
cdd6c482 | 5663 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5664 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5665 | comm_event->comm_size); |
c980d109 ACM |
5666 | |
5667 | perf_event__output_id_sample(event, &handle, &sample); | |
5668 | ||
8d1b2d93 | 5669 | perf_output_end(&handle); |
c980d109 ACM |
5670 | out: |
5671 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5672 | } |
5673 | ||
cdd6c482 | 5674 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5675 | { |
413ee3b4 | 5676 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5677 | unsigned int size; |
8d1b2d93 | 5678 | |
413ee3b4 | 5679 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5680 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5681 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5682 | |
5683 | comm_event->comm = comm; | |
5684 | comm_event->comm_size = size; | |
5685 | ||
cdd6c482 | 5686 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5687 | |
67516844 | 5688 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5689 | comm_event, |
5690 | NULL); | |
8d1b2d93 PZ |
5691 | } |
5692 | ||
82b89778 | 5693 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5694 | { |
9ee318a7 PZ |
5695 | struct perf_comm_event comm_event; |
5696 | ||
cdd6c482 | 5697 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5698 | return; |
a63eaf34 | 5699 | |
9ee318a7 | 5700 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5701 | .task = task, |
573402db PZ |
5702 | /* .comm */ |
5703 | /* .comm_size */ | |
cdd6c482 | 5704 | .event_id = { |
573402db | 5705 | .header = { |
cdd6c482 | 5706 | .type = PERF_RECORD_COMM, |
82b89778 | 5707 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5708 | /* .size */ |
5709 | }, | |
5710 | /* .pid */ | |
5711 | /* .tid */ | |
8d1b2d93 PZ |
5712 | }, |
5713 | }; | |
5714 | ||
cdd6c482 | 5715 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5716 | } |
5717 | ||
0a4a9391 PZ |
5718 | /* |
5719 | * mmap tracking | |
5720 | */ | |
5721 | ||
5722 | struct perf_mmap_event { | |
089dd79d PZ |
5723 | struct vm_area_struct *vma; |
5724 | ||
5725 | const char *file_name; | |
5726 | int file_size; | |
13d7a241 SE |
5727 | int maj, min; |
5728 | u64 ino; | |
5729 | u64 ino_generation; | |
f972eb63 | 5730 | u32 prot, flags; |
0a4a9391 PZ |
5731 | |
5732 | struct { | |
5733 | struct perf_event_header header; | |
5734 | ||
5735 | u32 pid; | |
5736 | u32 tid; | |
5737 | u64 start; | |
5738 | u64 len; | |
5739 | u64 pgoff; | |
cdd6c482 | 5740 | } event_id; |
0a4a9391 PZ |
5741 | }; |
5742 | ||
67516844 JO |
5743 | static int perf_event_mmap_match(struct perf_event *event, |
5744 | void *data) | |
5745 | { | |
5746 | struct perf_mmap_event *mmap_event = data; | |
5747 | struct vm_area_struct *vma = mmap_event->vma; | |
5748 | int executable = vma->vm_flags & VM_EXEC; | |
5749 | ||
5750 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5751 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5752 | } |
5753 | ||
cdd6c482 | 5754 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5755 | void *data) |
0a4a9391 | 5756 | { |
52d857a8 | 5757 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5758 | struct perf_output_handle handle; |
c980d109 | 5759 | struct perf_sample_data sample; |
cdd6c482 | 5760 | int size = mmap_event->event_id.header.size; |
c980d109 | 5761 | int ret; |
0a4a9391 | 5762 | |
67516844 JO |
5763 | if (!perf_event_mmap_match(event, data)) |
5764 | return; | |
5765 | ||
13d7a241 SE |
5766 | if (event->attr.mmap2) { |
5767 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5768 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5769 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5770 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5771 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
f972eb63 PZ |
5772 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
5773 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); | |
13d7a241 SE |
5774 | } |
5775 | ||
c980d109 ACM |
5776 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5777 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5778 | mmap_event->event_id.header.size); |
0a4a9391 | 5779 | if (ret) |
c980d109 | 5780 | goto out; |
0a4a9391 | 5781 | |
cdd6c482 IM |
5782 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5783 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5784 | |
cdd6c482 | 5785 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5786 | |
5787 | if (event->attr.mmap2) { | |
5788 | perf_output_put(&handle, mmap_event->maj); | |
5789 | perf_output_put(&handle, mmap_event->min); | |
5790 | perf_output_put(&handle, mmap_event->ino); | |
5791 | perf_output_put(&handle, mmap_event->ino_generation); | |
f972eb63 PZ |
5792 | perf_output_put(&handle, mmap_event->prot); |
5793 | perf_output_put(&handle, mmap_event->flags); | |
13d7a241 SE |
5794 | } |
5795 | ||
76369139 | 5796 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5797 | mmap_event->file_size); |
c980d109 ACM |
5798 | |
5799 | perf_event__output_id_sample(event, &handle, &sample); | |
5800 | ||
78d613eb | 5801 | perf_output_end(&handle); |
c980d109 ACM |
5802 | out: |
5803 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5804 | } |
5805 | ||
cdd6c482 | 5806 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5807 | { |
089dd79d PZ |
5808 | struct vm_area_struct *vma = mmap_event->vma; |
5809 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5810 | int maj = 0, min = 0; |
5811 | u64 ino = 0, gen = 0; | |
f972eb63 | 5812 | u32 prot = 0, flags = 0; |
0a4a9391 PZ |
5813 | unsigned int size; |
5814 | char tmp[16]; | |
5815 | char *buf = NULL; | |
2c42cfbf | 5816 | char *name; |
413ee3b4 | 5817 | |
0a4a9391 | 5818 | if (file) { |
13d7a241 SE |
5819 | struct inode *inode; |
5820 | dev_t dev; | |
3ea2f2b9 | 5821 | |
2c42cfbf | 5822 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5823 | if (!buf) { |
c7e548b4 ON |
5824 | name = "//enomem"; |
5825 | goto cpy_name; | |
0a4a9391 | 5826 | } |
413ee3b4 | 5827 | /* |
3ea2f2b9 | 5828 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5829 | * need to add enough zero bytes after the string to handle |
5830 | * the 64bit alignment we do later. | |
5831 | */ | |
9bf39ab2 | 5832 | name = file_path(file, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5833 | if (IS_ERR(name)) { |
c7e548b4 ON |
5834 | name = "//toolong"; |
5835 | goto cpy_name; | |
0a4a9391 | 5836 | } |
13d7a241 SE |
5837 | inode = file_inode(vma->vm_file); |
5838 | dev = inode->i_sb->s_dev; | |
5839 | ino = inode->i_ino; | |
5840 | gen = inode->i_generation; | |
5841 | maj = MAJOR(dev); | |
5842 | min = MINOR(dev); | |
f972eb63 PZ |
5843 | |
5844 | if (vma->vm_flags & VM_READ) | |
5845 | prot |= PROT_READ; | |
5846 | if (vma->vm_flags & VM_WRITE) | |
5847 | prot |= PROT_WRITE; | |
5848 | if (vma->vm_flags & VM_EXEC) | |
5849 | prot |= PROT_EXEC; | |
5850 | ||
5851 | if (vma->vm_flags & VM_MAYSHARE) | |
5852 | flags = MAP_SHARED; | |
5853 | else | |
5854 | flags = MAP_PRIVATE; | |
5855 | ||
5856 | if (vma->vm_flags & VM_DENYWRITE) | |
5857 | flags |= MAP_DENYWRITE; | |
5858 | if (vma->vm_flags & VM_MAYEXEC) | |
5859 | flags |= MAP_EXECUTABLE; | |
5860 | if (vma->vm_flags & VM_LOCKED) | |
5861 | flags |= MAP_LOCKED; | |
5862 | if (vma->vm_flags & VM_HUGETLB) | |
5863 | flags |= MAP_HUGETLB; | |
5864 | ||
c7e548b4 | 5865 | goto got_name; |
0a4a9391 | 5866 | } else { |
fbe26abe JO |
5867 | if (vma->vm_ops && vma->vm_ops->name) { |
5868 | name = (char *) vma->vm_ops->name(vma); | |
5869 | if (name) | |
5870 | goto cpy_name; | |
5871 | } | |
5872 | ||
2c42cfbf | 5873 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5874 | if (name) |
5875 | goto cpy_name; | |
089dd79d | 5876 | |
32c5fb7e | 5877 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5878 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5879 | name = "[heap]"; |
5880 | goto cpy_name; | |
32c5fb7e ON |
5881 | } |
5882 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5883 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5884 | name = "[stack]"; |
5885 | goto cpy_name; | |
089dd79d PZ |
5886 | } |
5887 | ||
c7e548b4 ON |
5888 | name = "//anon"; |
5889 | goto cpy_name; | |
0a4a9391 PZ |
5890 | } |
5891 | ||
c7e548b4 ON |
5892 | cpy_name: |
5893 | strlcpy(tmp, name, sizeof(tmp)); | |
5894 | name = tmp; | |
0a4a9391 | 5895 | got_name: |
2c42cfbf PZ |
5896 | /* |
5897 | * Since our buffer works in 8 byte units we need to align our string | |
5898 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5899 | * zero'd out to avoid leaking random bits to userspace. | |
5900 | */ | |
5901 | size = strlen(name)+1; | |
5902 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5903 | name[size++] = '\0'; | |
0a4a9391 PZ |
5904 | |
5905 | mmap_event->file_name = name; | |
5906 | mmap_event->file_size = size; | |
13d7a241 SE |
5907 | mmap_event->maj = maj; |
5908 | mmap_event->min = min; | |
5909 | mmap_event->ino = ino; | |
5910 | mmap_event->ino_generation = gen; | |
f972eb63 PZ |
5911 | mmap_event->prot = prot; |
5912 | mmap_event->flags = flags; | |
0a4a9391 | 5913 | |
2fe85427 SE |
5914 | if (!(vma->vm_flags & VM_EXEC)) |
5915 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5916 | ||
cdd6c482 | 5917 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5918 | |
67516844 | 5919 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5920 | mmap_event, |
5921 | NULL); | |
665c2142 | 5922 | |
0a4a9391 PZ |
5923 | kfree(buf); |
5924 | } | |
5925 | ||
3af9e859 | 5926 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5927 | { |
9ee318a7 PZ |
5928 | struct perf_mmap_event mmap_event; |
5929 | ||
cdd6c482 | 5930 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5931 | return; |
5932 | ||
5933 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5934 | .vma = vma, |
573402db PZ |
5935 | /* .file_name */ |
5936 | /* .file_size */ | |
cdd6c482 | 5937 | .event_id = { |
573402db | 5938 | .header = { |
cdd6c482 | 5939 | .type = PERF_RECORD_MMAP, |
39447b38 | 5940 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5941 | /* .size */ |
5942 | }, | |
5943 | /* .pid */ | |
5944 | /* .tid */ | |
089dd79d PZ |
5945 | .start = vma->vm_start, |
5946 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5947 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5948 | }, |
13d7a241 SE |
5949 | /* .maj (attr_mmap2 only) */ |
5950 | /* .min (attr_mmap2 only) */ | |
5951 | /* .ino (attr_mmap2 only) */ | |
5952 | /* .ino_generation (attr_mmap2 only) */ | |
f972eb63 PZ |
5953 | /* .prot (attr_mmap2 only) */ |
5954 | /* .flags (attr_mmap2 only) */ | |
0a4a9391 PZ |
5955 | }; |
5956 | ||
cdd6c482 | 5957 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5958 | } |
5959 | ||
68db7e98 AS |
5960 | void perf_event_aux_event(struct perf_event *event, unsigned long head, |
5961 | unsigned long size, u64 flags) | |
5962 | { | |
5963 | struct perf_output_handle handle; | |
5964 | struct perf_sample_data sample; | |
5965 | struct perf_aux_event { | |
5966 | struct perf_event_header header; | |
5967 | u64 offset; | |
5968 | u64 size; | |
5969 | u64 flags; | |
5970 | } rec = { | |
5971 | .header = { | |
5972 | .type = PERF_RECORD_AUX, | |
5973 | .misc = 0, | |
5974 | .size = sizeof(rec), | |
5975 | }, | |
5976 | .offset = head, | |
5977 | .size = size, | |
5978 | .flags = flags, | |
5979 | }; | |
5980 | int ret; | |
5981 | ||
5982 | perf_event_header__init_id(&rec.header, &sample, event); | |
5983 | ret = perf_output_begin(&handle, event, rec.header.size); | |
5984 | ||
5985 | if (ret) | |
5986 | return; | |
5987 | ||
5988 | perf_output_put(&handle, rec); | |
5989 | perf_event__output_id_sample(event, &handle, &sample); | |
5990 | ||
5991 | perf_output_end(&handle); | |
5992 | } | |
5993 | ||
f38b0dbb KL |
5994 | /* |
5995 | * Lost/dropped samples logging | |
5996 | */ | |
5997 | void perf_log_lost_samples(struct perf_event *event, u64 lost) | |
5998 | { | |
5999 | struct perf_output_handle handle; | |
6000 | struct perf_sample_data sample; | |
6001 | int ret; | |
6002 | ||
6003 | struct { | |
6004 | struct perf_event_header header; | |
6005 | u64 lost; | |
6006 | } lost_samples_event = { | |
6007 | .header = { | |
6008 | .type = PERF_RECORD_LOST_SAMPLES, | |
6009 | .misc = 0, | |
6010 | .size = sizeof(lost_samples_event), | |
6011 | }, | |
6012 | .lost = lost, | |
6013 | }; | |
6014 | ||
6015 | perf_event_header__init_id(&lost_samples_event.header, &sample, event); | |
6016 | ||
6017 | ret = perf_output_begin(&handle, event, | |
6018 | lost_samples_event.header.size); | |
6019 | if (ret) | |
6020 | return; | |
6021 | ||
6022 | perf_output_put(&handle, lost_samples_event); | |
6023 | perf_event__output_id_sample(event, &handle, &sample); | |
6024 | perf_output_end(&handle); | |
6025 | } | |
6026 | ||
a78ac325 PZ |
6027 | /* |
6028 | * IRQ throttle logging | |
6029 | */ | |
6030 | ||
cdd6c482 | 6031 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
6032 | { |
6033 | struct perf_output_handle handle; | |
c980d109 | 6034 | struct perf_sample_data sample; |
a78ac325 PZ |
6035 | int ret; |
6036 | ||
6037 | struct { | |
6038 | struct perf_event_header header; | |
6039 | u64 time; | |
cca3f454 | 6040 | u64 id; |
7f453c24 | 6041 | u64 stream_id; |
a78ac325 PZ |
6042 | } throttle_event = { |
6043 | .header = { | |
cdd6c482 | 6044 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
6045 | .misc = 0, |
6046 | .size = sizeof(throttle_event), | |
6047 | }, | |
34f43927 | 6048 | .time = perf_event_clock(event), |
cdd6c482 IM |
6049 | .id = primary_event_id(event), |
6050 | .stream_id = event->id, | |
a78ac325 PZ |
6051 | }; |
6052 | ||
966ee4d6 | 6053 | if (enable) |
cdd6c482 | 6054 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 6055 | |
c980d109 ACM |
6056 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
6057 | ||
6058 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 6059 | throttle_event.header.size); |
a78ac325 PZ |
6060 | if (ret) |
6061 | return; | |
6062 | ||
6063 | perf_output_put(&handle, throttle_event); | |
c980d109 | 6064 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
6065 | perf_output_end(&handle); |
6066 | } | |
6067 | ||
ec0d7729 AS |
6068 | static void perf_log_itrace_start(struct perf_event *event) |
6069 | { | |
6070 | struct perf_output_handle handle; | |
6071 | struct perf_sample_data sample; | |
6072 | struct perf_aux_event { | |
6073 | struct perf_event_header header; | |
6074 | u32 pid; | |
6075 | u32 tid; | |
6076 | } rec; | |
6077 | int ret; | |
6078 | ||
6079 | if (event->parent) | |
6080 | event = event->parent; | |
6081 | ||
6082 | if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) || | |
6083 | event->hw.itrace_started) | |
6084 | return; | |
6085 | ||
6086 | event->hw.itrace_started = 1; | |
6087 | ||
6088 | rec.header.type = PERF_RECORD_ITRACE_START; | |
6089 | rec.header.misc = 0; | |
6090 | rec.header.size = sizeof(rec); | |
6091 | rec.pid = perf_event_pid(event, current); | |
6092 | rec.tid = perf_event_tid(event, current); | |
6093 | ||
6094 | perf_event_header__init_id(&rec.header, &sample, event); | |
6095 | ret = perf_output_begin(&handle, event, rec.header.size); | |
6096 | ||
6097 | if (ret) | |
6098 | return; | |
6099 | ||
6100 | perf_output_put(&handle, rec); | |
6101 | perf_event__output_id_sample(event, &handle, &sample); | |
6102 | ||
6103 | perf_output_end(&handle); | |
6104 | } | |
6105 | ||
f6c7d5fe | 6106 | /* |
cdd6c482 | 6107 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
6108 | */ |
6109 | ||
a8b0ca17 | 6110 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
6111 | int throttle, struct perf_sample_data *data, |
6112 | struct pt_regs *regs) | |
f6c7d5fe | 6113 | { |
cdd6c482 IM |
6114 | int events = atomic_read(&event->event_limit); |
6115 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 6116 | u64 seq; |
79f14641 PZ |
6117 | int ret = 0; |
6118 | ||
96398826 PZ |
6119 | /* |
6120 | * Non-sampling counters might still use the PMI to fold short | |
6121 | * hardware counters, ignore those. | |
6122 | */ | |
6123 | if (unlikely(!is_sampling_event(event))) | |
6124 | return 0; | |
6125 | ||
e050e3f0 SE |
6126 | seq = __this_cpu_read(perf_throttled_seq); |
6127 | if (seq != hwc->interrupts_seq) { | |
6128 | hwc->interrupts_seq = seq; | |
6129 | hwc->interrupts = 1; | |
6130 | } else { | |
6131 | hwc->interrupts++; | |
6132 | if (unlikely(throttle | |
6133 | && hwc->interrupts >= max_samples_per_tick)) { | |
6134 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
6135 | hwc->interrupts = MAX_INTERRUPTS; |
6136 | perf_log_throttle(event, 0); | |
d84153d6 | 6137 | tick_nohz_full_kick(); |
a78ac325 PZ |
6138 | ret = 1; |
6139 | } | |
e050e3f0 | 6140 | } |
60db5e09 | 6141 | |
cdd6c482 | 6142 | if (event->attr.freq) { |
def0a9b2 | 6143 | u64 now = perf_clock(); |
abd50713 | 6144 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 6145 | |
abd50713 | 6146 | hwc->freq_time_stamp = now; |
bd2b5b12 | 6147 | |
abd50713 | 6148 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 6149 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
6150 | } |
6151 | ||
2023b359 PZ |
6152 | /* |
6153 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 6154 | * events |
2023b359 PZ |
6155 | */ |
6156 | ||
cdd6c482 IM |
6157 | event->pending_kill = POLL_IN; |
6158 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 6159 | ret = 1; |
cdd6c482 | 6160 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
6161 | event->pending_disable = 1; |
6162 | irq_work_queue(&event->pending); | |
79f14641 PZ |
6163 | } |
6164 | ||
453f19ee | 6165 | if (event->overflow_handler) |
a8b0ca17 | 6166 | event->overflow_handler(event, data, regs); |
453f19ee | 6167 | else |
a8b0ca17 | 6168 | perf_event_output(event, data, regs); |
453f19ee | 6169 | |
fed66e2c | 6170 | if (*perf_event_fasync(event) && event->pending_kill) { |
a8b0ca17 PZ |
6171 | event->pending_wakeup = 1; |
6172 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
6173 | } |
6174 | ||
79f14641 | 6175 | return ret; |
f6c7d5fe PZ |
6176 | } |
6177 | ||
a8b0ca17 | 6178 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
6179 | struct perf_sample_data *data, |
6180 | struct pt_regs *regs) | |
850bc73f | 6181 | { |
a8b0ca17 | 6182 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
6183 | } |
6184 | ||
15dbf27c | 6185 | /* |
cdd6c482 | 6186 | * Generic software event infrastructure |
15dbf27c PZ |
6187 | */ |
6188 | ||
b28ab83c PZ |
6189 | struct swevent_htable { |
6190 | struct swevent_hlist *swevent_hlist; | |
6191 | struct mutex hlist_mutex; | |
6192 | int hlist_refcount; | |
6193 | ||
6194 | /* Recursion avoidance in each contexts */ | |
6195 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
6196 | |
6197 | /* Keeps track of cpu being initialized/exited */ | |
6198 | bool online; | |
b28ab83c PZ |
6199 | }; |
6200 | ||
6201 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
6202 | ||
7b4b6658 | 6203 | /* |
cdd6c482 IM |
6204 | * We directly increment event->count and keep a second value in |
6205 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
6206 | * is kept in the range [-sample_period, 0] so that we can use the |
6207 | * sign as trigger. | |
6208 | */ | |
6209 | ||
ab573844 | 6210 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 6211 | { |
cdd6c482 | 6212 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
6213 | u64 period = hwc->last_period; |
6214 | u64 nr, offset; | |
6215 | s64 old, val; | |
6216 | ||
6217 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
6218 | |
6219 | again: | |
e7850595 | 6220 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
6221 | if (val < 0) |
6222 | return 0; | |
15dbf27c | 6223 | |
7b4b6658 PZ |
6224 | nr = div64_u64(period + val, period); |
6225 | offset = nr * period; | |
6226 | val -= offset; | |
e7850595 | 6227 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 6228 | goto again; |
15dbf27c | 6229 | |
7b4b6658 | 6230 | return nr; |
15dbf27c PZ |
6231 | } |
6232 | ||
0cff784a | 6233 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 6234 | struct perf_sample_data *data, |
5622f295 | 6235 | struct pt_regs *regs) |
15dbf27c | 6236 | { |
cdd6c482 | 6237 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 6238 | int throttle = 0; |
15dbf27c | 6239 | |
0cff784a PZ |
6240 | if (!overflow) |
6241 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 6242 | |
7b4b6658 PZ |
6243 | if (hwc->interrupts == MAX_INTERRUPTS) |
6244 | return; | |
15dbf27c | 6245 | |
7b4b6658 | 6246 | for (; overflow; overflow--) { |
a8b0ca17 | 6247 | if (__perf_event_overflow(event, throttle, |
5622f295 | 6248 | data, regs)) { |
7b4b6658 PZ |
6249 | /* |
6250 | * We inhibit the overflow from happening when | |
6251 | * hwc->interrupts == MAX_INTERRUPTS. | |
6252 | */ | |
6253 | break; | |
6254 | } | |
cf450a73 | 6255 | throttle = 1; |
7b4b6658 | 6256 | } |
15dbf27c PZ |
6257 | } |
6258 | ||
a4eaf7f1 | 6259 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 6260 | struct perf_sample_data *data, |
5622f295 | 6261 | struct pt_regs *regs) |
7b4b6658 | 6262 | { |
cdd6c482 | 6263 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 6264 | |
e7850595 | 6265 | local64_add(nr, &event->count); |
d6d020e9 | 6266 | |
0cff784a PZ |
6267 | if (!regs) |
6268 | return; | |
6269 | ||
6c7e550f | 6270 | if (!is_sampling_event(event)) |
7b4b6658 | 6271 | return; |
d6d020e9 | 6272 | |
5d81e5cf AV |
6273 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
6274 | data->period = nr; | |
6275 | return perf_swevent_overflow(event, 1, data, regs); | |
6276 | } else | |
6277 | data->period = event->hw.last_period; | |
6278 | ||
0cff784a | 6279 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 6280 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 6281 | |
e7850595 | 6282 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 6283 | return; |
df1a132b | 6284 | |
a8b0ca17 | 6285 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
6286 | } |
6287 | ||
f5ffe02e FW |
6288 | static int perf_exclude_event(struct perf_event *event, |
6289 | struct pt_regs *regs) | |
6290 | { | |
a4eaf7f1 | 6291 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 6292 | return 1; |
a4eaf7f1 | 6293 | |
f5ffe02e FW |
6294 | if (regs) { |
6295 | if (event->attr.exclude_user && user_mode(regs)) | |
6296 | return 1; | |
6297 | ||
6298 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
6299 | return 1; | |
6300 | } | |
6301 | ||
6302 | return 0; | |
6303 | } | |
6304 | ||
cdd6c482 | 6305 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 6306 | enum perf_type_id type, |
6fb2915d LZ |
6307 | u32 event_id, |
6308 | struct perf_sample_data *data, | |
6309 | struct pt_regs *regs) | |
15dbf27c | 6310 | { |
cdd6c482 | 6311 | if (event->attr.type != type) |
a21ca2ca | 6312 | return 0; |
f5ffe02e | 6313 | |
cdd6c482 | 6314 | if (event->attr.config != event_id) |
15dbf27c PZ |
6315 | return 0; |
6316 | ||
f5ffe02e FW |
6317 | if (perf_exclude_event(event, regs)) |
6318 | return 0; | |
15dbf27c PZ |
6319 | |
6320 | return 1; | |
6321 | } | |
6322 | ||
76e1d904 FW |
6323 | static inline u64 swevent_hash(u64 type, u32 event_id) |
6324 | { | |
6325 | u64 val = event_id | (type << 32); | |
6326 | ||
6327 | return hash_64(val, SWEVENT_HLIST_BITS); | |
6328 | } | |
6329 | ||
49f135ed FW |
6330 | static inline struct hlist_head * |
6331 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 6332 | { |
49f135ed FW |
6333 | u64 hash = swevent_hash(type, event_id); |
6334 | ||
6335 | return &hlist->heads[hash]; | |
6336 | } | |
76e1d904 | 6337 | |
49f135ed FW |
6338 | /* For the read side: events when they trigger */ |
6339 | static inline struct hlist_head * | |
b28ab83c | 6340 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
6341 | { |
6342 | struct swevent_hlist *hlist; | |
76e1d904 | 6343 | |
b28ab83c | 6344 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
6345 | if (!hlist) |
6346 | return NULL; | |
6347 | ||
49f135ed FW |
6348 | return __find_swevent_head(hlist, type, event_id); |
6349 | } | |
6350 | ||
6351 | /* For the event head insertion and removal in the hlist */ | |
6352 | static inline struct hlist_head * | |
b28ab83c | 6353 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
6354 | { |
6355 | struct swevent_hlist *hlist; | |
6356 | u32 event_id = event->attr.config; | |
6357 | u64 type = event->attr.type; | |
6358 | ||
6359 | /* | |
6360 | * Event scheduling is always serialized against hlist allocation | |
6361 | * and release. Which makes the protected version suitable here. | |
6362 | * The context lock guarantees that. | |
6363 | */ | |
b28ab83c | 6364 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
6365 | lockdep_is_held(&event->ctx->lock)); |
6366 | if (!hlist) | |
6367 | return NULL; | |
6368 | ||
6369 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
6370 | } |
6371 | ||
6372 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 6373 | u64 nr, |
76e1d904 FW |
6374 | struct perf_sample_data *data, |
6375 | struct pt_regs *regs) | |
15dbf27c | 6376 | { |
4a32fea9 | 6377 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6378 | struct perf_event *event; |
76e1d904 | 6379 | struct hlist_head *head; |
15dbf27c | 6380 | |
76e1d904 | 6381 | rcu_read_lock(); |
b28ab83c | 6382 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
6383 | if (!head) |
6384 | goto end; | |
6385 | ||
b67bfe0d | 6386 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 6387 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 6388 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 6389 | } |
76e1d904 FW |
6390 | end: |
6391 | rcu_read_unlock(); | |
15dbf27c PZ |
6392 | } |
6393 | ||
86038c5e PZI |
6394 | DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]); |
6395 | ||
4ed7c92d | 6396 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 6397 | { |
4a32fea9 | 6398 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
96f6d444 | 6399 | |
b28ab83c | 6400 | return get_recursion_context(swhash->recursion); |
96f6d444 | 6401 | } |
645e8cc0 | 6402 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 6403 | |
fa9f90be | 6404 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 6405 | { |
4a32fea9 | 6406 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
927c7a9e | 6407 | |
b28ab83c | 6408 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 6409 | } |
15dbf27c | 6410 | |
86038c5e | 6411 | void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 6412 | { |
a4234bfc | 6413 | struct perf_sample_data data; |
4ed7c92d | 6414 | |
86038c5e | 6415 | if (WARN_ON_ONCE(!regs)) |
4ed7c92d | 6416 | return; |
a4234bfc | 6417 | |
fd0d000b | 6418 | perf_sample_data_init(&data, addr, 0); |
a8b0ca17 | 6419 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
86038c5e PZI |
6420 | } |
6421 | ||
6422 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) | |
6423 | { | |
6424 | int rctx; | |
6425 | ||
6426 | preempt_disable_notrace(); | |
6427 | rctx = perf_swevent_get_recursion_context(); | |
6428 | if (unlikely(rctx < 0)) | |
6429 | goto fail; | |
6430 | ||
6431 | ___perf_sw_event(event_id, nr, regs, addr); | |
4ed7c92d PZ |
6432 | |
6433 | perf_swevent_put_recursion_context(rctx); | |
86038c5e | 6434 | fail: |
1c024eca | 6435 | preempt_enable_notrace(); |
b8e83514 PZ |
6436 | } |
6437 | ||
cdd6c482 | 6438 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 6439 | { |
15dbf27c PZ |
6440 | } |
6441 | ||
a4eaf7f1 | 6442 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 6443 | { |
4a32fea9 | 6444 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6445 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
6446 | struct hlist_head *head; |
6447 | ||
6c7e550f | 6448 | if (is_sampling_event(event)) { |
7b4b6658 | 6449 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 6450 | perf_swevent_set_period(event); |
7b4b6658 | 6451 | } |
76e1d904 | 6452 | |
a4eaf7f1 PZ |
6453 | hwc->state = !(flags & PERF_EF_START); |
6454 | ||
b28ab83c | 6455 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
6456 | if (!head) { |
6457 | /* | |
6458 | * We can race with cpu hotplug code. Do not | |
6459 | * WARN if the cpu just got unplugged. | |
6460 | */ | |
6461 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 6462 | return -EINVAL; |
39af6b16 | 6463 | } |
76e1d904 FW |
6464 | |
6465 | hlist_add_head_rcu(&event->hlist_entry, head); | |
6a694a60 | 6466 | perf_event_update_userpage(event); |
76e1d904 | 6467 | |
15dbf27c PZ |
6468 | return 0; |
6469 | } | |
6470 | ||
a4eaf7f1 | 6471 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 6472 | { |
76e1d904 | 6473 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
6474 | } |
6475 | ||
a4eaf7f1 | 6476 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 6477 | { |
a4eaf7f1 | 6478 | event->hw.state = 0; |
d6d020e9 | 6479 | } |
aa9c4c0f | 6480 | |
a4eaf7f1 | 6481 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 6482 | { |
a4eaf7f1 | 6483 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
6484 | } |
6485 | ||
49f135ed FW |
6486 | /* Deref the hlist from the update side */ |
6487 | static inline struct swevent_hlist * | |
b28ab83c | 6488 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 6489 | { |
b28ab83c PZ |
6490 | return rcu_dereference_protected(swhash->swevent_hlist, |
6491 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
6492 | } |
6493 | ||
b28ab83c | 6494 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 6495 | { |
b28ab83c | 6496 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 6497 | |
49f135ed | 6498 | if (!hlist) |
76e1d904 FW |
6499 | return; |
6500 | ||
70691d4a | 6501 | RCU_INIT_POINTER(swhash->swevent_hlist, NULL); |
fa4bbc4c | 6502 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
6503 | } |
6504 | ||
6505 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
6506 | { | |
b28ab83c | 6507 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 6508 | |
b28ab83c | 6509 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6510 | |
b28ab83c PZ |
6511 | if (!--swhash->hlist_refcount) |
6512 | swevent_hlist_release(swhash); | |
76e1d904 | 6513 | |
b28ab83c | 6514 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6515 | } |
6516 | ||
6517 | static void swevent_hlist_put(struct perf_event *event) | |
6518 | { | |
6519 | int cpu; | |
6520 | ||
76e1d904 FW |
6521 | for_each_possible_cpu(cpu) |
6522 | swevent_hlist_put_cpu(event, cpu); | |
6523 | } | |
6524 | ||
6525 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
6526 | { | |
b28ab83c | 6527 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
6528 | int err = 0; |
6529 | ||
b28ab83c | 6530 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6531 | |
b28ab83c | 6532 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
6533 | struct swevent_hlist *hlist; |
6534 | ||
6535 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
6536 | if (!hlist) { | |
6537 | err = -ENOMEM; | |
6538 | goto exit; | |
6539 | } | |
b28ab83c | 6540 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 6541 | } |
b28ab83c | 6542 | swhash->hlist_refcount++; |
9ed6060d | 6543 | exit: |
b28ab83c | 6544 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6545 | |
6546 | return err; | |
6547 | } | |
6548 | ||
6549 | static int swevent_hlist_get(struct perf_event *event) | |
6550 | { | |
6551 | int err; | |
6552 | int cpu, failed_cpu; | |
6553 | ||
76e1d904 FW |
6554 | get_online_cpus(); |
6555 | for_each_possible_cpu(cpu) { | |
6556 | err = swevent_hlist_get_cpu(event, cpu); | |
6557 | if (err) { | |
6558 | failed_cpu = cpu; | |
6559 | goto fail; | |
6560 | } | |
6561 | } | |
6562 | put_online_cpus(); | |
6563 | ||
6564 | return 0; | |
9ed6060d | 6565 | fail: |
76e1d904 FW |
6566 | for_each_possible_cpu(cpu) { |
6567 | if (cpu == failed_cpu) | |
6568 | break; | |
6569 | swevent_hlist_put_cpu(event, cpu); | |
6570 | } | |
6571 | ||
6572 | put_online_cpus(); | |
6573 | return err; | |
6574 | } | |
6575 | ||
c5905afb | 6576 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 6577 | |
b0a873eb PZ |
6578 | static void sw_perf_event_destroy(struct perf_event *event) |
6579 | { | |
6580 | u64 event_id = event->attr.config; | |
95476b64 | 6581 | |
b0a873eb PZ |
6582 | WARN_ON(event->parent); |
6583 | ||
c5905afb | 6584 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6585 | swevent_hlist_put(event); |
6586 | } | |
6587 | ||
6588 | static int perf_swevent_init(struct perf_event *event) | |
6589 | { | |
8176cced | 6590 | u64 event_id = event->attr.config; |
b0a873eb PZ |
6591 | |
6592 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6593 | return -ENOENT; | |
6594 | ||
2481c5fa SE |
6595 | /* |
6596 | * no branch sampling for software events | |
6597 | */ | |
6598 | if (has_branch_stack(event)) | |
6599 | return -EOPNOTSUPP; | |
6600 | ||
b0a873eb PZ |
6601 | switch (event_id) { |
6602 | case PERF_COUNT_SW_CPU_CLOCK: | |
6603 | case PERF_COUNT_SW_TASK_CLOCK: | |
6604 | return -ENOENT; | |
6605 | ||
6606 | default: | |
6607 | break; | |
6608 | } | |
6609 | ||
ce677831 | 6610 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
6611 | return -ENOENT; |
6612 | ||
6613 | if (!event->parent) { | |
6614 | int err; | |
6615 | ||
6616 | err = swevent_hlist_get(event); | |
6617 | if (err) | |
6618 | return err; | |
6619 | ||
c5905afb | 6620 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6621 | event->destroy = sw_perf_event_destroy; |
6622 | } | |
6623 | ||
6624 | return 0; | |
6625 | } | |
6626 | ||
6627 | static struct pmu perf_swevent = { | |
89a1e187 | 6628 | .task_ctx_nr = perf_sw_context, |
95476b64 | 6629 | |
34f43927 PZ |
6630 | .capabilities = PERF_PMU_CAP_NO_NMI, |
6631 | ||
b0a873eb | 6632 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
6633 | .add = perf_swevent_add, |
6634 | .del = perf_swevent_del, | |
6635 | .start = perf_swevent_start, | |
6636 | .stop = perf_swevent_stop, | |
1c024eca | 6637 | .read = perf_swevent_read, |
1c024eca PZ |
6638 | }; |
6639 | ||
b0a873eb PZ |
6640 | #ifdef CONFIG_EVENT_TRACING |
6641 | ||
1c024eca PZ |
6642 | static int perf_tp_filter_match(struct perf_event *event, |
6643 | struct perf_sample_data *data) | |
6644 | { | |
6645 | void *record = data->raw->data; | |
6646 | ||
6647 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
6648 | return 1; | |
6649 | return 0; | |
6650 | } | |
6651 | ||
6652 | static int perf_tp_event_match(struct perf_event *event, | |
6653 | struct perf_sample_data *data, | |
6654 | struct pt_regs *regs) | |
6655 | { | |
a0f7d0f7 FW |
6656 | if (event->hw.state & PERF_HES_STOPPED) |
6657 | return 0; | |
580d607c PZ |
6658 | /* |
6659 | * All tracepoints are from kernel-space. | |
6660 | */ | |
6661 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
6662 | return 0; |
6663 | ||
6664 | if (!perf_tp_filter_match(event, data)) | |
6665 | return 0; | |
6666 | ||
6667 | return 1; | |
6668 | } | |
6669 | ||
6670 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
6671 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
6672 | struct task_struct *task) | |
95476b64 FW |
6673 | { |
6674 | struct perf_sample_data data; | |
1c024eca | 6675 | struct perf_event *event; |
1c024eca | 6676 | |
95476b64 FW |
6677 | struct perf_raw_record raw = { |
6678 | .size = entry_size, | |
6679 | .data = record, | |
6680 | }; | |
6681 | ||
fd0d000b | 6682 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
6683 | data.raw = &raw; |
6684 | ||
b67bfe0d | 6685 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 6686 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 6687 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 6688 | } |
ecc55f84 | 6689 | |
e6dab5ff AV |
6690 | /* |
6691 | * If we got specified a target task, also iterate its context and | |
6692 | * deliver this event there too. | |
6693 | */ | |
6694 | if (task && task != current) { | |
6695 | struct perf_event_context *ctx; | |
6696 | struct trace_entry *entry = record; | |
6697 | ||
6698 | rcu_read_lock(); | |
6699 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
6700 | if (!ctx) | |
6701 | goto unlock; | |
6702 | ||
6703 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
6704 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6705 | continue; | |
6706 | if (event->attr.config != entry->type) | |
6707 | continue; | |
6708 | if (perf_tp_event_match(event, &data, regs)) | |
6709 | perf_swevent_event(event, count, &data, regs); | |
6710 | } | |
6711 | unlock: | |
6712 | rcu_read_unlock(); | |
6713 | } | |
6714 | ||
ecc55f84 | 6715 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
6716 | } |
6717 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
6718 | ||
cdd6c482 | 6719 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 6720 | { |
1c024eca | 6721 | perf_trace_destroy(event); |
e077df4f PZ |
6722 | } |
6723 | ||
b0a873eb | 6724 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 6725 | { |
76e1d904 FW |
6726 | int err; |
6727 | ||
b0a873eb PZ |
6728 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
6729 | return -ENOENT; | |
6730 | ||
2481c5fa SE |
6731 | /* |
6732 | * no branch sampling for tracepoint events | |
6733 | */ | |
6734 | if (has_branch_stack(event)) | |
6735 | return -EOPNOTSUPP; | |
6736 | ||
1c024eca PZ |
6737 | err = perf_trace_init(event); |
6738 | if (err) | |
b0a873eb | 6739 | return err; |
e077df4f | 6740 | |
cdd6c482 | 6741 | event->destroy = tp_perf_event_destroy; |
e077df4f | 6742 | |
b0a873eb PZ |
6743 | return 0; |
6744 | } | |
6745 | ||
6746 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
6747 | .task_ctx_nr = perf_sw_context, |
6748 | ||
b0a873eb | 6749 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6750 | .add = perf_trace_add, |
6751 | .del = perf_trace_del, | |
6752 | .start = perf_swevent_start, | |
6753 | .stop = perf_swevent_stop, | |
b0a873eb | 6754 | .read = perf_swevent_read, |
b0a873eb PZ |
6755 | }; |
6756 | ||
6757 | static inline void perf_tp_register(void) | |
6758 | { | |
2e80a82a | 6759 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6760 | } |
6fb2915d LZ |
6761 | |
6762 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6763 | { | |
6764 | char *filter_str; | |
6765 | int ret; | |
6766 | ||
6767 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6768 | return -EINVAL; | |
6769 | ||
6770 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6771 | if (IS_ERR(filter_str)) | |
6772 | return PTR_ERR(filter_str); | |
6773 | ||
6774 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6775 | ||
6776 | kfree(filter_str); | |
6777 | return ret; | |
6778 | } | |
6779 | ||
6780 | static void perf_event_free_filter(struct perf_event *event) | |
6781 | { | |
6782 | ftrace_profile_free_filter(event); | |
6783 | } | |
6784 | ||
2541517c AS |
6785 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd) |
6786 | { | |
6787 | struct bpf_prog *prog; | |
6788 | ||
6789 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6790 | return -EINVAL; | |
6791 | ||
6792 | if (event->tp_event->prog) | |
6793 | return -EEXIST; | |
6794 | ||
6795 | if (!(event->tp_event->flags & TRACE_EVENT_FL_KPROBE)) | |
6796 | /* bpf programs can only be attached to kprobes */ | |
6797 | return -EINVAL; | |
6798 | ||
6799 | prog = bpf_prog_get(prog_fd); | |
6800 | if (IS_ERR(prog)) | |
6801 | return PTR_ERR(prog); | |
6802 | ||
6c373ca8 | 6803 | if (prog->type != BPF_PROG_TYPE_KPROBE) { |
2541517c AS |
6804 | /* valid fd, but invalid bpf program type */ |
6805 | bpf_prog_put(prog); | |
6806 | return -EINVAL; | |
6807 | } | |
6808 | ||
6809 | event->tp_event->prog = prog; | |
6810 | ||
6811 | return 0; | |
6812 | } | |
6813 | ||
6814 | static void perf_event_free_bpf_prog(struct perf_event *event) | |
6815 | { | |
6816 | struct bpf_prog *prog; | |
6817 | ||
6818 | if (!event->tp_event) | |
6819 | return; | |
6820 | ||
6821 | prog = event->tp_event->prog; | |
6822 | if (prog) { | |
6823 | event->tp_event->prog = NULL; | |
6824 | bpf_prog_put(prog); | |
6825 | } | |
6826 | } | |
6827 | ||
e077df4f | 6828 | #else |
6fb2915d | 6829 | |
b0a873eb | 6830 | static inline void perf_tp_register(void) |
e077df4f | 6831 | { |
e077df4f | 6832 | } |
6fb2915d LZ |
6833 | |
6834 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6835 | { | |
6836 | return -ENOENT; | |
6837 | } | |
6838 | ||
6839 | static void perf_event_free_filter(struct perf_event *event) | |
6840 | { | |
6841 | } | |
6842 | ||
2541517c AS |
6843 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd) |
6844 | { | |
6845 | return -ENOENT; | |
6846 | } | |
6847 | ||
6848 | static void perf_event_free_bpf_prog(struct perf_event *event) | |
6849 | { | |
6850 | } | |
07b139c8 | 6851 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6852 | |
24f1e32c | 6853 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6854 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6855 | { |
f5ffe02e FW |
6856 | struct perf_sample_data sample; |
6857 | struct pt_regs *regs = data; | |
6858 | ||
fd0d000b | 6859 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6860 | |
a4eaf7f1 | 6861 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6862 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6863 | } |
6864 | #endif | |
6865 | ||
b0a873eb PZ |
6866 | /* |
6867 | * hrtimer based swevent callback | |
6868 | */ | |
f29ac756 | 6869 | |
b0a873eb | 6870 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6871 | { |
b0a873eb PZ |
6872 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6873 | struct perf_sample_data data; | |
6874 | struct pt_regs *regs; | |
6875 | struct perf_event *event; | |
6876 | u64 period; | |
f29ac756 | 6877 | |
b0a873eb | 6878 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6879 | |
6880 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6881 | return HRTIMER_NORESTART; | |
6882 | ||
b0a873eb | 6883 | event->pmu->read(event); |
f344011c | 6884 | |
fd0d000b | 6885 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6886 | regs = get_irq_regs(); |
6887 | ||
6888 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6889 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6890 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6891 | ret = HRTIMER_NORESTART; |
6892 | } | |
24f1e32c | 6893 | |
b0a873eb PZ |
6894 | period = max_t(u64, 10000, event->hw.sample_period); |
6895 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6896 | |
b0a873eb | 6897 | return ret; |
f29ac756 PZ |
6898 | } |
6899 | ||
b0a873eb | 6900 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6901 | { |
b0a873eb | 6902 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6903 | s64 period; |
6904 | ||
6905 | if (!is_sampling_event(event)) | |
6906 | return; | |
f5ffe02e | 6907 | |
5d508e82 FBH |
6908 | period = local64_read(&hwc->period_left); |
6909 | if (period) { | |
6910 | if (period < 0) | |
6911 | period = 10000; | |
fa407f35 | 6912 | |
5d508e82 FBH |
6913 | local64_set(&hwc->period_left, 0); |
6914 | } else { | |
6915 | period = max_t(u64, 10000, hwc->sample_period); | |
6916 | } | |
3497d206 TG |
6917 | hrtimer_start(&hwc->hrtimer, ns_to_ktime(period), |
6918 | HRTIMER_MODE_REL_PINNED); | |
24f1e32c | 6919 | } |
b0a873eb PZ |
6920 | |
6921 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6922 | { |
b0a873eb PZ |
6923 | struct hw_perf_event *hwc = &event->hw; |
6924 | ||
6c7e550f | 6925 | if (is_sampling_event(event)) { |
b0a873eb | 6926 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6927 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6928 | |
6929 | hrtimer_cancel(&hwc->hrtimer); | |
6930 | } | |
24f1e32c FW |
6931 | } |
6932 | ||
ba3dd36c PZ |
6933 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6934 | { | |
6935 | struct hw_perf_event *hwc = &event->hw; | |
6936 | ||
6937 | if (!is_sampling_event(event)) | |
6938 | return; | |
6939 | ||
6940 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6941 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6942 | ||
6943 | /* | |
6944 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6945 | * mapping and avoid the whole period adjust feedback stuff. | |
6946 | */ | |
6947 | if (event->attr.freq) { | |
6948 | long freq = event->attr.sample_freq; | |
6949 | ||
6950 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6951 | hwc->sample_period = event->attr.sample_period; | |
6952 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6953 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6954 | event->attr.freq = 0; |
6955 | } | |
6956 | } | |
6957 | ||
b0a873eb PZ |
6958 | /* |
6959 | * Software event: cpu wall time clock | |
6960 | */ | |
6961 | ||
6962 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6963 | { |
b0a873eb PZ |
6964 | s64 prev; |
6965 | u64 now; | |
6966 | ||
a4eaf7f1 | 6967 | now = local_clock(); |
b0a873eb PZ |
6968 | prev = local64_xchg(&event->hw.prev_count, now); |
6969 | local64_add(now - prev, &event->count); | |
24f1e32c | 6970 | } |
24f1e32c | 6971 | |
a4eaf7f1 | 6972 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6973 | { |
a4eaf7f1 | 6974 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6975 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6976 | } |
6977 | ||
a4eaf7f1 | 6978 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6979 | { |
b0a873eb PZ |
6980 | perf_swevent_cancel_hrtimer(event); |
6981 | cpu_clock_event_update(event); | |
6982 | } | |
f29ac756 | 6983 | |
a4eaf7f1 PZ |
6984 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6985 | { | |
6986 | if (flags & PERF_EF_START) | |
6987 | cpu_clock_event_start(event, flags); | |
6a694a60 | 6988 | perf_event_update_userpage(event); |
a4eaf7f1 PZ |
6989 | |
6990 | return 0; | |
6991 | } | |
6992 | ||
6993 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6994 | { | |
6995 | cpu_clock_event_stop(event, flags); | |
6996 | } | |
6997 | ||
b0a873eb PZ |
6998 | static void cpu_clock_event_read(struct perf_event *event) |
6999 | { | |
7000 | cpu_clock_event_update(event); | |
7001 | } | |
f344011c | 7002 | |
b0a873eb PZ |
7003 | static int cpu_clock_event_init(struct perf_event *event) |
7004 | { | |
7005 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
7006 | return -ENOENT; | |
7007 | ||
7008 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
7009 | return -ENOENT; | |
7010 | ||
2481c5fa SE |
7011 | /* |
7012 | * no branch sampling for software events | |
7013 | */ | |
7014 | if (has_branch_stack(event)) | |
7015 | return -EOPNOTSUPP; | |
7016 | ||
ba3dd36c PZ |
7017 | perf_swevent_init_hrtimer(event); |
7018 | ||
b0a873eb | 7019 | return 0; |
f29ac756 PZ |
7020 | } |
7021 | ||
b0a873eb | 7022 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
7023 | .task_ctx_nr = perf_sw_context, |
7024 | ||
34f43927 PZ |
7025 | .capabilities = PERF_PMU_CAP_NO_NMI, |
7026 | ||
b0a873eb | 7027 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
7028 | .add = cpu_clock_event_add, |
7029 | .del = cpu_clock_event_del, | |
7030 | .start = cpu_clock_event_start, | |
7031 | .stop = cpu_clock_event_stop, | |
b0a873eb PZ |
7032 | .read = cpu_clock_event_read, |
7033 | }; | |
7034 | ||
7035 | /* | |
7036 | * Software event: task time clock | |
7037 | */ | |
7038 | ||
7039 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 7040 | { |
b0a873eb PZ |
7041 | u64 prev; |
7042 | s64 delta; | |
5c92d124 | 7043 | |
b0a873eb PZ |
7044 | prev = local64_xchg(&event->hw.prev_count, now); |
7045 | delta = now - prev; | |
7046 | local64_add(delta, &event->count); | |
7047 | } | |
5c92d124 | 7048 | |
a4eaf7f1 | 7049 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 7050 | { |
a4eaf7f1 | 7051 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 7052 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
7053 | } |
7054 | ||
a4eaf7f1 | 7055 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
7056 | { |
7057 | perf_swevent_cancel_hrtimer(event); | |
7058 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
7059 | } |
7060 | ||
7061 | static int task_clock_event_add(struct perf_event *event, int flags) | |
7062 | { | |
7063 | if (flags & PERF_EF_START) | |
7064 | task_clock_event_start(event, flags); | |
6a694a60 | 7065 | perf_event_update_userpage(event); |
b0a873eb | 7066 | |
a4eaf7f1 PZ |
7067 | return 0; |
7068 | } | |
7069 | ||
7070 | static void task_clock_event_del(struct perf_event *event, int flags) | |
7071 | { | |
7072 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
7073 | } |
7074 | ||
7075 | static void task_clock_event_read(struct perf_event *event) | |
7076 | { | |
768a06e2 PZ |
7077 | u64 now = perf_clock(); |
7078 | u64 delta = now - event->ctx->timestamp; | |
7079 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
7080 | |
7081 | task_clock_event_update(event, time); | |
7082 | } | |
7083 | ||
7084 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 7085 | { |
b0a873eb PZ |
7086 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
7087 | return -ENOENT; | |
7088 | ||
7089 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
7090 | return -ENOENT; | |
7091 | ||
2481c5fa SE |
7092 | /* |
7093 | * no branch sampling for software events | |
7094 | */ | |
7095 | if (has_branch_stack(event)) | |
7096 | return -EOPNOTSUPP; | |
7097 | ||
ba3dd36c PZ |
7098 | perf_swevent_init_hrtimer(event); |
7099 | ||
b0a873eb | 7100 | return 0; |
6fb2915d LZ |
7101 | } |
7102 | ||
b0a873eb | 7103 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
7104 | .task_ctx_nr = perf_sw_context, |
7105 | ||
34f43927 PZ |
7106 | .capabilities = PERF_PMU_CAP_NO_NMI, |
7107 | ||
b0a873eb | 7108 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
7109 | .add = task_clock_event_add, |
7110 | .del = task_clock_event_del, | |
7111 | .start = task_clock_event_start, | |
7112 | .stop = task_clock_event_stop, | |
b0a873eb PZ |
7113 | .read = task_clock_event_read, |
7114 | }; | |
6fb2915d | 7115 | |
ad5133b7 | 7116 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 7117 | { |
e077df4f | 7118 | } |
6fb2915d | 7119 | |
ad5133b7 | 7120 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 7121 | { |
ad5133b7 | 7122 | return 0; |
6fb2915d LZ |
7123 | } |
7124 | ||
ad5133b7 | 7125 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 7126 | { |
ad5133b7 | 7127 | perf_pmu_disable(pmu); |
6fb2915d LZ |
7128 | } |
7129 | ||
ad5133b7 PZ |
7130 | static int perf_pmu_commit_txn(struct pmu *pmu) |
7131 | { | |
7132 | perf_pmu_enable(pmu); | |
7133 | return 0; | |
7134 | } | |
e077df4f | 7135 | |
ad5133b7 | 7136 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 7137 | { |
ad5133b7 | 7138 | perf_pmu_enable(pmu); |
24f1e32c FW |
7139 | } |
7140 | ||
35edc2a5 PZ |
7141 | static int perf_event_idx_default(struct perf_event *event) |
7142 | { | |
c719f560 | 7143 | return 0; |
35edc2a5 PZ |
7144 | } |
7145 | ||
8dc85d54 PZ |
7146 | /* |
7147 | * Ensures all contexts with the same task_ctx_nr have the same | |
7148 | * pmu_cpu_context too. | |
7149 | */ | |
9e317041 | 7150 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 7151 | { |
8dc85d54 | 7152 | struct pmu *pmu; |
b326e956 | 7153 | |
8dc85d54 PZ |
7154 | if (ctxn < 0) |
7155 | return NULL; | |
24f1e32c | 7156 | |
8dc85d54 PZ |
7157 | list_for_each_entry(pmu, &pmus, entry) { |
7158 | if (pmu->task_ctx_nr == ctxn) | |
7159 | return pmu->pmu_cpu_context; | |
7160 | } | |
24f1e32c | 7161 | |
8dc85d54 | 7162 | return NULL; |
24f1e32c FW |
7163 | } |
7164 | ||
51676957 | 7165 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 7166 | { |
51676957 PZ |
7167 | int cpu; |
7168 | ||
7169 | for_each_possible_cpu(cpu) { | |
7170 | struct perf_cpu_context *cpuctx; | |
7171 | ||
7172 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
7173 | ||
3f1f3320 PZ |
7174 | if (cpuctx->unique_pmu == old_pmu) |
7175 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
7176 | } |
7177 | } | |
7178 | ||
7179 | static void free_pmu_context(struct pmu *pmu) | |
7180 | { | |
7181 | struct pmu *i; | |
f5ffe02e | 7182 | |
8dc85d54 | 7183 | mutex_lock(&pmus_lock); |
0475f9ea | 7184 | /* |
8dc85d54 | 7185 | * Like a real lame refcount. |
0475f9ea | 7186 | */ |
51676957 PZ |
7187 | list_for_each_entry(i, &pmus, entry) { |
7188 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
7189 | update_pmu_context(i, pmu); | |
8dc85d54 | 7190 | goto out; |
51676957 | 7191 | } |
8dc85d54 | 7192 | } |
d6d020e9 | 7193 | |
51676957 | 7194 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
7195 | out: |
7196 | mutex_unlock(&pmus_lock); | |
24f1e32c | 7197 | } |
2e80a82a | 7198 | static struct idr pmu_idr; |
d6d020e9 | 7199 | |
abe43400 PZ |
7200 | static ssize_t |
7201 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
7202 | { | |
7203 | struct pmu *pmu = dev_get_drvdata(dev); | |
7204 | ||
7205 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
7206 | } | |
90826ca7 | 7207 | static DEVICE_ATTR_RO(type); |
abe43400 | 7208 | |
62b85639 SE |
7209 | static ssize_t |
7210 | perf_event_mux_interval_ms_show(struct device *dev, | |
7211 | struct device_attribute *attr, | |
7212 | char *page) | |
7213 | { | |
7214 | struct pmu *pmu = dev_get_drvdata(dev); | |
7215 | ||
7216 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
7217 | } | |
7218 | ||
272325c4 PZ |
7219 | static DEFINE_MUTEX(mux_interval_mutex); |
7220 | ||
62b85639 SE |
7221 | static ssize_t |
7222 | perf_event_mux_interval_ms_store(struct device *dev, | |
7223 | struct device_attribute *attr, | |
7224 | const char *buf, size_t count) | |
7225 | { | |
7226 | struct pmu *pmu = dev_get_drvdata(dev); | |
7227 | int timer, cpu, ret; | |
7228 | ||
7229 | ret = kstrtoint(buf, 0, &timer); | |
7230 | if (ret) | |
7231 | return ret; | |
7232 | ||
7233 | if (timer < 1) | |
7234 | return -EINVAL; | |
7235 | ||
7236 | /* same value, noting to do */ | |
7237 | if (timer == pmu->hrtimer_interval_ms) | |
7238 | return count; | |
7239 | ||
272325c4 | 7240 | mutex_lock(&mux_interval_mutex); |
62b85639 SE |
7241 | pmu->hrtimer_interval_ms = timer; |
7242 | ||
7243 | /* update all cpuctx for this PMU */ | |
272325c4 PZ |
7244 | get_online_cpus(); |
7245 | for_each_online_cpu(cpu) { | |
62b85639 SE |
7246 | struct perf_cpu_context *cpuctx; |
7247 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
7248 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
7249 | ||
272325c4 PZ |
7250 | cpu_function_call(cpu, |
7251 | (remote_function_f)perf_mux_hrtimer_restart, cpuctx); | |
62b85639 | 7252 | } |
272325c4 PZ |
7253 | put_online_cpus(); |
7254 | mutex_unlock(&mux_interval_mutex); | |
62b85639 SE |
7255 | |
7256 | return count; | |
7257 | } | |
90826ca7 | 7258 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 7259 | |
90826ca7 GKH |
7260 | static struct attribute *pmu_dev_attrs[] = { |
7261 | &dev_attr_type.attr, | |
7262 | &dev_attr_perf_event_mux_interval_ms.attr, | |
7263 | NULL, | |
abe43400 | 7264 | }; |
90826ca7 | 7265 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
7266 | |
7267 | static int pmu_bus_running; | |
7268 | static struct bus_type pmu_bus = { | |
7269 | .name = "event_source", | |
90826ca7 | 7270 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
7271 | }; |
7272 | ||
7273 | static void pmu_dev_release(struct device *dev) | |
7274 | { | |
7275 | kfree(dev); | |
7276 | } | |
7277 | ||
7278 | static int pmu_dev_alloc(struct pmu *pmu) | |
7279 | { | |
7280 | int ret = -ENOMEM; | |
7281 | ||
7282 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
7283 | if (!pmu->dev) | |
7284 | goto out; | |
7285 | ||
0c9d42ed | 7286 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
7287 | device_initialize(pmu->dev); |
7288 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
7289 | if (ret) | |
7290 | goto free_dev; | |
7291 | ||
7292 | dev_set_drvdata(pmu->dev, pmu); | |
7293 | pmu->dev->bus = &pmu_bus; | |
7294 | pmu->dev->release = pmu_dev_release; | |
7295 | ret = device_add(pmu->dev); | |
7296 | if (ret) | |
7297 | goto free_dev; | |
7298 | ||
7299 | out: | |
7300 | return ret; | |
7301 | ||
7302 | free_dev: | |
7303 | put_device(pmu->dev); | |
7304 | goto out; | |
7305 | } | |
7306 | ||
547e9fd7 | 7307 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 7308 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 7309 | |
03d8e80b | 7310 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 7311 | { |
108b02cf | 7312 | int cpu, ret; |
24f1e32c | 7313 | |
b0a873eb | 7314 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
7315 | ret = -ENOMEM; |
7316 | pmu->pmu_disable_count = alloc_percpu(int); | |
7317 | if (!pmu->pmu_disable_count) | |
7318 | goto unlock; | |
f29ac756 | 7319 | |
2e80a82a PZ |
7320 | pmu->type = -1; |
7321 | if (!name) | |
7322 | goto skip_type; | |
7323 | pmu->name = name; | |
7324 | ||
7325 | if (type < 0) { | |
0e9c3be2 TH |
7326 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
7327 | if (type < 0) { | |
7328 | ret = type; | |
2e80a82a PZ |
7329 | goto free_pdc; |
7330 | } | |
7331 | } | |
7332 | pmu->type = type; | |
7333 | ||
abe43400 PZ |
7334 | if (pmu_bus_running) { |
7335 | ret = pmu_dev_alloc(pmu); | |
7336 | if (ret) | |
7337 | goto free_idr; | |
7338 | } | |
7339 | ||
2e80a82a | 7340 | skip_type: |
8dc85d54 PZ |
7341 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
7342 | if (pmu->pmu_cpu_context) | |
7343 | goto got_cpu_context; | |
f29ac756 | 7344 | |
c4814202 | 7345 | ret = -ENOMEM; |
108b02cf PZ |
7346 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
7347 | if (!pmu->pmu_cpu_context) | |
abe43400 | 7348 | goto free_dev; |
f344011c | 7349 | |
108b02cf PZ |
7350 | for_each_possible_cpu(cpu) { |
7351 | struct perf_cpu_context *cpuctx; | |
7352 | ||
7353 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 7354 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 7355 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 7356 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
108b02cf | 7357 | cpuctx->ctx.pmu = pmu; |
9e630205 | 7358 | |
272325c4 | 7359 | __perf_mux_hrtimer_init(cpuctx, cpu); |
9e630205 | 7360 | |
3f1f3320 | 7361 | cpuctx->unique_pmu = pmu; |
108b02cf | 7362 | } |
76e1d904 | 7363 | |
8dc85d54 | 7364 | got_cpu_context: |
ad5133b7 PZ |
7365 | if (!pmu->start_txn) { |
7366 | if (pmu->pmu_enable) { | |
7367 | /* | |
7368 | * If we have pmu_enable/pmu_disable calls, install | |
7369 | * transaction stubs that use that to try and batch | |
7370 | * hardware accesses. | |
7371 | */ | |
7372 | pmu->start_txn = perf_pmu_start_txn; | |
7373 | pmu->commit_txn = perf_pmu_commit_txn; | |
7374 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
7375 | } else { | |
7376 | pmu->start_txn = perf_pmu_nop_void; | |
7377 | pmu->commit_txn = perf_pmu_nop_int; | |
7378 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 7379 | } |
5c92d124 | 7380 | } |
15dbf27c | 7381 | |
ad5133b7 PZ |
7382 | if (!pmu->pmu_enable) { |
7383 | pmu->pmu_enable = perf_pmu_nop_void; | |
7384 | pmu->pmu_disable = perf_pmu_nop_void; | |
7385 | } | |
7386 | ||
35edc2a5 PZ |
7387 | if (!pmu->event_idx) |
7388 | pmu->event_idx = perf_event_idx_default; | |
7389 | ||
b0a873eb | 7390 | list_add_rcu(&pmu->entry, &pmus); |
bed5b25a | 7391 | atomic_set(&pmu->exclusive_cnt, 0); |
33696fc0 PZ |
7392 | ret = 0; |
7393 | unlock: | |
b0a873eb PZ |
7394 | mutex_unlock(&pmus_lock); |
7395 | ||
33696fc0 | 7396 | return ret; |
108b02cf | 7397 | |
abe43400 PZ |
7398 | free_dev: |
7399 | device_del(pmu->dev); | |
7400 | put_device(pmu->dev); | |
7401 | ||
2e80a82a PZ |
7402 | free_idr: |
7403 | if (pmu->type >= PERF_TYPE_MAX) | |
7404 | idr_remove(&pmu_idr, pmu->type); | |
7405 | ||
108b02cf PZ |
7406 | free_pdc: |
7407 | free_percpu(pmu->pmu_disable_count); | |
7408 | goto unlock; | |
f29ac756 | 7409 | } |
c464c76e | 7410 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 7411 | |
b0a873eb | 7412 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 7413 | { |
b0a873eb PZ |
7414 | mutex_lock(&pmus_lock); |
7415 | list_del_rcu(&pmu->entry); | |
7416 | mutex_unlock(&pmus_lock); | |
5c92d124 | 7417 | |
0475f9ea | 7418 | /* |
cde8e884 PZ |
7419 | * We dereference the pmu list under both SRCU and regular RCU, so |
7420 | * synchronize against both of those. | |
0475f9ea | 7421 | */ |
b0a873eb | 7422 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 7423 | synchronize_rcu(); |
d6d020e9 | 7424 | |
33696fc0 | 7425 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
7426 | if (pmu->type >= PERF_TYPE_MAX) |
7427 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
7428 | device_del(pmu->dev); |
7429 | put_device(pmu->dev); | |
51676957 | 7430 | free_pmu_context(pmu); |
b0a873eb | 7431 | } |
c464c76e | 7432 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 7433 | |
cc34b98b MR |
7434 | static int perf_try_init_event(struct pmu *pmu, struct perf_event *event) |
7435 | { | |
ccd41c86 | 7436 | struct perf_event_context *ctx = NULL; |
cc34b98b MR |
7437 | int ret; |
7438 | ||
7439 | if (!try_module_get(pmu->module)) | |
7440 | return -ENODEV; | |
ccd41c86 PZ |
7441 | |
7442 | if (event->group_leader != event) { | |
8b10c5e2 PZ |
7443 | /* |
7444 | * This ctx->mutex can nest when we're called through | |
7445 | * inheritance. See the perf_event_ctx_lock_nested() comment. | |
7446 | */ | |
7447 | ctx = perf_event_ctx_lock_nested(event->group_leader, | |
7448 | SINGLE_DEPTH_NESTING); | |
ccd41c86 PZ |
7449 | BUG_ON(!ctx); |
7450 | } | |
7451 | ||
cc34b98b MR |
7452 | event->pmu = pmu; |
7453 | ret = pmu->event_init(event); | |
ccd41c86 PZ |
7454 | |
7455 | if (ctx) | |
7456 | perf_event_ctx_unlock(event->group_leader, ctx); | |
7457 | ||
cc34b98b MR |
7458 | if (ret) |
7459 | module_put(pmu->module); | |
7460 | ||
7461 | return ret; | |
7462 | } | |
7463 | ||
b0a873eb PZ |
7464 | struct pmu *perf_init_event(struct perf_event *event) |
7465 | { | |
7466 | struct pmu *pmu = NULL; | |
7467 | int idx; | |
940c5b29 | 7468 | int ret; |
b0a873eb PZ |
7469 | |
7470 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
7471 | |
7472 | rcu_read_lock(); | |
7473 | pmu = idr_find(&pmu_idr, event->attr.type); | |
7474 | rcu_read_unlock(); | |
940c5b29 | 7475 | if (pmu) { |
cc34b98b | 7476 | ret = perf_try_init_event(pmu, event); |
940c5b29 LM |
7477 | if (ret) |
7478 | pmu = ERR_PTR(ret); | |
2e80a82a | 7479 | goto unlock; |
940c5b29 | 7480 | } |
2e80a82a | 7481 | |
b0a873eb | 7482 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
cc34b98b | 7483 | ret = perf_try_init_event(pmu, event); |
b0a873eb | 7484 | if (!ret) |
e5f4d339 | 7485 | goto unlock; |
76e1d904 | 7486 | |
b0a873eb PZ |
7487 | if (ret != -ENOENT) { |
7488 | pmu = ERR_PTR(ret); | |
e5f4d339 | 7489 | goto unlock; |
f344011c | 7490 | } |
5c92d124 | 7491 | } |
e5f4d339 PZ |
7492 | pmu = ERR_PTR(-ENOENT); |
7493 | unlock: | |
b0a873eb | 7494 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 7495 | |
4aeb0b42 | 7496 | return pmu; |
5c92d124 IM |
7497 | } |
7498 | ||
4beb31f3 FW |
7499 | static void account_event_cpu(struct perf_event *event, int cpu) |
7500 | { | |
7501 | if (event->parent) | |
7502 | return; | |
7503 | ||
4beb31f3 FW |
7504 | if (is_cgroup_event(event)) |
7505 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
7506 | } | |
7507 | ||
766d6c07 FW |
7508 | static void account_event(struct perf_event *event) |
7509 | { | |
4beb31f3 FW |
7510 | if (event->parent) |
7511 | return; | |
7512 | ||
766d6c07 FW |
7513 | if (event->attach_state & PERF_ATTACH_TASK) |
7514 | static_key_slow_inc(&perf_sched_events.key); | |
7515 | if (event->attr.mmap || event->attr.mmap_data) | |
7516 | atomic_inc(&nr_mmap_events); | |
7517 | if (event->attr.comm) | |
7518 | atomic_inc(&nr_comm_events); | |
7519 | if (event->attr.task) | |
7520 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
7521 | if (event->attr.freq) { |
7522 | if (atomic_inc_return(&nr_freq_events) == 1) | |
7523 | tick_nohz_full_kick_all(); | |
7524 | } | |
4beb31f3 | 7525 | if (has_branch_stack(event)) |
766d6c07 | 7526 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 7527 | if (is_cgroup_event(event)) |
766d6c07 | 7528 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
7529 | |
7530 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
7531 | } |
7532 | ||
0793a61d | 7533 | /* |
cdd6c482 | 7534 | * Allocate and initialize a event structure |
0793a61d | 7535 | */ |
cdd6c482 | 7536 | static struct perf_event * |
c3f00c70 | 7537 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
7538 | struct task_struct *task, |
7539 | struct perf_event *group_leader, | |
7540 | struct perf_event *parent_event, | |
4dc0da86 | 7541 | perf_overflow_handler_t overflow_handler, |
79dff51e | 7542 | void *context, int cgroup_fd) |
0793a61d | 7543 | { |
51b0fe39 | 7544 | struct pmu *pmu; |
cdd6c482 IM |
7545 | struct perf_event *event; |
7546 | struct hw_perf_event *hwc; | |
90983b16 | 7547 | long err = -EINVAL; |
0793a61d | 7548 | |
66832eb4 ON |
7549 | if ((unsigned)cpu >= nr_cpu_ids) { |
7550 | if (!task || cpu != -1) | |
7551 | return ERR_PTR(-EINVAL); | |
7552 | } | |
7553 | ||
c3f00c70 | 7554 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 7555 | if (!event) |
d5d2bc0d | 7556 | return ERR_PTR(-ENOMEM); |
0793a61d | 7557 | |
04289bb9 | 7558 | /* |
cdd6c482 | 7559 | * Single events are their own group leaders, with an |
04289bb9 IM |
7560 | * empty sibling list: |
7561 | */ | |
7562 | if (!group_leader) | |
cdd6c482 | 7563 | group_leader = event; |
04289bb9 | 7564 | |
cdd6c482 IM |
7565 | mutex_init(&event->child_mutex); |
7566 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 7567 | |
cdd6c482 IM |
7568 | INIT_LIST_HEAD(&event->group_entry); |
7569 | INIT_LIST_HEAD(&event->event_entry); | |
7570 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 7571 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 7572 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
7573 | INIT_HLIST_NODE(&event->hlist_entry); |
7574 | ||
10c6db11 | 7575 | |
cdd6c482 | 7576 | init_waitqueue_head(&event->waitq); |
e360adbe | 7577 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 7578 | |
cdd6c482 | 7579 | mutex_init(&event->mmap_mutex); |
7b732a75 | 7580 | |
a6fa941d | 7581 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
7582 | event->cpu = cpu; |
7583 | event->attr = *attr; | |
7584 | event->group_leader = group_leader; | |
7585 | event->pmu = NULL; | |
cdd6c482 | 7586 | event->oncpu = -1; |
a96bbc16 | 7587 | |
cdd6c482 | 7588 | event->parent = parent_event; |
b84fbc9f | 7589 | |
17cf22c3 | 7590 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 7591 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 7592 | |
cdd6c482 | 7593 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 7594 | |
d580ff86 PZ |
7595 | if (task) { |
7596 | event->attach_state = PERF_ATTACH_TASK; | |
d580ff86 | 7597 | /* |
50f16a8b PZ |
7598 | * XXX pmu::event_init needs to know what task to account to |
7599 | * and we cannot use the ctx information because we need the | |
7600 | * pmu before we get a ctx. | |
d580ff86 | 7601 | */ |
50f16a8b | 7602 | event->hw.target = task; |
d580ff86 PZ |
7603 | } |
7604 | ||
34f43927 PZ |
7605 | event->clock = &local_clock; |
7606 | if (parent_event) | |
7607 | event->clock = parent_event->clock; | |
7608 | ||
4dc0da86 | 7609 | if (!overflow_handler && parent_event) { |
b326e956 | 7610 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
7611 | context = parent_event->overflow_handler_context; |
7612 | } | |
66832eb4 | 7613 | |
b326e956 | 7614 | event->overflow_handler = overflow_handler; |
4dc0da86 | 7615 | event->overflow_handler_context = context; |
97eaf530 | 7616 | |
0231bb53 | 7617 | perf_event__state_init(event); |
a86ed508 | 7618 | |
4aeb0b42 | 7619 | pmu = NULL; |
b8e83514 | 7620 | |
cdd6c482 | 7621 | hwc = &event->hw; |
bd2b5b12 | 7622 | hwc->sample_period = attr->sample_period; |
0d48696f | 7623 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 7624 | hwc->sample_period = 1; |
eced1dfc | 7625 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 7626 | |
e7850595 | 7627 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 7628 | |
2023b359 | 7629 | /* |
cdd6c482 | 7630 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 7631 | */ |
3dab77fb | 7632 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 7633 | goto err_ns; |
a46a2300 YZ |
7634 | |
7635 | if (!has_branch_stack(event)) | |
7636 | event->attr.branch_sample_type = 0; | |
2023b359 | 7637 | |
79dff51e MF |
7638 | if (cgroup_fd != -1) { |
7639 | err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader); | |
7640 | if (err) | |
7641 | goto err_ns; | |
7642 | } | |
7643 | ||
b0a873eb | 7644 | pmu = perf_init_event(event); |
4aeb0b42 | 7645 | if (!pmu) |
90983b16 FW |
7646 | goto err_ns; |
7647 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 7648 | err = PTR_ERR(pmu); |
90983b16 | 7649 | goto err_ns; |
621a01ea | 7650 | } |
d5d2bc0d | 7651 | |
bed5b25a AS |
7652 | err = exclusive_event_init(event); |
7653 | if (err) | |
7654 | goto err_pmu; | |
7655 | ||
cdd6c482 | 7656 | if (!event->parent) { |
927c7a9e FW |
7657 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
7658 | err = get_callchain_buffers(); | |
90983b16 | 7659 | if (err) |
bed5b25a | 7660 | goto err_per_task; |
d010b332 | 7661 | } |
f344011c | 7662 | } |
9ee318a7 | 7663 | |
cdd6c482 | 7664 | return event; |
90983b16 | 7665 | |
bed5b25a AS |
7666 | err_per_task: |
7667 | exclusive_event_destroy(event); | |
7668 | ||
90983b16 FW |
7669 | err_pmu: |
7670 | if (event->destroy) | |
7671 | event->destroy(event); | |
c464c76e | 7672 | module_put(pmu->module); |
90983b16 | 7673 | err_ns: |
79dff51e MF |
7674 | if (is_cgroup_event(event)) |
7675 | perf_detach_cgroup(event); | |
90983b16 FW |
7676 | if (event->ns) |
7677 | put_pid_ns(event->ns); | |
7678 | kfree(event); | |
7679 | ||
7680 | return ERR_PTR(err); | |
0793a61d TG |
7681 | } |
7682 | ||
cdd6c482 IM |
7683 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
7684 | struct perf_event_attr *attr) | |
974802ea | 7685 | { |
974802ea | 7686 | u32 size; |
cdf8073d | 7687 | int ret; |
974802ea PZ |
7688 | |
7689 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
7690 | return -EFAULT; | |
7691 | ||
7692 | /* | |
7693 | * zero the full structure, so that a short copy will be nice. | |
7694 | */ | |
7695 | memset(attr, 0, sizeof(*attr)); | |
7696 | ||
7697 | ret = get_user(size, &uattr->size); | |
7698 | if (ret) | |
7699 | return ret; | |
7700 | ||
7701 | if (size > PAGE_SIZE) /* silly large */ | |
7702 | goto err_size; | |
7703 | ||
7704 | if (!size) /* abi compat */ | |
7705 | size = PERF_ATTR_SIZE_VER0; | |
7706 | ||
7707 | if (size < PERF_ATTR_SIZE_VER0) | |
7708 | goto err_size; | |
7709 | ||
7710 | /* | |
7711 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
7712 | * ensure all the unknown bits are 0 - i.e. new |
7713 | * user-space does not rely on any kernel feature | |
7714 | * extensions we dont know about yet. | |
974802ea PZ |
7715 | */ |
7716 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
7717 | unsigned char __user *addr; |
7718 | unsigned char __user *end; | |
7719 | unsigned char val; | |
974802ea | 7720 | |
cdf8073d IS |
7721 | addr = (void __user *)uattr + sizeof(*attr); |
7722 | end = (void __user *)uattr + size; | |
974802ea | 7723 | |
cdf8073d | 7724 | for (; addr < end; addr++) { |
974802ea PZ |
7725 | ret = get_user(val, addr); |
7726 | if (ret) | |
7727 | return ret; | |
7728 | if (val) | |
7729 | goto err_size; | |
7730 | } | |
b3e62e35 | 7731 | size = sizeof(*attr); |
974802ea PZ |
7732 | } |
7733 | ||
7734 | ret = copy_from_user(attr, uattr, size); | |
7735 | if (ret) | |
7736 | return -EFAULT; | |
7737 | ||
cd757645 | 7738 | if (attr->__reserved_1) |
974802ea PZ |
7739 | return -EINVAL; |
7740 | ||
7741 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
7742 | return -EINVAL; | |
7743 | ||
7744 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
7745 | return -EINVAL; | |
7746 | ||
bce38cd5 SE |
7747 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
7748 | u64 mask = attr->branch_sample_type; | |
7749 | ||
7750 | /* only using defined bits */ | |
7751 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
7752 | return -EINVAL; | |
7753 | ||
7754 | /* at least one branch bit must be set */ | |
7755 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
7756 | return -EINVAL; | |
7757 | ||
bce38cd5 SE |
7758 | /* propagate priv level, when not set for branch */ |
7759 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
7760 | ||
7761 | /* exclude_kernel checked on syscall entry */ | |
7762 | if (!attr->exclude_kernel) | |
7763 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
7764 | ||
7765 | if (!attr->exclude_user) | |
7766 | mask |= PERF_SAMPLE_BRANCH_USER; | |
7767 | ||
7768 | if (!attr->exclude_hv) | |
7769 | mask |= PERF_SAMPLE_BRANCH_HV; | |
7770 | /* | |
7771 | * adjust user setting (for HW filter setup) | |
7772 | */ | |
7773 | attr->branch_sample_type = mask; | |
7774 | } | |
e712209a SE |
7775 | /* privileged levels capture (kernel, hv): check permissions */ |
7776 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
7777 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
7778 | return -EACCES; | |
bce38cd5 | 7779 | } |
4018994f | 7780 | |
c5ebcedb | 7781 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 7782 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
7783 | if (ret) |
7784 | return ret; | |
7785 | } | |
7786 | ||
7787 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
7788 | if (!arch_perf_have_user_stack_dump()) | |
7789 | return -ENOSYS; | |
7790 | ||
7791 | /* | |
7792 | * We have __u32 type for the size, but so far | |
7793 | * we can only use __u16 as maximum due to the | |
7794 | * __u16 sample size limit. | |
7795 | */ | |
7796 | if (attr->sample_stack_user >= USHRT_MAX) | |
7797 | ret = -EINVAL; | |
7798 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
7799 | ret = -EINVAL; | |
7800 | } | |
4018994f | 7801 | |
60e2364e SE |
7802 | if (attr->sample_type & PERF_SAMPLE_REGS_INTR) |
7803 | ret = perf_reg_validate(attr->sample_regs_intr); | |
974802ea PZ |
7804 | out: |
7805 | return ret; | |
7806 | ||
7807 | err_size: | |
7808 | put_user(sizeof(*attr), &uattr->size); | |
7809 | ret = -E2BIG; | |
7810 | goto out; | |
7811 | } | |
7812 | ||
ac9721f3 PZ |
7813 | static int |
7814 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 7815 | { |
b69cf536 | 7816 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
7817 | int ret = -EINVAL; |
7818 | ||
ac9721f3 | 7819 | if (!output_event) |
a4be7c27 PZ |
7820 | goto set; |
7821 | ||
ac9721f3 PZ |
7822 | /* don't allow circular references */ |
7823 | if (event == output_event) | |
a4be7c27 PZ |
7824 | goto out; |
7825 | ||
0f139300 PZ |
7826 | /* |
7827 | * Don't allow cross-cpu buffers | |
7828 | */ | |
7829 | if (output_event->cpu != event->cpu) | |
7830 | goto out; | |
7831 | ||
7832 | /* | |
76369139 | 7833 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
7834 | */ |
7835 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
7836 | goto out; | |
7837 | ||
34f43927 PZ |
7838 | /* |
7839 | * Mixing clocks in the same buffer is trouble you don't need. | |
7840 | */ | |
7841 | if (output_event->clock != event->clock) | |
7842 | goto out; | |
7843 | ||
45bfb2e5 PZ |
7844 | /* |
7845 | * If both events generate aux data, they must be on the same PMU | |
7846 | */ | |
7847 | if (has_aux(event) && has_aux(output_event) && | |
7848 | event->pmu != output_event->pmu) | |
7849 | goto out; | |
7850 | ||
a4be7c27 | 7851 | set: |
cdd6c482 | 7852 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7853 | /* Can't redirect output if we've got an active mmap() */ |
7854 | if (atomic_read(&event->mmap_count)) | |
7855 | goto unlock; | |
a4be7c27 | 7856 | |
ac9721f3 | 7857 | if (output_event) { |
76369139 FW |
7858 | /* get the rb we want to redirect to */ |
7859 | rb = ring_buffer_get(output_event); | |
7860 | if (!rb) | |
ac9721f3 | 7861 | goto unlock; |
a4be7c27 PZ |
7862 | } |
7863 | ||
b69cf536 | 7864 | ring_buffer_attach(event, rb); |
9bb5d40c | 7865 | |
a4be7c27 | 7866 | ret = 0; |
ac9721f3 PZ |
7867 | unlock: |
7868 | mutex_unlock(&event->mmap_mutex); | |
7869 | ||
a4be7c27 | 7870 | out: |
a4be7c27 PZ |
7871 | return ret; |
7872 | } | |
7873 | ||
f63a8daa PZ |
7874 | static void mutex_lock_double(struct mutex *a, struct mutex *b) |
7875 | { | |
7876 | if (b < a) | |
7877 | swap(a, b); | |
7878 | ||
7879 | mutex_lock(a); | |
7880 | mutex_lock_nested(b, SINGLE_DEPTH_NESTING); | |
7881 | } | |
7882 | ||
34f43927 PZ |
7883 | static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id) |
7884 | { | |
7885 | bool nmi_safe = false; | |
7886 | ||
7887 | switch (clk_id) { | |
7888 | case CLOCK_MONOTONIC: | |
7889 | event->clock = &ktime_get_mono_fast_ns; | |
7890 | nmi_safe = true; | |
7891 | break; | |
7892 | ||
7893 | case CLOCK_MONOTONIC_RAW: | |
7894 | event->clock = &ktime_get_raw_fast_ns; | |
7895 | nmi_safe = true; | |
7896 | break; | |
7897 | ||
7898 | case CLOCK_REALTIME: | |
7899 | event->clock = &ktime_get_real_ns; | |
7900 | break; | |
7901 | ||
7902 | case CLOCK_BOOTTIME: | |
7903 | event->clock = &ktime_get_boot_ns; | |
7904 | break; | |
7905 | ||
7906 | case CLOCK_TAI: | |
7907 | event->clock = &ktime_get_tai_ns; | |
7908 | break; | |
7909 | ||
7910 | default: | |
7911 | return -EINVAL; | |
7912 | } | |
7913 | ||
7914 | if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI)) | |
7915 | return -EINVAL; | |
7916 | ||
7917 | return 0; | |
7918 | } | |
7919 | ||
0793a61d | 7920 | /** |
cdd6c482 | 7921 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7922 | * |
cdd6c482 | 7923 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7924 | * @pid: target pid |
9f66a381 | 7925 | * @cpu: target cpu |
cdd6c482 | 7926 | * @group_fd: group leader event fd |
0793a61d | 7927 | */ |
cdd6c482 IM |
7928 | SYSCALL_DEFINE5(perf_event_open, |
7929 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7930 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7931 | { |
b04243ef PZ |
7932 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7933 | struct perf_event *event, *sibling; | |
cdd6c482 | 7934 | struct perf_event_attr attr; |
f63a8daa | 7935 | struct perf_event_context *ctx, *uninitialized_var(gctx); |
cdd6c482 | 7936 | struct file *event_file = NULL; |
2903ff01 | 7937 | struct fd group = {NULL, 0}; |
38a81da2 | 7938 | struct task_struct *task = NULL; |
89a1e187 | 7939 | struct pmu *pmu; |
ea635c64 | 7940 | int event_fd; |
b04243ef | 7941 | int move_group = 0; |
dc86cabe | 7942 | int err; |
a21b0b35 | 7943 | int f_flags = O_RDWR; |
79dff51e | 7944 | int cgroup_fd = -1; |
0793a61d | 7945 | |
2743a5b0 | 7946 | /* for future expandability... */ |
e5d1367f | 7947 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7948 | return -EINVAL; |
7949 | ||
dc86cabe IM |
7950 | err = perf_copy_attr(attr_uptr, &attr); |
7951 | if (err) | |
7952 | return err; | |
eab656ae | 7953 | |
0764771d PZ |
7954 | if (!attr.exclude_kernel) { |
7955 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7956 | return -EACCES; | |
7957 | } | |
7958 | ||
df58ab24 | 7959 | if (attr.freq) { |
cdd6c482 | 7960 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 7961 | return -EINVAL; |
0819b2e3 PZ |
7962 | } else { |
7963 | if (attr.sample_period & (1ULL << 63)) | |
7964 | return -EINVAL; | |
df58ab24 PZ |
7965 | } |
7966 | ||
e5d1367f SE |
7967 | /* |
7968 | * In cgroup mode, the pid argument is used to pass the fd | |
7969 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7970 | * designates the cpu on which to monitor threads from that | |
7971 | * cgroup. | |
7972 | */ | |
7973 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7974 | return -EINVAL; | |
7975 | ||
a21b0b35 YD |
7976 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7977 | f_flags |= O_CLOEXEC; | |
7978 | ||
7979 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7980 | if (event_fd < 0) |
7981 | return event_fd; | |
7982 | ||
ac9721f3 | 7983 | if (group_fd != -1) { |
2903ff01 AV |
7984 | err = perf_fget_light(group_fd, &group); |
7985 | if (err) | |
d14b12d7 | 7986 | goto err_fd; |
2903ff01 | 7987 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7988 | if (flags & PERF_FLAG_FD_OUTPUT) |
7989 | output_event = group_leader; | |
7990 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7991 | group_leader = NULL; | |
7992 | } | |
7993 | ||
e5d1367f | 7994 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7995 | task = find_lively_task_by_vpid(pid); |
7996 | if (IS_ERR(task)) { | |
7997 | err = PTR_ERR(task); | |
7998 | goto err_group_fd; | |
7999 | } | |
8000 | } | |
8001 | ||
1f4ee503 PZ |
8002 | if (task && group_leader && |
8003 | group_leader->attr.inherit != attr.inherit) { | |
8004 | err = -EINVAL; | |
8005 | goto err_task; | |
8006 | } | |
8007 | ||
fbfc623f YZ |
8008 | get_online_cpus(); |
8009 | ||
79dff51e MF |
8010 | if (flags & PERF_FLAG_PID_CGROUP) |
8011 | cgroup_fd = pid; | |
8012 | ||
4dc0da86 | 8013 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
79dff51e | 8014 | NULL, NULL, cgroup_fd); |
d14b12d7 SE |
8015 | if (IS_ERR(event)) { |
8016 | err = PTR_ERR(event); | |
1f4ee503 | 8017 | goto err_cpus; |
d14b12d7 SE |
8018 | } |
8019 | ||
53b25335 VW |
8020 | if (is_sampling_event(event)) { |
8021 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
8022 | err = -ENOTSUPP; | |
8023 | goto err_alloc; | |
8024 | } | |
8025 | } | |
8026 | ||
766d6c07 FW |
8027 | account_event(event); |
8028 | ||
89a1e187 PZ |
8029 | /* |
8030 | * Special case software events and allow them to be part of | |
8031 | * any hardware group. | |
8032 | */ | |
8033 | pmu = event->pmu; | |
b04243ef | 8034 | |
34f43927 PZ |
8035 | if (attr.use_clockid) { |
8036 | err = perf_event_set_clock(event, attr.clockid); | |
8037 | if (err) | |
8038 | goto err_alloc; | |
8039 | } | |
8040 | ||
b04243ef PZ |
8041 | if (group_leader && |
8042 | (is_software_event(event) != is_software_event(group_leader))) { | |
8043 | if (is_software_event(event)) { | |
8044 | /* | |
8045 | * If event and group_leader are not both a software | |
8046 | * event, and event is, then group leader is not. | |
8047 | * | |
8048 | * Allow the addition of software events to !software | |
8049 | * groups, this is safe because software events never | |
8050 | * fail to schedule. | |
8051 | */ | |
8052 | pmu = group_leader->pmu; | |
8053 | } else if (is_software_event(group_leader) && | |
8054 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
8055 | /* | |
8056 | * In case the group is a pure software group, and we | |
8057 | * try to add a hardware event, move the whole group to | |
8058 | * the hardware context. | |
8059 | */ | |
8060 | move_group = 1; | |
8061 | } | |
8062 | } | |
89a1e187 PZ |
8063 | |
8064 | /* | |
8065 | * Get the target context (task or percpu): | |
8066 | */ | |
4af57ef2 | 8067 | ctx = find_get_context(pmu, task, event); |
89a1e187 PZ |
8068 | if (IS_ERR(ctx)) { |
8069 | err = PTR_ERR(ctx); | |
c6be5a5c | 8070 | goto err_alloc; |
89a1e187 PZ |
8071 | } |
8072 | ||
bed5b25a AS |
8073 | if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) { |
8074 | err = -EBUSY; | |
8075 | goto err_context; | |
8076 | } | |
8077 | ||
fd1edb3a PZ |
8078 | if (task) { |
8079 | put_task_struct(task); | |
8080 | task = NULL; | |
8081 | } | |
8082 | ||
ccff286d | 8083 | /* |
cdd6c482 | 8084 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 8085 | */ |
ac9721f3 | 8086 | if (group_leader) { |
dc86cabe | 8087 | err = -EINVAL; |
04289bb9 | 8088 | |
04289bb9 | 8089 | /* |
ccff286d IM |
8090 | * Do not allow a recursive hierarchy (this new sibling |
8091 | * becoming part of another group-sibling): | |
8092 | */ | |
8093 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 8094 | goto err_context; |
34f43927 PZ |
8095 | |
8096 | /* All events in a group should have the same clock */ | |
8097 | if (group_leader->clock != event->clock) | |
8098 | goto err_context; | |
8099 | ||
ccff286d IM |
8100 | /* |
8101 | * Do not allow to attach to a group in a different | |
8102 | * task or CPU context: | |
04289bb9 | 8103 | */ |
b04243ef | 8104 | if (move_group) { |
c3c87e77 PZ |
8105 | /* |
8106 | * Make sure we're both on the same task, or both | |
8107 | * per-cpu events. | |
8108 | */ | |
8109 | if (group_leader->ctx->task != ctx->task) | |
8110 | goto err_context; | |
8111 | ||
8112 | /* | |
8113 | * Make sure we're both events for the same CPU; | |
8114 | * grouping events for different CPUs is broken; since | |
8115 | * you can never concurrently schedule them anyhow. | |
8116 | */ | |
8117 | if (group_leader->cpu != event->cpu) | |
b04243ef PZ |
8118 | goto err_context; |
8119 | } else { | |
8120 | if (group_leader->ctx != ctx) | |
8121 | goto err_context; | |
8122 | } | |
8123 | ||
3b6f9e5c PM |
8124 | /* |
8125 | * Only a group leader can be exclusive or pinned | |
8126 | */ | |
0d48696f | 8127 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 8128 | goto err_context; |
ac9721f3 PZ |
8129 | } |
8130 | ||
8131 | if (output_event) { | |
8132 | err = perf_event_set_output(event, output_event); | |
8133 | if (err) | |
c3f00c70 | 8134 | goto err_context; |
ac9721f3 | 8135 | } |
0793a61d | 8136 | |
a21b0b35 YD |
8137 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
8138 | f_flags); | |
ea635c64 AV |
8139 | if (IS_ERR(event_file)) { |
8140 | err = PTR_ERR(event_file); | |
c3f00c70 | 8141 | goto err_context; |
ea635c64 | 8142 | } |
9b51f66d | 8143 | |
b04243ef | 8144 | if (move_group) { |
f63a8daa PZ |
8145 | gctx = group_leader->ctx; |
8146 | ||
8147 | /* | |
8148 | * See perf_event_ctx_lock() for comments on the details | |
8149 | * of swizzling perf_event::ctx. | |
8150 | */ | |
8151 | mutex_lock_double(&gctx->mutex, &ctx->mutex); | |
b04243ef | 8152 | |
46ce0fe9 | 8153 | perf_remove_from_context(group_leader, false); |
0231bb53 | 8154 | |
b04243ef PZ |
8155 | list_for_each_entry(sibling, &group_leader->sibling_list, |
8156 | group_entry) { | |
46ce0fe9 | 8157 | perf_remove_from_context(sibling, false); |
b04243ef PZ |
8158 | put_ctx(gctx); |
8159 | } | |
f63a8daa PZ |
8160 | } else { |
8161 | mutex_lock(&ctx->mutex); | |
ea635c64 | 8162 | } |
9b51f66d | 8163 | |
ad3a37de | 8164 | WARN_ON_ONCE(ctx->parent_ctx); |
b04243ef PZ |
8165 | |
8166 | if (move_group) { | |
f63a8daa PZ |
8167 | /* |
8168 | * Wait for everybody to stop referencing the events through | |
8169 | * the old lists, before installing it on new lists. | |
8170 | */ | |
0cda4c02 | 8171 | synchronize_rcu(); |
f63a8daa | 8172 | |
8f95b435 PZI |
8173 | /* |
8174 | * Install the group siblings before the group leader. | |
8175 | * | |
8176 | * Because a group leader will try and install the entire group | |
8177 | * (through the sibling list, which is still in-tact), we can | |
8178 | * end up with siblings installed in the wrong context. | |
8179 | * | |
8180 | * By installing siblings first we NO-OP because they're not | |
8181 | * reachable through the group lists. | |
8182 | */ | |
b04243ef PZ |
8183 | list_for_each_entry(sibling, &group_leader->sibling_list, |
8184 | group_entry) { | |
8f95b435 | 8185 | perf_event__state_init(sibling); |
9fc81d87 | 8186 | perf_install_in_context(ctx, sibling, sibling->cpu); |
b04243ef PZ |
8187 | get_ctx(ctx); |
8188 | } | |
8f95b435 PZI |
8189 | |
8190 | /* | |
8191 | * Removing from the context ends up with disabled | |
8192 | * event. What we want here is event in the initial | |
8193 | * startup state, ready to be add into new context. | |
8194 | */ | |
8195 | perf_event__state_init(group_leader); | |
8196 | perf_install_in_context(ctx, group_leader, group_leader->cpu); | |
8197 | get_ctx(ctx); | |
b04243ef PZ |
8198 | } |
8199 | ||
bed5b25a AS |
8200 | if (!exclusive_event_installable(event, ctx)) { |
8201 | err = -EBUSY; | |
8202 | mutex_unlock(&ctx->mutex); | |
8203 | fput(event_file); | |
8204 | goto err_context; | |
8205 | } | |
8206 | ||
e2d37cd2 | 8207 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 8208 | perf_unpin_context(ctx); |
f63a8daa PZ |
8209 | |
8210 | if (move_group) { | |
8211 | mutex_unlock(&gctx->mutex); | |
8212 | put_ctx(gctx); | |
8213 | } | |
d859e29f | 8214 | mutex_unlock(&ctx->mutex); |
9b51f66d | 8215 | |
fbfc623f YZ |
8216 | put_online_cpus(); |
8217 | ||
cdd6c482 | 8218 | event->owner = current; |
8882135b | 8219 | |
cdd6c482 IM |
8220 | mutex_lock(¤t->perf_event_mutex); |
8221 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
8222 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 8223 | |
c320c7b7 ACM |
8224 | /* |
8225 | * Precalculate sample_data sizes | |
8226 | */ | |
8227 | perf_event__header_size(event); | |
6844c09d | 8228 | perf_event__id_header_size(event); |
c320c7b7 | 8229 | |
8a49542c PZ |
8230 | /* |
8231 | * Drop the reference on the group_event after placing the | |
8232 | * new event on the sibling_list. This ensures destruction | |
8233 | * of the group leader will find the pointer to itself in | |
8234 | * perf_group_detach(). | |
8235 | */ | |
2903ff01 | 8236 | fdput(group); |
ea635c64 AV |
8237 | fd_install(event_fd, event_file); |
8238 | return event_fd; | |
0793a61d | 8239 | |
c3f00c70 | 8240 | err_context: |
fe4b04fa | 8241 | perf_unpin_context(ctx); |
ea635c64 | 8242 | put_ctx(ctx); |
c6be5a5c | 8243 | err_alloc: |
ea635c64 | 8244 | free_event(event); |
1f4ee503 | 8245 | err_cpus: |
fbfc623f | 8246 | put_online_cpus(); |
1f4ee503 | 8247 | err_task: |
e7d0bc04 PZ |
8248 | if (task) |
8249 | put_task_struct(task); | |
89a1e187 | 8250 | err_group_fd: |
2903ff01 | 8251 | fdput(group); |
ea635c64 AV |
8252 | err_fd: |
8253 | put_unused_fd(event_fd); | |
dc86cabe | 8254 | return err; |
0793a61d TG |
8255 | } |
8256 | ||
fb0459d7 AV |
8257 | /** |
8258 | * perf_event_create_kernel_counter | |
8259 | * | |
8260 | * @attr: attributes of the counter to create | |
8261 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 8262 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
8263 | */ |
8264 | struct perf_event * | |
8265 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 8266 | struct task_struct *task, |
4dc0da86 AK |
8267 | perf_overflow_handler_t overflow_handler, |
8268 | void *context) | |
fb0459d7 | 8269 | { |
fb0459d7 | 8270 | struct perf_event_context *ctx; |
c3f00c70 | 8271 | struct perf_event *event; |
fb0459d7 | 8272 | int err; |
d859e29f | 8273 | |
fb0459d7 AV |
8274 | /* |
8275 | * Get the target context (task or percpu): | |
8276 | */ | |
d859e29f | 8277 | |
4dc0da86 | 8278 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
79dff51e | 8279 | overflow_handler, context, -1); |
c3f00c70 PZ |
8280 | if (IS_ERR(event)) { |
8281 | err = PTR_ERR(event); | |
8282 | goto err; | |
8283 | } | |
d859e29f | 8284 | |
f8697762 JO |
8285 | /* Mark owner so we could distinguish it from user events. */ |
8286 | event->owner = EVENT_OWNER_KERNEL; | |
8287 | ||
766d6c07 FW |
8288 | account_event(event); |
8289 | ||
4af57ef2 | 8290 | ctx = find_get_context(event->pmu, task, event); |
c6567f64 FW |
8291 | if (IS_ERR(ctx)) { |
8292 | err = PTR_ERR(ctx); | |
c3f00c70 | 8293 | goto err_free; |
d859e29f | 8294 | } |
fb0459d7 | 8295 | |
fb0459d7 AV |
8296 | WARN_ON_ONCE(ctx->parent_ctx); |
8297 | mutex_lock(&ctx->mutex); | |
bed5b25a AS |
8298 | if (!exclusive_event_installable(event, ctx)) { |
8299 | mutex_unlock(&ctx->mutex); | |
8300 | perf_unpin_context(ctx); | |
8301 | put_ctx(ctx); | |
8302 | err = -EBUSY; | |
8303 | goto err_free; | |
8304 | } | |
8305 | ||
fb0459d7 | 8306 | perf_install_in_context(ctx, event, cpu); |
fe4b04fa | 8307 | perf_unpin_context(ctx); |
fb0459d7 AV |
8308 | mutex_unlock(&ctx->mutex); |
8309 | ||
fb0459d7 AV |
8310 | return event; |
8311 | ||
c3f00c70 PZ |
8312 | err_free: |
8313 | free_event(event); | |
8314 | err: | |
c6567f64 | 8315 | return ERR_PTR(err); |
9b51f66d | 8316 | } |
fb0459d7 | 8317 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 8318 | |
0cda4c02 YZ |
8319 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
8320 | { | |
8321 | struct perf_event_context *src_ctx; | |
8322 | struct perf_event_context *dst_ctx; | |
8323 | struct perf_event *event, *tmp; | |
8324 | LIST_HEAD(events); | |
8325 | ||
8326 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
8327 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
8328 | ||
f63a8daa PZ |
8329 | /* |
8330 | * See perf_event_ctx_lock() for comments on the details | |
8331 | * of swizzling perf_event::ctx. | |
8332 | */ | |
8333 | mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); | |
0cda4c02 YZ |
8334 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, |
8335 | event_entry) { | |
46ce0fe9 | 8336 | perf_remove_from_context(event, false); |
9a545de0 | 8337 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 8338 | put_ctx(src_ctx); |
9886167d | 8339 | list_add(&event->migrate_entry, &events); |
0cda4c02 | 8340 | } |
0cda4c02 | 8341 | |
8f95b435 PZI |
8342 | /* |
8343 | * Wait for the events to quiesce before re-instating them. | |
8344 | */ | |
0cda4c02 YZ |
8345 | synchronize_rcu(); |
8346 | ||
8f95b435 PZI |
8347 | /* |
8348 | * Re-instate events in 2 passes. | |
8349 | * | |
8350 | * Skip over group leaders and only install siblings on this first | |
8351 | * pass, siblings will not get enabled without a leader, however a | |
8352 | * leader will enable its siblings, even if those are still on the old | |
8353 | * context. | |
8354 | */ | |
8355 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { | |
8356 | if (event->group_leader == event) | |
8357 | continue; | |
8358 | ||
8359 | list_del(&event->migrate_entry); | |
8360 | if (event->state >= PERF_EVENT_STATE_OFF) | |
8361 | event->state = PERF_EVENT_STATE_INACTIVE; | |
8362 | account_event_cpu(event, dst_cpu); | |
8363 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
8364 | get_ctx(dst_ctx); | |
8365 | } | |
8366 | ||
8367 | /* | |
8368 | * Once all the siblings are setup properly, install the group leaders | |
8369 | * to make it go. | |
8370 | */ | |
9886167d PZ |
8371 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
8372 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
8373 | if (event->state >= PERF_EVENT_STATE_OFF) |
8374 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 8375 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
8376 | perf_install_in_context(dst_ctx, event, dst_cpu); |
8377 | get_ctx(dst_ctx); | |
8378 | } | |
8379 | mutex_unlock(&dst_ctx->mutex); | |
f63a8daa | 8380 | mutex_unlock(&src_ctx->mutex); |
0cda4c02 YZ |
8381 | } |
8382 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
8383 | ||
cdd6c482 | 8384 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 8385 | struct task_struct *child) |
d859e29f | 8386 | { |
cdd6c482 | 8387 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 8388 | u64 child_val; |
d859e29f | 8389 | |
cdd6c482 IM |
8390 | if (child_event->attr.inherit_stat) |
8391 | perf_event_read_event(child_event, child); | |
38b200d6 | 8392 | |
b5e58793 | 8393 | child_val = perf_event_count(child_event); |
d859e29f PM |
8394 | |
8395 | /* | |
8396 | * Add back the child's count to the parent's count: | |
8397 | */ | |
a6e6dea6 | 8398 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
8399 | atomic64_add(child_event->total_time_enabled, |
8400 | &parent_event->child_total_time_enabled); | |
8401 | atomic64_add(child_event->total_time_running, | |
8402 | &parent_event->child_total_time_running); | |
d859e29f PM |
8403 | |
8404 | /* | |
cdd6c482 | 8405 | * Remove this event from the parent's list |
d859e29f | 8406 | */ |
cdd6c482 IM |
8407 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
8408 | mutex_lock(&parent_event->child_mutex); | |
8409 | list_del_init(&child_event->child_list); | |
8410 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f | 8411 | |
dc633982 JO |
8412 | /* |
8413 | * Make sure user/parent get notified, that we just | |
8414 | * lost one event. | |
8415 | */ | |
8416 | perf_event_wakeup(parent_event); | |
8417 | ||
d859e29f | 8418 | /* |
cdd6c482 | 8419 | * Release the parent event, if this was the last |
d859e29f PM |
8420 | * reference to it. |
8421 | */ | |
a6fa941d | 8422 | put_event(parent_event); |
d859e29f PM |
8423 | } |
8424 | ||
9b51f66d | 8425 | static void |
cdd6c482 IM |
8426 | __perf_event_exit_task(struct perf_event *child_event, |
8427 | struct perf_event_context *child_ctx, | |
38b200d6 | 8428 | struct task_struct *child) |
9b51f66d | 8429 | { |
1903d50c PZ |
8430 | /* |
8431 | * Do not destroy the 'original' grouping; because of the context | |
8432 | * switch optimization the original events could've ended up in a | |
8433 | * random child task. | |
8434 | * | |
8435 | * If we were to destroy the original group, all group related | |
8436 | * operations would cease to function properly after this random | |
8437 | * child dies. | |
8438 | * | |
8439 | * Do destroy all inherited groups, we don't care about those | |
8440 | * and being thorough is better. | |
8441 | */ | |
8442 | perf_remove_from_context(child_event, !!child_event->parent); | |
0cc0c027 | 8443 | |
9b51f66d | 8444 | /* |
38b435b1 | 8445 | * It can happen that the parent exits first, and has events |
9b51f66d | 8446 | * that are still around due to the child reference. These |
38b435b1 | 8447 | * events need to be zapped. |
9b51f66d | 8448 | */ |
38b435b1 | 8449 | if (child_event->parent) { |
cdd6c482 IM |
8450 | sync_child_event(child_event, child); |
8451 | free_event(child_event); | |
179033b3 JO |
8452 | } else { |
8453 | child_event->state = PERF_EVENT_STATE_EXIT; | |
8454 | perf_event_wakeup(child_event); | |
4bcf349a | 8455 | } |
9b51f66d IM |
8456 | } |
8457 | ||
8dc85d54 | 8458 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 8459 | { |
ebf905fc | 8460 | struct perf_event *child_event, *next; |
211de6eb | 8461 | struct perf_event_context *child_ctx, *clone_ctx = NULL; |
a63eaf34 | 8462 | unsigned long flags; |
9b51f66d | 8463 | |
8dc85d54 | 8464 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 8465 | perf_event_task(child, NULL, 0); |
9b51f66d | 8466 | return; |
9f498cc5 | 8467 | } |
9b51f66d | 8468 | |
a63eaf34 | 8469 | local_irq_save(flags); |
ad3a37de PM |
8470 | /* |
8471 | * We can't reschedule here because interrupts are disabled, | |
8472 | * and either child is current or it is a task that can't be | |
8473 | * scheduled, so we are now safe from rescheduling changing | |
8474 | * our context. | |
8475 | */ | |
806839b2 | 8476 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
8477 | |
8478 | /* | |
8479 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 8480 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
8481 | * incremented the context's refcount before we do put_ctx below. |
8482 | */ | |
e625cce1 | 8483 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 8484 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 8485 | child->perf_event_ctxp[ctxn] = NULL; |
4a1c0f26 | 8486 | |
71a851b4 PZ |
8487 | /* |
8488 | * If this context is a clone; unclone it so it can't get | |
8489 | * swapped to another process while we're removing all | |
cdd6c482 | 8490 | * the events from it. |
71a851b4 | 8491 | */ |
211de6eb | 8492 | clone_ctx = unclone_ctx(child_ctx); |
5e942bb3 | 8493 | update_context_time(child_ctx); |
e625cce1 | 8494 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 | 8495 | |
211de6eb PZ |
8496 | if (clone_ctx) |
8497 | put_ctx(clone_ctx); | |
4a1c0f26 | 8498 | |
9f498cc5 | 8499 | /* |
cdd6c482 IM |
8500 | * Report the task dead after unscheduling the events so that we |
8501 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
8502 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 8503 | */ |
cdd6c482 | 8504 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 8505 | |
66fff224 PZ |
8506 | /* |
8507 | * We can recurse on the same lock type through: | |
8508 | * | |
cdd6c482 IM |
8509 | * __perf_event_exit_task() |
8510 | * sync_child_event() | |
a6fa941d AV |
8511 | * put_event() |
8512 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
8513 | * |
8514 | * But since its the parent context it won't be the same instance. | |
8515 | */ | |
a0507c84 | 8516 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 8517 | |
ebf905fc | 8518 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 8519 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 8520 | |
a63eaf34 PM |
8521 | mutex_unlock(&child_ctx->mutex); |
8522 | ||
8523 | put_ctx(child_ctx); | |
9b51f66d IM |
8524 | } |
8525 | ||
8dc85d54 PZ |
8526 | /* |
8527 | * When a child task exits, feed back event values to parent events. | |
8528 | */ | |
8529 | void perf_event_exit_task(struct task_struct *child) | |
8530 | { | |
8882135b | 8531 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
8532 | int ctxn; |
8533 | ||
8882135b PZ |
8534 | mutex_lock(&child->perf_event_mutex); |
8535 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
8536 | owner_entry) { | |
8537 | list_del_init(&event->owner_entry); | |
8538 | ||
8539 | /* | |
8540 | * Ensure the list deletion is visible before we clear | |
8541 | * the owner, closes a race against perf_release() where | |
8542 | * we need to serialize on the owner->perf_event_mutex. | |
8543 | */ | |
8544 | smp_wmb(); | |
8545 | event->owner = NULL; | |
8546 | } | |
8547 | mutex_unlock(&child->perf_event_mutex); | |
8548 | ||
8dc85d54 PZ |
8549 | for_each_task_context_nr(ctxn) |
8550 | perf_event_exit_task_context(child, ctxn); | |
8551 | } | |
8552 | ||
889ff015 FW |
8553 | static void perf_free_event(struct perf_event *event, |
8554 | struct perf_event_context *ctx) | |
8555 | { | |
8556 | struct perf_event *parent = event->parent; | |
8557 | ||
8558 | if (WARN_ON_ONCE(!parent)) | |
8559 | return; | |
8560 | ||
8561 | mutex_lock(&parent->child_mutex); | |
8562 | list_del_init(&event->child_list); | |
8563 | mutex_unlock(&parent->child_mutex); | |
8564 | ||
a6fa941d | 8565 | put_event(parent); |
889ff015 | 8566 | |
652884fe | 8567 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 8568 | perf_group_detach(event); |
889ff015 | 8569 | list_del_event(event, ctx); |
652884fe | 8570 | raw_spin_unlock_irq(&ctx->lock); |
889ff015 FW |
8571 | free_event(event); |
8572 | } | |
8573 | ||
bbbee908 | 8574 | /* |
652884fe | 8575 | * Free an unexposed, unused context as created by inheritance by |
8dc85d54 | 8576 | * perf_event_init_task below, used by fork() in case of fail. |
652884fe PZ |
8577 | * |
8578 | * Not all locks are strictly required, but take them anyway to be nice and | |
8579 | * help out with the lockdep assertions. | |
bbbee908 | 8580 | */ |
cdd6c482 | 8581 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 8582 | { |
8dc85d54 | 8583 | struct perf_event_context *ctx; |
cdd6c482 | 8584 | struct perf_event *event, *tmp; |
8dc85d54 | 8585 | int ctxn; |
bbbee908 | 8586 | |
8dc85d54 PZ |
8587 | for_each_task_context_nr(ctxn) { |
8588 | ctx = task->perf_event_ctxp[ctxn]; | |
8589 | if (!ctx) | |
8590 | continue; | |
bbbee908 | 8591 | |
8dc85d54 | 8592 | mutex_lock(&ctx->mutex); |
bbbee908 | 8593 | again: |
8dc85d54 PZ |
8594 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
8595 | group_entry) | |
8596 | perf_free_event(event, ctx); | |
bbbee908 | 8597 | |
8dc85d54 PZ |
8598 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
8599 | group_entry) | |
8600 | perf_free_event(event, ctx); | |
bbbee908 | 8601 | |
8dc85d54 PZ |
8602 | if (!list_empty(&ctx->pinned_groups) || |
8603 | !list_empty(&ctx->flexible_groups)) | |
8604 | goto again; | |
bbbee908 | 8605 | |
8dc85d54 | 8606 | mutex_unlock(&ctx->mutex); |
bbbee908 | 8607 | |
8dc85d54 PZ |
8608 | put_ctx(ctx); |
8609 | } | |
889ff015 FW |
8610 | } |
8611 | ||
4e231c79 PZ |
8612 | void perf_event_delayed_put(struct task_struct *task) |
8613 | { | |
8614 | int ctxn; | |
8615 | ||
8616 | for_each_task_context_nr(ctxn) | |
8617 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
8618 | } | |
8619 | ||
97dee4f3 PZ |
8620 | /* |
8621 | * inherit a event from parent task to child task: | |
8622 | */ | |
8623 | static struct perf_event * | |
8624 | inherit_event(struct perf_event *parent_event, | |
8625 | struct task_struct *parent, | |
8626 | struct perf_event_context *parent_ctx, | |
8627 | struct task_struct *child, | |
8628 | struct perf_event *group_leader, | |
8629 | struct perf_event_context *child_ctx) | |
8630 | { | |
1929def9 | 8631 | enum perf_event_active_state parent_state = parent_event->state; |
97dee4f3 | 8632 | struct perf_event *child_event; |
cee010ec | 8633 | unsigned long flags; |
97dee4f3 PZ |
8634 | |
8635 | /* | |
8636 | * Instead of creating recursive hierarchies of events, | |
8637 | * we link inherited events back to the original parent, | |
8638 | * which has a filp for sure, which we use as the reference | |
8639 | * count: | |
8640 | */ | |
8641 | if (parent_event->parent) | |
8642 | parent_event = parent_event->parent; | |
8643 | ||
8644 | child_event = perf_event_alloc(&parent_event->attr, | |
8645 | parent_event->cpu, | |
d580ff86 | 8646 | child, |
97dee4f3 | 8647 | group_leader, parent_event, |
79dff51e | 8648 | NULL, NULL, -1); |
97dee4f3 PZ |
8649 | if (IS_ERR(child_event)) |
8650 | return child_event; | |
a6fa941d | 8651 | |
fadfe7be JO |
8652 | if (is_orphaned_event(parent_event) || |
8653 | !atomic_long_inc_not_zero(&parent_event->refcount)) { | |
a6fa941d AV |
8654 | free_event(child_event); |
8655 | return NULL; | |
8656 | } | |
8657 | ||
97dee4f3 PZ |
8658 | get_ctx(child_ctx); |
8659 | ||
8660 | /* | |
8661 | * Make the child state follow the state of the parent event, | |
8662 | * not its attr.disabled bit. We hold the parent's mutex, | |
8663 | * so we won't race with perf_event_{en, dis}able_family. | |
8664 | */ | |
1929def9 | 8665 | if (parent_state >= PERF_EVENT_STATE_INACTIVE) |
97dee4f3 PZ |
8666 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
8667 | else | |
8668 | child_event->state = PERF_EVENT_STATE_OFF; | |
8669 | ||
8670 | if (parent_event->attr.freq) { | |
8671 | u64 sample_period = parent_event->hw.sample_period; | |
8672 | struct hw_perf_event *hwc = &child_event->hw; | |
8673 | ||
8674 | hwc->sample_period = sample_period; | |
8675 | hwc->last_period = sample_period; | |
8676 | ||
8677 | local64_set(&hwc->period_left, sample_period); | |
8678 | } | |
8679 | ||
8680 | child_event->ctx = child_ctx; | |
8681 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
8682 | child_event->overflow_handler_context |
8683 | = parent_event->overflow_handler_context; | |
97dee4f3 | 8684 | |
614b6780 TG |
8685 | /* |
8686 | * Precalculate sample_data sizes | |
8687 | */ | |
8688 | perf_event__header_size(child_event); | |
6844c09d | 8689 | perf_event__id_header_size(child_event); |
614b6780 | 8690 | |
97dee4f3 PZ |
8691 | /* |
8692 | * Link it up in the child's context: | |
8693 | */ | |
cee010ec | 8694 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 8695 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 8696 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 8697 | |
97dee4f3 PZ |
8698 | /* |
8699 | * Link this into the parent event's child list | |
8700 | */ | |
8701 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
8702 | mutex_lock(&parent_event->child_mutex); | |
8703 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
8704 | mutex_unlock(&parent_event->child_mutex); | |
8705 | ||
8706 | return child_event; | |
8707 | } | |
8708 | ||
8709 | static int inherit_group(struct perf_event *parent_event, | |
8710 | struct task_struct *parent, | |
8711 | struct perf_event_context *parent_ctx, | |
8712 | struct task_struct *child, | |
8713 | struct perf_event_context *child_ctx) | |
8714 | { | |
8715 | struct perf_event *leader; | |
8716 | struct perf_event *sub; | |
8717 | struct perf_event *child_ctr; | |
8718 | ||
8719 | leader = inherit_event(parent_event, parent, parent_ctx, | |
8720 | child, NULL, child_ctx); | |
8721 | if (IS_ERR(leader)) | |
8722 | return PTR_ERR(leader); | |
8723 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
8724 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
8725 | child, leader, child_ctx); | |
8726 | if (IS_ERR(child_ctr)) | |
8727 | return PTR_ERR(child_ctr); | |
8728 | } | |
8729 | return 0; | |
889ff015 FW |
8730 | } |
8731 | ||
8732 | static int | |
8733 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
8734 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 8735 | struct task_struct *child, int ctxn, |
889ff015 FW |
8736 | int *inherited_all) |
8737 | { | |
8738 | int ret; | |
8dc85d54 | 8739 | struct perf_event_context *child_ctx; |
889ff015 FW |
8740 | |
8741 | if (!event->attr.inherit) { | |
8742 | *inherited_all = 0; | |
8743 | return 0; | |
bbbee908 PZ |
8744 | } |
8745 | ||
fe4b04fa | 8746 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
8747 | if (!child_ctx) { |
8748 | /* | |
8749 | * This is executed from the parent task context, so | |
8750 | * inherit events that have been marked for cloning. | |
8751 | * First allocate and initialize a context for the | |
8752 | * child. | |
8753 | */ | |
bbbee908 | 8754 | |
734df5ab | 8755 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
8756 | if (!child_ctx) |
8757 | return -ENOMEM; | |
bbbee908 | 8758 | |
8dc85d54 | 8759 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
8760 | } |
8761 | ||
8762 | ret = inherit_group(event, parent, parent_ctx, | |
8763 | child, child_ctx); | |
8764 | ||
8765 | if (ret) | |
8766 | *inherited_all = 0; | |
8767 | ||
8768 | return ret; | |
bbbee908 PZ |
8769 | } |
8770 | ||
9b51f66d | 8771 | /* |
cdd6c482 | 8772 | * Initialize the perf_event context in task_struct |
9b51f66d | 8773 | */ |
985c8dcb | 8774 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 8775 | { |
889ff015 | 8776 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
8777 | struct perf_event_context *cloned_ctx; |
8778 | struct perf_event *event; | |
9b51f66d | 8779 | struct task_struct *parent = current; |
564c2b21 | 8780 | int inherited_all = 1; |
dddd3379 | 8781 | unsigned long flags; |
6ab423e0 | 8782 | int ret = 0; |
9b51f66d | 8783 | |
8dc85d54 | 8784 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
8785 | return 0; |
8786 | ||
ad3a37de | 8787 | /* |
25346b93 PM |
8788 | * If the parent's context is a clone, pin it so it won't get |
8789 | * swapped under us. | |
ad3a37de | 8790 | */ |
8dc85d54 | 8791 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
8792 | if (!parent_ctx) |
8793 | return 0; | |
25346b93 | 8794 | |
ad3a37de PM |
8795 | /* |
8796 | * No need to check if parent_ctx != NULL here; since we saw | |
8797 | * it non-NULL earlier, the only reason for it to become NULL | |
8798 | * is if we exit, and since we're currently in the middle of | |
8799 | * a fork we can't be exiting at the same time. | |
8800 | */ | |
ad3a37de | 8801 | |
9b51f66d IM |
8802 | /* |
8803 | * Lock the parent list. No need to lock the child - not PID | |
8804 | * hashed yet and not running, so nobody can access it. | |
8805 | */ | |
d859e29f | 8806 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
8807 | |
8808 | /* | |
8809 | * We dont have to disable NMIs - we are only looking at | |
8810 | * the list, not manipulating it: | |
8811 | */ | |
889ff015 | 8812 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
8813 | ret = inherit_task_group(event, parent, parent_ctx, |
8814 | child, ctxn, &inherited_all); | |
889ff015 FW |
8815 | if (ret) |
8816 | break; | |
8817 | } | |
b93f7978 | 8818 | |
dddd3379 TG |
8819 | /* |
8820 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
8821 | * to allocations, but we need to prevent rotation because | |
8822 | * rotate_ctx() will change the list from interrupt context. | |
8823 | */ | |
8824 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
8825 | parent_ctx->rotate_disable = 1; | |
8826 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
8827 | ||
889ff015 | 8828 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
8829 | ret = inherit_task_group(event, parent, parent_ctx, |
8830 | child, ctxn, &inherited_all); | |
889ff015 | 8831 | if (ret) |
9b51f66d | 8832 | break; |
564c2b21 PM |
8833 | } |
8834 | ||
dddd3379 TG |
8835 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
8836 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 8837 | |
8dc85d54 | 8838 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 8839 | |
05cbaa28 | 8840 | if (child_ctx && inherited_all) { |
564c2b21 PM |
8841 | /* |
8842 | * Mark the child context as a clone of the parent | |
8843 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
8844 | * |
8845 | * Note that if the parent is a clone, the holding of | |
8846 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 8847 | */ |
c5ed5145 | 8848 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
8849 | if (cloned_ctx) { |
8850 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 8851 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
8852 | } else { |
8853 | child_ctx->parent_ctx = parent_ctx; | |
8854 | child_ctx->parent_gen = parent_ctx->generation; | |
8855 | } | |
8856 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
8857 | } |
8858 | ||
c5ed5145 | 8859 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 8860 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 8861 | |
25346b93 | 8862 | perf_unpin_context(parent_ctx); |
fe4b04fa | 8863 | put_ctx(parent_ctx); |
ad3a37de | 8864 | |
6ab423e0 | 8865 | return ret; |
9b51f66d IM |
8866 | } |
8867 | ||
8dc85d54 PZ |
8868 | /* |
8869 | * Initialize the perf_event context in task_struct | |
8870 | */ | |
8871 | int perf_event_init_task(struct task_struct *child) | |
8872 | { | |
8873 | int ctxn, ret; | |
8874 | ||
8550d7cb ON |
8875 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
8876 | mutex_init(&child->perf_event_mutex); | |
8877 | INIT_LIST_HEAD(&child->perf_event_list); | |
8878 | ||
8dc85d54 PZ |
8879 | for_each_task_context_nr(ctxn) { |
8880 | ret = perf_event_init_context(child, ctxn); | |
6c72e350 PZ |
8881 | if (ret) { |
8882 | perf_event_free_task(child); | |
8dc85d54 | 8883 | return ret; |
6c72e350 | 8884 | } |
8dc85d54 PZ |
8885 | } |
8886 | ||
8887 | return 0; | |
8888 | } | |
8889 | ||
220b140b PM |
8890 | static void __init perf_event_init_all_cpus(void) |
8891 | { | |
b28ab83c | 8892 | struct swevent_htable *swhash; |
220b140b | 8893 | int cpu; |
220b140b PM |
8894 | |
8895 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
8896 | swhash = &per_cpu(swevent_htable, cpu); |
8897 | mutex_init(&swhash->hlist_mutex); | |
2fde4f94 | 8898 | INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu)); |
220b140b PM |
8899 | } |
8900 | } | |
8901 | ||
0db0628d | 8902 | static void perf_event_init_cpu(int cpu) |
0793a61d | 8903 | { |
108b02cf | 8904 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 8905 | |
b28ab83c | 8906 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8907 | swhash->online = true; |
4536e4d1 | 8908 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
8909 | struct swevent_hlist *hlist; |
8910 | ||
b28ab83c PZ |
8911 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
8912 | WARN_ON(!hlist); | |
8913 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 8914 | } |
b28ab83c | 8915 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
8916 | } |
8917 | ||
c277443c | 8918 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
108b02cf | 8919 | static void __perf_event_exit_context(void *__info) |
0793a61d | 8920 | { |
226424ee | 8921 | struct remove_event re = { .detach_group = true }; |
108b02cf | 8922 | struct perf_event_context *ctx = __info; |
0793a61d | 8923 | |
e3703f8c | 8924 | rcu_read_lock(); |
46ce0fe9 PZ |
8925 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
8926 | __perf_remove_from_context(&re); | |
e3703f8c | 8927 | rcu_read_unlock(); |
0793a61d | 8928 | } |
108b02cf PZ |
8929 | |
8930 | static void perf_event_exit_cpu_context(int cpu) | |
8931 | { | |
8932 | struct perf_event_context *ctx; | |
8933 | struct pmu *pmu; | |
8934 | int idx; | |
8935 | ||
8936 | idx = srcu_read_lock(&pmus_srcu); | |
8937 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 8938 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
8939 | |
8940 | mutex_lock(&ctx->mutex); | |
8941 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
8942 | mutex_unlock(&ctx->mutex); | |
8943 | } | |
8944 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
8945 | } |
8946 | ||
cdd6c482 | 8947 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 8948 | { |
b28ab83c | 8949 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 8950 | |
e3703f8c PZ |
8951 | perf_event_exit_cpu_context(cpu); |
8952 | ||
b28ab83c | 8953 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8954 | swhash->online = false; |
b28ab83c PZ |
8955 | swevent_hlist_release(swhash); |
8956 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
8957 | } |
8958 | #else | |
cdd6c482 | 8959 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
8960 | #endif |
8961 | ||
c277443c PZ |
8962 | static int |
8963 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
8964 | { | |
8965 | int cpu; | |
8966 | ||
8967 | for_each_online_cpu(cpu) | |
8968 | perf_event_exit_cpu(cpu); | |
8969 | ||
8970 | return NOTIFY_OK; | |
8971 | } | |
8972 | ||
8973 | /* | |
8974 | * Run the perf reboot notifier at the very last possible moment so that | |
8975 | * the generic watchdog code runs as long as possible. | |
8976 | */ | |
8977 | static struct notifier_block perf_reboot_notifier = { | |
8978 | .notifier_call = perf_reboot, | |
8979 | .priority = INT_MIN, | |
8980 | }; | |
8981 | ||
0db0628d | 8982 | static int |
0793a61d TG |
8983 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
8984 | { | |
8985 | unsigned int cpu = (long)hcpu; | |
8986 | ||
4536e4d1 | 8987 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
8988 | |
8989 | case CPU_UP_PREPARE: | |
5e11637e | 8990 | case CPU_DOWN_FAILED: |
cdd6c482 | 8991 | perf_event_init_cpu(cpu); |
0793a61d TG |
8992 | break; |
8993 | ||
5e11637e | 8994 | case CPU_UP_CANCELED: |
0793a61d | 8995 | case CPU_DOWN_PREPARE: |
cdd6c482 | 8996 | perf_event_exit_cpu(cpu); |
0793a61d | 8997 | break; |
0793a61d TG |
8998 | default: |
8999 | break; | |
9000 | } | |
9001 | ||
9002 | return NOTIFY_OK; | |
9003 | } | |
9004 | ||
cdd6c482 | 9005 | void __init perf_event_init(void) |
0793a61d | 9006 | { |
3c502e7a JW |
9007 | int ret; |
9008 | ||
2e80a82a PZ |
9009 | idr_init(&pmu_idr); |
9010 | ||
220b140b | 9011 | perf_event_init_all_cpus(); |
b0a873eb | 9012 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
9013 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
9014 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
9015 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
9016 | perf_tp_register(); |
9017 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 9018 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
9019 | |
9020 | ret = init_hw_breakpoint(); | |
9021 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
9022 | |
9023 | /* do not patch jump label more than once per second */ | |
9024 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
9025 | |
9026 | /* | |
9027 | * Build time assertion that we keep the data_head at the intended | |
9028 | * location. IOW, validation we got the __reserved[] size right. | |
9029 | */ | |
9030 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
9031 | != 1024); | |
0793a61d | 9032 | } |
abe43400 | 9033 | |
fd979c01 CS |
9034 | ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr, |
9035 | char *page) | |
9036 | { | |
9037 | struct perf_pmu_events_attr *pmu_attr = | |
9038 | container_of(attr, struct perf_pmu_events_attr, attr); | |
9039 | ||
9040 | if (pmu_attr->event_str) | |
9041 | return sprintf(page, "%s\n", pmu_attr->event_str); | |
9042 | ||
9043 | return 0; | |
9044 | } | |
9045 | ||
abe43400 PZ |
9046 | static int __init perf_event_sysfs_init(void) |
9047 | { | |
9048 | struct pmu *pmu; | |
9049 | int ret; | |
9050 | ||
9051 | mutex_lock(&pmus_lock); | |
9052 | ||
9053 | ret = bus_register(&pmu_bus); | |
9054 | if (ret) | |
9055 | goto unlock; | |
9056 | ||
9057 | list_for_each_entry(pmu, &pmus, entry) { | |
9058 | if (!pmu->name || pmu->type < 0) | |
9059 | continue; | |
9060 | ||
9061 | ret = pmu_dev_alloc(pmu); | |
9062 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
9063 | } | |
9064 | pmu_bus_running = 1; | |
9065 | ret = 0; | |
9066 | ||
9067 | unlock: | |
9068 | mutex_unlock(&pmus_lock); | |
9069 | ||
9070 | return ret; | |
9071 | } | |
9072 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
9073 | |
9074 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
9075 | static struct cgroup_subsys_state * |
9076 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
9077 | { |
9078 | struct perf_cgroup *jc; | |
e5d1367f | 9079 | |
1b15d055 | 9080 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
9081 | if (!jc) |
9082 | return ERR_PTR(-ENOMEM); | |
9083 | ||
e5d1367f SE |
9084 | jc->info = alloc_percpu(struct perf_cgroup_info); |
9085 | if (!jc->info) { | |
9086 | kfree(jc); | |
9087 | return ERR_PTR(-ENOMEM); | |
9088 | } | |
9089 | ||
e5d1367f SE |
9090 | return &jc->css; |
9091 | } | |
9092 | ||
eb95419b | 9093 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 9094 | { |
eb95419b TH |
9095 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
9096 | ||
e5d1367f SE |
9097 | free_percpu(jc->info); |
9098 | kfree(jc); | |
9099 | } | |
9100 | ||
9101 | static int __perf_cgroup_move(void *info) | |
9102 | { | |
9103 | struct task_struct *task = info; | |
9104 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
9105 | return 0; | |
9106 | } | |
9107 | ||
eb95419b TH |
9108 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
9109 | struct cgroup_taskset *tset) | |
e5d1367f | 9110 | { |
bb9d97b6 TH |
9111 | struct task_struct *task; |
9112 | ||
924f0d9a | 9113 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 9114 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
9115 | } |
9116 | ||
eb95419b TH |
9117 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
9118 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 9119 | struct task_struct *task) |
e5d1367f SE |
9120 | { |
9121 | /* | |
9122 | * cgroup_exit() is called in the copy_process() failure path. | |
9123 | * Ignore this case since the task hasn't ran yet, this avoids | |
9124 | * trying to poke a half freed task state from generic code. | |
9125 | */ | |
9126 | if (!(task->flags & PF_EXITING)) | |
9127 | return; | |
9128 | ||
bb9d97b6 | 9129 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
9130 | } |
9131 | ||
073219e9 | 9132 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
9133 | .css_alloc = perf_cgroup_css_alloc, |
9134 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 9135 | .exit = perf_cgroup_exit, |
bb9d97b6 | 9136 | .attach = perf_cgroup_attach, |
e5d1367f SE |
9137 | }; |
9138 | #endif /* CONFIG_CGROUP_PERF */ |