]>
Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
cdd6c482 | 37 | #include <linux/perf_event.h> |
6fb2915d | 38 | #include <linux/ftrace_event.h> |
3c502e7a | 39 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 40 | #include <linux/mm_types.h> |
877c6856 | 41 | #include <linux/cgroup.h> |
c464c76e | 42 | #include <linux/module.h> |
f972eb63 | 43 | #include <linux/mman.h> |
b3f20785 | 44 | #include <linux/compat.h> |
0793a61d | 45 | |
76369139 FW |
46 | #include "internal.h" |
47 | ||
4e193bd4 TB |
48 | #include <asm/irq_regs.h> |
49 | ||
fadfe7be JO |
50 | static struct workqueue_struct *perf_wq; |
51 | ||
fe4b04fa | 52 | struct remote_function_call { |
e7e7ee2e IM |
53 | struct task_struct *p; |
54 | int (*func)(void *info); | |
55 | void *info; | |
56 | int ret; | |
fe4b04fa PZ |
57 | }; |
58 | ||
59 | static void remote_function(void *data) | |
60 | { | |
61 | struct remote_function_call *tfc = data; | |
62 | struct task_struct *p = tfc->p; | |
63 | ||
64 | if (p) { | |
65 | tfc->ret = -EAGAIN; | |
66 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
67 | return; | |
68 | } | |
69 | ||
70 | tfc->ret = tfc->func(tfc->info); | |
71 | } | |
72 | ||
73 | /** | |
74 | * task_function_call - call a function on the cpu on which a task runs | |
75 | * @p: the task to evaluate | |
76 | * @func: the function to be called | |
77 | * @info: the function call argument | |
78 | * | |
79 | * Calls the function @func when the task is currently running. This might | |
80 | * be on the current CPU, which just calls the function directly | |
81 | * | |
82 | * returns: @func return value, or | |
83 | * -ESRCH - when the process isn't running | |
84 | * -EAGAIN - when the process moved away | |
85 | */ | |
86 | static int | |
87 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
88 | { | |
89 | struct remote_function_call data = { | |
e7e7ee2e IM |
90 | .p = p, |
91 | .func = func, | |
92 | .info = info, | |
93 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
94 | }; |
95 | ||
96 | if (task_curr(p)) | |
97 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
98 | ||
99 | return data.ret; | |
100 | } | |
101 | ||
102 | /** | |
103 | * cpu_function_call - call a function on the cpu | |
104 | * @func: the function to be called | |
105 | * @info: the function call argument | |
106 | * | |
107 | * Calls the function @func on the remote cpu. | |
108 | * | |
109 | * returns: @func return value or -ENXIO when the cpu is offline | |
110 | */ | |
111 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
112 | { | |
113 | struct remote_function_call data = { | |
e7e7ee2e IM |
114 | .p = NULL, |
115 | .func = func, | |
116 | .info = info, | |
117 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
118 | }; |
119 | ||
120 | smp_call_function_single(cpu, remote_function, &data, 1); | |
121 | ||
122 | return data.ret; | |
123 | } | |
124 | ||
f8697762 JO |
125 | #define EVENT_OWNER_KERNEL ((void *) -1) |
126 | ||
127 | static bool is_kernel_event(struct perf_event *event) | |
128 | { | |
129 | return event->owner == EVENT_OWNER_KERNEL; | |
130 | } | |
131 | ||
e5d1367f SE |
132 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
133 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
134 | PERF_FLAG_PID_CGROUP |\ |
135 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 136 | |
bce38cd5 SE |
137 | /* |
138 | * branch priv levels that need permission checks | |
139 | */ | |
140 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
141 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
142 | PERF_SAMPLE_BRANCH_HV) | |
143 | ||
0b3fcf17 SE |
144 | enum event_type_t { |
145 | EVENT_FLEXIBLE = 0x1, | |
146 | EVENT_PINNED = 0x2, | |
147 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
148 | }; | |
149 | ||
e5d1367f SE |
150 | /* |
151 | * perf_sched_events : >0 events exist | |
152 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
153 | */ | |
c5905afb | 154 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 155 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
ba532500 | 156 | static DEFINE_PER_CPU(int, perf_sched_cb_usages); |
e5d1367f | 157 | |
cdd6c482 IM |
158 | static atomic_t nr_mmap_events __read_mostly; |
159 | static atomic_t nr_comm_events __read_mostly; | |
160 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 161 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 162 | |
108b02cf PZ |
163 | static LIST_HEAD(pmus); |
164 | static DEFINE_MUTEX(pmus_lock); | |
165 | static struct srcu_struct pmus_srcu; | |
166 | ||
0764771d | 167 | /* |
cdd6c482 | 168 | * perf event paranoia level: |
0fbdea19 IM |
169 | * -1 - not paranoid at all |
170 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 171 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 172 | * 2 - disallow kernel profiling for unpriv |
0764771d | 173 | */ |
cdd6c482 | 174 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 175 | |
20443384 FW |
176 | /* Minimum for 512 kiB + 1 user control page */ |
177 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
178 | |
179 | /* | |
cdd6c482 | 180 | * max perf event sample rate |
df58ab24 | 181 | */ |
14c63f17 DH |
182 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
183 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
184 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
185 | ||
186 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
187 | ||
188 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
189 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
190 | ||
d9494cb4 PZ |
191 | static int perf_sample_allowed_ns __read_mostly = |
192 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
193 | |
194 | void update_perf_cpu_limits(void) | |
195 | { | |
196 | u64 tmp = perf_sample_period_ns; | |
197 | ||
198 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 199 | do_div(tmp, 100); |
d9494cb4 | 200 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 201 | } |
163ec435 | 202 | |
9e630205 SE |
203 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
204 | ||
163ec435 PZ |
205 | int perf_proc_update_handler(struct ctl_table *table, int write, |
206 | void __user *buffer, size_t *lenp, | |
207 | loff_t *ppos) | |
208 | { | |
723478c8 | 209 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
210 | |
211 | if (ret || !write) | |
212 | return ret; | |
213 | ||
214 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
215 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
216 | update_perf_cpu_limits(); | |
217 | ||
218 | return 0; | |
219 | } | |
220 | ||
221 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
222 | ||
223 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
224 | void __user *buffer, size_t *lenp, | |
225 | loff_t *ppos) | |
226 | { | |
227 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
228 | ||
229 | if (ret || !write) | |
230 | return ret; | |
231 | ||
232 | update_perf_cpu_limits(); | |
163ec435 PZ |
233 | |
234 | return 0; | |
235 | } | |
1ccd1549 | 236 | |
14c63f17 DH |
237 | /* |
238 | * perf samples are done in some very critical code paths (NMIs). | |
239 | * If they take too much CPU time, the system can lock up and not | |
240 | * get any real work done. This will drop the sample rate when | |
241 | * we detect that events are taking too long. | |
242 | */ | |
243 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 244 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 245 | |
6a02ad66 | 246 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 247 | { |
6a02ad66 | 248 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 249 | u64 avg_local_sample_len; |
e5302920 | 250 | u64 local_samples_len; |
6a02ad66 | 251 | |
4a32fea9 | 252 | local_samples_len = __this_cpu_read(running_sample_length); |
6a02ad66 PZ |
253 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; |
254 | ||
255 | printk_ratelimited(KERN_WARNING | |
256 | "perf interrupt took too long (%lld > %lld), lowering " | |
257 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 258 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
259 | sysctl_perf_event_sample_rate); |
260 | } | |
261 | ||
262 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
263 | ||
264 | void perf_sample_event_took(u64 sample_len_ns) | |
265 | { | |
d9494cb4 | 266 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
267 | u64 avg_local_sample_len; |
268 | u64 local_samples_len; | |
14c63f17 | 269 | |
d9494cb4 | 270 | if (allowed_ns == 0) |
14c63f17 DH |
271 | return; |
272 | ||
273 | /* decay the counter by 1 average sample */ | |
4a32fea9 | 274 | local_samples_len = __this_cpu_read(running_sample_length); |
14c63f17 DH |
275 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; |
276 | local_samples_len += sample_len_ns; | |
4a32fea9 | 277 | __this_cpu_write(running_sample_length, local_samples_len); |
14c63f17 DH |
278 | |
279 | /* | |
280 | * note: this will be biased artifically low until we have | |
281 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
282 | * from having to maintain a count. | |
283 | */ | |
284 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
285 | ||
d9494cb4 | 286 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
287 | return; |
288 | ||
289 | if (max_samples_per_tick <= 1) | |
290 | return; | |
291 | ||
292 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
293 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
294 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
295 | ||
14c63f17 | 296 | update_perf_cpu_limits(); |
6a02ad66 | 297 | |
cd578abb PZ |
298 | if (!irq_work_queue(&perf_duration_work)) { |
299 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
300 | "kernel.perf_event_max_sample_rate to %d\n", | |
301 | avg_local_sample_len, allowed_ns >> 1, | |
302 | sysctl_perf_event_sample_rate); | |
303 | } | |
14c63f17 DH |
304 | } |
305 | ||
cdd6c482 | 306 | static atomic64_t perf_event_id; |
a96bbc16 | 307 | |
0b3fcf17 SE |
308 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
309 | enum event_type_t event_type); | |
310 | ||
311 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
312 | enum event_type_t event_type, |
313 | struct task_struct *task); | |
314 | ||
315 | static void update_context_time(struct perf_event_context *ctx); | |
316 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 317 | |
cdd6c482 | 318 | void __weak perf_event_print_debug(void) { } |
0793a61d | 319 | |
84c79910 | 320 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 321 | { |
84c79910 | 322 | return "pmu"; |
0793a61d TG |
323 | } |
324 | ||
0b3fcf17 SE |
325 | static inline u64 perf_clock(void) |
326 | { | |
327 | return local_clock(); | |
328 | } | |
329 | ||
e5d1367f SE |
330 | static inline struct perf_cpu_context * |
331 | __get_cpu_context(struct perf_event_context *ctx) | |
332 | { | |
333 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
334 | } | |
335 | ||
facc4307 PZ |
336 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
337 | struct perf_event_context *ctx) | |
338 | { | |
339 | raw_spin_lock(&cpuctx->ctx.lock); | |
340 | if (ctx) | |
341 | raw_spin_lock(&ctx->lock); | |
342 | } | |
343 | ||
344 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
345 | struct perf_event_context *ctx) | |
346 | { | |
347 | if (ctx) | |
348 | raw_spin_unlock(&ctx->lock); | |
349 | raw_spin_unlock(&cpuctx->ctx.lock); | |
350 | } | |
351 | ||
e5d1367f SE |
352 | #ifdef CONFIG_CGROUP_PERF |
353 | ||
877c6856 LZ |
354 | /* |
355 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
356 | * This is a per-cpu dynamically allocated data structure. | |
357 | */ | |
358 | struct perf_cgroup_info { | |
359 | u64 time; | |
360 | u64 timestamp; | |
361 | }; | |
362 | ||
363 | struct perf_cgroup { | |
364 | struct cgroup_subsys_state css; | |
86e213e1 | 365 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
366 | }; |
367 | ||
3f7cce3c SE |
368 | /* |
369 | * Must ensure cgroup is pinned (css_get) before calling | |
370 | * this function. In other words, we cannot call this function | |
371 | * if there is no cgroup event for the current CPU context. | |
372 | */ | |
e5d1367f SE |
373 | static inline struct perf_cgroup * |
374 | perf_cgroup_from_task(struct task_struct *task) | |
375 | { | |
073219e9 | 376 | return container_of(task_css(task, perf_event_cgrp_id), |
8af01f56 | 377 | struct perf_cgroup, css); |
e5d1367f SE |
378 | } |
379 | ||
380 | static inline bool | |
381 | perf_cgroup_match(struct perf_event *event) | |
382 | { | |
383 | struct perf_event_context *ctx = event->ctx; | |
384 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
385 | ||
ef824fa1 TH |
386 | /* @event doesn't care about cgroup */ |
387 | if (!event->cgrp) | |
388 | return true; | |
389 | ||
390 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
391 | if (!cpuctx->cgrp) | |
392 | return false; | |
393 | ||
394 | /* | |
395 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
396 | * also enabled for all its descendant cgroups. If @cpuctx's | |
397 | * cgroup is a descendant of @event's (the test covers identity | |
398 | * case), it's a match. | |
399 | */ | |
400 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
401 | event->cgrp->css.cgroup); | |
e5d1367f SE |
402 | } |
403 | ||
e5d1367f SE |
404 | static inline void perf_detach_cgroup(struct perf_event *event) |
405 | { | |
4e2ba650 | 406 | css_put(&event->cgrp->css); |
e5d1367f SE |
407 | event->cgrp = NULL; |
408 | } | |
409 | ||
410 | static inline int is_cgroup_event(struct perf_event *event) | |
411 | { | |
412 | return event->cgrp != NULL; | |
413 | } | |
414 | ||
415 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
416 | { | |
417 | struct perf_cgroup_info *t; | |
418 | ||
419 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
420 | return t->time; | |
421 | } | |
422 | ||
423 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
424 | { | |
425 | struct perf_cgroup_info *info; | |
426 | u64 now; | |
427 | ||
428 | now = perf_clock(); | |
429 | ||
430 | info = this_cpu_ptr(cgrp->info); | |
431 | ||
432 | info->time += now - info->timestamp; | |
433 | info->timestamp = now; | |
434 | } | |
435 | ||
436 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
437 | { | |
438 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
439 | if (cgrp_out) | |
440 | __update_cgrp_time(cgrp_out); | |
441 | } | |
442 | ||
443 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
444 | { | |
3f7cce3c SE |
445 | struct perf_cgroup *cgrp; |
446 | ||
e5d1367f | 447 | /* |
3f7cce3c SE |
448 | * ensure we access cgroup data only when needed and |
449 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 450 | */ |
3f7cce3c | 451 | if (!is_cgroup_event(event)) |
e5d1367f SE |
452 | return; |
453 | ||
3f7cce3c SE |
454 | cgrp = perf_cgroup_from_task(current); |
455 | /* | |
456 | * Do not update time when cgroup is not active | |
457 | */ | |
458 | if (cgrp == event->cgrp) | |
459 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
460 | } |
461 | ||
462 | static inline void | |
3f7cce3c SE |
463 | perf_cgroup_set_timestamp(struct task_struct *task, |
464 | struct perf_event_context *ctx) | |
e5d1367f SE |
465 | { |
466 | struct perf_cgroup *cgrp; | |
467 | struct perf_cgroup_info *info; | |
468 | ||
3f7cce3c SE |
469 | /* |
470 | * ctx->lock held by caller | |
471 | * ensure we do not access cgroup data | |
472 | * unless we have the cgroup pinned (css_get) | |
473 | */ | |
474 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
475 | return; |
476 | ||
477 | cgrp = perf_cgroup_from_task(task); | |
478 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 479 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
480 | } |
481 | ||
482 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
483 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
484 | ||
485 | /* | |
486 | * reschedule events based on the cgroup constraint of task. | |
487 | * | |
488 | * mode SWOUT : schedule out everything | |
489 | * mode SWIN : schedule in based on cgroup for next | |
490 | */ | |
491 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
492 | { | |
493 | struct perf_cpu_context *cpuctx; | |
494 | struct pmu *pmu; | |
495 | unsigned long flags; | |
496 | ||
497 | /* | |
498 | * disable interrupts to avoid geting nr_cgroup | |
499 | * changes via __perf_event_disable(). Also | |
500 | * avoids preemption. | |
501 | */ | |
502 | local_irq_save(flags); | |
503 | ||
504 | /* | |
505 | * we reschedule only in the presence of cgroup | |
506 | * constrained events. | |
507 | */ | |
508 | rcu_read_lock(); | |
509 | ||
510 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 511 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
512 | if (cpuctx->unique_pmu != pmu) |
513 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 514 | |
e5d1367f SE |
515 | /* |
516 | * perf_cgroup_events says at least one | |
517 | * context on this CPU has cgroup events. | |
518 | * | |
519 | * ctx->nr_cgroups reports the number of cgroup | |
520 | * events for a context. | |
521 | */ | |
522 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
523 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
524 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
525 | |
526 | if (mode & PERF_CGROUP_SWOUT) { | |
527 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
528 | /* | |
529 | * must not be done before ctxswout due | |
530 | * to event_filter_match() in event_sched_out() | |
531 | */ | |
532 | cpuctx->cgrp = NULL; | |
533 | } | |
534 | ||
535 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 536 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
537 | /* |
538 | * set cgrp before ctxsw in to allow | |
539 | * event_filter_match() to not have to pass | |
540 | * task around | |
e5d1367f SE |
541 | */ |
542 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
543 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
544 | } | |
facc4307 PZ |
545 | perf_pmu_enable(cpuctx->ctx.pmu); |
546 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 547 | } |
e5d1367f SE |
548 | } |
549 | ||
550 | rcu_read_unlock(); | |
551 | ||
552 | local_irq_restore(flags); | |
553 | } | |
554 | ||
a8d757ef SE |
555 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
556 | struct task_struct *next) | |
e5d1367f | 557 | { |
a8d757ef SE |
558 | struct perf_cgroup *cgrp1; |
559 | struct perf_cgroup *cgrp2 = NULL; | |
560 | ||
561 | /* | |
562 | * we come here when we know perf_cgroup_events > 0 | |
563 | */ | |
564 | cgrp1 = perf_cgroup_from_task(task); | |
565 | ||
566 | /* | |
567 | * next is NULL when called from perf_event_enable_on_exec() | |
568 | * that will systematically cause a cgroup_switch() | |
569 | */ | |
570 | if (next) | |
571 | cgrp2 = perf_cgroup_from_task(next); | |
572 | ||
573 | /* | |
574 | * only schedule out current cgroup events if we know | |
575 | * that we are switching to a different cgroup. Otherwise, | |
576 | * do no touch the cgroup events. | |
577 | */ | |
578 | if (cgrp1 != cgrp2) | |
579 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
580 | } |
581 | ||
a8d757ef SE |
582 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
583 | struct task_struct *task) | |
e5d1367f | 584 | { |
a8d757ef SE |
585 | struct perf_cgroup *cgrp1; |
586 | struct perf_cgroup *cgrp2 = NULL; | |
587 | ||
588 | /* | |
589 | * we come here when we know perf_cgroup_events > 0 | |
590 | */ | |
591 | cgrp1 = perf_cgroup_from_task(task); | |
592 | ||
593 | /* prev can never be NULL */ | |
594 | cgrp2 = perf_cgroup_from_task(prev); | |
595 | ||
596 | /* | |
597 | * only need to schedule in cgroup events if we are changing | |
598 | * cgroup during ctxsw. Cgroup events were not scheduled | |
599 | * out of ctxsw out if that was not the case. | |
600 | */ | |
601 | if (cgrp1 != cgrp2) | |
602 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
603 | } |
604 | ||
605 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
606 | struct perf_event_attr *attr, | |
607 | struct perf_event *group_leader) | |
608 | { | |
609 | struct perf_cgroup *cgrp; | |
610 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
611 | struct fd f = fdget(fd); |
612 | int ret = 0; | |
e5d1367f | 613 | |
2903ff01 | 614 | if (!f.file) |
e5d1367f SE |
615 | return -EBADF; |
616 | ||
b583043e | 617 | css = css_tryget_online_from_dir(f.file->f_path.dentry, |
ec903c0c | 618 | &perf_event_cgrp_subsys); |
3db272c0 LZ |
619 | if (IS_ERR(css)) { |
620 | ret = PTR_ERR(css); | |
621 | goto out; | |
622 | } | |
e5d1367f SE |
623 | |
624 | cgrp = container_of(css, struct perf_cgroup, css); | |
625 | event->cgrp = cgrp; | |
626 | ||
627 | /* | |
628 | * all events in a group must monitor | |
629 | * the same cgroup because a task belongs | |
630 | * to only one perf cgroup at a time | |
631 | */ | |
632 | if (group_leader && group_leader->cgrp != cgrp) { | |
633 | perf_detach_cgroup(event); | |
634 | ret = -EINVAL; | |
e5d1367f | 635 | } |
3db272c0 | 636 | out: |
2903ff01 | 637 | fdput(f); |
e5d1367f SE |
638 | return ret; |
639 | } | |
640 | ||
641 | static inline void | |
642 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
643 | { | |
644 | struct perf_cgroup_info *t; | |
645 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
646 | event->shadow_ctx_time = now - t->timestamp; | |
647 | } | |
648 | ||
649 | static inline void | |
650 | perf_cgroup_defer_enabled(struct perf_event *event) | |
651 | { | |
652 | /* | |
653 | * when the current task's perf cgroup does not match | |
654 | * the event's, we need to remember to call the | |
655 | * perf_mark_enable() function the first time a task with | |
656 | * a matching perf cgroup is scheduled in. | |
657 | */ | |
658 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
659 | event->cgrp_defer_enabled = 1; | |
660 | } | |
661 | ||
662 | static inline void | |
663 | perf_cgroup_mark_enabled(struct perf_event *event, | |
664 | struct perf_event_context *ctx) | |
665 | { | |
666 | struct perf_event *sub; | |
667 | u64 tstamp = perf_event_time(event); | |
668 | ||
669 | if (!event->cgrp_defer_enabled) | |
670 | return; | |
671 | ||
672 | event->cgrp_defer_enabled = 0; | |
673 | ||
674 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
675 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
676 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
677 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
678 | sub->cgrp_defer_enabled = 0; | |
679 | } | |
680 | } | |
681 | } | |
682 | #else /* !CONFIG_CGROUP_PERF */ | |
683 | ||
684 | static inline bool | |
685 | perf_cgroup_match(struct perf_event *event) | |
686 | { | |
687 | return true; | |
688 | } | |
689 | ||
690 | static inline void perf_detach_cgroup(struct perf_event *event) | |
691 | {} | |
692 | ||
693 | static inline int is_cgroup_event(struct perf_event *event) | |
694 | { | |
695 | return 0; | |
696 | } | |
697 | ||
698 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
699 | { | |
700 | return 0; | |
701 | } | |
702 | ||
703 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
704 | { | |
705 | } | |
706 | ||
707 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
708 | { | |
709 | } | |
710 | ||
a8d757ef SE |
711 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
712 | struct task_struct *next) | |
e5d1367f SE |
713 | { |
714 | } | |
715 | ||
a8d757ef SE |
716 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
717 | struct task_struct *task) | |
e5d1367f SE |
718 | { |
719 | } | |
720 | ||
721 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
722 | struct perf_event_attr *attr, | |
723 | struct perf_event *group_leader) | |
724 | { | |
725 | return -EINVAL; | |
726 | } | |
727 | ||
728 | static inline void | |
3f7cce3c SE |
729 | perf_cgroup_set_timestamp(struct task_struct *task, |
730 | struct perf_event_context *ctx) | |
e5d1367f SE |
731 | { |
732 | } | |
733 | ||
734 | void | |
735 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
736 | { | |
737 | } | |
738 | ||
739 | static inline void | |
740 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
741 | { | |
742 | } | |
743 | ||
744 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
745 | { | |
746 | return 0; | |
747 | } | |
748 | ||
749 | static inline void | |
750 | perf_cgroup_defer_enabled(struct perf_event *event) | |
751 | { | |
752 | } | |
753 | ||
754 | static inline void | |
755 | perf_cgroup_mark_enabled(struct perf_event *event, | |
756 | struct perf_event_context *ctx) | |
757 | { | |
758 | } | |
759 | #endif | |
760 | ||
9e630205 SE |
761 | /* |
762 | * set default to be dependent on timer tick just | |
763 | * like original code | |
764 | */ | |
765 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
766 | /* | |
767 | * function must be called with interrupts disbled | |
768 | */ | |
769 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
770 | { | |
771 | struct perf_cpu_context *cpuctx; | |
772 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
773 | int rotations = 0; | |
774 | ||
775 | WARN_ON(!irqs_disabled()); | |
776 | ||
777 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
778 | ||
779 | rotations = perf_rotate_context(cpuctx); | |
780 | ||
781 | /* | |
782 | * arm timer if needed | |
783 | */ | |
784 | if (rotations) { | |
785 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
786 | ret = HRTIMER_RESTART; | |
787 | } | |
788 | ||
789 | return ret; | |
790 | } | |
791 | ||
792 | /* CPU is going down */ | |
793 | void perf_cpu_hrtimer_cancel(int cpu) | |
794 | { | |
795 | struct perf_cpu_context *cpuctx; | |
796 | struct pmu *pmu; | |
797 | unsigned long flags; | |
798 | ||
799 | if (WARN_ON(cpu != smp_processor_id())) | |
800 | return; | |
801 | ||
802 | local_irq_save(flags); | |
803 | ||
804 | rcu_read_lock(); | |
805 | ||
806 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
807 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
808 | ||
809 | if (pmu->task_ctx_nr == perf_sw_context) | |
810 | continue; | |
811 | ||
812 | hrtimer_cancel(&cpuctx->hrtimer); | |
813 | } | |
814 | ||
815 | rcu_read_unlock(); | |
816 | ||
817 | local_irq_restore(flags); | |
818 | } | |
819 | ||
820 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
821 | { | |
822 | struct hrtimer *hr = &cpuctx->hrtimer; | |
823 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 824 | int timer; |
9e630205 SE |
825 | |
826 | /* no multiplexing needed for SW PMU */ | |
827 | if (pmu->task_ctx_nr == perf_sw_context) | |
828 | return; | |
829 | ||
62b85639 SE |
830 | /* |
831 | * check default is sane, if not set then force to | |
832 | * default interval (1/tick) | |
833 | */ | |
834 | timer = pmu->hrtimer_interval_ms; | |
835 | if (timer < 1) | |
836 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
837 | ||
838 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
839 | |
840 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
841 | hr->function = perf_cpu_hrtimer_handler; | |
842 | } | |
843 | ||
844 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
845 | { | |
846 | struct hrtimer *hr = &cpuctx->hrtimer; | |
847 | struct pmu *pmu = cpuctx->ctx.pmu; | |
848 | ||
849 | /* not for SW PMU */ | |
850 | if (pmu->task_ctx_nr == perf_sw_context) | |
851 | return; | |
852 | ||
853 | if (hrtimer_active(hr)) | |
854 | return; | |
855 | ||
856 | if (!hrtimer_callback_running(hr)) | |
857 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
858 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
859 | } | |
860 | ||
33696fc0 | 861 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 862 | { |
33696fc0 PZ |
863 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
864 | if (!(*count)++) | |
865 | pmu->pmu_disable(pmu); | |
9e35ad38 | 866 | } |
9e35ad38 | 867 | |
33696fc0 | 868 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 869 | { |
33696fc0 PZ |
870 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
871 | if (!--(*count)) | |
872 | pmu->pmu_enable(pmu); | |
9e35ad38 | 873 | } |
9e35ad38 | 874 | |
2fde4f94 | 875 | static DEFINE_PER_CPU(struct list_head, active_ctx_list); |
e9d2b064 PZ |
876 | |
877 | /* | |
2fde4f94 MR |
878 | * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and |
879 | * perf_event_task_tick() are fully serialized because they're strictly cpu | |
880 | * affine and perf_event_ctx{activate,deactivate} are called with IRQs | |
881 | * disabled, while perf_event_task_tick is called from IRQ context. | |
e9d2b064 | 882 | */ |
2fde4f94 | 883 | static void perf_event_ctx_activate(struct perf_event_context *ctx) |
9e35ad38 | 884 | { |
2fde4f94 | 885 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
b5ab4cd5 | 886 | |
e9d2b064 | 887 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 888 | |
2fde4f94 MR |
889 | WARN_ON(!list_empty(&ctx->active_ctx_list)); |
890 | ||
891 | list_add(&ctx->active_ctx_list, head); | |
892 | } | |
893 | ||
894 | static void perf_event_ctx_deactivate(struct perf_event_context *ctx) | |
895 | { | |
896 | WARN_ON(!irqs_disabled()); | |
897 | ||
898 | WARN_ON(list_empty(&ctx->active_ctx_list)); | |
899 | ||
900 | list_del_init(&ctx->active_ctx_list); | |
9e35ad38 | 901 | } |
9e35ad38 | 902 | |
cdd6c482 | 903 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 904 | { |
e5289d4a | 905 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
906 | } |
907 | ||
4af57ef2 YZ |
908 | static void free_ctx(struct rcu_head *head) |
909 | { | |
910 | struct perf_event_context *ctx; | |
911 | ||
912 | ctx = container_of(head, struct perf_event_context, rcu_head); | |
913 | kfree(ctx->task_ctx_data); | |
914 | kfree(ctx); | |
915 | } | |
916 | ||
cdd6c482 | 917 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 918 | { |
564c2b21 PM |
919 | if (atomic_dec_and_test(&ctx->refcount)) { |
920 | if (ctx->parent_ctx) | |
921 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
922 | if (ctx->task) |
923 | put_task_struct(ctx->task); | |
4af57ef2 | 924 | call_rcu(&ctx->rcu_head, free_ctx); |
564c2b21 | 925 | } |
a63eaf34 PM |
926 | } |
927 | ||
f63a8daa PZ |
928 | /* |
929 | * Because of perf_event::ctx migration in sys_perf_event_open::move_group and | |
930 | * perf_pmu_migrate_context() we need some magic. | |
931 | * | |
932 | * Those places that change perf_event::ctx will hold both | |
933 | * perf_event_ctx::mutex of the 'old' and 'new' ctx value. | |
934 | * | |
935 | * Lock ordering is by mutex address. There is one other site where | |
936 | * perf_event_context::mutex nests and that is put_event(). But remember that | |
937 | * that is a parent<->child context relation, and migration does not affect | |
938 | * children, therefore these two orderings should not interact. | |
939 | * | |
940 | * The change in perf_event::ctx does not affect children (as claimed above) | |
941 | * because the sys_perf_event_open() case will install a new event and break | |
942 | * the ctx parent<->child relation, and perf_pmu_migrate_context() is only | |
943 | * concerned with cpuctx and that doesn't have children. | |
944 | * | |
945 | * The places that change perf_event::ctx will issue: | |
946 | * | |
947 | * perf_remove_from_context(); | |
948 | * synchronize_rcu(); | |
949 | * perf_install_in_context(); | |
950 | * | |
951 | * to affect the change. The remove_from_context() + synchronize_rcu() should | |
952 | * quiesce the event, after which we can install it in the new location. This | |
953 | * means that only external vectors (perf_fops, prctl) can perturb the event | |
954 | * while in transit. Therefore all such accessors should also acquire | |
955 | * perf_event_context::mutex to serialize against this. | |
956 | * | |
957 | * However; because event->ctx can change while we're waiting to acquire | |
958 | * ctx->mutex we must be careful and use the below perf_event_ctx_lock() | |
959 | * function. | |
960 | * | |
961 | * Lock order: | |
962 | * task_struct::perf_event_mutex | |
963 | * perf_event_context::mutex | |
964 | * perf_event_context::lock | |
965 | * perf_event::child_mutex; | |
966 | * perf_event::mmap_mutex | |
967 | * mmap_sem | |
968 | */ | |
a83fe28e PZ |
969 | static struct perf_event_context * |
970 | perf_event_ctx_lock_nested(struct perf_event *event, int nesting) | |
f63a8daa PZ |
971 | { |
972 | struct perf_event_context *ctx; | |
973 | ||
974 | again: | |
975 | rcu_read_lock(); | |
976 | ctx = ACCESS_ONCE(event->ctx); | |
977 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
978 | rcu_read_unlock(); | |
979 | goto again; | |
980 | } | |
981 | rcu_read_unlock(); | |
982 | ||
a83fe28e | 983 | mutex_lock_nested(&ctx->mutex, nesting); |
f63a8daa PZ |
984 | if (event->ctx != ctx) { |
985 | mutex_unlock(&ctx->mutex); | |
986 | put_ctx(ctx); | |
987 | goto again; | |
988 | } | |
989 | ||
990 | return ctx; | |
991 | } | |
992 | ||
a83fe28e PZ |
993 | static inline struct perf_event_context * |
994 | perf_event_ctx_lock(struct perf_event *event) | |
995 | { | |
996 | return perf_event_ctx_lock_nested(event, 0); | |
997 | } | |
998 | ||
f63a8daa PZ |
999 | static void perf_event_ctx_unlock(struct perf_event *event, |
1000 | struct perf_event_context *ctx) | |
1001 | { | |
1002 | mutex_unlock(&ctx->mutex); | |
1003 | put_ctx(ctx); | |
1004 | } | |
1005 | ||
211de6eb PZ |
1006 | /* |
1007 | * This must be done under the ctx->lock, such as to serialize against | |
1008 | * context_equiv(), therefore we cannot call put_ctx() since that might end up | |
1009 | * calling scheduler related locks and ctx->lock nests inside those. | |
1010 | */ | |
1011 | static __must_check struct perf_event_context * | |
1012 | unclone_ctx(struct perf_event_context *ctx) | |
71a851b4 | 1013 | { |
211de6eb PZ |
1014 | struct perf_event_context *parent_ctx = ctx->parent_ctx; |
1015 | ||
1016 | lockdep_assert_held(&ctx->lock); | |
1017 | ||
1018 | if (parent_ctx) | |
71a851b4 | 1019 | ctx->parent_ctx = NULL; |
5a3126d4 | 1020 | ctx->generation++; |
211de6eb PZ |
1021 | |
1022 | return parent_ctx; | |
71a851b4 PZ |
1023 | } |
1024 | ||
6844c09d ACM |
1025 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
1026 | { | |
1027 | /* | |
1028 | * only top level events have the pid namespace they were created in | |
1029 | */ | |
1030 | if (event->parent) | |
1031 | event = event->parent; | |
1032 | ||
1033 | return task_tgid_nr_ns(p, event->ns); | |
1034 | } | |
1035 | ||
1036 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
1037 | { | |
1038 | /* | |
1039 | * only top level events have the pid namespace they were created in | |
1040 | */ | |
1041 | if (event->parent) | |
1042 | event = event->parent; | |
1043 | ||
1044 | return task_pid_nr_ns(p, event->ns); | |
1045 | } | |
1046 | ||
7f453c24 | 1047 | /* |
cdd6c482 | 1048 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
1049 | * to userspace. |
1050 | */ | |
cdd6c482 | 1051 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 1052 | { |
cdd6c482 | 1053 | u64 id = event->id; |
7f453c24 | 1054 | |
cdd6c482 IM |
1055 | if (event->parent) |
1056 | id = event->parent->id; | |
7f453c24 PZ |
1057 | |
1058 | return id; | |
1059 | } | |
1060 | ||
25346b93 | 1061 | /* |
cdd6c482 | 1062 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
1063 | * This has to cope with with the fact that until it is locked, |
1064 | * the context could get moved to another task. | |
1065 | */ | |
cdd6c482 | 1066 | static struct perf_event_context * |
8dc85d54 | 1067 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 1068 | { |
cdd6c482 | 1069 | struct perf_event_context *ctx; |
25346b93 | 1070 | |
9ed6060d | 1071 | retry: |
058ebd0e PZ |
1072 | /* |
1073 | * One of the few rules of preemptible RCU is that one cannot do | |
1074 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
1075 | * part of the read side critical section was preemptible -- see | |
1076 | * rcu_read_unlock_special(). | |
1077 | * | |
1078 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
1079 | * side critical section is non-preemptible. | |
1080 | */ | |
1081 | preempt_disable(); | |
1082 | rcu_read_lock(); | |
8dc85d54 | 1083 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
1084 | if (ctx) { |
1085 | /* | |
1086 | * If this context is a clone of another, it might | |
1087 | * get swapped for another underneath us by | |
cdd6c482 | 1088 | * perf_event_task_sched_out, though the |
25346b93 PM |
1089 | * rcu_read_lock() protects us from any context |
1090 | * getting freed. Lock the context and check if it | |
1091 | * got swapped before we could get the lock, and retry | |
1092 | * if so. If we locked the right context, then it | |
1093 | * can't get swapped on us any more. | |
1094 | */ | |
e625cce1 | 1095 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 1096 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 1097 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
1098 | rcu_read_unlock(); |
1099 | preempt_enable(); | |
25346b93 PM |
1100 | goto retry; |
1101 | } | |
b49a9e7e PZ |
1102 | |
1103 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 1104 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1105 | ctx = NULL; |
1106 | } | |
25346b93 PM |
1107 | } |
1108 | rcu_read_unlock(); | |
058ebd0e | 1109 | preempt_enable(); |
25346b93 PM |
1110 | return ctx; |
1111 | } | |
1112 | ||
1113 | /* | |
1114 | * Get the context for a task and increment its pin_count so it | |
1115 | * can't get swapped to another task. This also increments its | |
1116 | * reference count so that the context can't get freed. | |
1117 | */ | |
8dc85d54 PZ |
1118 | static struct perf_event_context * |
1119 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1120 | { |
cdd6c482 | 1121 | struct perf_event_context *ctx; |
25346b93 PM |
1122 | unsigned long flags; |
1123 | ||
8dc85d54 | 1124 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1125 | if (ctx) { |
1126 | ++ctx->pin_count; | |
e625cce1 | 1127 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1128 | } |
1129 | return ctx; | |
1130 | } | |
1131 | ||
cdd6c482 | 1132 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1133 | { |
1134 | unsigned long flags; | |
1135 | ||
e625cce1 | 1136 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1137 | --ctx->pin_count; |
e625cce1 | 1138 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1139 | } |
1140 | ||
f67218c3 PZ |
1141 | /* |
1142 | * Update the record of the current time in a context. | |
1143 | */ | |
1144 | static void update_context_time(struct perf_event_context *ctx) | |
1145 | { | |
1146 | u64 now = perf_clock(); | |
1147 | ||
1148 | ctx->time += now - ctx->timestamp; | |
1149 | ctx->timestamp = now; | |
1150 | } | |
1151 | ||
4158755d SE |
1152 | static u64 perf_event_time(struct perf_event *event) |
1153 | { | |
1154 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1155 | |
1156 | if (is_cgroup_event(event)) | |
1157 | return perf_cgroup_event_time(event); | |
1158 | ||
4158755d SE |
1159 | return ctx ? ctx->time : 0; |
1160 | } | |
1161 | ||
f67218c3 PZ |
1162 | /* |
1163 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1164 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1165 | */ |
1166 | static void update_event_times(struct perf_event *event) | |
1167 | { | |
1168 | struct perf_event_context *ctx = event->ctx; | |
1169 | u64 run_end; | |
1170 | ||
1171 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1172 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1173 | return; | |
e5d1367f SE |
1174 | /* |
1175 | * in cgroup mode, time_enabled represents | |
1176 | * the time the event was enabled AND active | |
1177 | * tasks were in the monitored cgroup. This is | |
1178 | * independent of the activity of the context as | |
1179 | * there may be a mix of cgroup and non-cgroup events. | |
1180 | * | |
1181 | * That is why we treat cgroup events differently | |
1182 | * here. | |
1183 | */ | |
1184 | if (is_cgroup_event(event)) | |
46cd6a7f | 1185 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1186 | else if (ctx->is_active) |
1187 | run_end = ctx->time; | |
acd1d7c1 PZ |
1188 | else |
1189 | run_end = event->tstamp_stopped; | |
1190 | ||
1191 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1192 | |
1193 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1194 | run_end = event->tstamp_stopped; | |
1195 | else | |
4158755d | 1196 | run_end = perf_event_time(event); |
f67218c3 PZ |
1197 | |
1198 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1199 | |
f67218c3 PZ |
1200 | } |
1201 | ||
96c21a46 PZ |
1202 | /* |
1203 | * Update total_time_enabled and total_time_running for all events in a group. | |
1204 | */ | |
1205 | static void update_group_times(struct perf_event *leader) | |
1206 | { | |
1207 | struct perf_event *event; | |
1208 | ||
1209 | update_event_times(leader); | |
1210 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1211 | update_event_times(event); | |
1212 | } | |
1213 | ||
889ff015 FW |
1214 | static struct list_head * |
1215 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1216 | { | |
1217 | if (event->attr.pinned) | |
1218 | return &ctx->pinned_groups; | |
1219 | else | |
1220 | return &ctx->flexible_groups; | |
1221 | } | |
1222 | ||
fccc714b | 1223 | /* |
cdd6c482 | 1224 | * Add a event from the lists for its context. |
fccc714b PZ |
1225 | * Must be called with ctx->mutex and ctx->lock held. |
1226 | */ | |
04289bb9 | 1227 | static void |
cdd6c482 | 1228 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1229 | { |
8a49542c PZ |
1230 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1231 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1232 | |
1233 | /* | |
8a49542c PZ |
1234 | * If we're a stand alone event or group leader, we go to the context |
1235 | * list, group events are kept attached to the group so that | |
1236 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1237 | */ |
8a49542c | 1238 | if (event->group_leader == event) { |
889ff015 FW |
1239 | struct list_head *list; |
1240 | ||
d6f962b5 FW |
1241 | if (is_software_event(event)) |
1242 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1243 | ||
889ff015 FW |
1244 | list = ctx_group_list(event, ctx); |
1245 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1246 | } |
592903cd | 1247 | |
08309379 | 1248 | if (is_cgroup_event(event)) |
e5d1367f | 1249 | ctx->nr_cgroups++; |
e5d1367f | 1250 | |
cdd6c482 IM |
1251 | list_add_rcu(&event->event_entry, &ctx->event_list); |
1252 | ctx->nr_events++; | |
1253 | if (event->attr.inherit_stat) | |
bfbd3381 | 1254 | ctx->nr_stat++; |
5a3126d4 PZ |
1255 | |
1256 | ctx->generation++; | |
04289bb9 IM |
1257 | } |
1258 | ||
0231bb53 JO |
1259 | /* |
1260 | * Initialize event state based on the perf_event_attr::disabled. | |
1261 | */ | |
1262 | static inline void perf_event__state_init(struct perf_event *event) | |
1263 | { | |
1264 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1265 | PERF_EVENT_STATE_INACTIVE; | |
1266 | } | |
1267 | ||
c320c7b7 ACM |
1268 | /* |
1269 | * Called at perf_event creation and when events are attached/detached from a | |
1270 | * group. | |
1271 | */ | |
1272 | static void perf_event__read_size(struct perf_event *event) | |
1273 | { | |
1274 | int entry = sizeof(u64); /* value */ | |
1275 | int size = 0; | |
1276 | int nr = 1; | |
1277 | ||
1278 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1279 | size += sizeof(u64); | |
1280 | ||
1281 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1282 | size += sizeof(u64); | |
1283 | ||
1284 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1285 | entry += sizeof(u64); | |
1286 | ||
1287 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1288 | nr += event->group_leader->nr_siblings; | |
1289 | size += sizeof(u64); | |
1290 | } | |
1291 | ||
1292 | size += entry * nr; | |
1293 | event->read_size = size; | |
1294 | } | |
1295 | ||
1296 | static void perf_event__header_size(struct perf_event *event) | |
1297 | { | |
1298 | struct perf_sample_data *data; | |
1299 | u64 sample_type = event->attr.sample_type; | |
1300 | u16 size = 0; | |
1301 | ||
1302 | perf_event__read_size(event); | |
1303 | ||
1304 | if (sample_type & PERF_SAMPLE_IP) | |
1305 | size += sizeof(data->ip); | |
1306 | ||
6844c09d ACM |
1307 | if (sample_type & PERF_SAMPLE_ADDR) |
1308 | size += sizeof(data->addr); | |
1309 | ||
1310 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1311 | size += sizeof(data->period); | |
1312 | ||
c3feedf2 AK |
1313 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1314 | size += sizeof(data->weight); | |
1315 | ||
6844c09d ACM |
1316 | if (sample_type & PERF_SAMPLE_READ) |
1317 | size += event->read_size; | |
1318 | ||
d6be9ad6 SE |
1319 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1320 | size += sizeof(data->data_src.val); | |
1321 | ||
fdfbbd07 AK |
1322 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1323 | size += sizeof(data->txn); | |
1324 | ||
6844c09d ACM |
1325 | event->header_size = size; |
1326 | } | |
1327 | ||
1328 | static void perf_event__id_header_size(struct perf_event *event) | |
1329 | { | |
1330 | struct perf_sample_data *data; | |
1331 | u64 sample_type = event->attr.sample_type; | |
1332 | u16 size = 0; | |
1333 | ||
c320c7b7 ACM |
1334 | if (sample_type & PERF_SAMPLE_TID) |
1335 | size += sizeof(data->tid_entry); | |
1336 | ||
1337 | if (sample_type & PERF_SAMPLE_TIME) | |
1338 | size += sizeof(data->time); | |
1339 | ||
ff3d527c AH |
1340 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1341 | size += sizeof(data->id); | |
1342 | ||
c320c7b7 ACM |
1343 | if (sample_type & PERF_SAMPLE_ID) |
1344 | size += sizeof(data->id); | |
1345 | ||
1346 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1347 | size += sizeof(data->stream_id); | |
1348 | ||
1349 | if (sample_type & PERF_SAMPLE_CPU) | |
1350 | size += sizeof(data->cpu_entry); | |
1351 | ||
6844c09d | 1352 | event->id_header_size = size; |
c320c7b7 ACM |
1353 | } |
1354 | ||
8a49542c PZ |
1355 | static void perf_group_attach(struct perf_event *event) |
1356 | { | |
c320c7b7 | 1357 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1358 | |
74c3337c PZ |
1359 | /* |
1360 | * We can have double attach due to group movement in perf_event_open. | |
1361 | */ | |
1362 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1363 | return; | |
1364 | ||
8a49542c PZ |
1365 | event->attach_state |= PERF_ATTACH_GROUP; |
1366 | ||
1367 | if (group_leader == event) | |
1368 | return; | |
1369 | ||
652884fe PZ |
1370 | WARN_ON_ONCE(group_leader->ctx != event->ctx); |
1371 | ||
8a49542c PZ |
1372 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
1373 | !is_software_event(event)) | |
1374 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1375 | ||
1376 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1377 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1378 | |
1379 | perf_event__header_size(group_leader); | |
1380 | ||
1381 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1382 | perf_event__header_size(pos); | |
8a49542c PZ |
1383 | } |
1384 | ||
a63eaf34 | 1385 | /* |
cdd6c482 | 1386 | * Remove a event from the lists for its context. |
fccc714b | 1387 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1388 | */ |
04289bb9 | 1389 | static void |
cdd6c482 | 1390 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1391 | { |
68cacd29 | 1392 | struct perf_cpu_context *cpuctx; |
652884fe PZ |
1393 | |
1394 | WARN_ON_ONCE(event->ctx != ctx); | |
1395 | lockdep_assert_held(&ctx->lock); | |
1396 | ||
8a49542c PZ |
1397 | /* |
1398 | * We can have double detach due to exit/hot-unplug + close. | |
1399 | */ | |
1400 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1401 | return; |
8a49542c PZ |
1402 | |
1403 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1404 | ||
68cacd29 | 1405 | if (is_cgroup_event(event)) { |
e5d1367f | 1406 | ctx->nr_cgroups--; |
68cacd29 SE |
1407 | cpuctx = __get_cpu_context(ctx); |
1408 | /* | |
1409 | * if there are no more cgroup events | |
1410 | * then cler cgrp to avoid stale pointer | |
1411 | * in update_cgrp_time_from_cpuctx() | |
1412 | */ | |
1413 | if (!ctx->nr_cgroups) | |
1414 | cpuctx->cgrp = NULL; | |
1415 | } | |
e5d1367f | 1416 | |
cdd6c482 IM |
1417 | ctx->nr_events--; |
1418 | if (event->attr.inherit_stat) | |
bfbd3381 | 1419 | ctx->nr_stat--; |
8bc20959 | 1420 | |
cdd6c482 | 1421 | list_del_rcu(&event->event_entry); |
04289bb9 | 1422 | |
8a49542c PZ |
1423 | if (event->group_leader == event) |
1424 | list_del_init(&event->group_entry); | |
5c148194 | 1425 | |
96c21a46 | 1426 | update_group_times(event); |
b2e74a26 SE |
1427 | |
1428 | /* | |
1429 | * If event was in error state, then keep it | |
1430 | * that way, otherwise bogus counts will be | |
1431 | * returned on read(). The only way to get out | |
1432 | * of error state is by explicit re-enabling | |
1433 | * of the event | |
1434 | */ | |
1435 | if (event->state > PERF_EVENT_STATE_OFF) | |
1436 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1437 | |
1438 | ctx->generation++; | |
050735b0 PZ |
1439 | } |
1440 | ||
8a49542c | 1441 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1442 | { |
1443 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1444 | struct list_head *list = NULL; |
1445 | ||
1446 | /* | |
1447 | * We can have double detach due to exit/hot-unplug + close. | |
1448 | */ | |
1449 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1450 | return; | |
1451 | ||
1452 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1453 | ||
1454 | /* | |
1455 | * If this is a sibling, remove it from its group. | |
1456 | */ | |
1457 | if (event->group_leader != event) { | |
1458 | list_del_init(&event->group_entry); | |
1459 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1460 | goto out; |
8a49542c PZ |
1461 | } |
1462 | ||
1463 | if (!list_empty(&event->group_entry)) | |
1464 | list = &event->group_entry; | |
2e2af50b | 1465 | |
04289bb9 | 1466 | /* |
cdd6c482 IM |
1467 | * If this was a group event with sibling events then |
1468 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1469 | * to whatever list we are on. |
04289bb9 | 1470 | */ |
cdd6c482 | 1471 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1472 | if (list) |
1473 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1474 | sibling->group_leader = sibling; |
d6f962b5 FW |
1475 | |
1476 | /* Inherit group flags from the previous leader */ | |
1477 | sibling->group_flags = event->group_flags; | |
652884fe PZ |
1478 | |
1479 | WARN_ON_ONCE(sibling->ctx != event->ctx); | |
04289bb9 | 1480 | } |
c320c7b7 ACM |
1481 | |
1482 | out: | |
1483 | perf_event__header_size(event->group_leader); | |
1484 | ||
1485 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1486 | perf_event__header_size(tmp); | |
04289bb9 IM |
1487 | } |
1488 | ||
fadfe7be JO |
1489 | /* |
1490 | * User event without the task. | |
1491 | */ | |
1492 | static bool is_orphaned_event(struct perf_event *event) | |
1493 | { | |
1494 | return event && !is_kernel_event(event) && !event->owner; | |
1495 | } | |
1496 | ||
1497 | /* | |
1498 | * Event has a parent but parent's task finished and it's | |
1499 | * alive only because of children holding refference. | |
1500 | */ | |
1501 | static bool is_orphaned_child(struct perf_event *event) | |
1502 | { | |
1503 | return is_orphaned_event(event->parent); | |
1504 | } | |
1505 | ||
1506 | static void orphans_remove_work(struct work_struct *work); | |
1507 | ||
1508 | static void schedule_orphans_remove(struct perf_event_context *ctx) | |
1509 | { | |
1510 | if (!ctx->task || ctx->orphans_remove_sched || !perf_wq) | |
1511 | return; | |
1512 | ||
1513 | if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) { | |
1514 | get_ctx(ctx); | |
1515 | ctx->orphans_remove_sched = true; | |
1516 | } | |
1517 | } | |
1518 | ||
1519 | static int __init perf_workqueue_init(void) | |
1520 | { | |
1521 | perf_wq = create_singlethread_workqueue("perf"); | |
1522 | WARN(!perf_wq, "failed to create perf workqueue\n"); | |
1523 | return perf_wq ? 0 : -1; | |
1524 | } | |
1525 | ||
1526 | core_initcall(perf_workqueue_init); | |
1527 | ||
fa66f07a SE |
1528 | static inline int |
1529 | event_filter_match(struct perf_event *event) | |
1530 | { | |
e5d1367f SE |
1531 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1532 | && perf_cgroup_match(event); | |
fa66f07a SE |
1533 | } |
1534 | ||
9ffcfa6f SE |
1535 | static void |
1536 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1537 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1538 | struct perf_event_context *ctx) |
3b6f9e5c | 1539 | { |
4158755d | 1540 | u64 tstamp = perf_event_time(event); |
fa66f07a | 1541 | u64 delta; |
652884fe PZ |
1542 | |
1543 | WARN_ON_ONCE(event->ctx != ctx); | |
1544 | lockdep_assert_held(&ctx->lock); | |
1545 | ||
fa66f07a SE |
1546 | /* |
1547 | * An event which could not be activated because of | |
1548 | * filter mismatch still needs to have its timings | |
1549 | * maintained, otherwise bogus information is return | |
1550 | * via read() for time_enabled, time_running: | |
1551 | */ | |
1552 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1553 | && !event_filter_match(event)) { | |
e5d1367f | 1554 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1555 | event->tstamp_running += delta; |
4158755d | 1556 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1557 | } |
1558 | ||
cdd6c482 | 1559 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1560 | return; |
3b6f9e5c | 1561 | |
44377277 AS |
1562 | perf_pmu_disable(event->pmu); |
1563 | ||
cdd6c482 IM |
1564 | event->state = PERF_EVENT_STATE_INACTIVE; |
1565 | if (event->pending_disable) { | |
1566 | event->pending_disable = 0; | |
1567 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1568 | } |
4158755d | 1569 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1570 | event->pmu->del(event, 0); |
cdd6c482 | 1571 | event->oncpu = -1; |
3b6f9e5c | 1572 | |
cdd6c482 | 1573 | if (!is_software_event(event)) |
3b6f9e5c | 1574 | cpuctx->active_oncpu--; |
2fde4f94 MR |
1575 | if (!--ctx->nr_active) |
1576 | perf_event_ctx_deactivate(ctx); | |
0f5a2601 PZ |
1577 | if (event->attr.freq && event->attr.sample_freq) |
1578 | ctx->nr_freq--; | |
cdd6c482 | 1579 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1580 | cpuctx->exclusive = 0; |
44377277 | 1581 | |
fadfe7be JO |
1582 | if (is_orphaned_child(event)) |
1583 | schedule_orphans_remove(ctx); | |
1584 | ||
44377277 | 1585 | perf_pmu_enable(event->pmu); |
3b6f9e5c PM |
1586 | } |
1587 | ||
d859e29f | 1588 | static void |
cdd6c482 | 1589 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1590 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1591 | struct perf_event_context *ctx) |
d859e29f | 1592 | { |
cdd6c482 | 1593 | struct perf_event *event; |
fa66f07a | 1594 | int state = group_event->state; |
d859e29f | 1595 | |
cdd6c482 | 1596 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1597 | |
1598 | /* | |
1599 | * Schedule out siblings (if any): | |
1600 | */ | |
cdd6c482 IM |
1601 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1602 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1603 | |
fa66f07a | 1604 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1605 | cpuctx->exclusive = 0; |
1606 | } | |
1607 | ||
46ce0fe9 PZ |
1608 | struct remove_event { |
1609 | struct perf_event *event; | |
1610 | bool detach_group; | |
1611 | }; | |
1612 | ||
0793a61d | 1613 | /* |
cdd6c482 | 1614 | * Cross CPU call to remove a performance event |
0793a61d | 1615 | * |
cdd6c482 | 1616 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1617 | * remove it from the context list. |
1618 | */ | |
fe4b04fa | 1619 | static int __perf_remove_from_context(void *info) |
0793a61d | 1620 | { |
46ce0fe9 PZ |
1621 | struct remove_event *re = info; |
1622 | struct perf_event *event = re->event; | |
cdd6c482 | 1623 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1624 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1625 | |
e625cce1 | 1626 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1627 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1628 | if (re->detach_group) |
1629 | perf_group_detach(event); | |
cdd6c482 | 1630 | list_del_event(event, ctx); |
64ce3126 PZ |
1631 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1632 | ctx->is_active = 0; | |
1633 | cpuctx->task_ctx = NULL; | |
1634 | } | |
e625cce1 | 1635 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1636 | |
1637 | return 0; | |
0793a61d TG |
1638 | } |
1639 | ||
1640 | ||
1641 | /* | |
cdd6c482 | 1642 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1643 | * |
cdd6c482 | 1644 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1645 | * call when the task is on a CPU. |
c93f7669 | 1646 | * |
cdd6c482 IM |
1647 | * If event->ctx is a cloned context, callers must make sure that |
1648 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1649 | * remains valid. This is OK when called from perf_release since |
1650 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1651 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1652 | * context has been detached from its task. |
0793a61d | 1653 | */ |
46ce0fe9 | 1654 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1655 | { |
cdd6c482 | 1656 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1657 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1658 | struct remove_event re = { |
1659 | .event = event, | |
1660 | .detach_group = detach_group, | |
1661 | }; | |
0793a61d | 1662 | |
fe4b04fa PZ |
1663 | lockdep_assert_held(&ctx->mutex); |
1664 | ||
0793a61d TG |
1665 | if (!task) { |
1666 | /* | |
226424ee MR |
1667 | * Per cpu events are removed via an smp call. The removal can |
1668 | * fail if the CPU is currently offline, but in that case we | |
1669 | * already called __perf_remove_from_context from | |
1670 | * perf_event_exit_cpu. | |
0793a61d | 1671 | */ |
46ce0fe9 | 1672 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1673 | return; |
1674 | } | |
1675 | ||
1676 | retry: | |
46ce0fe9 | 1677 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1678 | return; |
0793a61d | 1679 | |
e625cce1 | 1680 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1681 | /* |
fe4b04fa PZ |
1682 | * If we failed to find a running task, but find the context active now |
1683 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1684 | */ |
fe4b04fa | 1685 | if (ctx->is_active) { |
e625cce1 | 1686 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
1687 | /* |
1688 | * Reload the task pointer, it might have been changed by | |
1689 | * a concurrent perf_event_context_sched_out(). | |
1690 | */ | |
1691 | task = ctx->task; | |
0793a61d TG |
1692 | goto retry; |
1693 | } | |
1694 | ||
1695 | /* | |
fe4b04fa PZ |
1696 | * Since the task isn't running, its safe to remove the event, us |
1697 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1698 | */ |
46ce0fe9 PZ |
1699 | if (detach_group) |
1700 | perf_group_detach(event); | |
fe4b04fa | 1701 | list_del_event(event, ctx); |
e625cce1 | 1702 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1703 | } |
1704 | ||
d859e29f | 1705 | /* |
cdd6c482 | 1706 | * Cross CPU call to disable a performance event |
d859e29f | 1707 | */ |
500ad2d8 | 1708 | int __perf_event_disable(void *info) |
d859e29f | 1709 | { |
cdd6c482 | 1710 | struct perf_event *event = info; |
cdd6c482 | 1711 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1712 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1713 | |
1714 | /* | |
cdd6c482 IM |
1715 | * If this is a per-task event, need to check whether this |
1716 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1717 | * |
1718 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1719 | * flipping contexts around. | |
d859e29f | 1720 | */ |
665c2142 | 1721 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1722 | return -EINVAL; |
d859e29f | 1723 | |
e625cce1 | 1724 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1725 | |
1726 | /* | |
cdd6c482 | 1727 | * If the event is on, turn it off. |
d859e29f PM |
1728 | * If it is in error state, leave it in error state. |
1729 | */ | |
cdd6c482 | 1730 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1731 | update_context_time(ctx); |
e5d1367f | 1732 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1733 | update_group_times(event); |
1734 | if (event == event->group_leader) | |
1735 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1736 | else |
cdd6c482 IM |
1737 | event_sched_out(event, cpuctx, ctx); |
1738 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1739 | } |
1740 | ||
e625cce1 | 1741 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1742 | |
1743 | return 0; | |
d859e29f PM |
1744 | } |
1745 | ||
1746 | /* | |
cdd6c482 | 1747 | * Disable a event. |
c93f7669 | 1748 | * |
cdd6c482 IM |
1749 | * If event->ctx is a cloned context, callers must make sure that |
1750 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1751 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1752 | * perf_event_for_each_child or perf_event_for_each because they |
1753 | * hold the top-level event's child_mutex, so any descendant that | |
1754 | * goes to exit will block in sync_child_event. | |
1755 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1756 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1757 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1758 | */ |
f63a8daa | 1759 | static void _perf_event_disable(struct perf_event *event) |
d859e29f | 1760 | { |
cdd6c482 | 1761 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1762 | struct task_struct *task = ctx->task; |
1763 | ||
1764 | if (!task) { | |
1765 | /* | |
cdd6c482 | 1766 | * Disable the event on the cpu that it's on |
d859e29f | 1767 | */ |
fe4b04fa | 1768 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1769 | return; |
1770 | } | |
1771 | ||
9ed6060d | 1772 | retry: |
fe4b04fa PZ |
1773 | if (!task_function_call(task, __perf_event_disable, event)) |
1774 | return; | |
d859e29f | 1775 | |
e625cce1 | 1776 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1777 | /* |
cdd6c482 | 1778 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1779 | */ |
cdd6c482 | 1780 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1781 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1782 | /* |
1783 | * Reload the task pointer, it might have been changed by | |
1784 | * a concurrent perf_event_context_sched_out(). | |
1785 | */ | |
1786 | task = ctx->task; | |
d859e29f PM |
1787 | goto retry; |
1788 | } | |
1789 | ||
1790 | /* | |
1791 | * Since we have the lock this context can't be scheduled | |
1792 | * in, so we can change the state safely. | |
1793 | */ | |
cdd6c482 IM |
1794 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1795 | update_group_times(event); | |
1796 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1797 | } |
e625cce1 | 1798 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1799 | } |
f63a8daa PZ |
1800 | |
1801 | /* | |
1802 | * Strictly speaking kernel users cannot create groups and therefore this | |
1803 | * interface does not need the perf_event_ctx_lock() magic. | |
1804 | */ | |
1805 | void perf_event_disable(struct perf_event *event) | |
1806 | { | |
1807 | struct perf_event_context *ctx; | |
1808 | ||
1809 | ctx = perf_event_ctx_lock(event); | |
1810 | _perf_event_disable(event); | |
1811 | perf_event_ctx_unlock(event, ctx); | |
1812 | } | |
dcfce4a0 | 1813 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1814 | |
e5d1367f SE |
1815 | static void perf_set_shadow_time(struct perf_event *event, |
1816 | struct perf_event_context *ctx, | |
1817 | u64 tstamp) | |
1818 | { | |
1819 | /* | |
1820 | * use the correct time source for the time snapshot | |
1821 | * | |
1822 | * We could get by without this by leveraging the | |
1823 | * fact that to get to this function, the caller | |
1824 | * has most likely already called update_context_time() | |
1825 | * and update_cgrp_time_xx() and thus both timestamp | |
1826 | * are identical (or very close). Given that tstamp is, | |
1827 | * already adjusted for cgroup, we could say that: | |
1828 | * tstamp - ctx->timestamp | |
1829 | * is equivalent to | |
1830 | * tstamp - cgrp->timestamp. | |
1831 | * | |
1832 | * Then, in perf_output_read(), the calculation would | |
1833 | * work with no changes because: | |
1834 | * - event is guaranteed scheduled in | |
1835 | * - no scheduled out in between | |
1836 | * - thus the timestamp would be the same | |
1837 | * | |
1838 | * But this is a bit hairy. | |
1839 | * | |
1840 | * So instead, we have an explicit cgroup call to remain | |
1841 | * within the time time source all along. We believe it | |
1842 | * is cleaner and simpler to understand. | |
1843 | */ | |
1844 | if (is_cgroup_event(event)) | |
1845 | perf_cgroup_set_shadow_time(event, tstamp); | |
1846 | else | |
1847 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1848 | } | |
1849 | ||
4fe757dd PZ |
1850 | #define MAX_INTERRUPTS (~0ULL) |
1851 | ||
1852 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1853 | ||
235c7fc7 | 1854 | static int |
9ffcfa6f | 1855 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1856 | struct perf_cpu_context *cpuctx, |
6e37738a | 1857 | struct perf_event_context *ctx) |
235c7fc7 | 1858 | { |
4158755d | 1859 | u64 tstamp = perf_event_time(event); |
44377277 | 1860 | int ret = 0; |
4158755d | 1861 | |
63342411 PZ |
1862 | lockdep_assert_held(&ctx->lock); |
1863 | ||
cdd6c482 | 1864 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1865 | return 0; |
1866 | ||
cdd6c482 | 1867 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1868 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1869 | |
1870 | /* | |
1871 | * Unthrottle events, since we scheduled we might have missed several | |
1872 | * ticks already, also for a heavily scheduling task there is little | |
1873 | * guarantee it'll get a tick in a timely manner. | |
1874 | */ | |
1875 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1876 | perf_log_throttle(event, 1); | |
1877 | event->hw.interrupts = 0; | |
1878 | } | |
1879 | ||
235c7fc7 IM |
1880 | /* |
1881 | * The new state must be visible before we turn it on in the hardware: | |
1882 | */ | |
1883 | smp_wmb(); | |
1884 | ||
44377277 AS |
1885 | perf_pmu_disable(event->pmu); |
1886 | ||
72f669c0 SL |
1887 | event->tstamp_running += tstamp - event->tstamp_stopped; |
1888 | ||
1889 | perf_set_shadow_time(event, ctx, tstamp); | |
1890 | ||
a4eaf7f1 | 1891 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1892 | event->state = PERF_EVENT_STATE_INACTIVE; |
1893 | event->oncpu = -1; | |
44377277 AS |
1894 | ret = -EAGAIN; |
1895 | goto out; | |
235c7fc7 IM |
1896 | } |
1897 | ||
cdd6c482 | 1898 | if (!is_software_event(event)) |
3b6f9e5c | 1899 | cpuctx->active_oncpu++; |
2fde4f94 MR |
1900 | if (!ctx->nr_active++) |
1901 | perf_event_ctx_activate(ctx); | |
0f5a2601 PZ |
1902 | if (event->attr.freq && event->attr.sample_freq) |
1903 | ctx->nr_freq++; | |
235c7fc7 | 1904 | |
cdd6c482 | 1905 | if (event->attr.exclusive) |
3b6f9e5c PM |
1906 | cpuctx->exclusive = 1; |
1907 | ||
fadfe7be JO |
1908 | if (is_orphaned_child(event)) |
1909 | schedule_orphans_remove(ctx); | |
1910 | ||
44377277 AS |
1911 | out: |
1912 | perf_pmu_enable(event->pmu); | |
1913 | ||
1914 | return ret; | |
235c7fc7 IM |
1915 | } |
1916 | ||
6751b71e | 1917 | static int |
cdd6c482 | 1918 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1919 | struct perf_cpu_context *cpuctx, |
6e37738a | 1920 | struct perf_event_context *ctx) |
6751b71e | 1921 | { |
6bde9b6c | 1922 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1923 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1924 | u64 now = ctx->time; |
1925 | bool simulate = false; | |
6751b71e | 1926 | |
cdd6c482 | 1927 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1928 | return 0; |
1929 | ||
ad5133b7 | 1930 | pmu->start_txn(pmu); |
6bde9b6c | 1931 | |
9ffcfa6f | 1932 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1933 | pmu->cancel_txn(pmu); |
9e630205 | 1934 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1935 | return -EAGAIN; |
90151c35 | 1936 | } |
6751b71e PM |
1937 | |
1938 | /* | |
1939 | * Schedule in siblings as one group (if any): | |
1940 | */ | |
cdd6c482 | 1941 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1942 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1943 | partial_group = event; |
6751b71e PM |
1944 | goto group_error; |
1945 | } | |
1946 | } | |
1947 | ||
9ffcfa6f | 1948 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1949 | return 0; |
9ffcfa6f | 1950 | |
6751b71e PM |
1951 | group_error: |
1952 | /* | |
1953 | * Groups can be scheduled in as one unit only, so undo any | |
1954 | * partial group before returning: | |
d7842da4 SE |
1955 | * The events up to the failed event are scheduled out normally, |
1956 | * tstamp_stopped will be updated. | |
1957 | * | |
1958 | * The failed events and the remaining siblings need to have | |
1959 | * their timings updated as if they had gone thru event_sched_in() | |
1960 | * and event_sched_out(). This is required to get consistent timings | |
1961 | * across the group. This also takes care of the case where the group | |
1962 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1963 | * the time the event was actually stopped, such that time delta | |
1964 | * calculation in update_event_times() is correct. | |
6751b71e | 1965 | */ |
cdd6c482 IM |
1966 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1967 | if (event == partial_group) | |
d7842da4 SE |
1968 | simulate = true; |
1969 | ||
1970 | if (simulate) { | |
1971 | event->tstamp_running += now - event->tstamp_stopped; | |
1972 | event->tstamp_stopped = now; | |
1973 | } else { | |
1974 | event_sched_out(event, cpuctx, ctx); | |
1975 | } | |
6751b71e | 1976 | } |
9ffcfa6f | 1977 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1978 | |
ad5133b7 | 1979 | pmu->cancel_txn(pmu); |
90151c35 | 1980 | |
9e630205 SE |
1981 | perf_cpu_hrtimer_restart(cpuctx); |
1982 | ||
6751b71e PM |
1983 | return -EAGAIN; |
1984 | } | |
1985 | ||
3b6f9e5c | 1986 | /* |
cdd6c482 | 1987 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1988 | */ |
cdd6c482 | 1989 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1990 | struct perf_cpu_context *cpuctx, |
1991 | int can_add_hw) | |
1992 | { | |
1993 | /* | |
cdd6c482 | 1994 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1995 | */ |
d6f962b5 | 1996 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1997 | return 1; |
1998 | /* | |
1999 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 2000 | * events can go on. |
3b6f9e5c PM |
2001 | */ |
2002 | if (cpuctx->exclusive) | |
2003 | return 0; | |
2004 | /* | |
2005 | * If this group is exclusive and there are already | |
cdd6c482 | 2006 | * events on the CPU, it can't go on. |
3b6f9e5c | 2007 | */ |
cdd6c482 | 2008 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
2009 | return 0; |
2010 | /* | |
2011 | * Otherwise, try to add it if all previous groups were able | |
2012 | * to go on. | |
2013 | */ | |
2014 | return can_add_hw; | |
2015 | } | |
2016 | ||
cdd6c482 IM |
2017 | static void add_event_to_ctx(struct perf_event *event, |
2018 | struct perf_event_context *ctx) | |
53cfbf59 | 2019 | { |
4158755d SE |
2020 | u64 tstamp = perf_event_time(event); |
2021 | ||
cdd6c482 | 2022 | list_add_event(event, ctx); |
8a49542c | 2023 | perf_group_attach(event); |
4158755d SE |
2024 | event->tstamp_enabled = tstamp; |
2025 | event->tstamp_running = tstamp; | |
2026 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
2027 | } |
2028 | ||
2c29ef0f PZ |
2029 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
2030 | static void | |
2031 | ctx_sched_in(struct perf_event_context *ctx, | |
2032 | struct perf_cpu_context *cpuctx, | |
2033 | enum event_type_t event_type, | |
2034 | struct task_struct *task); | |
fe4b04fa | 2035 | |
dce5855b PZ |
2036 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
2037 | struct perf_event_context *ctx, | |
2038 | struct task_struct *task) | |
2039 | { | |
2040 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
2041 | if (ctx) | |
2042 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
2043 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
2044 | if (ctx) | |
2045 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
2046 | } | |
2047 | ||
0793a61d | 2048 | /* |
cdd6c482 | 2049 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
2050 | * |
2051 | * Must be called with ctx->mutex held | |
0793a61d | 2052 | */ |
fe4b04fa | 2053 | static int __perf_install_in_context(void *info) |
0793a61d | 2054 | { |
cdd6c482 IM |
2055 | struct perf_event *event = info; |
2056 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2057 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
2058 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
2059 | struct task_struct *task = current; | |
2060 | ||
b58f6b0d | 2061 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 2062 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
2063 | |
2064 | /* | |
2c29ef0f | 2065 | * If there was an active task_ctx schedule it out. |
0793a61d | 2066 | */ |
b58f6b0d | 2067 | if (task_ctx) |
2c29ef0f | 2068 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
2069 | |
2070 | /* | |
2071 | * If the context we're installing events in is not the | |
2072 | * active task_ctx, flip them. | |
2073 | */ | |
2074 | if (ctx->task && task_ctx != ctx) { | |
2075 | if (task_ctx) | |
2076 | raw_spin_unlock(&task_ctx->lock); | |
2077 | raw_spin_lock(&ctx->lock); | |
2078 | task_ctx = ctx; | |
2079 | } | |
2080 | ||
2081 | if (task_ctx) { | |
2082 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
2083 | task = task_ctx->task; |
2084 | } | |
b58f6b0d | 2085 | |
2c29ef0f | 2086 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 2087 | |
4af4998b | 2088 | update_context_time(ctx); |
e5d1367f SE |
2089 | /* |
2090 | * update cgrp time only if current cgrp | |
2091 | * matches event->cgrp. Must be done before | |
2092 | * calling add_event_to_ctx() | |
2093 | */ | |
2094 | update_cgrp_time_from_event(event); | |
0793a61d | 2095 | |
cdd6c482 | 2096 | add_event_to_ctx(event, ctx); |
0793a61d | 2097 | |
d859e29f | 2098 | /* |
2c29ef0f | 2099 | * Schedule everything back in |
d859e29f | 2100 | */ |
dce5855b | 2101 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
2102 | |
2103 | perf_pmu_enable(cpuctx->ctx.pmu); | |
2104 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
2105 | |
2106 | return 0; | |
0793a61d TG |
2107 | } |
2108 | ||
2109 | /* | |
cdd6c482 | 2110 | * Attach a performance event to a context |
0793a61d | 2111 | * |
cdd6c482 IM |
2112 | * First we add the event to the list with the hardware enable bit |
2113 | * in event->hw_config cleared. | |
0793a61d | 2114 | * |
cdd6c482 | 2115 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
2116 | * call to enable it in the task context. The task might have been |
2117 | * scheduled away, but we check this in the smp call again. | |
2118 | */ | |
2119 | static void | |
cdd6c482 IM |
2120 | perf_install_in_context(struct perf_event_context *ctx, |
2121 | struct perf_event *event, | |
0793a61d TG |
2122 | int cpu) |
2123 | { | |
2124 | struct task_struct *task = ctx->task; | |
2125 | ||
fe4b04fa PZ |
2126 | lockdep_assert_held(&ctx->mutex); |
2127 | ||
c3f00c70 | 2128 | event->ctx = ctx; |
0cda4c02 YZ |
2129 | if (event->cpu != -1) |
2130 | event->cpu = cpu; | |
c3f00c70 | 2131 | |
0793a61d TG |
2132 | if (!task) { |
2133 | /* | |
cdd6c482 | 2134 | * Per cpu events are installed via an smp call and |
af901ca1 | 2135 | * the install is always successful. |
0793a61d | 2136 | */ |
fe4b04fa | 2137 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
2138 | return; |
2139 | } | |
2140 | ||
0793a61d | 2141 | retry: |
fe4b04fa PZ |
2142 | if (!task_function_call(task, __perf_install_in_context, event)) |
2143 | return; | |
0793a61d | 2144 | |
e625cce1 | 2145 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 2146 | /* |
fe4b04fa PZ |
2147 | * If we failed to find a running task, but find the context active now |
2148 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 2149 | */ |
fe4b04fa | 2150 | if (ctx->is_active) { |
e625cce1 | 2151 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
2152 | /* |
2153 | * Reload the task pointer, it might have been changed by | |
2154 | * a concurrent perf_event_context_sched_out(). | |
2155 | */ | |
2156 | task = ctx->task; | |
0793a61d TG |
2157 | goto retry; |
2158 | } | |
2159 | ||
2160 | /* | |
fe4b04fa PZ |
2161 | * Since the task isn't running, its safe to add the event, us holding |
2162 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 2163 | */ |
fe4b04fa | 2164 | add_event_to_ctx(event, ctx); |
e625cce1 | 2165 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
2166 | } |
2167 | ||
fa289bec | 2168 | /* |
cdd6c482 | 2169 | * Put a event into inactive state and update time fields. |
fa289bec PM |
2170 | * Enabling the leader of a group effectively enables all |
2171 | * the group members that aren't explicitly disabled, so we | |
2172 | * have to update their ->tstamp_enabled also. | |
2173 | * Note: this works for group members as well as group leaders | |
2174 | * since the non-leader members' sibling_lists will be empty. | |
2175 | */ | |
1d9b482e | 2176 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 2177 | { |
cdd6c482 | 2178 | struct perf_event *sub; |
4158755d | 2179 | u64 tstamp = perf_event_time(event); |
fa289bec | 2180 | |
cdd6c482 | 2181 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 2182 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 2183 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
2184 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
2185 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 2186 | } |
fa289bec PM |
2187 | } |
2188 | ||
d859e29f | 2189 | /* |
cdd6c482 | 2190 | * Cross CPU call to enable a performance event |
d859e29f | 2191 | */ |
fe4b04fa | 2192 | static int __perf_event_enable(void *info) |
04289bb9 | 2193 | { |
cdd6c482 | 2194 | struct perf_event *event = info; |
cdd6c482 IM |
2195 | struct perf_event_context *ctx = event->ctx; |
2196 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2197 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2198 | int err; |
04289bb9 | 2199 | |
06f41796 JO |
2200 | /* |
2201 | * There's a time window between 'ctx->is_active' check | |
2202 | * in perf_event_enable function and this place having: | |
2203 | * - IRQs on | |
2204 | * - ctx->lock unlocked | |
2205 | * | |
2206 | * where the task could be killed and 'ctx' deactivated | |
2207 | * by perf_event_exit_task. | |
2208 | */ | |
2209 | if (!ctx->is_active) | |
fe4b04fa | 2210 | return -EINVAL; |
3cbed429 | 2211 | |
e625cce1 | 2212 | raw_spin_lock(&ctx->lock); |
4af4998b | 2213 | update_context_time(ctx); |
d859e29f | 2214 | |
cdd6c482 | 2215 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2216 | goto unlock; |
e5d1367f SE |
2217 | |
2218 | /* | |
2219 | * set current task's cgroup time reference point | |
2220 | */ | |
3f7cce3c | 2221 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2222 | |
1d9b482e | 2223 | __perf_event_mark_enabled(event); |
04289bb9 | 2224 | |
e5d1367f SE |
2225 | if (!event_filter_match(event)) { |
2226 | if (is_cgroup_event(event)) | |
2227 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2228 | goto unlock; |
e5d1367f | 2229 | } |
f4c4176f | 2230 | |
04289bb9 | 2231 | /* |
cdd6c482 | 2232 | * If the event is in a group and isn't the group leader, |
d859e29f | 2233 | * then don't put it on unless the group is on. |
04289bb9 | 2234 | */ |
cdd6c482 | 2235 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2236 | goto unlock; |
3b6f9e5c | 2237 | |
cdd6c482 | 2238 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2239 | err = -EEXIST; |
e758a33d | 2240 | } else { |
cdd6c482 | 2241 | if (event == leader) |
6e37738a | 2242 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2243 | else |
6e37738a | 2244 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2245 | } |
d859e29f PM |
2246 | |
2247 | if (err) { | |
2248 | /* | |
cdd6c482 | 2249 | * If this event can't go on and it's part of a |
d859e29f PM |
2250 | * group, then the whole group has to come off. |
2251 | */ | |
9e630205 | 2252 | if (leader != event) { |
d859e29f | 2253 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2254 | perf_cpu_hrtimer_restart(cpuctx); |
2255 | } | |
0d48696f | 2256 | if (leader->attr.pinned) { |
53cfbf59 | 2257 | update_group_times(leader); |
cdd6c482 | 2258 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2259 | } |
d859e29f PM |
2260 | } |
2261 | ||
9ed6060d | 2262 | unlock: |
e625cce1 | 2263 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2264 | |
2265 | return 0; | |
d859e29f PM |
2266 | } |
2267 | ||
2268 | /* | |
cdd6c482 | 2269 | * Enable a event. |
c93f7669 | 2270 | * |
cdd6c482 IM |
2271 | * If event->ctx is a cloned context, callers must make sure that |
2272 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2273 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2274 | * perf_event_for_each_child or perf_event_for_each as described |
2275 | * for perf_event_disable. | |
d859e29f | 2276 | */ |
f63a8daa | 2277 | static void _perf_event_enable(struct perf_event *event) |
d859e29f | 2278 | { |
cdd6c482 | 2279 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2280 | struct task_struct *task = ctx->task; |
2281 | ||
2282 | if (!task) { | |
2283 | /* | |
cdd6c482 | 2284 | * Enable the event on the cpu that it's on |
d859e29f | 2285 | */ |
fe4b04fa | 2286 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2287 | return; |
2288 | } | |
2289 | ||
e625cce1 | 2290 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2291 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2292 | goto out; |
2293 | ||
2294 | /* | |
cdd6c482 IM |
2295 | * If the event is in error state, clear that first. |
2296 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2297 | * know that it has gone back into error state, as distinct |
2298 | * from the task having been scheduled away before the | |
2299 | * cross-call arrived. | |
2300 | */ | |
cdd6c482 IM |
2301 | if (event->state == PERF_EVENT_STATE_ERROR) |
2302 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2303 | |
9ed6060d | 2304 | retry: |
fe4b04fa | 2305 | if (!ctx->is_active) { |
1d9b482e | 2306 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2307 | goto out; |
2308 | } | |
2309 | ||
e625cce1 | 2310 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2311 | |
2312 | if (!task_function_call(task, __perf_event_enable, event)) | |
2313 | return; | |
d859e29f | 2314 | |
e625cce1 | 2315 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2316 | |
2317 | /* | |
cdd6c482 | 2318 | * If the context is active and the event is still off, |
d859e29f PM |
2319 | * we need to retry the cross-call. |
2320 | */ | |
fe4b04fa PZ |
2321 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2322 | /* | |
2323 | * task could have been flipped by a concurrent | |
2324 | * perf_event_context_sched_out() | |
2325 | */ | |
2326 | task = ctx->task; | |
d859e29f | 2327 | goto retry; |
fe4b04fa | 2328 | } |
fa289bec | 2329 | |
9ed6060d | 2330 | out: |
e625cce1 | 2331 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2332 | } |
f63a8daa PZ |
2333 | |
2334 | /* | |
2335 | * See perf_event_disable(); | |
2336 | */ | |
2337 | void perf_event_enable(struct perf_event *event) | |
2338 | { | |
2339 | struct perf_event_context *ctx; | |
2340 | ||
2341 | ctx = perf_event_ctx_lock(event); | |
2342 | _perf_event_enable(event); | |
2343 | perf_event_ctx_unlock(event, ctx); | |
2344 | } | |
dcfce4a0 | 2345 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2346 | |
f63a8daa | 2347 | static int _perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2348 | { |
2023b359 | 2349 | /* |
cdd6c482 | 2350 | * not supported on inherited events |
2023b359 | 2351 | */ |
2e939d1d | 2352 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2353 | return -EINVAL; |
2354 | ||
cdd6c482 | 2355 | atomic_add(refresh, &event->event_limit); |
f63a8daa | 2356 | _perf_event_enable(event); |
2023b359 PZ |
2357 | |
2358 | return 0; | |
79f14641 | 2359 | } |
f63a8daa PZ |
2360 | |
2361 | /* | |
2362 | * See perf_event_disable() | |
2363 | */ | |
2364 | int perf_event_refresh(struct perf_event *event, int refresh) | |
2365 | { | |
2366 | struct perf_event_context *ctx; | |
2367 | int ret; | |
2368 | ||
2369 | ctx = perf_event_ctx_lock(event); | |
2370 | ret = _perf_event_refresh(event, refresh); | |
2371 | perf_event_ctx_unlock(event, ctx); | |
2372 | ||
2373 | return ret; | |
2374 | } | |
26ca5c11 | 2375 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2376 | |
5b0311e1 FW |
2377 | static void ctx_sched_out(struct perf_event_context *ctx, |
2378 | struct perf_cpu_context *cpuctx, | |
2379 | enum event_type_t event_type) | |
235c7fc7 | 2380 | { |
cdd6c482 | 2381 | struct perf_event *event; |
db24d33e | 2382 | int is_active = ctx->is_active; |
235c7fc7 | 2383 | |
db24d33e | 2384 | ctx->is_active &= ~event_type; |
cdd6c482 | 2385 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2386 | return; |
2387 | ||
4af4998b | 2388 | update_context_time(ctx); |
e5d1367f | 2389 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2390 | if (!ctx->nr_active) |
facc4307 | 2391 | return; |
5b0311e1 | 2392 | |
075e0b00 | 2393 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2394 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2395 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2396 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2397 | } |
889ff015 | 2398 | |
db24d33e | 2399 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2400 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2401 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2402 | } |
1b9a644f | 2403 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2404 | } |
2405 | ||
564c2b21 | 2406 | /* |
5a3126d4 PZ |
2407 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2408 | * cloned from the same version of the same context. | |
2409 | * | |
2410 | * Equivalence is measured using a generation number in the context that is | |
2411 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2412 | * and list_del_event(). | |
564c2b21 | 2413 | */ |
cdd6c482 IM |
2414 | static int context_equiv(struct perf_event_context *ctx1, |
2415 | struct perf_event_context *ctx2) | |
564c2b21 | 2416 | { |
211de6eb PZ |
2417 | lockdep_assert_held(&ctx1->lock); |
2418 | lockdep_assert_held(&ctx2->lock); | |
2419 | ||
5a3126d4 PZ |
2420 | /* Pinning disables the swap optimization */ |
2421 | if (ctx1->pin_count || ctx2->pin_count) | |
2422 | return 0; | |
2423 | ||
2424 | /* If ctx1 is the parent of ctx2 */ | |
2425 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2426 | return 1; | |
2427 | ||
2428 | /* If ctx2 is the parent of ctx1 */ | |
2429 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2430 | return 1; | |
2431 | ||
2432 | /* | |
2433 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2434 | * hierarchy, see perf_event_init_context(). | |
2435 | */ | |
2436 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2437 | ctx1->parent_gen == ctx2->parent_gen) | |
2438 | return 1; | |
2439 | ||
2440 | /* Unmatched */ | |
2441 | return 0; | |
564c2b21 PM |
2442 | } |
2443 | ||
cdd6c482 IM |
2444 | static void __perf_event_sync_stat(struct perf_event *event, |
2445 | struct perf_event *next_event) | |
bfbd3381 PZ |
2446 | { |
2447 | u64 value; | |
2448 | ||
cdd6c482 | 2449 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2450 | return; |
2451 | ||
2452 | /* | |
cdd6c482 | 2453 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2454 | * because we're in the middle of a context switch and have IRQs |
2455 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2456 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2457 | * don't need to use it. |
2458 | */ | |
cdd6c482 IM |
2459 | switch (event->state) { |
2460 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2461 | event->pmu->read(event); |
2462 | /* fall-through */ | |
bfbd3381 | 2463 | |
cdd6c482 IM |
2464 | case PERF_EVENT_STATE_INACTIVE: |
2465 | update_event_times(event); | |
bfbd3381 PZ |
2466 | break; |
2467 | ||
2468 | default: | |
2469 | break; | |
2470 | } | |
2471 | ||
2472 | /* | |
cdd6c482 | 2473 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2474 | * values when we flip the contexts. |
2475 | */ | |
e7850595 PZ |
2476 | value = local64_read(&next_event->count); |
2477 | value = local64_xchg(&event->count, value); | |
2478 | local64_set(&next_event->count, value); | |
bfbd3381 | 2479 | |
cdd6c482 IM |
2480 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2481 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2482 | |
bfbd3381 | 2483 | /* |
19d2e755 | 2484 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2485 | */ |
cdd6c482 IM |
2486 | perf_event_update_userpage(event); |
2487 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2488 | } |
2489 | ||
cdd6c482 IM |
2490 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2491 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2492 | { |
cdd6c482 | 2493 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2494 | |
2495 | if (!ctx->nr_stat) | |
2496 | return; | |
2497 | ||
02ffdbc8 PZ |
2498 | update_context_time(ctx); |
2499 | ||
cdd6c482 IM |
2500 | event = list_first_entry(&ctx->event_list, |
2501 | struct perf_event, event_entry); | |
bfbd3381 | 2502 | |
cdd6c482 IM |
2503 | next_event = list_first_entry(&next_ctx->event_list, |
2504 | struct perf_event, event_entry); | |
bfbd3381 | 2505 | |
cdd6c482 IM |
2506 | while (&event->event_entry != &ctx->event_list && |
2507 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2508 | |
cdd6c482 | 2509 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2510 | |
cdd6c482 IM |
2511 | event = list_next_entry(event, event_entry); |
2512 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2513 | } |
2514 | } | |
2515 | ||
fe4b04fa PZ |
2516 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2517 | struct task_struct *next) | |
0793a61d | 2518 | { |
8dc85d54 | 2519 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2520 | struct perf_event_context *next_ctx; |
5a3126d4 | 2521 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2522 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2523 | int do_switch = 1; |
0793a61d | 2524 | |
108b02cf PZ |
2525 | if (likely(!ctx)) |
2526 | return; | |
10989fb2 | 2527 | |
108b02cf PZ |
2528 | cpuctx = __get_cpu_context(ctx); |
2529 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2530 | return; |
2531 | ||
c93f7669 | 2532 | rcu_read_lock(); |
8dc85d54 | 2533 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2534 | if (!next_ctx) |
2535 | goto unlock; | |
2536 | ||
2537 | parent = rcu_dereference(ctx->parent_ctx); | |
2538 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2539 | ||
2540 | /* If neither context have a parent context; they cannot be clones. */ | |
802c8a61 | 2541 | if (!parent && !next_parent) |
5a3126d4 PZ |
2542 | goto unlock; |
2543 | ||
2544 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2545 | /* |
2546 | * Looks like the two contexts are clones, so we might be | |
2547 | * able to optimize the context switch. We lock both | |
2548 | * contexts and check that they are clones under the | |
2549 | * lock (including re-checking that neither has been | |
2550 | * uncloned in the meantime). It doesn't matter which | |
2551 | * order we take the locks because no other cpu could | |
2552 | * be trying to lock both of these tasks. | |
2553 | */ | |
e625cce1 TG |
2554 | raw_spin_lock(&ctx->lock); |
2555 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2556 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2557 | /* |
2558 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2559 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2560 | */ |
8dc85d54 PZ |
2561 | task->perf_event_ctxp[ctxn] = next_ctx; |
2562 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2563 | ctx->task = next; |
2564 | next_ctx->task = task; | |
5a158c3c YZ |
2565 | |
2566 | swap(ctx->task_ctx_data, next_ctx->task_ctx_data); | |
2567 | ||
c93f7669 | 2568 | do_switch = 0; |
bfbd3381 | 2569 | |
cdd6c482 | 2570 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2571 | } |
e625cce1 TG |
2572 | raw_spin_unlock(&next_ctx->lock); |
2573 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2574 | } |
5a3126d4 | 2575 | unlock: |
c93f7669 | 2576 | rcu_read_unlock(); |
564c2b21 | 2577 | |
c93f7669 | 2578 | if (do_switch) { |
facc4307 | 2579 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2580 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2581 | cpuctx->task_ctx = NULL; |
facc4307 | 2582 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2583 | } |
0793a61d TG |
2584 | } |
2585 | ||
ba532500 YZ |
2586 | void perf_sched_cb_dec(struct pmu *pmu) |
2587 | { | |
2588 | this_cpu_dec(perf_sched_cb_usages); | |
2589 | } | |
2590 | ||
2591 | void perf_sched_cb_inc(struct pmu *pmu) | |
2592 | { | |
2593 | this_cpu_inc(perf_sched_cb_usages); | |
2594 | } | |
2595 | ||
2596 | /* | |
2597 | * This function provides the context switch callback to the lower code | |
2598 | * layer. It is invoked ONLY when the context switch callback is enabled. | |
2599 | */ | |
2600 | static void perf_pmu_sched_task(struct task_struct *prev, | |
2601 | struct task_struct *next, | |
2602 | bool sched_in) | |
2603 | { | |
2604 | struct perf_cpu_context *cpuctx; | |
2605 | struct pmu *pmu; | |
2606 | unsigned long flags; | |
2607 | ||
2608 | if (prev == next) | |
2609 | return; | |
2610 | ||
2611 | local_irq_save(flags); | |
2612 | ||
2613 | rcu_read_lock(); | |
2614 | ||
2615 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2616 | if (pmu->sched_task) { | |
2617 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2618 | ||
2619 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2620 | ||
2621 | perf_pmu_disable(pmu); | |
2622 | ||
2623 | pmu->sched_task(cpuctx->task_ctx, sched_in); | |
2624 | ||
2625 | perf_pmu_enable(pmu); | |
2626 | ||
2627 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2628 | } | |
2629 | } | |
2630 | ||
2631 | rcu_read_unlock(); | |
2632 | ||
2633 | local_irq_restore(flags); | |
2634 | } | |
2635 | ||
8dc85d54 PZ |
2636 | #define for_each_task_context_nr(ctxn) \ |
2637 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2638 | ||
2639 | /* | |
2640 | * Called from scheduler to remove the events of the current task, | |
2641 | * with interrupts disabled. | |
2642 | * | |
2643 | * We stop each event and update the event value in event->count. | |
2644 | * | |
2645 | * This does not protect us against NMI, but disable() | |
2646 | * sets the disabled bit in the control field of event _before_ | |
2647 | * accessing the event control register. If a NMI hits, then it will | |
2648 | * not restart the event. | |
2649 | */ | |
ab0cce56 JO |
2650 | void __perf_event_task_sched_out(struct task_struct *task, |
2651 | struct task_struct *next) | |
8dc85d54 PZ |
2652 | { |
2653 | int ctxn; | |
2654 | ||
ba532500 YZ |
2655 | if (__this_cpu_read(perf_sched_cb_usages)) |
2656 | perf_pmu_sched_task(task, next, false); | |
2657 | ||
8dc85d54 PZ |
2658 | for_each_task_context_nr(ctxn) |
2659 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2660 | |
2661 | /* | |
2662 | * if cgroup events exist on this CPU, then we need | |
2663 | * to check if we have to switch out PMU state. | |
2664 | * cgroup event are system-wide mode only | |
2665 | */ | |
4a32fea9 | 2666 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2667 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2668 | } |
2669 | ||
04dc2dbb | 2670 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2671 | { |
108b02cf | 2672 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2673 | |
a63eaf34 PM |
2674 | if (!cpuctx->task_ctx) |
2675 | return; | |
012b84da IM |
2676 | |
2677 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2678 | return; | |
2679 | ||
04dc2dbb | 2680 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2681 | cpuctx->task_ctx = NULL; |
2682 | } | |
2683 | ||
5b0311e1 FW |
2684 | /* |
2685 | * Called with IRQs disabled | |
2686 | */ | |
2687 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2688 | enum event_type_t event_type) | |
2689 | { | |
2690 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2691 | } |
2692 | ||
235c7fc7 | 2693 | static void |
5b0311e1 | 2694 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2695 | struct perf_cpu_context *cpuctx) |
0793a61d | 2696 | { |
cdd6c482 | 2697 | struct perf_event *event; |
0793a61d | 2698 | |
889ff015 FW |
2699 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2700 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2701 | continue; |
5632ab12 | 2702 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2703 | continue; |
2704 | ||
e5d1367f SE |
2705 | /* may need to reset tstamp_enabled */ |
2706 | if (is_cgroup_event(event)) | |
2707 | perf_cgroup_mark_enabled(event, ctx); | |
2708 | ||
8c9ed8e1 | 2709 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2710 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2711 | |
2712 | /* | |
2713 | * If this pinned group hasn't been scheduled, | |
2714 | * put it in error state. | |
2715 | */ | |
cdd6c482 IM |
2716 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2717 | update_group_times(event); | |
2718 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2719 | } |
3b6f9e5c | 2720 | } |
5b0311e1 FW |
2721 | } |
2722 | ||
2723 | static void | |
2724 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2725 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2726 | { |
2727 | struct perf_event *event; | |
2728 | int can_add_hw = 1; | |
3b6f9e5c | 2729 | |
889ff015 FW |
2730 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2731 | /* Ignore events in OFF or ERROR state */ | |
2732 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2733 | continue; |
04289bb9 IM |
2734 | /* |
2735 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2736 | * of events: |
04289bb9 | 2737 | */ |
5632ab12 | 2738 | if (!event_filter_match(event)) |
0793a61d TG |
2739 | continue; |
2740 | ||
e5d1367f SE |
2741 | /* may need to reset tstamp_enabled */ |
2742 | if (is_cgroup_event(event)) | |
2743 | perf_cgroup_mark_enabled(event, ctx); | |
2744 | ||
9ed6060d | 2745 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2746 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2747 | can_add_hw = 0; |
9ed6060d | 2748 | } |
0793a61d | 2749 | } |
5b0311e1 FW |
2750 | } |
2751 | ||
2752 | static void | |
2753 | ctx_sched_in(struct perf_event_context *ctx, | |
2754 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2755 | enum event_type_t event_type, |
2756 | struct task_struct *task) | |
5b0311e1 | 2757 | { |
e5d1367f | 2758 | u64 now; |
db24d33e | 2759 | int is_active = ctx->is_active; |
e5d1367f | 2760 | |
db24d33e | 2761 | ctx->is_active |= event_type; |
5b0311e1 | 2762 | if (likely(!ctx->nr_events)) |
facc4307 | 2763 | return; |
5b0311e1 | 2764 | |
e5d1367f SE |
2765 | now = perf_clock(); |
2766 | ctx->timestamp = now; | |
3f7cce3c | 2767 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2768 | /* |
2769 | * First go through the list and put on any pinned groups | |
2770 | * in order to give them the best chance of going on. | |
2771 | */ | |
db24d33e | 2772 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2773 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2774 | |
2775 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2776 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2777 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2778 | } |
2779 | ||
329c0e01 | 2780 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2781 | enum event_type_t event_type, |
2782 | struct task_struct *task) | |
329c0e01 FW |
2783 | { |
2784 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2785 | ||
e5d1367f | 2786 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2787 | } |
2788 | ||
e5d1367f SE |
2789 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2790 | struct task_struct *task) | |
235c7fc7 | 2791 | { |
108b02cf | 2792 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2793 | |
108b02cf | 2794 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2795 | if (cpuctx->task_ctx == ctx) |
2796 | return; | |
2797 | ||
facc4307 | 2798 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2799 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2800 | /* |
2801 | * We want to keep the following priority order: | |
2802 | * cpu pinned (that don't need to move), task pinned, | |
2803 | * cpu flexible, task flexible. | |
2804 | */ | |
2805 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2806 | ||
1d5f003f GN |
2807 | if (ctx->nr_events) |
2808 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2809 | |
86b47c25 GN |
2810 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2811 | ||
facc4307 PZ |
2812 | perf_pmu_enable(ctx->pmu); |
2813 | perf_ctx_unlock(cpuctx, ctx); | |
235c7fc7 IM |
2814 | } |
2815 | ||
8dc85d54 PZ |
2816 | /* |
2817 | * Called from scheduler to add the events of the current task | |
2818 | * with interrupts disabled. | |
2819 | * | |
2820 | * We restore the event value and then enable it. | |
2821 | * | |
2822 | * This does not protect us against NMI, but enable() | |
2823 | * sets the enabled bit in the control field of event _before_ | |
2824 | * accessing the event control register. If a NMI hits, then it will | |
2825 | * keep the event running. | |
2826 | */ | |
ab0cce56 JO |
2827 | void __perf_event_task_sched_in(struct task_struct *prev, |
2828 | struct task_struct *task) | |
8dc85d54 PZ |
2829 | { |
2830 | struct perf_event_context *ctx; | |
2831 | int ctxn; | |
2832 | ||
2833 | for_each_task_context_nr(ctxn) { | |
2834 | ctx = task->perf_event_ctxp[ctxn]; | |
2835 | if (likely(!ctx)) | |
2836 | continue; | |
2837 | ||
e5d1367f | 2838 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2839 | } |
e5d1367f SE |
2840 | /* |
2841 | * if cgroup events exist on this CPU, then we need | |
2842 | * to check if we have to switch in PMU state. | |
2843 | * cgroup event are system-wide mode only | |
2844 | */ | |
4a32fea9 | 2845 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2846 | perf_cgroup_sched_in(prev, task); |
d010b332 | 2847 | |
ba532500 YZ |
2848 | if (__this_cpu_read(perf_sched_cb_usages)) |
2849 | perf_pmu_sched_task(prev, task, true); | |
235c7fc7 IM |
2850 | } |
2851 | ||
abd50713 PZ |
2852 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2853 | { | |
2854 | u64 frequency = event->attr.sample_freq; | |
2855 | u64 sec = NSEC_PER_SEC; | |
2856 | u64 divisor, dividend; | |
2857 | ||
2858 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2859 | ||
2860 | count_fls = fls64(count); | |
2861 | nsec_fls = fls64(nsec); | |
2862 | frequency_fls = fls64(frequency); | |
2863 | sec_fls = 30; | |
2864 | ||
2865 | /* | |
2866 | * We got @count in @nsec, with a target of sample_freq HZ | |
2867 | * the target period becomes: | |
2868 | * | |
2869 | * @count * 10^9 | |
2870 | * period = ------------------- | |
2871 | * @nsec * sample_freq | |
2872 | * | |
2873 | */ | |
2874 | ||
2875 | /* | |
2876 | * Reduce accuracy by one bit such that @a and @b converge | |
2877 | * to a similar magnitude. | |
2878 | */ | |
fe4b04fa | 2879 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2880 | do { \ |
2881 | if (a##_fls > b##_fls) { \ | |
2882 | a >>= 1; \ | |
2883 | a##_fls--; \ | |
2884 | } else { \ | |
2885 | b >>= 1; \ | |
2886 | b##_fls--; \ | |
2887 | } \ | |
2888 | } while (0) | |
2889 | ||
2890 | /* | |
2891 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2892 | * the other, so that finally we can do a u64/u64 division. | |
2893 | */ | |
2894 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2895 | REDUCE_FLS(nsec, frequency); | |
2896 | REDUCE_FLS(sec, count); | |
2897 | } | |
2898 | ||
2899 | if (count_fls + sec_fls > 64) { | |
2900 | divisor = nsec * frequency; | |
2901 | ||
2902 | while (count_fls + sec_fls > 64) { | |
2903 | REDUCE_FLS(count, sec); | |
2904 | divisor >>= 1; | |
2905 | } | |
2906 | ||
2907 | dividend = count * sec; | |
2908 | } else { | |
2909 | dividend = count * sec; | |
2910 | ||
2911 | while (nsec_fls + frequency_fls > 64) { | |
2912 | REDUCE_FLS(nsec, frequency); | |
2913 | dividend >>= 1; | |
2914 | } | |
2915 | ||
2916 | divisor = nsec * frequency; | |
2917 | } | |
2918 | ||
f6ab91ad PZ |
2919 | if (!divisor) |
2920 | return dividend; | |
2921 | ||
abd50713 PZ |
2922 | return div64_u64(dividend, divisor); |
2923 | } | |
2924 | ||
e050e3f0 SE |
2925 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2926 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2927 | ||
f39d47ff | 2928 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2929 | { |
cdd6c482 | 2930 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2931 | s64 period, sample_period; |
bd2b5b12 PZ |
2932 | s64 delta; |
2933 | ||
abd50713 | 2934 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2935 | |
2936 | delta = (s64)(period - hwc->sample_period); | |
2937 | delta = (delta + 7) / 8; /* low pass filter */ | |
2938 | ||
2939 | sample_period = hwc->sample_period + delta; | |
2940 | ||
2941 | if (!sample_period) | |
2942 | sample_period = 1; | |
2943 | ||
bd2b5b12 | 2944 | hwc->sample_period = sample_period; |
abd50713 | 2945 | |
e7850595 | 2946 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2947 | if (disable) |
2948 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2949 | ||
e7850595 | 2950 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2951 | |
2952 | if (disable) | |
2953 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2954 | } |
bd2b5b12 PZ |
2955 | } |
2956 | ||
e050e3f0 SE |
2957 | /* |
2958 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2959 | * events. At the same time, make sure, having freq events does not change | |
2960 | * the rate of unthrottling as that would introduce bias. | |
2961 | */ | |
2962 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2963 | int needs_unthr) | |
60db5e09 | 2964 | { |
cdd6c482 IM |
2965 | struct perf_event *event; |
2966 | struct hw_perf_event *hwc; | |
e050e3f0 | 2967 | u64 now, period = TICK_NSEC; |
abd50713 | 2968 | s64 delta; |
60db5e09 | 2969 | |
e050e3f0 SE |
2970 | /* |
2971 | * only need to iterate over all events iff: | |
2972 | * - context have events in frequency mode (needs freq adjust) | |
2973 | * - there are events to unthrottle on this cpu | |
2974 | */ | |
2975 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2976 | return; |
2977 | ||
e050e3f0 | 2978 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2979 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2980 | |
03541f8b | 2981 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2982 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2983 | continue; |
2984 | ||
5632ab12 | 2985 | if (!event_filter_match(event)) |
5d27c23d PZ |
2986 | continue; |
2987 | ||
44377277 AS |
2988 | perf_pmu_disable(event->pmu); |
2989 | ||
cdd6c482 | 2990 | hwc = &event->hw; |
6a24ed6c | 2991 | |
ae23bff1 | 2992 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2993 | hwc->interrupts = 0; |
cdd6c482 | 2994 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2995 | event->pmu->start(event, 0); |
a78ac325 PZ |
2996 | } |
2997 | ||
cdd6c482 | 2998 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2999 | goto next; |
60db5e09 | 3000 | |
e050e3f0 SE |
3001 | /* |
3002 | * stop the event and update event->count | |
3003 | */ | |
3004 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3005 | ||
e7850595 | 3006 | now = local64_read(&event->count); |
abd50713 PZ |
3007 | delta = now - hwc->freq_count_stamp; |
3008 | hwc->freq_count_stamp = now; | |
60db5e09 | 3009 | |
e050e3f0 SE |
3010 | /* |
3011 | * restart the event | |
3012 | * reload only if value has changed | |
f39d47ff SE |
3013 | * we have stopped the event so tell that |
3014 | * to perf_adjust_period() to avoid stopping it | |
3015 | * twice. | |
e050e3f0 | 3016 | */ |
abd50713 | 3017 | if (delta > 0) |
f39d47ff | 3018 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
3019 | |
3020 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
3021 | next: |
3022 | perf_pmu_enable(event->pmu); | |
60db5e09 | 3023 | } |
e050e3f0 | 3024 | |
f39d47ff | 3025 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 3026 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
3027 | } |
3028 | ||
235c7fc7 | 3029 | /* |
cdd6c482 | 3030 | * Round-robin a context's events: |
235c7fc7 | 3031 | */ |
cdd6c482 | 3032 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 3033 | { |
dddd3379 TG |
3034 | /* |
3035 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
3036 | * disabled by the inheritance code. | |
3037 | */ | |
3038 | if (!ctx->rotate_disable) | |
3039 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
3040 | } |
3041 | ||
9e630205 | 3042 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 3043 | { |
8dc85d54 | 3044 | struct perf_event_context *ctx = NULL; |
2fde4f94 | 3045 | int rotate = 0; |
7fc23a53 | 3046 | |
b5ab4cd5 | 3047 | if (cpuctx->ctx.nr_events) { |
b5ab4cd5 PZ |
3048 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
3049 | rotate = 1; | |
3050 | } | |
235c7fc7 | 3051 | |
8dc85d54 | 3052 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 3053 | if (ctx && ctx->nr_events) { |
b5ab4cd5 PZ |
3054 | if (ctx->nr_events != ctx->nr_active) |
3055 | rotate = 1; | |
3056 | } | |
9717e6cd | 3057 | |
e050e3f0 | 3058 | if (!rotate) |
0f5a2601 PZ |
3059 | goto done; |
3060 | ||
facc4307 | 3061 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 3062 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 3063 | |
e050e3f0 SE |
3064 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
3065 | if (ctx) | |
3066 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 3067 | |
e050e3f0 SE |
3068 | rotate_ctx(&cpuctx->ctx); |
3069 | if (ctx) | |
3070 | rotate_ctx(ctx); | |
235c7fc7 | 3071 | |
e050e3f0 | 3072 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 3073 | |
0f5a2601 PZ |
3074 | perf_pmu_enable(cpuctx->ctx.pmu); |
3075 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 3076 | done: |
9e630205 SE |
3077 | |
3078 | return rotate; | |
e9d2b064 PZ |
3079 | } |
3080 | ||
026249ef FW |
3081 | #ifdef CONFIG_NO_HZ_FULL |
3082 | bool perf_event_can_stop_tick(void) | |
3083 | { | |
948b26b6 | 3084 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 3085 | __this_cpu_read(perf_throttled_count)) |
026249ef | 3086 | return false; |
d84153d6 FW |
3087 | else |
3088 | return true; | |
026249ef FW |
3089 | } |
3090 | #endif | |
3091 | ||
e9d2b064 PZ |
3092 | void perf_event_task_tick(void) |
3093 | { | |
2fde4f94 MR |
3094 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
3095 | struct perf_event_context *ctx, *tmp; | |
e050e3f0 | 3096 | int throttled; |
b5ab4cd5 | 3097 | |
e9d2b064 PZ |
3098 | WARN_ON(!irqs_disabled()); |
3099 | ||
e050e3f0 SE |
3100 | __this_cpu_inc(perf_throttled_seq); |
3101 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
3102 | ||
2fde4f94 | 3103 | list_for_each_entry_safe(ctx, tmp, head, active_ctx_list) |
e050e3f0 | 3104 | perf_adjust_freq_unthr_context(ctx, throttled); |
0793a61d TG |
3105 | } |
3106 | ||
889ff015 FW |
3107 | static int event_enable_on_exec(struct perf_event *event, |
3108 | struct perf_event_context *ctx) | |
3109 | { | |
3110 | if (!event->attr.enable_on_exec) | |
3111 | return 0; | |
3112 | ||
3113 | event->attr.enable_on_exec = 0; | |
3114 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
3115 | return 0; | |
3116 | ||
1d9b482e | 3117 | __perf_event_mark_enabled(event); |
889ff015 FW |
3118 | |
3119 | return 1; | |
3120 | } | |
3121 | ||
57e7986e | 3122 | /* |
cdd6c482 | 3123 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
3124 | * This expects task == current. |
3125 | */ | |
8dc85d54 | 3126 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 3127 | { |
211de6eb | 3128 | struct perf_event_context *clone_ctx = NULL; |
cdd6c482 | 3129 | struct perf_event *event; |
57e7986e PM |
3130 | unsigned long flags; |
3131 | int enabled = 0; | |
889ff015 | 3132 | int ret; |
57e7986e PM |
3133 | |
3134 | local_irq_save(flags); | |
cdd6c482 | 3135 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
3136 | goto out; |
3137 | ||
e566b76e SE |
3138 | /* |
3139 | * We must ctxsw out cgroup events to avoid conflict | |
3140 | * when invoking perf_task_event_sched_in() later on | |
3141 | * in this function. Otherwise we end up trying to | |
3142 | * ctxswin cgroup events which are already scheduled | |
3143 | * in. | |
3144 | */ | |
a8d757ef | 3145 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 3146 | |
e625cce1 | 3147 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 3148 | task_ctx_sched_out(ctx); |
57e7986e | 3149 | |
b79387ef | 3150 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
3151 | ret = event_enable_on_exec(event, ctx); |
3152 | if (ret) | |
3153 | enabled = 1; | |
57e7986e PM |
3154 | } |
3155 | ||
3156 | /* | |
cdd6c482 | 3157 | * Unclone this context if we enabled any event. |
57e7986e | 3158 | */ |
71a851b4 | 3159 | if (enabled) |
211de6eb | 3160 | clone_ctx = unclone_ctx(ctx); |
57e7986e | 3161 | |
e625cce1 | 3162 | raw_spin_unlock(&ctx->lock); |
57e7986e | 3163 | |
e566b76e SE |
3164 | /* |
3165 | * Also calls ctxswin for cgroup events, if any: | |
3166 | */ | |
e5d1367f | 3167 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 3168 | out: |
57e7986e | 3169 | local_irq_restore(flags); |
211de6eb PZ |
3170 | |
3171 | if (clone_ctx) | |
3172 | put_ctx(clone_ctx); | |
57e7986e PM |
3173 | } |
3174 | ||
e041e328 PZ |
3175 | void perf_event_exec(void) |
3176 | { | |
3177 | struct perf_event_context *ctx; | |
3178 | int ctxn; | |
3179 | ||
3180 | rcu_read_lock(); | |
3181 | for_each_task_context_nr(ctxn) { | |
3182 | ctx = current->perf_event_ctxp[ctxn]; | |
3183 | if (!ctx) | |
3184 | continue; | |
3185 | ||
3186 | perf_event_enable_on_exec(ctx); | |
3187 | } | |
3188 | rcu_read_unlock(); | |
3189 | } | |
3190 | ||
0793a61d | 3191 | /* |
cdd6c482 | 3192 | * Cross CPU call to read the hardware event |
0793a61d | 3193 | */ |
cdd6c482 | 3194 | static void __perf_event_read(void *info) |
0793a61d | 3195 | { |
cdd6c482 IM |
3196 | struct perf_event *event = info; |
3197 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 3198 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3199 | |
e1ac3614 PM |
3200 | /* |
3201 | * If this is a task context, we need to check whether it is | |
3202 | * the current task context of this cpu. If not it has been | |
3203 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3204 | * event->count would have been updated to a recent sample |
3205 | * when the event was scheduled out. | |
e1ac3614 PM |
3206 | */ |
3207 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3208 | return; | |
3209 | ||
e625cce1 | 3210 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3211 | if (ctx->is_active) { |
542e72fc | 3212 | update_context_time(ctx); |
e5d1367f SE |
3213 | update_cgrp_time_from_event(event); |
3214 | } | |
cdd6c482 | 3215 | update_event_times(event); |
542e72fc PZ |
3216 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3217 | event->pmu->read(event); | |
e625cce1 | 3218 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3219 | } |
3220 | ||
b5e58793 PZ |
3221 | static inline u64 perf_event_count(struct perf_event *event) |
3222 | { | |
e7850595 | 3223 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
3224 | } |
3225 | ||
cdd6c482 | 3226 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3227 | { |
3228 | /* | |
cdd6c482 IM |
3229 | * If event is enabled and currently active on a CPU, update the |
3230 | * value in the event structure: | |
0793a61d | 3231 | */ |
cdd6c482 IM |
3232 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3233 | smp_call_function_single(event->oncpu, | |
3234 | __perf_event_read, event, 1); | |
3235 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3236 | struct perf_event_context *ctx = event->ctx; |
3237 | unsigned long flags; | |
3238 | ||
e625cce1 | 3239 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3240 | /* |
3241 | * may read while context is not active | |
3242 | * (e.g., thread is blocked), in that case | |
3243 | * we cannot update context time | |
3244 | */ | |
e5d1367f | 3245 | if (ctx->is_active) { |
c530ccd9 | 3246 | update_context_time(ctx); |
e5d1367f SE |
3247 | update_cgrp_time_from_event(event); |
3248 | } | |
cdd6c482 | 3249 | update_event_times(event); |
e625cce1 | 3250 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3251 | } |
3252 | ||
b5e58793 | 3253 | return perf_event_count(event); |
0793a61d TG |
3254 | } |
3255 | ||
a63eaf34 | 3256 | /* |
cdd6c482 | 3257 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3258 | */ |
eb184479 | 3259 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3260 | { |
e625cce1 | 3261 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3262 | mutex_init(&ctx->mutex); |
2fde4f94 | 3263 | INIT_LIST_HEAD(&ctx->active_ctx_list); |
889ff015 FW |
3264 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3265 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3266 | INIT_LIST_HEAD(&ctx->event_list); |
3267 | atomic_set(&ctx->refcount, 1); | |
fadfe7be | 3268 | INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work); |
eb184479 PZ |
3269 | } |
3270 | ||
3271 | static struct perf_event_context * | |
3272 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3273 | { | |
3274 | struct perf_event_context *ctx; | |
3275 | ||
3276 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3277 | if (!ctx) | |
3278 | return NULL; | |
3279 | ||
3280 | __perf_event_init_context(ctx); | |
3281 | if (task) { | |
3282 | ctx->task = task; | |
3283 | get_task_struct(task); | |
0793a61d | 3284 | } |
eb184479 PZ |
3285 | ctx->pmu = pmu; |
3286 | ||
3287 | return ctx; | |
a63eaf34 PM |
3288 | } |
3289 | ||
2ebd4ffb MH |
3290 | static struct task_struct * |
3291 | find_lively_task_by_vpid(pid_t vpid) | |
3292 | { | |
3293 | struct task_struct *task; | |
3294 | int err; | |
0793a61d TG |
3295 | |
3296 | rcu_read_lock(); | |
2ebd4ffb | 3297 | if (!vpid) |
0793a61d TG |
3298 | task = current; |
3299 | else | |
2ebd4ffb | 3300 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3301 | if (task) |
3302 | get_task_struct(task); | |
3303 | rcu_read_unlock(); | |
3304 | ||
3305 | if (!task) | |
3306 | return ERR_PTR(-ESRCH); | |
3307 | ||
0793a61d | 3308 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3309 | err = -EACCES; |
3310 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3311 | goto errout; | |
3312 | ||
2ebd4ffb MH |
3313 | return task; |
3314 | errout: | |
3315 | put_task_struct(task); | |
3316 | return ERR_PTR(err); | |
3317 | ||
3318 | } | |
3319 | ||
fe4b04fa PZ |
3320 | /* |
3321 | * Returns a matching context with refcount and pincount. | |
3322 | */ | |
108b02cf | 3323 | static struct perf_event_context * |
4af57ef2 YZ |
3324 | find_get_context(struct pmu *pmu, struct task_struct *task, |
3325 | struct perf_event *event) | |
0793a61d | 3326 | { |
211de6eb | 3327 | struct perf_event_context *ctx, *clone_ctx = NULL; |
22a4f650 | 3328 | struct perf_cpu_context *cpuctx; |
4af57ef2 | 3329 | void *task_ctx_data = NULL; |
25346b93 | 3330 | unsigned long flags; |
8dc85d54 | 3331 | int ctxn, err; |
4af57ef2 | 3332 | int cpu = event->cpu; |
0793a61d | 3333 | |
22a4ec72 | 3334 | if (!task) { |
cdd6c482 | 3335 | /* Must be root to operate on a CPU event: */ |
0764771d | 3336 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3337 | return ERR_PTR(-EACCES); |
3338 | ||
0793a61d | 3339 | /* |
cdd6c482 | 3340 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3341 | * offline CPU and activate it when the CPU comes up, but |
3342 | * that's for later. | |
3343 | */ | |
f6325e30 | 3344 | if (!cpu_online(cpu)) |
0793a61d TG |
3345 | return ERR_PTR(-ENODEV); |
3346 | ||
108b02cf | 3347 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3348 | ctx = &cpuctx->ctx; |
c93f7669 | 3349 | get_ctx(ctx); |
fe4b04fa | 3350 | ++ctx->pin_count; |
0793a61d | 3351 | |
0793a61d TG |
3352 | return ctx; |
3353 | } | |
3354 | ||
8dc85d54 PZ |
3355 | err = -EINVAL; |
3356 | ctxn = pmu->task_ctx_nr; | |
3357 | if (ctxn < 0) | |
3358 | goto errout; | |
3359 | ||
4af57ef2 YZ |
3360 | if (event->attach_state & PERF_ATTACH_TASK_DATA) { |
3361 | task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL); | |
3362 | if (!task_ctx_data) { | |
3363 | err = -ENOMEM; | |
3364 | goto errout; | |
3365 | } | |
3366 | } | |
3367 | ||
9ed6060d | 3368 | retry: |
8dc85d54 | 3369 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3370 | if (ctx) { |
211de6eb | 3371 | clone_ctx = unclone_ctx(ctx); |
fe4b04fa | 3372 | ++ctx->pin_count; |
4af57ef2 YZ |
3373 | |
3374 | if (task_ctx_data && !ctx->task_ctx_data) { | |
3375 | ctx->task_ctx_data = task_ctx_data; | |
3376 | task_ctx_data = NULL; | |
3377 | } | |
e625cce1 | 3378 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
211de6eb PZ |
3379 | |
3380 | if (clone_ctx) | |
3381 | put_ctx(clone_ctx); | |
9137fb28 | 3382 | } else { |
eb184479 | 3383 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3384 | err = -ENOMEM; |
3385 | if (!ctx) | |
3386 | goto errout; | |
eb184479 | 3387 | |
4af57ef2 YZ |
3388 | if (task_ctx_data) { |
3389 | ctx->task_ctx_data = task_ctx_data; | |
3390 | task_ctx_data = NULL; | |
3391 | } | |
3392 | ||
dbe08d82 ON |
3393 | err = 0; |
3394 | mutex_lock(&task->perf_event_mutex); | |
3395 | /* | |
3396 | * If it has already passed perf_event_exit_task(). | |
3397 | * we must see PF_EXITING, it takes this mutex too. | |
3398 | */ | |
3399 | if (task->flags & PF_EXITING) | |
3400 | err = -ESRCH; | |
3401 | else if (task->perf_event_ctxp[ctxn]) | |
3402 | err = -EAGAIN; | |
fe4b04fa | 3403 | else { |
9137fb28 | 3404 | get_ctx(ctx); |
fe4b04fa | 3405 | ++ctx->pin_count; |
dbe08d82 | 3406 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3407 | } |
dbe08d82 ON |
3408 | mutex_unlock(&task->perf_event_mutex); |
3409 | ||
3410 | if (unlikely(err)) { | |
9137fb28 | 3411 | put_ctx(ctx); |
dbe08d82 ON |
3412 | |
3413 | if (err == -EAGAIN) | |
3414 | goto retry; | |
3415 | goto errout; | |
a63eaf34 PM |
3416 | } |
3417 | } | |
3418 | ||
4af57ef2 | 3419 | kfree(task_ctx_data); |
0793a61d | 3420 | return ctx; |
c93f7669 | 3421 | |
9ed6060d | 3422 | errout: |
4af57ef2 | 3423 | kfree(task_ctx_data); |
c93f7669 | 3424 | return ERR_PTR(err); |
0793a61d TG |
3425 | } |
3426 | ||
6fb2915d LZ |
3427 | static void perf_event_free_filter(struct perf_event *event); |
3428 | ||
cdd6c482 | 3429 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3430 | { |
cdd6c482 | 3431 | struct perf_event *event; |
592903cd | 3432 | |
cdd6c482 IM |
3433 | event = container_of(head, struct perf_event, rcu_head); |
3434 | if (event->ns) | |
3435 | put_pid_ns(event->ns); | |
6fb2915d | 3436 | perf_event_free_filter(event); |
cdd6c482 | 3437 | kfree(event); |
592903cd PZ |
3438 | } |
3439 | ||
76369139 | 3440 | static void ring_buffer_put(struct ring_buffer *rb); |
b69cf536 PZ |
3441 | static void ring_buffer_attach(struct perf_event *event, |
3442 | struct ring_buffer *rb); | |
925d519a | 3443 | |
4beb31f3 | 3444 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3445 | { |
4beb31f3 FW |
3446 | if (event->parent) |
3447 | return; | |
3448 | ||
4beb31f3 FW |
3449 | if (is_cgroup_event(event)) |
3450 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3451 | } | |
925d519a | 3452 | |
4beb31f3 FW |
3453 | static void unaccount_event(struct perf_event *event) |
3454 | { | |
3455 | if (event->parent) | |
3456 | return; | |
3457 | ||
3458 | if (event->attach_state & PERF_ATTACH_TASK) | |
3459 | static_key_slow_dec_deferred(&perf_sched_events); | |
3460 | if (event->attr.mmap || event->attr.mmap_data) | |
3461 | atomic_dec(&nr_mmap_events); | |
3462 | if (event->attr.comm) | |
3463 | atomic_dec(&nr_comm_events); | |
3464 | if (event->attr.task) | |
3465 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3466 | if (event->attr.freq) |
3467 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3468 | if (is_cgroup_event(event)) |
3469 | static_key_slow_dec_deferred(&perf_sched_events); | |
3470 | if (has_branch_stack(event)) | |
3471 | static_key_slow_dec_deferred(&perf_sched_events); | |
3472 | ||
3473 | unaccount_event_cpu(event, event->cpu); | |
3474 | } | |
925d519a | 3475 | |
766d6c07 FW |
3476 | static void __free_event(struct perf_event *event) |
3477 | { | |
cdd6c482 | 3478 | if (!event->parent) { |
927c7a9e FW |
3479 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3480 | put_callchain_buffers(); | |
f344011c | 3481 | } |
9ee318a7 | 3482 | |
766d6c07 FW |
3483 | if (event->destroy) |
3484 | event->destroy(event); | |
3485 | ||
3486 | if (event->ctx) | |
3487 | put_ctx(event->ctx); | |
3488 | ||
c464c76e YZ |
3489 | if (event->pmu) |
3490 | module_put(event->pmu->module); | |
3491 | ||
766d6c07 FW |
3492 | call_rcu(&event->rcu_head, free_event_rcu); |
3493 | } | |
683ede43 PZ |
3494 | |
3495 | static void _free_event(struct perf_event *event) | |
f1600952 | 3496 | { |
e360adbe | 3497 | irq_work_sync(&event->pending); |
925d519a | 3498 | |
4beb31f3 | 3499 | unaccount_event(event); |
9ee318a7 | 3500 | |
76369139 | 3501 | if (event->rb) { |
9bb5d40c PZ |
3502 | /* |
3503 | * Can happen when we close an event with re-directed output. | |
3504 | * | |
3505 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3506 | * over us; possibly making our ring_buffer_put() the last. | |
3507 | */ | |
3508 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3509 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3510 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3511 | } |
3512 | ||
e5d1367f SE |
3513 | if (is_cgroup_event(event)) |
3514 | perf_detach_cgroup(event); | |
3515 | ||
766d6c07 | 3516 | __free_event(event); |
f1600952 PZ |
3517 | } |
3518 | ||
683ede43 PZ |
3519 | /* |
3520 | * Used to free events which have a known refcount of 1, such as in error paths | |
3521 | * where the event isn't exposed yet and inherited events. | |
3522 | */ | |
3523 | static void free_event(struct perf_event *event) | |
0793a61d | 3524 | { |
683ede43 PZ |
3525 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3526 | "unexpected event refcount: %ld; ptr=%p\n", | |
3527 | atomic_long_read(&event->refcount), event)) { | |
3528 | /* leak to avoid use-after-free */ | |
3529 | return; | |
3530 | } | |
0793a61d | 3531 | |
683ede43 | 3532 | _free_event(event); |
0793a61d TG |
3533 | } |
3534 | ||
a66a3052 | 3535 | /* |
f8697762 | 3536 | * Remove user event from the owner task. |
a66a3052 | 3537 | */ |
f8697762 | 3538 | static void perf_remove_from_owner(struct perf_event *event) |
fb0459d7 | 3539 | { |
8882135b | 3540 | struct task_struct *owner; |
fb0459d7 | 3541 | |
8882135b PZ |
3542 | rcu_read_lock(); |
3543 | owner = ACCESS_ONCE(event->owner); | |
3544 | /* | |
3545 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3546 | * !owner it means the list deletion is complete and we can indeed | |
3547 | * free this event, otherwise we need to serialize on | |
3548 | * owner->perf_event_mutex. | |
3549 | */ | |
3550 | smp_read_barrier_depends(); | |
3551 | if (owner) { | |
3552 | /* | |
3553 | * Since delayed_put_task_struct() also drops the last | |
3554 | * task reference we can safely take a new reference | |
3555 | * while holding the rcu_read_lock(). | |
3556 | */ | |
3557 | get_task_struct(owner); | |
3558 | } | |
3559 | rcu_read_unlock(); | |
3560 | ||
3561 | if (owner) { | |
f63a8daa PZ |
3562 | /* |
3563 | * If we're here through perf_event_exit_task() we're already | |
3564 | * holding ctx->mutex which would be an inversion wrt. the | |
3565 | * normal lock order. | |
3566 | * | |
3567 | * However we can safely take this lock because its the child | |
3568 | * ctx->mutex. | |
3569 | */ | |
3570 | mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); | |
3571 | ||
8882135b PZ |
3572 | /* |
3573 | * We have to re-check the event->owner field, if it is cleared | |
3574 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3575 | * ensured they're done, and we can proceed with freeing the | |
3576 | * event. | |
3577 | */ | |
3578 | if (event->owner) | |
3579 | list_del_init(&event->owner_entry); | |
3580 | mutex_unlock(&owner->perf_event_mutex); | |
3581 | put_task_struct(owner); | |
3582 | } | |
f8697762 JO |
3583 | } |
3584 | ||
3585 | /* | |
3586 | * Called when the last reference to the file is gone. | |
3587 | */ | |
3588 | static void put_event(struct perf_event *event) | |
3589 | { | |
a83fe28e | 3590 | struct perf_event_context *ctx; |
f8697762 JO |
3591 | |
3592 | if (!atomic_long_dec_and_test(&event->refcount)) | |
3593 | return; | |
3594 | ||
3595 | if (!is_kernel_event(event)) | |
3596 | perf_remove_from_owner(event); | |
8882135b | 3597 | |
683ede43 PZ |
3598 | /* |
3599 | * There are two ways this annotation is useful: | |
3600 | * | |
3601 | * 1) there is a lock recursion from perf_event_exit_task | |
3602 | * see the comment there. | |
3603 | * | |
3604 | * 2) there is a lock-inversion with mmap_sem through | |
3605 | * perf_event_read_group(), which takes faults while | |
3606 | * holding ctx->mutex, however this is called after | |
3607 | * the last filedesc died, so there is no possibility | |
3608 | * to trigger the AB-BA case. | |
3609 | */ | |
a83fe28e PZ |
3610 | ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING); |
3611 | WARN_ON_ONCE(ctx->parent_ctx); | |
683ede43 PZ |
3612 | perf_remove_from_context(event, true); |
3613 | mutex_unlock(&ctx->mutex); | |
3614 | ||
3615 | _free_event(event); | |
a6fa941d AV |
3616 | } |
3617 | ||
683ede43 PZ |
3618 | int perf_event_release_kernel(struct perf_event *event) |
3619 | { | |
3620 | put_event(event); | |
3621 | return 0; | |
3622 | } | |
3623 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3624 | ||
a6fa941d AV |
3625 | static int perf_release(struct inode *inode, struct file *file) |
3626 | { | |
3627 | put_event(file->private_data); | |
3628 | return 0; | |
fb0459d7 | 3629 | } |
fb0459d7 | 3630 | |
fadfe7be JO |
3631 | /* |
3632 | * Remove all orphanes events from the context. | |
3633 | */ | |
3634 | static void orphans_remove_work(struct work_struct *work) | |
3635 | { | |
3636 | struct perf_event_context *ctx; | |
3637 | struct perf_event *event, *tmp; | |
3638 | ||
3639 | ctx = container_of(work, struct perf_event_context, | |
3640 | orphans_remove.work); | |
3641 | ||
3642 | mutex_lock(&ctx->mutex); | |
3643 | list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) { | |
3644 | struct perf_event *parent_event = event->parent; | |
3645 | ||
3646 | if (!is_orphaned_child(event)) | |
3647 | continue; | |
3648 | ||
3649 | perf_remove_from_context(event, true); | |
3650 | ||
3651 | mutex_lock(&parent_event->child_mutex); | |
3652 | list_del_init(&event->child_list); | |
3653 | mutex_unlock(&parent_event->child_mutex); | |
3654 | ||
3655 | free_event(event); | |
3656 | put_event(parent_event); | |
3657 | } | |
3658 | ||
3659 | raw_spin_lock_irq(&ctx->lock); | |
3660 | ctx->orphans_remove_sched = false; | |
3661 | raw_spin_unlock_irq(&ctx->lock); | |
3662 | mutex_unlock(&ctx->mutex); | |
3663 | ||
3664 | put_ctx(ctx); | |
3665 | } | |
3666 | ||
59ed446f | 3667 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3668 | { |
cdd6c482 | 3669 | struct perf_event *child; |
e53c0994 PZ |
3670 | u64 total = 0; |
3671 | ||
59ed446f PZ |
3672 | *enabled = 0; |
3673 | *running = 0; | |
3674 | ||
6f10581a | 3675 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3676 | total += perf_event_read(event); |
59ed446f PZ |
3677 | *enabled += event->total_time_enabled + |
3678 | atomic64_read(&event->child_total_time_enabled); | |
3679 | *running += event->total_time_running + | |
3680 | atomic64_read(&event->child_total_time_running); | |
3681 | ||
3682 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3683 | total += perf_event_read(child); |
59ed446f PZ |
3684 | *enabled += child->total_time_enabled; |
3685 | *running += child->total_time_running; | |
3686 | } | |
6f10581a | 3687 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3688 | |
3689 | return total; | |
3690 | } | |
fb0459d7 | 3691 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3692 | |
cdd6c482 | 3693 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3694 | u64 read_format, char __user *buf) |
3695 | { | |
cdd6c482 | 3696 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a | 3697 | struct perf_event_context *ctx = leader->ctx; |
f63a8daa | 3698 | int n = 0, size = 0, ret; |
59ed446f | 3699 | u64 count, enabled, running; |
f63a8daa PZ |
3700 | u64 values[5]; |
3701 | ||
3702 | lockdep_assert_held(&ctx->mutex); | |
abf4868b | 3703 | |
59ed446f | 3704 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3705 | |
3706 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3707 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3708 | values[n++] = enabled; | |
3709 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3710 | values[n++] = running; | |
abf4868b PZ |
3711 | values[n++] = count; |
3712 | if (read_format & PERF_FORMAT_ID) | |
3713 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3714 | |
3715 | size = n * sizeof(u64); | |
3716 | ||
3717 | if (copy_to_user(buf, values, size)) | |
f63a8daa | 3718 | return -EFAULT; |
3dab77fb | 3719 | |
6f10581a | 3720 | ret = size; |
3dab77fb | 3721 | |
65abc865 | 3722 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3723 | n = 0; |
3dab77fb | 3724 | |
59ed446f | 3725 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3726 | if (read_format & PERF_FORMAT_ID) |
3727 | values[n++] = primary_event_id(sub); | |
3728 | ||
3729 | size = n * sizeof(u64); | |
3730 | ||
184d3da8 | 3731 | if (copy_to_user(buf + ret, values, size)) { |
f63a8daa | 3732 | return -EFAULT; |
6f10581a | 3733 | } |
abf4868b PZ |
3734 | |
3735 | ret += size; | |
3dab77fb PZ |
3736 | } |
3737 | ||
abf4868b | 3738 | return ret; |
3dab77fb PZ |
3739 | } |
3740 | ||
cdd6c482 | 3741 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3742 | u64 read_format, char __user *buf) |
3743 | { | |
59ed446f | 3744 | u64 enabled, running; |
3dab77fb PZ |
3745 | u64 values[4]; |
3746 | int n = 0; | |
3747 | ||
59ed446f PZ |
3748 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3749 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3750 | values[n++] = enabled; | |
3751 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3752 | values[n++] = running; | |
3dab77fb | 3753 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3754 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3755 | |
3756 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3757 | return -EFAULT; | |
3758 | ||
3759 | return n * sizeof(u64); | |
3760 | } | |
3761 | ||
dc633982 JO |
3762 | static bool is_event_hup(struct perf_event *event) |
3763 | { | |
3764 | bool no_children; | |
3765 | ||
3766 | if (event->state != PERF_EVENT_STATE_EXIT) | |
3767 | return false; | |
3768 | ||
3769 | mutex_lock(&event->child_mutex); | |
3770 | no_children = list_empty(&event->child_list); | |
3771 | mutex_unlock(&event->child_mutex); | |
3772 | return no_children; | |
3773 | } | |
3774 | ||
0793a61d | 3775 | /* |
cdd6c482 | 3776 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3777 | */ |
3778 | static ssize_t | |
cdd6c482 | 3779 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3780 | { |
cdd6c482 | 3781 | u64 read_format = event->attr.read_format; |
3dab77fb | 3782 | int ret; |
0793a61d | 3783 | |
3b6f9e5c | 3784 | /* |
cdd6c482 | 3785 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3786 | * error state (i.e. because it was pinned but it couldn't be |
3787 | * scheduled on to the CPU at some point). | |
3788 | */ | |
cdd6c482 | 3789 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3790 | return 0; |
3791 | ||
c320c7b7 | 3792 | if (count < event->read_size) |
3dab77fb PZ |
3793 | return -ENOSPC; |
3794 | ||
cdd6c482 | 3795 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3796 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3797 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3798 | else |
cdd6c482 | 3799 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3800 | |
3dab77fb | 3801 | return ret; |
0793a61d TG |
3802 | } |
3803 | ||
0793a61d TG |
3804 | static ssize_t |
3805 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3806 | { | |
cdd6c482 | 3807 | struct perf_event *event = file->private_data; |
f63a8daa PZ |
3808 | struct perf_event_context *ctx; |
3809 | int ret; | |
0793a61d | 3810 | |
f63a8daa PZ |
3811 | ctx = perf_event_ctx_lock(event); |
3812 | ret = perf_read_hw(event, buf, count); | |
3813 | perf_event_ctx_unlock(event, ctx); | |
3814 | ||
3815 | return ret; | |
0793a61d TG |
3816 | } |
3817 | ||
3818 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3819 | { | |
cdd6c482 | 3820 | struct perf_event *event = file->private_data; |
76369139 | 3821 | struct ring_buffer *rb; |
61b67684 | 3822 | unsigned int events = POLLHUP; |
c7138f37 | 3823 | |
e708d7ad | 3824 | poll_wait(file, &event->waitq, wait); |
179033b3 | 3825 | |
dc633982 | 3826 | if (is_event_hup(event)) |
179033b3 | 3827 | return events; |
c7138f37 | 3828 | |
10c6db11 | 3829 | /* |
9bb5d40c PZ |
3830 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3831 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3832 | */ |
3833 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3834 | rb = event->rb; |
3835 | if (rb) | |
76369139 | 3836 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3837 | mutex_unlock(&event->mmap_mutex); |
0793a61d TG |
3838 | return events; |
3839 | } | |
3840 | ||
f63a8daa | 3841 | static void _perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3842 | { |
cdd6c482 | 3843 | (void)perf_event_read(event); |
e7850595 | 3844 | local64_set(&event->count, 0); |
cdd6c482 | 3845 | perf_event_update_userpage(event); |
3df5edad PZ |
3846 | } |
3847 | ||
c93f7669 | 3848 | /* |
cdd6c482 IM |
3849 | * Holding the top-level event's child_mutex means that any |
3850 | * descendant process that has inherited this event will block | |
3851 | * in sync_child_event if it goes to exit, thus satisfying the | |
3852 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3853 | */ |
cdd6c482 IM |
3854 | static void perf_event_for_each_child(struct perf_event *event, |
3855 | void (*func)(struct perf_event *)) | |
3df5edad | 3856 | { |
cdd6c482 | 3857 | struct perf_event *child; |
3df5edad | 3858 | |
cdd6c482 | 3859 | WARN_ON_ONCE(event->ctx->parent_ctx); |
f63a8daa | 3860 | |
cdd6c482 IM |
3861 | mutex_lock(&event->child_mutex); |
3862 | func(event); | |
3863 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3864 | func(child); |
cdd6c482 | 3865 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3866 | } |
3867 | ||
cdd6c482 IM |
3868 | static void perf_event_for_each(struct perf_event *event, |
3869 | void (*func)(struct perf_event *)) | |
3df5edad | 3870 | { |
cdd6c482 IM |
3871 | struct perf_event_context *ctx = event->ctx; |
3872 | struct perf_event *sibling; | |
3df5edad | 3873 | |
f63a8daa PZ |
3874 | lockdep_assert_held(&ctx->mutex); |
3875 | ||
cdd6c482 | 3876 | event = event->group_leader; |
75f937f2 | 3877 | |
cdd6c482 | 3878 | perf_event_for_each_child(event, func); |
cdd6c482 | 3879 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3880 | perf_event_for_each_child(sibling, func); |
6de6a7b9 PZ |
3881 | } |
3882 | ||
cdd6c482 | 3883 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3884 | { |
cdd6c482 | 3885 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3886 | int ret = 0, active; |
08247e31 PZ |
3887 | u64 value; |
3888 | ||
6c7e550f | 3889 | if (!is_sampling_event(event)) |
08247e31 PZ |
3890 | return -EINVAL; |
3891 | ||
ad0cf347 | 3892 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3893 | return -EFAULT; |
3894 | ||
3895 | if (!value) | |
3896 | return -EINVAL; | |
3897 | ||
e625cce1 | 3898 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3899 | if (event->attr.freq) { |
3900 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3901 | ret = -EINVAL; |
3902 | goto unlock; | |
3903 | } | |
3904 | ||
cdd6c482 | 3905 | event->attr.sample_freq = value; |
08247e31 | 3906 | } else { |
cdd6c482 IM |
3907 | event->attr.sample_period = value; |
3908 | event->hw.sample_period = value; | |
08247e31 | 3909 | } |
bad7192b PZ |
3910 | |
3911 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3912 | if (active) { | |
3913 | perf_pmu_disable(ctx->pmu); | |
3914 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3915 | } | |
3916 | ||
3917 | local64_set(&event->hw.period_left, 0); | |
3918 | ||
3919 | if (active) { | |
3920 | event->pmu->start(event, PERF_EF_RELOAD); | |
3921 | perf_pmu_enable(ctx->pmu); | |
3922 | } | |
3923 | ||
08247e31 | 3924 | unlock: |
e625cce1 | 3925 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3926 | |
3927 | return ret; | |
3928 | } | |
3929 | ||
ac9721f3 PZ |
3930 | static const struct file_operations perf_fops; |
3931 | ||
2903ff01 | 3932 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3933 | { |
2903ff01 AV |
3934 | struct fd f = fdget(fd); |
3935 | if (!f.file) | |
3936 | return -EBADF; | |
ac9721f3 | 3937 | |
2903ff01 AV |
3938 | if (f.file->f_op != &perf_fops) { |
3939 | fdput(f); | |
3940 | return -EBADF; | |
ac9721f3 | 3941 | } |
2903ff01 AV |
3942 | *p = f; |
3943 | return 0; | |
ac9721f3 PZ |
3944 | } |
3945 | ||
3946 | static int perf_event_set_output(struct perf_event *event, | |
3947 | struct perf_event *output_event); | |
6fb2915d | 3948 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3949 | |
f63a8daa | 3950 | static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg) |
d859e29f | 3951 | { |
cdd6c482 | 3952 | void (*func)(struct perf_event *); |
3df5edad | 3953 | u32 flags = arg; |
d859e29f PM |
3954 | |
3955 | switch (cmd) { | |
cdd6c482 | 3956 | case PERF_EVENT_IOC_ENABLE: |
f63a8daa | 3957 | func = _perf_event_enable; |
d859e29f | 3958 | break; |
cdd6c482 | 3959 | case PERF_EVENT_IOC_DISABLE: |
f63a8daa | 3960 | func = _perf_event_disable; |
79f14641 | 3961 | break; |
cdd6c482 | 3962 | case PERF_EVENT_IOC_RESET: |
f63a8daa | 3963 | func = _perf_event_reset; |
6de6a7b9 | 3964 | break; |
3df5edad | 3965 | |
cdd6c482 | 3966 | case PERF_EVENT_IOC_REFRESH: |
f63a8daa | 3967 | return _perf_event_refresh(event, arg); |
08247e31 | 3968 | |
cdd6c482 IM |
3969 | case PERF_EVENT_IOC_PERIOD: |
3970 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3971 | |
cf4957f1 JO |
3972 | case PERF_EVENT_IOC_ID: |
3973 | { | |
3974 | u64 id = primary_event_id(event); | |
3975 | ||
3976 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3977 | return -EFAULT; | |
3978 | return 0; | |
3979 | } | |
3980 | ||
cdd6c482 | 3981 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3982 | { |
ac9721f3 | 3983 | int ret; |
ac9721f3 | 3984 | if (arg != -1) { |
2903ff01 AV |
3985 | struct perf_event *output_event; |
3986 | struct fd output; | |
3987 | ret = perf_fget_light(arg, &output); | |
3988 | if (ret) | |
3989 | return ret; | |
3990 | output_event = output.file->private_data; | |
3991 | ret = perf_event_set_output(event, output_event); | |
3992 | fdput(output); | |
3993 | } else { | |
3994 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3995 | } |
ac9721f3 PZ |
3996 | return ret; |
3997 | } | |
a4be7c27 | 3998 | |
6fb2915d LZ |
3999 | case PERF_EVENT_IOC_SET_FILTER: |
4000 | return perf_event_set_filter(event, (void __user *)arg); | |
4001 | ||
d859e29f | 4002 | default: |
3df5edad | 4003 | return -ENOTTY; |
d859e29f | 4004 | } |
3df5edad PZ |
4005 | |
4006 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 4007 | perf_event_for_each(event, func); |
3df5edad | 4008 | else |
cdd6c482 | 4009 | perf_event_for_each_child(event, func); |
3df5edad PZ |
4010 | |
4011 | return 0; | |
d859e29f PM |
4012 | } |
4013 | ||
f63a8daa PZ |
4014 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
4015 | { | |
4016 | struct perf_event *event = file->private_data; | |
4017 | struct perf_event_context *ctx; | |
4018 | long ret; | |
4019 | ||
4020 | ctx = perf_event_ctx_lock(event); | |
4021 | ret = _perf_ioctl(event, cmd, arg); | |
4022 | perf_event_ctx_unlock(event, ctx); | |
4023 | ||
4024 | return ret; | |
4025 | } | |
4026 | ||
b3f20785 PM |
4027 | #ifdef CONFIG_COMPAT |
4028 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, | |
4029 | unsigned long arg) | |
4030 | { | |
4031 | switch (_IOC_NR(cmd)) { | |
4032 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): | |
4033 | case _IOC_NR(PERF_EVENT_IOC_ID): | |
4034 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ | |
4035 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { | |
4036 | cmd &= ~IOCSIZE_MASK; | |
4037 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; | |
4038 | } | |
4039 | break; | |
4040 | } | |
4041 | return perf_ioctl(file, cmd, arg); | |
4042 | } | |
4043 | #else | |
4044 | # define perf_compat_ioctl NULL | |
4045 | #endif | |
4046 | ||
cdd6c482 | 4047 | int perf_event_task_enable(void) |
771d7cde | 4048 | { |
f63a8daa | 4049 | struct perf_event_context *ctx; |
cdd6c482 | 4050 | struct perf_event *event; |
771d7cde | 4051 | |
cdd6c482 | 4052 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4053 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4054 | ctx = perf_event_ctx_lock(event); | |
4055 | perf_event_for_each_child(event, _perf_event_enable); | |
4056 | perf_event_ctx_unlock(event, ctx); | |
4057 | } | |
cdd6c482 | 4058 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4059 | |
4060 | return 0; | |
4061 | } | |
4062 | ||
cdd6c482 | 4063 | int perf_event_task_disable(void) |
771d7cde | 4064 | { |
f63a8daa | 4065 | struct perf_event_context *ctx; |
cdd6c482 | 4066 | struct perf_event *event; |
771d7cde | 4067 | |
cdd6c482 | 4068 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4069 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4070 | ctx = perf_event_ctx_lock(event); | |
4071 | perf_event_for_each_child(event, _perf_event_disable); | |
4072 | perf_event_ctx_unlock(event, ctx); | |
4073 | } | |
cdd6c482 | 4074 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4075 | |
4076 | return 0; | |
4077 | } | |
4078 | ||
cdd6c482 | 4079 | static int perf_event_index(struct perf_event *event) |
194002b2 | 4080 | { |
a4eaf7f1 PZ |
4081 | if (event->hw.state & PERF_HES_STOPPED) |
4082 | return 0; | |
4083 | ||
cdd6c482 | 4084 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
4085 | return 0; |
4086 | ||
35edc2a5 | 4087 | return event->pmu->event_idx(event); |
194002b2 PZ |
4088 | } |
4089 | ||
c4794295 | 4090 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 4091 | u64 *now, |
7f310a5d EM |
4092 | u64 *enabled, |
4093 | u64 *running) | |
c4794295 | 4094 | { |
e3f3541c | 4095 | u64 ctx_time; |
c4794295 | 4096 | |
e3f3541c PZ |
4097 | *now = perf_clock(); |
4098 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
4099 | *enabled = ctx_time - event->tstamp_enabled; |
4100 | *running = ctx_time - event->tstamp_running; | |
4101 | } | |
4102 | ||
fa731587 PZ |
4103 | static void perf_event_init_userpage(struct perf_event *event) |
4104 | { | |
4105 | struct perf_event_mmap_page *userpg; | |
4106 | struct ring_buffer *rb; | |
4107 | ||
4108 | rcu_read_lock(); | |
4109 | rb = rcu_dereference(event->rb); | |
4110 | if (!rb) | |
4111 | goto unlock; | |
4112 | ||
4113 | userpg = rb->user_page; | |
4114 | ||
4115 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
4116 | userpg->cap_bit0_is_deprecated = 1; | |
4117 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
4118 | ||
4119 | unlock: | |
4120 | rcu_read_unlock(); | |
4121 | } | |
4122 | ||
c1317ec2 AL |
4123 | void __weak arch_perf_update_userpage( |
4124 | struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now) | |
e3f3541c PZ |
4125 | { |
4126 | } | |
4127 | ||
38ff667b PZ |
4128 | /* |
4129 | * Callers need to ensure there can be no nesting of this function, otherwise | |
4130 | * the seqlock logic goes bad. We can not serialize this because the arch | |
4131 | * code calls this from NMI context. | |
4132 | */ | |
cdd6c482 | 4133 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 4134 | { |
cdd6c482 | 4135 | struct perf_event_mmap_page *userpg; |
76369139 | 4136 | struct ring_buffer *rb; |
e3f3541c | 4137 | u64 enabled, running, now; |
38ff667b PZ |
4138 | |
4139 | rcu_read_lock(); | |
5ec4c599 PZ |
4140 | rb = rcu_dereference(event->rb); |
4141 | if (!rb) | |
4142 | goto unlock; | |
4143 | ||
0d641208 EM |
4144 | /* |
4145 | * compute total_time_enabled, total_time_running | |
4146 | * based on snapshot values taken when the event | |
4147 | * was last scheduled in. | |
4148 | * | |
4149 | * we cannot simply called update_context_time() | |
4150 | * because of locking issue as we can be called in | |
4151 | * NMI context | |
4152 | */ | |
e3f3541c | 4153 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 4154 | |
76369139 | 4155 | userpg = rb->user_page; |
7b732a75 PZ |
4156 | /* |
4157 | * Disable preemption so as to not let the corresponding user-space | |
4158 | * spin too long if we get preempted. | |
4159 | */ | |
4160 | preempt_disable(); | |
37d81828 | 4161 | ++userpg->lock; |
92f22a38 | 4162 | barrier(); |
cdd6c482 | 4163 | userpg->index = perf_event_index(event); |
b5e58793 | 4164 | userpg->offset = perf_event_count(event); |
365a4038 | 4165 | if (userpg->index) |
e7850595 | 4166 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 4167 | |
0d641208 | 4168 | userpg->time_enabled = enabled + |
cdd6c482 | 4169 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 4170 | |
0d641208 | 4171 | userpg->time_running = running + |
cdd6c482 | 4172 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 4173 | |
c1317ec2 | 4174 | arch_perf_update_userpage(event, userpg, now); |
e3f3541c | 4175 | |
92f22a38 | 4176 | barrier(); |
37d81828 | 4177 | ++userpg->lock; |
7b732a75 | 4178 | preempt_enable(); |
38ff667b | 4179 | unlock: |
7b732a75 | 4180 | rcu_read_unlock(); |
37d81828 PM |
4181 | } |
4182 | ||
906010b2 PZ |
4183 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
4184 | { | |
4185 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 4186 | struct ring_buffer *rb; |
906010b2 PZ |
4187 | int ret = VM_FAULT_SIGBUS; |
4188 | ||
4189 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
4190 | if (vmf->pgoff == 0) | |
4191 | ret = 0; | |
4192 | return ret; | |
4193 | } | |
4194 | ||
4195 | rcu_read_lock(); | |
76369139 FW |
4196 | rb = rcu_dereference(event->rb); |
4197 | if (!rb) | |
906010b2 PZ |
4198 | goto unlock; |
4199 | ||
4200 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
4201 | goto unlock; | |
4202 | ||
76369139 | 4203 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
4204 | if (!vmf->page) |
4205 | goto unlock; | |
4206 | ||
4207 | get_page(vmf->page); | |
4208 | vmf->page->mapping = vma->vm_file->f_mapping; | |
4209 | vmf->page->index = vmf->pgoff; | |
4210 | ||
4211 | ret = 0; | |
4212 | unlock: | |
4213 | rcu_read_unlock(); | |
4214 | ||
4215 | return ret; | |
4216 | } | |
4217 | ||
10c6db11 PZ |
4218 | static void ring_buffer_attach(struct perf_event *event, |
4219 | struct ring_buffer *rb) | |
4220 | { | |
b69cf536 | 4221 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
4222 | unsigned long flags; |
4223 | ||
b69cf536 PZ |
4224 | if (event->rb) { |
4225 | /* | |
4226 | * Should be impossible, we set this when removing | |
4227 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
4228 | */ | |
4229 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 4230 | |
b69cf536 PZ |
4231 | old_rb = event->rb; |
4232 | event->rcu_batches = get_state_synchronize_rcu(); | |
4233 | event->rcu_pending = 1; | |
10c6db11 | 4234 | |
b69cf536 PZ |
4235 | spin_lock_irqsave(&old_rb->event_lock, flags); |
4236 | list_del_rcu(&event->rb_entry); | |
4237 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
4238 | } | |
10c6db11 | 4239 | |
b69cf536 PZ |
4240 | if (event->rcu_pending && rb) { |
4241 | cond_synchronize_rcu(event->rcu_batches); | |
4242 | event->rcu_pending = 0; | |
4243 | } | |
10c6db11 | 4244 | |
b69cf536 PZ |
4245 | if (rb) { |
4246 | spin_lock_irqsave(&rb->event_lock, flags); | |
4247 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
4248 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
4249 | } | |
4250 | ||
4251 | rcu_assign_pointer(event->rb, rb); | |
4252 | ||
4253 | if (old_rb) { | |
4254 | ring_buffer_put(old_rb); | |
4255 | /* | |
4256 | * Since we detached before setting the new rb, so that we | |
4257 | * could attach the new rb, we could have missed a wakeup. | |
4258 | * Provide it now. | |
4259 | */ | |
4260 | wake_up_all(&event->waitq); | |
4261 | } | |
10c6db11 PZ |
4262 | } |
4263 | ||
4264 | static void ring_buffer_wakeup(struct perf_event *event) | |
4265 | { | |
4266 | struct ring_buffer *rb; | |
4267 | ||
4268 | rcu_read_lock(); | |
4269 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
4270 | if (rb) { |
4271 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
4272 | wake_up_all(&event->waitq); | |
4273 | } | |
10c6db11 PZ |
4274 | rcu_read_unlock(); |
4275 | } | |
4276 | ||
76369139 | 4277 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 4278 | { |
76369139 | 4279 | struct ring_buffer *rb; |
906010b2 | 4280 | |
76369139 FW |
4281 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
4282 | rb_free(rb); | |
7b732a75 PZ |
4283 | } |
4284 | ||
76369139 | 4285 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 4286 | { |
76369139 | 4287 | struct ring_buffer *rb; |
7b732a75 | 4288 | |
ac9721f3 | 4289 | rcu_read_lock(); |
76369139 FW |
4290 | rb = rcu_dereference(event->rb); |
4291 | if (rb) { | |
4292 | if (!atomic_inc_not_zero(&rb->refcount)) | |
4293 | rb = NULL; | |
ac9721f3 PZ |
4294 | } |
4295 | rcu_read_unlock(); | |
4296 | ||
76369139 | 4297 | return rb; |
ac9721f3 PZ |
4298 | } |
4299 | ||
76369139 | 4300 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 4301 | { |
76369139 | 4302 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 4303 | return; |
7b732a75 | 4304 | |
9bb5d40c | 4305 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 4306 | |
76369139 | 4307 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
4308 | } |
4309 | ||
4310 | static void perf_mmap_open(struct vm_area_struct *vma) | |
4311 | { | |
cdd6c482 | 4312 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4313 | |
cdd6c482 | 4314 | atomic_inc(&event->mmap_count); |
9bb5d40c | 4315 | atomic_inc(&event->rb->mmap_count); |
1e0fb9ec AL |
4316 | |
4317 | if (event->pmu->event_mapped) | |
4318 | event->pmu->event_mapped(event); | |
7b732a75 PZ |
4319 | } |
4320 | ||
9bb5d40c PZ |
4321 | /* |
4322 | * A buffer can be mmap()ed multiple times; either directly through the same | |
4323 | * event, or through other events by use of perf_event_set_output(). | |
4324 | * | |
4325 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
4326 | * the buffer here, where we still have a VM context. This means we need | |
4327 | * to detach all events redirecting to us. | |
4328 | */ | |
7b732a75 PZ |
4329 | static void perf_mmap_close(struct vm_area_struct *vma) |
4330 | { | |
cdd6c482 | 4331 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4332 | |
b69cf536 | 4333 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
4334 | struct user_struct *mmap_user = rb->mmap_user; |
4335 | int mmap_locked = rb->mmap_locked; | |
4336 | unsigned long size = perf_data_size(rb); | |
789f90fc | 4337 | |
1e0fb9ec AL |
4338 | if (event->pmu->event_unmapped) |
4339 | event->pmu->event_unmapped(event); | |
4340 | ||
9bb5d40c PZ |
4341 | atomic_dec(&rb->mmap_count); |
4342 | ||
4343 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4344 | goto out_put; |
9bb5d40c | 4345 | |
b69cf536 | 4346 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4347 | mutex_unlock(&event->mmap_mutex); |
4348 | ||
4349 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4350 | if (atomic_read(&rb->mmap_count)) |
4351 | goto out_put; | |
ac9721f3 | 4352 | |
9bb5d40c PZ |
4353 | /* |
4354 | * No other mmap()s, detach from all other events that might redirect | |
4355 | * into the now unreachable buffer. Somewhat complicated by the | |
4356 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4357 | */ | |
4358 | again: | |
4359 | rcu_read_lock(); | |
4360 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4361 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4362 | /* | |
4363 | * This event is en-route to free_event() which will | |
4364 | * detach it and remove it from the list. | |
4365 | */ | |
4366 | continue; | |
4367 | } | |
4368 | rcu_read_unlock(); | |
789f90fc | 4369 | |
9bb5d40c PZ |
4370 | mutex_lock(&event->mmap_mutex); |
4371 | /* | |
4372 | * Check we didn't race with perf_event_set_output() which can | |
4373 | * swizzle the rb from under us while we were waiting to | |
4374 | * acquire mmap_mutex. | |
4375 | * | |
4376 | * If we find a different rb; ignore this event, a next | |
4377 | * iteration will no longer find it on the list. We have to | |
4378 | * still restart the iteration to make sure we're not now | |
4379 | * iterating the wrong list. | |
4380 | */ | |
b69cf536 PZ |
4381 | if (event->rb == rb) |
4382 | ring_buffer_attach(event, NULL); | |
4383 | ||
cdd6c482 | 4384 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4385 | put_event(event); |
ac9721f3 | 4386 | |
9bb5d40c PZ |
4387 | /* |
4388 | * Restart the iteration; either we're on the wrong list or | |
4389 | * destroyed its integrity by doing a deletion. | |
4390 | */ | |
4391 | goto again; | |
7b732a75 | 4392 | } |
9bb5d40c PZ |
4393 | rcu_read_unlock(); |
4394 | ||
4395 | /* | |
4396 | * It could be there's still a few 0-ref events on the list; they'll | |
4397 | * get cleaned up by free_event() -- they'll also still have their | |
4398 | * ref on the rb and will free it whenever they are done with it. | |
4399 | * | |
4400 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4401 | * undo the VM accounting. | |
4402 | */ | |
4403 | ||
4404 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4405 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4406 | free_uid(mmap_user); | |
4407 | ||
b69cf536 | 4408 | out_put: |
9bb5d40c | 4409 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4410 | } |
4411 | ||
f0f37e2f | 4412 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4413 | .open = perf_mmap_open, |
4414 | .close = perf_mmap_close, | |
4415 | .fault = perf_mmap_fault, | |
4416 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4417 | }; |
4418 | ||
4419 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4420 | { | |
cdd6c482 | 4421 | struct perf_event *event = file->private_data; |
22a4f650 | 4422 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4423 | struct user_struct *user = current_user(); |
22a4f650 | 4424 | unsigned long locked, lock_limit; |
76369139 | 4425 | struct ring_buffer *rb; |
7b732a75 PZ |
4426 | unsigned long vma_size; |
4427 | unsigned long nr_pages; | |
789f90fc | 4428 | long user_extra, extra; |
d57e34fd | 4429 | int ret = 0, flags = 0; |
37d81828 | 4430 | |
c7920614 PZ |
4431 | /* |
4432 | * Don't allow mmap() of inherited per-task counters. This would | |
4433 | * create a performance issue due to all children writing to the | |
76369139 | 4434 | * same rb. |
c7920614 PZ |
4435 | */ |
4436 | if (event->cpu == -1 && event->attr.inherit) | |
4437 | return -EINVAL; | |
4438 | ||
43a21ea8 | 4439 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4440 | return -EINVAL; |
7b732a75 PZ |
4441 | |
4442 | vma_size = vma->vm_end - vma->vm_start; | |
4443 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4444 | ||
7730d865 | 4445 | /* |
76369139 | 4446 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4447 | * can do bitmasks instead of modulo. |
4448 | */ | |
4449 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
4450 | return -EINVAL; |
4451 | ||
7b732a75 | 4452 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4453 | return -EINVAL; |
4454 | ||
7b732a75 PZ |
4455 | if (vma->vm_pgoff != 0) |
4456 | return -EINVAL; | |
37d81828 | 4457 | |
cdd6c482 | 4458 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4459 | again: |
cdd6c482 | 4460 | mutex_lock(&event->mmap_mutex); |
76369139 | 4461 | if (event->rb) { |
9bb5d40c | 4462 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4463 | ret = -EINVAL; |
9bb5d40c PZ |
4464 | goto unlock; |
4465 | } | |
4466 | ||
4467 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4468 | /* | |
4469 | * Raced against perf_mmap_close() through | |
4470 | * perf_event_set_output(). Try again, hope for better | |
4471 | * luck. | |
4472 | */ | |
4473 | mutex_unlock(&event->mmap_mutex); | |
4474 | goto again; | |
4475 | } | |
4476 | ||
ebb3c4c4 PZ |
4477 | goto unlock; |
4478 | } | |
4479 | ||
789f90fc | 4480 | user_extra = nr_pages + 1; |
cdd6c482 | 4481 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4482 | |
4483 | /* | |
4484 | * Increase the limit linearly with more CPUs: | |
4485 | */ | |
4486 | user_lock_limit *= num_online_cpus(); | |
4487 | ||
789f90fc | 4488 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4489 | |
789f90fc PZ |
4490 | extra = 0; |
4491 | if (user_locked > user_lock_limit) | |
4492 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4493 | |
78d7d407 | 4494 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4495 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4496 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4497 | |
459ec28a IM |
4498 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4499 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4500 | ret = -EPERM; |
4501 | goto unlock; | |
4502 | } | |
7b732a75 | 4503 | |
76369139 | 4504 | WARN_ON(event->rb); |
906010b2 | 4505 | |
d57e34fd | 4506 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4507 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4508 | |
4ec8363d VW |
4509 | rb = rb_alloc(nr_pages, |
4510 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4511 | event->cpu, flags); | |
4512 | ||
76369139 | 4513 | if (!rb) { |
ac9721f3 | 4514 | ret = -ENOMEM; |
ebb3c4c4 | 4515 | goto unlock; |
ac9721f3 | 4516 | } |
26cb63ad | 4517 | |
9bb5d40c | 4518 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4519 | rb->mmap_locked = extra; |
4520 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4521 | |
ac9721f3 | 4522 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4523 | vma->vm_mm->pinned_vm += extra; |
4524 | ||
9bb5d40c | 4525 | ring_buffer_attach(event, rb); |
ac9721f3 | 4526 | |
fa731587 | 4527 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4528 | perf_event_update_userpage(event); |
4529 | ||
ebb3c4c4 | 4530 | unlock: |
ac9721f3 PZ |
4531 | if (!ret) |
4532 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4533 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4534 | |
9bb5d40c PZ |
4535 | /* |
4536 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4537 | * vma. | |
4538 | */ | |
26cb63ad | 4539 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4540 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 | 4541 | |
1e0fb9ec AL |
4542 | if (event->pmu->event_mapped) |
4543 | event->pmu->event_mapped(event); | |
4544 | ||
7b732a75 | 4545 | return ret; |
37d81828 PM |
4546 | } |
4547 | ||
3c446b3d PZ |
4548 | static int perf_fasync(int fd, struct file *filp, int on) |
4549 | { | |
496ad9aa | 4550 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4551 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4552 | int retval; |
4553 | ||
4554 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4555 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4556 | mutex_unlock(&inode->i_mutex); |
4557 | ||
4558 | if (retval < 0) | |
4559 | return retval; | |
4560 | ||
4561 | return 0; | |
4562 | } | |
4563 | ||
0793a61d | 4564 | static const struct file_operations perf_fops = { |
3326c1ce | 4565 | .llseek = no_llseek, |
0793a61d TG |
4566 | .release = perf_release, |
4567 | .read = perf_read, | |
4568 | .poll = perf_poll, | |
d859e29f | 4569 | .unlocked_ioctl = perf_ioctl, |
b3f20785 | 4570 | .compat_ioctl = perf_compat_ioctl, |
37d81828 | 4571 | .mmap = perf_mmap, |
3c446b3d | 4572 | .fasync = perf_fasync, |
0793a61d TG |
4573 | }; |
4574 | ||
925d519a | 4575 | /* |
cdd6c482 | 4576 | * Perf event wakeup |
925d519a PZ |
4577 | * |
4578 | * If there's data, ensure we set the poll() state and publish everything | |
4579 | * to user-space before waking everybody up. | |
4580 | */ | |
4581 | ||
cdd6c482 | 4582 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4583 | { |
10c6db11 | 4584 | ring_buffer_wakeup(event); |
4c9e2542 | 4585 | |
cdd6c482 IM |
4586 | if (event->pending_kill) { |
4587 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4588 | event->pending_kill = 0; | |
4c9e2542 | 4589 | } |
925d519a PZ |
4590 | } |
4591 | ||
e360adbe | 4592 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4593 | { |
cdd6c482 IM |
4594 | struct perf_event *event = container_of(entry, |
4595 | struct perf_event, pending); | |
79f14641 | 4596 | |
cdd6c482 IM |
4597 | if (event->pending_disable) { |
4598 | event->pending_disable = 0; | |
4599 | __perf_event_disable(event); | |
79f14641 PZ |
4600 | } |
4601 | ||
cdd6c482 IM |
4602 | if (event->pending_wakeup) { |
4603 | event->pending_wakeup = 0; | |
4604 | perf_event_wakeup(event); | |
79f14641 PZ |
4605 | } |
4606 | } | |
4607 | ||
39447b38 ZY |
4608 | /* |
4609 | * We assume there is only KVM supporting the callbacks. | |
4610 | * Later on, we might change it to a list if there is | |
4611 | * another virtualization implementation supporting the callbacks. | |
4612 | */ | |
4613 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4614 | ||
4615 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4616 | { | |
4617 | perf_guest_cbs = cbs; | |
4618 | return 0; | |
4619 | } | |
4620 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4621 | ||
4622 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4623 | { | |
4624 | perf_guest_cbs = NULL; | |
4625 | return 0; | |
4626 | } | |
4627 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4628 | ||
4018994f JO |
4629 | static void |
4630 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4631 | struct pt_regs *regs, u64 mask) | |
4632 | { | |
4633 | int bit; | |
4634 | ||
4635 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4636 | sizeof(mask) * BITS_PER_BYTE) { | |
4637 | u64 val; | |
4638 | ||
4639 | val = perf_reg_value(regs, bit); | |
4640 | perf_output_put(handle, val); | |
4641 | } | |
4642 | } | |
4643 | ||
60e2364e | 4644 | static void perf_sample_regs_user(struct perf_regs *regs_user, |
88a7c26a AL |
4645 | struct pt_regs *regs, |
4646 | struct pt_regs *regs_user_copy) | |
4018994f | 4647 | { |
88a7c26a AL |
4648 | if (user_mode(regs)) { |
4649 | regs_user->abi = perf_reg_abi(current); | |
2565711f | 4650 | regs_user->regs = regs; |
88a7c26a AL |
4651 | } else if (current->mm) { |
4652 | perf_get_regs_user(regs_user, regs, regs_user_copy); | |
2565711f PZ |
4653 | } else { |
4654 | regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; | |
4655 | regs_user->regs = NULL; | |
4018994f JO |
4656 | } |
4657 | } | |
4658 | ||
60e2364e SE |
4659 | static void perf_sample_regs_intr(struct perf_regs *regs_intr, |
4660 | struct pt_regs *regs) | |
4661 | { | |
4662 | regs_intr->regs = regs; | |
4663 | regs_intr->abi = perf_reg_abi(current); | |
4664 | } | |
4665 | ||
4666 | ||
c5ebcedb JO |
4667 | /* |
4668 | * Get remaining task size from user stack pointer. | |
4669 | * | |
4670 | * It'd be better to take stack vma map and limit this more | |
4671 | * precisly, but there's no way to get it safely under interrupt, | |
4672 | * so using TASK_SIZE as limit. | |
4673 | */ | |
4674 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4675 | { | |
4676 | unsigned long addr = perf_user_stack_pointer(regs); | |
4677 | ||
4678 | if (!addr || addr >= TASK_SIZE) | |
4679 | return 0; | |
4680 | ||
4681 | return TASK_SIZE - addr; | |
4682 | } | |
4683 | ||
4684 | static u16 | |
4685 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4686 | struct pt_regs *regs) | |
4687 | { | |
4688 | u64 task_size; | |
4689 | ||
4690 | /* No regs, no stack pointer, no dump. */ | |
4691 | if (!regs) | |
4692 | return 0; | |
4693 | ||
4694 | /* | |
4695 | * Check if we fit in with the requested stack size into the: | |
4696 | * - TASK_SIZE | |
4697 | * If we don't, we limit the size to the TASK_SIZE. | |
4698 | * | |
4699 | * - remaining sample size | |
4700 | * If we don't, we customize the stack size to | |
4701 | * fit in to the remaining sample size. | |
4702 | */ | |
4703 | ||
4704 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4705 | stack_size = min(stack_size, (u16) task_size); | |
4706 | ||
4707 | /* Current header size plus static size and dynamic size. */ | |
4708 | header_size += 2 * sizeof(u64); | |
4709 | ||
4710 | /* Do we fit in with the current stack dump size? */ | |
4711 | if ((u16) (header_size + stack_size) < header_size) { | |
4712 | /* | |
4713 | * If we overflow the maximum size for the sample, | |
4714 | * we customize the stack dump size to fit in. | |
4715 | */ | |
4716 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4717 | stack_size = round_up(stack_size, sizeof(u64)); | |
4718 | } | |
4719 | ||
4720 | return stack_size; | |
4721 | } | |
4722 | ||
4723 | static void | |
4724 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4725 | struct pt_regs *regs) | |
4726 | { | |
4727 | /* Case of a kernel thread, nothing to dump */ | |
4728 | if (!regs) { | |
4729 | u64 size = 0; | |
4730 | perf_output_put(handle, size); | |
4731 | } else { | |
4732 | unsigned long sp; | |
4733 | unsigned int rem; | |
4734 | u64 dyn_size; | |
4735 | ||
4736 | /* | |
4737 | * We dump: | |
4738 | * static size | |
4739 | * - the size requested by user or the best one we can fit | |
4740 | * in to the sample max size | |
4741 | * data | |
4742 | * - user stack dump data | |
4743 | * dynamic size | |
4744 | * - the actual dumped size | |
4745 | */ | |
4746 | ||
4747 | /* Static size. */ | |
4748 | perf_output_put(handle, dump_size); | |
4749 | ||
4750 | /* Data. */ | |
4751 | sp = perf_user_stack_pointer(regs); | |
4752 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4753 | dyn_size = dump_size - rem; | |
4754 | ||
4755 | perf_output_skip(handle, rem); | |
4756 | ||
4757 | /* Dynamic size. */ | |
4758 | perf_output_put(handle, dyn_size); | |
4759 | } | |
4760 | } | |
4761 | ||
c980d109 ACM |
4762 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4763 | struct perf_sample_data *data, | |
4764 | struct perf_event *event) | |
6844c09d ACM |
4765 | { |
4766 | u64 sample_type = event->attr.sample_type; | |
4767 | ||
4768 | data->type = sample_type; | |
4769 | header->size += event->id_header_size; | |
4770 | ||
4771 | if (sample_type & PERF_SAMPLE_TID) { | |
4772 | /* namespace issues */ | |
4773 | data->tid_entry.pid = perf_event_pid(event, current); | |
4774 | data->tid_entry.tid = perf_event_tid(event, current); | |
4775 | } | |
4776 | ||
4777 | if (sample_type & PERF_SAMPLE_TIME) | |
4778 | data->time = perf_clock(); | |
4779 | ||
ff3d527c | 4780 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4781 | data->id = primary_event_id(event); |
4782 | ||
4783 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4784 | data->stream_id = event->id; | |
4785 | ||
4786 | if (sample_type & PERF_SAMPLE_CPU) { | |
4787 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4788 | data->cpu_entry.reserved = 0; | |
4789 | } | |
4790 | } | |
4791 | ||
76369139 FW |
4792 | void perf_event_header__init_id(struct perf_event_header *header, |
4793 | struct perf_sample_data *data, | |
4794 | struct perf_event *event) | |
c980d109 ACM |
4795 | { |
4796 | if (event->attr.sample_id_all) | |
4797 | __perf_event_header__init_id(header, data, event); | |
4798 | } | |
4799 | ||
4800 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4801 | struct perf_sample_data *data) | |
4802 | { | |
4803 | u64 sample_type = data->type; | |
4804 | ||
4805 | if (sample_type & PERF_SAMPLE_TID) | |
4806 | perf_output_put(handle, data->tid_entry); | |
4807 | ||
4808 | if (sample_type & PERF_SAMPLE_TIME) | |
4809 | perf_output_put(handle, data->time); | |
4810 | ||
4811 | if (sample_type & PERF_SAMPLE_ID) | |
4812 | perf_output_put(handle, data->id); | |
4813 | ||
4814 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4815 | perf_output_put(handle, data->stream_id); | |
4816 | ||
4817 | if (sample_type & PERF_SAMPLE_CPU) | |
4818 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4819 | |
4820 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4821 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4822 | } |
4823 | ||
76369139 FW |
4824 | void perf_event__output_id_sample(struct perf_event *event, |
4825 | struct perf_output_handle *handle, | |
4826 | struct perf_sample_data *sample) | |
c980d109 ACM |
4827 | { |
4828 | if (event->attr.sample_id_all) | |
4829 | __perf_event__output_id_sample(handle, sample); | |
4830 | } | |
4831 | ||
3dab77fb | 4832 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4833 | struct perf_event *event, |
4834 | u64 enabled, u64 running) | |
3dab77fb | 4835 | { |
cdd6c482 | 4836 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4837 | u64 values[4]; |
4838 | int n = 0; | |
4839 | ||
b5e58793 | 4840 | values[n++] = perf_event_count(event); |
3dab77fb | 4841 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4842 | values[n++] = enabled + |
cdd6c482 | 4843 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4844 | } |
4845 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4846 | values[n++] = running + |
cdd6c482 | 4847 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4848 | } |
4849 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4850 | values[n++] = primary_event_id(event); |
3dab77fb | 4851 | |
76369139 | 4852 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4853 | } |
4854 | ||
4855 | /* | |
cdd6c482 | 4856 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4857 | */ |
4858 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4859 | struct perf_event *event, |
4860 | u64 enabled, u64 running) | |
3dab77fb | 4861 | { |
cdd6c482 IM |
4862 | struct perf_event *leader = event->group_leader, *sub; |
4863 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4864 | u64 values[5]; |
4865 | int n = 0; | |
4866 | ||
4867 | values[n++] = 1 + leader->nr_siblings; | |
4868 | ||
4869 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4870 | values[n++] = enabled; |
3dab77fb PZ |
4871 | |
4872 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4873 | values[n++] = running; |
3dab77fb | 4874 | |
cdd6c482 | 4875 | if (leader != event) |
3dab77fb PZ |
4876 | leader->pmu->read(leader); |
4877 | ||
b5e58793 | 4878 | values[n++] = perf_event_count(leader); |
3dab77fb | 4879 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4880 | values[n++] = primary_event_id(leader); |
3dab77fb | 4881 | |
76369139 | 4882 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4883 | |
65abc865 | 4884 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4885 | n = 0; |
4886 | ||
6f5ab001 JO |
4887 | if ((sub != event) && |
4888 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4889 | sub->pmu->read(sub); |
4890 | ||
b5e58793 | 4891 | values[n++] = perf_event_count(sub); |
3dab77fb | 4892 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4893 | values[n++] = primary_event_id(sub); |
3dab77fb | 4894 | |
76369139 | 4895 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4896 | } |
4897 | } | |
4898 | ||
eed01528 SE |
4899 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4900 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4901 | ||
3dab77fb | 4902 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4903 | struct perf_event *event) |
3dab77fb | 4904 | { |
e3f3541c | 4905 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4906 | u64 read_format = event->attr.read_format; |
4907 | ||
4908 | /* | |
4909 | * compute total_time_enabled, total_time_running | |
4910 | * based on snapshot values taken when the event | |
4911 | * was last scheduled in. | |
4912 | * | |
4913 | * we cannot simply called update_context_time() | |
4914 | * because of locking issue as we are called in | |
4915 | * NMI context | |
4916 | */ | |
c4794295 | 4917 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4918 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4919 | |
cdd6c482 | 4920 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4921 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4922 | else |
eed01528 | 4923 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4924 | } |
4925 | ||
5622f295 MM |
4926 | void perf_output_sample(struct perf_output_handle *handle, |
4927 | struct perf_event_header *header, | |
4928 | struct perf_sample_data *data, | |
cdd6c482 | 4929 | struct perf_event *event) |
5622f295 MM |
4930 | { |
4931 | u64 sample_type = data->type; | |
4932 | ||
4933 | perf_output_put(handle, *header); | |
4934 | ||
ff3d527c AH |
4935 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4936 | perf_output_put(handle, data->id); | |
4937 | ||
5622f295 MM |
4938 | if (sample_type & PERF_SAMPLE_IP) |
4939 | perf_output_put(handle, data->ip); | |
4940 | ||
4941 | if (sample_type & PERF_SAMPLE_TID) | |
4942 | perf_output_put(handle, data->tid_entry); | |
4943 | ||
4944 | if (sample_type & PERF_SAMPLE_TIME) | |
4945 | perf_output_put(handle, data->time); | |
4946 | ||
4947 | if (sample_type & PERF_SAMPLE_ADDR) | |
4948 | perf_output_put(handle, data->addr); | |
4949 | ||
4950 | if (sample_type & PERF_SAMPLE_ID) | |
4951 | perf_output_put(handle, data->id); | |
4952 | ||
4953 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4954 | perf_output_put(handle, data->stream_id); | |
4955 | ||
4956 | if (sample_type & PERF_SAMPLE_CPU) | |
4957 | perf_output_put(handle, data->cpu_entry); | |
4958 | ||
4959 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4960 | perf_output_put(handle, data->period); | |
4961 | ||
4962 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4963 | perf_output_read(handle, event); |
5622f295 MM |
4964 | |
4965 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4966 | if (data->callchain) { | |
4967 | int size = 1; | |
4968 | ||
4969 | if (data->callchain) | |
4970 | size += data->callchain->nr; | |
4971 | ||
4972 | size *= sizeof(u64); | |
4973 | ||
76369139 | 4974 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4975 | } else { |
4976 | u64 nr = 0; | |
4977 | perf_output_put(handle, nr); | |
4978 | } | |
4979 | } | |
4980 | ||
4981 | if (sample_type & PERF_SAMPLE_RAW) { | |
4982 | if (data->raw) { | |
4983 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4984 | __output_copy(handle, data->raw->data, |
4985 | data->raw->size); | |
5622f295 MM |
4986 | } else { |
4987 | struct { | |
4988 | u32 size; | |
4989 | u32 data; | |
4990 | } raw = { | |
4991 | .size = sizeof(u32), | |
4992 | .data = 0, | |
4993 | }; | |
4994 | perf_output_put(handle, raw); | |
4995 | } | |
4996 | } | |
a7ac67ea | 4997 | |
bce38cd5 SE |
4998 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4999 | if (data->br_stack) { | |
5000 | size_t size; | |
5001 | ||
5002 | size = data->br_stack->nr | |
5003 | * sizeof(struct perf_branch_entry); | |
5004 | ||
5005 | perf_output_put(handle, data->br_stack->nr); | |
5006 | perf_output_copy(handle, data->br_stack->entries, size); | |
5007 | } else { | |
5008 | /* | |
5009 | * we always store at least the value of nr | |
5010 | */ | |
5011 | u64 nr = 0; | |
5012 | perf_output_put(handle, nr); | |
5013 | } | |
5014 | } | |
4018994f JO |
5015 | |
5016 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
5017 | u64 abi = data->regs_user.abi; | |
5018 | ||
5019 | /* | |
5020 | * If there are no regs to dump, notice it through | |
5021 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5022 | */ | |
5023 | perf_output_put(handle, abi); | |
5024 | ||
5025 | if (abi) { | |
5026 | u64 mask = event->attr.sample_regs_user; | |
5027 | perf_output_sample_regs(handle, | |
5028 | data->regs_user.regs, | |
5029 | mask); | |
5030 | } | |
5031 | } | |
c5ebcedb | 5032 | |
a5cdd40c | 5033 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
5034 | perf_output_sample_ustack(handle, |
5035 | data->stack_user_size, | |
5036 | data->regs_user.regs); | |
a5cdd40c | 5037 | } |
c3feedf2 AK |
5038 | |
5039 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
5040 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
5041 | |
5042 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
5043 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 5044 | |
fdfbbd07 AK |
5045 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
5046 | perf_output_put(handle, data->txn); | |
5047 | ||
60e2364e SE |
5048 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
5049 | u64 abi = data->regs_intr.abi; | |
5050 | /* | |
5051 | * If there are no regs to dump, notice it through | |
5052 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5053 | */ | |
5054 | perf_output_put(handle, abi); | |
5055 | ||
5056 | if (abi) { | |
5057 | u64 mask = event->attr.sample_regs_intr; | |
5058 | ||
5059 | perf_output_sample_regs(handle, | |
5060 | data->regs_intr.regs, | |
5061 | mask); | |
5062 | } | |
5063 | } | |
5064 | ||
a5cdd40c PZ |
5065 | if (!event->attr.watermark) { |
5066 | int wakeup_events = event->attr.wakeup_events; | |
5067 | ||
5068 | if (wakeup_events) { | |
5069 | struct ring_buffer *rb = handle->rb; | |
5070 | int events = local_inc_return(&rb->events); | |
5071 | ||
5072 | if (events >= wakeup_events) { | |
5073 | local_sub(wakeup_events, &rb->events); | |
5074 | local_inc(&rb->wakeup); | |
5075 | } | |
5076 | } | |
5077 | } | |
5622f295 MM |
5078 | } |
5079 | ||
5080 | void perf_prepare_sample(struct perf_event_header *header, | |
5081 | struct perf_sample_data *data, | |
cdd6c482 | 5082 | struct perf_event *event, |
5622f295 | 5083 | struct pt_regs *regs) |
7b732a75 | 5084 | { |
cdd6c482 | 5085 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 5086 | |
cdd6c482 | 5087 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 5088 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
5089 | |
5090 | header->misc = 0; | |
5091 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 5092 | |
c980d109 | 5093 | __perf_event_header__init_id(header, data, event); |
6844c09d | 5094 | |
c320c7b7 | 5095 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
5096 | data->ip = perf_instruction_pointer(regs); |
5097 | ||
b23f3325 | 5098 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 5099 | int size = 1; |
394ee076 | 5100 | |
e6dab5ff | 5101 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
5102 | |
5103 | if (data->callchain) | |
5104 | size += data->callchain->nr; | |
5105 | ||
5106 | header->size += size * sizeof(u64); | |
394ee076 PZ |
5107 | } |
5108 | ||
3a43ce68 | 5109 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
5110 | int size = sizeof(u32); |
5111 | ||
5112 | if (data->raw) | |
5113 | size += data->raw->size; | |
5114 | else | |
5115 | size += sizeof(u32); | |
5116 | ||
5117 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 5118 | header->size += size; |
7f453c24 | 5119 | } |
bce38cd5 SE |
5120 | |
5121 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
5122 | int size = sizeof(u64); /* nr */ | |
5123 | if (data->br_stack) { | |
5124 | size += data->br_stack->nr | |
5125 | * sizeof(struct perf_branch_entry); | |
5126 | } | |
5127 | header->size += size; | |
5128 | } | |
4018994f | 5129 | |
2565711f | 5130 | if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER)) |
88a7c26a AL |
5131 | perf_sample_regs_user(&data->regs_user, regs, |
5132 | &data->regs_user_copy); | |
2565711f | 5133 | |
4018994f JO |
5134 | if (sample_type & PERF_SAMPLE_REGS_USER) { |
5135 | /* regs dump ABI info */ | |
5136 | int size = sizeof(u64); | |
5137 | ||
4018994f JO |
5138 | if (data->regs_user.regs) { |
5139 | u64 mask = event->attr.sample_regs_user; | |
5140 | size += hweight64(mask) * sizeof(u64); | |
5141 | } | |
5142 | ||
5143 | header->size += size; | |
5144 | } | |
c5ebcedb JO |
5145 | |
5146 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
5147 | /* | |
5148 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
5149 | * processed as the last one or have additional check added | |
5150 | * in case new sample type is added, because we could eat | |
5151 | * up the rest of the sample size. | |
5152 | */ | |
c5ebcedb JO |
5153 | u16 stack_size = event->attr.sample_stack_user; |
5154 | u16 size = sizeof(u64); | |
5155 | ||
c5ebcedb | 5156 | stack_size = perf_sample_ustack_size(stack_size, header->size, |
2565711f | 5157 | data->regs_user.regs); |
c5ebcedb JO |
5158 | |
5159 | /* | |
5160 | * If there is something to dump, add space for the dump | |
5161 | * itself and for the field that tells the dynamic size, | |
5162 | * which is how many have been actually dumped. | |
5163 | */ | |
5164 | if (stack_size) | |
5165 | size += sizeof(u64) + stack_size; | |
5166 | ||
5167 | data->stack_user_size = stack_size; | |
5168 | header->size += size; | |
5169 | } | |
60e2364e SE |
5170 | |
5171 | if (sample_type & PERF_SAMPLE_REGS_INTR) { | |
5172 | /* regs dump ABI info */ | |
5173 | int size = sizeof(u64); | |
5174 | ||
5175 | perf_sample_regs_intr(&data->regs_intr, regs); | |
5176 | ||
5177 | if (data->regs_intr.regs) { | |
5178 | u64 mask = event->attr.sample_regs_intr; | |
5179 | ||
5180 | size += hweight64(mask) * sizeof(u64); | |
5181 | } | |
5182 | ||
5183 | header->size += size; | |
5184 | } | |
5622f295 | 5185 | } |
7f453c24 | 5186 | |
a8b0ca17 | 5187 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
5188 | struct perf_sample_data *data, |
5189 | struct pt_regs *regs) | |
5190 | { | |
5191 | struct perf_output_handle handle; | |
5192 | struct perf_event_header header; | |
689802b2 | 5193 | |
927c7a9e FW |
5194 | /* protect the callchain buffers */ |
5195 | rcu_read_lock(); | |
5196 | ||
cdd6c482 | 5197 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 5198 | |
a7ac67ea | 5199 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 5200 | goto exit; |
0322cd6e | 5201 | |
cdd6c482 | 5202 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 5203 | |
8a057d84 | 5204 | perf_output_end(&handle); |
927c7a9e FW |
5205 | |
5206 | exit: | |
5207 | rcu_read_unlock(); | |
0322cd6e PZ |
5208 | } |
5209 | ||
38b200d6 | 5210 | /* |
cdd6c482 | 5211 | * read event_id |
38b200d6 PZ |
5212 | */ |
5213 | ||
5214 | struct perf_read_event { | |
5215 | struct perf_event_header header; | |
5216 | ||
5217 | u32 pid; | |
5218 | u32 tid; | |
38b200d6 PZ |
5219 | }; |
5220 | ||
5221 | static void | |
cdd6c482 | 5222 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
5223 | struct task_struct *task) |
5224 | { | |
5225 | struct perf_output_handle handle; | |
c980d109 | 5226 | struct perf_sample_data sample; |
dfc65094 | 5227 | struct perf_read_event read_event = { |
38b200d6 | 5228 | .header = { |
cdd6c482 | 5229 | .type = PERF_RECORD_READ, |
38b200d6 | 5230 | .misc = 0, |
c320c7b7 | 5231 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 5232 | }, |
cdd6c482 IM |
5233 | .pid = perf_event_pid(event, task), |
5234 | .tid = perf_event_tid(event, task), | |
38b200d6 | 5235 | }; |
3dab77fb | 5236 | int ret; |
38b200d6 | 5237 | |
c980d109 | 5238 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 5239 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
5240 | if (ret) |
5241 | return; | |
5242 | ||
dfc65094 | 5243 | perf_output_put(&handle, read_event); |
cdd6c482 | 5244 | perf_output_read(&handle, event); |
c980d109 | 5245 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 5246 | |
38b200d6 PZ |
5247 | perf_output_end(&handle); |
5248 | } | |
5249 | ||
52d857a8 JO |
5250 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
5251 | ||
5252 | static void | |
5253 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
5254 | perf_event_aux_output_cb output, |
5255 | void *data) | |
5256 | { | |
5257 | struct perf_event *event; | |
5258 | ||
5259 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5260 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
5261 | continue; | |
5262 | if (!event_filter_match(event)) | |
5263 | continue; | |
67516844 | 5264 | output(event, data); |
52d857a8 JO |
5265 | } |
5266 | } | |
5267 | ||
5268 | static void | |
67516844 | 5269 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
5270 | struct perf_event_context *task_ctx) |
5271 | { | |
5272 | struct perf_cpu_context *cpuctx; | |
5273 | struct perf_event_context *ctx; | |
5274 | struct pmu *pmu; | |
5275 | int ctxn; | |
5276 | ||
5277 | rcu_read_lock(); | |
5278 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
5279 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
5280 | if (cpuctx->unique_pmu != pmu) | |
5281 | goto next; | |
67516844 | 5282 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
5283 | if (task_ctx) |
5284 | goto next; | |
5285 | ctxn = pmu->task_ctx_nr; | |
5286 | if (ctxn < 0) | |
5287 | goto next; | |
5288 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
5289 | if (ctx) | |
67516844 | 5290 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
5291 | next: |
5292 | put_cpu_ptr(pmu->pmu_cpu_context); | |
5293 | } | |
5294 | ||
5295 | if (task_ctx) { | |
5296 | preempt_disable(); | |
67516844 | 5297 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
5298 | preempt_enable(); |
5299 | } | |
5300 | rcu_read_unlock(); | |
5301 | } | |
5302 | ||
60313ebe | 5303 | /* |
9f498cc5 PZ |
5304 | * task tracking -- fork/exit |
5305 | * | |
13d7a241 | 5306 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
5307 | */ |
5308 | ||
9f498cc5 | 5309 | struct perf_task_event { |
3a80b4a3 | 5310 | struct task_struct *task; |
cdd6c482 | 5311 | struct perf_event_context *task_ctx; |
60313ebe PZ |
5312 | |
5313 | struct { | |
5314 | struct perf_event_header header; | |
5315 | ||
5316 | u32 pid; | |
5317 | u32 ppid; | |
9f498cc5 PZ |
5318 | u32 tid; |
5319 | u32 ptid; | |
393b2ad8 | 5320 | u64 time; |
cdd6c482 | 5321 | } event_id; |
60313ebe PZ |
5322 | }; |
5323 | ||
67516844 JO |
5324 | static int perf_event_task_match(struct perf_event *event) |
5325 | { | |
13d7a241 SE |
5326 | return event->attr.comm || event->attr.mmap || |
5327 | event->attr.mmap2 || event->attr.mmap_data || | |
5328 | event->attr.task; | |
67516844 JO |
5329 | } |
5330 | ||
cdd6c482 | 5331 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 5332 | void *data) |
60313ebe | 5333 | { |
52d857a8 | 5334 | struct perf_task_event *task_event = data; |
60313ebe | 5335 | struct perf_output_handle handle; |
c980d109 | 5336 | struct perf_sample_data sample; |
9f498cc5 | 5337 | struct task_struct *task = task_event->task; |
c980d109 | 5338 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 5339 | |
67516844 JO |
5340 | if (!perf_event_task_match(event)) |
5341 | return; | |
5342 | ||
c980d109 | 5343 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 5344 | |
c980d109 | 5345 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 5346 | task_event->event_id.header.size); |
ef60777c | 5347 | if (ret) |
c980d109 | 5348 | goto out; |
60313ebe | 5349 | |
cdd6c482 IM |
5350 | task_event->event_id.pid = perf_event_pid(event, task); |
5351 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 5352 | |
cdd6c482 IM |
5353 | task_event->event_id.tid = perf_event_tid(event, task); |
5354 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 5355 | |
cdd6c482 | 5356 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 5357 | |
c980d109 ACM |
5358 | perf_event__output_id_sample(event, &handle, &sample); |
5359 | ||
60313ebe | 5360 | perf_output_end(&handle); |
c980d109 ACM |
5361 | out: |
5362 | task_event->event_id.header.size = size; | |
60313ebe PZ |
5363 | } |
5364 | ||
cdd6c482 IM |
5365 | static void perf_event_task(struct task_struct *task, |
5366 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 5367 | int new) |
60313ebe | 5368 | { |
9f498cc5 | 5369 | struct perf_task_event task_event; |
60313ebe | 5370 | |
cdd6c482 IM |
5371 | if (!atomic_read(&nr_comm_events) && |
5372 | !atomic_read(&nr_mmap_events) && | |
5373 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
5374 | return; |
5375 | ||
9f498cc5 | 5376 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
5377 | .task = task, |
5378 | .task_ctx = task_ctx, | |
cdd6c482 | 5379 | .event_id = { |
60313ebe | 5380 | .header = { |
cdd6c482 | 5381 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 5382 | .misc = 0, |
cdd6c482 | 5383 | .size = sizeof(task_event.event_id), |
60313ebe | 5384 | }, |
573402db PZ |
5385 | /* .pid */ |
5386 | /* .ppid */ | |
9f498cc5 PZ |
5387 | /* .tid */ |
5388 | /* .ptid */ | |
6f93d0a7 | 5389 | .time = perf_clock(), |
60313ebe PZ |
5390 | }, |
5391 | }; | |
5392 | ||
67516844 | 5393 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5394 | &task_event, |
5395 | task_ctx); | |
9f498cc5 PZ |
5396 | } |
5397 | ||
cdd6c482 | 5398 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5399 | { |
cdd6c482 | 5400 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5401 | } |
5402 | ||
8d1b2d93 PZ |
5403 | /* |
5404 | * comm tracking | |
5405 | */ | |
5406 | ||
5407 | struct perf_comm_event { | |
22a4f650 IM |
5408 | struct task_struct *task; |
5409 | char *comm; | |
8d1b2d93 PZ |
5410 | int comm_size; |
5411 | ||
5412 | struct { | |
5413 | struct perf_event_header header; | |
5414 | ||
5415 | u32 pid; | |
5416 | u32 tid; | |
cdd6c482 | 5417 | } event_id; |
8d1b2d93 PZ |
5418 | }; |
5419 | ||
67516844 JO |
5420 | static int perf_event_comm_match(struct perf_event *event) |
5421 | { | |
5422 | return event->attr.comm; | |
5423 | } | |
5424 | ||
cdd6c482 | 5425 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5426 | void *data) |
8d1b2d93 | 5427 | { |
52d857a8 | 5428 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5429 | struct perf_output_handle handle; |
c980d109 | 5430 | struct perf_sample_data sample; |
cdd6c482 | 5431 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5432 | int ret; |
5433 | ||
67516844 JO |
5434 | if (!perf_event_comm_match(event)) |
5435 | return; | |
5436 | ||
c980d109 ACM |
5437 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5438 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5439 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5440 | |
5441 | if (ret) | |
c980d109 | 5442 | goto out; |
8d1b2d93 | 5443 | |
cdd6c482 IM |
5444 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5445 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5446 | |
cdd6c482 | 5447 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5448 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5449 | comm_event->comm_size); |
c980d109 ACM |
5450 | |
5451 | perf_event__output_id_sample(event, &handle, &sample); | |
5452 | ||
8d1b2d93 | 5453 | perf_output_end(&handle); |
c980d109 ACM |
5454 | out: |
5455 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5456 | } |
5457 | ||
cdd6c482 | 5458 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5459 | { |
413ee3b4 | 5460 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5461 | unsigned int size; |
8d1b2d93 | 5462 | |
413ee3b4 | 5463 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5464 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5465 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5466 | |
5467 | comm_event->comm = comm; | |
5468 | comm_event->comm_size = size; | |
5469 | ||
cdd6c482 | 5470 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5471 | |
67516844 | 5472 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5473 | comm_event, |
5474 | NULL); | |
8d1b2d93 PZ |
5475 | } |
5476 | ||
82b89778 | 5477 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5478 | { |
9ee318a7 PZ |
5479 | struct perf_comm_event comm_event; |
5480 | ||
cdd6c482 | 5481 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5482 | return; |
a63eaf34 | 5483 | |
9ee318a7 | 5484 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5485 | .task = task, |
573402db PZ |
5486 | /* .comm */ |
5487 | /* .comm_size */ | |
cdd6c482 | 5488 | .event_id = { |
573402db | 5489 | .header = { |
cdd6c482 | 5490 | .type = PERF_RECORD_COMM, |
82b89778 | 5491 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5492 | /* .size */ |
5493 | }, | |
5494 | /* .pid */ | |
5495 | /* .tid */ | |
8d1b2d93 PZ |
5496 | }, |
5497 | }; | |
5498 | ||
cdd6c482 | 5499 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5500 | } |
5501 | ||
0a4a9391 PZ |
5502 | /* |
5503 | * mmap tracking | |
5504 | */ | |
5505 | ||
5506 | struct perf_mmap_event { | |
089dd79d PZ |
5507 | struct vm_area_struct *vma; |
5508 | ||
5509 | const char *file_name; | |
5510 | int file_size; | |
13d7a241 SE |
5511 | int maj, min; |
5512 | u64 ino; | |
5513 | u64 ino_generation; | |
f972eb63 | 5514 | u32 prot, flags; |
0a4a9391 PZ |
5515 | |
5516 | struct { | |
5517 | struct perf_event_header header; | |
5518 | ||
5519 | u32 pid; | |
5520 | u32 tid; | |
5521 | u64 start; | |
5522 | u64 len; | |
5523 | u64 pgoff; | |
cdd6c482 | 5524 | } event_id; |
0a4a9391 PZ |
5525 | }; |
5526 | ||
67516844 JO |
5527 | static int perf_event_mmap_match(struct perf_event *event, |
5528 | void *data) | |
5529 | { | |
5530 | struct perf_mmap_event *mmap_event = data; | |
5531 | struct vm_area_struct *vma = mmap_event->vma; | |
5532 | int executable = vma->vm_flags & VM_EXEC; | |
5533 | ||
5534 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5535 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5536 | } |
5537 | ||
cdd6c482 | 5538 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5539 | void *data) |
0a4a9391 | 5540 | { |
52d857a8 | 5541 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5542 | struct perf_output_handle handle; |
c980d109 | 5543 | struct perf_sample_data sample; |
cdd6c482 | 5544 | int size = mmap_event->event_id.header.size; |
c980d109 | 5545 | int ret; |
0a4a9391 | 5546 | |
67516844 JO |
5547 | if (!perf_event_mmap_match(event, data)) |
5548 | return; | |
5549 | ||
13d7a241 SE |
5550 | if (event->attr.mmap2) { |
5551 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5552 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5553 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5554 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5555 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
f972eb63 PZ |
5556 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
5557 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); | |
13d7a241 SE |
5558 | } |
5559 | ||
c980d109 ACM |
5560 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5561 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5562 | mmap_event->event_id.header.size); |
0a4a9391 | 5563 | if (ret) |
c980d109 | 5564 | goto out; |
0a4a9391 | 5565 | |
cdd6c482 IM |
5566 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5567 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5568 | |
cdd6c482 | 5569 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5570 | |
5571 | if (event->attr.mmap2) { | |
5572 | perf_output_put(&handle, mmap_event->maj); | |
5573 | perf_output_put(&handle, mmap_event->min); | |
5574 | perf_output_put(&handle, mmap_event->ino); | |
5575 | perf_output_put(&handle, mmap_event->ino_generation); | |
f972eb63 PZ |
5576 | perf_output_put(&handle, mmap_event->prot); |
5577 | perf_output_put(&handle, mmap_event->flags); | |
13d7a241 SE |
5578 | } |
5579 | ||
76369139 | 5580 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5581 | mmap_event->file_size); |
c980d109 ACM |
5582 | |
5583 | perf_event__output_id_sample(event, &handle, &sample); | |
5584 | ||
78d613eb | 5585 | perf_output_end(&handle); |
c980d109 ACM |
5586 | out: |
5587 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5588 | } |
5589 | ||
cdd6c482 | 5590 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5591 | { |
089dd79d PZ |
5592 | struct vm_area_struct *vma = mmap_event->vma; |
5593 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5594 | int maj = 0, min = 0; |
5595 | u64 ino = 0, gen = 0; | |
f972eb63 | 5596 | u32 prot = 0, flags = 0; |
0a4a9391 PZ |
5597 | unsigned int size; |
5598 | char tmp[16]; | |
5599 | char *buf = NULL; | |
2c42cfbf | 5600 | char *name; |
413ee3b4 | 5601 | |
0a4a9391 | 5602 | if (file) { |
13d7a241 SE |
5603 | struct inode *inode; |
5604 | dev_t dev; | |
3ea2f2b9 | 5605 | |
2c42cfbf | 5606 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5607 | if (!buf) { |
c7e548b4 ON |
5608 | name = "//enomem"; |
5609 | goto cpy_name; | |
0a4a9391 | 5610 | } |
413ee3b4 | 5611 | /* |
3ea2f2b9 | 5612 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5613 | * need to add enough zero bytes after the string to handle |
5614 | * the 64bit alignment we do later. | |
5615 | */ | |
3ea2f2b9 | 5616 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5617 | if (IS_ERR(name)) { |
c7e548b4 ON |
5618 | name = "//toolong"; |
5619 | goto cpy_name; | |
0a4a9391 | 5620 | } |
13d7a241 SE |
5621 | inode = file_inode(vma->vm_file); |
5622 | dev = inode->i_sb->s_dev; | |
5623 | ino = inode->i_ino; | |
5624 | gen = inode->i_generation; | |
5625 | maj = MAJOR(dev); | |
5626 | min = MINOR(dev); | |
f972eb63 PZ |
5627 | |
5628 | if (vma->vm_flags & VM_READ) | |
5629 | prot |= PROT_READ; | |
5630 | if (vma->vm_flags & VM_WRITE) | |
5631 | prot |= PROT_WRITE; | |
5632 | if (vma->vm_flags & VM_EXEC) | |
5633 | prot |= PROT_EXEC; | |
5634 | ||
5635 | if (vma->vm_flags & VM_MAYSHARE) | |
5636 | flags = MAP_SHARED; | |
5637 | else | |
5638 | flags = MAP_PRIVATE; | |
5639 | ||
5640 | if (vma->vm_flags & VM_DENYWRITE) | |
5641 | flags |= MAP_DENYWRITE; | |
5642 | if (vma->vm_flags & VM_MAYEXEC) | |
5643 | flags |= MAP_EXECUTABLE; | |
5644 | if (vma->vm_flags & VM_LOCKED) | |
5645 | flags |= MAP_LOCKED; | |
5646 | if (vma->vm_flags & VM_HUGETLB) | |
5647 | flags |= MAP_HUGETLB; | |
5648 | ||
c7e548b4 | 5649 | goto got_name; |
0a4a9391 | 5650 | } else { |
fbe26abe JO |
5651 | if (vma->vm_ops && vma->vm_ops->name) { |
5652 | name = (char *) vma->vm_ops->name(vma); | |
5653 | if (name) | |
5654 | goto cpy_name; | |
5655 | } | |
5656 | ||
2c42cfbf | 5657 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5658 | if (name) |
5659 | goto cpy_name; | |
089dd79d | 5660 | |
32c5fb7e | 5661 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5662 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5663 | name = "[heap]"; |
5664 | goto cpy_name; | |
32c5fb7e ON |
5665 | } |
5666 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5667 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5668 | name = "[stack]"; |
5669 | goto cpy_name; | |
089dd79d PZ |
5670 | } |
5671 | ||
c7e548b4 ON |
5672 | name = "//anon"; |
5673 | goto cpy_name; | |
0a4a9391 PZ |
5674 | } |
5675 | ||
c7e548b4 ON |
5676 | cpy_name: |
5677 | strlcpy(tmp, name, sizeof(tmp)); | |
5678 | name = tmp; | |
0a4a9391 | 5679 | got_name: |
2c42cfbf PZ |
5680 | /* |
5681 | * Since our buffer works in 8 byte units we need to align our string | |
5682 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5683 | * zero'd out to avoid leaking random bits to userspace. | |
5684 | */ | |
5685 | size = strlen(name)+1; | |
5686 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5687 | name[size++] = '\0'; | |
0a4a9391 PZ |
5688 | |
5689 | mmap_event->file_name = name; | |
5690 | mmap_event->file_size = size; | |
13d7a241 SE |
5691 | mmap_event->maj = maj; |
5692 | mmap_event->min = min; | |
5693 | mmap_event->ino = ino; | |
5694 | mmap_event->ino_generation = gen; | |
f972eb63 PZ |
5695 | mmap_event->prot = prot; |
5696 | mmap_event->flags = flags; | |
0a4a9391 | 5697 | |
2fe85427 SE |
5698 | if (!(vma->vm_flags & VM_EXEC)) |
5699 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5700 | ||
cdd6c482 | 5701 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5702 | |
67516844 | 5703 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5704 | mmap_event, |
5705 | NULL); | |
665c2142 | 5706 | |
0a4a9391 PZ |
5707 | kfree(buf); |
5708 | } | |
5709 | ||
3af9e859 | 5710 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5711 | { |
9ee318a7 PZ |
5712 | struct perf_mmap_event mmap_event; |
5713 | ||
cdd6c482 | 5714 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5715 | return; |
5716 | ||
5717 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5718 | .vma = vma, |
573402db PZ |
5719 | /* .file_name */ |
5720 | /* .file_size */ | |
cdd6c482 | 5721 | .event_id = { |
573402db | 5722 | .header = { |
cdd6c482 | 5723 | .type = PERF_RECORD_MMAP, |
39447b38 | 5724 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5725 | /* .size */ |
5726 | }, | |
5727 | /* .pid */ | |
5728 | /* .tid */ | |
089dd79d PZ |
5729 | .start = vma->vm_start, |
5730 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5731 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5732 | }, |
13d7a241 SE |
5733 | /* .maj (attr_mmap2 only) */ |
5734 | /* .min (attr_mmap2 only) */ | |
5735 | /* .ino (attr_mmap2 only) */ | |
5736 | /* .ino_generation (attr_mmap2 only) */ | |
f972eb63 PZ |
5737 | /* .prot (attr_mmap2 only) */ |
5738 | /* .flags (attr_mmap2 only) */ | |
0a4a9391 PZ |
5739 | }; |
5740 | ||
cdd6c482 | 5741 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5742 | } |
5743 | ||
a78ac325 PZ |
5744 | /* |
5745 | * IRQ throttle logging | |
5746 | */ | |
5747 | ||
cdd6c482 | 5748 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5749 | { |
5750 | struct perf_output_handle handle; | |
c980d109 | 5751 | struct perf_sample_data sample; |
a78ac325 PZ |
5752 | int ret; |
5753 | ||
5754 | struct { | |
5755 | struct perf_event_header header; | |
5756 | u64 time; | |
cca3f454 | 5757 | u64 id; |
7f453c24 | 5758 | u64 stream_id; |
a78ac325 PZ |
5759 | } throttle_event = { |
5760 | .header = { | |
cdd6c482 | 5761 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5762 | .misc = 0, |
5763 | .size = sizeof(throttle_event), | |
5764 | }, | |
def0a9b2 | 5765 | .time = perf_clock(), |
cdd6c482 IM |
5766 | .id = primary_event_id(event), |
5767 | .stream_id = event->id, | |
a78ac325 PZ |
5768 | }; |
5769 | ||
966ee4d6 | 5770 | if (enable) |
cdd6c482 | 5771 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5772 | |
c980d109 ACM |
5773 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5774 | ||
5775 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5776 | throttle_event.header.size); |
a78ac325 PZ |
5777 | if (ret) |
5778 | return; | |
5779 | ||
5780 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5781 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5782 | perf_output_end(&handle); |
5783 | } | |
5784 | ||
f6c7d5fe | 5785 | /* |
cdd6c482 | 5786 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5787 | */ |
5788 | ||
a8b0ca17 | 5789 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5790 | int throttle, struct perf_sample_data *data, |
5791 | struct pt_regs *regs) | |
f6c7d5fe | 5792 | { |
cdd6c482 IM |
5793 | int events = atomic_read(&event->event_limit); |
5794 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5795 | u64 seq; |
79f14641 PZ |
5796 | int ret = 0; |
5797 | ||
96398826 PZ |
5798 | /* |
5799 | * Non-sampling counters might still use the PMI to fold short | |
5800 | * hardware counters, ignore those. | |
5801 | */ | |
5802 | if (unlikely(!is_sampling_event(event))) | |
5803 | return 0; | |
5804 | ||
e050e3f0 SE |
5805 | seq = __this_cpu_read(perf_throttled_seq); |
5806 | if (seq != hwc->interrupts_seq) { | |
5807 | hwc->interrupts_seq = seq; | |
5808 | hwc->interrupts = 1; | |
5809 | } else { | |
5810 | hwc->interrupts++; | |
5811 | if (unlikely(throttle | |
5812 | && hwc->interrupts >= max_samples_per_tick)) { | |
5813 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5814 | hwc->interrupts = MAX_INTERRUPTS; |
5815 | perf_log_throttle(event, 0); | |
d84153d6 | 5816 | tick_nohz_full_kick(); |
a78ac325 PZ |
5817 | ret = 1; |
5818 | } | |
e050e3f0 | 5819 | } |
60db5e09 | 5820 | |
cdd6c482 | 5821 | if (event->attr.freq) { |
def0a9b2 | 5822 | u64 now = perf_clock(); |
abd50713 | 5823 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5824 | |
abd50713 | 5825 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5826 | |
abd50713 | 5827 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5828 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5829 | } |
5830 | ||
2023b359 PZ |
5831 | /* |
5832 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5833 | * events |
2023b359 PZ |
5834 | */ |
5835 | ||
cdd6c482 IM |
5836 | event->pending_kill = POLL_IN; |
5837 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5838 | ret = 1; |
cdd6c482 | 5839 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5840 | event->pending_disable = 1; |
5841 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5842 | } |
5843 | ||
453f19ee | 5844 | if (event->overflow_handler) |
a8b0ca17 | 5845 | event->overflow_handler(event, data, regs); |
453f19ee | 5846 | else |
a8b0ca17 | 5847 | perf_event_output(event, data, regs); |
453f19ee | 5848 | |
f506b3dc | 5849 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5850 | event->pending_wakeup = 1; |
5851 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5852 | } |
5853 | ||
79f14641 | 5854 | return ret; |
f6c7d5fe PZ |
5855 | } |
5856 | ||
a8b0ca17 | 5857 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5858 | struct perf_sample_data *data, |
5859 | struct pt_regs *regs) | |
850bc73f | 5860 | { |
a8b0ca17 | 5861 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5862 | } |
5863 | ||
15dbf27c | 5864 | /* |
cdd6c482 | 5865 | * Generic software event infrastructure |
15dbf27c PZ |
5866 | */ |
5867 | ||
b28ab83c PZ |
5868 | struct swevent_htable { |
5869 | struct swevent_hlist *swevent_hlist; | |
5870 | struct mutex hlist_mutex; | |
5871 | int hlist_refcount; | |
5872 | ||
5873 | /* Recursion avoidance in each contexts */ | |
5874 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
5875 | |
5876 | /* Keeps track of cpu being initialized/exited */ | |
5877 | bool online; | |
b28ab83c PZ |
5878 | }; |
5879 | ||
5880 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5881 | ||
7b4b6658 | 5882 | /* |
cdd6c482 IM |
5883 | * We directly increment event->count and keep a second value in |
5884 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5885 | * is kept in the range [-sample_period, 0] so that we can use the |
5886 | * sign as trigger. | |
5887 | */ | |
5888 | ||
ab573844 | 5889 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5890 | { |
cdd6c482 | 5891 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5892 | u64 period = hwc->last_period; |
5893 | u64 nr, offset; | |
5894 | s64 old, val; | |
5895 | ||
5896 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5897 | |
5898 | again: | |
e7850595 | 5899 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5900 | if (val < 0) |
5901 | return 0; | |
15dbf27c | 5902 | |
7b4b6658 PZ |
5903 | nr = div64_u64(period + val, period); |
5904 | offset = nr * period; | |
5905 | val -= offset; | |
e7850595 | 5906 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5907 | goto again; |
15dbf27c | 5908 | |
7b4b6658 | 5909 | return nr; |
15dbf27c PZ |
5910 | } |
5911 | ||
0cff784a | 5912 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5913 | struct perf_sample_data *data, |
5622f295 | 5914 | struct pt_regs *regs) |
15dbf27c | 5915 | { |
cdd6c482 | 5916 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5917 | int throttle = 0; |
15dbf27c | 5918 | |
0cff784a PZ |
5919 | if (!overflow) |
5920 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5921 | |
7b4b6658 PZ |
5922 | if (hwc->interrupts == MAX_INTERRUPTS) |
5923 | return; | |
15dbf27c | 5924 | |
7b4b6658 | 5925 | for (; overflow; overflow--) { |
a8b0ca17 | 5926 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5927 | data, regs)) { |
7b4b6658 PZ |
5928 | /* |
5929 | * We inhibit the overflow from happening when | |
5930 | * hwc->interrupts == MAX_INTERRUPTS. | |
5931 | */ | |
5932 | break; | |
5933 | } | |
cf450a73 | 5934 | throttle = 1; |
7b4b6658 | 5935 | } |
15dbf27c PZ |
5936 | } |
5937 | ||
a4eaf7f1 | 5938 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5939 | struct perf_sample_data *data, |
5622f295 | 5940 | struct pt_regs *regs) |
7b4b6658 | 5941 | { |
cdd6c482 | 5942 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5943 | |
e7850595 | 5944 | local64_add(nr, &event->count); |
d6d020e9 | 5945 | |
0cff784a PZ |
5946 | if (!regs) |
5947 | return; | |
5948 | ||
6c7e550f | 5949 | if (!is_sampling_event(event)) |
7b4b6658 | 5950 | return; |
d6d020e9 | 5951 | |
5d81e5cf AV |
5952 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5953 | data->period = nr; | |
5954 | return perf_swevent_overflow(event, 1, data, regs); | |
5955 | } else | |
5956 | data->period = event->hw.last_period; | |
5957 | ||
0cff784a | 5958 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5959 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5960 | |
e7850595 | 5961 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5962 | return; |
df1a132b | 5963 | |
a8b0ca17 | 5964 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5965 | } |
5966 | ||
f5ffe02e FW |
5967 | static int perf_exclude_event(struct perf_event *event, |
5968 | struct pt_regs *regs) | |
5969 | { | |
a4eaf7f1 | 5970 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5971 | return 1; |
a4eaf7f1 | 5972 | |
f5ffe02e FW |
5973 | if (regs) { |
5974 | if (event->attr.exclude_user && user_mode(regs)) | |
5975 | return 1; | |
5976 | ||
5977 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5978 | return 1; | |
5979 | } | |
5980 | ||
5981 | return 0; | |
5982 | } | |
5983 | ||
cdd6c482 | 5984 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5985 | enum perf_type_id type, |
6fb2915d LZ |
5986 | u32 event_id, |
5987 | struct perf_sample_data *data, | |
5988 | struct pt_regs *regs) | |
15dbf27c | 5989 | { |
cdd6c482 | 5990 | if (event->attr.type != type) |
a21ca2ca | 5991 | return 0; |
f5ffe02e | 5992 | |
cdd6c482 | 5993 | if (event->attr.config != event_id) |
15dbf27c PZ |
5994 | return 0; |
5995 | ||
f5ffe02e FW |
5996 | if (perf_exclude_event(event, regs)) |
5997 | return 0; | |
15dbf27c PZ |
5998 | |
5999 | return 1; | |
6000 | } | |
6001 | ||
76e1d904 FW |
6002 | static inline u64 swevent_hash(u64 type, u32 event_id) |
6003 | { | |
6004 | u64 val = event_id | (type << 32); | |
6005 | ||
6006 | return hash_64(val, SWEVENT_HLIST_BITS); | |
6007 | } | |
6008 | ||
49f135ed FW |
6009 | static inline struct hlist_head * |
6010 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 6011 | { |
49f135ed FW |
6012 | u64 hash = swevent_hash(type, event_id); |
6013 | ||
6014 | return &hlist->heads[hash]; | |
6015 | } | |
76e1d904 | 6016 | |
49f135ed FW |
6017 | /* For the read side: events when they trigger */ |
6018 | static inline struct hlist_head * | |
b28ab83c | 6019 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
6020 | { |
6021 | struct swevent_hlist *hlist; | |
76e1d904 | 6022 | |
b28ab83c | 6023 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
6024 | if (!hlist) |
6025 | return NULL; | |
6026 | ||
49f135ed FW |
6027 | return __find_swevent_head(hlist, type, event_id); |
6028 | } | |
6029 | ||
6030 | /* For the event head insertion and removal in the hlist */ | |
6031 | static inline struct hlist_head * | |
b28ab83c | 6032 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
6033 | { |
6034 | struct swevent_hlist *hlist; | |
6035 | u32 event_id = event->attr.config; | |
6036 | u64 type = event->attr.type; | |
6037 | ||
6038 | /* | |
6039 | * Event scheduling is always serialized against hlist allocation | |
6040 | * and release. Which makes the protected version suitable here. | |
6041 | * The context lock guarantees that. | |
6042 | */ | |
b28ab83c | 6043 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
6044 | lockdep_is_held(&event->ctx->lock)); |
6045 | if (!hlist) | |
6046 | return NULL; | |
6047 | ||
6048 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
6049 | } |
6050 | ||
6051 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 6052 | u64 nr, |
76e1d904 FW |
6053 | struct perf_sample_data *data, |
6054 | struct pt_regs *regs) | |
15dbf27c | 6055 | { |
4a32fea9 | 6056 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6057 | struct perf_event *event; |
76e1d904 | 6058 | struct hlist_head *head; |
15dbf27c | 6059 | |
76e1d904 | 6060 | rcu_read_lock(); |
b28ab83c | 6061 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
6062 | if (!head) |
6063 | goto end; | |
6064 | ||
b67bfe0d | 6065 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 6066 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 6067 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 6068 | } |
76e1d904 FW |
6069 | end: |
6070 | rcu_read_unlock(); | |
15dbf27c PZ |
6071 | } |
6072 | ||
86038c5e PZI |
6073 | DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]); |
6074 | ||
4ed7c92d | 6075 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 6076 | { |
4a32fea9 | 6077 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
96f6d444 | 6078 | |
b28ab83c | 6079 | return get_recursion_context(swhash->recursion); |
96f6d444 | 6080 | } |
645e8cc0 | 6081 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 6082 | |
fa9f90be | 6083 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 6084 | { |
4a32fea9 | 6085 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
927c7a9e | 6086 | |
b28ab83c | 6087 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 6088 | } |
15dbf27c | 6089 | |
86038c5e | 6090 | void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 6091 | { |
a4234bfc | 6092 | struct perf_sample_data data; |
4ed7c92d | 6093 | |
86038c5e | 6094 | if (WARN_ON_ONCE(!regs)) |
4ed7c92d | 6095 | return; |
a4234bfc | 6096 | |
fd0d000b | 6097 | perf_sample_data_init(&data, addr, 0); |
a8b0ca17 | 6098 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
86038c5e PZI |
6099 | } |
6100 | ||
6101 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) | |
6102 | { | |
6103 | int rctx; | |
6104 | ||
6105 | preempt_disable_notrace(); | |
6106 | rctx = perf_swevent_get_recursion_context(); | |
6107 | if (unlikely(rctx < 0)) | |
6108 | goto fail; | |
6109 | ||
6110 | ___perf_sw_event(event_id, nr, regs, addr); | |
4ed7c92d PZ |
6111 | |
6112 | perf_swevent_put_recursion_context(rctx); | |
86038c5e | 6113 | fail: |
1c024eca | 6114 | preempt_enable_notrace(); |
b8e83514 PZ |
6115 | } |
6116 | ||
cdd6c482 | 6117 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 6118 | { |
15dbf27c PZ |
6119 | } |
6120 | ||
a4eaf7f1 | 6121 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 6122 | { |
4a32fea9 | 6123 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6124 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
6125 | struct hlist_head *head; |
6126 | ||
6c7e550f | 6127 | if (is_sampling_event(event)) { |
7b4b6658 | 6128 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 6129 | perf_swevent_set_period(event); |
7b4b6658 | 6130 | } |
76e1d904 | 6131 | |
a4eaf7f1 PZ |
6132 | hwc->state = !(flags & PERF_EF_START); |
6133 | ||
b28ab83c | 6134 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
6135 | if (!head) { |
6136 | /* | |
6137 | * We can race with cpu hotplug code. Do not | |
6138 | * WARN if the cpu just got unplugged. | |
6139 | */ | |
6140 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 6141 | return -EINVAL; |
39af6b16 | 6142 | } |
76e1d904 FW |
6143 | |
6144 | hlist_add_head_rcu(&event->hlist_entry, head); | |
6a694a60 | 6145 | perf_event_update_userpage(event); |
76e1d904 | 6146 | |
15dbf27c PZ |
6147 | return 0; |
6148 | } | |
6149 | ||
a4eaf7f1 | 6150 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 6151 | { |
76e1d904 | 6152 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
6153 | } |
6154 | ||
a4eaf7f1 | 6155 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 6156 | { |
a4eaf7f1 | 6157 | event->hw.state = 0; |
d6d020e9 | 6158 | } |
aa9c4c0f | 6159 | |
a4eaf7f1 | 6160 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 6161 | { |
a4eaf7f1 | 6162 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
6163 | } |
6164 | ||
49f135ed FW |
6165 | /* Deref the hlist from the update side */ |
6166 | static inline struct swevent_hlist * | |
b28ab83c | 6167 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 6168 | { |
b28ab83c PZ |
6169 | return rcu_dereference_protected(swhash->swevent_hlist, |
6170 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
6171 | } |
6172 | ||
b28ab83c | 6173 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 6174 | { |
b28ab83c | 6175 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 6176 | |
49f135ed | 6177 | if (!hlist) |
76e1d904 FW |
6178 | return; |
6179 | ||
70691d4a | 6180 | RCU_INIT_POINTER(swhash->swevent_hlist, NULL); |
fa4bbc4c | 6181 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
6182 | } |
6183 | ||
6184 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
6185 | { | |
b28ab83c | 6186 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 6187 | |
b28ab83c | 6188 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6189 | |
b28ab83c PZ |
6190 | if (!--swhash->hlist_refcount) |
6191 | swevent_hlist_release(swhash); | |
76e1d904 | 6192 | |
b28ab83c | 6193 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6194 | } |
6195 | ||
6196 | static void swevent_hlist_put(struct perf_event *event) | |
6197 | { | |
6198 | int cpu; | |
6199 | ||
76e1d904 FW |
6200 | for_each_possible_cpu(cpu) |
6201 | swevent_hlist_put_cpu(event, cpu); | |
6202 | } | |
6203 | ||
6204 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
6205 | { | |
b28ab83c | 6206 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
6207 | int err = 0; |
6208 | ||
b28ab83c | 6209 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6210 | |
b28ab83c | 6211 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
6212 | struct swevent_hlist *hlist; |
6213 | ||
6214 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
6215 | if (!hlist) { | |
6216 | err = -ENOMEM; | |
6217 | goto exit; | |
6218 | } | |
b28ab83c | 6219 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 6220 | } |
b28ab83c | 6221 | swhash->hlist_refcount++; |
9ed6060d | 6222 | exit: |
b28ab83c | 6223 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6224 | |
6225 | return err; | |
6226 | } | |
6227 | ||
6228 | static int swevent_hlist_get(struct perf_event *event) | |
6229 | { | |
6230 | int err; | |
6231 | int cpu, failed_cpu; | |
6232 | ||
76e1d904 FW |
6233 | get_online_cpus(); |
6234 | for_each_possible_cpu(cpu) { | |
6235 | err = swevent_hlist_get_cpu(event, cpu); | |
6236 | if (err) { | |
6237 | failed_cpu = cpu; | |
6238 | goto fail; | |
6239 | } | |
6240 | } | |
6241 | put_online_cpus(); | |
6242 | ||
6243 | return 0; | |
9ed6060d | 6244 | fail: |
76e1d904 FW |
6245 | for_each_possible_cpu(cpu) { |
6246 | if (cpu == failed_cpu) | |
6247 | break; | |
6248 | swevent_hlist_put_cpu(event, cpu); | |
6249 | } | |
6250 | ||
6251 | put_online_cpus(); | |
6252 | return err; | |
6253 | } | |
6254 | ||
c5905afb | 6255 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 6256 | |
b0a873eb PZ |
6257 | static void sw_perf_event_destroy(struct perf_event *event) |
6258 | { | |
6259 | u64 event_id = event->attr.config; | |
95476b64 | 6260 | |
b0a873eb PZ |
6261 | WARN_ON(event->parent); |
6262 | ||
c5905afb | 6263 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6264 | swevent_hlist_put(event); |
6265 | } | |
6266 | ||
6267 | static int perf_swevent_init(struct perf_event *event) | |
6268 | { | |
8176cced | 6269 | u64 event_id = event->attr.config; |
b0a873eb PZ |
6270 | |
6271 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6272 | return -ENOENT; | |
6273 | ||
2481c5fa SE |
6274 | /* |
6275 | * no branch sampling for software events | |
6276 | */ | |
6277 | if (has_branch_stack(event)) | |
6278 | return -EOPNOTSUPP; | |
6279 | ||
b0a873eb PZ |
6280 | switch (event_id) { |
6281 | case PERF_COUNT_SW_CPU_CLOCK: | |
6282 | case PERF_COUNT_SW_TASK_CLOCK: | |
6283 | return -ENOENT; | |
6284 | ||
6285 | default: | |
6286 | break; | |
6287 | } | |
6288 | ||
ce677831 | 6289 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
6290 | return -ENOENT; |
6291 | ||
6292 | if (!event->parent) { | |
6293 | int err; | |
6294 | ||
6295 | err = swevent_hlist_get(event); | |
6296 | if (err) | |
6297 | return err; | |
6298 | ||
c5905afb | 6299 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6300 | event->destroy = sw_perf_event_destroy; |
6301 | } | |
6302 | ||
6303 | return 0; | |
6304 | } | |
6305 | ||
6306 | static struct pmu perf_swevent = { | |
89a1e187 | 6307 | .task_ctx_nr = perf_sw_context, |
95476b64 | 6308 | |
b0a873eb | 6309 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
6310 | .add = perf_swevent_add, |
6311 | .del = perf_swevent_del, | |
6312 | .start = perf_swevent_start, | |
6313 | .stop = perf_swevent_stop, | |
1c024eca | 6314 | .read = perf_swevent_read, |
1c024eca PZ |
6315 | }; |
6316 | ||
b0a873eb PZ |
6317 | #ifdef CONFIG_EVENT_TRACING |
6318 | ||
1c024eca PZ |
6319 | static int perf_tp_filter_match(struct perf_event *event, |
6320 | struct perf_sample_data *data) | |
6321 | { | |
6322 | void *record = data->raw->data; | |
6323 | ||
6324 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
6325 | return 1; | |
6326 | return 0; | |
6327 | } | |
6328 | ||
6329 | static int perf_tp_event_match(struct perf_event *event, | |
6330 | struct perf_sample_data *data, | |
6331 | struct pt_regs *regs) | |
6332 | { | |
a0f7d0f7 FW |
6333 | if (event->hw.state & PERF_HES_STOPPED) |
6334 | return 0; | |
580d607c PZ |
6335 | /* |
6336 | * All tracepoints are from kernel-space. | |
6337 | */ | |
6338 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
6339 | return 0; |
6340 | ||
6341 | if (!perf_tp_filter_match(event, data)) | |
6342 | return 0; | |
6343 | ||
6344 | return 1; | |
6345 | } | |
6346 | ||
6347 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
6348 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
6349 | struct task_struct *task) | |
95476b64 FW |
6350 | { |
6351 | struct perf_sample_data data; | |
1c024eca | 6352 | struct perf_event *event; |
1c024eca | 6353 | |
95476b64 FW |
6354 | struct perf_raw_record raw = { |
6355 | .size = entry_size, | |
6356 | .data = record, | |
6357 | }; | |
6358 | ||
fd0d000b | 6359 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
6360 | data.raw = &raw; |
6361 | ||
b67bfe0d | 6362 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 6363 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 6364 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 6365 | } |
ecc55f84 | 6366 | |
e6dab5ff AV |
6367 | /* |
6368 | * If we got specified a target task, also iterate its context and | |
6369 | * deliver this event there too. | |
6370 | */ | |
6371 | if (task && task != current) { | |
6372 | struct perf_event_context *ctx; | |
6373 | struct trace_entry *entry = record; | |
6374 | ||
6375 | rcu_read_lock(); | |
6376 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
6377 | if (!ctx) | |
6378 | goto unlock; | |
6379 | ||
6380 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
6381 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6382 | continue; | |
6383 | if (event->attr.config != entry->type) | |
6384 | continue; | |
6385 | if (perf_tp_event_match(event, &data, regs)) | |
6386 | perf_swevent_event(event, count, &data, regs); | |
6387 | } | |
6388 | unlock: | |
6389 | rcu_read_unlock(); | |
6390 | } | |
6391 | ||
ecc55f84 | 6392 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
6393 | } |
6394 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
6395 | ||
cdd6c482 | 6396 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 6397 | { |
1c024eca | 6398 | perf_trace_destroy(event); |
e077df4f PZ |
6399 | } |
6400 | ||
b0a873eb | 6401 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 6402 | { |
76e1d904 FW |
6403 | int err; |
6404 | ||
b0a873eb PZ |
6405 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
6406 | return -ENOENT; | |
6407 | ||
2481c5fa SE |
6408 | /* |
6409 | * no branch sampling for tracepoint events | |
6410 | */ | |
6411 | if (has_branch_stack(event)) | |
6412 | return -EOPNOTSUPP; | |
6413 | ||
1c024eca PZ |
6414 | err = perf_trace_init(event); |
6415 | if (err) | |
b0a873eb | 6416 | return err; |
e077df4f | 6417 | |
cdd6c482 | 6418 | event->destroy = tp_perf_event_destroy; |
e077df4f | 6419 | |
b0a873eb PZ |
6420 | return 0; |
6421 | } | |
6422 | ||
6423 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
6424 | .task_ctx_nr = perf_sw_context, |
6425 | ||
b0a873eb | 6426 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6427 | .add = perf_trace_add, |
6428 | .del = perf_trace_del, | |
6429 | .start = perf_swevent_start, | |
6430 | .stop = perf_swevent_stop, | |
b0a873eb | 6431 | .read = perf_swevent_read, |
b0a873eb PZ |
6432 | }; |
6433 | ||
6434 | static inline void perf_tp_register(void) | |
6435 | { | |
2e80a82a | 6436 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6437 | } |
6fb2915d LZ |
6438 | |
6439 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6440 | { | |
6441 | char *filter_str; | |
6442 | int ret; | |
6443 | ||
6444 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6445 | return -EINVAL; | |
6446 | ||
6447 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6448 | if (IS_ERR(filter_str)) | |
6449 | return PTR_ERR(filter_str); | |
6450 | ||
6451 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6452 | ||
6453 | kfree(filter_str); | |
6454 | return ret; | |
6455 | } | |
6456 | ||
6457 | static void perf_event_free_filter(struct perf_event *event) | |
6458 | { | |
6459 | ftrace_profile_free_filter(event); | |
6460 | } | |
6461 | ||
e077df4f | 6462 | #else |
6fb2915d | 6463 | |
b0a873eb | 6464 | static inline void perf_tp_register(void) |
e077df4f | 6465 | { |
e077df4f | 6466 | } |
6fb2915d LZ |
6467 | |
6468 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6469 | { | |
6470 | return -ENOENT; | |
6471 | } | |
6472 | ||
6473 | static void perf_event_free_filter(struct perf_event *event) | |
6474 | { | |
6475 | } | |
6476 | ||
07b139c8 | 6477 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6478 | |
24f1e32c | 6479 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6480 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6481 | { |
f5ffe02e FW |
6482 | struct perf_sample_data sample; |
6483 | struct pt_regs *regs = data; | |
6484 | ||
fd0d000b | 6485 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6486 | |
a4eaf7f1 | 6487 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6488 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6489 | } |
6490 | #endif | |
6491 | ||
b0a873eb PZ |
6492 | /* |
6493 | * hrtimer based swevent callback | |
6494 | */ | |
f29ac756 | 6495 | |
b0a873eb | 6496 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6497 | { |
b0a873eb PZ |
6498 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6499 | struct perf_sample_data data; | |
6500 | struct pt_regs *regs; | |
6501 | struct perf_event *event; | |
6502 | u64 period; | |
f29ac756 | 6503 | |
b0a873eb | 6504 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6505 | |
6506 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6507 | return HRTIMER_NORESTART; | |
6508 | ||
b0a873eb | 6509 | event->pmu->read(event); |
f344011c | 6510 | |
fd0d000b | 6511 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6512 | regs = get_irq_regs(); |
6513 | ||
6514 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6515 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6516 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6517 | ret = HRTIMER_NORESTART; |
6518 | } | |
24f1e32c | 6519 | |
b0a873eb PZ |
6520 | period = max_t(u64, 10000, event->hw.sample_period); |
6521 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6522 | |
b0a873eb | 6523 | return ret; |
f29ac756 PZ |
6524 | } |
6525 | ||
b0a873eb | 6526 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6527 | { |
b0a873eb | 6528 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6529 | s64 period; |
6530 | ||
6531 | if (!is_sampling_event(event)) | |
6532 | return; | |
f5ffe02e | 6533 | |
5d508e82 FBH |
6534 | period = local64_read(&hwc->period_left); |
6535 | if (period) { | |
6536 | if (period < 0) | |
6537 | period = 10000; | |
fa407f35 | 6538 | |
5d508e82 FBH |
6539 | local64_set(&hwc->period_left, 0); |
6540 | } else { | |
6541 | period = max_t(u64, 10000, hwc->sample_period); | |
6542 | } | |
6543 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6544 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6545 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6546 | } |
b0a873eb PZ |
6547 | |
6548 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6549 | { |
b0a873eb PZ |
6550 | struct hw_perf_event *hwc = &event->hw; |
6551 | ||
6c7e550f | 6552 | if (is_sampling_event(event)) { |
b0a873eb | 6553 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6554 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6555 | |
6556 | hrtimer_cancel(&hwc->hrtimer); | |
6557 | } | |
24f1e32c FW |
6558 | } |
6559 | ||
ba3dd36c PZ |
6560 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6561 | { | |
6562 | struct hw_perf_event *hwc = &event->hw; | |
6563 | ||
6564 | if (!is_sampling_event(event)) | |
6565 | return; | |
6566 | ||
6567 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6568 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6569 | ||
6570 | /* | |
6571 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6572 | * mapping and avoid the whole period adjust feedback stuff. | |
6573 | */ | |
6574 | if (event->attr.freq) { | |
6575 | long freq = event->attr.sample_freq; | |
6576 | ||
6577 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6578 | hwc->sample_period = event->attr.sample_period; | |
6579 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6580 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6581 | event->attr.freq = 0; |
6582 | } | |
6583 | } | |
6584 | ||
b0a873eb PZ |
6585 | /* |
6586 | * Software event: cpu wall time clock | |
6587 | */ | |
6588 | ||
6589 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6590 | { |
b0a873eb PZ |
6591 | s64 prev; |
6592 | u64 now; | |
6593 | ||
a4eaf7f1 | 6594 | now = local_clock(); |
b0a873eb PZ |
6595 | prev = local64_xchg(&event->hw.prev_count, now); |
6596 | local64_add(now - prev, &event->count); | |
24f1e32c | 6597 | } |
24f1e32c | 6598 | |
a4eaf7f1 | 6599 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6600 | { |
a4eaf7f1 | 6601 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6602 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6603 | } |
6604 | ||
a4eaf7f1 | 6605 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6606 | { |
b0a873eb PZ |
6607 | perf_swevent_cancel_hrtimer(event); |
6608 | cpu_clock_event_update(event); | |
6609 | } | |
f29ac756 | 6610 | |
a4eaf7f1 PZ |
6611 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6612 | { | |
6613 | if (flags & PERF_EF_START) | |
6614 | cpu_clock_event_start(event, flags); | |
6a694a60 | 6615 | perf_event_update_userpage(event); |
a4eaf7f1 PZ |
6616 | |
6617 | return 0; | |
6618 | } | |
6619 | ||
6620 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6621 | { | |
6622 | cpu_clock_event_stop(event, flags); | |
6623 | } | |
6624 | ||
b0a873eb PZ |
6625 | static void cpu_clock_event_read(struct perf_event *event) |
6626 | { | |
6627 | cpu_clock_event_update(event); | |
6628 | } | |
f344011c | 6629 | |
b0a873eb PZ |
6630 | static int cpu_clock_event_init(struct perf_event *event) |
6631 | { | |
6632 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6633 | return -ENOENT; | |
6634 | ||
6635 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6636 | return -ENOENT; | |
6637 | ||
2481c5fa SE |
6638 | /* |
6639 | * no branch sampling for software events | |
6640 | */ | |
6641 | if (has_branch_stack(event)) | |
6642 | return -EOPNOTSUPP; | |
6643 | ||
ba3dd36c PZ |
6644 | perf_swevent_init_hrtimer(event); |
6645 | ||
b0a873eb | 6646 | return 0; |
f29ac756 PZ |
6647 | } |
6648 | ||
b0a873eb | 6649 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6650 | .task_ctx_nr = perf_sw_context, |
6651 | ||
b0a873eb | 6652 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6653 | .add = cpu_clock_event_add, |
6654 | .del = cpu_clock_event_del, | |
6655 | .start = cpu_clock_event_start, | |
6656 | .stop = cpu_clock_event_stop, | |
b0a873eb PZ |
6657 | .read = cpu_clock_event_read, |
6658 | }; | |
6659 | ||
6660 | /* | |
6661 | * Software event: task time clock | |
6662 | */ | |
6663 | ||
6664 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6665 | { |
b0a873eb PZ |
6666 | u64 prev; |
6667 | s64 delta; | |
5c92d124 | 6668 | |
b0a873eb PZ |
6669 | prev = local64_xchg(&event->hw.prev_count, now); |
6670 | delta = now - prev; | |
6671 | local64_add(delta, &event->count); | |
6672 | } | |
5c92d124 | 6673 | |
a4eaf7f1 | 6674 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6675 | { |
a4eaf7f1 | 6676 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6677 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6678 | } |
6679 | ||
a4eaf7f1 | 6680 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6681 | { |
6682 | perf_swevent_cancel_hrtimer(event); | |
6683 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6684 | } |
6685 | ||
6686 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6687 | { | |
6688 | if (flags & PERF_EF_START) | |
6689 | task_clock_event_start(event, flags); | |
6a694a60 | 6690 | perf_event_update_userpage(event); |
b0a873eb | 6691 | |
a4eaf7f1 PZ |
6692 | return 0; |
6693 | } | |
6694 | ||
6695 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6696 | { | |
6697 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6698 | } |
6699 | ||
6700 | static void task_clock_event_read(struct perf_event *event) | |
6701 | { | |
768a06e2 PZ |
6702 | u64 now = perf_clock(); |
6703 | u64 delta = now - event->ctx->timestamp; | |
6704 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6705 | |
6706 | task_clock_event_update(event, time); | |
6707 | } | |
6708 | ||
6709 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6710 | { |
b0a873eb PZ |
6711 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6712 | return -ENOENT; | |
6713 | ||
6714 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6715 | return -ENOENT; | |
6716 | ||
2481c5fa SE |
6717 | /* |
6718 | * no branch sampling for software events | |
6719 | */ | |
6720 | if (has_branch_stack(event)) | |
6721 | return -EOPNOTSUPP; | |
6722 | ||
ba3dd36c PZ |
6723 | perf_swevent_init_hrtimer(event); |
6724 | ||
b0a873eb | 6725 | return 0; |
6fb2915d LZ |
6726 | } |
6727 | ||
b0a873eb | 6728 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6729 | .task_ctx_nr = perf_sw_context, |
6730 | ||
b0a873eb | 6731 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6732 | .add = task_clock_event_add, |
6733 | .del = task_clock_event_del, | |
6734 | .start = task_clock_event_start, | |
6735 | .stop = task_clock_event_stop, | |
b0a873eb PZ |
6736 | .read = task_clock_event_read, |
6737 | }; | |
6fb2915d | 6738 | |
ad5133b7 | 6739 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6740 | { |
e077df4f | 6741 | } |
6fb2915d | 6742 | |
ad5133b7 | 6743 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6744 | { |
ad5133b7 | 6745 | return 0; |
6fb2915d LZ |
6746 | } |
6747 | ||
ad5133b7 | 6748 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6749 | { |
ad5133b7 | 6750 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6751 | } |
6752 | ||
ad5133b7 PZ |
6753 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6754 | { | |
6755 | perf_pmu_enable(pmu); | |
6756 | return 0; | |
6757 | } | |
e077df4f | 6758 | |
ad5133b7 | 6759 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6760 | { |
ad5133b7 | 6761 | perf_pmu_enable(pmu); |
24f1e32c FW |
6762 | } |
6763 | ||
35edc2a5 PZ |
6764 | static int perf_event_idx_default(struct perf_event *event) |
6765 | { | |
c719f560 | 6766 | return 0; |
35edc2a5 PZ |
6767 | } |
6768 | ||
8dc85d54 PZ |
6769 | /* |
6770 | * Ensures all contexts with the same task_ctx_nr have the same | |
6771 | * pmu_cpu_context too. | |
6772 | */ | |
9e317041 | 6773 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 6774 | { |
8dc85d54 | 6775 | struct pmu *pmu; |
b326e956 | 6776 | |
8dc85d54 PZ |
6777 | if (ctxn < 0) |
6778 | return NULL; | |
24f1e32c | 6779 | |
8dc85d54 PZ |
6780 | list_for_each_entry(pmu, &pmus, entry) { |
6781 | if (pmu->task_ctx_nr == ctxn) | |
6782 | return pmu->pmu_cpu_context; | |
6783 | } | |
24f1e32c | 6784 | |
8dc85d54 | 6785 | return NULL; |
24f1e32c FW |
6786 | } |
6787 | ||
51676957 | 6788 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6789 | { |
51676957 PZ |
6790 | int cpu; |
6791 | ||
6792 | for_each_possible_cpu(cpu) { | |
6793 | struct perf_cpu_context *cpuctx; | |
6794 | ||
6795 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6796 | ||
3f1f3320 PZ |
6797 | if (cpuctx->unique_pmu == old_pmu) |
6798 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6799 | } |
6800 | } | |
6801 | ||
6802 | static void free_pmu_context(struct pmu *pmu) | |
6803 | { | |
6804 | struct pmu *i; | |
f5ffe02e | 6805 | |
8dc85d54 | 6806 | mutex_lock(&pmus_lock); |
0475f9ea | 6807 | /* |
8dc85d54 | 6808 | * Like a real lame refcount. |
0475f9ea | 6809 | */ |
51676957 PZ |
6810 | list_for_each_entry(i, &pmus, entry) { |
6811 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6812 | update_pmu_context(i, pmu); | |
8dc85d54 | 6813 | goto out; |
51676957 | 6814 | } |
8dc85d54 | 6815 | } |
d6d020e9 | 6816 | |
51676957 | 6817 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6818 | out: |
6819 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6820 | } |
2e80a82a | 6821 | static struct idr pmu_idr; |
d6d020e9 | 6822 | |
abe43400 PZ |
6823 | static ssize_t |
6824 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6825 | { | |
6826 | struct pmu *pmu = dev_get_drvdata(dev); | |
6827 | ||
6828 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6829 | } | |
90826ca7 | 6830 | static DEVICE_ATTR_RO(type); |
abe43400 | 6831 | |
62b85639 SE |
6832 | static ssize_t |
6833 | perf_event_mux_interval_ms_show(struct device *dev, | |
6834 | struct device_attribute *attr, | |
6835 | char *page) | |
6836 | { | |
6837 | struct pmu *pmu = dev_get_drvdata(dev); | |
6838 | ||
6839 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6840 | } | |
6841 | ||
6842 | static ssize_t | |
6843 | perf_event_mux_interval_ms_store(struct device *dev, | |
6844 | struct device_attribute *attr, | |
6845 | const char *buf, size_t count) | |
6846 | { | |
6847 | struct pmu *pmu = dev_get_drvdata(dev); | |
6848 | int timer, cpu, ret; | |
6849 | ||
6850 | ret = kstrtoint(buf, 0, &timer); | |
6851 | if (ret) | |
6852 | return ret; | |
6853 | ||
6854 | if (timer < 1) | |
6855 | return -EINVAL; | |
6856 | ||
6857 | /* same value, noting to do */ | |
6858 | if (timer == pmu->hrtimer_interval_ms) | |
6859 | return count; | |
6860 | ||
6861 | pmu->hrtimer_interval_ms = timer; | |
6862 | ||
6863 | /* update all cpuctx for this PMU */ | |
6864 | for_each_possible_cpu(cpu) { | |
6865 | struct perf_cpu_context *cpuctx; | |
6866 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6867 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6868 | ||
6869 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6870 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6871 | } | |
6872 | ||
6873 | return count; | |
6874 | } | |
90826ca7 | 6875 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6876 | |
90826ca7 GKH |
6877 | static struct attribute *pmu_dev_attrs[] = { |
6878 | &dev_attr_type.attr, | |
6879 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6880 | NULL, | |
abe43400 | 6881 | }; |
90826ca7 | 6882 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6883 | |
6884 | static int pmu_bus_running; | |
6885 | static struct bus_type pmu_bus = { | |
6886 | .name = "event_source", | |
90826ca7 | 6887 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6888 | }; |
6889 | ||
6890 | static void pmu_dev_release(struct device *dev) | |
6891 | { | |
6892 | kfree(dev); | |
6893 | } | |
6894 | ||
6895 | static int pmu_dev_alloc(struct pmu *pmu) | |
6896 | { | |
6897 | int ret = -ENOMEM; | |
6898 | ||
6899 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6900 | if (!pmu->dev) | |
6901 | goto out; | |
6902 | ||
0c9d42ed | 6903 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6904 | device_initialize(pmu->dev); |
6905 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6906 | if (ret) | |
6907 | goto free_dev; | |
6908 | ||
6909 | dev_set_drvdata(pmu->dev, pmu); | |
6910 | pmu->dev->bus = &pmu_bus; | |
6911 | pmu->dev->release = pmu_dev_release; | |
6912 | ret = device_add(pmu->dev); | |
6913 | if (ret) | |
6914 | goto free_dev; | |
6915 | ||
6916 | out: | |
6917 | return ret; | |
6918 | ||
6919 | free_dev: | |
6920 | put_device(pmu->dev); | |
6921 | goto out; | |
6922 | } | |
6923 | ||
547e9fd7 | 6924 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6925 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6926 | |
03d8e80b | 6927 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6928 | { |
108b02cf | 6929 | int cpu, ret; |
24f1e32c | 6930 | |
b0a873eb | 6931 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6932 | ret = -ENOMEM; |
6933 | pmu->pmu_disable_count = alloc_percpu(int); | |
6934 | if (!pmu->pmu_disable_count) | |
6935 | goto unlock; | |
f29ac756 | 6936 | |
2e80a82a PZ |
6937 | pmu->type = -1; |
6938 | if (!name) | |
6939 | goto skip_type; | |
6940 | pmu->name = name; | |
6941 | ||
6942 | if (type < 0) { | |
0e9c3be2 TH |
6943 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6944 | if (type < 0) { | |
6945 | ret = type; | |
2e80a82a PZ |
6946 | goto free_pdc; |
6947 | } | |
6948 | } | |
6949 | pmu->type = type; | |
6950 | ||
abe43400 PZ |
6951 | if (pmu_bus_running) { |
6952 | ret = pmu_dev_alloc(pmu); | |
6953 | if (ret) | |
6954 | goto free_idr; | |
6955 | } | |
6956 | ||
2e80a82a | 6957 | skip_type: |
8dc85d54 PZ |
6958 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6959 | if (pmu->pmu_cpu_context) | |
6960 | goto got_cpu_context; | |
f29ac756 | 6961 | |
c4814202 | 6962 | ret = -ENOMEM; |
108b02cf PZ |
6963 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6964 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6965 | goto free_dev; |
f344011c | 6966 | |
108b02cf PZ |
6967 | for_each_possible_cpu(cpu) { |
6968 | struct perf_cpu_context *cpuctx; | |
6969 | ||
6970 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6971 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6972 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6973 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
108b02cf | 6974 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6975 | |
6976 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6977 | ||
3f1f3320 | 6978 | cpuctx->unique_pmu = pmu; |
108b02cf | 6979 | } |
76e1d904 | 6980 | |
8dc85d54 | 6981 | got_cpu_context: |
ad5133b7 PZ |
6982 | if (!pmu->start_txn) { |
6983 | if (pmu->pmu_enable) { | |
6984 | /* | |
6985 | * If we have pmu_enable/pmu_disable calls, install | |
6986 | * transaction stubs that use that to try and batch | |
6987 | * hardware accesses. | |
6988 | */ | |
6989 | pmu->start_txn = perf_pmu_start_txn; | |
6990 | pmu->commit_txn = perf_pmu_commit_txn; | |
6991 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6992 | } else { | |
6993 | pmu->start_txn = perf_pmu_nop_void; | |
6994 | pmu->commit_txn = perf_pmu_nop_int; | |
6995 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6996 | } |
5c92d124 | 6997 | } |
15dbf27c | 6998 | |
ad5133b7 PZ |
6999 | if (!pmu->pmu_enable) { |
7000 | pmu->pmu_enable = perf_pmu_nop_void; | |
7001 | pmu->pmu_disable = perf_pmu_nop_void; | |
7002 | } | |
7003 | ||
35edc2a5 PZ |
7004 | if (!pmu->event_idx) |
7005 | pmu->event_idx = perf_event_idx_default; | |
7006 | ||
b0a873eb | 7007 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
7008 | ret = 0; |
7009 | unlock: | |
b0a873eb PZ |
7010 | mutex_unlock(&pmus_lock); |
7011 | ||
33696fc0 | 7012 | return ret; |
108b02cf | 7013 | |
abe43400 PZ |
7014 | free_dev: |
7015 | device_del(pmu->dev); | |
7016 | put_device(pmu->dev); | |
7017 | ||
2e80a82a PZ |
7018 | free_idr: |
7019 | if (pmu->type >= PERF_TYPE_MAX) | |
7020 | idr_remove(&pmu_idr, pmu->type); | |
7021 | ||
108b02cf PZ |
7022 | free_pdc: |
7023 | free_percpu(pmu->pmu_disable_count); | |
7024 | goto unlock; | |
f29ac756 | 7025 | } |
c464c76e | 7026 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 7027 | |
b0a873eb | 7028 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 7029 | { |
b0a873eb PZ |
7030 | mutex_lock(&pmus_lock); |
7031 | list_del_rcu(&pmu->entry); | |
7032 | mutex_unlock(&pmus_lock); | |
5c92d124 | 7033 | |
0475f9ea | 7034 | /* |
cde8e884 PZ |
7035 | * We dereference the pmu list under both SRCU and regular RCU, so |
7036 | * synchronize against both of those. | |
0475f9ea | 7037 | */ |
b0a873eb | 7038 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 7039 | synchronize_rcu(); |
d6d020e9 | 7040 | |
33696fc0 | 7041 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
7042 | if (pmu->type >= PERF_TYPE_MAX) |
7043 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
7044 | device_del(pmu->dev); |
7045 | put_device(pmu->dev); | |
51676957 | 7046 | free_pmu_context(pmu); |
b0a873eb | 7047 | } |
c464c76e | 7048 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 7049 | |
cc34b98b MR |
7050 | static int perf_try_init_event(struct pmu *pmu, struct perf_event *event) |
7051 | { | |
7052 | int ret; | |
7053 | ||
7054 | if (!try_module_get(pmu->module)) | |
7055 | return -ENODEV; | |
7056 | event->pmu = pmu; | |
7057 | ret = pmu->event_init(event); | |
7058 | if (ret) | |
7059 | module_put(pmu->module); | |
7060 | ||
7061 | return ret; | |
7062 | } | |
7063 | ||
b0a873eb PZ |
7064 | struct pmu *perf_init_event(struct perf_event *event) |
7065 | { | |
7066 | struct pmu *pmu = NULL; | |
7067 | int idx; | |
940c5b29 | 7068 | int ret; |
b0a873eb PZ |
7069 | |
7070 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
7071 | |
7072 | rcu_read_lock(); | |
7073 | pmu = idr_find(&pmu_idr, event->attr.type); | |
7074 | rcu_read_unlock(); | |
940c5b29 | 7075 | if (pmu) { |
cc34b98b | 7076 | ret = perf_try_init_event(pmu, event); |
940c5b29 LM |
7077 | if (ret) |
7078 | pmu = ERR_PTR(ret); | |
2e80a82a | 7079 | goto unlock; |
940c5b29 | 7080 | } |
2e80a82a | 7081 | |
b0a873eb | 7082 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
cc34b98b | 7083 | ret = perf_try_init_event(pmu, event); |
b0a873eb | 7084 | if (!ret) |
e5f4d339 | 7085 | goto unlock; |
76e1d904 | 7086 | |
b0a873eb PZ |
7087 | if (ret != -ENOENT) { |
7088 | pmu = ERR_PTR(ret); | |
e5f4d339 | 7089 | goto unlock; |
f344011c | 7090 | } |
5c92d124 | 7091 | } |
e5f4d339 PZ |
7092 | pmu = ERR_PTR(-ENOENT); |
7093 | unlock: | |
b0a873eb | 7094 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 7095 | |
4aeb0b42 | 7096 | return pmu; |
5c92d124 IM |
7097 | } |
7098 | ||
4beb31f3 FW |
7099 | static void account_event_cpu(struct perf_event *event, int cpu) |
7100 | { | |
7101 | if (event->parent) | |
7102 | return; | |
7103 | ||
4beb31f3 FW |
7104 | if (is_cgroup_event(event)) |
7105 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
7106 | } | |
7107 | ||
766d6c07 FW |
7108 | static void account_event(struct perf_event *event) |
7109 | { | |
4beb31f3 FW |
7110 | if (event->parent) |
7111 | return; | |
7112 | ||
766d6c07 FW |
7113 | if (event->attach_state & PERF_ATTACH_TASK) |
7114 | static_key_slow_inc(&perf_sched_events.key); | |
7115 | if (event->attr.mmap || event->attr.mmap_data) | |
7116 | atomic_inc(&nr_mmap_events); | |
7117 | if (event->attr.comm) | |
7118 | atomic_inc(&nr_comm_events); | |
7119 | if (event->attr.task) | |
7120 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
7121 | if (event->attr.freq) { |
7122 | if (atomic_inc_return(&nr_freq_events) == 1) | |
7123 | tick_nohz_full_kick_all(); | |
7124 | } | |
4beb31f3 | 7125 | if (has_branch_stack(event)) |
766d6c07 | 7126 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 7127 | if (is_cgroup_event(event)) |
766d6c07 | 7128 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
7129 | |
7130 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
7131 | } |
7132 | ||
0793a61d | 7133 | /* |
cdd6c482 | 7134 | * Allocate and initialize a event structure |
0793a61d | 7135 | */ |
cdd6c482 | 7136 | static struct perf_event * |
c3f00c70 | 7137 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
7138 | struct task_struct *task, |
7139 | struct perf_event *group_leader, | |
7140 | struct perf_event *parent_event, | |
4dc0da86 AK |
7141 | perf_overflow_handler_t overflow_handler, |
7142 | void *context) | |
0793a61d | 7143 | { |
51b0fe39 | 7144 | struct pmu *pmu; |
cdd6c482 IM |
7145 | struct perf_event *event; |
7146 | struct hw_perf_event *hwc; | |
90983b16 | 7147 | long err = -EINVAL; |
0793a61d | 7148 | |
66832eb4 ON |
7149 | if ((unsigned)cpu >= nr_cpu_ids) { |
7150 | if (!task || cpu != -1) | |
7151 | return ERR_PTR(-EINVAL); | |
7152 | } | |
7153 | ||
c3f00c70 | 7154 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 7155 | if (!event) |
d5d2bc0d | 7156 | return ERR_PTR(-ENOMEM); |
0793a61d | 7157 | |
04289bb9 | 7158 | /* |
cdd6c482 | 7159 | * Single events are their own group leaders, with an |
04289bb9 IM |
7160 | * empty sibling list: |
7161 | */ | |
7162 | if (!group_leader) | |
cdd6c482 | 7163 | group_leader = event; |
04289bb9 | 7164 | |
cdd6c482 IM |
7165 | mutex_init(&event->child_mutex); |
7166 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 7167 | |
cdd6c482 IM |
7168 | INIT_LIST_HEAD(&event->group_entry); |
7169 | INIT_LIST_HEAD(&event->event_entry); | |
7170 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 7171 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 7172 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
7173 | INIT_HLIST_NODE(&event->hlist_entry); |
7174 | ||
10c6db11 | 7175 | |
cdd6c482 | 7176 | init_waitqueue_head(&event->waitq); |
e360adbe | 7177 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 7178 | |
cdd6c482 | 7179 | mutex_init(&event->mmap_mutex); |
7b732a75 | 7180 | |
a6fa941d | 7181 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
7182 | event->cpu = cpu; |
7183 | event->attr = *attr; | |
7184 | event->group_leader = group_leader; | |
7185 | event->pmu = NULL; | |
cdd6c482 | 7186 | event->oncpu = -1; |
a96bbc16 | 7187 | |
cdd6c482 | 7188 | event->parent = parent_event; |
b84fbc9f | 7189 | |
17cf22c3 | 7190 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 7191 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 7192 | |
cdd6c482 | 7193 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 7194 | |
d580ff86 PZ |
7195 | if (task) { |
7196 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
7197 | |
7198 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
7199 | event->hw.tp_target = task; | |
d580ff86 PZ |
7200 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
7201 | /* | |
7202 | * hw_breakpoint is a bit difficult here.. | |
7203 | */ | |
f22c1bb6 | 7204 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
7205 | event->hw.bp_target = task; |
7206 | #endif | |
7207 | } | |
7208 | ||
4dc0da86 | 7209 | if (!overflow_handler && parent_event) { |
b326e956 | 7210 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
7211 | context = parent_event->overflow_handler_context; |
7212 | } | |
66832eb4 | 7213 | |
b326e956 | 7214 | event->overflow_handler = overflow_handler; |
4dc0da86 | 7215 | event->overflow_handler_context = context; |
97eaf530 | 7216 | |
0231bb53 | 7217 | perf_event__state_init(event); |
a86ed508 | 7218 | |
4aeb0b42 | 7219 | pmu = NULL; |
b8e83514 | 7220 | |
cdd6c482 | 7221 | hwc = &event->hw; |
bd2b5b12 | 7222 | hwc->sample_period = attr->sample_period; |
0d48696f | 7223 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 7224 | hwc->sample_period = 1; |
eced1dfc | 7225 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 7226 | |
e7850595 | 7227 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 7228 | |
2023b359 | 7229 | /* |
cdd6c482 | 7230 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 7231 | */ |
3dab77fb | 7232 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 7233 | goto err_ns; |
2023b359 | 7234 | |
b0a873eb | 7235 | pmu = perf_init_event(event); |
4aeb0b42 | 7236 | if (!pmu) |
90983b16 FW |
7237 | goto err_ns; |
7238 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 7239 | err = PTR_ERR(pmu); |
90983b16 | 7240 | goto err_ns; |
621a01ea | 7241 | } |
d5d2bc0d | 7242 | |
cdd6c482 | 7243 | if (!event->parent) { |
927c7a9e FW |
7244 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
7245 | err = get_callchain_buffers(); | |
90983b16 FW |
7246 | if (err) |
7247 | goto err_pmu; | |
d010b332 | 7248 | } |
f344011c | 7249 | } |
9ee318a7 | 7250 | |
cdd6c482 | 7251 | return event; |
90983b16 FW |
7252 | |
7253 | err_pmu: | |
7254 | if (event->destroy) | |
7255 | event->destroy(event); | |
c464c76e | 7256 | module_put(pmu->module); |
90983b16 FW |
7257 | err_ns: |
7258 | if (event->ns) | |
7259 | put_pid_ns(event->ns); | |
7260 | kfree(event); | |
7261 | ||
7262 | return ERR_PTR(err); | |
0793a61d TG |
7263 | } |
7264 | ||
cdd6c482 IM |
7265 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
7266 | struct perf_event_attr *attr) | |
974802ea | 7267 | { |
974802ea | 7268 | u32 size; |
cdf8073d | 7269 | int ret; |
974802ea PZ |
7270 | |
7271 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
7272 | return -EFAULT; | |
7273 | ||
7274 | /* | |
7275 | * zero the full structure, so that a short copy will be nice. | |
7276 | */ | |
7277 | memset(attr, 0, sizeof(*attr)); | |
7278 | ||
7279 | ret = get_user(size, &uattr->size); | |
7280 | if (ret) | |
7281 | return ret; | |
7282 | ||
7283 | if (size > PAGE_SIZE) /* silly large */ | |
7284 | goto err_size; | |
7285 | ||
7286 | if (!size) /* abi compat */ | |
7287 | size = PERF_ATTR_SIZE_VER0; | |
7288 | ||
7289 | if (size < PERF_ATTR_SIZE_VER0) | |
7290 | goto err_size; | |
7291 | ||
7292 | /* | |
7293 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
7294 | * ensure all the unknown bits are 0 - i.e. new |
7295 | * user-space does not rely on any kernel feature | |
7296 | * extensions we dont know about yet. | |
974802ea PZ |
7297 | */ |
7298 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
7299 | unsigned char __user *addr; |
7300 | unsigned char __user *end; | |
7301 | unsigned char val; | |
974802ea | 7302 | |
cdf8073d IS |
7303 | addr = (void __user *)uattr + sizeof(*attr); |
7304 | end = (void __user *)uattr + size; | |
974802ea | 7305 | |
cdf8073d | 7306 | for (; addr < end; addr++) { |
974802ea PZ |
7307 | ret = get_user(val, addr); |
7308 | if (ret) | |
7309 | return ret; | |
7310 | if (val) | |
7311 | goto err_size; | |
7312 | } | |
b3e62e35 | 7313 | size = sizeof(*attr); |
974802ea PZ |
7314 | } |
7315 | ||
7316 | ret = copy_from_user(attr, uattr, size); | |
7317 | if (ret) | |
7318 | return -EFAULT; | |
7319 | ||
cd757645 | 7320 | if (attr->__reserved_1) |
974802ea PZ |
7321 | return -EINVAL; |
7322 | ||
7323 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
7324 | return -EINVAL; | |
7325 | ||
7326 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
7327 | return -EINVAL; | |
7328 | ||
bce38cd5 SE |
7329 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
7330 | u64 mask = attr->branch_sample_type; | |
7331 | ||
7332 | /* only using defined bits */ | |
7333 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
7334 | return -EINVAL; | |
7335 | ||
7336 | /* at least one branch bit must be set */ | |
7337 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
7338 | return -EINVAL; | |
7339 | ||
bce38cd5 SE |
7340 | /* propagate priv level, when not set for branch */ |
7341 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
7342 | ||
7343 | /* exclude_kernel checked on syscall entry */ | |
7344 | if (!attr->exclude_kernel) | |
7345 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
7346 | ||
7347 | if (!attr->exclude_user) | |
7348 | mask |= PERF_SAMPLE_BRANCH_USER; | |
7349 | ||
7350 | if (!attr->exclude_hv) | |
7351 | mask |= PERF_SAMPLE_BRANCH_HV; | |
7352 | /* | |
7353 | * adjust user setting (for HW filter setup) | |
7354 | */ | |
7355 | attr->branch_sample_type = mask; | |
7356 | } | |
e712209a SE |
7357 | /* privileged levels capture (kernel, hv): check permissions */ |
7358 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
7359 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
7360 | return -EACCES; | |
bce38cd5 | 7361 | } |
4018994f | 7362 | |
c5ebcedb | 7363 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 7364 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
7365 | if (ret) |
7366 | return ret; | |
7367 | } | |
7368 | ||
7369 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
7370 | if (!arch_perf_have_user_stack_dump()) | |
7371 | return -ENOSYS; | |
7372 | ||
7373 | /* | |
7374 | * We have __u32 type for the size, but so far | |
7375 | * we can only use __u16 as maximum due to the | |
7376 | * __u16 sample size limit. | |
7377 | */ | |
7378 | if (attr->sample_stack_user >= USHRT_MAX) | |
7379 | ret = -EINVAL; | |
7380 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
7381 | ret = -EINVAL; | |
7382 | } | |
4018994f | 7383 | |
60e2364e SE |
7384 | if (attr->sample_type & PERF_SAMPLE_REGS_INTR) |
7385 | ret = perf_reg_validate(attr->sample_regs_intr); | |
974802ea PZ |
7386 | out: |
7387 | return ret; | |
7388 | ||
7389 | err_size: | |
7390 | put_user(sizeof(*attr), &uattr->size); | |
7391 | ret = -E2BIG; | |
7392 | goto out; | |
7393 | } | |
7394 | ||
ac9721f3 PZ |
7395 | static int |
7396 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 7397 | { |
b69cf536 | 7398 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
7399 | int ret = -EINVAL; |
7400 | ||
ac9721f3 | 7401 | if (!output_event) |
a4be7c27 PZ |
7402 | goto set; |
7403 | ||
ac9721f3 PZ |
7404 | /* don't allow circular references */ |
7405 | if (event == output_event) | |
a4be7c27 PZ |
7406 | goto out; |
7407 | ||
0f139300 PZ |
7408 | /* |
7409 | * Don't allow cross-cpu buffers | |
7410 | */ | |
7411 | if (output_event->cpu != event->cpu) | |
7412 | goto out; | |
7413 | ||
7414 | /* | |
76369139 | 7415 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
7416 | */ |
7417 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
7418 | goto out; | |
7419 | ||
a4be7c27 | 7420 | set: |
cdd6c482 | 7421 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7422 | /* Can't redirect output if we've got an active mmap() */ |
7423 | if (atomic_read(&event->mmap_count)) | |
7424 | goto unlock; | |
a4be7c27 | 7425 | |
ac9721f3 | 7426 | if (output_event) { |
76369139 FW |
7427 | /* get the rb we want to redirect to */ |
7428 | rb = ring_buffer_get(output_event); | |
7429 | if (!rb) | |
ac9721f3 | 7430 | goto unlock; |
a4be7c27 PZ |
7431 | } |
7432 | ||
b69cf536 | 7433 | ring_buffer_attach(event, rb); |
9bb5d40c | 7434 | |
a4be7c27 | 7435 | ret = 0; |
ac9721f3 PZ |
7436 | unlock: |
7437 | mutex_unlock(&event->mmap_mutex); | |
7438 | ||
a4be7c27 | 7439 | out: |
a4be7c27 PZ |
7440 | return ret; |
7441 | } | |
7442 | ||
f63a8daa PZ |
7443 | static void mutex_lock_double(struct mutex *a, struct mutex *b) |
7444 | { | |
7445 | if (b < a) | |
7446 | swap(a, b); | |
7447 | ||
7448 | mutex_lock(a); | |
7449 | mutex_lock_nested(b, SINGLE_DEPTH_NESTING); | |
7450 | } | |
7451 | ||
0793a61d | 7452 | /** |
cdd6c482 | 7453 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7454 | * |
cdd6c482 | 7455 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7456 | * @pid: target pid |
9f66a381 | 7457 | * @cpu: target cpu |
cdd6c482 | 7458 | * @group_fd: group leader event fd |
0793a61d | 7459 | */ |
cdd6c482 IM |
7460 | SYSCALL_DEFINE5(perf_event_open, |
7461 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7462 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7463 | { |
b04243ef PZ |
7464 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7465 | struct perf_event *event, *sibling; | |
cdd6c482 | 7466 | struct perf_event_attr attr; |
f63a8daa | 7467 | struct perf_event_context *ctx, *uninitialized_var(gctx); |
cdd6c482 | 7468 | struct file *event_file = NULL; |
2903ff01 | 7469 | struct fd group = {NULL, 0}; |
38a81da2 | 7470 | struct task_struct *task = NULL; |
89a1e187 | 7471 | struct pmu *pmu; |
ea635c64 | 7472 | int event_fd; |
b04243ef | 7473 | int move_group = 0; |
dc86cabe | 7474 | int err; |
a21b0b35 | 7475 | int f_flags = O_RDWR; |
0793a61d | 7476 | |
2743a5b0 | 7477 | /* for future expandability... */ |
e5d1367f | 7478 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7479 | return -EINVAL; |
7480 | ||
dc86cabe IM |
7481 | err = perf_copy_attr(attr_uptr, &attr); |
7482 | if (err) | |
7483 | return err; | |
eab656ae | 7484 | |
0764771d PZ |
7485 | if (!attr.exclude_kernel) { |
7486 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7487 | return -EACCES; | |
7488 | } | |
7489 | ||
df58ab24 | 7490 | if (attr.freq) { |
cdd6c482 | 7491 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 7492 | return -EINVAL; |
0819b2e3 PZ |
7493 | } else { |
7494 | if (attr.sample_period & (1ULL << 63)) | |
7495 | return -EINVAL; | |
df58ab24 PZ |
7496 | } |
7497 | ||
e5d1367f SE |
7498 | /* |
7499 | * In cgroup mode, the pid argument is used to pass the fd | |
7500 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7501 | * designates the cpu on which to monitor threads from that | |
7502 | * cgroup. | |
7503 | */ | |
7504 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7505 | return -EINVAL; | |
7506 | ||
a21b0b35 YD |
7507 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7508 | f_flags |= O_CLOEXEC; | |
7509 | ||
7510 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7511 | if (event_fd < 0) |
7512 | return event_fd; | |
7513 | ||
ac9721f3 | 7514 | if (group_fd != -1) { |
2903ff01 AV |
7515 | err = perf_fget_light(group_fd, &group); |
7516 | if (err) | |
d14b12d7 | 7517 | goto err_fd; |
2903ff01 | 7518 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7519 | if (flags & PERF_FLAG_FD_OUTPUT) |
7520 | output_event = group_leader; | |
7521 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7522 | group_leader = NULL; | |
7523 | } | |
7524 | ||
e5d1367f | 7525 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7526 | task = find_lively_task_by_vpid(pid); |
7527 | if (IS_ERR(task)) { | |
7528 | err = PTR_ERR(task); | |
7529 | goto err_group_fd; | |
7530 | } | |
7531 | } | |
7532 | ||
1f4ee503 PZ |
7533 | if (task && group_leader && |
7534 | group_leader->attr.inherit != attr.inherit) { | |
7535 | err = -EINVAL; | |
7536 | goto err_task; | |
7537 | } | |
7538 | ||
fbfc623f YZ |
7539 | get_online_cpus(); |
7540 | ||
4dc0da86 AK |
7541 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
7542 | NULL, NULL); | |
d14b12d7 SE |
7543 | if (IS_ERR(event)) { |
7544 | err = PTR_ERR(event); | |
1f4ee503 | 7545 | goto err_cpus; |
d14b12d7 SE |
7546 | } |
7547 | ||
e5d1367f SE |
7548 | if (flags & PERF_FLAG_PID_CGROUP) { |
7549 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
7550 | if (err) { |
7551 | __free_event(event); | |
1f4ee503 | 7552 | goto err_cpus; |
766d6c07 | 7553 | } |
e5d1367f SE |
7554 | } |
7555 | ||
53b25335 VW |
7556 | if (is_sampling_event(event)) { |
7557 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
7558 | err = -ENOTSUPP; | |
7559 | goto err_alloc; | |
7560 | } | |
7561 | } | |
7562 | ||
766d6c07 FW |
7563 | account_event(event); |
7564 | ||
89a1e187 PZ |
7565 | /* |
7566 | * Special case software events and allow them to be part of | |
7567 | * any hardware group. | |
7568 | */ | |
7569 | pmu = event->pmu; | |
b04243ef PZ |
7570 | |
7571 | if (group_leader && | |
7572 | (is_software_event(event) != is_software_event(group_leader))) { | |
7573 | if (is_software_event(event)) { | |
7574 | /* | |
7575 | * If event and group_leader are not both a software | |
7576 | * event, and event is, then group leader is not. | |
7577 | * | |
7578 | * Allow the addition of software events to !software | |
7579 | * groups, this is safe because software events never | |
7580 | * fail to schedule. | |
7581 | */ | |
7582 | pmu = group_leader->pmu; | |
7583 | } else if (is_software_event(group_leader) && | |
7584 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7585 | /* | |
7586 | * In case the group is a pure software group, and we | |
7587 | * try to add a hardware event, move the whole group to | |
7588 | * the hardware context. | |
7589 | */ | |
7590 | move_group = 1; | |
7591 | } | |
7592 | } | |
89a1e187 PZ |
7593 | |
7594 | /* | |
7595 | * Get the target context (task or percpu): | |
7596 | */ | |
4af57ef2 | 7597 | ctx = find_get_context(pmu, task, event); |
89a1e187 PZ |
7598 | if (IS_ERR(ctx)) { |
7599 | err = PTR_ERR(ctx); | |
c6be5a5c | 7600 | goto err_alloc; |
89a1e187 PZ |
7601 | } |
7602 | ||
fd1edb3a PZ |
7603 | if (task) { |
7604 | put_task_struct(task); | |
7605 | task = NULL; | |
7606 | } | |
7607 | ||
ccff286d | 7608 | /* |
cdd6c482 | 7609 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7610 | */ |
ac9721f3 | 7611 | if (group_leader) { |
dc86cabe | 7612 | err = -EINVAL; |
04289bb9 | 7613 | |
04289bb9 | 7614 | /* |
ccff286d IM |
7615 | * Do not allow a recursive hierarchy (this new sibling |
7616 | * becoming part of another group-sibling): | |
7617 | */ | |
7618 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7619 | goto err_context; |
ccff286d IM |
7620 | /* |
7621 | * Do not allow to attach to a group in a different | |
7622 | * task or CPU context: | |
04289bb9 | 7623 | */ |
b04243ef | 7624 | if (move_group) { |
c3c87e77 PZ |
7625 | /* |
7626 | * Make sure we're both on the same task, or both | |
7627 | * per-cpu events. | |
7628 | */ | |
7629 | if (group_leader->ctx->task != ctx->task) | |
7630 | goto err_context; | |
7631 | ||
7632 | /* | |
7633 | * Make sure we're both events for the same CPU; | |
7634 | * grouping events for different CPUs is broken; since | |
7635 | * you can never concurrently schedule them anyhow. | |
7636 | */ | |
7637 | if (group_leader->cpu != event->cpu) | |
b04243ef PZ |
7638 | goto err_context; |
7639 | } else { | |
7640 | if (group_leader->ctx != ctx) | |
7641 | goto err_context; | |
7642 | } | |
7643 | ||
3b6f9e5c PM |
7644 | /* |
7645 | * Only a group leader can be exclusive or pinned | |
7646 | */ | |
0d48696f | 7647 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7648 | goto err_context; |
ac9721f3 PZ |
7649 | } |
7650 | ||
7651 | if (output_event) { | |
7652 | err = perf_event_set_output(event, output_event); | |
7653 | if (err) | |
c3f00c70 | 7654 | goto err_context; |
ac9721f3 | 7655 | } |
0793a61d | 7656 | |
a21b0b35 YD |
7657 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7658 | f_flags); | |
ea635c64 AV |
7659 | if (IS_ERR(event_file)) { |
7660 | err = PTR_ERR(event_file); | |
c3f00c70 | 7661 | goto err_context; |
ea635c64 | 7662 | } |
9b51f66d | 7663 | |
b04243ef | 7664 | if (move_group) { |
f63a8daa PZ |
7665 | gctx = group_leader->ctx; |
7666 | ||
7667 | /* | |
7668 | * See perf_event_ctx_lock() for comments on the details | |
7669 | * of swizzling perf_event::ctx. | |
7670 | */ | |
7671 | mutex_lock_double(&gctx->mutex, &ctx->mutex); | |
b04243ef | 7672 | |
46ce0fe9 | 7673 | perf_remove_from_context(group_leader, false); |
0231bb53 | 7674 | |
b04243ef PZ |
7675 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7676 | group_entry) { | |
46ce0fe9 | 7677 | perf_remove_from_context(sibling, false); |
b04243ef PZ |
7678 | put_ctx(gctx); |
7679 | } | |
f63a8daa PZ |
7680 | } else { |
7681 | mutex_lock(&ctx->mutex); | |
ea635c64 | 7682 | } |
9b51f66d | 7683 | |
ad3a37de | 7684 | WARN_ON_ONCE(ctx->parent_ctx); |
b04243ef PZ |
7685 | |
7686 | if (move_group) { | |
f63a8daa PZ |
7687 | /* |
7688 | * Wait for everybody to stop referencing the events through | |
7689 | * the old lists, before installing it on new lists. | |
7690 | */ | |
0cda4c02 | 7691 | synchronize_rcu(); |
f63a8daa | 7692 | |
8f95b435 PZI |
7693 | /* |
7694 | * Install the group siblings before the group leader. | |
7695 | * | |
7696 | * Because a group leader will try and install the entire group | |
7697 | * (through the sibling list, which is still in-tact), we can | |
7698 | * end up with siblings installed in the wrong context. | |
7699 | * | |
7700 | * By installing siblings first we NO-OP because they're not | |
7701 | * reachable through the group lists. | |
7702 | */ | |
b04243ef PZ |
7703 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7704 | group_entry) { | |
8f95b435 | 7705 | perf_event__state_init(sibling); |
9fc81d87 | 7706 | perf_install_in_context(ctx, sibling, sibling->cpu); |
b04243ef PZ |
7707 | get_ctx(ctx); |
7708 | } | |
8f95b435 PZI |
7709 | |
7710 | /* | |
7711 | * Removing from the context ends up with disabled | |
7712 | * event. What we want here is event in the initial | |
7713 | * startup state, ready to be add into new context. | |
7714 | */ | |
7715 | perf_event__state_init(group_leader); | |
7716 | perf_install_in_context(ctx, group_leader, group_leader->cpu); | |
7717 | get_ctx(ctx); | |
b04243ef PZ |
7718 | } |
7719 | ||
e2d37cd2 | 7720 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7721 | perf_unpin_context(ctx); |
f63a8daa PZ |
7722 | |
7723 | if (move_group) { | |
7724 | mutex_unlock(&gctx->mutex); | |
7725 | put_ctx(gctx); | |
7726 | } | |
d859e29f | 7727 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7728 | |
fbfc623f YZ |
7729 | put_online_cpus(); |
7730 | ||
cdd6c482 | 7731 | event->owner = current; |
8882135b | 7732 | |
cdd6c482 IM |
7733 | mutex_lock(¤t->perf_event_mutex); |
7734 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7735 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7736 | |
c320c7b7 ACM |
7737 | /* |
7738 | * Precalculate sample_data sizes | |
7739 | */ | |
7740 | perf_event__header_size(event); | |
6844c09d | 7741 | perf_event__id_header_size(event); |
c320c7b7 | 7742 | |
8a49542c PZ |
7743 | /* |
7744 | * Drop the reference on the group_event after placing the | |
7745 | * new event on the sibling_list. This ensures destruction | |
7746 | * of the group leader will find the pointer to itself in | |
7747 | * perf_group_detach(). | |
7748 | */ | |
2903ff01 | 7749 | fdput(group); |
ea635c64 AV |
7750 | fd_install(event_fd, event_file); |
7751 | return event_fd; | |
0793a61d | 7752 | |
c3f00c70 | 7753 | err_context: |
fe4b04fa | 7754 | perf_unpin_context(ctx); |
ea635c64 | 7755 | put_ctx(ctx); |
c6be5a5c | 7756 | err_alloc: |
ea635c64 | 7757 | free_event(event); |
1f4ee503 | 7758 | err_cpus: |
fbfc623f | 7759 | put_online_cpus(); |
1f4ee503 | 7760 | err_task: |
e7d0bc04 PZ |
7761 | if (task) |
7762 | put_task_struct(task); | |
89a1e187 | 7763 | err_group_fd: |
2903ff01 | 7764 | fdput(group); |
ea635c64 AV |
7765 | err_fd: |
7766 | put_unused_fd(event_fd); | |
dc86cabe | 7767 | return err; |
0793a61d TG |
7768 | } |
7769 | ||
fb0459d7 AV |
7770 | /** |
7771 | * perf_event_create_kernel_counter | |
7772 | * | |
7773 | * @attr: attributes of the counter to create | |
7774 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7775 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7776 | */ |
7777 | struct perf_event * | |
7778 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7779 | struct task_struct *task, |
4dc0da86 AK |
7780 | perf_overflow_handler_t overflow_handler, |
7781 | void *context) | |
fb0459d7 | 7782 | { |
fb0459d7 | 7783 | struct perf_event_context *ctx; |
c3f00c70 | 7784 | struct perf_event *event; |
fb0459d7 | 7785 | int err; |
d859e29f | 7786 | |
fb0459d7 AV |
7787 | /* |
7788 | * Get the target context (task or percpu): | |
7789 | */ | |
d859e29f | 7790 | |
4dc0da86 AK |
7791 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7792 | overflow_handler, context); | |
c3f00c70 PZ |
7793 | if (IS_ERR(event)) { |
7794 | err = PTR_ERR(event); | |
7795 | goto err; | |
7796 | } | |
d859e29f | 7797 | |
f8697762 JO |
7798 | /* Mark owner so we could distinguish it from user events. */ |
7799 | event->owner = EVENT_OWNER_KERNEL; | |
7800 | ||
766d6c07 FW |
7801 | account_event(event); |
7802 | ||
4af57ef2 | 7803 | ctx = find_get_context(event->pmu, task, event); |
c6567f64 FW |
7804 | if (IS_ERR(ctx)) { |
7805 | err = PTR_ERR(ctx); | |
c3f00c70 | 7806 | goto err_free; |
d859e29f | 7807 | } |
fb0459d7 | 7808 | |
fb0459d7 AV |
7809 | WARN_ON_ONCE(ctx->parent_ctx); |
7810 | mutex_lock(&ctx->mutex); | |
7811 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7812 | perf_unpin_context(ctx); |
fb0459d7 AV |
7813 | mutex_unlock(&ctx->mutex); |
7814 | ||
fb0459d7 AV |
7815 | return event; |
7816 | ||
c3f00c70 PZ |
7817 | err_free: |
7818 | free_event(event); | |
7819 | err: | |
c6567f64 | 7820 | return ERR_PTR(err); |
9b51f66d | 7821 | } |
fb0459d7 | 7822 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7823 | |
0cda4c02 YZ |
7824 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7825 | { | |
7826 | struct perf_event_context *src_ctx; | |
7827 | struct perf_event_context *dst_ctx; | |
7828 | struct perf_event *event, *tmp; | |
7829 | LIST_HEAD(events); | |
7830 | ||
7831 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7832 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7833 | ||
f63a8daa PZ |
7834 | /* |
7835 | * See perf_event_ctx_lock() for comments on the details | |
7836 | * of swizzling perf_event::ctx. | |
7837 | */ | |
7838 | mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); | |
0cda4c02 YZ |
7839 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, |
7840 | event_entry) { | |
46ce0fe9 | 7841 | perf_remove_from_context(event, false); |
9a545de0 | 7842 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7843 | put_ctx(src_ctx); |
9886167d | 7844 | list_add(&event->migrate_entry, &events); |
0cda4c02 | 7845 | } |
0cda4c02 | 7846 | |
8f95b435 PZI |
7847 | /* |
7848 | * Wait for the events to quiesce before re-instating them. | |
7849 | */ | |
0cda4c02 YZ |
7850 | synchronize_rcu(); |
7851 | ||
8f95b435 PZI |
7852 | /* |
7853 | * Re-instate events in 2 passes. | |
7854 | * | |
7855 | * Skip over group leaders and only install siblings on this first | |
7856 | * pass, siblings will not get enabled without a leader, however a | |
7857 | * leader will enable its siblings, even if those are still on the old | |
7858 | * context. | |
7859 | */ | |
7860 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { | |
7861 | if (event->group_leader == event) | |
7862 | continue; | |
7863 | ||
7864 | list_del(&event->migrate_entry); | |
7865 | if (event->state >= PERF_EVENT_STATE_OFF) | |
7866 | event->state = PERF_EVENT_STATE_INACTIVE; | |
7867 | account_event_cpu(event, dst_cpu); | |
7868 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
7869 | get_ctx(dst_ctx); | |
7870 | } | |
7871 | ||
7872 | /* | |
7873 | * Once all the siblings are setup properly, install the group leaders | |
7874 | * to make it go. | |
7875 | */ | |
9886167d PZ |
7876 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7877 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7878 | if (event->state >= PERF_EVENT_STATE_OFF) |
7879 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7880 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7881 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7882 | get_ctx(dst_ctx); | |
7883 | } | |
7884 | mutex_unlock(&dst_ctx->mutex); | |
f63a8daa | 7885 | mutex_unlock(&src_ctx->mutex); |
0cda4c02 YZ |
7886 | } |
7887 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7888 | ||
cdd6c482 | 7889 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7890 | struct task_struct *child) |
d859e29f | 7891 | { |
cdd6c482 | 7892 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7893 | u64 child_val; |
d859e29f | 7894 | |
cdd6c482 IM |
7895 | if (child_event->attr.inherit_stat) |
7896 | perf_event_read_event(child_event, child); | |
38b200d6 | 7897 | |
b5e58793 | 7898 | child_val = perf_event_count(child_event); |
d859e29f PM |
7899 | |
7900 | /* | |
7901 | * Add back the child's count to the parent's count: | |
7902 | */ | |
a6e6dea6 | 7903 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7904 | atomic64_add(child_event->total_time_enabled, |
7905 | &parent_event->child_total_time_enabled); | |
7906 | atomic64_add(child_event->total_time_running, | |
7907 | &parent_event->child_total_time_running); | |
d859e29f PM |
7908 | |
7909 | /* | |
cdd6c482 | 7910 | * Remove this event from the parent's list |
d859e29f | 7911 | */ |
cdd6c482 IM |
7912 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7913 | mutex_lock(&parent_event->child_mutex); | |
7914 | list_del_init(&child_event->child_list); | |
7915 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f | 7916 | |
dc633982 JO |
7917 | /* |
7918 | * Make sure user/parent get notified, that we just | |
7919 | * lost one event. | |
7920 | */ | |
7921 | perf_event_wakeup(parent_event); | |
7922 | ||
d859e29f | 7923 | /* |
cdd6c482 | 7924 | * Release the parent event, if this was the last |
d859e29f PM |
7925 | * reference to it. |
7926 | */ | |
a6fa941d | 7927 | put_event(parent_event); |
d859e29f PM |
7928 | } |
7929 | ||
9b51f66d | 7930 | static void |
cdd6c482 IM |
7931 | __perf_event_exit_task(struct perf_event *child_event, |
7932 | struct perf_event_context *child_ctx, | |
38b200d6 | 7933 | struct task_struct *child) |
9b51f66d | 7934 | { |
1903d50c PZ |
7935 | /* |
7936 | * Do not destroy the 'original' grouping; because of the context | |
7937 | * switch optimization the original events could've ended up in a | |
7938 | * random child task. | |
7939 | * | |
7940 | * If we were to destroy the original group, all group related | |
7941 | * operations would cease to function properly after this random | |
7942 | * child dies. | |
7943 | * | |
7944 | * Do destroy all inherited groups, we don't care about those | |
7945 | * and being thorough is better. | |
7946 | */ | |
7947 | perf_remove_from_context(child_event, !!child_event->parent); | |
0cc0c027 | 7948 | |
9b51f66d | 7949 | /* |
38b435b1 | 7950 | * It can happen that the parent exits first, and has events |
9b51f66d | 7951 | * that are still around due to the child reference. These |
38b435b1 | 7952 | * events need to be zapped. |
9b51f66d | 7953 | */ |
38b435b1 | 7954 | if (child_event->parent) { |
cdd6c482 IM |
7955 | sync_child_event(child_event, child); |
7956 | free_event(child_event); | |
179033b3 JO |
7957 | } else { |
7958 | child_event->state = PERF_EVENT_STATE_EXIT; | |
7959 | perf_event_wakeup(child_event); | |
4bcf349a | 7960 | } |
9b51f66d IM |
7961 | } |
7962 | ||
8dc85d54 | 7963 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7964 | { |
ebf905fc | 7965 | struct perf_event *child_event, *next; |
211de6eb | 7966 | struct perf_event_context *child_ctx, *clone_ctx = NULL; |
a63eaf34 | 7967 | unsigned long flags; |
9b51f66d | 7968 | |
8dc85d54 | 7969 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7970 | perf_event_task(child, NULL, 0); |
9b51f66d | 7971 | return; |
9f498cc5 | 7972 | } |
9b51f66d | 7973 | |
a63eaf34 | 7974 | local_irq_save(flags); |
ad3a37de PM |
7975 | /* |
7976 | * We can't reschedule here because interrupts are disabled, | |
7977 | * and either child is current or it is a task that can't be | |
7978 | * scheduled, so we are now safe from rescheduling changing | |
7979 | * our context. | |
7980 | */ | |
806839b2 | 7981 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7982 | |
7983 | /* | |
7984 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7985 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7986 | * incremented the context's refcount before we do put_ctx below. |
7987 | */ | |
e625cce1 | 7988 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7989 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7990 | child->perf_event_ctxp[ctxn] = NULL; |
4a1c0f26 | 7991 | |
71a851b4 PZ |
7992 | /* |
7993 | * If this context is a clone; unclone it so it can't get | |
7994 | * swapped to another process while we're removing all | |
cdd6c482 | 7995 | * the events from it. |
71a851b4 | 7996 | */ |
211de6eb | 7997 | clone_ctx = unclone_ctx(child_ctx); |
5e942bb3 | 7998 | update_context_time(child_ctx); |
e625cce1 | 7999 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 | 8000 | |
211de6eb PZ |
8001 | if (clone_ctx) |
8002 | put_ctx(clone_ctx); | |
4a1c0f26 | 8003 | |
9f498cc5 | 8004 | /* |
cdd6c482 IM |
8005 | * Report the task dead after unscheduling the events so that we |
8006 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
8007 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 8008 | */ |
cdd6c482 | 8009 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 8010 | |
66fff224 PZ |
8011 | /* |
8012 | * We can recurse on the same lock type through: | |
8013 | * | |
cdd6c482 IM |
8014 | * __perf_event_exit_task() |
8015 | * sync_child_event() | |
a6fa941d AV |
8016 | * put_event() |
8017 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
8018 | * |
8019 | * But since its the parent context it won't be the same instance. | |
8020 | */ | |
a0507c84 | 8021 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 8022 | |
ebf905fc | 8023 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 8024 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 8025 | |
a63eaf34 PM |
8026 | mutex_unlock(&child_ctx->mutex); |
8027 | ||
8028 | put_ctx(child_ctx); | |
9b51f66d IM |
8029 | } |
8030 | ||
8dc85d54 PZ |
8031 | /* |
8032 | * When a child task exits, feed back event values to parent events. | |
8033 | */ | |
8034 | void perf_event_exit_task(struct task_struct *child) | |
8035 | { | |
8882135b | 8036 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
8037 | int ctxn; |
8038 | ||
8882135b PZ |
8039 | mutex_lock(&child->perf_event_mutex); |
8040 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
8041 | owner_entry) { | |
8042 | list_del_init(&event->owner_entry); | |
8043 | ||
8044 | /* | |
8045 | * Ensure the list deletion is visible before we clear | |
8046 | * the owner, closes a race against perf_release() where | |
8047 | * we need to serialize on the owner->perf_event_mutex. | |
8048 | */ | |
8049 | smp_wmb(); | |
8050 | event->owner = NULL; | |
8051 | } | |
8052 | mutex_unlock(&child->perf_event_mutex); | |
8053 | ||
8dc85d54 PZ |
8054 | for_each_task_context_nr(ctxn) |
8055 | perf_event_exit_task_context(child, ctxn); | |
8056 | } | |
8057 | ||
889ff015 FW |
8058 | static void perf_free_event(struct perf_event *event, |
8059 | struct perf_event_context *ctx) | |
8060 | { | |
8061 | struct perf_event *parent = event->parent; | |
8062 | ||
8063 | if (WARN_ON_ONCE(!parent)) | |
8064 | return; | |
8065 | ||
8066 | mutex_lock(&parent->child_mutex); | |
8067 | list_del_init(&event->child_list); | |
8068 | mutex_unlock(&parent->child_mutex); | |
8069 | ||
a6fa941d | 8070 | put_event(parent); |
889ff015 | 8071 | |
652884fe | 8072 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 8073 | perf_group_detach(event); |
889ff015 | 8074 | list_del_event(event, ctx); |
652884fe | 8075 | raw_spin_unlock_irq(&ctx->lock); |
889ff015 FW |
8076 | free_event(event); |
8077 | } | |
8078 | ||
bbbee908 | 8079 | /* |
652884fe | 8080 | * Free an unexposed, unused context as created by inheritance by |
8dc85d54 | 8081 | * perf_event_init_task below, used by fork() in case of fail. |
652884fe PZ |
8082 | * |
8083 | * Not all locks are strictly required, but take them anyway to be nice and | |
8084 | * help out with the lockdep assertions. | |
bbbee908 | 8085 | */ |
cdd6c482 | 8086 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 8087 | { |
8dc85d54 | 8088 | struct perf_event_context *ctx; |
cdd6c482 | 8089 | struct perf_event *event, *tmp; |
8dc85d54 | 8090 | int ctxn; |
bbbee908 | 8091 | |
8dc85d54 PZ |
8092 | for_each_task_context_nr(ctxn) { |
8093 | ctx = task->perf_event_ctxp[ctxn]; | |
8094 | if (!ctx) | |
8095 | continue; | |
bbbee908 | 8096 | |
8dc85d54 | 8097 | mutex_lock(&ctx->mutex); |
bbbee908 | 8098 | again: |
8dc85d54 PZ |
8099 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
8100 | group_entry) | |
8101 | perf_free_event(event, ctx); | |
bbbee908 | 8102 | |
8dc85d54 PZ |
8103 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
8104 | group_entry) | |
8105 | perf_free_event(event, ctx); | |
bbbee908 | 8106 | |
8dc85d54 PZ |
8107 | if (!list_empty(&ctx->pinned_groups) || |
8108 | !list_empty(&ctx->flexible_groups)) | |
8109 | goto again; | |
bbbee908 | 8110 | |
8dc85d54 | 8111 | mutex_unlock(&ctx->mutex); |
bbbee908 | 8112 | |
8dc85d54 PZ |
8113 | put_ctx(ctx); |
8114 | } | |
889ff015 FW |
8115 | } |
8116 | ||
4e231c79 PZ |
8117 | void perf_event_delayed_put(struct task_struct *task) |
8118 | { | |
8119 | int ctxn; | |
8120 | ||
8121 | for_each_task_context_nr(ctxn) | |
8122 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
8123 | } | |
8124 | ||
97dee4f3 PZ |
8125 | /* |
8126 | * inherit a event from parent task to child task: | |
8127 | */ | |
8128 | static struct perf_event * | |
8129 | inherit_event(struct perf_event *parent_event, | |
8130 | struct task_struct *parent, | |
8131 | struct perf_event_context *parent_ctx, | |
8132 | struct task_struct *child, | |
8133 | struct perf_event *group_leader, | |
8134 | struct perf_event_context *child_ctx) | |
8135 | { | |
1929def9 | 8136 | enum perf_event_active_state parent_state = parent_event->state; |
97dee4f3 | 8137 | struct perf_event *child_event; |
cee010ec | 8138 | unsigned long flags; |
97dee4f3 PZ |
8139 | |
8140 | /* | |
8141 | * Instead of creating recursive hierarchies of events, | |
8142 | * we link inherited events back to the original parent, | |
8143 | * which has a filp for sure, which we use as the reference | |
8144 | * count: | |
8145 | */ | |
8146 | if (parent_event->parent) | |
8147 | parent_event = parent_event->parent; | |
8148 | ||
8149 | child_event = perf_event_alloc(&parent_event->attr, | |
8150 | parent_event->cpu, | |
d580ff86 | 8151 | child, |
97dee4f3 | 8152 | group_leader, parent_event, |
4dc0da86 | 8153 | NULL, NULL); |
97dee4f3 PZ |
8154 | if (IS_ERR(child_event)) |
8155 | return child_event; | |
a6fa941d | 8156 | |
fadfe7be JO |
8157 | if (is_orphaned_event(parent_event) || |
8158 | !atomic_long_inc_not_zero(&parent_event->refcount)) { | |
a6fa941d AV |
8159 | free_event(child_event); |
8160 | return NULL; | |
8161 | } | |
8162 | ||
97dee4f3 PZ |
8163 | get_ctx(child_ctx); |
8164 | ||
8165 | /* | |
8166 | * Make the child state follow the state of the parent event, | |
8167 | * not its attr.disabled bit. We hold the parent's mutex, | |
8168 | * so we won't race with perf_event_{en, dis}able_family. | |
8169 | */ | |
1929def9 | 8170 | if (parent_state >= PERF_EVENT_STATE_INACTIVE) |
97dee4f3 PZ |
8171 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
8172 | else | |
8173 | child_event->state = PERF_EVENT_STATE_OFF; | |
8174 | ||
8175 | if (parent_event->attr.freq) { | |
8176 | u64 sample_period = parent_event->hw.sample_period; | |
8177 | struct hw_perf_event *hwc = &child_event->hw; | |
8178 | ||
8179 | hwc->sample_period = sample_period; | |
8180 | hwc->last_period = sample_period; | |
8181 | ||
8182 | local64_set(&hwc->period_left, sample_period); | |
8183 | } | |
8184 | ||
8185 | child_event->ctx = child_ctx; | |
8186 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
8187 | child_event->overflow_handler_context |
8188 | = parent_event->overflow_handler_context; | |
97dee4f3 | 8189 | |
614b6780 TG |
8190 | /* |
8191 | * Precalculate sample_data sizes | |
8192 | */ | |
8193 | perf_event__header_size(child_event); | |
6844c09d | 8194 | perf_event__id_header_size(child_event); |
614b6780 | 8195 | |
97dee4f3 PZ |
8196 | /* |
8197 | * Link it up in the child's context: | |
8198 | */ | |
cee010ec | 8199 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 8200 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 8201 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 8202 | |
97dee4f3 PZ |
8203 | /* |
8204 | * Link this into the parent event's child list | |
8205 | */ | |
8206 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
8207 | mutex_lock(&parent_event->child_mutex); | |
8208 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
8209 | mutex_unlock(&parent_event->child_mutex); | |
8210 | ||
8211 | return child_event; | |
8212 | } | |
8213 | ||
8214 | static int inherit_group(struct perf_event *parent_event, | |
8215 | struct task_struct *parent, | |
8216 | struct perf_event_context *parent_ctx, | |
8217 | struct task_struct *child, | |
8218 | struct perf_event_context *child_ctx) | |
8219 | { | |
8220 | struct perf_event *leader; | |
8221 | struct perf_event *sub; | |
8222 | struct perf_event *child_ctr; | |
8223 | ||
8224 | leader = inherit_event(parent_event, parent, parent_ctx, | |
8225 | child, NULL, child_ctx); | |
8226 | if (IS_ERR(leader)) | |
8227 | return PTR_ERR(leader); | |
8228 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
8229 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
8230 | child, leader, child_ctx); | |
8231 | if (IS_ERR(child_ctr)) | |
8232 | return PTR_ERR(child_ctr); | |
8233 | } | |
8234 | return 0; | |
889ff015 FW |
8235 | } |
8236 | ||
8237 | static int | |
8238 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
8239 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 8240 | struct task_struct *child, int ctxn, |
889ff015 FW |
8241 | int *inherited_all) |
8242 | { | |
8243 | int ret; | |
8dc85d54 | 8244 | struct perf_event_context *child_ctx; |
889ff015 FW |
8245 | |
8246 | if (!event->attr.inherit) { | |
8247 | *inherited_all = 0; | |
8248 | return 0; | |
bbbee908 PZ |
8249 | } |
8250 | ||
fe4b04fa | 8251 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
8252 | if (!child_ctx) { |
8253 | /* | |
8254 | * This is executed from the parent task context, so | |
8255 | * inherit events that have been marked for cloning. | |
8256 | * First allocate and initialize a context for the | |
8257 | * child. | |
8258 | */ | |
bbbee908 | 8259 | |
734df5ab | 8260 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
8261 | if (!child_ctx) |
8262 | return -ENOMEM; | |
bbbee908 | 8263 | |
8dc85d54 | 8264 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
8265 | } |
8266 | ||
8267 | ret = inherit_group(event, parent, parent_ctx, | |
8268 | child, child_ctx); | |
8269 | ||
8270 | if (ret) | |
8271 | *inherited_all = 0; | |
8272 | ||
8273 | return ret; | |
bbbee908 PZ |
8274 | } |
8275 | ||
9b51f66d | 8276 | /* |
cdd6c482 | 8277 | * Initialize the perf_event context in task_struct |
9b51f66d | 8278 | */ |
985c8dcb | 8279 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 8280 | { |
889ff015 | 8281 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
8282 | struct perf_event_context *cloned_ctx; |
8283 | struct perf_event *event; | |
9b51f66d | 8284 | struct task_struct *parent = current; |
564c2b21 | 8285 | int inherited_all = 1; |
dddd3379 | 8286 | unsigned long flags; |
6ab423e0 | 8287 | int ret = 0; |
9b51f66d | 8288 | |
8dc85d54 | 8289 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
8290 | return 0; |
8291 | ||
ad3a37de | 8292 | /* |
25346b93 PM |
8293 | * If the parent's context is a clone, pin it so it won't get |
8294 | * swapped under us. | |
ad3a37de | 8295 | */ |
8dc85d54 | 8296 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
8297 | if (!parent_ctx) |
8298 | return 0; | |
25346b93 | 8299 | |
ad3a37de PM |
8300 | /* |
8301 | * No need to check if parent_ctx != NULL here; since we saw | |
8302 | * it non-NULL earlier, the only reason for it to become NULL | |
8303 | * is if we exit, and since we're currently in the middle of | |
8304 | * a fork we can't be exiting at the same time. | |
8305 | */ | |
ad3a37de | 8306 | |
9b51f66d IM |
8307 | /* |
8308 | * Lock the parent list. No need to lock the child - not PID | |
8309 | * hashed yet and not running, so nobody can access it. | |
8310 | */ | |
d859e29f | 8311 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
8312 | |
8313 | /* | |
8314 | * We dont have to disable NMIs - we are only looking at | |
8315 | * the list, not manipulating it: | |
8316 | */ | |
889ff015 | 8317 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
8318 | ret = inherit_task_group(event, parent, parent_ctx, |
8319 | child, ctxn, &inherited_all); | |
889ff015 FW |
8320 | if (ret) |
8321 | break; | |
8322 | } | |
b93f7978 | 8323 | |
dddd3379 TG |
8324 | /* |
8325 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
8326 | * to allocations, but we need to prevent rotation because | |
8327 | * rotate_ctx() will change the list from interrupt context. | |
8328 | */ | |
8329 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
8330 | parent_ctx->rotate_disable = 1; | |
8331 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
8332 | ||
889ff015 | 8333 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
8334 | ret = inherit_task_group(event, parent, parent_ctx, |
8335 | child, ctxn, &inherited_all); | |
889ff015 | 8336 | if (ret) |
9b51f66d | 8337 | break; |
564c2b21 PM |
8338 | } |
8339 | ||
dddd3379 TG |
8340 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
8341 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 8342 | |
8dc85d54 | 8343 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 8344 | |
05cbaa28 | 8345 | if (child_ctx && inherited_all) { |
564c2b21 PM |
8346 | /* |
8347 | * Mark the child context as a clone of the parent | |
8348 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
8349 | * |
8350 | * Note that if the parent is a clone, the holding of | |
8351 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 8352 | */ |
c5ed5145 | 8353 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
8354 | if (cloned_ctx) { |
8355 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 8356 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
8357 | } else { |
8358 | child_ctx->parent_ctx = parent_ctx; | |
8359 | child_ctx->parent_gen = parent_ctx->generation; | |
8360 | } | |
8361 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
8362 | } |
8363 | ||
c5ed5145 | 8364 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 8365 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 8366 | |
25346b93 | 8367 | perf_unpin_context(parent_ctx); |
fe4b04fa | 8368 | put_ctx(parent_ctx); |
ad3a37de | 8369 | |
6ab423e0 | 8370 | return ret; |
9b51f66d IM |
8371 | } |
8372 | ||
8dc85d54 PZ |
8373 | /* |
8374 | * Initialize the perf_event context in task_struct | |
8375 | */ | |
8376 | int perf_event_init_task(struct task_struct *child) | |
8377 | { | |
8378 | int ctxn, ret; | |
8379 | ||
8550d7cb ON |
8380 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
8381 | mutex_init(&child->perf_event_mutex); | |
8382 | INIT_LIST_HEAD(&child->perf_event_list); | |
8383 | ||
8dc85d54 PZ |
8384 | for_each_task_context_nr(ctxn) { |
8385 | ret = perf_event_init_context(child, ctxn); | |
6c72e350 PZ |
8386 | if (ret) { |
8387 | perf_event_free_task(child); | |
8dc85d54 | 8388 | return ret; |
6c72e350 | 8389 | } |
8dc85d54 PZ |
8390 | } |
8391 | ||
8392 | return 0; | |
8393 | } | |
8394 | ||
220b140b PM |
8395 | static void __init perf_event_init_all_cpus(void) |
8396 | { | |
b28ab83c | 8397 | struct swevent_htable *swhash; |
220b140b | 8398 | int cpu; |
220b140b PM |
8399 | |
8400 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
8401 | swhash = &per_cpu(swevent_htable, cpu); |
8402 | mutex_init(&swhash->hlist_mutex); | |
2fde4f94 | 8403 | INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu)); |
220b140b PM |
8404 | } |
8405 | } | |
8406 | ||
0db0628d | 8407 | static void perf_event_init_cpu(int cpu) |
0793a61d | 8408 | { |
108b02cf | 8409 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 8410 | |
b28ab83c | 8411 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8412 | swhash->online = true; |
4536e4d1 | 8413 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
8414 | struct swevent_hlist *hlist; |
8415 | ||
b28ab83c PZ |
8416 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
8417 | WARN_ON(!hlist); | |
8418 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 8419 | } |
b28ab83c | 8420 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
8421 | } |
8422 | ||
c277443c | 8423 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
108b02cf | 8424 | static void __perf_event_exit_context(void *__info) |
0793a61d | 8425 | { |
226424ee | 8426 | struct remove_event re = { .detach_group = true }; |
108b02cf | 8427 | struct perf_event_context *ctx = __info; |
0793a61d | 8428 | |
e3703f8c | 8429 | rcu_read_lock(); |
46ce0fe9 PZ |
8430 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
8431 | __perf_remove_from_context(&re); | |
e3703f8c | 8432 | rcu_read_unlock(); |
0793a61d | 8433 | } |
108b02cf PZ |
8434 | |
8435 | static void perf_event_exit_cpu_context(int cpu) | |
8436 | { | |
8437 | struct perf_event_context *ctx; | |
8438 | struct pmu *pmu; | |
8439 | int idx; | |
8440 | ||
8441 | idx = srcu_read_lock(&pmus_srcu); | |
8442 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 8443 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
8444 | |
8445 | mutex_lock(&ctx->mutex); | |
8446 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
8447 | mutex_unlock(&ctx->mutex); | |
8448 | } | |
8449 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
8450 | } |
8451 | ||
cdd6c482 | 8452 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 8453 | { |
b28ab83c | 8454 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 8455 | |
e3703f8c PZ |
8456 | perf_event_exit_cpu_context(cpu); |
8457 | ||
b28ab83c | 8458 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8459 | swhash->online = false; |
b28ab83c PZ |
8460 | swevent_hlist_release(swhash); |
8461 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
8462 | } |
8463 | #else | |
cdd6c482 | 8464 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
8465 | #endif |
8466 | ||
c277443c PZ |
8467 | static int |
8468 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
8469 | { | |
8470 | int cpu; | |
8471 | ||
8472 | for_each_online_cpu(cpu) | |
8473 | perf_event_exit_cpu(cpu); | |
8474 | ||
8475 | return NOTIFY_OK; | |
8476 | } | |
8477 | ||
8478 | /* | |
8479 | * Run the perf reboot notifier at the very last possible moment so that | |
8480 | * the generic watchdog code runs as long as possible. | |
8481 | */ | |
8482 | static struct notifier_block perf_reboot_notifier = { | |
8483 | .notifier_call = perf_reboot, | |
8484 | .priority = INT_MIN, | |
8485 | }; | |
8486 | ||
0db0628d | 8487 | static int |
0793a61d TG |
8488 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
8489 | { | |
8490 | unsigned int cpu = (long)hcpu; | |
8491 | ||
4536e4d1 | 8492 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
8493 | |
8494 | case CPU_UP_PREPARE: | |
5e11637e | 8495 | case CPU_DOWN_FAILED: |
cdd6c482 | 8496 | perf_event_init_cpu(cpu); |
0793a61d TG |
8497 | break; |
8498 | ||
5e11637e | 8499 | case CPU_UP_CANCELED: |
0793a61d | 8500 | case CPU_DOWN_PREPARE: |
cdd6c482 | 8501 | perf_event_exit_cpu(cpu); |
0793a61d | 8502 | break; |
0793a61d TG |
8503 | default: |
8504 | break; | |
8505 | } | |
8506 | ||
8507 | return NOTIFY_OK; | |
8508 | } | |
8509 | ||
cdd6c482 | 8510 | void __init perf_event_init(void) |
0793a61d | 8511 | { |
3c502e7a JW |
8512 | int ret; |
8513 | ||
2e80a82a PZ |
8514 | idr_init(&pmu_idr); |
8515 | ||
220b140b | 8516 | perf_event_init_all_cpus(); |
b0a873eb | 8517 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
8518 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
8519 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
8520 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
8521 | perf_tp_register(); |
8522 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 8523 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
8524 | |
8525 | ret = init_hw_breakpoint(); | |
8526 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
8527 | |
8528 | /* do not patch jump label more than once per second */ | |
8529 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
8530 | |
8531 | /* | |
8532 | * Build time assertion that we keep the data_head at the intended | |
8533 | * location. IOW, validation we got the __reserved[] size right. | |
8534 | */ | |
8535 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
8536 | != 1024); | |
0793a61d | 8537 | } |
abe43400 PZ |
8538 | |
8539 | static int __init perf_event_sysfs_init(void) | |
8540 | { | |
8541 | struct pmu *pmu; | |
8542 | int ret; | |
8543 | ||
8544 | mutex_lock(&pmus_lock); | |
8545 | ||
8546 | ret = bus_register(&pmu_bus); | |
8547 | if (ret) | |
8548 | goto unlock; | |
8549 | ||
8550 | list_for_each_entry(pmu, &pmus, entry) { | |
8551 | if (!pmu->name || pmu->type < 0) | |
8552 | continue; | |
8553 | ||
8554 | ret = pmu_dev_alloc(pmu); | |
8555 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8556 | } | |
8557 | pmu_bus_running = 1; | |
8558 | ret = 0; | |
8559 | ||
8560 | unlock: | |
8561 | mutex_unlock(&pmus_lock); | |
8562 | ||
8563 | return ret; | |
8564 | } | |
8565 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8566 | |
8567 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8568 | static struct cgroup_subsys_state * |
8569 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8570 | { |
8571 | struct perf_cgroup *jc; | |
e5d1367f | 8572 | |
1b15d055 | 8573 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8574 | if (!jc) |
8575 | return ERR_PTR(-ENOMEM); | |
8576 | ||
e5d1367f SE |
8577 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8578 | if (!jc->info) { | |
8579 | kfree(jc); | |
8580 | return ERR_PTR(-ENOMEM); | |
8581 | } | |
8582 | ||
e5d1367f SE |
8583 | return &jc->css; |
8584 | } | |
8585 | ||
eb95419b | 8586 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8587 | { |
eb95419b TH |
8588 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8589 | ||
e5d1367f SE |
8590 | free_percpu(jc->info); |
8591 | kfree(jc); | |
8592 | } | |
8593 | ||
8594 | static int __perf_cgroup_move(void *info) | |
8595 | { | |
8596 | struct task_struct *task = info; | |
8597 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8598 | return 0; | |
8599 | } | |
8600 | ||
eb95419b TH |
8601 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8602 | struct cgroup_taskset *tset) | |
e5d1367f | 8603 | { |
bb9d97b6 TH |
8604 | struct task_struct *task; |
8605 | ||
924f0d9a | 8606 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 8607 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8608 | } |
8609 | ||
eb95419b TH |
8610 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8611 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8612 | struct task_struct *task) |
e5d1367f SE |
8613 | { |
8614 | /* | |
8615 | * cgroup_exit() is called in the copy_process() failure path. | |
8616 | * Ignore this case since the task hasn't ran yet, this avoids | |
8617 | * trying to poke a half freed task state from generic code. | |
8618 | */ | |
8619 | if (!(task->flags & PF_EXITING)) | |
8620 | return; | |
8621 | ||
bb9d97b6 | 8622 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8623 | } |
8624 | ||
073219e9 | 8625 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
8626 | .css_alloc = perf_cgroup_css_alloc, |
8627 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8628 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8629 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8630 | }; |
8631 | #endif /* CONFIG_CGROUP_PERF */ |