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