Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
cdd6c482 | 37 | #include <linux/perf_event.h> |
6fb2915d | 38 | #include <linux/ftrace_event.h> |
3c502e7a | 39 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 40 | #include <linux/mm_types.h> |
877c6856 | 41 | #include <linux/cgroup.h> |
c464c76e | 42 | #include <linux/module.h> |
0793a61d | 43 | |
76369139 FW |
44 | #include "internal.h" |
45 | ||
4e193bd4 TB |
46 | #include <asm/irq_regs.h> |
47 | ||
fe4b04fa | 48 | struct remote_function_call { |
e7e7ee2e IM |
49 | struct task_struct *p; |
50 | int (*func)(void *info); | |
51 | void *info; | |
52 | int ret; | |
fe4b04fa PZ |
53 | }; |
54 | ||
55 | static void remote_function(void *data) | |
56 | { | |
57 | struct remote_function_call *tfc = data; | |
58 | struct task_struct *p = tfc->p; | |
59 | ||
60 | if (p) { | |
61 | tfc->ret = -EAGAIN; | |
62 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
63 | return; | |
64 | } | |
65 | ||
66 | tfc->ret = tfc->func(tfc->info); | |
67 | } | |
68 | ||
69 | /** | |
70 | * task_function_call - call a function on the cpu on which a task runs | |
71 | * @p: the task to evaluate | |
72 | * @func: the function to be called | |
73 | * @info: the function call argument | |
74 | * | |
75 | * Calls the function @func when the task is currently running. This might | |
76 | * be on the current CPU, which just calls the function directly | |
77 | * | |
78 | * returns: @func return value, or | |
79 | * -ESRCH - when the process isn't running | |
80 | * -EAGAIN - when the process moved away | |
81 | */ | |
82 | static int | |
83 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
84 | { | |
85 | struct remote_function_call data = { | |
e7e7ee2e IM |
86 | .p = p, |
87 | .func = func, | |
88 | .info = info, | |
89 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
90 | }; |
91 | ||
92 | if (task_curr(p)) | |
93 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
94 | ||
95 | return data.ret; | |
96 | } | |
97 | ||
98 | /** | |
99 | * cpu_function_call - call a function on the cpu | |
100 | * @func: the function to be called | |
101 | * @info: the function call argument | |
102 | * | |
103 | * Calls the function @func on the remote cpu. | |
104 | * | |
105 | * returns: @func return value or -ENXIO when the cpu is offline | |
106 | */ | |
107 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
108 | { | |
109 | struct remote_function_call data = { | |
e7e7ee2e IM |
110 | .p = NULL, |
111 | .func = func, | |
112 | .info = info, | |
113 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
114 | }; |
115 | ||
116 | smp_call_function_single(cpu, remote_function, &data, 1); | |
117 | ||
118 | return data.ret; | |
119 | } | |
120 | ||
e5d1367f SE |
121 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
122 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
123 | PERF_FLAG_PID_CGROUP |\ |
124 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 125 | |
bce38cd5 SE |
126 | /* |
127 | * branch priv levels that need permission checks | |
128 | */ | |
129 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
130 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
131 | PERF_SAMPLE_BRANCH_HV) | |
132 | ||
0b3fcf17 SE |
133 | enum event_type_t { |
134 | EVENT_FLEXIBLE = 0x1, | |
135 | EVENT_PINNED = 0x2, | |
136 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
137 | }; | |
138 | ||
e5d1367f SE |
139 | /* |
140 | * perf_sched_events : >0 events exist | |
141 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
142 | */ | |
c5905afb | 143 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 144 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 145 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 146 | |
cdd6c482 IM |
147 | static atomic_t nr_mmap_events __read_mostly; |
148 | static atomic_t nr_comm_events __read_mostly; | |
149 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 150 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 151 | |
108b02cf PZ |
152 | static LIST_HEAD(pmus); |
153 | static DEFINE_MUTEX(pmus_lock); | |
154 | static struct srcu_struct pmus_srcu; | |
155 | ||
0764771d | 156 | /* |
cdd6c482 | 157 | * perf event paranoia level: |
0fbdea19 IM |
158 | * -1 - not paranoid at all |
159 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 160 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 161 | * 2 - disallow kernel profiling for unpriv |
0764771d | 162 | */ |
cdd6c482 | 163 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 164 | |
20443384 FW |
165 | /* Minimum for 512 kiB + 1 user control page */ |
166 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
167 | |
168 | /* | |
cdd6c482 | 169 | * max perf event sample rate |
df58ab24 | 170 | */ |
14c63f17 DH |
171 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
172 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
173 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
174 | ||
175 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
176 | ||
177 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
178 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
179 | ||
d9494cb4 PZ |
180 | static int perf_sample_allowed_ns __read_mostly = |
181 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
182 | |
183 | void update_perf_cpu_limits(void) | |
184 | { | |
185 | u64 tmp = perf_sample_period_ns; | |
186 | ||
187 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 188 | do_div(tmp, 100); |
d9494cb4 | 189 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 190 | } |
163ec435 | 191 | |
9e630205 SE |
192 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
193 | ||
163ec435 PZ |
194 | int perf_proc_update_handler(struct ctl_table *table, int write, |
195 | void __user *buffer, size_t *lenp, | |
196 | loff_t *ppos) | |
197 | { | |
723478c8 | 198 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
199 | |
200 | if (ret || !write) | |
201 | return ret; | |
202 | ||
203 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
204 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
205 | update_perf_cpu_limits(); | |
206 | ||
207 | return 0; | |
208 | } | |
209 | ||
210 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
211 | ||
212 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
213 | void __user *buffer, size_t *lenp, | |
214 | loff_t *ppos) | |
215 | { | |
216 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
217 | ||
218 | if (ret || !write) | |
219 | return ret; | |
220 | ||
221 | update_perf_cpu_limits(); | |
163ec435 PZ |
222 | |
223 | return 0; | |
224 | } | |
1ccd1549 | 225 | |
14c63f17 DH |
226 | /* |
227 | * perf samples are done in some very critical code paths (NMIs). | |
228 | * If they take too much CPU time, the system can lock up and not | |
229 | * get any real work done. This will drop the sample rate when | |
230 | * we detect that events are taking too long. | |
231 | */ | |
232 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 233 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 234 | |
6a02ad66 | 235 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 236 | { |
6a02ad66 | 237 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 238 | u64 avg_local_sample_len; |
e5302920 | 239 | u64 local_samples_len; |
6a02ad66 PZ |
240 | |
241 | local_samples_len = __get_cpu_var(running_sample_length); | |
242 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
243 | ||
244 | printk_ratelimited(KERN_WARNING | |
245 | "perf interrupt took too long (%lld > %lld), lowering " | |
246 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 247 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
248 | sysctl_perf_event_sample_rate); |
249 | } | |
250 | ||
251 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
252 | ||
253 | void perf_sample_event_took(u64 sample_len_ns) | |
254 | { | |
d9494cb4 | 255 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
256 | u64 avg_local_sample_len; |
257 | u64 local_samples_len; | |
14c63f17 | 258 | |
d9494cb4 | 259 | if (allowed_ns == 0) |
14c63f17 DH |
260 | return; |
261 | ||
262 | /* decay the counter by 1 average sample */ | |
263 | local_samples_len = __get_cpu_var(running_sample_length); | |
264 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
265 | local_samples_len += sample_len_ns; | |
266 | __get_cpu_var(running_sample_length) = local_samples_len; | |
267 | ||
268 | /* | |
269 | * note: this will be biased artifically low until we have | |
270 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
271 | * from having to maintain a count. | |
272 | */ | |
273 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
274 | ||
d9494cb4 | 275 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
276 | return; |
277 | ||
278 | if (max_samples_per_tick <= 1) | |
279 | return; | |
280 | ||
281 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
282 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
283 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
284 | ||
14c63f17 | 285 | update_perf_cpu_limits(); |
6a02ad66 | 286 | |
cd578abb PZ |
287 | if (!irq_work_queue(&perf_duration_work)) { |
288 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
289 | "kernel.perf_event_max_sample_rate to %d\n", | |
290 | avg_local_sample_len, allowed_ns >> 1, | |
291 | sysctl_perf_event_sample_rate); | |
292 | } | |
14c63f17 DH |
293 | } |
294 | ||
cdd6c482 | 295 | static atomic64_t perf_event_id; |
a96bbc16 | 296 | |
0b3fcf17 SE |
297 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
298 | enum event_type_t event_type); | |
299 | ||
300 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
301 | enum event_type_t event_type, |
302 | struct task_struct *task); | |
303 | ||
304 | static void update_context_time(struct perf_event_context *ctx); | |
305 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 306 | |
cdd6c482 | 307 | void __weak perf_event_print_debug(void) { } |
0793a61d | 308 | |
84c79910 | 309 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 310 | { |
84c79910 | 311 | return "pmu"; |
0793a61d TG |
312 | } |
313 | ||
0b3fcf17 SE |
314 | static inline u64 perf_clock(void) |
315 | { | |
316 | return local_clock(); | |
317 | } | |
318 | ||
e5d1367f SE |
319 | static inline struct perf_cpu_context * |
320 | __get_cpu_context(struct perf_event_context *ctx) | |
321 | { | |
322 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
323 | } | |
324 | ||
facc4307 PZ |
325 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
326 | struct perf_event_context *ctx) | |
327 | { | |
328 | raw_spin_lock(&cpuctx->ctx.lock); | |
329 | if (ctx) | |
330 | raw_spin_lock(&ctx->lock); | |
331 | } | |
332 | ||
333 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
334 | struct perf_event_context *ctx) | |
335 | { | |
336 | if (ctx) | |
337 | raw_spin_unlock(&ctx->lock); | |
338 | raw_spin_unlock(&cpuctx->ctx.lock); | |
339 | } | |
340 | ||
e5d1367f SE |
341 | #ifdef CONFIG_CGROUP_PERF |
342 | ||
877c6856 LZ |
343 | /* |
344 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
345 | * This is a per-cpu dynamically allocated data structure. | |
346 | */ | |
347 | struct perf_cgroup_info { | |
348 | u64 time; | |
349 | u64 timestamp; | |
350 | }; | |
351 | ||
352 | struct perf_cgroup { | |
353 | struct cgroup_subsys_state css; | |
86e213e1 | 354 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
355 | }; |
356 | ||
3f7cce3c SE |
357 | /* |
358 | * Must ensure cgroup is pinned (css_get) before calling | |
359 | * this function. In other words, we cannot call this function | |
360 | * if there is no cgroup event for the current CPU context. | |
361 | */ | |
e5d1367f SE |
362 | static inline struct perf_cgroup * |
363 | perf_cgroup_from_task(struct task_struct *task) | |
364 | { | |
073219e9 | 365 | return container_of(task_css(task, perf_event_cgrp_id), |
8af01f56 | 366 | struct perf_cgroup, css); |
e5d1367f SE |
367 | } |
368 | ||
369 | static inline bool | |
370 | perf_cgroup_match(struct perf_event *event) | |
371 | { | |
372 | struct perf_event_context *ctx = event->ctx; | |
373 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
374 | ||
ef824fa1 TH |
375 | /* @event doesn't care about cgroup */ |
376 | if (!event->cgrp) | |
377 | return true; | |
378 | ||
379 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
380 | if (!cpuctx->cgrp) | |
381 | return false; | |
382 | ||
383 | /* | |
384 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
385 | * also enabled for all its descendant cgroups. If @cpuctx's | |
386 | * cgroup is a descendant of @event's (the test covers identity | |
387 | * case), it's a match. | |
388 | */ | |
389 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
390 | event->cgrp->css.cgroup); | |
e5d1367f SE |
391 | } |
392 | ||
e5d1367f SE |
393 | static inline void perf_put_cgroup(struct perf_event *event) |
394 | { | |
395 | css_put(&event->cgrp->css); | |
396 | } | |
397 | ||
398 | static inline void perf_detach_cgroup(struct perf_event *event) | |
399 | { | |
400 | perf_put_cgroup(event); | |
401 | event->cgrp = NULL; | |
402 | } | |
403 | ||
404 | static inline int is_cgroup_event(struct perf_event *event) | |
405 | { | |
406 | return event->cgrp != NULL; | |
407 | } | |
408 | ||
409 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
410 | { | |
411 | struct perf_cgroup_info *t; | |
412 | ||
413 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
414 | return t->time; | |
415 | } | |
416 | ||
417 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
418 | { | |
419 | struct perf_cgroup_info *info; | |
420 | u64 now; | |
421 | ||
422 | now = perf_clock(); | |
423 | ||
424 | info = this_cpu_ptr(cgrp->info); | |
425 | ||
426 | info->time += now - info->timestamp; | |
427 | info->timestamp = now; | |
428 | } | |
429 | ||
430 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
431 | { | |
432 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
433 | if (cgrp_out) | |
434 | __update_cgrp_time(cgrp_out); | |
435 | } | |
436 | ||
437 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
438 | { | |
3f7cce3c SE |
439 | struct perf_cgroup *cgrp; |
440 | ||
e5d1367f | 441 | /* |
3f7cce3c SE |
442 | * ensure we access cgroup data only when needed and |
443 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 444 | */ |
3f7cce3c | 445 | if (!is_cgroup_event(event)) |
e5d1367f SE |
446 | return; |
447 | ||
3f7cce3c SE |
448 | cgrp = perf_cgroup_from_task(current); |
449 | /* | |
450 | * Do not update time when cgroup is not active | |
451 | */ | |
452 | if (cgrp == event->cgrp) | |
453 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
454 | } |
455 | ||
456 | static inline void | |
3f7cce3c SE |
457 | perf_cgroup_set_timestamp(struct task_struct *task, |
458 | struct perf_event_context *ctx) | |
e5d1367f SE |
459 | { |
460 | struct perf_cgroup *cgrp; | |
461 | struct perf_cgroup_info *info; | |
462 | ||
3f7cce3c SE |
463 | /* |
464 | * ctx->lock held by caller | |
465 | * ensure we do not access cgroup data | |
466 | * unless we have the cgroup pinned (css_get) | |
467 | */ | |
468 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
469 | return; |
470 | ||
471 | cgrp = perf_cgroup_from_task(task); | |
472 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 473 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
474 | } |
475 | ||
476 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
477 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
478 | ||
479 | /* | |
480 | * reschedule events based on the cgroup constraint of task. | |
481 | * | |
482 | * mode SWOUT : schedule out everything | |
483 | * mode SWIN : schedule in based on cgroup for next | |
484 | */ | |
485 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
486 | { | |
487 | struct perf_cpu_context *cpuctx; | |
488 | struct pmu *pmu; | |
489 | unsigned long flags; | |
490 | ||
491 | /* | |
492 | * disable interrupts to avoid geting nr_cgroup | |
493 | * changes via __perf_event_disable(). Also | |
494 | * avoids preemption. | |
495 | */ | |
496 | local_irq_save(flags); | |
497 | ||
498 | /* | |
499 | * we reschedule only in the presence of cgroup | |
500 | * constrained events. | |
501 | */ | |
502 | rcu_read_lock(); | |
503 | ||
504 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 505 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
506 | if (cpuctx->unique_pmu != pmu) |
507 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 508 | |
e5d1367f SE |
509 | /* |
510 | * perf_cgroup_events says at least one | |
511 | * context on this CPU has cgroup events. | |
512 | * | |
513 | * ctx->nr_cgroups reports the number of cgroup | |
514 | * events for a context. | |
515 | */ | |
516 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
517 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
518 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
519 | |
520 | if (mode & PERF_CGROUP_SWOUT) { | |
521 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
522 | /* | |
523 | * must not be done before ctxswout due | |
524 | * to event_filter_match() in event_sched_out() | |
525 | */ | |
526 | cpuctx->cgrp = NULL; | |
527 | } | |
528 | ||
529 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 530 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
531 | /* |
532 | * set cgrp before ctxsw in to allow | |
533 | * event_filter_match() to not have to pass | |
534 | * task around | |
e5d1367f SE |
535 | */ |
536 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
537 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
538 | } | |
facc4307 PZ |
539 | perf_pmu_enable(cpuctx->ctx.pmu); |
540 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 541 | } |
e5d1367f SE |
542 | } |
543 | ||
544 | rcu_read_unlock(); | |
545 | ||
546 | local_irq_restore(flags); | |
547 | } | |
548 | ||
a8d757ef SE |
549 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
550 | struct task_struct *next) | |
e5d1367f | 551 | { |
a8d757ef SE |
552 | struct perf_cgroup *cgrp1; |
553 | struct perf_cgroup *cgrp2 = NULL; | |
554 | ||
555 | /* | |
556 | * we come here when we know perf_cgroup_events > 0 | |
557 | */ | |
558 | cgrp1 = perf_cgroup_from_task(task); | |
559 | ||
560 | /* | |
561 | * next is NULL when called from perf_event_enable_on_exec() | |
562 | * that will systematically cause a cgroup_switch() | |
563 | */ | |
564 | if (next) | |
565 | cgrp2 = perf_cgroup_from_task(next); | |
566 | ||
567 | /* | |
568 | * only schedule out current cgroup events if we know | |
569 | * that we are switching to a different cgroup. Otherwise, | |
570 | * do no touch the cgroup events. | |
571 | */ | |
572 | if (cgrp1 != cgrp2) | |
573 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
574 | } |
575 | ||
a8d757ef SE |
576 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
577 | struct task_struct *task) | |
e5d1367f | 578 | { |
a8d757ef SE |
579 | struct perf_cgroup *cgrp1; |
580 | struct perf_cgroup *cgrp2 = NULL; | |
581 | ||
582 | /* | |
583 | * we come here when we know perf_cgroup_events > 0 | |
584 | */ | |
585 | cgrp1 = perf_cgroup_from_task(task); | |
586 | ||
587 | /* prev can never be NULL */ | |
588 | cgrp2 = perf_cgroup_from_task(prev); | |
589 | ||
590 | /* | |
591 | * only need to schedule in cgroup events if we are changing | |
592 | * cgroup during ctxsw. Cgroup events were not scheduled | |
593 | * out of ctxsw out if that was not the case. | |
594 | */ | |
595 | if (cgrp1 != cgrp2) | |
596 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
597 | } |
598 | ||
599 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
600 | struct perf_event_attr *attr, | |
601 | struct perf_event *group_leader) | |
602 | { | |
603 | struct perf_cgroup *cgrp; | |
604 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
605 | struct fd f = fdget(fd); |
606 | int ret = 0; | |
e5d1367f | 607 | |
2903ff01 | 608 | if (!f.file) |
e5d1367f SE |
609 | return -EBADF; |
610 | ||
5a17f543 | 611 | css = css_tryget_from_dir(f.file->f_dentry, &perf_event_cgrp_subsys); |
3db272c0 LZ |
612 | if (IS_ERR(css)) { |
613 | ret = PTR_ERR(css); | |
614 | goto out; | |
615 | } | |
e5d1367f SE |
616 | |
617 | cgrp = container_of(css, struct perf_cgroup, css); | |
618 | event->cgrp = cgrp; | |
619 | ||
620 | /* | |
621 | * all events in a group must monitor | |
622 | * the same cgroup because a task belongs | |
623 | * to only one perf cgroup at a time | |
624 | */ | |
625 | if (group_leader && group_leader->cgrp != cgrp) { | |
626 | perf_detach_cgroup(event); | |
627 | ret = -EINVAL; | |
e5d1367f | 628 | } |
3db272c0 | 629 | out: |
2903ff01 | 630 | fdput(f); |
e5d1367f SE |
631 | return ret; |
632 | } | |
633 | ||
634 | static inline void | |
635 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
636 | { | |
637 | struct perf_cgroup_info *t; | |
638 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
639 | event->shadow_ctx_time = now - t->timestamp; | |
640 | } | |
641 | ||
642 | static inline void | |
643 | perf_cgroup_defer_enabled(struct perf_event *event) | |
644 | { | |
645 | /* | |
646 | * when the current task's perf cgroup does not match | |
647 | * the event's, we need to remember to call the | |
648 | * perf_mark_enable() function the first time a task with | |
649 | * a matching perf cgroup is scheduled in. | |
650 | */ | |
651 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
652 | event->cgrp_defer_enabled = 1; | |
653 | } | |
654 | ||
655 | static inline void | |
656 | perf_cgroup_mark_enabled(struct perf_event *event, | |
657 | struct perf_event_context *ctx) | |
658 | { | |
659 | struct perf_event *sub; | |
660 | u64 tstamp = perf_event_time(event); | |
661 | ||
662 | if (!event->cgrp_defer_enabled) | |
663 | return; | |
664 | ||
665 | event->cgrp_defer_enabled = 0; | |
666 | ||
667 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
668 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
669 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
670 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
671 | sub->cgrp_defer_enabled = 0; | |
672 | } | |
673 | } | |
674 | } | |
675 | #else /* !CONFIG_CGROUP_PERF */ | |
676 | ||
677 | static inline bool | |
678 | perf_cgroup_match(struct perf_event *event) | |
679 | { | |
680 | return true; | |
681 | } | |
682 | ||
683 | static inline void perf_detach_cgroup(struct perf_event *event) | |
684 | {} | |
685 | ||
686 | static inline int is_cgroup_event(struct perf_event *event) | |
687 | { | |
688 | return 0; | |
689 | } | |
690 | ||
691 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
692 | { | |
693 | return 0; | |
694 | } | |
695 | ||
696 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
697 | { | |
698 | } | |
699 | ||
700 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
701 | { | |
702 | } | |
703 | ||
a8d757ef SE |
704 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
705 | struct task_struct *next) | |
e5d1367f SE |
706 | { |
707 | } | |
708 | ||
a8d757ef SE |
709 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
710 | struct task_struct *task) | |
e5d1367f SE |
711 | { |
712 | } | |
713 | ||
714 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
715 | struct perf_event_attr *attr, | |
716 | struct perf_event *group_leader) | |
717 | { | |
718 | return -EINVAL; | |
719 | } | |
720 | ||
721 | static inline void | |
3f7cce3c SE |
722 | perf_cgroup_set_timestamp(struct task_struct *task, |
723 | struct perf_event_context *ctx) | |
e5d1367f SE |
724 | { |
725 | } | |
726 | ||
727 | void | |
728 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
729 | { | |
730 | } | |
731 | ||
732 | static inline void | |
733 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
734 | { | |
735 | } | |
736 | ||
737 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
738 | { | |
739 | return 0; | |
740 | } | |
741 | ||
742 | static inline void | |
743 | perf_cgroup_defer_enabled(struct perf_event *event) | |
744 | { | |
745 | } | |
746 | ||
747 | static inline void | |
748 | perf_cgroup_mark_enabled(struct perf_event *event, | |
749 | struct perf_event_context *ctx) | |
750 | { | |
751 | } | |
752 | #endif | |
753 | ||
9e630205 SE |
754 | /* |
755 | * set default to be dependent on timer tick just | |
756 | * like original code | |
757 | */ | |
758 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
759 | /* | |
760 | * function must be called with interrupts disbled | |
761 | */ | |
762 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
763 | { | |
764 | struct perf_cpu_context *cpuctx; | |
765 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
766 | int rotations = 0; | |
767 | ||
768 | WARN_ON(!irqs_disabled()); | |
769 | ||
770 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
771 | ||
772 | rotations = perf_rotate_context(cpuctx); | |
773 | ||
774 | /* | |
775 | * arm timer if needed | |
776 | */ | |
777 | if (rotations) { | |
778 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
779 | ret = HRTIMER_RESTART; | |
780 | } | |
781 | ||
782 | return ret; | |
783 | } | |
784 | ||
785 | /* CPU is going down */ | |
786 | void perf_cpu_hrtimer_cancel(int cpu) | |
787 | { | |
788 | struct perf_cpu_context *cpuctx; | |
789 | struct pmu *pmu; | |
790 | unsigned long flags; | |
791 | ||
792 | if (WARN_ON(cpu != smp_processor_id())) | |
793 | return; | |
794 | ||
795 | local_irq_save(flags); | |
796 | ||
797 | rcu_read_lock(); | |
798 | ||
799 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
800 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
801 | ||
802 | if (pmu->task_ctx_nr == perf_sw_context) | |
803 | continue; | |
804 | ||
805 | hrtimer_cancel(&cpuctx->hrtimer); | |
806 | } | |
807 | ||
808 | rcu_read_unlock(); | |
809 | ||
810 | local_irq_restore(flags); | |
811 | } | |
812 | ||
813 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
814 | { | |
815 | struct hrtimer *hr = &cpuctx->hrtimer; | |
816 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 817 | int timer; |
9e630205 SE |
818 | |
819 | /* no multiplexing needed for SW PMU */ | |
820 | if (pmu->task_ctx_nr == perf_sw_context) | |
821 | return; | |
822 | ||
62b85639 SE |
823 | /* |
824 | * check default is sane, if not set then force to | |
825 | * default interval (1/tick) | |
826 | */ | |
827 | timer = pmu->hrtimer_interval_ms; | |
828 | if (timer < 1) | |
829 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
830 | ||
831 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
832 | |
833 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
834 | hr->function = perf_cpu_hrtimer_handler; | |
835 | } | |
836 | ||
837 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
838 | { | |
839 | struct hrtimer *hr = &cpuctx->hrtimer; | |
840 | struct pmu *pmu = cpuctx->ctx.pmu; | |
841 | ||
842 | /* not for SW PMU */ | |
843 | if (pmu->task_ctx_nr == perf_sw_context) | |
844 | return; | |
845 | ||
846 | if (hrtimer_active(hr)) | |
847 | return; | |
848 | ||
849 | if (!hrtimer_callback_running(hr)) | |
850 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
851 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
852 | } | |
853 | ||
33696fc0 | 854 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 855 | { |
33696fc0 PZ |
856 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
857 | if (!(*count)++) | |
858 | pmu->pmu_disable(pmu); | |
9e35ad38 | 859 | } |
9e35ad38 | 860 | |
33696fc0 | 861 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 862 | { |
33696fc0 PZ |
863 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
864 | if (!--(*count)) | |
865 | pmu->pmu_enable(pmu); | |
9e35ad38 | 866 | } |
9e35ad38 | 867 | |
e9d2b064 PZ |
868 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
869 | ||
870 | /* | |
871 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
872 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
873 | * disabled, while rotate_context is called from IRQ context. | |
874 | */ | |
108b02cf | 875 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 876 | { |
108b02cf | 877 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 878 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 879 | |
e9d2b064 | 880 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 881 | |
d84153d6 | 882 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 883 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 884 | } |
9e35ad38 | 885 | |
cdd6c482 | 886 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 887 | { |
e5289d4a | 888 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
889 | } |
890 | ||
cdd6c482 | 891 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 892 | { |
564c2b21 PM |
893 | if (atomic_dec_and_test(&ctx->refcount)) { |
894 | if (ctx->parent_ctx) | |
895 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
896 | if (ctx->task) |
897 | put_task_struct(ctx->task); | |
cb796ff3 | 898 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 899 | } |
a63eaf34 PM |
900 | } |
901 | ||
cdd6c482 | 902 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
903 | { |
904 | if (ctx->parent_ctx) { | |
905 | put_ctx(ctx->parent_ctx); | |
906 | ctx->parent_ctx = NULL; | |
907 | } | |
5a3126d4 | 908 | ctx->generation++; |
71a851b4 PZ |
909 | } |
910 | ||
6844c09d ACM |
911 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
912 | { | |
913 | /* | |
914 | * only top level events have the pid namespace they were created in | |
915 | */ | |
916 | if (event->parent) | |
917 | event = event->parent; | |
918 | ||
919 | return task_tgid_nr_ns(p, event->ns); | |
920 | } | |
921 | ||
922 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
923 | { | |
924 | /* | |
925 | * only top level events have the pid namespace they were created in | |
926 | */ | |
927 | if (event->parent) | |
928 | event = event->parent; | |
929 | ||
930 | return task_pid_nr_ns(p, event->ns); | |
931 | } | |
932 | ||
7f453c24 | 933 | /* |
cdd6c482 | 934 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
935 | * to userspace. |
936 | */ | |
cdd6c482 | 937 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 938 | { |
cdd6c482 | 939 | u64 id = event->id; |
7f453c24 | 940 | |
cdd6c482 IM |
941 | if (event->parent) |
942 | id = event->parent->id; | |
7f453c24 PZ |
943 | |
944 | return id; | |
945 | } | |
946 | ||
25346b93 | 947 | /* |
cdd6c482 | 948 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
949 | * This has to cope with with the fact that until it is locked, |
950 | * the context could get moved to another task. | |
951 | */ | |
cdd6c482 | 952 | static struct perf_event_context * |
8dc85d54 | 953 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 954 | { |
cdd6c482 | 955 | struct perf_event_context *ctx; |
25346b93 | 956 | |
9ed6060d | 957 | retry: |
058ebd0e PZ |
958 | /* |
959 | * One of the few rules of preemptible RCU is that one cannot do | |
960 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
961 | * part of the read side critical section was preemptible -- see | |
962 | * rcu_read_unlock_special(). | |
963 | * | |
964 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
965 | * side critical section is non-preemptible. | |
966 | */ | |
967 | preempt_disable(); | |
968 | rcu_read_lock(); | |
8dc85d54 | 969 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
970 | if (ctx) { |
971 | /* | |
972 | * If this context is a clone of another, it might | |
973 | * get swapped for another underneath us by | |
cdd6c482 | 974 | * perf_event_task_sched_out, though the |
25346b93 PM |
975 | * rcu_read_lock() protects us from any context |
976 | * getting freed. Lock the context and check if it | |
977 | * got swapped before we could get the lock, and retry | |
978 | * if so. If we locked the right context, then it | |
979 | * can't get swapped on us any more. | |
980 | */ | |
e625cce1 | 981 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 982 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 983 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
984 | rcu_read_unlock(); |
985 | preempt_enable(); | |
25346b93 PM |
986 | goto retry; |
987 | } | |
b49a9e7e PZ |
988 | |
989 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 990 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
991 | ctx = NULL; |
992 | } | |
25346b93 PM |
993 | } |
994 | rcu_read_unlock(); | |
058ebd0e | 995 | preempt_enable(); |
25346b93 PM |
996 | return ctx; |
997 | } | |
998 | ||
999 | /* | |
1000 | * Get the context for a task and increment its pin_count so it | |
1001 | * can't get swapped to another task. This also increments its | |
1002 | * reference count so that the context can't get freed. | |
1003 | */ | |
8dc85d54 PZ |
1004 | static struct perf_event_context * |
1005 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1006 | { |
cdd6c482 | 1007 | struct perf_event_context *ctx; |
25346b93 PM |
1008 | unsigned long flags; |
1009 | ||
8dc85d54 | 1010 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1011 | if (ctx) { |
1012 | ++ctx->pin_count; | |
e625cce1 | 1013 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1014 | } |
1015 | return ctx; | |
1016 | } | |
1017 | ||
cdd6c482 | 1018 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1019 | { |
1020 | unsigned long flags; | |
1021 | ||
e625cce1 | 1022 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1023 | --ctx->pin_count; |
e625cce1 | 1024 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1025 | } |
1026 | ||
f67218c3 PZ |
1027 | /* |
1028 | * Update the record of the current time in a context. | |
1029 | */ | |
1030 | static void update_context_time(struct perf_event_context *ctx) | |
1031 | { | |
1032 | u64 now = perf_clock(); | |
1033 | ||
1034 | ctx->time += now - ctx->timestamp; | |
1035 | ctx->timestamp = now; | |
1036 | } | |
1037 | ||
4158755d SE |
1038 | static u64 perf_event_time(struct perf_event *event) |
1039 | { | |
1040 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1041 | |
1042 | if (is_cgroup_event(event)) | |
1043 | return perf_cgroup_event_time(event); | |
1044 | ||
4158755d SE |
1045 | return ctx ? ctx->time : 0; |
1046 | } | |
1047 | ||
f67218c3 PZ |
1048 | /* |
1049 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1050 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1051 | */ |
1052 | static void update_event_times(struct perf_event *event) | |
1053 | { | |
1054 | struct perf_event_context *ctx = event->ctx; | |
1055 | u64 run_end; | |
1056 | ||
1057 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1058 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1059 | return; | |
e5d1367f SE |
1060 | /* |
1061 | * in cgroup mode, time_enabled represents | |
1062 | * the time the event was enabled AND active | |
1063 | * tasks were in the monitored cgroup. This is | |
1064 | * independent of the activity of the context as | |
1065 | * there may be a mix of cgroup and non-cgroup events. | |
1066 | * | |
1067 | * That is why we treat cgroup events differently | |
1068 | * here. | |
1069 | */ | |
1070 | if (is_cgroup_event(event)) | |
46cd6a7f | 1071 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1072 | else if (ctx->is_active) |
1073 | run_end = ctx->time; | |
acd1d7c1 PZ |
1074 | else |
1075 | run_end = event->tstamp_stopped; | |
1076 | ||
1077 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1078 | |
1079 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1080 | run_end = event->tstamp_stopped; | |
1081 | else | |
4158755d | 1082 | run_end = perf_event_time(event); |
f67218c3 PZ |
1083 | |
1084 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1085 | |
f67218c3 PZ |
1086 | } |
1087 | ||
96c21a46 PZ |
1088 | /* |
1089 | * Update total_time_enabled and total_time_running for all events in a group. | |
1090 | */ | |
1091 | static void update_group_times(struct perf_event *leader) | |
1092 | { | |
1093 | struct perf_event *event; | |
1094 | ||
1095 | update_event_times(leader); | |
1096 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1097 | update_event_times(event); | |
1098 | } | |
1099 | ||
889ff015 FW |
1100 | static struct list_head * |
1101 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1102 | { | |
1103 | if (event->attr.pinned) | |
1104 | return &ctx->pinned_groups; | |
1105 | else | |
1106 | return &ctx->flexible_groups; | |
1107 | } | |
1108 | ||
fccc714b | 1109 | /* |
cdd6c482 | 1110 | * Add a event from the lists for its context. |
fccc714b PZ |
1111 | * Must be called with ctx->mutex and ctx->lock held. |
1112 | */ | |
04289bb9 | 1113 | static void |
cdd6c482 | 1114 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1115 | { |
8a49542c PZ |
1116 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1117 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1118 | |
1119 | /* | |
8a49542c PZ |
1120 | * If we're a stand alone event or group leader, we go to the context |
1121 | * list, group events are kept attached to the group so that | |
1122 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1123 | */ |
8a49542c | 1124 | if (event->group_leader == event) { |
889ff015 FW |
1125 | struct list_head *list; |
1126 | ||
d6f962b5 FW |
1127 | if (is_software_event(event)) |
1128 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1129 | ||
889ff015 FW |
1130 | list = ctx_group_list(event, ctx); |
1131 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1132 | } |
592903cd | 1133 | |
08309379 | 1134 | if (is_cgroup_event(event)) |
e5d1367f | 1135 | ctx->nr_cgroups++; |
e5d1367f | 1136 | |
d010b332 SE |
1137 | if (has_branch_stack(event)) |
1138 | ctx->nr_branch_stack++; | |
1139 | ||
cdd6c482 | 1140 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1141 | if (!ctx->nr_events) |
108b02cf | 1142 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1143 | ctx->nr_events++; |
1144 | if (event->attr.inherit_stat) | |
bfbd3381 | 1145 | ctx->nr_stat++; |
5a3126d4 PZ |
1146 | |
1147 | ctx->generation++; | |
04289bb9 IM |
1148 | } |
1149 | ||
0231bb53 JO |
1150 | /* |
1151 | * Initialize event state based on the perf_event_attr::disabled. | |
1152 | */ | |
1153 | static inline void perf_event__state_init(struct perf_event *event) | |
1154 | { | |
1155 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1156 | PERF_EVENT_STATE_INACTIVE; | |
1157 | } | |
1158 | ||
c320c7b7 ACM |
1159 | /* |
1160 | * Called at perf_event creation and when events are attached/detached from a | |
1161 | * group. | |
1162 | */ | |
1163 | static void perf_event__read_size(struct perf_event *event) | |
1164 | { | |
1165 | int entry = sizeof(u64); /* value */ | |
1166 | int size = 0; | |
1167 | int nr = 1; | |
1168 | ||
1169 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1170 | size += sizeof(u64); | |
1171 | ||
1172 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1173 | size += sizeof(u64); | |
1174 | ||
1175 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1176 | entry += sizeof(u64); | |
1177 | ||
1178 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1179 | nr += event->group_leader->nr_siblings; | |
1180 | size += sizeof(u64); | |
1181 | } | |
1182 | ||
1183 | size += entry * nr; | |
1184 | event->read_size = size; | |
1185 | } | |
1186 | ||
1187 | static void perf_event__header_size(struct perf_event *event) | |
1188 | { | |
1189 | struct perf_sample_data *data; | |
1190 | u64 sample_type = event->attr.sample_type; | |
1191 | u16 size = 0; | |
1192 | ||
1193 | perf_event__read_size(event); | |
1194 | ||
1195 | if (sample_type & PERF_SAMPLE_IP) | |
1196 | size += sizeof(data->ip); | |
1197 | ||
6844c09d ACM |
1198 | if (sample_type & PERF_SAMPLE_ADDR) |
1199 | size += sizeof(data->addr); | |
1200 | ||
1201 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1202 | size += sizeof(data->period); | |
1203 | ||
c3feedf2 AK |
1204 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1205 | size += sizeof(data->weight); | |
1206 | ||
6844c09d ACM |
1207 | if (sample_type & PERF_SAMPLE_READ) |
1208 | size += event->read_size; | |
1209 | ||
d6be9ad6 SE |
1210 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1211 | size += sizeof(data->data_src.val); | |
1212 | ||
fdfbbd07 AK |
1213 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1214 | size += sizeof(data->txn); | |
1215 | ||
6844c09d ACM |
1216 | event->header_size = size; |
1217 | } | |
1218 | ||
1219 | static void perf_event__id_header_size(struct perf_event *event) | |
1220 | { | |
1221 | struct perf_sample_data *data; | |
1222 | u64 sample_type = event->attr.sample_type; | |
1223 | u16 size = 0; | |
1224 | ||
c320c7b7 ACM |
1225 | if (sample_type & PERF_SAMPLE_TID) |
1226 | size += sizeof(data->tid_entry); | |
1227 | ||
1228 | if (sample_type & PERF_SAMPLE_TIME) | |
1229 | size += sizeof(data->time); | |
1230 | ||
ff3d527c AH |
1231 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1232 | size += sizeof(data->id); | |
1233 | ||
c320c7b7 ACM |
1234 | if (sample_type & PERF_SAMPLE_ID) |
1235 | size += sizeof(data->id); | |
1236 | ||
1237 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1238 | size += sizeof(data->stream_id); | |
1239 | ||
1240 | if (sample_type & PERF_SAMPLE_CPU) | |
1241 | size += sizeof(data->cpu_entry); | |
1242 | ||
6844c09d | 1243 | event->id_header_size = size; |
c320c7b7 ACM |
1244 | } |
1245 | ||
8a49542c PZ |
1246 | static void perf_group_attach(struct perf_event *event) |
1247 | { | |
c320c7b7 | 1248 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1249 | |
74c3337c PZ |
1250 | /* |
1251 | * We can have double attach due to group movement in perf_event_open. | |
1252 | */ | |
1253 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1254 | return; | |
1255 | ||
8a49542c PZ |
1256 | event->attach_state |= PERF_ATTACH_GROUP; |
1257 | ||
1258 | if (group_leader == event) | |
1259 | return; | |
1260 | ||
1261 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1262 | !is_software_event(event)) | |
1263 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1264 | ||
1265 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1266 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1267 | |
1268 | perf_event__header_size(group_leader); | |
1269 | ||
1270 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1271 | perf_event__header_size(pos); | |
8a49542c PZ |
1272 | } |
1273 | ||
a63eaf34 | 1274 | /* |
cdd6c482 | 1275 | * Remove a event from the lists for its context. |
fccc714b | 1276 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1277 | */ |
04289bb9 | 1278 | static void |
cdd6c482 | 1279 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1280 | { |
68cacd29 | 1281 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1282 | /* |
1283 | * We can have double detach due to exit/hot-unplug + close. | |
1284 | */ | |
1285 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1286 | return; |
8a49542c PZ |
1287 | |
1288 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1289 | ||
68cacd29 | 1290 | if (is_cgroup_event(event)) { |
e5d1367f | 1291 | ctx->nr_cgroups--; |
68cacd29 SE |
1292 | cpuctx = __get_cpu_context(ctx); |
1293 | /* | |
1294 | * if there are no more cgroup events | |
1295 | * then cler cgrp to avoid stale pointer | |
1296 | * in update_cgrp_time_from_cpuctx() | |
1297 | */ | |
1298 | if (!ctx->nr_cgroups) | |
1299 | cpuctx->cgrp = NULL; | |
1300 | } | |
e5d1367f | 1301 | |
d010b332 SE |
1302 | if (has_branch_stack(event)) |
1303 | ctx->nr_branch_stack--; | |
1304 | ||
cdd6c482 IM |
1305 | ctx->nr_events--; |
1306 | if (event->attr.inherit_stat) | |
bfbd3381 | 1307 | ctx->nr_stat--; |
8bc20959 | 1308 | |
cdd6c482 | 1309 | list_del_rcu(&event->event_entry); |
04289bb9 | 1310 | |
8a49542c PZ |
1311 | if (event->group_leader == event) |
1312 | list_del_init(&event->group_entry); | |
5c148194 | 1313 | |
96c21a46 | 1314 | update_group_times(event); |
b2e74a26 SE |
1315 | |
1316 | /* | |
1317 | * If event was in error state, then keep it | |
1318 | * that way, otherwise bogus counts will be | |
1319 | * returned on read(). The only way to get out | |
1320 | * of error state is by explicit re-enabling | |
1321 | * of the event | |
1322 | */ | |
1323 | if (event->state > PERF_EVENT_STATE_OFF) | |
1324 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1325 | |
1326 | ctx->generation++; | |
050735b0 PZ |
1327 | } |
1328 | ||
8a49542c | 1329 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1330 | { |
1331 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1332 | struct list_head *list = NULL; |
1333 | ||
1334 | /* | |
1335 | * We can have double detach due to exit/hot-unplug + close. | |
1336 | */ | |
1337 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1338 | return; | |
1339 | ||
1340 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1341 | ||
1342 | /* | |
1343 | * If this is a sibling, remove it from its group. | |
1344 | */ | |
1345 | if (event->group_leader != event) { | |
1346 | list_del_init(&event->group_entry); | |
1347 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1348 | goto out; |
8a49542c PZ |
1349 | } |
1350 | ||
1351 | if (!list_empty(&event->group_entry)) | |
1352 | list = &event->group_entry; | |
2e2af50b | 1353 | |
04289bb9 | 1354 | /* |
cdd6c482 IM |
1355 | * If this was a group event with sibling events then |
1356 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1357 | * to whatever list we are on. |
04289bb9 | 1358 | */ |
cdd6c482 | 1359 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1360 | if (list) |
1361 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1362 | sibling->group_leader = sibling; |
d6f962b5 FW |
1363 | |
1364 | /* Inherit group flags from the previous leader */ | |
1365 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1366 | } |
c320c7b7 ACM |
1367 | |
1368 | out: | |
1369 | perf_event__header_size(event->group_leader); | |
1370 | ||
1371 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1372 | perf_event__header_size(tmp); | |
04289bb9 IM |
1373 | } |
1374 | ||
fa66f07a SE |
1375 | static inline int |
1376 | event_filter_match(struct perf_event *event) | |
1377 | { | |
e5d1367f SE |
1378 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1379 | && perf_cgroup_match(event); | |
fa66f07a SE |
1380 | } |
1381 | ||
9ffcfa6f SE |
1382 | static void |
1383 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1384 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1385 | struct perf_event_context *ctx) |
3b6f9e5c | 1386 | { |
4158755d | 1387 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1388 | u64 delta; |
1389 | /* | |
1390 | * An event which could not be activated because of | |
1391 | * filter mismatch still needs to have its timings | |
1392 | * maintained, otherwise bogus information is return | |
1393 | * via read() for time_enabled, time_running: | |
1394 | */ | |
1395 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1396 | && !event_filter_match(event)) { | |
e5d1367f | 1397 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1398 | event->tstamp_running += delta; |
4158755d | 1399 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1400 | } |
1401 | ||
cdd6c482 | 1402 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1403 | return; |
3b6f9e5c | 1404 | |
44377277 AS |
1405 | perf_pmu_disable(event->pmu); |
1406 | ||
cdd6c482 IM |
1407 | event->state = PERF_EVENT_STATE_INACTIVE; |
1408 | if (event->pending_disable) { | |
1409 | event->pending_disable = 0; | |
1410 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1411 | } |
4158755d | 1412 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1413 | event->pmu->del(event, 0); |
cdd6c482 | 1414 | event->oncpu = -1; |
3b6f9e5c | 1415 | |
cdd6c482 | 1416 | if (!is_software_event(event)) |
3b6f9e5c PM |
1417 | cpuctx->active_oncpu--; |
1418 | ctx->nr_active--; | |
0f5a2601 PZ |
1419 | if (event->attr.freq && event->attr.sample_freq) |
1420 | ctx->nr_freq--; | |
cdd6c482 | 1421 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1422 | cpuctx->exclusive = 0; |
44377277 AS |
1423 | |
1424 | perf_pmu_enable(event->pmu); | |
3b6f9e5c PM |
1425 | } |
1426 | ||
d859e29f | 1427 | static void |
cdd6c482 | 1428 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1429 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1430 | struct perf_event_context *ctx) |
d859e29f | 1431 | { |
cdd6c482 | 1432 | struct perf_event *event; |
fa66f07a | 1433 | int state = group_event->state; |
d859e29f | 1434 | |
cdd6c482 | 1435 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1436 | |
1437 | /* | |
1438 | * Schedule out siblings (if any): | |
1439 | */ | |
cdd6c482 IM |
1440 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1441 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1442 | |
fa66f07a | 1443 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1444 | cpuctx->exclusive = 0; |
1445 | } | |
1446 | ||
46ce0fe9 PZ |
1447 | struct remove_event { |
1448 | struct perf_event *event; | |
1449 | bool detach_group; | |
1450 | }; | |
1451 | ||
0793a61d | 1452 | /* |
cdd6c482 | 1453 | * Cross CPU call to remove a performance event |
0793a61d | 1454 | * |
cdd6c482 | 1455 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1456 | * remove it from the context list. |
1457 | */ | |
fe4b04fa | 1458 | static int __perf_remove_from_context(void *info) |
0793a61d | 1459 | { |
46ce0fe9 PZ |
1460 | struct remove_event *re = info; |
1461 | struct perf_event *event = re->event; | |
cdd6c482 | 1462 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1463 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1464 | |
e625cce1 | 1465 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1466 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1467 | if (re->detach_group) |
1468 | perf_group_detach(event); | |
cdd6c482 | 1469 | list_del_event(event, ctx); |
64ce3126 PZ |
1470 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1471 | ctx->is_active = 0; | |
1472 | cpuctx->task_ctx = NULL; | |
1473 | } | |
e625cce1 | 1474 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1475 | |
1476 | return 0; | |
0793a61d TG |
1477 | } |
1478 | ||
1479 | ||
1480 | /* | |
cdd6c482 | 1481 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1482 | * |
cdd6c482 | 1483 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1484 | * call when the task is on a CPU. |
c93f7669 | 1485 | * |
cdd6c482 IM |
1486 | * If event->ctx is a cloned context, callers must make sure that |
1487 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1488 | * remains valid. This is OK when called from perf_release since |
1489 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1490 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1491 | * context has been detached from its task. |
0793a61d | 1492 | */ |
46ce0fe9 | 1493 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1494 | { |
cdd6c482 | 1495 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1496 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1497 | struct remove_event re = { |
1498 | .event = event, | |
1499 | .detach_group = detach_group, | |
1500 | }; | |
0793a61d | 1501 | |
fe4b04fa PZ |
1502 | lockdep_assert_held(&ctx->mutex); |
1503 | ||
0793a61d TG |
1504 | if (!task) { |
1505 | /* | |
cdd6c482 | 1506 | * Per cpu events are removed via an smp call and |
af901ca1 | 1507 | * the removal is always successful. |
0793a61d | 1508 | */ |
46ce0fe9 | 1509 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1510 | return; |
1511 | } | |
1512 | ||
1513 | retry: | |
46ce0fe9 | 1514 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1515 | return; |
0793a61d | 1516 | |
e625cce1 | 1517 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1518 | /* |
fe4b04fa PZ |
1519 | * If we failed to find a running task, but find the context active now |
1520 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1521 | */ |
fe4b04fa | 1522 | if (ctx->is_active) { |
e625cce1 | 1523 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1524 | goto retry; |
1525 | } | |
1526 | ||
1527 | /* | |
fe4b04fa PZ |
1528 | * Since the task isn't running, its safe to remove the event, us |
1529 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1530 | */ |
46ce0fe9 PZ |
1531 | if (detach_group) |
1532 | perf_group_detach(event); | |
fe4b04fa | 1533 | list_del_event(event, ctx); |
e625cce1 | 1534 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1535 | } |
1536 | ||
d859e29f | 1537 | /* |
cdd6c482 | 1538 | * Cross CPU call to disable a performance event |
d859e29f | 1539 | */ |
500ad2d8 | 1540 | int __perf_event_disable(void *info) |
d859e29f | 1541 | { |
cdd6c482 | 1542 | struct perf_event *event = info; |
cdd6c482 | 1543 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1544 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1545 | |
1546 | /* | |
cdd6c482 IM |
1547 | * If this is a per-task event, need to check whether this |
1548 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1549 | * |
1550 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1551 | * flipping contexts around. | |
d859e29f | 1552 | */ |
665c2142 | 1553 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1554 | return -EINVAL; |
d859e29f | 1555 | |
e625cce1 | 1556 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1557 | |
1558 | /* | |
cdd6c482 | 1559 | * If the event is on, turn it off. |
d859e29f PM |
1560 | * If it is in error state, leave it in error state. |
1561 | */ | |
cdd6c482 | 1562 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1563 | update_context_time(ctx); |
e5d1367f | 1564 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1565 | update_group_times(event); |
1566 | if (event == event->group_leader) | |
1567 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1568 | else |
cdd6c482 IM |
1569 | event_sched_out(event, cpuctx, ctx); |
1570 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1571 | } |
1572 | ||
e625cce1 | 1573 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1574 | |
1575 | return 0; | |
d859e29f PM |
1576 | } |
1577 | ||
1578 | /* | |
cdd6c482 | 1579 | * Disable a event. |
c93f7669 | 1580 | * |
cdd6c482 IM |
1581 | * If event->ctx is a cloned context, callers must make sure that |
1582 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1583 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1584 | * perf_event_for_each_child or perf_event_for_each because they |
1585 | * hold the top-level event's child_mutex, so any descendant that | |
1586 | * goes to exit will block in sync_child_event. | |
1587 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1588 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1589 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1590 | */ |
44234adc | 1591 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1592 | { |
cdd6c482 | 1593 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1594 | struct task_struct *task = ctx->task; |
1595 | ||
1596 | if (!task) { | |
1597 | /* | |
cdd6c482 | 1598 | * Disable the event on the cpu that it's on |
d859e29f | 1599 | */ |
fe4b04fa | 1600 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1601 | return; |
1602 | } | |
1603 | ||
9ed6060d | 1604 | retry: |
fe4b04fa PZ |
1605 | if (!task_function_call(task, __perf_event_disable, event)) |
1606 | return; | |
d859e29f | 1607 | |
e625cce1 | 1608 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1609 | /* |
cdd6c482 | 1610 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1611 | */ |
cdd6c482 | 1612 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1613 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1614 | /* |
1615 | * Reload the task pointer, it might have been changed by | |
1616 | * a concurrent perf_event_context_sched_out(). | |
1617 | */ | |
1618 | task = ctx->task; | |
d859e29f PM |
1619 | goto retry; |
1620 | } | |
1621 | ||
1622 | /* | |
1623 | * Since we have the lock this context can't be scheduled | |
1624 | * in, so we can change the state safely. | |
1625 | */ | |
cdd6c482 IM |
1626 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1627 | update_group_times(event); | |
1628 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1629 | } |
e625cce1 | 1630 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1631 | } |
dcfce4a0 | 1632 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1633 | |
e5d1367f SE |
1634 | static void perf_set_shadow_time(struct perf_event *event, |
1635 | struct perf_event_context *ctx, | |
1636 | u64 tstamp) | |
1637 | { | |
1638 | /* | |
1639 | * use the correct time source for the time snapshot | |
1640 | * | |
1641 | * We could get by without this by leveraging the | |
1642 | * fact that to get to this function, the caller | |
1643 | * has most likely already called update_context_time() | |
1644 | * and update_cgrp_time_xx() and thus both timestamp | |
1645 | * are identical (or very close). Given that tstamp is, | |
1646 | * already adjusted for cgroup, we could say that: | |
1647 | * tstamp - ctx->timestamp | |
1648 | * is equivalent to | |
1649 | * tstamp - cgrp->timestamp. | |
1650 | * | |
1651 | * Then, in perf_output_read(), the calculation would | |
1652 | * work with no changes because: | |
1653 | * - event is guaranteed scheduled in | |
1654 | * - no scheduled out in between | |
1655 | * - thus the timestamp would be the same | |
1656 | * | |
1657 | * But this is a bit hairy. | |
1658 | * | |
1659 | * So instead, we have an explicit cgroup call to remain | |
1660 | * within the time time source all along. We believe it | |
1661 | * is cleaner and simpler to understand. | |
1662 | */ | |
1663 | if (is_cgroup_event(event)) | |
1664 | perf_cgroup_set_shadow_time(event, tstamp); | |
1665 | else | |
1666 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1667 | } | |
1668 | ||
4fe757dd PZ |
1669 | #define MAX_INTERRUPTS (~0ULL) |
1670 | ||
1671 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1672 | ||
235c7fc7 | 1673 | static int |
9ffcfa6f | 1674 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1675 | struct perf_cpu_context *cpuctx, |
6e37738a | 1676 | struct perf_event_context *ctx) |
235c7fc7 | 1677 | { |
4158755d | 1678 | u64 tstamp = perf_event_time(event); |
44377277 | 1679 | int ret = 0; |
4158755d | 1680 | |
63342411 PZ |
1681 | lockdep_assert_held(&ctx->lock); |
1682 | ||
cdd6c482 | 1683 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1684 | return 0; |
1685 | ||
cdd6c482 | 1686 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1687 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1688 | |
1689 | /* | |
1690 | * Unthrottle events, since we scheduled we might have missed several | |
1691 | * ticks already, also for a heavily scheduling task there is little | |
1692 | * guarantee it'll get a tick in a timely manner. | |
1693 | */ | |
1694 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1695 | perf_log_throttle(event, 1); | |
1696 | event->hw.interrupts = 0; | |
1697 | } | |
1698 | ||
235c7fc7 IM |
1699 | /* |
1700 | * The new state must be visible before we turn it on in the hardware: | |
1701 | */ | |
1702 | smp_wmb(); | |
1703 | ||
44377277 AS |
1704 | perf_pmu_disable(event->pmu); |
1705 | ||
a4eaf7f1 | 1706 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1707 | event->state = PERF_EVENT_STATE_INACTIVE; |
1708 | event->oncpu = -1; | |
44377277 AS |
1709 | ret = -EAGAIN; |
1710 | goto out; | |
235c7fc7 IM |
1711 | } |
1712 | ||
4158755d | 1713 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1714 | |
e5d1367f | 1715 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1716 | |
cdd6c482 | 1717 | if (!is_software_event(event)) |
3b6f9e5c | 1718 | cpuctx->active_oncpu++; |
235c7fc7 | 1719 | ctx->nr_active++; |
0f5a2601 PZ |
1720 | if (event->attr.freq && event->attr.sample_freq) |
1721 | ctx->nr_freq++; | |
235c7fc7 | 1722 | |
cdd6c482 | 1723 | if (event->attr.exclusive) |
3b6f9e5c PM |
1724 | cpuctx->exclusive = 1; |
1725 | ||
44377277 AS |
1726 | out: |
1727 | perf_pmu_enable(event->pmu); | |
1728 | ||
1729 | return ret; | |
235c7fc7 IM |
1730 | } |
1731 | ||
6751b71e | 1732 | static int |
cdd6c482 | 1733 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1734 | struct perf_cpu_context *cpuctx, |
6e37738a | 1735 | struct perf_event_context *ctx) |
6751b71e | 1736 | { |
6bde9b6c | 1737 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1738 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1739 | u64 now = ctx->time; |
1740 | bool simulate = false; | |
6751b71e | 1741 | |
cdd6c482 | 1742 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1743 | return 0; |
1744 | ||
ad5133b7 | 1745 | pmu->start_txn(pmu); |
6bde9b6c | 1746 | |
9ffcfa6f | 1747 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1748 | pmu->cancel_txn(pmu); |
9e630205 | 1749 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1750 | return -EAGAIN; |
90151c35 | 1751 | } |
6751b71e PM |
1752 | |
1753 | /* | |
1754 | * Schedule in siblings as one group (if any): | |
1755 | */ | |
cdd6c482 | 1756 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1757 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1758 | partial_group = event; |
6751b71e PM |
1759 | goto group_error; |
1760 | } | |
1761 | } | |
1762 | ||
9ffcfa6f | 1763 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1764 | return 0; |
9ffcfa6f | 1765 | |
6751b71e PM |
1766 | group_error: |
1767 | /* | |
1768 | * Groups can be scheduled in as one unit only, so undo any | |
1769 | * partial group before returning: | |
d7842da4 SE |
1770 | * The events up to the failed event are scheduled out normally, |
1771 | * tstamp_stopped will be updated. | |
1772 | * | |
1773 | * The failed events and the remaining siblings need to have | |
1774 | * their timings updated as if they had gone thru event_sched_in() | |
1775 | * and event_sched_out(). This is required to get consistent timings | |
1776 | * across the group. This also takes care of the case where the group | |
1777 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1778 | * the time the event was actually stopped, such that time delta | |
1779 | * calculation in update_event_times() is correct. | |
6751b71e | 1780 | */ |
cdd6c482 IM |
1781 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1782 | if (event == partial_group) | |
d7842da4 SE |
1783 | simulate = true; |
1784 | ||
1785 | if (simulate) { | |
1786 | event->tstamp_running += now - event->tstamp_stopped; | |
1787 | event->tstamp_stopped = now; | |
1788 | } else { | |
1789 | event_sched_out(event, cpuctx, ctx); | |
1790 | } | |
6751b71e | 1791 | } |
9ffcfa6f | 1792 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1793 | |
ad5133b7 | 1794 | pmu->cancel_txn(pmu); |
90151c35 | 1795 | |
9e630205 SE |
1796 | perf_cpu_hrtimer_restart(cpuctx); |
1797 | ||
6751b71e PM |
1798 | return -EAGAIN; |
1799 | } | |
1800 | ||
3b6f9e5c | 1801 | /* |
cdd6c482 | 1802 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1803 | */ |
cdd6c482 | 1804 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1805 | struct perf_cpu_context *cpuctx, |
1806 | int can_add_hw) | |
1807 | { | |
1808 | /* | |
cdd6c482 | 1809 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1810 | */ |
d6f962b5 | 1811 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1812 | return 1; |
1813 | /* | |
1814 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1815 | * events can go on. |
3b6f9e5c PM |
1816 | */ |
1817 | if (cpuctx->exclusive) | |
1818 | return 0; | |
1819 | /* | |
1820 | * If this group is exclusive and there are already | |
cdd6c482 | 1821 | * events on the CPU, it can't go on. |
3b6f9e5c | 1822 | */ |
cdd6c482 | 1823 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1824 | return 0; |
1825 | /* | |
1826 | * Otherwise, try to add it if all previous groups were able | |
1827 | * to go on. | |
1828 | */ | |
1829 | return can_add_hw; | |
1830 | } | |
1831 | ||
cdd6c482 IM |
1832 | static void add_event_to_ctx(struct perf_event *event, |
1833 | struct perf_event_context *ctx) | |
53cfbf59 | 1834 | { |
4158755d SE |
1835 | u64 tstamp = perf_event_time(event); |
1836 | ||
cdd6c482 | 1837 | list_add_event(event, ctx); |
8a49542c | 1838 | perf_group_attach(event); |
4158755d SE |
1839 | event->tstamp_enabled = tstamp; |
1840 | event->tstamp_running = tstamp; | |
1841 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1842 | } |
1843 | ||
2c29ef0f PZ |
1844 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1845 | static void | |
1846 | ctx_sched_in(struct perf_event_context *ctx, | |
1847 | struct perf_cpu_context *cpuctx, | |
1848 | enum event_type_t event_type, | |
1849 | struct task_struct *task); | |
fe4b04fa | 1850 | |
dce5855b PZ |
1851 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1852 | struct perf_event_context *ctx, | |
1853 | struct task_struct *task) | |
1854 | { | |
1855 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1856 | if (ctx) | |
1857 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1858 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1859 | if (ctx) | |
1860 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1861 | } | |
1862 | ||
0793a61d | 1863 | /* |
cdd6c482 | 1864 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1865 | * |
1866 | * Must be called with ctx->mutex held | |
0793a61d | 1867 | */ |
fe4b04fa | 1868 | static int __perf_install_in_context(void *info) |
0793a61d | 1869 | { |
cdd6c482 IM |
1870 | struct perf_event *event = info; |
1871 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1872 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1873 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1874 | struct task_struct *task = current; | |
1875 | ||
b58f6b0d | 1876 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1877 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1878 | |
1879 | /* | |
2c29ef0f | 1880 | * If there was an active task_ctx schedule it out. |
0793a61d | 1881 | */ |
b58f6b0d | 1882 | if (task_ctx) |
2c29ef0f | 1883 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1884 | |
1885 | /* | |
1886 | * If the context we're installing events in is not the | |
1887 | * active task_ctx, flip them. | |
1888 | */ | |
1889 | if (ctx->task && task_ctx != ctx) { | |
1890 | if (task_ctx) | |
1891 | raw_spin_unlock(&task_ctx->lock); | |
1892 | raw_spin_lock(&ctx->lock); | |
1893 | task_ctx = ctx; | |
1894 | } | |
1895 | ||
1896 | if (task_ctx) { | |
1897 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1898 | task = task_ctx->task; |
1899 | } | |
b58f6b0d | 1900 | |
2c29ef0f | 1901 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1902 | |
4af4998b | 1903 | update_context_time(ctx); |
e5d1367f SE |
1904 | /* |
1905 | * update cgrp time only if current cgrp | |
1906 | * matches event->cgrp. Must be done before | |
1907 | * calling add_event_to_ctx() | |
1908 | */ | |
1909 | update_cgrp_time_from_event(event); | |
0793a61d | 1910 | |
cdd6c482 | 1911 | add_event_to_ctx(event, ctx); |
0793a61d | 1912 | |
d859e29f | 1913 | /* |
2c29ef0f | 1914 | * Schedule everything back in |
d859e29f | 1915 | */ |
dce5855b | 1916 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1917 | |
1918 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1919 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1920 | |
1921 | return 0; | |
0793a61d TG |
1922 | } |
1923 | ||
1924 | /* | |
cdd6c482 | 1925 | * Attach a performance event to a context |
0793a61d | 1926 | * |
cdd6c482 IM |
1927 | * First we add the event to the list with the hardware enable bit |
1928 | * in event->hw_config cleared. | |
0793a61d | 1929 | * |
cdd6c482 | 1930 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1931 | * call to enable it in the task context. The task might have been |
1932 | * scheduled away, but we check this in the smp call again. | |
1933 | */ | |
1934 | static void | |
cdd6c482 IM |
1935 | perf_install_in_context(struct perf_event_context *ctx, |
1936 | struct perf_event *event, | |
0793a61d TG |
1937 | int cpu) |
1938 | { | |
1939 | struct task_struct *task = ctx->task; | |
1940 | ||
fe4b04fa PZ |
1941 | lockdep_assert_held(&ctx->mutex); |
1942 | ||
c3f00c70 | 1943 | event->ctx = ctx; |
0cda4c02 YZ |
1944 | if (event->cpu != -1) |
1945 | event->cpu = cpu; | |
c3f00c70 | 1946 | |
0793a61d TG |
1947 | if (!task) { |
1948 | /* | |
cdd6c482 | 1949 | * Per cpu events are installed via an smp call and |
af901ca1 | 1950 | * the install is always successful. |
0793a61d | 1951 | */ |
fe4b04fa | 1952 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1953 | return; |
1954 | } | |
1955 | ||
0793a61d | 1956 | retry: |
fe4b04fa PZ |
1957 | if (!task_function_call(task, __perf_install_in_context, event)) |
1958 | return; | |
0793a61d | 1959 | |
e625cce1 | 1960 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1961 | /* |
fe4b04fa PZ |
1962 | * If we failed to find a running task, but find the context active now |
1963 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1964 | */ |
fe4b04fa | 1965 | if (ctx->is_active) { |
e625cce1 | 1966 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1967 | goto retry; |
1968 | } | |
1969 | ||
1970 | /* | |
fe4b04fa PZ |
1971 | * Since the task isn't running, its safe to add the event, us holding |
1972 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1973 | */ |
fe4b04fa | 1974 | add_event_to_ctx(event, ctx); |
e625cce1 | 1975 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1976 | } |
1977 | ||
fa289bec | 1978 | /* |
cdd6c482 | 1979 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1980 | * Enabling the leader of a group effectively enables all |
1981 | * the group members that aren't explicitly disabled, so we | |
1982 | * have to update their ->tstamp_enabled also. | |
1983 | * Note: this works for group members as well as group leaders | |
1984 | * since the non-leader members' sibling_lists will be empty. | |
1985 | */ | |
1d9b482e | 1986 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1987 | { |
cdd6c482 | 1988 | struct perf_event *sub; |
4158755d | 1989 | u64 tstamp = perf_event_time(event); |
fa289bec | 1990 | |
cdd6c482 | 1991 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1992 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1993 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1994 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1995 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1996 | } |
fa289bec PM |
1997 | } |
1998 | ||
d859e29f | 1999 | /* |
cdd6c482 | 2000 | * Cross CPU call to enable a performance event |
d859e29f | 2001 | */ |
fe4b04fa | 2002 | static int __perf_event_enable(void *info) |
04289bb9 | 2003 | { |
cdd6c482 | 2004 | struct perf_event *event = info; |
cdd6c482 IM |
2005 | struct perf_event_context *ctx = event->ctx; |
2006 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2007 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2008 | int err; |
04289bb9 | 2009 | |
06f41796 JO |
2010 | /* |
2011 | * There's a time window between 'ctx->is_active' check | |
2012 | * in perf_event_enable function and this place having: | |
2013 | * - IRQs on | |
2014 | * - ctx->lock unlocked | |
2015 | * | |
2016 | * where the task could be killed and 'ctx' deactivated | |
2017 | * by perf_event_exit_task. | |
2018 | */ | |
2019 | if (!ctx->is_active) | |
fe4b04fa | 2020 | return -EINVAL; |
3cbed429 | 2021 | |
e625cce1 | 2022 | raw_spin_lock(&ctx->lock); |
4af4998b | 2023 | update_context_time(ctx); |
d859e29f | 2024 | |
cdd6c482 | 2025 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2026 | goto unlock; |
e5d1367f SE |
2027 | |
2028 | /* | |
2029 | * set current task's cgroup time reference point | |
2030 | */ | |
3f7cce3c | 2031 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2032 | |
1d9b482e | 2033 | __perf_event_mark_enabled(event); |
04289bb9 | 2034 | |
e5d1367f SE |
2035 | if (!event_filter_match(event)) { |
2036 | if (is_cgroup_event(event)) | |
2037 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2038 | goto unlock; |
e5d1367f | 2039 | } |
f4c4176f | 2040 | |
04289bb9 | 2041 | /* |
cdd6c482 | 2042 | * If the event is in a group and isn't the group leader, |
d859e29f | 2043 | * then don't put it on unless the group is on. |
04289bb9 | 2044 | */ |
cdd6c482 | 2045 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2046 | goto unlock; |
3b6f9e5c | 2047 | |
cdd6c482 | 2048 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2049 | err = -EEXIST; |
e758a33d | 2050 | } else { |
cdd6c482 | 2051 | if (event == leader) |
6e37738a | 2052 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2053 | else |
6e37738a | 2054 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2055 | } |
d859e29f PM |
2056 | |
2057 | if (err) { | |
2058 | /* | |
cdd6c482 | 2059 | * If this event can't go on and it's part of a |
d859e29f PM |
2060 | * group, then the whole group has to come off. |
2061 | */ | |
9e630205 | 2062 | if (leader != event) { |
d859e29f | 2063 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2064 | perf_cpu_hrtimer_restart(cpuctx); |
2065 | } | |
0d48696f | 2066 | if (leader->attr.pinned) { |
53cfbf59 | 2067 | update_group_times(leader); |
cdd6c482 | 2068 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2069 | } |
d859e29f PM |
2070 | } |
2071 | ||
9ed6060d | 2072 | unlock: |
e625cce1 | 2073 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2074 | |
2075 | return 0; | |
d859e29f PM |
2076 | } |
2077 | ||
2078 | /* | |
cdd6c482 | 2079 | * Enable a event. |
c93f7669 | 2080 | * |
cdd6c482 IM |
2081 | * If event->ctx is a cloned context, callers must make sure that |
2082 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2083 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2084 | * perf_event_for_each_child or perf_event_for_each as described |
2085 | * for perf_event_disable. | |
d859e29f | 2086 | */ |
44234adc | 2087 | void perf_event_enable(struct perf_event *event) |
d859e29f | 2088 | { |
cdd6c482 | 2089 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2090 | struct task_struct *task = ctx->task; |
2091 | ||
2092 | if (!task) { | |
2093 | /* | |
cdd6c482 | 2094 | * Enable the event on the cpu that it's on |
d859e29f | 2095 | */ |
fe4b04fa | 2096 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2097 | return; |
2098 | } | |
2099 | ||
e625cce1 | 2100 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2101 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2102 | goto out; |
2103 | ||
2104 | /* | |
cdd6c482 IM |
2105 | * If the event is in error state, clear that first. |
2106 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2107 | * know that it has gone back into error state, as distinct |
2108 | * from the task having been scheduled away before the | |
2109 | * cross-call arrived. | |
2110 | */ | |
cdd6c482 IM |
2111 | if (event->state == PERF_EVENT_STATE_ERROR) |
2112 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2113 | |
9ed6060d | 2114 | retry: |
fe4b04fa | 2115 | if (!ctx->is_active) { |
1d9b482e | 2116 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2117 | goto out; |
2118 | } | |
2119 | ||
e625cce1 | 2120 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2121 | |
2122 | if (!task_function_call(task, __perf_event_enable, event)) | |
2123 | return; | |
d859e29f | 2124 | |
e625cce1 | 2125 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2126 | |
2127 | /* | |
cdd6c482 | 2128 | * If the context is active and the event is still off, |
d859e29f PM |
2129 | * we need to retry the cross-call. |
2130 | */ | |
fe4b04fa PZ |
2131 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2132 | /* | |
2133 | * task could have been flipped by a concurrent | |
2134 | * perf_event_context_sched_out() | |
2135 | */ | |
2136 | task = ctx->task; | |
d859e29f | 2137 | goto retry; |
fe4b04fa | 2138 | } |
fa289bec | 2139 | |
9ed6060d | 2140 | out: |
e625cce1 | 2141 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2142 | } |
dcfce4a0 | 2143 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2144 | |
26ca5c11 | 2145 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2146 | { |
2023b359 | 2147 | /* |
cdd6c482 | 2148 | * not supported on inherited events |
2023b359 | 2149 | */ |
2e939d1d | 2150 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2151 | return -EINVAL; |
2152 | ||
cdd6c482 IM |
2153 | atomic_add(refresh, &event->event_limit); |
2154 | perf_event_enable(event); | |
2023b359 PZ |
2155 | |
2156 | return 0; | |
79f14641 | 2157 | } |
26ca5c11 | 2158 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2159 | |
5b0311e1 FW |
2160 | static void ctx_sched_out(struct perf_event_context *ctx, |
2161 | struct perf_cpu_context *cpuctx, | |
2162 | enum event_type_t event_type) | |
235c7fc7 | 2163 | { |
cdd6c482 | 2164 | struct perf_event *event; |
db24d33e | 2165 | int is_active = ctx->is_active; |
235c7fc7 | 2166 | |
db24d33e | 2167 | ctx->is_active &= ~event_type; |
cdd6c482 | 2168 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2169 | return; |
2170 | ||
4af4998b | 2171 | update_context_time(ctx); |
e5d1367f | 2172 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2173 | if (!ctx->nr_active) |
facc4307 | 2174 | return; |
5b0311e1 | 2175 | |
075e0b00 | 2176 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2177 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2178 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2179 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2180 | } |
889ff015 | 2181 | |
db24d33e | 2182 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2183 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2184 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2185 | } |
1b9a644f | 2186 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2187 | } |
2188 | ||
564c2b21 | 2189 | /* |
5a3126d4 PZ |
2190 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2191 | * cloned from the same version of the same context. | |
2192 | * | |
2193 | * Equivalence is measured using a generation number in the context that is | |
2194 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2195 | * and list_del_event(). | |
564c2b21 | 2196 | */ |
cdd6c482 IM |
2197 | static int context_equiv(struct perf_event_context *ctx1, |
2198 | struct perf_event_context *ctx2) | |
564c2b21 | 2199 | { |
5a3126d4 PZ |
2200 | /* Pinning disables the swap optimization */ |
2201 | if (ctx1->pin_count || ctx2->pin_count) | |
2202 | return 0; | |
2203 | ||
2204 | /* If ctx1 is the parent of ctx2 */ | |
2205 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2206 | return 1; | |
2207 | ||
2208 | /* If ctx2 is the parent of ctx1 */ | |
2209 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2210 | return 1; | |
2211 | ||
2212 | /* | |
2213 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2214 | * hierarchy, see perf_event_init_context(). | |
2215 | */ | |
2216 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2217 | ctx1->parent_gen == ctx2->parent_gen) | |
2218 | return 1; | |
2219 | ||
2220 | /* Unmatched */ | |
2221 | return 0; | |
564c2b21 PM |
2222 | } |
2223 | ||
cdd6c482 IM |
2224 | static void __perf_event_sync_stat(struct perf_event *event, |
2225 | struct perf_event *next_event) | |
bfbd3381 PZ |
2226 | { |
2227 | u64 value; | |
2228 | ||
cdd6c482 | 2229 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2230 | return; |
2231 | ||
2232 | /* | |
cdd6c482 | 2233 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2234 | * because we're in the middle of a context switch and have IRQs |
2235 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2236 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2237 | * don't need to use it. |
2238 | */ | |
cdd6c482 IM |
2239 | switch (event->state) { |
2240 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2241 | event->pmu->read(event); |
2242 | /* fall-through */ | |
bfbd3381 | 2243 | |
cdd6c482 IM |
2244 | case PERF_EVENT_STATE_INACTIVE: |
2245 | update_event_times(event); | |
bfbd3381 PZ |
2246 | break; |
2247 | ||
2248 | default: | |
2249 | break; | |
2250 | } | |
2251 | ||
2252 | /* | |
cdd6c482 | 2253 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2254 | * values when we flip the contexts. |
2255 | */ | |
e7850595 PZ |
2256 | value = local64_read(&next_event->count); |
2257 | value = local64_xchg(&event->count, value); | |
2258 | local64_set(&next_event->count, value); | |
bfbd3381 | 2259 | |
cdd6c482 IM |
2260 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2261 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2262 | |
bfbd3381 | 2263 | /* |
19d2e755 | 2264 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2265 | */ |
cdd6c482 IM |
2266 | perf_event_update_userpage(event); |
2267 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2268 | } |
2269 | ||
cdd6c482 IM |
2270 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2271 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2272 | { |
cdd6c482 | 2273 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2274 | |
2275 | if (!ctx->nr_stat) | |
2276 | return; | |
2277 | ||
02ffdbc8 PZ |
2278 | update_context_time(ctx); |
2279 | ||
cdd6c482 IM |
2280 | event = list_first_entry(&ctx->event_list, |
2281 | struct perf_event, event_entry); | |
bfbd3381 | 2282 | |
cdd6c482 IM |
2283 | next_event = list_first_entry(&next_ctx->event_list, |
2284 | struct perf_event, event_entry); | |
bfbd3381 | 2285 | |
cdd6c482 IM |
2286 | while (&event->event_entry != &ctx->event_list && |
2287 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2288 | |
cdd6c482 | 2289 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2290 | |
cdd6c482 IM |
2291 | event = list_next_entry(event, event_entry); |
2292 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2293 | } |
2294 | } | |
2295 | ||
fe4b04fa PZ |
2296 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2297 | struct task_struct *next) | |
0793a61d | 2298 | { |
8dc85d54 | 2299 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2300 | struct perf_event_context *next_ctx; |
5a3126d4 | 2301 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2302 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2303 | int do_switch = 1; |
0793a61d | 2304 | |
108b02cf PZ |
2305 | if (likely(!ctx)) |
2306 | return; | |
10989fb2 | 2307 | |
108b02cf PZ |
2308 | cpuctx = __get_cpu_context(ctx); |
2309 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2310 | return; |
2311 | ||
c93f7669 | 2312 | rcu_read_lock(); |
8dc85d54 | 2313 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2314 | if (!next_ctx) |
2315 | goto unlock; | |
2316 | ||
2317 | parent = rcu_dereference(ctx->parent_ctx); | |
2318 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2319 | ||
2320 | /* If neither context have a parent context; they cannot be clones. */ | |
2321 | if (!parent && !next_parent) | |
2322 | goto unlock; | |
2323 | ||
2324 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2325 | /* |
2326 | * Looks like the two contexts are clones, so we might be | |
2327 | * able to optimize the context switch. We lock both | |
2328 | * contexts and check that they are clones under the | |
2329 | * lock (including re-checking that neither has been | |
2330 | * uncloned in the meantime). It doesn't matter which | |
2331 | * order we take the locks because no other cpu could | |
2332 | * be trying to lock both of these tasks. | |
2333 | */ | |
e625cce1 TG |
2334 | raw_spin_lock(&ctx->lock); |
2335 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2336 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2337 | /* |
2338 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2339 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2340 | */ |
8dc85d54 PZ |
2341 | task->perf_event_ctxp[ctxn] = next_ctx; |
2342 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2343 | ctx->task = next; |
2344 | next_ctx->task = task; | |
2345 | do_switch = 0; | |
bfbd3381 | 2346 | |
cdd6c482 | 2347 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2348 | } |
e625cce1 TG |
2349 | raw_spin_unlock(&next_ctx->lock); |
2350 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2351 | } |
5a3126d4 | 2352 | unlock: |
c93f7669 | 2353 | rcu_read_unlock(); |
564c2b21 | 2354 | |
c93f7669 | 2355 | if (do_switch) { |
facc4307 | 2356 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2357 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2358 | cpuctx->task_ctx = NULL; |
facc4307 | 2359 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2360 | } |
0793a61d TG |
2361 | } |
2362 | ||
8dc85d54 PZ |
2363 | #define for_each_task_context_nr(ctxn) \ |
2364 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2365 | ||
2366 | /* | |
2367 | * Called from scheduler to remove the events of the current task, | |
2368 | * with interrupts disabled. | |
2369 | * | |
2370 | * We stop each event and update the event value in event->count. | |
2371 | * | |
2372 | * This does not protect us against NMI, but disable() | |
2373 | * sets the disabled bit in the control field of event _before_ | |
2374 | * accessing the event control register. If a NMI hits, then it will | |
2375 | * not restart the event. | |
2376 | */ | |
ab0cce56 JO |
2377 | void __perf_event_task_sched_out(struct task_struct *task, |
2378 | struct task_struct *next) | |
8dc85d54 PZ |
2379 | { |
2380 | int ctxn; | |
2381 | ||
8dc85d54 PZ |
2382 | for_each_task_context_nr(ctxn) |
2383 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2384 | |
2385 | /* | |
2386 | * if cgroup events exist on this CPU, then we need | |
2387 | * to check if we have to switch out PMU state. | |
2388 | * cgroup event are system-wide mode only | |
2389 | */ | |
2390 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2391 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2392 | } |
2393 | ||
04dc2dbb | 2394 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2395 | { |
108b02cf | 2396 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2397 | |
a63eaf34 PM |
2398 | if (!cpuctx->task_ctx) |
2399 | return; | |
012b84da IM |
2400 | |
2401 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2402 | return; | |
2403 | ||
04dc2dbb | 2404 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2405 | cpuctx->task_ctx = NULL; |
2406 | } | |
2407 | ||
5b0311e1 FW |
2408 | /* |
2409 | * Called with IRQs disabled | |
2410 | */ | |
2411 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2412 | enum event_type_t event_type) | |
2413 | { | |
2414 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2415 | } |
2416 | ||
235c7fc7 | 2417 | static void |
5b0311e1 | 2418 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2419 | struct perf_cpu_context *cpuctx) |
0793a61d | 2420 | { |
cdd6c482 | 2421 | struct perf_event *event; |
0793a61d | 2422 | |
889ff015 FW |
2423 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2424 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2425 | continue; |
5632ab12 | 2426 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2427 | continue; |
2428 | ||
e5d1367f SE |
2429 | /* may need to reset tstamp_enabled */ |
2430 | if (is_cgroup_event(event)) | |
2431 | perf_cgroup_mark_enabled(event, ctx); | |
2432 | ||
8c9ed8e1 | 2433 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2434 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2435 | |
2436 | /* | |
2437 | * If this pinned group hasn't been scheduled, | |
2438 | * put it in error state. | |
2439 | */ | |
cdd6c482 IM |
2440 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2441 | update_group_times(event); | |
2442 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2443 | } |
3b6f9e5c | 2444 | } |
5b0311e1 FW |
2445 | } |
2446 | ||
2447 | static void | |
2448 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2449 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2450 | { |
2451 | struct perf_event *event; | |
2452 | int can_add_hw = 1; | |
3b6f9e5c | 2453 | |
889ff015 FW |
2454 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2455 | /* Ignore events in OFF or ERROR state */ | |
2456 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2457 | continue; |
04289bb9 IM |
2458 | /* |
2459 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2460 | * of events: |
04289bb9 | 2461 | */ |
5632ab12 | 2462 | if (!event_filter_match(event)) |
0793a61d TG |
2463 | continue; |
2464 | ||
e5d1367f SE |
2465 | /* may need to reset tstamp_enabled */ |
2466 | if (is_cgroup_event(event)) | |
2467 | perf_cgroup_mark_enabled(event, ctx); | |
2468 | ||
9ed6060d | 2469 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2470 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2471 | can_add_hw = 0; |
9ed6060d | 2472 | } |
0793a61d | 2473 | } |
5b0311e1 FW |
2474 | } |
2475 | ||
2476 | static void | |
2477 | ctx_sched_in(struct perf_event_context *ctx, | |
2478 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2479 | enum event_type_t event_type, |
2480 | struct task_struct *task) | |
5b0311e1 | 2481 | { |
e5d1367f | 2482 | u64 now; |
db24d33e | 2483 | int is_active = ctx->is_active; |
e5d1367f | 2484 | |
db24d33e | 2485 | ctx->is_active |= event_type; |
5b0311e1 | 2486 | if (likely(!ctx->nr_events)) |
facc4307 | 2487 | return; |
5b0311e1 | 2488 | |
e5d1367f SE |
2489 | now = perf_clock(); |
2490 | ctx->timestamp = now; | |
3f7cce3c | 2491 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2492 | /* |
2493 | * First go through the list and put on any pinned groups | |
2494 | * in order to give them the best chance of going on. | |
2495 | */ | |
db24d33e | 2496 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2497 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2498 | |
2499 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2500 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2501 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2502 | } |
2503 | ||
329c0e01 | 2504 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2505 | enum event_type_t event_type, |
2506 | struct task_struct *task) | |
329c0e01 FW |
2507 | { |
2508 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2509 | ||
e5d1367f | 2510 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2511 | } |
2512 | ||
e5d1367f SE |
2513 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2514 | struct task_struct *task) | |
235c7fc7 | 2515 | { |
108b02cf | 2516 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2517 | |
108b02cf | 2518 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2519 | if (cpuctx->task_ctx == ctx) |
2520 | return; | |
2521 | ||
facc4307 | 2522 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2523 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2524 | /* |
2525 | * We want to keep the following priority order: | |
2526 | * cpu pinned (that don't need to move), task pinned, | |
2527 | * cpu flexible, task flexible. | |
2528 | */ | |
2529 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2530 | ||
1d5f003f GN |
2531 | if (ctx->nr_events) |
2532 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2533 | |
86b47c25 GN |
2534 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2535 | ||
facc4307 PZ |
2536 | perf_pmu_enable(ctx->pmu); |
2537 | perf_ctx_unlock(cpuctx, ctx); | |
2538 | ||
b5ab4cd5 PZ |
2539 | /* |
2540 | * Since these rotations are per-cpu, we need to ensure the | |
2541 | * cpu-context we got scheduled on is actually rotating. | |
2542 | */ | |
108b02cf | 2543 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2544 | } |
2545 | ||
d010b332 SE |
2546 | /* |
2547 | * When sampling the branck stack in system-wide, it may be necessary | |
2548 | * to flush the stack on context switch. This happens when the branch | |
2549 | * stack does not tag its entries with the pid of the current task. | |
2550 | * Otherwise it becomes impossible to associate a branch entry with a | |
2551 | * task. This ambiguity is more likely to appear when the branch stack | |
2552 | * supports priv level filtering and the user sets it to monitor only | |
2553 | * at the user level (which could be a useful measurement in system-wide | |
2554 | * mode). In that case, the risk is high of having a branch stack with | |
2555 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2556 | * entries or stashing them somewhere in the PMU specific code layer. | |
2557 | * | |
2558 | * This function provides the context switch callback to the lower code | |
2559 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2560 | * with at least one active event using taken branch sampling. | |
2561 | */ | |
2562 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2563 | struct task_struct *task) | |
2564 | { | |
2565 | struct perf_cpu_context *cpuctx; | |
2566 | struct pmu *pmu; | |
2567 | unsigned long flags; | |
2568 | ||
2569 | /* no need to flush branch stack if not changing task */ | |
2570 | if (prev == task) | |
2571 | return; | |
2572 | ||
2573 | local_irq_save(flags); | |
2574 | ||
2575 | rcu_read_lock(); | |
2576 | ||
2577 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2578 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2579 | ||
2580 | /* | |
2581 | * check if the context has at least one | |
2582 | * event using PERF_SAMPLE_BRANCH_STACK | |
2583 | */ | |
2584 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2585 | && pmu->flush_branch_stack) { | |
2586 | ||
d010b332 SE |
2587 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
2588 | ||
2589 | perf_pmu_disable(pmu); | |
2590 | ||
2591 | pmu->flush_branch_stack(); | |
2592 | ||
2593 | perf_pmu_enable(pmu); | |
2594 | ||
2595 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2596 | } | |
2597 | } | |
2598 | ||
2599 | rcu_read_unlock(); | |
2600 | ||
2601 | local_irq_restore(flags); | |
2602 | } | |
2603 | ||
8dc85d54 PZ |
2604 | /* |
2605 | * Called from scheduler to add the events of the current task | |
2606 | * with interrupts disabled. | |
2607 | * | |
2608 | * We restore the event value and then enable it. | |
2609 | * | |
2610 | * This does not protect us against NMI, but enable() | |
2611 | * sets the enabled bit in the control field of event _before_ | |
2612 | * accessing the event control register. If a NMI hits, then it will | |
2613 | * keep the event running. | |
2614 | */ | |
ab0cce56 JO |
2615 | void __perf_event_task_sched_in(struct task_struct *prev, |
2616 | struct task_struct *task) | |
8dc85d54 PZ |
2617 | { |
2618 | struct perf_event_context *ctx; | |
2619 | int ctxn; | |
2620 | ||
2621 | for_each_task_context_nr(ctxn) { | |
2622 | ctx = task->perf_event_ctxp[ctxn]; | |
2623 | if (likely(!ctx)) | |
2624 | continue; | |
2625 | ||
e5d1367f | 2626 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2627 | } |
e5d1367f SE |
2628 | /* |
2629 | * if cgroup events exist on this CPU, then we need | |
2630 | * to check if we have to switch in PMU state. | |
2631 | * cgroup event are system-wide mode only | |
2632 | */ | |
2633 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2634 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2635 | |
2636 | /* check for system-wide branch_stack events */ | |
2637 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2638 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2639 | } |
2640 | ||
abd50713 PZ |
2641 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2642 | { | |
2643 | u64 frequency = event->attr.sample_freq; | |
2644 | u64 sec = NSEC_PER_SEC; | |
2645 | u64 divisor, dividend; | |
2646 | ||
2647 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2648 | ||
2649 | count_fls = fls64(count); | |
2650 | nsec_fls = fls64(nsec); | |
2651 | frequency_fls = fls64(frequency); | |
2652 | sec_fls = 30; | |
2653 | ||
2654 | /* | |
2655 | * We got @count in @nsec, with a target of sample_freq HZ | |
2656 | * the target period becomes: | |
2657 | * | |
2658 | * @count * 10^9 | |
2659 | * period = ------------------- | |
2660 | * @nsec * sample_freq | |
2661 | * | |
2662 | */ | |
2663 | ||
2664 | /* | |
2665 | * Reduce accuracy by one bit such that @a and @b converge | |
2666 | * to a similar magnitude. | |
2667 | */ | |
fe4b04fa | 2668 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2669 | do { \ |
2670 | if (a##_fls > b##_fls) { \ | |
2671 | a >>= 1; \ | |
2672 | a##_fls--; \ | |
2673 | } else { \ | |
2674 | b >>= 1; \ | |
2675 | b##_fls--; \ | |
2676 | } \ | |
2677 | } while (0) | |
2678 | ||
2679 | /* | |
2680 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2681 | * the other, so that finally we can do a u64/u64 division. | |
2682 | */ | |
2683 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2684 | REDUCE_FLS(nsec, frequency); | |
2685 | REDUCE_FLS(sec, count); | |
2686 | } | |
2687 | ||
2688 | if (count_fls + sec_fls > 64) { | |
2689 | divisor = nsec * frequency; | |
2690 | ||
2691 | while (count_fls + sec_fls > 64) { | |
2692 | REDUCE_FLS(count, sec); | |
2693 | divisor >>= 1; | |
2694 | } | |
2695 | ||
2696 | dividend = count * sec; | |
2697 | } else { | |
2698 | dividend = count * sec; | |
2699 | ||
2700 | while (nsec_fls + frequency_fls > 64) { | |
2701 | REDUCE_FLS(nsec, frequency); | |
2702 | dividend >>= 1; | |
2703 | } | |
2704 | ||
2705 | divisor = nsec * frequency; | |
2706 | } | |
2707 | ||
f6ab91ad PZ |
2708 | if (!divisor) |
2709 | return dividend; | |
2710 | ||
abd50713 PZ |
2711 | return div64_u64(dividend, divisor); |
2712 | } | |
2713 | ||
e050e3f0 SE |
2714 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2715 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2716 | ||
f39d47ff | 2717 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2718 | { |
cdd6c482 | 2719 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2720 | s64 period, sample_period; |
bd2b5b12 PZ |
2721 | s64 delta; |
2722 | ||
abd50713 | 2723 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2724 | |
2725 | delta = (s64)(period - hwc->sample_period); | |
2726 | delta = (delta + 7) / 8; /* low pass filter */ | |
2727 | ||
2728 | sample_period = hwc->sample_period + delta; | |
2729 | ||
2730 | if (!sample_period) | |
2731 | sample_period = 1; | |
2732 | ||
bd2b5b12 | 2733 | hwc->sample_period = sample_period; |
abd50713 | 2734 | |
e7850595 | 2735 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2736 | if (disable) |
2737 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2738 | ||
e7850595 | 2739 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2740 | |
2741 | if (disable) | |
2742 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2743 | } |
bd2b5b12 PZ |
2744 | } |
2745 | ||
e050e3f0 SE |
2746 | /* |
2747 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2748 | * events. At the same time, make sure, having freq events does not change | |
2749 | * the rate of unthrottling as that would introduce bias. | |
2750 | */ | |
2751 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2752 | int needs_unthr) | |
60db5e09 | 2753 | { |
cdd6c482 IM |
2754 | struct perf_event *event; |
2755 | struct hw_perf_event *hwc; | |
e050e3f0 | 2756 | u64 now, period = TICK_NSEC; |
abd50713 | 2757 | s64 delta; |
60db5e09 | 2758 | |
e050e3f0 SE |
2759 | /* |
2760 | * only need to iterate over all events iff: | |
2761 | * - context have events in frequency mode (needs freq adjust) | |
2762 | * - there are events to unthrottle on this cpu | |
2763 | */ | |
2764 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2765 | return; |
2766 | ||
e050e3f0 | 2767 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2768 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2769 | |
03541f8b | 2770 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2771 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2772 | continue; |
2773 | ||
5632ab12 | 2774 | if (!event_filter_match(event)) |
5d27c23d PZ |
2775 | continue; |
2776 | ||
44377277 AS |
2777 | perf_pmu_disable(event->pmu); |
2778 | ||
cdd6c482 | 2779 | hwc = &event->hw; |
6a24ed6c | 2780 | |
ae23bff1 | 2781 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2782 | hwc->interrupts = 0; |
cdd6c482 | 2783 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2784 | event->pmu->start(event, 0); |
a78ac325 PZ |
2785 | } |
2786 | ||
cdd6c482 | 2787 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2788 | goto next; |
60db5e09 | 2789 | |
e050e3f0 SE |
2790 | /* |
2791 | * stop the event and update event->count | |
2792 | */ | |
2793 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2794 | ||
e7850595 | 2795 | now = local64_read(&event->count); |
abd50713 PZ |
2796 | delta = now - hwc->freq_count_stamp; |
2797 | hwc->freq_count_stamp = now; | |
60db5e09 | 2798 | |
e050e3f0 SE |
2799 | /* |
2800 | * restart the event | |
2801 | * reload only if value has changed | |
f39d47ff SE |
2802 | * we have stopped the event so tell that |
2803 | * to perf_adjust_period() to avoid stopping it | |
2804 | * twice. | |
e050e3f0 | 2805 | */ |
abd50713 | 2806 | if (delta > 0) |
f39d47ff | 2807 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2808 | |
2809 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
2810 | next: |
2811 | perf_pmu_enable(event->pmu); | |
60db5e09 | 2812 | } |
e050e3f0 | 2813 | |
f39d47ff | 2814 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2815 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2816 | } |
2817 | ||
235c7fc7 | 2818 | /* |
cdd6c482 | 2819 | * Round-robin a context's events: |
235c7fc7 | 2820 | */ |
cdd6c482 | 2821 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2822 | { |
dddd3379 TG |
2823 | /* |
2824 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2825 | * disabled by the inheritance code. | |
2826 | */ | |
2827 | if (!ctx->rotate_disable) | |
2828 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2829 | } |
2830 | ||
b5ab4cd5 | 2831 | /* |
e9d2b064 PZ |
2832 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2833 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2834 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2835 | */ |
9e630205 | 2836 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2837 | { |
8dc85d54 | 2838 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2839 | int rotate = 0, remove = 1; |
7fc23a53 | 2840 | |
b5ab4cd5 | 2841 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2842 | remove = 0; |
b5ab4cd5 PZ |
2843 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2844 | rotate = 1; | |
2845 | } | |
235c7fc7 | 2846 | |
8dc85d54 | 2847 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2848 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2849 | remove = 0; |
b5ab4cd5 PZ |
2850 | if (ctx->nr_events != ctx->nr_active) |
2851 | rotate = 1; | |
2852 | } | |
9717e6cd | 2853 | |
e050e3f0 | 2854 | if (!rotate) |
0f5a2601 PZ |
2855 | goto done; |
2856 | ||
facc4307 | 2857 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2858 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2859 | |
e050e3f0 SE |
2860 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2861 | if (ctx) | |
2862 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2863 | |
e050e3f0 SE |
2864 | rotate_ctx(&cpuctx->ctx); |
2865 | if (ctx) | |
2866 | rotate_ctx(ctx); | |
235c7fc7 | 2867 | |
e050e3f0 | 2868 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2869 | |
0f5a2601 PZ |
2870 | perf_pmu_enable(cpuctx->ctx.pmu); |
2871 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2872 | done: |
e9d2b064 PZ |
2873 | if (remove) |
2874 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2875 | |
2876 | return rotate; | |
e9d2b064 PZ |
2877 | } |
2878 | ||
026249ef FW |
2879 | #ifdef CONFIG_NO_HZ_FULL |
2880 | bool perf_event_can_stop_tick(void) | |
2881 | { | |
948b26b6 | 2882 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 2883 | __this_cpu_read(perf_throttled_count)) |
026249ef | 2884 | return false; |
d84153d6 FW |
2885 | else |
2886 | return true; | |
026249ef FW |
2887 | } |
2888 | #endif | |
2889 | ||
e9d2b064 PZ |
2890 | void perf_event_task_tick(void) |
2891 | { | |
2892 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2893 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2894 | struct perf_event_context *ctx; |
2895 | int throttled; | |
b5ab4cd5 | 2896 | |
e9d2b064 PZ |
2897 | WARN_ON(!irqs_disabled()); |
2898 | ||
e050e3f0 SE |
2899 | __this_cpu_inc(perf_throttled_seq); |
2900 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2901 | ||
e9d2b064 | 2902 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2903 | ctx = &cpuctx->ctx; |
2904 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2905 | ||
2906 | ctx = cpuctx->task_ctx; | |
2907 | if (ctx) | |
2908 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2909 | } |
0793a61d TG |
2910 | } |
2911 | ||
889ff015 FW |
2912 | static int event_enable_on_exec(struct perf_event *event, |
2913 | struct perf_event_context *ctx) | |
2914 | { | |
2915 | if (!event->attr.enable_on_exec) | |
2916 | return 0; | |
2917 | ||
2918 | event->attr.enable_on_exec = 0; | |
2919 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2920 | return 0; | |
2921 | ||
1d9b482e | 2922 | __perf_event_mark_enabled(event); |
889ff015 FW |
2923 | |
2924 | return 1; | |
2925 | } | |
2926 | ||
57e7986e | 2927 | /* |
cdd6c482 | 2928 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2929 | * This expects task == current. |
2930 | */ | |
8dc85d54 | 2931 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2932 | { |
cdd6c482 | 2933 | struct perf_event *event; |
57e7986e PM |
2934 | unsigned long flags; |
2935 | int enabled = 0; | |
889ff015 | 2936 | int ret; |
57e7986e PM |
2937 | |
2938 | local_irq_save(flags); | |
cdd6c482 | 2939 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2940 | goto out; |
2941 | ||
e566b76e SE |
2942 | /* |
2943 | * We must ctxsw out cgroup events to avoid conflict | |
2944 | * when invoking perf_task_event_sched_in() later on | |
2945 | * in this function. Otherwise we end up trying to | |
2946 | * ctxswin cgroup events which are already scheduled | |
2947 | * in. | |
2948 | */ | |
a8d757ef | 2949 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2950 | |
e625cce1 | 2951 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2952 | task_ctx_sched_out(ctx); |
57e7986e | 2953 | |
b79387ef | 2954 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2955 | ret = event_enable_on_exec(event, ctx); |
2956 | if (ret) | |
2957 | enabled = 1; | |
57e7986e PM |
2958 | } |
2959 | ||
2960 | /* | |
cdd6c482 | 2961 | * Unclone this context if we enabled any event. |
57e7986e | 2962 | */ |
71a851b4 PZ |
2963 | if (enabled) |
2964 | unclone_ctx(ctx); | |
57e7986e | 2965 | |
e625cce1 | 2966 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2967 | |
e566b76e SE |
2968 | /* |
2969 | * Also calls ctxswin for cgroup events, if any: | |
2970 | */ | |
e5d1367f | 2971 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2972 | out: |
57e7986e PM |
2973 | local_irq_restore(flags); |
2974 | } | |
2975 | ||
e041e328 PZ |
2976 | void perf_event_exec(void) |
2977 | { | |
2978 | struct perf_event_context *ctx; | |
2979 | int ctxn; | |
2980 | ||
2981 | rcu_read_lock(); | |
2982 | for_each_task_context_nr(ctxn) { | |
2983 | ctx = current->perf_event_ctxp[ctxn]; | |
2984 | if (!ctx) | |
2985 | continue; | |
2986 | ||
2987 | perf_event_enable_on_exec(ctx); | |
2988 | } | |
2989 | rcu_read_unlock(); | |
2990 | } | |
2991 | ||
0793a61d | 2992 | /* |
cdd6c482 | 2993 | * Cross CPU call to read the hardware event |
0793a61d | 2994 | */ |
cdd6c482 | 2995 | static void __perf_event_read(void *info) |
0793a61d | 2996 | { |
cdd6c482 IM |
2997 | struct perf_event *event = info; |
2998 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2999 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3000 | |
e1ac3614 PM |
3001 | /* |
3002 | * If this is a task context, we need to check whether it is | |
3003 | * the current task context of this cpu. If not it has been | |
3004 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3005 | * event->count would have been updated to a recent sample |
3006 | * when the event was scheduled out. | |
e1ac3614 PM |
3007 | */ |
3008 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3009 | return; | |
3010 | ||
e625cce1 | 3011 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3012 | if (ctx->is_active) { |
542e72fc | 3013 | update_context_time(ctx); |
e5d1367f SE |
3014 | update_cgrp_time_from_event(event); |
3015 | } | |
cdd6c482 | 3016 | update_event_times(event); |
542e72fc PZ |
3017 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3018 | event->pmu->read(event); | |
e625cce1 | 3019 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3020 | } |
3021 | ||
b5e58793 PZ |
3022 | static inline u64 perf_event_count(struct perf_event *event) |
3023 | { | |
e7850595 | 3024 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
3025 | } |
3026 | ||
cdd6c482 | 3027 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3028 | { |
3029 | /* | |
cdd6c482 IM |
3030 | * If event is enabled and currently active on a CPU, update the |
3031 | * value in the event structure: | |
0793a61d | 3032 | */ |
cdd6c482 IM |
3033 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3034 | smp_call_function_single(event->oncpu, | |
3035 | __perf_event_read, event, 1); | |
3036 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3037 | struct perf_event_context *ctx = event->ctx; |
3038 | unsigned long flags; | |
3039 | ||
e625cce1 | 3040 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3041 | /* |
3042 | * may read while context is not active | |
3043 | * (e.g., thread is blocked), in that case | |
3044 | * we cannot update context time | |
3045 | */ | |
e5d1367f | 3046 | if (ctx->is_active) { |
c530ccd9 | 3047 | update_context_time(ctx); |
e5d1367f SE |
3048 | update_cgrp_time_from_event(event); |
3049 | } | |
cdd6c482 | 3050 | update_event_times(event); |
e625cce1 | 3051 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3052 | } |
3053 | ||
b5e58793 | 3054 | return perf_event_count(event); |
0793a61d TG |
3055 | } |
3056 | ||
a63eaf34 | 3057 | /* |
cdd6c482 | 3058 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3059 | */ |
eb184479 | 3060 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3061 | { |
e625cce1 | 3062 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3063 | mutex_init(&ctx->mutex); |
889ff015 FW |
3064 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3065 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3066 | INIT_LIST_HEAD(&ctx->event_list); |
3067 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
3068 | } |
3069 | ||
3070 | static struct perf_event_context * | |
3071 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3072 | { | |
3073 | struct perf_event_context *ctx; | |
3074 | ||
3075 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3076 | if (!ctx) | |
3077 | return NULL; | |
3078 | ||
3079 | __perf_event_init_context(ctx); | |
3080 | if (task) { | |
3081 | ctx->task = task; | |
3082 | get_task_struct(task); | |
0793a61d | 3083 | } |
eb184479 PZ |
3084 | ctx->pmu = pmu; |
3085 | ||
3086 | return ctx; | |
a63eaf34 PM |
3087 | } |
3088 | ||
2ebd4ffb MH |
3089 | static struct task_struct * |
3090 | find_lively_task_by_vpid(pid_t vpid) | |
3091 | { | |
3092 | struct task_struct *task; | |
3093 | int err; | |
0793a61d TG |
3094 | |
3095 | rcu_read_lock(); | |
2ebd4ffb | 3096 | if (!vpid) |
0793a61d TG |
3097 | task = current; |
3098 | else | |
2ebd4ffb | 3099 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3100 | if (task) |
3101 | get_task_struct(task); | |
3102 | rcu_read_unlock(); | |
3103 | ||
3104 | if (!task) | |
3105 | return ERR_PTR(-ESRCH); | |
3106 | ||
0793a61d | 3107 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3108 | err = -EACCES; |
3109 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3110 | goto errout; | |
3111 | ||
2ebd4ffb MH |
3112 | return task; |
3113 | errout: | |
3114 | put_task_struct(task); | |
3115 | return ERR_PTR(err); | |
3116 | ||
3117 | } | |
3118 | ||
fe4b04fa PZ |
3119 | /* |
3120 | * Returns a matching context with refcount and pincount. | |
3121 | */ | |
108b02cf | 3122 | static struct perf_event_context * |
38a81da2 | 3123 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3124 | { |
cdd6c482 | 3125 | struct perf_event_context *ctx; |
22a4f650 | 3126 | struct perf_cpu_context *cpuctx; |
25346b93 | 3127 | unsigned long flags; |
8dc85d54 | 3128 | int ctxn, err; |
0793a61d | 3129 | |
22a4ec72 | 3130 | if (!task) { |
cdd6c482 | 3131 | /* Must be root to operate on a CPU event: */ |
0764771d | 3132 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3133 | return ERR_PTR(-EACCES); |
3134 | ||
0793a61d | 3135 | /* |
cdd6c482 | 3136 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3137 | * offline CPU and activate it when the CPU comes up, but |
3138 | * that's for later. | |
3139 | */ | |
f6325e30 | 3140 | if (!cpu_online(cpu)) |
0793a61d TG |
3141 | return ERR_PTR(-ENODEV); |
3142 | ||
108b02cf | 3143 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3144 | ctx = &cpuctx->ctx; |
c93f7669 | 3145 | get_ctx(ctx); |
fe4b04fa | 3146 | ++ctx->pin_count; |
0793a61d | 3147 | |
0793a61d TG |
3148 | return ctx; |
3149 | } | |
3150 | ||
8dc85d54 PZ |
3151 | err = -EINVAL; |
3152 | ctxn = pmu->task_ctx_nr; | |
3153 | if (ctxn < 0) | |
3154 | goto errout; | |
3155 | ||
9ed6060d | 3156 | retry: |
8dc85d54 | 3157 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3158 | if (ctx) { |
71a851b4 | 3159 | unclone_ctx(ctx); |
fe4b04fa | 3160 | ++ctx->pin_count; |
e625cce1 | 3161 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3162 | } else { |
eb184479 | 3163 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3164 | err = -ENOMEM; |
3165 | if (!ctx) | |
3166 | goto errout; | |
eb184479 | 3167 | |
dbe08d82 ON |
3168 | err = 0; |
3169 | mutex_lock(&task->perf_event_mutex); | |
3170 | /* | |
3171 | * If it has already passed perf_event_exit_task(). | |
3172 | * we must see PF_EXITING, it takes this mutex too. | |
3173 | */ | |
3174 | if (task->flags & PF_EXITING) | |
3175 | err = -ESRCH; | |
3176 | else if (task->perf_event_ctxp[ctxn]) | |
3177 | err = -EAGAIN; | |
fe4b04fa | 3178 | else { |
9137fb28 | 3179 | get_ctx(ctx); |
fe4b04fa | 3180 | ++ctx->pin_count; |
dbe08d82 | 3181 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3182 | } |
dbe08d82 ON |
3183 | mutex_unlock(&task->perf_event_mutex); |
3184 | ||
3185 | if (unlikely(err)) { | |
9137fb28 | 3186 | put_ctx(ctx); |
dbe08d82 ON |
3187 | |
3188 | if (err == -EAGAIN) | |
3189 | goto retry; | |
3190 | goto errout; | |
a63eaf34 PM |
3191 | } |
3192 | } | |
3193 | ||
0793a61d | 3194 | return ctx; |
c93f7669 | 3195 | |
9ed6060d | 3196 | errout: |
c93f7669 | 3197 | return ERR_PTR(err); |
0793a61d TG |
3198 | } |
3199 | ||
6fb2915d LZ |
3200 | static void perf_event_free_filter(struct perf_event *event); |
3201 | ||
cdd6c482 | 3202 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3203 | { |
cdd6c482 | 3204 | struct perf_event *event; |
592903cd | 3205 | |
cdd6c482 IM |
3206 | event = container_of(head, struct perf_event, rcu_head); |
3207 | if (event->ns) | |
3208 | put_pid_ns(event->ns); | |
6fb2915d | 3209 | perf_event_free_filter(event); |
cdd6c482 | 3210 | kfree(event); |
592903cd PZ |
3211 | } |
3212 | ||
76369139 | 3213 | static void ring_buffer_put(struct ring_buffer *rb); |
b69cf536 PZ |
3214 | static void ring_buffer_attach(struct perf_event *event, |
3215 | struct ring_buffer *rb); | |
925d519a | 3216 | |
4beb31f3 | 3217 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3218 | { |
4beb31f3 FW |
3219 | if (event->parent) |
3220 | return; | |
3221 | ||
3222 | if (has_branch_stack(event)) { | |
3223 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3224 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3225 | } | |
3226 | if (is_cgroup_event(event)) | |
3227 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3228 | } | |
925d519a | 3229 | |
4beb31f3 FW |
3230 | static void unaccount_event(struct perf_event *event) |
3231 | { | |
3232 | if (event->parent) | |
3233 | return; | |
3234 | ||
3235 | if (event->attach_state & PERF_ATTACH_TASK) | |
3236 | static_key_slow_dec_deferred(&perf_sched_events); | |
3237 | if (event->attr.mmap || event->attr.mmap_data) | |
3238 | atomic_dec(&nr_mmap_events); | |
3239 | if (event->attr.comm) | |
3240 | atomic_dec(&nr_comm_events); | |
3241 | if (event->attr.task) | |
3242 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3243 | if (event->attr.freq) |
3244 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3245 | if (is_cgroup_event(event)) |
3246 | static_key_slow_dec_deferred(&perf_sched_events); | |
3247 | if (has_branch_stack(event)) | |
3248 | static_key_slow_dec_deferred(&perf_sched_events); | |
3249 | ||
3250 | unaccount_event_cpu(event, event->cpu); | |
3251 | } | |
925d519a | 3252 | |
766d6c07 FW |
3253 | static void __free_event(struct perf_event *event) |
3254 | { | |
cdd6c482 | 3255 | if (!event->parent) { |
927c7a9e FW |
3256 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3257 | put_callchain_buffers(); | |
f344011c | 3258 | } |
9ee318a7 | 3259 | |
766d6c07 FW |
3260 | if (event->destroy) |
3261 | event->destroy(event); | |
3262 | ||
3263 | if (event->ctx) | |
3264 | put_ctx(event->ctx); | |
3265 | ||
c464c76e YZ |
3266 | if (event->pmu) |
3267 | module_put(event->pmu->module); | |
3268 | ||
766d6c07 FW |
3269 | call_rcu(&event->rcu_head, free_event_rcu); |
3270 | } | |
683ede43 PZ |
3271 | |
3272 | static void _free_event(struct perf_event *event) | |
f1600952 | 3273 | { |
e360adbe | 3274 | irq_work_sync(&event->pending); |
925d519a | 3275 | |
4beb31f3 | 3276 | unaccount_event(event); |
9ee318a7 | 3277 | |
76369139 | 3278 | if (event->rb) { |
9bb5d40c PZ |
3279 | /* |
3280 | * Can happen when we close an event with re-directed output. | |
3281 | * | |
3282 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3283 | * over us; possibly making our ring_buffer_put() the last. | |
3284 | */ | |
3285 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3286 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3287 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3288 | } |
3289 | ||
e5d1367f SE |
3290 | if (is_cgroup_event(event)) |
3291 | perf_detach_cgroup(event); | |
3292 | ||
766d6c07 | 3293 | __free_event(event); |
f1600952 PZ |
3294 | } |
3295 | ||
683ede43 PZ |
3296 | /* |
3297 | * Used to free events which have a known refcount of 1, such as in error paths | |
3298 | * where the event isn't exposed yet and inherited events. | |
3299 | */ | |
3300 | static void free_event(struct perf_event *event) | |
0793a61d | 3301 | { |
683ede43 PZ |
3302 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3303 | "unexpected event refcount: %ld; ptr=%p\n", | |
3304 | atomic_long_read(&event->refcount), event)) { | |
3305 | /* leak to avoid use-after-free */ | |
3306 | return; | |
3307 | } | |
0793a61d | 3308 | |
683ede43 | 3309 | _free_event(event); |
0793a61d TG |
3310 | } |
3311 | ||
a66a3052 PZ |
3312 | /* |
3313 | * Called when the last reference to the file is gone. | |
3314 | */ | |
a6fa941d | 3315 | static void put_event(struct perf_event *event) |
fb0459d7 | 3316 | { |
683ede43 | 3317 | struct perf_event_context *ctx = event->ctx; |
8882135b | 3318 | struct task_struct *owner; |
fb0459d7 | 3319 | |
a6fa941d AV |
3320 | if (!atomic_long_dec_and_test(&event->refcount)) |
3321 | return; | |
fb0459d7 | 3322 | |
8882135b PZ |
3323 | rcu_read_lock(); |
3324 | owner = ACCESS_ONCE(event->owner); | |
3325 | /* | |
3326 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3327 | * !owner it means the list deletion is complete and we can indeed | |
3328 | * free this event, otherwise we need to serialize on | |
3329 | * owner->perf_event_mutex. | |
3330 | */ | |
3331 | smp_read_barrier_depends(); | |
3332 | if (owner) { | |
3333 | /* | |
3334 | * Since delayed_put_task_struct() also drops the last | |
3335 | * task reference we can safely take a new reference | |
3336 | * while holding the rcu_read_lock(). | |
3337 | */ | |
3338 | get_task_struct(owner); | |
3339 | } | |
3340 | rcu_read_unlock(); | |
3341 | ||
3342 | if (owner) { | |
3343 | mutex_lock(&owner->perf_event_mutex); | |
3344 | /* | |
3345 | * We have to re-check the event->owner field, if it is cleared | |
3346 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3347 | * ensured they're done, and we can proceed with freeing the | |
3348 | * event. | |
3349 | */ | |
3350 | if (event->owner) | |
3351 | list_del_init(&event->owner_entry); | |
3352 | mutex_unlock(&owner->perf_event_mutex); | |
3353 | put_task_struct(owner); | |
3354 | } | |
3355 | ||
683ede43 PZ |
3356 | WARN_ON_ONCE(ctx->parent_ctx); |
3357 | /* | |
3358 | * There are two ways this annotation is useful: | |
3359 | * | |
3360 | * 1) there is a lock recursion from perf_event_exit_task | |
3361 | * see the comment there. | |
3362 | * | |
3363 | * 2) there is a lock-inversion with mmap_sem through | |
3364 | * perf_event_read_group(), which takes faults while | |
3365 | * holding ctx->mutex, however this is called after | |
3366 | * the last filedesc died, so there is no possibility | |
3367 | * to trigger the AB-BA case. | |
3368 | */ | |
3369 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
3370 | perf_remove_from_context(event, true); | |
3371 | mutex_unlock(&ctx->mutex); | |
3372 | ||
3373 | _free_event(event); | |
a6fa941d AV |
3374 | } |
3375 | ||
683ede43 PZ |
3376 | int perf_event_release_kernel(struct perf_event *event) |
3377 | { | |
3378 | put_event(event); | |
3379 | return 0; | |
3380 | } | |
3381 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3382 | ||
a6fa941d AV |
3383 | static int perf_release(struct inode *inode, struct file *file) |
3384 | { | |
3385 | put_event(file->private_data); | |
3386 | return 0; | |
fb0459d7 | 3387 | } |
fb0459d7 | 3388 | |
59ed446f | 3389 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3390 | { |
cdd6c482 | 3391 | struct perf_event *child; |
e53c0994 PZ |
3392 | u64 total = 0; |
3393 | ||
59ed446f PZ |
3394 | *enabled = 0; |
3395 | *running = 0; | |
3396 | ||
6f10581a | 3397 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3398 | total += perf_event_read(event); |
59ed446f PZ |
3399 | *enabled += event->total_time_enabled + |
3400 | atomic64_read(&event->child_total_time_enabled); | |
3401 | *running += event->total_time_running + | |
3402 | atomic64_read(&event->child_total_time_running); | |
3403 | ||
3404 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3405 | total += perf_event_read(child); |
59ed446f PZ |
3406 | *enabled += child->total_time_enabled; |
3407 | *running += child->total_time_running; | |
3408 | } | |
6f10581a | 3409 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3410 | |
3411 | return total; | |
3412 | } | |
fb0459d7 | 3413 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3414 | |
cdd6c482 | 3415 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3416 | u64 read_format, char __user *buf) |
3417 | { | |
cdd6c482 | 3418 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3419 | int n = 0, size = 0, ret = -EFAULT; |
3420 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3421 | u64 values[5]; |
59ed446f | 3422 | u64 count, enabled, running; |
abf4868b | 3423 | |
6f10581a | 3424 | mutex_lock(&ctx->mutex); |
59ed446f | 3425 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3426 | |
3427 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3428 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3429 | values[n++] = enabled; | |
3430 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3431 | values[n++] = running; | |
abf4868b PZ |
3432 | values[n++] = count; |
3433 | if (read_format & PERF_FORMAT_ID) | |
3434 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3435 | |
3436 | size = n * sizeof(u64); | |
3437 | ||
3438 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3439 | goto unlock; |
3dab77fb | 3440 | |
6f10581a | 3441 | ret = size; |
3dab77fb | 3442 | |
65abc865 | 3443 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3444 | n = 0; |
3dab77fb | 3445 | |
59ed446f | 3446 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3447 | if (read_format & PERF_FORMAT_ID) |
3448 | values[n++] = primary_event_id(sub); | |
3449 | ||
3450 | size = n * sizeof(u64); | |
3451 | ||
184d3da8 | 3452 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3453 | ret = -EFAULT; |
3454 | goto unlock; | |
3455 | } | |
abf4868b PZ |
3456 | |
3457 | ret += size; | |
3dab77fb | 3458 | } |
6f10581a PZ |
3459 | unlock: |
3460 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3461 | |
abf4868b | 3462 | return ret; |
3dab77fb PZ |
3463 | } |
3464 | ||
cdd6c482 | 3465 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3466 | u64 read_format, char __user *buf) |
3467 | { | |
59ed446f | 3468 | u64 enabled, running; |
3dab77fb PZ |
3469 | u64 values[4]; |
3470 | int n = 0; | |
3471 | ||
59ed446f PZ |
3472 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3473 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3474 | values[n++] = enabled; | |
3475 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3476 | values[n++] = running; | |
3dab77fb | 3477 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3478 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3479 | |
3480 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3481 | return -EFAULT; | |
3482 | ||
3483 | return n * sizeof(u64); | |
3484 | } | |
3485 | ||
0793a61d | 3486 | /* |
cdd6c482 | 3487 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3488 | */ |
3489 | static ssize_t | |
cdd6c482 | 3490 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3491 | { |
cdd6c482 | 3492 | u64 read_format = event->attr.read_format; |
3dab77fb | 3493 | int ret; |
0793a61d | 3494 | |
3b6f9e5c | 3495 | /* |
cdd6c482 | 3496 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3497 | * error state (i.e. because it was pinned but it couldn't be |
3498 | * scheduled on to the CPU at some point). | |
3499 | */ | |
cdd6c482 | 3500 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3501 | return 0; |
3502 | ||
c320c7b7 | 3503 | if (count < event->read_size) |
3dab77fb PZ |
3504 | return -ENOSPC; |
3505 | ||
cdd6c482 | 3506 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3507 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3508 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3509 | else |
cdd6c482 | 3510 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3511 | |
3dab77fb | 3512 | return ret; |
0793a61d TG |
3513 | } |
3514 | ||
0793a61d TG |
3515 | static ssize_t |
3516 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3517 | { | |
cdd6c482 | 3518 | struct perf_event *event = file->private_data; |
0793a61d | 3519 | |
cdd6c482 | 3520 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3521 | } |
3522 | ||
3523 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3524 | { | |
cdd6c482 | 3525 | struct perf_event *event = file->private_data; |
76369139 | 3526 | struct ring_buffer *rb; |
c33a0bc4 | 3527 | unsigned int events = POLL_HUP; |
c7138f37 | 3528 | |
10c6db11 | 3529 | /* |
9bb5d40c PZ |
3530 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3531 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3532 | */ |
3533 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3534 | rb = event->rb; |
3535 | if (rb) | |
76369139 | 3536 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3537 | mutex_unlock(&event->mmap_mutex); |
3538 | ||
cdd6c482 | 3539 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3540 | |
0793a61d TG |
3541 | return events; |
3542 | } | |
3543 | ||
cdd6c482 | 3544 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3545 | { |
cdd6c482 | 3546 | (void)perf_event_read(event); |
e7850595 | 3547 | local64_set(&event->count, 0); |
cdd6c482 | 3548 | perf_event_update_userpage(event); |
3df5edad PZ |
3549 | } |
3550 | ||
c93f7669 | 3551 | /* |
cdd6c482 IM |
3552 | * Holding the top-level event's child_mutex means that any |
3553 | * descendant process that has inherited this event will block | |
3554 | * in sync_child_event if it goes to exit, thus satisfying the | |
3555 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3556 | */ |
cdd6c482 IM |
3557 | static void perf_event_for_each_child(struct perf_event *event, |
3558 | void (*func)(struct perf_event *)) | |
3df5edad | 3559 | { |
cdd6c482 | 3560 | struct perf_event *child; |
3df5edad | 3561 | |
cdd6c482 IM |
3562 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3563 | mutex_lock(&event->child_mutex); | |
3564 | func(event); | |
3565 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3566 | func(child); |
cdd6c482 | 3567 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3568 | } |
3569 | ||
cdd6c482 IM |
3570 | static void perf_event_for_each(struct perf_event *event, |
3571 | void (*func)(struct perf_event *)) | |
3df5edad | 3572 | { |
cdd6c482 IM |
3573 | struct perf_event_context *ctx = event->ctx; |
3574 | struct perf_event *sibling; | |
3df5edad | 3575 | |
75f937f2 PZ |
3576 | WARN_ON_ONCE(ctx->parent_ctx); |
3577 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3578 | event = event->group_leader; |
75f937f2 | 3579 | |
cdd6c482 | 3580 | perf_event_for_each_child(event, func); |
cdd6c482 | 3581 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3582 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3583 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3584 | } |
3585 | ||
cdd6c482 | 3586 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3587 | { |
cdd6c482 | 3588 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3589 | int ret = 0, active; |
08247e31 PZ |
3590 | u64 value; |
3591 | ||
6c7e550f | 3592 | if (!is_sampling_event(event)) |
08247e31 PZ |
3593 | return -EINVAL; |
3594 | ||
ad0cf347 | 3595 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3596 | return -EFAULT; |
3597 | ||
3598 | if (!value) | |
3599 | return -EINVAL; | |
3600 | ||
e625cce1 | 3601 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3602 | if (event->attr.freq) { |
3603 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3604 | ret = -EINVAL; |
3605 | goto unlock; | |
3606 | } | |
3607 | ||
cdd6c482 | 3608 | event->attr.sample_freq = value; |
08247e31 | 3609 | } else { |
cdd6c482 IM |
3610 | event->attr.sample_period = value; |
3611 | event->hw.sample_period = value; | |
08247e31 | 3612 | } |
bad7192b PZ |
3613 | |
3614 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3615 | if (active) { | |
3616 | perf_pmu_disable(ctx->pmu); | |
3617 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3618 | } | |
3619 | ||
3620 | local64_set(&event->hw.period_left, 0); | |
3621 | ||
3622 | if (active) { | |
3623 | event->pmu->start(event, PERF_EF_RELOAD); | |
3624 | perf_pmu_enable(ctx->pmu); | |
3625 | } | |
3626 | ||
08247e31 | 3627 | unlock: |
e625cce1 | 3628 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3629 | |
3630 | return ret; | |
3631 | } | |
3632 | ||
ac9721f3 PZ |
3633 | static const struct file_operations perf_fops; |
3634 | ||
2903ff01 | 3635 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3636 | { |
2903ff01 AV |
3637 | struct fd f = fdget(fd); |
3638 | if (!f.file) | |
3639 | return -EBADF; | |
ac9721f3 | 3640 | |
2903ff01 AV |
3641 | if (f.file->f_op != &perf_fops) { |
3642 | fdput(f); | |
3643 | return -EBADF; | |
ac9721f3 | 3644 | } |
2903ff01 AV |
3645 | *p = f; |
3646 | return 0; | |
ac9721f3 PZ |
3647 | } |
3648 | ||
3649 | static int perf_event_set_output(struct perf_event *event, | |
3650 | struct perf_event *output_event); | |
6fb2915d | 3651 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3652 | |
d859e29f PM |
3653 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3654 | { | |
cdd6c482 IM |
3655 | struct perf_event *event = file->private_data; |
3656 | void (*func)(struct perf_event *); | |
3df5edad | 3657 | u32 flags = arg; |
d859e29f PM |
3658 | |
3659 | switch (cmd) { | |
cdd6c482 IM |
3660 | case PERF_EVENT_IOC_ENABLE: |
3661 | func = perf_event_enable; | |
d859e29f | 3662 | break; |
cdd6c482 IM |
3663 | case PERF_EVENT_IOC_DISABLE: |
3664 | func = perf_event_disable; | |
79f14641 | 3665 | break; |
cdd6c482 IM |
3666 | case PERF_EVENT_IOC_RESET: |
3667 | func = perf_event_reset; | |
6de6a7b9 | 3668 | break; |
3df5edad | 3669 | |
cdd6c482 IM |
3670 | case PERF_EVENT_IOC_REFRESH: |
3671 | return perf_event_refresh(event, arg); | |
08247e31 | 3672 | |
cdd6c482 IM |
3673 | case PERF_EVENT_IOC_PERIOD: |
3674 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3675 | |
cf4957f1 JO |
3676 | case PERF_EVENT_IOC_ID: |
3677 | { | |
3678 | u64 id = primary_event_id(event); | |
3679 | ||
3680 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3681 | return -EFAULT; | |
3682 | return 0; | |
3683 | } | |
3684 | ||
cdd6c482 | 3685 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3686 | { |
ac9721f3 | 3687 | int ret; |
ac9721f3 | 3688 | if (arg != -1) { |
2903ff01 AV |
3689 | struct perf_event *output_event; |
3690 | struct fd output; | |
3691 | ret = perf_fget_light(arg, &output); | |
3692 | if (ret) | |
3693 | return ret; | |
3694 | output_event = output.file->private_data; | |
3695 | ret = perf_event_set_output(event, output_event); | |
3696 | fdput(output); | |
3697 | } else { | |
3698 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3699 | } |
ac9721f3 PZ |
3700 | return ret; |
3701 | } | |
a4be7c27 | 3702 | |
6fb2915d LZ |
3703 | case PERF_EVENT_IOC_SET_FILTER: |
3704 | return perf_event_set_filter(event, (void __user *)arg); | |
3705 | ||
d859e29f | 3706 | default: |
3df5edad | 3707 | return -ENOTTY; |
d859e29f | 3708 | } |
3df5edad PZ |
3709 | |
3710 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3711 | perf_event_for_each(event, func); |
3df5edad | 3712 | else |
cdd6c482 | 3713 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3714 | |
3715 | return 0; | |
d859e29f PM |
3716 | } |
3717 | ||
cdd6c482 | 3718 | int perf_event_task_enable(void) |
771d7cde | 3719 | { |
cdd6c482 | 3720 | struct perf_event *event; |
771d7cde | 3721 | |
cdd6c482 IM |
3722 | mutex_lock(¤t->perf_event_mutex); |
3723 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3724 | perf_event_for_each_child(event, perf_event_enable); | |
3725 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3726 | |
3727 | return 0; | |
3728 | } | |
3729 | ||
cdd6c482 | 3730 | int perf_event_task_disable(void) |
771d7cde | 3731 | { |
cdd6c482 | 3732 | struct perf_event *event; |
771d7cde | 3733 | |
cdd6c482 IM |
3734 | mutex_lock(¤t->perf_event_mutex); |
3735 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3736 | perf_event_for_each_child(event, perf_event_disable); | |
3737 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3738 | |
3739 | return 0; | |
3740 | } | |
3741 | ||
cdd6c482 | 3742 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3743 | { |
a4eaf7f1 PZ |
3744 | if (event->hw.state & PERF_HES_STOPPED) |
3745 | return 0; | |
3746 | ||
cdd6c482 | 3747 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3748 | return 0; |
3749 | ||
35edc2a5 | 3750 | return event->pmu->event_idx(event); |
194002b2 PZ |
3751 | } |
3752 | ||
c4794295 | 3753 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3754 | u64 *now, |
7f310a5d EM |
3755 | u64 *enabled, |
3756 | u64 *running) | |
c4794295 | 3757 | { |
e3f3541c | 3758 | u64 ctx_time; |
c4794295 | 3759 | |
e3f3541c PZ |
3760 | *now = perf_clock(); |
3761 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3762 | *enabled = ctx_time - event->tstamp_enabled; |
3763 | *running = ctx_time - event->tstamp_running; | |
3764 | } | |
3765 | ||
fa731587 PZ |
3766 | static void perf_event_init_userpage(struct perf_event *event) |
3767 | { | |
3768 | struct perf_event_mmap_page *userpg; | |
3769 | struct ring_buffer *rb; | |
3770 | ||
3771 | rcu_read_lock(); | |
3772 | rb = rcu_dereference(event->rb); | |
3773 | if (!rb) | |
3774 | goto unlock; | |
3775 | ||
3776 | userpg = rb->user_page; | |
3777 | ||
3778 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
3779 | userpg->cap_bit0_is_deprecated = 1; | |
3780 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
3781 | ||
3782 | unlock: | |
3783 | rcu_read_unlock(); | |
3784 | } | |
3785 | ||
c7206205 | 3786 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3787 | { |
3788 | } | |
3789 | ||
38ff667b PZ |
3790 | /* |
3791 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3792 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3793 | * code calls this from NMI context. | |
3794 | */ | |
cdd6c482 | 3795 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3796 | { |
cdd6c482 | 3797 | struct perf_event_mmap_page *userpg; |
76369139 | 3798 | struct ring_buffer *rb; |
e3f3541c | 3799 | u64 enabled, running, now; |
38ff667b PZ |
3800 | |
3801 | rcu_read_lock(); | |
5ec4c599 PZ |
3802 | rb = rcu_dereference(event->rb); |
3803 | if (!rb) | |
3804 | goto unlock; | |
3805 | ||
0d641208 EM |
3806 | /* |
3807 | * compute total_time_enabled, total_time_running | |
3808 | * based on snapshot values taken when the event | |
3809 | * was last scheduled in. | |
3810 | * | |
3811 | * we cannot simply called update_context_time() | |
3812 | * because of locking issue as we can be called in | |
3813 | * NMI context | |
3814 | */ | |
e3f3541c | 3815 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 3816 | |
76369139 | 3817 | userpg = rb->user_page; |
7b732a75 PZ |
3818 | /* |
3819 | * Disable preemption so as to not let the corresponding user-space | |
3820 | * spin too long if we get preempted. | |
3821 | */ | |
3822 | preempt_disable(); | |
37d81828 | 3823 | ++userpg->lock; |
92f22a38 | 3824 | barrier(); |
cdd6c482 | 3825 | userpg->index = perf_event_index(event); |
b5e58793 | 3826 | userpg->offset = perf_event_count(event); |
365a4038 | 3827 | if (userpg->index) |
e7850595 | 3828 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3829 | |
0d641208 | 3830 | userpg->time_enabled = enabled + |
cdd6c482 | 3831 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3832 | |
0d641208 | 3833 | userpg->time_running = running + |
cdd6c482 | 3834 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3835 | |
c7206205 | 3836 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3837 | |
92f22a38 | 3838 | barrier(); |
37d81828 | 3839 | ++userpg->lock; |
7b732a75 | 3840 | preempt_enable(); |
38ff667b | 3841 | unlock: |
7b732a75 | 3842 | rcu_read_unlock(); |
37d81828 PM |
3843 | } |
3844 | ||
906010b2 PZ |
3845 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3846 | { | |
3847 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3848 | struct ring_buffer *rb; |
906010b2 PZ |
3849 | int ret = VM_FAULT_SIGBUS; |
3850 | ||
3851 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3852 | if (vmf->pgoff == 0) | |
3853 | ret = 0; | |
3854 | return ret; | |
3855 | } | |
3856 | ||
3857 | rcu_read_lock(); | |
76369139 FW |
3858 | rb = rcu_dereference(event->rb); |
3859 | if (!rb) | |
906010b2 PZ |
3860 | goto unlock; |
3861 | ||
3862 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3863 | goto unlock; | |
3864 | ||
76369139 | 3865 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3866 | if (!vmf->page) |
3867 | goto unlock; | |
3868 | ||
3869 | get_page(vmf->page); | |
3870 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3871 | vmf->page->index = vmf->pgoff; | |
3872 | ||
3873 | ret = 0; | |
3874 | unlock: | |
3875 | rcu_read_unlock(); | |
3876 | ||
3877 | return ret; | |
3878 | } | |
3879 | ||
10c6db11 PZ |
3880 | static void ring_buffer_attach(struct perf_event *event, |
3881 | struct ring_buffer *rb) | |
3882 | { | |
b69cf536 | 3883 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
3884 | unsigned long flags; |
3885 | ||
b69cf536 PZ |
3886 | if (event->rb) { |
3887 | /* | |
3888 | * Should be impossible, we set this when removing | |
3889 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
3890 | */ | |
3891 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 3892 | |
b69cf536 PZ |
3893 | old_rb = event->rb; |
3894 | event->rcu_batches = get_state_synchronize_rcu(); | |
3895 | event->rcu_pending = 1; | |
10c6db11 | 3896 | |
b69cf536 PZ |
3897 | spin_lock_irqsave(&old_rb->event_lock, flags); |
3898 | list_del_rcu(&event->rb_entry); | |
3899 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
3900 | } | |
10c6db11 | 3901 | |
b69cf536 PZ |
3902 | if (event->rcu_pending && rb) { |
3903 | cond_synchronize_rcu(event->rcu_batches); | |
3904 | event->rcu_pending = 0; | |
3905 | } | |
10c6db11 | 3906 | |
b69cf536 PZ |
3907 | if (rb) { |
3908 | spin_lock_irqsave(&rb->event_lock, flags); | |
3909 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
3910 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3911 | } | |
3912 | ||
3913 | rcu_assign_pointer(event->rb, rb); | |
10c6db11 | 3914 | |
b69cf536 PZ |
3915 | if (old_rb) { |
3916 | ring_buffer_put(old_rb); | |
3917 | /* | |
3918 | * Since we detached before setting the new rb, so that we | |
3919 | * could attach the new rb, we could have missed a wakeup. | |
3920 | * Provide it now. | |
3921 | */ | |
3922 | wake_up_all(&event->waitq); | |
3923 | } | |
10c6db11 PZ |
3924 | } |
3925 | ||
3926 | static void ring_buffer_wakeup(struct perf_event *event) | |
3927 | { | |
3928 | struct ring_buffer *rb; | |
3929 | ||
3930 | rcu_read_lock(); | |
3931 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3932 | if (rb) { |
3933 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3934 | wake_up_all(&event->waitq); | |
3935 | } | |
10c6db11 PZ |
3936 | rcu_read_unlock(); |
3937 | } | |
3938 | ||
76369139 | 3939 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3940 | { |
76369139 | 3941 | struct ring_buffer *rb; |
906010b2 | 3942 | |
76369139 FW |
3943 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3944 | rb_free(rb); | |
7b732a75 PZ |
3945 | } |
3946 | ||
76369139 | 3947 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3948 | { |
76369139 | 3949 | struct ring_buffer *rb; |
7b732a75 | 3950 | |
ac9721f3 | 3951 | rcu_read_lock(); |
76369139 FW |
3952 | rb = rcu_dereference(event->rb); |
3953 | if (rb) { | |
3954 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3955 | rb = NULL; | |
ac9721f3 PZ |
3956 | } |
3957 | rcu_read_unlock(); | |
3958 | ||
76369139 | 3959 | return rb; |
ac9721f3 PZ |
3960 | } |
3961 | ||
76369139 | 3962 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3963 | { |
76369139 | 3964 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3965 | return; |
7b732a75 | 3966 | |
9bb5d40c | 3967 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3968 | |
76369139 | 3969 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3970 | } |
3971 | ||
3972 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3973 | { | |
cdd6c482 | 3974 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3975 | |
cdd6c482 | 3976 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3977 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3978 | } |
3979 | ||
9bb5d40c PZ |
3980 | /* |
3981 | * A buffer can be mmap()ed multiple times; either directly through the same | |
3982 | * event, or through other events by use of perf_event_set_output(). | |
3983 | * | |
3984 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
3985 | * the buffer here, where we still have a VM context. This means we need | |
3986 | * to detach all events redirecting to us. | |
3987 | */ | |
7b732a75 PZ |
3988 | static void perf_mmap_close(struct vm_area_struct *vma) |
3989 | { | |
cdd6c482 | 3990 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3991 | |
b69cf536 | 3992 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
3993 | struct user_struct *mmap_user = rb->mmap_user; |
3994 | int mmap_locked = rb->mmap_locked; | |
3995 | unsigned long size = perf_data_size(rb); | |
789f90fc | 3996 | |
9bb5d40c PZ |
3997 | atomic_dec(&rb->mmap_count); |
3998 | ||
3999 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4000 | goto out_put; |
9bb5d40c | 4001 | |
b69cf536 | 4002 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4003 | mutex_unlock(&event->mmap_mutex); |
4004 | ||
4005 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4006 | if (atomic_read(&rb->mmap_count)) |
4007 | goto out_put; | |
ac9721f3 | 4008 | |
9bb5d40c PZ |
4009 | /* |
4010 | * No other mmap()s, detach from all other events that might redirect | |
4011 | * into the now unreachable buffer. Somewhat complicated by the | |
4012 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4013 | */ | |
4014 | again: | |
4015 | rcu_read_lock(); | |
4016 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4017 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4018 | /* | |
4019 | * This event is en-route to free_event() which will | |
4020 | * detach it and remove it from the list. | |
4021 | */ | |
4022 | continue; | |
4023 | } | |
4024 | rcu_read_unlock(); | |
789f90fc | 4025 | |
9bb5d40c PZ |
4026 | mutex_lock(&event->mmap_mutex); |
4027 | /* | |
4028 | * Check we didn't race with perf_event_set_output() which can | |
4029 | * swizzle the rb from under us while we were waiting to | |
4030 | * acquire mmap_mutex. | |
4031 | * | |
4032 | * If we find a different rb; ignore this event, a next | |
4033 | * iteration will no longer find it on the list. We have to | |
4034 | * still restart the iteration to make sure we're not now | |
4035 | * iterating the wrong list. | |
4036 | */ | |
b69cf536 PZ |
4037 | if (event->rb == rb) |
4038 | ring_buffer_attach(event, NULL); | |
4039 | ||
cdd6c482 | 4040 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4041 | put_event(event); |
ac9721f3 | 4042 | |
9bb5d40c PZ |
4043 | /* |
4044 | * Restart the iteration; either we're on the wrong list or | |
4045 | * destroyed its integrity by doing a deletion. | |
4046 | */ | |
4047 | goto again; | |
7b732a75 | 4048 | } |
9bb5d40c PZ |
4049 | rcu_read_unlock(); |
4050 | ||
4051 | /* | |
4052 | * It could be there's still a few 0-ref events on the list; they'll | |
4053 | * get cleaned up by free_event() -- they'll also still have their | |
4054 | * ref on the rb and will free it whenever they are done with it. | |
4055 | * | |
4056 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4057 | * undo the VM accounting. | |
4058 | */ | |
4059 | ||
4060 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4061 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4062 | free_uid(mmap_user); | |
4063 | ||
b69cf536 | 4064 | out_put: |
9bb5d40c | 4065 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4066 | } |
4067 | ||
f0f37e2f | 4068 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4069 | .open = perf_mmap_open, |
4070 | .close = perf_mmap_close, | |
4071 | .fault = perf_mmap_fault, | |
4072 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4073 | }; |
4074 | ||
4075 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4076 | { | |
cdd6c482 | 4077 | struct perf_event *event = file->private_data; |
22a4f650 | 4078 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4079 | struct user_struct *user = current_user(); |
22a4f650 | 4080 | unsigned long locked, lock_limit; |
76369139 | 4081 | struct ring_buffer *rb; |
7b732a75 PZ |
4082 | unsigned long vma_size; |
4083 | unsigned long nr_pages; | |
789f90fc | 4084 | long user_extra, extra; |
d57e34fd | 4085 | int ret = 0, flags = 0; |
37d81828 | 4086 | |
c7920614 PZ |
4087 | /* |
4088 | * Don't allow mmap() of inherited per-task counters. This would | |
4089 | * create a performance issue due to all children writing to the | |
76369139 | 4090 | * same rb. |
c7920614 PZ |
4091 | */ |
4092 | if (event->cpu == -1 && event->attr.inherit) | |
4093 | return -EINVAL; | |
4094 | ||
43a21ea8 | 4095 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4096 | return -EINVAL; |
7b732a75 PZ |
4097 | |
4098 | vma_size = vma->vm_end - vma->vm_start; | |
4099 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4100 | ||
7730d865 | 4101 | /* |
76369139 | 4102 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4103 | * can do bitmasks instead of modulo. |
4104 | */ | |
4105 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
4106 | return -EINVAL; |
4107 | ||
7b732a75 | 4108 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4109 | return -EINVAL; |
4110 | ||
7b732a75 PZ |
4111 | if (vma->vm_pgoff != 0) |
4112 | return -EINVAL; | |
37d81828 | 4113 | |
cdd6c482 | 4114 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4115 | again: |
cdd6c482 | 4116 | mutex_lock(&event->mmap_mutex); |
76369139 | 4117 | if (event->rb) { |
9bb5d40c | 4118 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4119 | ret = -EINVAL; |
9bb5d40c PZ |
4120 | goto unlock; |
4121 | } | |
4122 | ||
4123 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4124 | /* | |
4125 | * Raced against perf_mmap_close() through | |
4126 | * perf_event_set_output(). Try again, hope for better | |
4127 | * luck. | |
4128 | */ | |
4129 | mutex_unlock(&event->mmap_mutex); | |
4130 | goto again; | |
4131 | } | |
4132 | ||
ebb3c4c4 PZ |
4133 | goto unlock; |
4134 | } | |
4135 | ||
789f90fc | 4136 | user_extra = nr_pages + 1; |
cdd6c482 | 4137 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4138 | |
4139 | /* | |
4140 | * Increase the limit linearly with more CPUs: | |
4141 | */ | |
4142 | user_lock_limit *= num_online_cpus(); | |
4143 | ||
789f90fc | 4144 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4145 | |
789f90fc PZ |
4146 | extra = 0; |
4147 | if (user_locked > user_lock_limit) | |
4148 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4149 | |
78d7d407 | 4150 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4151 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4152 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4153 | |
459ec28a IM |
4154 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4155 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4156 | ret = -EPERM; |
4157 | goto unlock; | |
4158 | } | |
7b732a75 | 4159 | |
76369139 | 4160 | WARN_ON(event->rb); |
906010b2 | 4161 | |
d57e34fd | 4162 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4163 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4164 | |
4ec8363d VW |
4165 | rb = rb_alloc(nr_pages, |
4166 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4167 | event->cpu, flags); | |
4168 | ||
76369139 | 4169 | if (!rb) { |
ac9721f3 | 4170 | ret = -ENOMEM; |
ebb3c4c4 | 4171 | goto unlock; |
ac9721f3 | 4172 | } |
26cb63ad | 4173 | |
9bb5d40c | 4174 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4175 | rb->mmap_locked = extra; |
4176 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4177 | |
ac9721f3 | 4178 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4179 | vma->vm_mm->pinned_vm += extra; |
4180 | ||
9bb5d40c | 4181 | ring_buffer_attach(event, rb); |
ac9721f3 | 4182 | |
fa731587 | 4183 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4184 | perf_event_update_userpage(event); |
4185 | ||
ebb3c4c4 | 4186 | unlock: |
ac9721f3 PZ |
4187 | if (!ret) |
4188 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4189 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4190 | |
9bb5d40c PZ |
4191 | /* |
4192 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4193 | * vma. | |
4194 | */ | |
26cb63ad | 4195 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4196 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4197 | |
4198 | return ret; | |
37d81828 PM |
4199 | } |
4200 | ||
3c446b3d PZ |
4201 | static int perf_fasync(int fd, struct file *filp, int on) |
4202 | { | |
496ad9aa | 4203 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4204 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4205 | int retval; |
4206 | ||
4207 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4208 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4209 | mutex_unlock(&inode->i_mutex); |
4210 | ||
4211 | if (retval < 0) | |
4212 | return retval; | |
4213 | ||
4214 | return 0; | |
4215 | } | |
4216 | ||
0793a61d | 4217 | static const struct file_operations perf_fops = { |
3326c1ce | 4218 | .llseek = no_llseek, |
0793a61d TG |
4219 | .release = perf_release, |
4220 | .read = perf_read, | |
4221 | .poll = perf_poll, | |
d859e29f PM |
4222 | .unlocked_ioctl = perf_ioctl, |
4223 | .compat_ioctl = perf_ioctl, | |
37d81828 | 4224 | .mmap = perf_mmap, |
3c446b3d | 4225 | .fasync = perf_fasync, |
0793a61d TG |
4226 | }; |
4227 | ||
925d519a | 4228 | /* |
cdd6c482 | 4229 | * Perf event wakeup |
925d519a PZ |
4230 | * |
4231 | * If there's data, ensure we set the poll() state and publish everything | |
4232 | * to user-space before waking everybody up. | |
4233 | */ | |
4234 | ||
cdd6c482 | 4235 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4236 | { |
10c6db11 | 4237 | ring_buffer_wakeup(event); |
4c9e2542 | 4238 | |
cdd6c482 IM |
4239 | if (event->pending_kill) { |
4240 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4241 | event->pending_kill = 0; | |
4c9e2542 | 4242 | } |
925d519a PZ |
4243 | } |
4244 | ||
e360adbe | 4245 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4246 | { |
cdd6c482 IM |
4247 | struct perf_event *event = container_of(entry, |
4248 | struct perf_event, pending); | |
79f14641 | 4249 | |
cdd6c482 IM |
4250 | if (event->pending_disable) { |
4251 | event->pending_disable = 0; | |
4252 | __perf_event_disable(event); | |
79f14641 PZ |
4253 | } |
4254 | ||
cdd6c482 IM |
4255 | if (event->pending_wakeup) { |
4256 | event->pending_wakeup = 0; | |
4257 | perf_event_wakeup(event); | |
79f14641 PZ |
4258 | } |
4259 | } | |
4260 | ||
39447b38 ZY |
4261 | /* |
4262 | * We assume there is only KVM supporting the callbacks. | |
4263 | * Later on, we might change it to a list if there is | |
4264 | * another virtualization implementation supporting the callbacks. | |
4265 | */ | |
4266 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4267 | ||
4268 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4269 | { | |
4270 | perf_guest_cbs = cbs; | |
4271 | return 0; | |
4272 | } | |
4273 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4274 | ||
4275 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4276 | { | |
4277 | perf_guest_cbs = NULL; | |
4278 | return 0; | |
4279 | } | |
4280 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4281 | ||
4018994f JO |
4282 | static void |
4283 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4284 | struct pt_regs *regs, u64 mask) | |
4285 | { | |
4286 | int bit; | |
4287 | ||
4288 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4289 | sizeof(mask) * BITS_PER_BYTE) { | |
4290 | u64 val; | |
4291 | ||
4292 | val = perf_reg_value(regs, bit); | |
4293 | perf_output_put(handle, val); | |
4294 | } | |
4295 | } | |
4296 | ||
4297 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4298 | struct pt_regs *regs) | |
4299 | { | |
4300 | if (!user_mode(regs)) { | |
4301 | if (current->mm) | |
4302 | regs = task_pt_regs(current); | |
4303 | else | |
4304 | regs = NULL; | |
4305 | } | |
4306 | ||
4307 | if (regs) { | |
4308 | regs_user->regs = regs; | |
4309 | regs_user->abi = perf_reg_abi(current); | |
4310 | } | |
4311 | } | |
4312 | ||
c5ebcedb JO |
4313 | /* |
4314 | * Get remaining task size from user stack pointer. | |
4315 | * | |
4316 | * It'd be better to take stack vma map and limit this more | |
4317 | * precisly, but there's no way to get it safely under interrupt, | |
4318 | * so using TASK_SIZE as limit. | |
4319 | */ | |
4320 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4321 | { | |
4322 | unsigned long addr = perf_user_stack_pointer(regs); | |
4323 | ||
4324 | if (!addr || addr >= TASK_SIZE) | |
4325 | return 0; | |
4326 | ||
4327 | return TASK_SIZE - addr; | |
4328 | } | |
4329 | ||
4330 | static u16 | |
4331 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4332 | struct pt_regs *regs) | |
4333 | { | |
4334 | u64 task_size; | |
4335 | ||
4336 | /* No regs, no stack pointer, no dump. */ | |
4337 | if (!regs) | |
4338 | return 0; | |
4339 | ||
4340 | /* | |
4341 | * Check if we fit in with the requested stack size into the: | |
4342 | * - TASK_SIZE | |
4343 | * If we don't, we limit the size to the TASK_SIZE. | |
4344 | * | |
4345 | * - remaining sample size | |
4346 | * If we don't, we customize the stack size to | |
4347 | * fit in to the remaining sample size. | |
4348 | */ | |
4349 | ||
4350 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4351 | stack_size = min(stack_size, (u16) task_size); | |
4352 | ||
4353 | /* Current header size plus static size and dynamic size. */ | |
4354 | header_size += 2 * sizeof(u64); | |
4355 | ||
4356 | /* Do we fit in with the current stack dump size? */ | |
4357 | if ((u16) (header_size + stack_size) < header_size) { | |
4358 | /* | |
4359 | * If we overflow the maximum size for the sample, | |
4360 | * we customize the stack dump size to fit in. | |
4361 | */ | |
4362 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4363 | stack_size = round_up(stack_size, sizeof(u64)); | |
4364 | } | |
4365 | ||
4366 | return stack_size; | |
4367 | } | |
4368 | ||
4369 | static void | |
4370 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4371 | struct pt_regs *regs) | |
4372 | { | |
4373 | /* Case of a kernel thread, nothing to dump */ | |
4374 | if (!regs) { | |
4375 | u64 size = 0; | |
4376 | perf_output_put(handle, size); | |
4377 | } else { | |
4378 | unsigned long sp; | |
4379 | unsigned int rem; | |
4380 | u64 dyn_size; | |
4381 | ||
4382 | /* | |
4383 | * We dump: | |
4384 | * static size | |
4385 | * - the size requested by user or the best one we can fit | |
4386 | * in to the sample max size | |
4387 | * data | |
4388 | * - user stack dump data | |
4389 | * dynamic size | |
4390 | * - the actual dumped size | |
4391 | */ | |
4392 | ||
4393 | /* Static size. */ | |
4394 | perf_output_put(handle, dump_size); | |
4395 | ||
4396 | /* Data. */ | |
4397 | sp = perf_user_stack_pointer(regs); | |
4398 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4399 | dyn_size = dump_size - rem; | |
4400 | ||
4401 | perf_output_skip(handle, rem); | |
4402 | ||
4403 | /* Dynamic size. */ | |
4404 | perf_output_put(handle, dyn_size); | |
4405 | } | |
4406 | } | |
4407 | ||
c980d109 ACM |
4408 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4409 | struct perf_sample_data *data, | |
4410 | struct perf_event *event) | |
6844c09d ACM |
4411 | { |
4412 | u64 sample_type = event->attr.sample_type; | |
4413 | ||
4414 | data->type = sample_type; | |
4415 | header->size += event->id_header_size; | |
4416 | ||
4417 | if (sample_type & PERF_SAMPLE_TID) { | |
4418 | /* namespace issues */ | |
4419 | data->tid_entry.pid = perf_event_pid(event, current); | |
4420 | data->tid_entry.tid = perf_event_tid(event, current); | |
4421 | } | |
4422 | ||
4423 | if (sample_type & PERF_SAMPLE_TIME) | |
4424 | data->time = perf_clock(); | |
4425 | ||
ff3d527c | 4426 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4427 | data->id = primary_event_id(event); |
4428 | ||
4429 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4430 | data->stream_id = event->id; | |
4431 | ||
4432 | if (sample_type & PERF_SAMPLE_CPU) { | |
4433 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4434 | data->cpu_entry.reserved = 0; | |
4435 | } | |
4436 | } | |
4437 | ||
76369139 FW |
4438 | void perf_event_header__init_id(struct perf_event_header *header, |
4439 | struct perf_sample_data *data, | |
4440 | struct perf_event *event) | |
c980d109 ACM |
4441 | { |
4442 | if (event->attr.sample_id_all) | |
4443 | __perf_event_header__init_id(header, data, event); | |
4444 | } | |
4445 | ||
4446 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4447 | struct perf_sample_data *data) | |
4448 | { | |
4449 | u64 sample_type = data->type; | |
4450 | ||
4451 | if (sample_type & PERF_SAMPLE_TID) | |
4452 | perf_output_put(handle, data->tid_entry); | |
4453 | ||
4454 | if (sample_type & PERF_SAMPLE_TIME) | |
4455 | perf_output_put(handle, data->time); | |
4456 | ||
4457 | if (sample_type & PERF_SAMPLE_ID) | |
4458 | perf_output_put(handle, data->id); | |
4459 | ||
4460 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4461 | perf_output_put(handle, data->stream_id); | |
4462 | ||
4463 | if (sample_type & PERF_SAMPLE_CPU) | |
4464 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4465 | |
4466 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4467 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4468 | } |
4469 | ||
76369139 FW |
4470 | void perf_event__output_id_sample(struct perf_event *event, |
4471 | struct perf_output_handle *handle, | |
4472 | struct perf_sample_data *sample) | |
c980d109 ACM |
4473 | { |
4474 | if (event->attr.sample_id_all) | |
4475 | __perf_event__output_id_sample(handle, sample); | |
4476 | } | |
4477 | ||
3dab77fb | 4478 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4479 | struct perf_event *event, |
4480 | u64 enabled, u64 running) | |
3dab77fb | 4481 | { |
cdd6c482 | 4482 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4483 | u64 values[4]; |
4484 | int n = 0; | |
4485 | ||
b5e58793 | 4486 | values[n++] = perf_event_count(event); |
3dab77fb | 4487 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4488 | values[n++] = enabled + |
cdd6c482 | 4489 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4490 | } |
4491 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4492 | values[n++] = running + |
cdd6c482 | 4493 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4494 | } |
4495 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4496 | values[n++] = primary_event_id(event); |
3dab77fb | 4497 | |
76369139 | 4498 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4499 | } |
4500 | ||
4501 | /* | |
cdd6c482 | 4502 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4503 | */ |
4504 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4505 | struct perf_event *event, |
4506 | u64 enabled, u64 running) | |
3dab77fb | 4507 | { |
cdd6c482 IM |
4508 | struct perf_event *leader = event->group_leader, *sub; |
4509 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4510 | u64 values[5]; |
4511 | int n = 0; | |
4512 | ||
4513 | values[n++] = 1 + leader->nr_siblings; | |
4514 | ||
4515 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4516 | values[n++] = enabled; |
3dab77fb PZ |
4517 | |
4518 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4519 | values[n++] = running; |
3dab77fb | 4520 | |
cdd6c482 | 4521 | if (leader != event) |
3dab77fb PZ |
4522 | leader->pmu->read(leader); |
4523 | ||
b5e58793 | 4524 | values[n++] = perf_event_count(leader); |
3dab77fb | 4525 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4526 | values[n++] = primary_event_id(leader); |
3dab77fb | 4527 | |
76369139 | 4528 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4529 | |
65abc865 | 4530 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4531 | n = 0; |
4532 | ||
6f5ab001 JO |
4533 | if ((sub != event) && |
4534 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4535 | sub->pmu->read(sub); |
4536 | ||
b5e58793 | 4537 | values[n++] = perf_event_count(sub); |
3dab77fb | 4538 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4539 | values[n++] = primary_event_id(sub); |
3dab77fb | 4540 | |
76369139 | 4541 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4542 | } |
4543 | } | |
4544 | ||
eed01528 SE |
4545 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4546 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4547 | ||
3dab77fb | 4548 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4549 | struct perf_event *event) |
3dab77fb | 4550 | { |
e3f3541c | 4551 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4552 | u64 read_format = event->attr.read_format; |
4553 | ||
4554 | /* | |
4555 | * compute total_time_enabled, total_time_running | |
4556 | * based on snapshot values taken when the event | |
4557 | * was last scheduled in. | |
4558 | * | |
4559 | * we cannot simply called update_context_time() | |
4560 | * because of locking issue as we are called in | |
4561 | * NMI context | |
4562 | */ | |
c4794295 | 4563 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4564 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4565 | |
cdd6c482 | 4566 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4567 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4568 | else |
eed01528 | 4569 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4570 | } |
4571 | ||
5622f295 MM |
4572 | void perf_output_sample(struct perf_output_handle *handle, |
4573 | struct perf_event_header *header, | |
4574 | struct perf_sample_data *data, | |
cdd6c482 | 4575 | struct perf_event *event) |
5622f295 MM |
4576 | { |
4577 | u64 sample_type = data->type; | |
4578 | ||
4579 | perf_output_put(handle, *header); | |
4580 | ||
ff3d527c AH |
4581 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4582 | perf_output_put(handle, data->id); | |
4583 | ||
5622f295 MM |
4584 | if (sample_type & PERF_SAMPLE_IP) |
4585 | perf_output_put(handle, data->ip); | |
4586 | ||
4587 | if (sample_type & PERF_SAMPLE_TID) | |
4588 | perf_output_put(handle, data->tid_entry); | |
4589 | ||
4590 | if (sample_type & PERF_SAMPLE_TIME) | |
4591 | perf_output_put(handle, data->time); | |
4592 | ||
4593 | if (sample_type & PERF_SAMPLE_ADDR) | |
4594 | perf_output_put(handle, data->addr); | |
4595 | ||
4596 | if (sample_type & PERF_SAMPLE_ID) | |
4597 | perf_output_put(handle, data->id); | |
4598 | ||
4599 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4600 | perf_output_put(handle, data->stream_id); | |
4601 | ||
4602 | if (sample_type & PERF_SAMPLE_CPU) | |
4603 | perf_output_put(handle, data->cpu_entry); | |
4604 | ||
4605 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4606 | perf_output_put(handle, data->period); | |
4607 | ||
4608 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4609 | perf_output_read(handle, event); |
5622f295 MM |
4610 | |
4611 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4612 | if (data->callchain) { | |
4613 | int size = 1; | |
4614 | ||
4615 | if (data->callchain) | |
4616 | size += data->callchain->nr; | |
4617 | ||
4618 | size *= sizeof(u64); | |
4619 | ||
76369139 | 4620 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4621 | } else { |
4622 | u64 nr = 0; | |
4623 | perf_output_put(handle, nr); | |
4624 | } | |
4625 | } | |
4626 | ||
4627 | if (sample_type & PERF_SAMPLE_RAW) { | |
4628 | if (data->raw) { | |
4629 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4630 | __output_copy(handle, data->raw->data, |
4631 | data->raw->size); | |
5622f295 MM |
4632 | } else { |
4633 | struct { | |
4634 | u32 size; | |
4635 | u32 data; | |
4636 | } raw = { | |
4637 | .size = sizeof(u32), | |
4638 | .data = 0, | |
4639 | }; | |
4640 | perf_output_put(handle, raw); | |
4641 | } | |
4642 | } | |
a7ac67ea | 4643 | |
bce38cd5 SE |
4644 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4645 | if (data->br_stack) { | |
4646 | size_t size; | |
4647 | ||
4648 | size = data->br_stack->nr | |
4649 | * sizeof(struct perf_branch_entry); | |
4650 | ||
4651 | perf_output_put(handle, data->br_stack->nr); | |
4652 | perf_output_copy(handle, data->br_stack->entries, size); | |
4653 | } else { | |
4654 | /* | |
4655 | * we always store at least the value of nr | |
4656 | */ | |
4657 | u64 nr = 0; | |
4658 | perf_output_put(handle, nr); | |
4659 | } | |
4660 | } | |
4018994f JO |
4661 | |
4662 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4663 | u64 abi = data->regs_user.abi; | |
4664 | ||
4665 | /* | |
4666 | * If there are no regs to dump, notice it through | |
4667 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4668 | */ | |
4669 | perf_output_put(handle, abi); | |
4670 | ||
4671 | if (abi) { | |
4672 | u64 mask = event->attr.sample_regs_user; | |
4673 | perf_output_sample_regs(handle, | |
4674 | data->regs_user.regs, | |
4675 | mask); | |
4676 | } | |
4677 | } | |
c5ebcedb | 4678 | |
a5cdd40c | 4679 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
4680 | perf_output_sample_ustack(handle, |
4681 | data->stack_user_size, | |
4682 | data->regs_user.regs); | |
a5cdd40c | 4683 | } |
c3feedf2 AK |
4684 | |
4685 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4686 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4687 | |
4688 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4689 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 4690 | |
fdfbbd07 AK |
4691 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
4692 | perf_output_put(handle, data->txn); | |
4693 | ||
a5cdd40c PZ |
4694 | if (!event->attr.watermark) { |
4695 | int wakeup_events = event->attr.wakeup_events; | |
4696 | ||
4697 | if (wakeup_events) { | |
4698 | struct ring_buffer *rb = handle->rb; | |
4699 | int events = local_inc_return(&rb->events); | |
4700 | ||
4701 | if (events >= wakeup_events) { | |
4702 | local_sub(wakeup_events, &rb->events); | |
4703 | local_inc(&rb->wakeup); | |
4704 | } | |
4705 | } | |
4706 | } | |
5622f295 MM |
4707 | } |
4708 | ||
4709 | void perf_prepare_sample(struct perf_event_header *header, | |
4710 | struct perf_sample_data *data, | |
cdd6c482 | 4711 | struct perf_event *event, |
5622f295 | 4712 | struct pt_regs *regs) |
7b732a75 | 4713 | { |
cdd6c482 | 4714 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4715 | |
cdd6c482 | 4716 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4717 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4718 | |
4719 | header->misc = 0; | |
4720 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4721 | |
c980d109 | 4722 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4723 | |
c320c7b7 | 4724 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4725 | data->ip = perf_instruction_pointer(regs); |
4726 | ||
b23f3325 | 4727 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4728 | int size = 1; |
394ee076 | 4729 | |
e6dab5ff | 4730 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4731 | |
4732 | if (data->callchain) | |
4733 | size += data->callchain->nr; | |
4734 | ||
4735 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4736 | } |
4737 | ||
3a43ce68 | 4738 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4739 | int size = sizeof(u32); |
4740 | ||
4741 | if (data->raw) | |
4742 | size += data->raw->size; | |
4743 | else | |
4744 | size += sizeof(u32); | |
4745 | ||
4746 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4747 | header->size += size; |
7f453c24 | 4748 | } |
bce38cd5 SE |
4749 | |
4750 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4751 | int size = sizeof(u64); /* nr */ | |
4752 | if (data->br_stack) { | |
4753 | size += data->br_stack->nr | |
4754 | * sizeof(struct perf_branch_entry); | |
4755 | } | |
4756 | header->size += size; | |
4757 | } | |
4018994f JO |
4758 | |
4759 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4760 | /* regs dump ABI info */ | |
4761 | int size = sizeof(u64); | |
4762 | ||
4763 | perf_sample_regs_user(&data->regs_user, regs); | |
4764 | ||
4765 | if (data->regs_user.regs) { | |
4766 | u64 mask = event->attr.sample_regs_user; | |
4767 | size += hweight64(mask) * sizeof(u64); | |
4768 | } | |
4769 | ||
4770 | header->size += size; | |
4771 | } | |
c5ebcedb JO |
4772 | |
4773 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4774 | /* | |
4775 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4776 | * processed as the last one or have additional check added | |
4777 | * in case new sample type is added, because we could eat | |
4778 | * up the rest of the sample size. | |
4779 | */ | |
4780 | struct perf_regs_user *uregs = &data->regs_user; | |
4781 | u16 stack_size = event->attr.sample_stack_user; | |
4782 | u16 size = sizeof(u64); | |
4783 | ||
4784 | if (!uregs->abi) | |
4785 | perf_sample_regs_user(uregs, regs); | |
4786 | ||
4787 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4788 | uregs->regs); | |
4789 | ||
4790 | /* | |
4791 | * If there is something to dump, add space for the dump | |
4792 | * itself and for the field that tells the dynamic size, | |
4793 | * which is how many have been actually dumped. | |
4794 | */ | |
4795 | if (stack_size) | |
4796 | size += sizeof(u64) + stack_size; | |
4797 | ||
4798 | data->stack_user_size = stack_size; | |
4799 | header->size += size; | |
4800 | } | |
5622f295 | 4801 | } |
7f453c24 | 4802 | |
a8b0ca17 | 4803 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4804 | struct perf_sample_data *data, |
4805 | struct pt_regs *regs) | |
4806 | { | |
4807 | struct perf_output_handle handle; | |
4808 | struct perf_event_header header; | |
689802b2 | 4809 | |
927c7a9e FW |
4810 | /* protect the callchain buffers */ |
4811 | rcu_read_lock(); | |
4812 | ||
cdd6c482 | 4813 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4814 | |
a7ac67ea | 4815 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4816 | goto exit; |
0322cd6e | 4817 | |
cdd6c482 | 4818 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4819 | |
8a057d84 | 4820 | perf_output_end(&handle); |
927c7a9e FW |
4821 | |
4822 | exit: | |
4823 | rcu_read_unlock(); | |
0322cd6e PZ |
4824 | } |
4825 | ||
38b200d6 | 4826 | /* |
cdd6c482 | 4827 | * read event_id |
38b200d6 PZ |
4828 | */ |
4829 | ||
4830 | struct perf_read_event { | |
4831 | struct perf_event_header header; | |
4832 | ||
4833 | u32 pid; | |
4834 | u32 tid; | |
38b200d6 PZ |
4835 | }; |
4836 | ||
4837 | static void | |
cdd6c482 | 4838 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4839 | struct task_struct *task) |
4840 | { | |
4841 | struct perf_output_handle handle; | |
c980d109 | 4842 | struct perf_sample_data sample; |
dfc65094 | 4843 | struct perf_read_event read_event = { |
38b200d6 | 4844 | .header = { |
cdd6c482 | 4845 | .type = PERF_RECORD_READ, |
38b200d6 | 4846 | .misc = 0, |
c320c7b7 | 4847 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4848 | }, |
cdd6c482 IM |
4849 | .pid = perf_event_pid(event, task), |
4850 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4851 | }; |
3dab77fb | 4852 | int ret; |
38b200d6 | 4853 | |
c980d109 | 4854 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4855 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4856 | if (ret) |
4857 | return; | |
4858 | ||
dfc65094 | 4859 | perf_output_put(&handle, read_event); |
cdd6c482 | 4860 | perf_output_read(&handle, event); |
c980d109 | 4861 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4862 | |
38b200d6 PZ |
4863 | perf_output_end(&handle); |
4864 | } | |
4865 | ||
52d857a8 JO |
4866 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
4867 | ||
4868 | static void | |
4869 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
4870 | perf_event_aux_output_cb output, |
4871 | void *data) | |
4872 | { | |
4873 | struct perf_event *event; | |
4874 | ||
4875 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4876 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4877 | continue; | |
4878 | if (!event_filter_match(event)) | |
4879 | continue; | |
67516844 | 4880 | output(event, data); |
52d857a8 JO |
4881 | } |
4882 | } | |
4883 | ||
4884 | static void | |
67516844 | 4885 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
4886 | struct perf_event_context *task_ctx) |
4887 | { | |
4888 | struct perf_cpu_context *cpuctx; | |
4889 | struct perf_event_context *ctx; | |
4890 | struct pmu *pmu; | |
4891 | int ctxn; | |
4892 | ||
4893 | rcu_read_lock(); | |
4894 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4895 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4896 | if (cpuctx->unique_pmu != pmu) | |
4897 | goto next; | |
67516844 | 4898 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
4899 | if (task_ctx) |
4900 | goto next; | |
4901 | ctxn = pmu->task_ctx_nr; | |
4902 | if (ctxn < 0) | |
4903 | goto next; | |
4904 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4905 | if (ctx) | |
67516844 | 4906 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
4907 | next: |
4908 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4909 | } | |
4910 | ||
4911 | if (task_ctx) { | |
4912 | preempt_disable(); | |
67516844 | 4913 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
4914 | preempt_enable(); |
4915 | } | |
4916 | rcu_read_unlock(); | |
4917 | } | |
4918 | ||
60313ebe | 4919 | /* |
9f498cc5 PZ |
4920 | * task tracking -- fork/exit |
4921 | * | |
13d7a241 | 4922 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
4923 | */ |
4924 | ||
9f498cc5 | 4925 | struct perf_task_event { |
3a80b4a3 | 4926 | struct task_struct *task; |
cdd6c482 | 4927 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4928 | |
4929 | struct { | |
4930 | struct perf_event_header header; | |
4931 | ||
4932 | u32 pid; | |
4933 | u32 ppid; | |
9f498cc5 PZ |
4934 | u32 tid; |
4935 | u32 ptid; | |
393b2ad8 | 4936 | u64 time; |
cdd6c482 | 4937 | } event_id; |
60313ebe PZ |
4938 | }; |
4939 | ||
67516844 JO |
4940 | static int perf_event_task_match(struct perf_event *event) |
4941 | { | |
13d7a241 SE |
4942 | return event->attr.comm || event->attr.mmap || |
4943 | event->attr.mmap2 || event->attr.mmap_data || | |
4944 | event->attr.task; | |
67516844 JO |
4945 | } |
4946 | ||
cdd6c482 | 4947 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4948 | void *data) |
60313ebe | 4949 | { |
52d857a8 | 4950 | struct perf_task_event *task_event = data; |
60313ebe | 4951 | struct perf_output_handle handle; |
c980d109 | 4952 | struct perf_sample_data sample; |
9f498cc5 | 4953 | struct task_struct *task = task_event->task; |
c980d109 | 4954 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4955 | |
67516844 JO |
4956 | if (!perf_event_task_match(event)) |
4957 | return; | |
4958 | ||
c980d109 | 4959 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4960 | |
c980d109 | 4961 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4962 | task_event->event_id.header.size); |
ef60777c | 4963 | if (ret) |
c980d109 | 4964 | goto out; |
60313ebe | 4965 | |
cdd6c482 IM |
4966 | task_event->event_id.pid = perf_event_pid(event, task); |
4967 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4968 | |
cdd6c482 IM |
4969 | task_event->event_id.tid = perf_event_tid(event, task); |
4970 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4971 | |
cdd6c482 | 4972 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4973 | |
c980d109 ACM |
4974 | perf_event__output_id_sample(event, &handle, &sample); |
4975 | ||
60313ebe | 4976 | perf_output_end(&handle); |
c980d109 ACM |
4977 | out: |
4978 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4979 | } |
4980 | ||
cdd6c482 IM |
4981 | static void perf_event_task(struct task_struct *task, |
4982 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4983 | int new) |
60313ebe | 4984 | { |
9f498cc5 | 4985 | struct perf_task_event task_event; |
60313ebe | 4986 | |
cdd6c482 IM |
4987 | if (!atomic_read(&nr_comm_events) && |
4988 | !atomic_read(&nr_mmap_events) && | |
4989 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4990 | return; |
4991 | ||
9f498cc5 | 4992 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4993 | .task = task, |
4994 | .task_ctx = task_ctx, | |
cdd6c482 | 4995 | .event_id = { |
60313ebe | 4996 | .header = { |
cdd6c482 | 4997 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4998 | .misc = 0, |
cdd6c482 | 4999 | .size = sizeof(task_event.event_id), |
60313ebe | 5000 | }, |
573402db PZ |
5001 | /* .pid */ |
5002 | /* .ppid */ | |
9f498cc5 PZ |
5003 | /* .tid */ |
5004 | /* .ptid */ | |
6f93d0a7 | 5005 | .time = perf_clock(), |
60313ebe PZ |
5006 | }, |
5007 | }; | |
5008 | ||
67516844 | 5009 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5010 | &task_event, |
5011 | task_ctx); | |
9f498cc5 PZ |
5012 | } |
5013 | ||
cdd6c482 | 5014 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5015 | { |
cdd6c482 | 5016 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5017 | } |
5018 | ||
8d1b2d93 PZ |
5019 | /* |
5020 | * comm tracking | |
5021 | */ | |
5022 | ||
5023 | struct perf_comm_event { | |
22a4f650 IM |
5024 | struct task_struct *task; |
5025 | char *comm; | |
8d1b2d93 PZ |
5026 | int comm_size; |
5027 | ||
5028 | struct { | |
5029 | struct perf_event_header header; | |
5030 | ||
5031 | u32 pid; | |
5032 | u32 tid; | |
cdd6c482 | 5033 | } event_id; |
8d1b2d93 PZ |
5034 | }; |
5035 | ||
67516844 JO |
5036 | static int perf_event_comm_match(struct perf_event *event) |
5037 | { | |
5038 | return event->attr.comm; | |
5039 | } | |
5040 | ||
cdd6c482 | 5041 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5042 | void *data) |
8d1b2d93 | 5043 | { |
52d857a8 | 5044 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5045 | struct perf_output_handle handle; |
c980d109 | 5046 | struct perf_sample_data sample; |
cdd6c482 | 5047 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5048 | int ret; |
5049 | ||
67516844 JO |
5050 | if (!perf_event_comm_match(event)) |
5051 | return; | |
5052 | ||
c980d109 ACM |
5053 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5054 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5055 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5056 | |
5057 | if (ret) | |
c980d109 | 5058 | goto out; |
8d1b2d93 | 5059 | |
cdd6c482 IM |
5060 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5061 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5062 | |
cdd6c482 | 5063 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5064 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5065 | comm_event->comm_size); |
c980d109 ACM |
5066 | |
5067 | perf_event__output_id_sample(event, &handle, &sample); | |
5068 | ||
8d1b2d93 | 5069 | perf_output_end(&handle); |
c980d109 ACM |
5070 | out: |
5071 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5072 | } |
5073 | ||
cdd6c482 | 5074 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5075 | { |
413ee3b4 | 5076 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5077 | unsigned int size; |
8d1b2d93 | 5078 | |
413ee3b4 | 5079 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5080 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5081 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5082 | |
5083 | comm_event->comm = comm; | |
5084 | comm_event->comm_size = size; | |
5085 | ||
cdd6c482 | 5086 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5087 | |
67516844 | 5088 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5089 | comm_event, |
5090 | NULL); | |
8d1b2d93 PZ |
5091 | } |
5092 | ||
82b89778 | 5093 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5094 | { |
9ee318a7 PZ |
5095 | struct perf_comm_event comm_event; |
5096 | ||
cdd6c482 | 5097 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5098 | return; |
a63eaf34 | 5099 | |
9ee318a7 | 5100 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5101 | .task = task, |
573402db PZ |
5102 | /* .comm */ |
5103 | /* .comm_size */ | |
cdd6c482 | 5104 | .event_id = { |
573402db | 5105 | .header = { |
cdd6c482 | 5106 | .type = PERF_RECORD_COMM, |
82b89778 | 5107 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5108 | /* .size */ |
5109 | }, | |
5110 | /* .pid */ | |
5111 | /* .tid */ | |
8d1b2d93 PZ |
5112 | }, |
5113 | }; | |
5114 | ||
cdd6c482 | 5115 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5116 | } |
5117 | ||
0a4a9391 PZ |
5118 | /* |
5119 | * mmap tracking | |
5120 | */ | |
5121 | ||
5122 | struct perf_mmap_event { | |
089dd79d PZ |
5123 | struct vm_area_struct *vma; |
5124 | ||
5125 | const char *file_name; | |
5126 | int file_size; | |
13d7a241 SE |
5127 | int maj, min; |
5128 | u64 ino; | |
5129 | u64 ino_generation; | |
0a4a9391 PZ |
5130 | |
5131 | struct { | |
5132 | struct perf_event_header header; | |
5133 | ||
5134 | u32 pid; | |
5135 | u32 tid; | |
5136 | u64 start; | |
5137 | u64 len; | |
5138 | u64 pgoff; | |
cdd6c482 | 5139 | } event_id; |
0a4a9391 PZ |
5140 | }; |
5141 | ||
67516844 JO |
5142 | static int perf_event_mmap_match(struct perf_event *event, |
5143 | void *data) | |
5144 | { | |
5145 | struct perf_mmap_event *mmap_event = data; | |
5146 | struct vm_area_struct *vma = mmap_event->vma; | |
5147 | int executable = vma->vm_flags & VM_EXEC; | |
5148 | ||
5149 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5150 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5151 | } |
5152 | ||
cdd6c482 | 5153 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5154 | void *data) |
0a4a9391 | 5155 | { |
52d857a8 | 5156 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5157 | struct perf_output_handle handle; |
c980d109 | 5158 | struct perf_sample_data sample; |
cdd6c482 | 5159 | int size = mmap_event->event_id.header.size; |
c980d109 | 5160 | int ret; |
0a4a9391 | 5161 | |
67516844 JO |
5162 | if (!perf_event_mmap_match(event, data)) |
5163 | return; | |
5164 | ||
13d7a241 SE |
5165 | if (event->attr.mmap2) { |
5166 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5167 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5168 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5169 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5170 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
13d7a241 SE |
5171 | } |
5172 | ||
c980d109 ACM |
5173 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5174 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5175 | mmap_event->event_id.header.size); |
0a4a9391 | 5176 | if (ret) |
c980d109 | 5177 | goto out; |
0a4a9391 | 5178 | |
cdd6c482 IM |
5179 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5180 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5181 | |
cdd6c482 | 5182 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5183 | |
5184 | if (event->attr.mmap2) { | |
5185 | perf_output_put(&handle, mmap_event->maj); | |
5186 | perf_output_put(&handle, mmap_event->min); | |
5187 | perf_output_put(&handle, mmap_event->ino); | |
5188 | perf_output_put(&handle, mmap_event->ino_generation); | |
5189 | } | |
5190 | ||
76369139 | 5191 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5192 | mmap_event->file_size); |
c980d109 ACM |
5193 | |
5194 | perf_event__output_id_sample(event, &handle, &sample); | |
5195 | ||
78d613eb | 5196 | perf_output_end(&handle); |
c980d109 ACM |
5197 | out: |
5198 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5199 | } |
5200 | ||
cdd6c482 | 5201 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5202 | { |
089dd79d PZ |
5203 | struct vm_area_struct *vma = mmap_event->vma; |
5204 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5205 | int maj = 0, min = 0; |
5206 | u64 ino = 0, gen = 0; | |
0a4a9391 PZ |
5207 | unsigned int size; |
5208 | char tmp[16]; | |
5209 | char *buf = NULL; | |
2c42cfbf | 5210 | char *name; |
413ee3b4 | 5211 | |
0a4a9391 | 5212 | if (file) { |
13d7a241 SE |
5213 | struct inode *inode; |
5214 | dev_t dev; | |
3ea2f2b9 | 5215 | |
2c42cfbf | 5216 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5217 | if (!buf) { |
c7e548b4 ON |
5218 | name = "//enomem"; |
5219 | goto cpy_name; | |
0a4a9391 | 5220 | } |
413ee3b4 | 5221 | /* |
3ea2f2b9 | 5222 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5223 | * need to add enough zero bytes after the string to handle |
5224 | * the 64bit alignment we do later. | |
5225 | */ | |
3ea2f2b9 | 5226 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5227 | if (IS_ERR(name)) { |
c7e548b4 ON |
5228 | name = "//toolong"; |
5229 | goto cpy_name; | |
0a4a9391 | 5230 | } |
13d7a241 SE |
5231 | inode = file_inode(vma->vm_file); |
5232 | dev = inode->i_sb->s_dev; | |
5233 | ino = inode->i_ino; | |
5234 | gen = inode->i_generation; | |
5235 | maj = MAJOR(dev); | |
5236 | min = MINOR(dev); | |
c7e548b4 | 5237 | goto got_name; |
0a4a9391 | 5238 | } else { |
2c42cfbf | 5239 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5240 | if (name) |
5241 | goto cpy_name; | |
089dd79d | 5242 | |
32c5fb7e | 5243 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5244 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5245 | name = "[heap]"; |
5246 | goto cpy_name; | |
32c5fb7e ON |
5247 | } |
5248 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5249 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5250 | name = "[stack]"; |
5251 | goto cpy_name; | |
089dd79d PZ |
5252 | } |
5253 | ||
c7e548b4 ON |
5254 | name = "//anon"; |
5255 | goto cpy_name; | |
0a4a9391 PZ |
5256 | } |
5257 | ||
c7e548b4 ON |
5258 | cpy_name: |
5259 | strlcpy(tmp, name, sizeof(tmp)); | |
5260 | name = tmp; | |
0a4a9391 | 5261 | got_name: |
2c42cfbf PZ |
5262 | /* |
5263 | * Since our buffer works in 8 byte units we need to align our string | |
5264 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5265 | * zero'd out to avoid leaking random bits to userspace. | |
5266 | */ | |
5267 | size = strlen(name)+1; | |
5268 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5269 | name[size++] = '\0'; | |
0a4a9391 PZ |
5270 | |
5271 | mmap_event->file_name = name; | |
5272 | mmap_event->file_size = size; | |
13d7a241 SE |
5273 | mmap_event->maj = maj; |
5274 | mmap_event->min = min; | |
5275 | mmap_event->ino = ino; | |
5276 | mmap_event->ino_generation = gen; | |
0a4a9391 | 5277 | |
2fe85427 SE |
5278 | if (!(vma->vm_flags & VM_EXEC)) |
5279 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5280 | ||
cdd6c482 | 5281 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5282 | |
67516844 | 5283 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5284 | mmap_event, |
5285 | NULL); | |
665c2142 | 5286 | |
0a4a9391 PZ |
5287 | kfree(buf); |
5288 | } | |
5289 | ||
3af9e859 | 5290 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5291 | { |
9ee318a7 PZ |
5292 | struct perf_mmap_event mmap_event; |
5293 | ||
cdd6c482 | 5294 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5295 | return; |
5296 | ||
5297 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5298 | .vma = vma, |
573402db PZ |
5299 | /* .file_name */ |
5300 | /* .file_size */ | |
cdd6c482 | 5301 | .event_id = { |
573402db | 5302 | .header = { |
cdd6c482 | 5303 | .type = PERF_RECORD_MMAP, |
39447b38 | 5304 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5305 | /* .size */ |
5306 | }, | |
5307 | /* .pid */ | |
5308 | /* .tid */ | |
089dd79d PZ |
5309 | .start = vma->vm_start, |
5310 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5311 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5312 | }, |
13d7a241 SE |
5313 | /* .maj (attr_mmap2 only) */ |
5314 | /* .min (attr_mmap2 only) */ | |
5315 | /* .ino (attr_mmap2 only) */ | |
5316 | /* .ino_generation (attr_mmap2 only) */ | |
0a4a9391 PZ |
5317 | }; |
5318 | ||
cdd6c482 | 5319 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5320 | } |
5321 | ||
a78ac325 PZ |
5322 | /* |
5323 | * IRQ throttle logging | |
5324 | */ | |
5325 | ||
cdd6c482 | 5326 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5327 | { |
5328 | struct perf_output_handle handle; | |
c980d109 | 5329 | struct perf_sample_data sample; |
a78ac325 PZ |
5330 | int ret; |
5331 | ||
5332 | struct { | |
5333 | struct perf_event_header header; | |
5334 | u64 time; | |
cca3f454 | 5335 | u64 id; |
7f453c24 | 5336 | u64 stream_id; |
a78ac325 PZ |
5337 | } throttle_event = { |
5338 | .header = { | |
cdd6c482 | 5339 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5340 | .misc = 0, |
5341 | .size = sizeof(throttle_event), | |
5342 | }, | |
def0a9b2 | 5343 | .time = perf_clock(), |
cdd6c482 IM |
5344 | .id = primary_event_id(event), |
5345 | .stream_id = event->id, | |
a78ac325 PZ |
5346 | }; |
5347 | ||
966ee4d6 | 5348 | if (enable) |
cdd6c482 | 5349 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5350 | |
c980d109 ACM |
5351 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5352 | ||
5353 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5354 | throttle_event.header.size); |
a78ac325 PZ |
5355 | if (ret) |
5356 | return; | |
5357 | ||
5358 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5359 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5360 | perf_output_end(&handle); |
5361 | } | |
5362 | ||
f6c7d5fe | 5363 | /* |
cdd6c482 | 5364 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5365 | */ |
5366 | ||
a8b0ca17 | 5367 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5368 | int throttle, struct perf_sample_data *data, |
5369 | struct pt_regs *regs) | |
f6c7d5fe | 5370 | { |
cdd6c482 IM |
5371 | int events = atomic_read(&event->event_limit); |
5372 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5373 | u64 seq; |
79f14641 PZ |
5374 | int ret = 0; |
5375 | ||
96398826 PZ |
5376 | /* |
5377 | * Non-sampling counters might still use the PMI to fold short | |
5378 | * hardware counters, ignore those. | |
5379 | */ | |
5380 | if (unlikely(!is_sampling_event(event))) | |
5381 | return 0; | |
5382 | ||
e050e3f0 SE |
5383 | seq = __this_cpu_read(perf_throttled_seq); |
5384 | if (seq != hwc->interrupts_seq) { | |
5385 | hwc->interrupts_seq = seq; | |
5386 | hwc->interrupts = 1; | |
5387 | } else { | |
5388 | hwc->interrupts++; | |
5389 | if (unlikely(throttle | |
5390 | && hwc->interrupts >= max_samples_per_tick)) { | |
5391 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5392 | hwc->interrupts = MAX_INTERRUPTS; |
5393 | perf_log_throttle(event, 0); | |
d84153d6 | 5394 | tick_nohz_full_kick(); |
a78ac325 PZ |
5395 | ret = 1; |
5396 | } | |
e050e3f0 | 5397 | } |
60db5e09 | 5398 | |
cdd6c482 | 5399 | if (event->attr.freq) { |
def0a9b2 | 5400 | u64 now = perf_clock(); |
abd50713 | 5401 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5402 | |
abd50713 | 5403 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5404 | |
abd50713 | 5405 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5406 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5407 | } |
5408 | ||
2023b359 PZ |
5409 | /* |
5410 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5411 | * events |
2023b359 PZ |
5412 | */ |
5413 | ||
cdd6c482 IM |
5414 | event->pending_kill = POLL_IN; |
5415 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5416 | ret = 1; |
cdd6c482 | 5417 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5418 | event->pending_disable = 1; |
5419 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5420 | } |
5421 | ||
453f19ee | 5422 | if (event->overflow_handler) |
a8b0ca17 | 5423 | event->overflow_handler(event, data, regs); |
453f19ee | 5424 | else |
a8b0ca17 | 5425 | perf_event_output(event, data, regs); |
453f19ee | 5426 | |
f506b3dc | 5427 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5428 | event->pending_wakeup = 1; |
5429 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5430 | } |
5431 | ||
79f14641 | 5432 | return ret; |
f6c7d5fe PZ |
5433 | } |
5434 | ||
a8b0ca17 | 5435 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5436 | struct perf_sample_data *data, |
5437 | struct pt_regs *regs) | |
850bc73f | 5438 | { |
a8b0ca17 | 5439 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5440 | } |
5441 | ||
15dbf27c | 5442 | /* |
cdd6c482 | 5443 | * Generic software event infrastructure |
15dbf27c PZ |
5444 | */ |
5445 | ||
b28ab83c PZ |
5446 | struct swevent_htable { |
5447 | struct swevent_hlist *swevent_hlist; | |
5448 | struct mutex hlist_mutex; | |
5449 | int hlist_refcount; | |
5450 | ||
5451 | /* Recursion avoidance in each contexts */ | |
5452 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
5453 | |
5454 | /* Keeps track of cpu being initialized/exited */ | |
5455 | bool online; | |
b28ab83c PZ |
5456 | }; |
5457 | ||
5458 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5459 | ||
7b4b6658 | 5460 | /* |
cdd6c482 IM |
5461 | * We directly increment event->count and keep a second value in |
5462 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5463 | * is kept in the range [-sample_period, 0] so that we can use the |
5464 | * sign as trigger. | |
5465 | */ | |
5466 | ||
ab573844 | 5467 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5468 | { |
cdd6c482 | 5469 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5470 | u64 period = hwc->last_period; |
5471 | u64 nr, offset; | |
5472 | s64 old, val; | |
5473 | ||
5474 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5475 | |
5476 | again: | |
e7850595 | 5477 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5478 | if (val < 0) |
5479 | return 0; | |
15dbf27c | 5480 | |
7b4b6658 PZ |
5481 | nr = div64_u64(period + val, period); |
5482 | offset = nr * period; | |
5483 | val -= offset; | |
e7850595 | 5484 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5485 | goto again; |
15dbf27c | 5486 | |
7b4b6658 | 5487 | return nr; |
15dbf27c PZ |
5488 | } |
5489 | ||
0cff784a | 5490 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5491 | struct perf_sample_data *data, |
5622f295 | 5492 | struct pt_regs *regs) |
15dbf27c | 5493 | { |
cdd6c482 | 5494 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5495 | int throttle = 0; |
15dbf27c | 5496 | |
0cff784a PZ |
5497 | if (!overflow) |
5498 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5499 | |
7b4b6658 PZ |
5500 | if (hwc->interrupts == MAX_INTERRUPTS) |
5501 | return; | |
15dbf27c | 5502 | |
7b4b6658 | 5503 | for (; overflow; overflow--) { |
a8b0ca17 | 5504 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5505 | data, regs)) { |
7b4b6658 PZ |
5506 | /* |
5507 | * We inhibit the overflow from happening when | |
5508 | * hwc->interrupts == MAX_INTERRUPTS. | |
5509 | */ | |
5510 | break; | |
5511 | } | |
cf450a73 | 5512 | throttle = 1; |
7b4b6658 | 5513 | } |
15dbf27c PZ |
5514 | } |
5515 | ||
a4eaf7f1 | 5516 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5517 | struct perf_sample_data *data, |
5622f295 | 5518 | struct pt_regs *regs) |
7b4b6658 | 5519 | { |
cdd6c482 | 5520 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5521 | |
e7850595 | 5522 | local64_add(nr, &event->count); |
d6d020e9 | 5523 | |
0cff784a PZ |
5524 | if (!regs) |
5525 | return; | |
5526 | ||
6c7e550f | 5527 | if (!is_sampling_event(event)) |
7b4b6658 | 5528 | return; |
d6d020e9 | 5529 | |
5d81e5cf AV |
5530 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5531 | data->period = nr; | |
5532 | return perf_swevent_overflow(event, 1, data, regs); | |
5533 | } else | |
5534 | data->period = event->hw.last_period; | |
5535 | ||
0cff784a | 5536 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5537 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5538 | |
e7850595 | 5539 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5540 | return; |
df1a132b | 5541 | |
a8b0ca17 | 5542 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5543 | } |
5544 | ||
f5ffe02e FW |
5545 | static int perf_exclude_event(struct perf_event *event, |
5546 | struct pt_regs *regs) | |
5547 | { | |
a4eaf7f1 | 5548 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5549 | return 1; |
a4eaf7f1 | 5550 | |
f5ffe02e FW |
5551 | if (regs) { |
5552 | if (event->attr.exclude_user && user_mode(regs)) | |
5553 | return 1; | |
5554 | ||
5555 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5556 | return 1; | |
5557 | } | |
5558 | ||
5559 | return 0; | |
5560 | } | |
5561 | ||
cdd6c482 | 5562 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5563 | enum perf_type_id type, |
6fb2915d LZ |
5564 | u32 event_id, |
5565 | struct perf_sample_data *data, | |
5566 | struct pt_regs *regs) | |
15dbf27c | 5567 | { |
cdd6c482 | 5568 | if (event->attr.type != type) |
a21ca2ca | 5569 | return 0; |
f5ffe02e | 5570 | |
cdd6c482 | 5571 | if (event->attr.config != event_id) |
15dbf27c PZ |
5572 | return 0; |
5573 | ||
f5ffe02e FW |
5574 | if (perf_exclude_event(event, regs)) |
5575 | return 0; | |
15dbf27c PZ |
5576 | |
5577 | return 1; | |
5578 | } | |
5579 | ||
76e1d904 FW |
5580 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5581 | { | |
5582 | u64 val = event_id | (type << 32); | |
5583 | ||
5584 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5585 | } | |
5586 | ||
49f135ed FW |
5587 | static inline struct hlist_head * |
5588 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5589 | { |
49f135ed FW |
5590 | u64 hash = swevent_hash(type, event_id); |
5591 | ||
5592 | return &hlist->heads[hash]; | |
5593 | } | |
76e1d904 | 5594 | |
49f135ed FW |
5595 | /* For the read side: events when they trigger */ |
5596 | static inline struct hlist_head * | |
b28ab83c | 5597 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5598 | { |
5599 | struct swevent_hlist *hlist; | |
76e1d904 | 5600 | |
b28ab83c | 5601 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5602 | if (!hlist) |
5603 | return NULL; | |
5604 | ||
49f135ed FW |
5605 | return __find_swevent_head(hlist, type, event_id); |
5606 | } | |
5607 | ||
5608 | /* For the event head insertion and removal in the hlist */ | |
5609 | static inline struct hlist_head * | |
b28ab83c | 5610 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5611 | { |
5612 | struct swevent_hlist *hlist; | |
5613 | u32 event_id = event->attr.config; | |
5614 | u64 type = event->attr.type; | |
5615 | ||
5616 | /* | |
5617 | * Event scheduling is always serialized against hlist allocation | |
5618 | * and release. Which makes the protected version suitable here. | |
5619 | * The context lock guarantees that. | |
5620 | */ | |
b28ab83c | 5621 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5622 | lockdep_is_held(&event->ctx->lock)); |
5623 | if (!hlist) | |
5624 | return NULL; | |
5625 | ||
5626 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5627 | } |
5628 | ||
5629 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5630 | u64 nr, |
76e1d904 FW |
5631 | struct perf_sample_data *data, |
5632 | struct pt_regs *regs) | |
15dbf27c | 5633 | { |
b28ab83c | 5634 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5635 | struct perf_event *event; |
76e1d904 | 5636 | struct hlist_head *head; |
15dbf27c | 5637 | |
76e1d904 | 5638 | rcu_read_lock(); |
b28ab83c | 5639 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5640 | if (!head) |
5641 | goto end; | |
5642 | ||
b67bfe0d | 5643 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5644 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5645 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5646 | } |
76e1d904 FW |
5647 | end: |
5648 | rcu_read_unlock(); | |
15dbf27c PZ |
5649 | } |
5650 | ||
4ed7c92d | 5651 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5652 | { |
b28ab83c | 5653 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5654 | |
b28ab83c | 5655 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5656 | } |
645e8cc0 | 5657 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5658 | |
fa9f90be | 5659 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5660 | { |
b28ab83c | 5661 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5662 | |
b28ab83c | 5663 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5664 | } |
15dbf27c | 5665 | |
a8b0ca17 | 5666 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5667 | { |
a4234bfc | 5668 | struct perf_sample_data data; |
4ed7c92d PZ |
5669 | int rctx; |
5670 | ||
1c024eca | 5671 | preempt_disable_notrace(); |
4ed7c92d PZ |
5672 | rctx = perf_swevent_get_recursion_context(); |
5673 | if (rctx < 0) | |
5674 | return; | |
a4234bfc | 5675 | |
fd0d000b | 5676 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5677 | |
a8b0ca17 | 5678 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5679 | |
5680 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5681 | preempt_enable_notrace(); |
b8e83514 PZ |
5682 | } |
5683 | ||
cdd6c482 | 5684 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5685 | { |
15dbf27c PZ |
5686 | } |
5687 | ||
a4eaf7f1 | 5688 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5689 | { |
b28ab83c | 5690 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5691 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5692 | struct hlist_head *head; |
5693 | ||
6c7e550f | 5694 | if (is_sampling_event(event)) { |
7b4b6658 | 5695 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5696 | perf_swevent_set_period(event); |
7b4b6658 | 5697 | } |
76e1d904 | 5698 | |
a4eaf7f1 PZ |
5699 | hwc->state = !(flags & PERF_EF_START); |
5700 | ||
b28ab83c | 5701 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
5702 | if (!head) { |
5703 | /* | |
5704 | * We can race with cpu hotplug code. Do not | |
5705 | * WARN if the cpu just got unplugged. | |
5706 | */ | |
5707 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 5708 | return -EINVAL; |
39af6b16 | 5709 | } |
76e1d904 FW |
5710 | |
5711 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5712 | ||
15dbf27c PZ |
5713 | return 0; |
5714 | } | |
5715 | ||
a4eaf7f1 | 5716 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5717 | { |
76e1d904 | 5718 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5719 | } |
5720 | ||
a4eaf7f1 | 5721 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5722 | { |
a4eaf7f1 | 5723 | event->hw.state = 0; |
d6d020e9 | 5724 | } |
aa9c4c0f | 5725 | |
a4eaf7f1 | 5726 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5727 | { |
a4eaf7f1 | 5728 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5729 | } |
5730 | ||
49f135ed FW |
5731 | /* Deref the hlist from the update side */ |
5732 | static inline struct swevent_hlist * | |
b28ab83c | 5733 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5734 | { |
b28ab83c PZ |
5735 | return rcu_dereference_protected(swhash->swevent_hlist, |
5736 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5737 | } |
5738 | ||
b28ab83c | 5739 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5740 | { |
b28ab83c | 5741 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5742 | |
49f135ed | 5743 | if (!hlist) |
76e1d904 FW |
5744 | return; |
5745 | ||
b28ab83c | 5746 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5747 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5748 | } |
5749 | ||
5750 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5751 | { | |
b28ab83c | 5752 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5753 | |
b28ab83c | 5754 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5755 | |
b28ab83c PZ |
5756 | if (!--swhash->hlist_refcount) |
5757 | swevent_hlist_release(swhash); | |
76e1d904 | 5758 | |
b28ab83c | 5759 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5760 | } |
5761 | ||
5762 | static void swevent_hlist_put(struct perf_event *event) | |
5763 | { | |
5764 | int cpu; | |
5765 | ||
76e1d904 FW |
5766 | for_each_possible_cpu(cpu) |
5767 | swevent_hlist_put_cpu(event, cpu); | |
5768 | } | |
5769 | ||
5770 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5771 | { | |
b28ab83c | 5772 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5773 | int err = 0; |
5774 | ||
b28ab83c | 5775 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5776 | |
b28ab83c | 5777 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5778 | struct swevent_hlist *hlist; |
5779 | ||
5780 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5781 | if (!hlist) { | |
5782 | err = -ENOMEM; | |
5783 | goto exit; | |
5784 | } | |
b28ab83c | 5785 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5786 | } |
b28ab83c | 5787 | swhash->hlist_refcount++; |
9ed6060d | 5788 | exit: |
b28ab83c | 5789 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5790 | |
5791 | return err; | |
5792 | } | |
5793 | ||
5794 | static int swevent_hlist_get(struct perf_event *event) | |
5795 | { | |
5796 | int err; | |
5797 | int cpu, failed_cpu; | |
5798 | ||
76e1d904 FW |
5799 | get_online_cpus(); |
5800 | for_each_possible_cpu(cpu) { | |
5801 | err = swevent_hlist_get_cpu(event, cpu); | |
5802 | if (err) { | |
5803 | failed_cpu = cpu; | |
5804 | goto fail; | |
5805 | } | |
5806 | } | |
5807 | put_online_cpus(); | |
5808 | ||
5809 | return 0; | |
9ed6060d | 5810 | fail: |
76e1d904 FW |
5811 | for_each_possible_cpu(cpu) { |
5812 | if (cpu == failed_cpu) | |
5813 | break; | |
5814 | swevent_hlist_put_cpu(event, cpu); | |
5815 | } | |
5816 | ||
5817 | put_online_cpus(); | |
5818 | return err; | |
5819 | } | |
5820 | ||
c5905afb | 5821 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5822 | |
b0a873eb PZ |
5823 | static void sw_perf_event_destroy(struct perf_event *event) |
5824 | { | |
5825 | u64 event_id = event->attr.config; | |
95476b64 | 5826 | |
b0a873eb PZ |
5827 | WARN_ON(event->parent); |
5828 | ||
c5905afb | 5829 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5830 | swevent_hlist_put(event); |
5831 | } | |
5832 | ||
5833 | static int perf_swevent_init(struct perf_event *event) | |
5834 | { | |
8176cced | 5835 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5836 | |
5837 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5838 | return -ENOENT; | |
5839 | ||
2481c5fa SE |
5840 | /* |
5841 | * no branch sampling for software events | |
5842 | */ | |
5843 | if (has_branch_stack(event)) | |
5844 | return -EOPNOTSUPP; | |
5845 | ||
b0a873eb PZ |
5846 | switch (event_id) { |
5847 | case PERF_COUNT_SW_CPU_CLOCK: | |
5848 | case PERF_COUNT_SW_TASK_CLOCK: | |
5849 | return -ENOENT; | |
5850 | ||
5851 | default: | |
5852 | break; | |
5853 | } | |
5854 | ||
ce677831 | 5855 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5856 | return -ENOENT; |
5857 | ||
5858 | if (!event->parent) { | |
5859 | int err; | |
5860 | ||
5861 | err = swevent_hlist_get(event); | |
5862 | if (err) | |
5863 | return err; | |
5864 | ||
c5905afb | 5865 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5866 | event->destroy = sw_perf_event_destroy; |
5867 | } | |
5868 | ||
5869 | return 0; | |
5870 | } | |
5871 | ||
35edc2a5 PZ |
5872 | static int perf_swevent_event_idx(struct perf_event *event) |
5873 | { | |
5874 | return 0; | |
5875 | } | |
5876 | ||
b0a873eb | 5877 | static struct pmu perf_swevent = { |
89a1e187 | 5878 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5879 | |
b0a873eb | 5880 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5881 | .add = perf_swevent_add, |
5882 | .del = perf_swevent_del, | |
5883 | .start = perf_swevent_start, | |
5884 | .stop = perf_swevent_stop, | |
1c024eca | 5885 | .read = perf_swevent_read, |
35edc2a5 PZ |
5886 | |
5887 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5888 | }; |
5889 | ||
b0a873eb PZ |
5890 | #ifdef CONFIG_EVENT_TRACING |
5891 | ||
1c024eca PZ |
5892 | static int perf_tp_filter_match(struct perf_event *event, |
5893 | struct perf_sample_data *data) | |
5894 | { | |
5895 | void *record = data->raw->data; | |
5896 | ||
5897 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5898 | return 1; | |
5899 | return 0; | |
5900 | } | |
5901 | ||
5902 | static int perf_tp_event_match(struct perf_event *event, | |
5903 | struct perf_sample_data *data, | |
5904 | struct pt_regs *regs) | |
5905 | { | |
a0f7d0f7 FW |
5906 | if (event->hw.state & PERF_HES_STOPPED) |
5907 | return 0; | |
580d607c PZ |
5908 | /* |
5909 | * All tracepoints are from kernel-space. | |
5910 | */ | |
5911 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5912 | return 0; |
5913 | ||
5914 | if (!perf_tp_filter_match(event, data)) | |
5915 | return 0; | |
5916 | ||
5917 | return 1; | |
5918 | } | |
5919 | ||
5920 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5921 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5922 | struct task_struct *task) | |
95476b64 FW |
5923 | { |
5924 | struct perf_sample_data data; | |
1c024eca | 5925 | struct perf_event *event; |
1c024eca | 5926 | |
95476b64 FW |
5927 | struct perf_raw_record raw = { |
5928 | .size = entry_size, | |
5929 | .data = record, | |
5930 | }; | |
5931 | ||
fd0d000b | 5932 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5933 | data.raw = &raw; |
5934 | ||
b67bfe0d | 5935 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5936 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5937 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5938 | } |
ecc55f84 | 5939 | |
e6dab5ff AV |
5940 | /* |
5941 | * If we got specified a target task, also iterate its context and | |
5942 | * deliver this event there too. | |
5943 | */ | |
5944 | if (task && task != current) { | |
5945 | struct perf_event_context *ctx; | |
5946 | struct trace_entry *entry = record; | |
5947 | ||
5948 | rcu_read_lock(); | |
5949 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5950 | if (!ctx) | |
5951 | goto unlock; | |
5952 | ||
5953 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5954 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5955 | continue; | |
5956 | if (event->attr.config != entry->type) | |
5957 | continue; | |
5958 | if (perf_tp_event_match(event, &data, regs)) | |
5959 | perf_swevent_event(event, count, &data, regs); | |
5960 | } | |
5961 | unlock: | |
5962 | rcu_read_unlock(); | |
5963 | } | |
5964 | ||
ecc55f84 | 5965 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5966 | } |
5967 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5968 | ||
cdd6c482 | 5969 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5970 | { |
1c024eca | 5971 | perf_trace_destroy(event); |
e077df4f PZ |
5972 | } |
5973 | ||
b0a873eb | 5974 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5975 | { |
76e1d904 FW |
5976 | int err; |
5977 | ||
b0a873eb PZ |
5978 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5979 | return -ENOENT; | |
5980 | ||
2481c5fa SE |
5981 | /* |
5982 | * no branch sampling for tracepoint events | |
5983 | */ | |
5984 | if (has_branch_stack(event)) | |
5985 | return -EOPNOTSUPP; | |
5986 | ||
1c024eca PZ |
5987 | err = perf_trace_init(event); |
5988 | if (err) | |
b0a873eb | 5989 | return err; |
e077df4f | 5990 | |
cdd6c482 | 5991 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5992 | |
b0a873eb PZ |
5993 | return 0; |
5994 | } | |
5995 | ||
5996 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5997 | .task_ctx_nr = perf_sw_context, |
5998 | ||
b0a873eb | 5999 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6000 | .add = perf_trace_add, |
6001 | .del = perf_trace_del, | |
6002 | .start = perf_swevent_start, | |
6003 | .stop = perf_swevent_stop, | |
b0a873eb | 6004 | .read = perf_swevent_read, |
35edc2a5 PZ |
6005 | |
6006 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6007 | }; |
6008 | ||
6009 | static inline void perf_tp_register(void) | |
6010 | { | |
2e80a82a | 6011 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6012 | } |
6fb2915d LZ |
6013 | |
6014 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6015 | { | |
6016 | char *filter_str; | |
6017 | int ret; | |
6018 | ||
6019 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6020 | return -EINVAL; | |
6021 | ||
6022 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6023 | if (IS_ERR(filter_str)) | |
6024 | return PTR_ERR(filter_str); | |
6025 | ||
6026 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6027 | ||
6028 | kfree(filter_str); | |
6029 | return ret; | |
6030 | } | |
6031 | ||
6032 | static void perf_event_free_filter(struct perf_event *event) | |
6033 | { | |
6034 | ftrace_profile_free_filter(event); | |
6035 | } | |
6036 | ||
e077df4f | 6037 | #else |
6fb2915d | 6038 | |
b0a873eb | 6039 | static inline void perf_tp_register(void) |
e077df4f | 6040 | { |
e077df4f | 6041 | } |
6fb2915d LZ |
6042 | |
6043 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6044 | { | |
6045 | return -ENOENT; | |
6046 | } | |
6047 | ||
6048 | static void perf_event_free_filter(struct perf_event *event) | |
6049 | { | |
6050 | } | |
6051 | ||
07b139c8 | 6052 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6053 | |
24f1e32c | 6054 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6055 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6056 | { |
f5ffe02e FW |
6057 | struct perf_sample_data sample; |
6058 | struct pt_regs *regs = data; | |
6059 | ||
fd0d000b | 6060 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6061 | |
a4eaf7f1 | 6062 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6063 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6064 | } |
6065 | #endif | |
6066 | ||
b0a873eb PZ |
6067 | /* |
6068 | * hrtimer based swevent callback | |
6069 | */ | |
f29ac756 | 6070 | |
b0a873eb | 6071 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6072 | { |
b0a873eb PZ |
6073 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6074 | struct perf_sample_data data; | |
6075 | struct pt_regs *regs; | |
6076 | struct perf_event *event; | |
6077 | u64 period; | |
f29ac756 | 6078 | |
b0a873eb | 6079 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6080 | |
6081 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6082 | return HRTIMER_NORESTART; | |
6083 | ||
b0a873eb | 6084 | event->pmu->read(event); |
f344011c | 6085 | |
fd0d000b | 6086 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6087 | regs = get_irq_regs(); |
6088 | ||
6089 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6090 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6091 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6092 | ret = HRTIMER_NORESTART; |
6093 | } | |
24f1e32c | 6094 | |
b0a873eb PZ |
6095 | period = max_t(u64, 10000, event->hw.sample_period); |
6096 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6097 | |
b0a873eb | 6098 | return ret; |
f29ac756 PZ |
6099 | } |
6100 | ||
b0a873eb | 6101 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6102 | { |
b0a873eb | 6103 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6104 | s64 period; |
6105 | ||
6106 | if (!is_sampling_event(event)) | |
6107 | return; | |
f5ffe02e | 6108 | |
5d508e82 FBH |
6109 | period = local64_read(&hwc->period_left); |
6110 | if (period) { | |
6111 | if (period < 0) | |
6112 | period = 10000; | |
fa407f35 | 6113 | |
5d508e82 FBH |
6114 | local64_set(&hwc->period_left, 0); |
6115 | } else { | |
6116 | period = max_t(u64, 10000, hwc->sample_period); | |
6117 | } | |
6118 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6119 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6120 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6121 | } |
b0a873eb PZ |
6122 | |
6123 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6124 | { |
b0a873eb PZ |
6125 | struct hw_perf_event *hwc = &event->hw; |
6126 | ||
6c7e550f | 6127 | if (is_sampling_event(event)) { |
b0a873eb | 6128 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6129 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6130 | |
6131 | hrtimer_cancel(&hwc->hrtimer); | |
6132 | } | |
24f1e32c FW |
6133 | } |
6134 | ||
ba3dd36c PZ |
6135 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6136 | { | |
6137 | struct hw_perf_event *hwc = &event->hw; | |
6138 | ||
6139 | if (!is_sampling_event(event)) | |
6140 | return; | |
6141 | ||
6142 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6143 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6144 | ||
6145 | /* | |
6146 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6147 | * mapping and avoid the whole period adjust feedback stuff. | |
6148 | */ | |
6149 | if (event->attr.freq) { | |
6150 | long freq = event->attr.sample_freq; | |
6151 | ||
6152 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6153 | hwc->sample_period = event->attr.sample_period; | |
6154 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6155 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6156 | event->attr.freq = 0; |
6157 | } | |
6158 | } | |
6159 | ||
b0a873eb PZ |
6160 | /* |
6161 | * Software event: cpu wall time clock | |
6162 | */ | |
6163 | ||
6164 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6165 | { |
b0a873eb PZ |
6166 | s64 prev; |
6167 | u64 now; | |
6168 | ||
a4eaf7f1 | 6169 | now = local_clock(); |
b0a873eb PZ |
6170 | prev = local64_xchg(&event->hw.prev_count, now); |
6171 | local64_add(now - prev, &event->count); | |
24f1e32c | 6172 | } |
24f1e32c | 6173 | |
a4eaf7f1 | 6174 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6175 | { |
a4eaf7f1 | 6176 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6177 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6178 | } |
6179 | ||
a4eaf7f1 | 6180 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6181 | { |
b0a873eb PZ |
6182 | perf_swevent_cancel_hrtimer(event); |
6183 | cpu_clock_event_update(event); | |
6184 | } | |
f29ac756 | 6185 | |
a4eaf7f1 PZ |
6186 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6187 | { | |
6188 | if (flags & PERF_EF_START) | |
6189 | cpu_clock_event_start(event, flags); | |
6190 | ||
6191 | return 0; | |
6192 | } | |
6193 | ||
6194 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6195 | { | |
6196 | cpu_clock_event_stop(event, flags); | |
6197 | } | |
6198 | ||
b0a873eb PZ |
6199 | static void cpu_clock_event_read(struct perf_event *event) |
6200 | { | |
6201 | cpu_clock_event_update(event); | |
6202 | } | |
f344011c | 6203 | |
b0a873eb PZ |
6204 | static int cpu_clock_event_init(struct perf_event *event) |
6205 | { | |
6206 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6207 | return -ENOENT; | |
6208 | ||
6209 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6210 | return -ENOENT; | |
6211 | ||
2481c5fa SE |
6212 | /* |
6213 | * no branch sampling for software events | |
6214 | */ | |
6215 | if (has_branch_stack(event)) | |
6216 | return -EOPNOTSUPP; | |
6217 | ||
ba3dd36c PZ |
6218 | perf_swevent_init_hrtimer(event); |
6219 | ||
b0a873eb | 6220 | return 0; |
f29ac756 PZ |
6221 | } |
6222 | ||
b0a873eb | 6223 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6224 | .task_ctx_nr = perf_sw_context, |
6225 | ||
b0a873eb | 6226 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6227 | .add = cpu_clock_event_add, |
6228 | .del = cpu_clock_event_del, | |
6229 | .start = cpu_clock_event_start, | |
6230 | .stop = cpu_clock_event_stop, | |
b0a873eb | 6231 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
6232 | |
6233 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6234 | }; |
6235 | ||
6236 | /* | |
6237 | * Software event: task time clock | |
6238 | */ | |
6239 | ||
6240 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6241 | { |
b0a873eb PZ |
6242 | u64 prev; |
6243 | s64 delta; | |
5c92d124 | 6244 | |
b0a873eb PZ |
6245 | prev = local64_xchg(&event->hw.prev_count, now); |
6246 | delta = now - prev; | |
6247 | local64_add(delta, &event->count); | |
6248 | } | |
5c92d124 | 6249 | |
a4eaf7f1 | 6250 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6251 | { |
a4eaf7f1 | 6252 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6253 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6254 | } |
6255 | ||
a4eaf7f1 | 6256 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6257 | { |
6258 | perf_swevent_cancel_hrtimer(event); | |
6259 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6260 | } |
6261 | ||
6262 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6263 | { | |
6264 | if (flags & PERF_EF_START) | |
6265 | task_clock_event_start(event, flags); | |
b0a873eb | 6266 | |
a4eaf7f1 PZ |
6267 | return 0; |
6268 | } | |
6269 | ||
6270 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6271 | { | |
6272 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6273 | } |
6274 | ||
6275 | static void task_clock_event_read(struct perf_event *event) | |
6276 | { | |
768a06e2 PZ |
6277 | u64 now = perf_clock(); |
6278 | u64 delta = now - event->ctx->timestamp; | |
6279 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6280 | |
6281 | task_clock_event_update(event, time); | |
6282 | } | |
6283 | ||
6284 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6285 | { |
b0a873eb PZ |
6286 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6287 | return -ENOENT; | |
6288 | ||
6289 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6290 | return -ENOENT; | |
6291 | ||
2481c5fa SE |
6292 | /* |
6293 | * no branch sampling for software events | |
6294 | */ | |
6295 | if (has_branch_stack(event)) | |
6296 | return -EOPNOTSUPP; | |
6297 | ||
ba3dd36c PZ |
6298 | perf_swevent_init_hrtimer(event); |
6299 | ||
b0a873eb | 6300 | return 0; |
6fb2915d LZ |
6301 | } |
6302 | ||
b0a873eb | 6303 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6304 | .task_ctx_nr = perf_sw_context, |
6305 | ||
b0a873eb | 6306 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6307 | .add = task_clock_event_add, |
6308 | .del = task_clock_event_del, | |
6309 | .start = task_clock_event_start, | |
6310 | .stop = task_clock_event_stop, | |
b0a873eb | 6311 | .read = task_clock_event_read, |
35edc2a5 PZ |
6312 | |
6313 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6314 | }; |
6fb2915d | 6315 | |
ad5133b7 | 6316 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6317 | { |
e077df4f | 6318 | } |
6fb2915d | 6319 | |
ad5133b7 | 6320 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6321 | { |
ad5133b7 | 6322 | return 0; |
6fb2915d LZ |
6323 | } |
6324 | ||
ad5133b7 | 6325 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6326 | { |
ad5133b7 | 6327 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6328 | } |
6329 | ||
ad5133b7 PZ |
6330 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6331 | { | |
6332 | perf_pmu_enable(pmu); | |
6333 | return 0; | |
6334 | } | |
e077df4f | 6335 | |
ad5133b7 | 6336 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6337 | { |
ad5133b7 | 6338 | perf_pmu_enable(pmu); |
24f1e32c FW |
6339 | } |
6340 | ||
35edc2a5 PZ |
6341 | static int perf_event_idx_default(struct perf_event *event) |
6342 | { | |
6343 | return event->hw.idx + 1; | |
6344 | } | |
6345 | ||
8dc85d54 PZ |
6346 | /* |
6347 | * Ensures all contexts with the same task_ctx_nr have the same | |
6348 | * pmu_cpu_context too. | |
6349 | */ | |
9e317041 | 6350 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 6351 | { |
8dc85d54 | 6352 | struct pmu *pmu; |
b326e956 | 6353 | |
8dc85d54 PZ |
6354 | if (ctxn < 0) |
6355 | return NULL; | |
24f1e32c | 6356 | |
8dc85d54 PZ |
6357 | list_for_each_entry(pmu, &pmus, entry) { |
6358 | if (pmu->task_ctx_nr == ctxn) | |
6359 | return pmu->pmu_cpu_context; | |
6360 | } | |
24f1e32c | 6361 | |
8dc85d54 | 6362 | return NULL; |
24f1e32c FW |
6363 | } |
6364 | ||
51676957 | 6365 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6366 | { |
51676957 PZ |
6367 | int cpu; |
6368 | ||
6369 | for_each_possible_cpu(cpu) { | |
6370 | struct perf_cpu_context *cpuctx; | |
6371 | ||
6372 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6373 | ||
3f1f3320 PZ |
6374 | if (cpuctx->unique_pmu == old_pmu) |
6375 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6376 | } |
6377 | } | |
6378 | ||
6379 | static void free_pmu_context(struct pmu *pmu) | |
6380 | { | |
6381 | struct pmu *i; | |
f5ffe02e | 6382 | |
8dc85d54 | 6383 | mutex_lock(&pmus_lock); |
0475f9ea | 6384 | /* |
8dc85d54 | 6385 | * Like a real lame refcount. |
0475f9ea | 6386 | */ |
51676957 PZ |
6387 | list_for_each_entry(i, &pmus, entry) { |
6388 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6389 | update_pmu_context(i, pmu); | |
8dc85d54 | 6390 | goto out; |
51676957 | 6391 | } |
8dc85d54 | 6392 | } |
d6d020e9 | 6393 | |
51676957 | 6394 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6395 | out: |
6396 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6397 | } |
2e80a82a | 6398 | static struct idr pmu_idr; |
d6d020e9 | 6399 | |
abe43400 PZ |
6400 | static ssize_t |
6401 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6402 | { | |
6403 | struct pmu *pmu = dev_get_drvdata(dev); | |
6404 | ||
6405 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6406 | } | |
90826ca7 | 6407 | static DEVICE_ATTR_RO(type); |
abe43400 | 6408 | |
62b85639 SE |
6409 | static ssize_t |
6410 | perf_event_mux_interval_ms_show(struct device *dev, | |
6411 | struct device_attribute *attr, | |
6412 | char *page) | |
6413 | { | |
6414 | struct pmu *pmu = dev_get_drvdata(dev); | |
6415 | ||
6416 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6417 | } | |
6418 | ||
6419 | static ssize_t | |
6420 | perf_event_mux_interval_ms_store(struct device *dev, | |
6421 | struct device_attribute *attr, | |
6422 | const char *buf, size_t count) | |
6423 | { | |
6424 | struct pmu *pmu = dev_get_drvdata(dev); | |
6425 | int timer, cpu, ret; | |
6426 | ||
6427 | ret = kstrtoint(buf, 0, &timer); | |
6428 | if (ret) | |
6429 | return ret; | |
6430 | ||
6431 | if (timer < 1) | |
6432 | return -EINVAL; | |
6433 | ||
6434 | /* same value, noting to do */ | |
6435 | if (timer == pmu->hrtimer_interval_ms) | |
6436 | return count; | |
6437 | ||
6438 | pmu->hrtimer_interval_ms = timer; | |
6439 | ||
6440 | /* update all cpuctx for this PMU */ | |
6441 | for_each_possible_cpu(cpu) { | |
6442 | struct perf_cpu_context *cpuctx; | |
6443 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6444 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6445 | ||
6446 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6447 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6448 | } | |
6449 | ||
6450 | return count; | |
6451 | } | |
90826ca7 | 6452 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6453 | |
90826ca7 GKH |
6454 | static struct attribute *pmu_dev_attrs[] = { |
6455 | &dev_attr_type.attr, | |
6456 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6457 | NULL, | |
abe43400 | 6458 | }; |
90826ca7 | 6459 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6460 | |
6461 | static int pmu_bus_running; | |
6462 | static struct bus_type pmu_bus = { | |
6463 | .name = "event_source", | |
90826ca7 | 6464 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6465 | }; |
6466 | ||
6467 | static void pmu_dev_release(struct device *dev) | |
6468 | { | |
6469 | kfree(dev); | |
6470 | } | |
6471 | ||
6472 | static int pmu_dev_alloc(struct pmu *pmu) | |
6473 | { | |
6474 | int ret = -ENOMEM; | |
6475 | ||
6476 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6477 | if (!pmu->dev) | |
6478 | goto out; | |
6479 | ||
0c9d42ed | 6480 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6481 | device_initialize(pmu->dev); |
6482 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6483 | if (ret) | |
6484 | goto free_dev; | |
6485 | ||
6486 | dev_set_drvdata(pmu->dev, pmu); | |
6487 | pmu->dev->bus = &pmu_bus; | |
6488 | pmu->dev->release = pmu_dev_release; | |
6489 | ret = device_add(pmu->dev); | |
6490 | if (ret) | |
6491 | goto free_dev; | |
6492 | ||
6493 | out: | |
6494 | return ret; | |
6495 | ||
6496 | free_dev: | |
6497 | put_device(pmu->dev); | |
6498 | goto out; | |
6499 | } | |
6500 | ||
547e9fd7 | 6501 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6502 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6503 | |
03d8e80b | 6504 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6505 | { |
108b02cf | 6506 | int cpu, ret; |
24f1e32c | 6507 | |
b0a873eb | 6508 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6509 | ret = -ENOMEM; |
6510 | pmu->pmu_disable_count = alloc_percpu(int); | |
6511 | if (!pmu->pmu_disable_count) | |
6512 | goto unlock; | |
f29ac756 | 6513 | |
2e80a82a PZ |
6514 | pmu->type = -1; |
6515 | if (!name) | |
6516 | goto skip_type; | |
6517 | pmu->name = name; | |
6518 | ||
6519 | if (type < 0) { | |
0e9c3be2 TH |
6520 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6521 | if (type < 0) { | |
6522 | ret = type; | |
2e80a82a PZ |
6523 | goto free_pdc; |
6524 | } | |
6525 | } | |
6526 | pmu->type = type; | |
6527 | ||
abe43400 PZ |
6528 | if (pmu_bus_running) { |
6529 | ret = pmu_dev_alloc(pmu); | |
6530 | if (ret) | |
6531 | goto free_idr; | |
6532 | } | |
6533 | ||
2e80a82a | 6534 | skip_type: |
8dc85d54 PZ |
6535 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6536 | if (pmu->pmu_cpu_context) | |
6537 | goto got_cpu_context; | |
f29ac756 | 6538 | |
c4814202 | 6539 | ret = -ENOMEM; |
108b02cf PZ |
6540 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6541 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6542 | goto free_dev; |
f344011c | 6543 | |
108b02cf PZ |
6544 | for_each_possible_cpu(cpu) { |
6545 | struct perf_cpu_context *cpuctx; | |
6546 | ||
6547 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6548 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6549 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6550 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6551 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6552 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6553 | |
6554 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6555 | ||
e9d2b064 | 6556 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6557 | cpuctx->unique_pmu = pmu; |
108b02cf | 6558 | } |
76e1d904 | 6559 | |
8dc85d54 | 6560 | got_cpu_context: |
ad5133b7 PZ |
6561 | if (!pmu->start_txn) { |
6562 | if (pmu->pmu_enable) { | |
6563 | /* | |
6564 | * If we have pmu_enable/pmu_disable calls, install | |
6565 | * transaction stubs that use that to try and batch | |
6566 | * hardware accesses. | |
6567 | */ | |
6568 | pmu->start_txn = perf_pmu_start_txn; | |
6569 | pmu->commit_txn = perf_pmu_commit_txn; | |
6570 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6571 | } else { | |
6572 | pmu->start_txn = perf_pmu_nop_void; | |
6573 | pmu->commit_txn = perf_pmu_nop_int; | |
6574 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6575 | } |
5c92d124 | 6576 | } |
15dbf27c | 6577 | |
ad5133b7 PZ |
6578 | if (!pmu->pmu_enable) { |
6579 | pmu->pmu_enable = perf_pmu_nop_void; | |
6580 | pmu->pmu_disable = perf_pmu_nop_void; | |
6581 | } | |
6582 | ||
35edc2a5 PZ |
6583 | if (!pmu->event_idx) |
6584 | pmu->event_idx = perf_event_idx_default; | |
6585 | ||
b0a873eb | 6586 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6587 | ret = 0; |
6588 | unlock: | |
b0a873eb PZ |
6589 | mutex_unlock(&pmus_lock); |
6590 | ||
33696fc0 | 6591 | return ret; |
108b02cf | 6592 | |
abe43400 PZ |
6593 | free_dev: |
6594 | device_del(pmu->dev); | |
6595 | put_device(pmu->dev); | |
6596 | ||
2e80a82a PZ |
6597 | free_idr: |
6598 | if (pmu->type >= PERF_TYPE_MAX) | |
6599 | idr_remove(&pmu_idr, pmu->type); | |
6600 | ||
108b02cf PZ |
6601 | free_pdc: |
6602 | free_percpu(pmu->pmu_disable_count); | |
6603 | goto unlock; | |
f29ac756 | 6604 | } |
c464c76e | 6605 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 6606 | |
b0a873eb | 6607 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6608 | { |
b0a873eb PZ |
6609 | mutex_lock(&pmus_lock); |
6610 | list_del_rcu(&pmu->entry); | |
6611 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6612 | |
0475f9ea | 6613 | /* |
cde8e884 PZ |
6614 | * We dereference the pmu list under both SRCU and regular RCU, so |
6615 | * synchronize against both of those. | |
0475f9ea | 6616 | */ |
b0a873eb | 6617 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6618 | synchronize_rcu(); |
d6d020e9 | 6619 | |
33696fc0 | 6620 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6621 | if (pmu->type >= PERF_TYPE_MAX) |
6622 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6623 | device_del(pmu->dev); |
6624 | put_device(pmu->dev); | |
51676957 | 6625 | free_pmu_context(pmu); |
b0a873eb | 6626 | } |
c464c76e | 6627 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 6628 | |
b0a873eb PZ |
6629 | struct pmu *perf_init_event(struct perf_event *event) |
6630 | { | |
6631 | struct pmu *pmu = NULL; | |
6632 | int idx; | |
940c5b29 | 6633 | int ret; |
b0a873eb PZ |
6634 | |
6635 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6636 | |
6637 | rcu_read_lock(); | |
6638 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6639 | rcu_read_unlock(); | |
940c5b29 | 6640 | if (pmu) { |
c464c76e YZ |
6641 | if (!try_module_get(pmu->module)) { |
6642 | pmu = ERR_PTR(-ENODEV); | |
6643 | goto unlock; | |
6644 | } | |
7e5b2a01 | 6645 | event->pmu = pmu; |
940c5b29 LM |
6646 | ret = pmu->event_init(event); |
6647 | if (ret) | |
6648 | pmu = ERR_PTR(ret); | |
2e80a82a | 6649 | goto unlock; |
940c5b29 | 6650 | } |
2e80a82a | 6651 | |
b0a873eb | 6652 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
c464c76e YZ |
6653 | if (!try_module_get(pmu->module)) { |
6654 | pmu = ERR_PTR(-ENODEV); | |
6655 | goto unlock; | |
6656 | } | |
7e5b2a01 | 6657 | event->pmu = pmu; |
940c5b29 | 6658 | ret = pmu->event_init(event); |
b0a873eb | 6659 | if (!ret) |
e5f4d339 | 6660 | goto unlock; |
76e1d904 | 6661 | |
b0a873eb PZ |
6662 | if (ret != -ENOENT) { |
6663 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6664 | goto unlock; |
f344011c | 6665 | } |
5c92d124 | 6666 | } |
e5f4d339 PZ |
6667 | pmu = ERR_PTR(-ENOENT); |
6668 | unlock: | |
b0a873eb | 6669 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6670 | |
4aeb0b42 | 6671 | return pmu; |
5c92d124 IM |
6672 | } |
6673 | ||
4beb31f3 FW |
6674 | static void account_event_cpu(struct perf_event *event, int cpu) |
6675 | { | |
6676 | if (event->parent) | |
6677 | return; | |
6678 | ||
6679 | if (has_branch_stack(event)) { | |
6680 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6681 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
6682 | } | |
6683 | if (is_cgroup_event(event)) | |
6684 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
6685 | } | |
6686 | ||
766d6c07 FW |
6687 | static void account_event(struct perf_event *event) |
6688 | { | |
4beb31f3 FW |
6689 | if (event->parent) |
6690 | return; | |
6691 | ||
766d6c07 FW |
6692 | if (event->attach_state & PERF_ATTACH_TASK) |
6693 | static_key_slow_inc(&perf_sched_events.key); | |
6694 | if (event->attr.mmap || event->attr.mmap_data) | |
6695 | atomic_inc(&nr_mmap_events); | |
6696 | if (event->attr.comm) | |
6697 | atomic_inc(&nr_comm_events); | |
6698 | if (event->attr.task) | |
6699 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
6700 | if (event->attr.freq) { |
6701 | if (atomic_inc_return(&nr_freq_events) == 1) | |
6702 | tick_nohz_full_kick_all(); | |
6703 | } | |
4beb31f3 | 6704 | if (has_branch_stack(event)) |
766d6c07 | 6705 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 6706 | if (is_cgroup_event(event)) |
766d6c07 | 6707 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
6708 | |
6709 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
6710 | } |
6711 | ||
0793a61d | 6712 | /* |
cdd6c482 | 6713 | * Allocate and initialize a event structure |
0793a61d | 6714 | */ |
cdd6c482 | 6715 | static struct perf_event * |
c3f00c70 | 6716 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6717 | struct task_struct *task, |
6718 | struct perf_event *group_leader, | |
6719 | struct perf_event *parent_event, | |
4dc0da86 AK |
6720 | perf_overflow_handler_t overflow_handler, |
6721 | void *context) | |
0793a61d | 6722 | { |
51b0fe39 | 6723 | struct pmu *pmu; |
cdd6c482 IM |
6724 | struct perf_event *event; |
6725 | struct hw_perf_event *hwc; | |
90983b16 | 6726 | long err = -EINVAL; |
0793a61d | 6727 | |
66832eb4 ON |
6728 | if ((unsigned)cpu >= nr_cpu_ids) { |
6729 | if (!task || cpu != -1) | |
6730 | return ERR_PTR(-EINVAL); | |
6731 | } | |
6732 | ||
c3f00c70 | 6733 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6734 | if (!event) |
d5d2bc0d | 6735 | return ERR_PTR(-ENOMEM); |
0793a61d | 6736 | |
04289bb9 | 6737 | /* |
cdd6c482 | 6738 | * Single events are their own group leaders, with an |
04289bb9 IM |
6739 | * empty sibling list: |
6740 | */ | |
6741 | if (!group_leader) | |
cdd6c482 | 6742 | group_leader = event; |
04289bb9 | 6743 | |
cdd6c482 IM |
6744 | mutex_init(&event->child_mutex); |
6745 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6746 | |
cdd6c482 IM |
6747 | INIT_LIST_HEAD(&event->group_entry); |
6748 | INIT_LIST_HEAD(&event->event_entry); | |
6749 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 6750 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 6751 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
6752 | INIT_HLIST_NODE(&event->hlist_entry); |
6753 | ||
10c6db11 | 6754 | |
cdd6c482 | 6755 | init_waitqueue_head(&event->waitq); |
e360adbe | 6756 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6757 | |
cdd6c482 | 6758 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6759 | |
a6fa941d | 6760 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6761 | event->cpu = cpu; |
6762 | event->attr = *attr; | |
6763 | event->group_leader = group_leader; | |
6764 | event->pmu = NULL; | |
cdd6c482 | 6765 | event->oncpu = -1; |
a96bbc16 | 6766 | |
cdd6c482 | 6767 | event->parent = parent_event; |
b84fbc9f | 6768 | |
17cf22c3 | 6769 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6770 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6771 | |
cdd6c482 | 6772 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6773 | |
d580ff86 PZ |
6774 | if (task) { |
6775 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6776 | |
6777 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6778 | event->hw.tp_target = task; | |
d580ff86 PZ |
6779 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6780 | /* | |
6781 | * hw_breakpoint is a bit difficult here.. | |
6782 | */ | |
f22c1bb6 | 6783 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6784 | event->hw.bp_target = task; |
6785 | #endif | |
6786 | } | |
6787 | ||
4dc0da86 | 6788 | if (!overflow_handler && parent_event) { |
b326e956 | 6789 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6790 | context = parent_event->overflow_handler_context; |
6791 | } | |
66832eb4 | 6792 | |
b326e956 | 6793 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6794 | event->overflow_handler_context = context; |
97eaf530 | 6795 | |
0231bb53 | 6796 | perf_event__state_init(event); |
a86ed508 | 6797 | |
4aeb0b42 | 6798 | pmu = NULL; |
b8e83514 | 6799 | |
cdd6c482 | 6800 | hwc = &event->hw; |
bd2b5b12 | 6801 | hwc->sample_period = attr->sample_period; |
0d48696f | 6802 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6803 | hwc->sample_period = 1; |
eced1dfc | 6804 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6805 | |
e7850595 | 6806 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6807 | |
2023b359 | 6808 | /* |
cdd6c482 | 6809 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6810 | */ |
3dab77fb | 6811 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 6812 | goto err_ns; |
2023b359 | 6813 | |
b0a873eb | 6814 | pmu = perf_init_event(event); |
4aeb0b42 | 6815 | if (!pmu) |
90983b16 FW |
6816 | goto err_ns; |
6817 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 6818 | err = PTR_ERR(pmu); |
90983b16 | 6819 | goto err_ns; |
621a01ea | 6820 | } |
d5d2bc0d | 6821 | |
cdd6c482 | 6822 | if (!event->parent) { |
927c7a9e FW |
6823 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6824 | err = get_callchain_buffers(); | |
90983b16 FW |
6825 | if (err) |
6826 | goto err_pmu; | |
d010b332 | 6827 | } |
f344011c | 6828 | } |
9ee318a7 | 6829 | |
cdd6c482 | 6830 | return event; |
90983b16 FW |
6831 | |
6832 | err_pmu: | |
6833 | if (event->destroy) | |
6834 | event->destroy(event); | |
c464c76e | 6835 | module_put(pmu->module); |
90983b16 FW |
6836 | err_ns: |
6837 | if (event->ns) | |
6838 | put_pid_ns(event->ns); | |
6839 | kfree(event); | |
6840 | ||
6841 | return ERR_PTR(err); | |
0793a61d TG |
6842 | } |
6843 | ||
cdd6c482 IM |
6844 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6845 | struct perf_event_attr *attr) | |
974802ea | 6846 | { |
974802ea | 6847 | u32 size; |
cdf8073d | 6848 | int ret; |
974802ea PZ |
6849 | |
6850 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6851 | return -EFAULT; | |
6852 | ||
6853 | /* | |
6854 | * zero the full structure, so that a short copy will be nice. | |
6855 | */ | |
6856 | memset(attr, 0, sizeof(*attr)); | |
6857 | ||
6858 | ret = get_user(size, &uattr->size); | |
6859 | if (ret) | |
6860 | return ret; | |
6861 | ||
6862 | if (size > PAGE_SIZE) /* silly large */ | |
6863 | goto err_size; | |
6864 | ||
6865 | if (!size) /* abi compat */ | |
6866 | size = PERF_ATTR_SIZE_VER0; | |
6867 | ||
6868 | if (size < PERF_ATTR_SIZE_VER0) | |
6869 | goto err_size; | |
6870 | ||
6871 | /* | |
6872 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6873 | * ensure all the unknown bits are 0 - i.e. new |
6874 | * user-space does not rely on any kernel feature | |
6875 | * extensions we dont know about yet. | |
974802ea PZ |
6876 | */ |
6877 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6878 | unsigned char __user *addr; |
6879 | unsigned char __user *end; | |
6880 | unsigned char val; | |
974802ea | 6881 | |
cdf8073d IS |
6882 | addr = (void __user *)uattr + sizeof(*attr); |
6883 | end = (void __user *)uattr + size; | |
974802ea | 6884 | |
cdf8073d | 6885 | for (; addr < end; addr++) { |
974802ea PZ |
6886 | ret = get_user(val, addr); |
6887 | if (ret) | |
6888 | return ret; | |
6889 | if (val) | |
6890 | goto err_size; | |
6891 | } | |
b3e62e35 | 6892 | size = sizeof(*attr); |
974802ea PZ |
6893 | } |
6894 | ||
6895 | ret = copy_from_user(attr, uattr, size); | |
6896 | if (ret) | |
6897 | return -EFAULT; | |
6898 | ||
3090ffb5 SE |
6899 | /* disabled for now */ |
6900 | if (attr->mmap2) | |
6901 | return -EINVAL; | |
6902 | ||
cd757645 | 6903 | if (attr->__reserved_1) |
974802ea PZ |
6904 | return -EINVAL; |
6905 | ||
6906 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6907 | return -EINVAL; | |
6908 | ||
6909 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6910 | return -EINVAL; | |
6911 | ||
bce38cd5 SE |
6912 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6913 | u64 mask = attr->branch_sample_type; | |
6914 | ||
6915 | /* only using defined bits */ | |
6916 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6917 | return -EINVAL; | |
6918 | ||
6919 | /* at least one branch bit must be set */ | |
6920 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6921 | return -EINVAL; | |
6922 | ||
bce38cd5 SE |
6923 | /* propagate priv level, when not set for branch */ |
6924 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6925 | ||
6926 | /* exclude_kernel checked on syscall entry */ | |
6927 | if (!attr->exclude_kernel) | |
6928 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6929 | ||
6930 | if (!attr->exclude_user) | |
6931 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6932 | ||
6933 | if (!attr->exclude_hv) | |
6934 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6935 | /* | |
6936 | * adjust user setting (for HW filter setup) | |
6937 | */ | |
6938 | attr->branch_sample_type = mask; | |
6939 | } | |
e712209a SE |
6940 | /* privileged levels capture (kernel, hv): check permissions */ |
6941 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
6942 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6943 | return -EACCES; | |
bce38cd5 | 6944 | } |
4018994f | 6945 | |
c5ebcedb | 6946 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6947 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6948 | if (ret) |
6949 | return ret; | |
6950 | } | |
6951 | ||
6952 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6953 | if (!arch_perf_have_user_stack_dump()) | |
6954 | return -ENOSYS; | |
6955 | ||
6956 | /* | |
6957 | * We have __u32 type for the size, but so far | |
6958 | * we can only use __u16 as maximum due to the | |
6959 | * __u16 sample size limit. | |
6960 | */ | |
6961 | if (attr->sample_stack_user >= USHRT_MAX) | |
6962 | ret = -EINVAL; | |
6963 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6964 | ret = -EINVAL; | |
6965 | } | |
4018994f | 6966 | |
974802ea PZ |
6967 | out: |
6968 | return ret; | |
6969 | ||
6970 | err_size: | |
6971 | put_user(sizeof(*attr), &uattr->size); | |
6972 | ret = -E2BIG; | |
6973 | goto out; | |
6974 | } | |
6975 | ||
ac9721f3 PZ |
6976 | static int |
6977 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6978 | { |
b69cf536 | 6979 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
6980 | int ret = -EINVAL; |
6981 | ||
ac9721f3 | 6982 | if (!output_event) |
a4be7c27 PZ |
6983 | goto set; |
6984 | ||
ac9721f3 PZ |
6985 | /* don't allow circular references */ |
6986 | if (event == output_event) | |
a4be7c27 PZ |
6987 | goto out; |
6988 | ||
0f139300 PZ |
6989 | /* |
6990 | * Don't allow cross-cpu buffers | |
6991 | */ | |
6992 | if (output_event->cpu != event->cpu) | |
6993 | goto out; | |
6994 | ||
6995 | /* | |
76369139 | 6996 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6997 | */ |
6998 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6999 | goto out; | |
7000 | ||
a4be7c27 | 7001 | set: |
cdd6c482 | 7002 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7003 | /* Can't redirect output if we've got an active mmap() */ |
7004 | if (atomic_read(&event->mmap_count)) | |
7005 | goto unlock; | |
a4be7c27 | 7006 | |
ac9721f3 | 7007 | if (output_event) { |
76369139 FW |
7008 | /* get the rb we want to redirect to */ |
7009 | rb = ring_buffer_get(output_event); | |
7010 | if (!rb) | |
ac9721f3 | 7011 | goto unlock; |
a4be7c27 PZ |
7012 | } |
7013 | ||
b69cf536 | 7014 | ring_buffer_attach(event, rb); |
9bb5d40c | 7015 | |
a4be7c27 | 7016 | ret = 0; |
ac9721f3 PZ |
7017 | unlock: |
7018 | mutex_unlock(&event->mmap_mutex); | |
7019 | ||
a4be7c27 | 7020 | out: |
a4be7c27 PZ |
7021 | return ret; |
7022 | } | |
7023 | ||
0793a61d | 7024 | /** |
cdd6c482 | 7025 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7026 | * |
cdd6c482 | 7027 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7028 | * @pid: target pid |
9f66a381 | 7029 | * @cpu: target cpu |
cdd6c482 | 7030 | * @group_fd: group leader event fd |
0793a61d | 7031 | */ |
cdd6c482 IM |
7032 | SYSCALL_DEFINE5(perf_event_open, |
7033 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7034 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7035 | { |
b04243ef PZ |
7036 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7037 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
7038 | struct perf_event_attr attr; |
7039 | struct perf_event_context *ctx; | |
7040 | struct file *event_file = NULL; | |
2903ff01 | 7041 | struct fd group = {NULL, 0}; |
38a81da2 | 7042 | struct task_struct *task = NULL; |
89a1e187 | 7043 | struct pmu *pmu; |
ea635c64 | 7044 | int event_fd; |
b04243ef | 7045 | int move_group = 0; |
dc86cabe | 7046 | int err; |
a21b0b35 | 7047 | int f_flags = O_RDWR; |
0793a61d | 7048 | |
2743a5b0 | 7049 | /* for future expandability... */ |
e5d1367f | 7050 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7051 | return -EINVAL; |
7052 | ||
dc86cabe IM |
7053 | err = perf_copy_attr(attr_uptr, &attr); |
7054 | if (err) | |
7055 | return err; | |
eab656ae | 7056 | |
0764771d PZ |
7057 | if (!attr.exclude_kernel) { |
7058 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7059 | return -EACCES; | |
7060 | } | |
7061 | ||
df58ab24 | 7062 | if (attr.freq) { |
cdd6c482 | 7063 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 7064 | return -EINVAL; |
0819b2e3 PZ |
7065 | } else { |
7066 | if (attr.sample_period & (1ULL << 63)) | |
7067 | return -EINVAL; | |
df58ab24 PZ |
7068 | } |
7069 | ||
e5d1367f SE |
7070 | /* |
7071 | * In cgroup mode, the pid argument is used to pass the fd | |
7072 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7073 | * designates the cpu on which to monitor threads from that | |
7074 | * cgroup. | |
7075 | */ | |
7076 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7077 | return -EINVAL; | |
7078 | ||
a21b0b35 YD |
7079 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7080 | f_flags |= O_CLOEXEC; | |
7081 | ||
7082 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7083 | if (event_fd < 0) |
7084 | return event_fd; | |
7085 | ||
ac9721f3 | 7086 | if (group_fd != -1) { |
2903ff01 AV |
7087 | err = perf_fget_light(group_fd, &group); |
7088 | if (err) | |
d14b12d7 | 7089 | goto err_fd; |
2903ff01 | 7090 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7091 | if (flags & PERF_FLAG_FD_OUTPUT) |
7092 | output_event = group_leader; | |
7093 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7094 | group_leader = NULL; | |
7095 | } | |
7096 | ||
e5d1367f | 7097 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7098 | task = find_lively_task_by_vpid(pid); |
7099 | if (IS_ERR(task)) { | |
7100 | err = PTR_ERR(task); | |
7101 | goto err_group_fd; | |
7102 | } | |
7103 | } | |
7104 | ||
1f4ee503 PZ |
7105 | if (task && group_leader && |
7106 | group_leader->attr.inherit != attr.inherit) { | |
7107 | err = -EINVAL; | |
7108 | goto err_task; | |
7109 | } | |
7110 | ||
fbfc623f YZ |
7111 | get_online_cpus(); |
7112 | ||
4dc0da86 AK |
7113 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
7114 | NULL, NULL); | |
d14b12d7 SE |
7115 | if (IS_ERR(event)) { |
7116 | err = PTR_ERR(event); | |
1f4ee503 | 7117 | goto err_cpus; |
d14b12d7 SE |
7118 | } |
7119 | ||
e5d1367f SE |
7120 | if (flags & PERF_FLAG_PID_CGROUP) { |
7121 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
7122 | if (err) { |
7123 | __free_event(event); | |
1f4ee503 | 7124 | goto err_cpus; |
766d6c07 | 7125 | } |
e5d1367f SE |
7126 | } |
7127 | ||
53b25335 VW |
7128 | if (is_sampling_event(event)) { |
7129 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
7130 | err = -ENOTSUPP; | |
7131 | goto err_alloc; | |
7132 | } | |
7133 | } | |
7134 | ||
766d6c07 FW |
7135 | account_event(event); |
7136 | ||
89a1e187 PZ |
7137 | /* |
7138 | * Special case software events and allow them to be part of | |
7139 | * any hardware group. | |
7140 | */ | |
7141 | pmu = event->pmu; | |
b04243ef PZ |
7142 | |
7143 | if (group_leader && | |
7144 | (is_software_event(event) != is_software_event(group_leader))) { | |
7145 | if (is_software_event(event)) { | |
7146 | /* | |
7147 | * If event and group_leader are not both a software | |
7148 | * event, and event is, then group leader is not. | |
7149 | * | |
7150 | * Allow the addition of software events to !software | |
7151 | * groups, this is safe because software events never | |
7152 | * fail to schedule. | |
7153 | */ | |
7154 | pmu = group_leader->pmu; | |
7155 | } else if (is_software_event(group_leader) && | |
7156 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7157 | /* | |
7158 | * In case the group is a pure software group, and we | |
7159 | * try to add a hardware event, move the whole group to | |
7160 | * the hardware context. | |
7161 | */ | |
7162 | move_group = 1; | |
7163 | } | |
7164 | } | |
89a1e187 PZ |
7165 | |
7166 | /* | |
7167 | * Get the target context (task or percpu): | |
7168 | */ | |
e2d37cd2 | 7169 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
7170 | if (IS_ERR(ctx)) { |
7171 | err = PTR_ERR(ctx); | |
c6be5a5c | 7172 | goto err_alloc; |
89a1e187 PZ |
7173 | } |
7174 | ||
fd1edb3a PZ |
7175 | if (task) { |
7176 | put_task_struct(task); | |
7177 | task = NULL; | |
7178 | } | |
7179 | ||
ccff286d | 7180 | /* |
cdd6c482 | 7181 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7182 | */ |
ac9721f3 | 7183 | if (group_leader) { |
dc86cabe | 7184 | err = -EINVAL; |
04289bb9 | 7185 | |
04289bb9 | 7186 | /* |
ccff286d IM |
7187 | * Do not allow a recursive hierarchy (this new sibling |
7188 | * becoming part of another group-sibling): | |
7189 | */ | |
7190 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7191 | goto err_context; |
ccff286d IM |
7192 | /* |
7193 | * Do not allow to attach to a group in a different | |
7194 | * task or CPU context: | |
04289bb9 | 7195 | */ |
b04243ef PZ |
7196 | if (move_group) { |
7197 | if (group_leader->ctx->type != ctx->type) | |
7198 | goto err_context; | |
7199 | } else { | |
7200 | if (group_leader->ctx != ctx) | |
7201 | goto err_context; | |
7202 | } | |
7203 | ||
3b6f9e5c PM |
7204 | /* |
7205 | * Only a group leader can be exclusive or pinned | |
7206 | */ | |
0d48696f | 7207 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7208 | goto err_context; |
ac9721f3 PZ |
7209 | } |
7210 | ||
7211 | if (output_event) { | |
7212 | err = perf_event_set_output(event, output_event); | |
7213 | if (err) | |
c3f00c70 | 7214 | goto err_context; |
ac9721f3 | 7215 | } |
0793a61d | 7216 | |
a21b0b35 YD |
7217 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7218 | f_flags); | |
ea635c64 AV |
7219 | if (IS_ERR(event_file)) { |
7220 | err = PTR_ERR(event_file); | |
c3f00c70 | 7221 | goto err_context; |
ea635c64 | 7222 | } |
9b51f66d | 7223 | |
b04243ef PZ |
7224 | if (move_group) { |
7225 | struct perf_event_context *gctx = group_leader->ctx; | |
7226 | ||
7227 | mutex_lock(&gctx->mutex); | |
46ce0fe9 | 7228 | perf_remove_from_context(group_leader, false); |
0231bb53 JO |
7229 | |
7230 | /* | |
7231 | * Removing from the context ends up with disabled | |
7232 | * event. What we want here is event in the initial | |
7233 | * startup state, ready to be add into new context. | |
7234 | */ | |
7235 | perf_event__state_init(group_leader); | |
b04243ef PZ |
7236 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7237 | group_entry) { | |
46ce0fe9 | 7238 | perf_remove_from_context(sibling, false); |
0231bb53 | 7239 | perf_event__state_init(sibling); |
b04243ef PZ |
7240 | put_ctx(gctx); |
7241 | } | |
7242 | mutex_unlock(&gctx->mutex); | |
7243 | put_ctx(gctx); | |
ea635c64 | 7244 | } |
9b51f66d | 7245 | |
ad3a37de | 7246 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 7247 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
7248 | |
7249 | if (move_group) { | |
0cda4c02 | 7250 | synchronize_rcu(); |
e2d37cd2 | 7251 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
7252 | get_ctx(ctx); |
7253 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
7254 | group_entry) { | |
e2d37cd2 | 7255 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
7256 | get_ctx(ctx); |
7257 | } | |
7258 | } | |
7259 | ||
e2d37cd2 | 7260 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7261 | perf_unpin_context(ctx); |
d859e29f | 7262 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7263 | |
fbfc623f YZ |
7264 | put_online_cpus(); |
7265 | ||
cdd6c482 | 7266 | event->owner = current; |
8882135b | 7267 | |
cdd6c482 IM |
7268 | mutex_lock(¤t->perf_event_mutex); |
7269 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7270 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7271 | |
c320c7b7 ACM |
7272 | /* |
7273 | * Precalculate sample_data sizes | |
7274 | */ | |
7275 | perf_event__header_size(event); | |
6844c09d | 7276 | perf_event__id_header_size(event); |
c320c7b7 | 7277 | |
8a49542c PZ |
7278 | /* |
7279 | * Drop the reference on the group_event after placing the | |
7280 | * new event on the sibling_list. This ensures destruction | |
7281 | * of the group leader will find the pointer to itself in | |
7282 | * perf_group_detach(). | |
7283 | */ | |
2903ff01 | 7284 | fdput(group); |
ea635c64 AV |
7285 | fd_install(event_fd, event_file); |
7286 | return event_fd; | |
0793a61d | 7287 | |
c3f00c70 | 7288 | err_context: |
fe4b04fa | 7289 | perf_unpin_context(ctx); |
ea635c64 | 7290 | put_ctx(ctx); |
c6be5a5c | 7291 | err_alloc: |
ea635c64 | 7292 | free_event(event); |
1f4ee503 | 7293 | err_cpus: |
fbfc623f | 7294 | put_online_cpus(); |
1f4ee503 | 7295 | err_task: |
e7d0bc04 PZ |
7296 | if (task) |
7297 | put_task_struct(task); | |
89a1e187 | 7298 | err_group_fd: |
2903ff01 | 7299 | fdput(group); |
ea635c64 AV |
7300 | err_fd: |
7301 | put_unused_fd(event_fd); | |
dc86cabe | 7302 | return err; |
0793a61d TG |
7303 | } |
7304 | ||
fb0459d7 AV |
7305 | /** |
7306 | * perf_event_create_kernel_counter | |
7307 | * | |
7308 | * @attr: attributes of the counter to create | |
7309 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7310 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7311 | */ |
7312 | struct perf_event * | |
7313 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7314 | struct task_struct *task, |
4dc0da86 AK |
7315 | perf_overflow_handler_t overflow_handler, |
7316 | void *context) | |
fb0459d7 | 7317 | { |
fb0459d7 | 7318 | struct perf_event_context *ctx; |
c3f00c70 | 7319 | struct perf_event *event; |
fb0459d7 | 7320 | int err; |
d859e29f | 7321 | |
fb0459d7 AV |
7322 | /* |
7323 | * Get the target context (task or percpu): | |
7324 | */ | |
d859e29f | 7325 | |
4dc0da86 AK |
7326 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7327 | overflow_handler, context); | |
c3f00c70 PZ |
7328 | if (IS_ERR(event)) { |
7329 | err = PTR_ERR(event); | |
7330 | goto err; | |
7331 | } | |
d859e29f | 7332 | |
766d6c07 FW |
7333 | account_event(event); |
7334 | ||
38a81da2 | 7335 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7336 | if (IS_ERR(ctx)) { |
7337 | err = PTR_ERR(ctx); | |
c3f00c70 | 7338 | goto err_free; |
d859e29f | 7339 | } |
fb0459d7 | 7340 | |
fb0459d7 AV |
7341 | WARN_ON_ONCE(ctx->parent_ctx); |
7342 | mutex_lock(&ctx->mutex); | |
7343 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7344 | perf_unpin_context(ctx); |
fb0459d7 AV |
7345 | mutex_unlock(&ctx->mutex); |
7346 | ||
fb0459d7 AV |
7347 | return event; |
7348 | ||
c3f00c70 PZ |
7349 | err_free: |
7350 | free_event(event); | |
7351 | err: | |
c6567f64 | 7352 | return ERR_PTR(err); |
9b51f66d | 7353 | } |
fb0459d7 | 7354 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7355 | |
0cda4c02 YZ |
7356 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7357 | { | |
7358 | struct perf_event_context *src_ctx; | |
7359 | struct perf_event_context *dst_ctx; | |
7360 | struct perf_event *event, *tmp; | |
7361 | LIST_HEAD(events); | |
7362 | ||
7363 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7364 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7365 | ||
7366 | mutex_lock(&src_ctx->mutex); | |
7367 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7368 | event_entry) { | |
46ce0fe9 | 7369 | perf_remove_from_context(event, false); |
9a545de0 | 7370 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7371 | put_ctx(src_ctx); |
9886167d | 7372 | list_add(&event->migrate_entry, &events); |
0cda4c02 YZ |
7373 | } |
7374 | mutex_unlock(&src_ctx->mutex); | |
7375 | ||
7376 | synchronize_rcu(); | |
7377 | ||
7378 | mutex_lock(&dst_ctx->mutex); | |
9886167d PZ |
7379 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7380 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7381 | if (event->state >= PERF_EVENT_STATE_OFF) |
7382 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7383 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7384 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7385 | get_ctx(dst_ctx); | |
7386 | } | |
7387 | mutex_unlock(&dst_ctx->mutex); | |
7388 | } | |
7389 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7390 | ||
cdd6c482 | 7391 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7392 | struct task_struct *child) |
d859e29f | 7393 | { |
cdd6c482 | 7394 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7395 | u64 child_val; |
d859e29f | 7396 | |
cdd6c482 IM |
7397 | if (child_event->attr.inherit_stat) |
7398 | perf_event_read_event(child_event, child); | |
38b200d6 | 7399 | |
b5e58793 | 7400 | child_val = perf_event_count(child_event); |
d859e29f PM |
7401 | |
7402 | /* | |
7403 | * Add back the child's count to the parent's count: | |
7404 | */ | |
a6e6dea6 | 7405 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7406 | atomic64_add(child_event->total_time_enabled, |
7407 | &parent_event->child_total_time_enabled); | |
7408 | atomic64_add(child_event->total_time_running, | |
7409 | &parent_event->child_total_time_running); | |
d859e29f PM |
7410 | |
7411 | /* | |
cdd6c482 | 7412 | * Remove this event from the parent's list |
d859e29f | 7413 | */ |
cdd6c482 IM |
7414 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7415 | mutex_lock(&parent_event->child_mutex); | |
7416 | list_del_init(&child_event->child_list); | |
7417 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7418 | |
7419 | /* | |
cdd6c482 | 7420 | * Release the parent event, if this was the last |
d859e29f PM |
7421 | * reference to it. |
7422 | */ | |
a6fa941d | 7423 | put_event(parent_event); |
d859e29f PM |
7424 | } |
7425 | ||
9b51f66d | 7426 | static void |
cdd6c482 IM |
7427 | __perf_event_exit_task(struct perf_event *child_event, |
7428 | struct perf_event_context *child_ctx, | |
38b200d6 | 7429 | struct task_struct *child) |
9b51f66d | 7430 | { |
15a2d4de | 7431 | perf_remove_from_context(child_event, true); |
0cc0c027 | 7432 | |
9b51f66d | 7433 | /* |
38b435b1 | 7434 | * It can happen that the parent exits first, and has events |
9b51f66d | 7435 | * that are still around due to the child reference. These |
38b435b1 | 7436 | * events need to be zapped. |
9b51f66d | 7437 | */ |
38b435b1 | 7438 | if (child_event->parent) { |
cdd6c482 IM |
7439 | sync_child_event(child_event, child); |
7440 | free_event(child_event); | |
4bcf349a | 7441 | } |
9b51f66d IM |
7442 | } |
7443 | ||
8dc85d54 | 7444 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7445 | { |
ebf905fc | 7446 | struct perf_event *child_event, *next; |
cdd6c482 | 7447 | struct perf_event_context *child_ctx; |
a63eaf34 | 7448 | unsigned long flags; |
9b51f66d | 7449 | |
8dc85d54 | 7450 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7451 | perf_event_task(child, NULL, 0); |
9b51f66d | 7452 | return; |
9f498cc5 | 7453 | } |
9b51f66d | 7454 | |
a63eaf34 | 7455 | local_irq_save(flags); |
ad3a37de PM |
7456 | /* |
7457 | * We can't reschedule here because interrupts are disabled, | |
7458 | * and either child is current or it is a task that can't be | |
7459 | * scheduled, so we are now safe from rescheduling changing | |
7460 | * our context. | |
7461 | */ | |
806839b2 | 7462 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7463 | |
7464 | /* | |
7465 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7466 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7467 | * incremented the context's refcount before we do put_ctx below. |
7468 | */ | |
e625cce1 | 7469 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7470 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7471 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
7472 | /* |
7473 | * If this context is a clone; unclone it so it can't get | |
7474 | * swapped to another process while we're removing all | |
cdd6c482 | 7475 | * the events from it. |
71a851b4 PZ |
7476 | */ |
7477 | unclone_ctx(child_ctx); | |
5e942bb3 | 7478 | update_context_time(child_ctx); |
e625cce1 | 7479 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
7480 | |
7481 | /* | |
cdd6c482 IM |
7482 | * Report the task dead after unscheduling the events so that we |
7483 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7484 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7485 | */ |
cdd6c482 | 7486 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7487 | |
66fff224 PZ |
7488 | /* |
7489 | * We can recurse on the same lock type through: | |
7490 | * | |
cdd6c482 IM |
7491 | * __perf_event_exit_task() |
7492 | * sync_child_event() | |
a6fa941d AV |
7493 | * put_event() |
7494 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7495 | * |
7496 | * But since its the parent context it won't be the same instance. | |
7497 | */ | |
a0507c84 | 7498 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7499 | |
ebf905fc | 7500 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 7501 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 7502 | |
a63eaf34 PM |
7503 | mutex_unlock(&child_ctx->mutex); |
7504 | ||
7505 | put_ctx(child_ctx); | |
9b51f66d IM |
7506 | } |
7507 | ||
8dc85d54 PZ |
7508 | /* |
7509 | * When a child task exits, feed back event values to parent events. | |
7510 | */ | |
7511 | void perf_event_exit_task(struct task_struct *child) | |
7512 | { | |
8882135b | 7513 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7514 | int ctxn; |
7515 | ||
8882135b PZ |
7516 | mutex_lock(&child->perf_event_mutex); |
7517 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7518 | owner_entry) { | |
7519 | list_del_init(&event->owner_entry); | |
7520 | ||
7521 | /* | |
7522 | * Ensure the list deletion is visible before we clear | |
7523 | * the owner, closes a race against perf_release() where | |
7524 | * we need to serialize on the owner->perf_event_mutex. | |
7525 | */ | |
7526 | smp_wmb(); | |
7527 | event->owner = NULL; | |
7528 | } | |
7529 | mutex_unlock(&child->perf_event_mutex); | |
7530 | ||
8dc85d54 PZ |
7531 | for_each_task_context_nr(ctxn) |
7532 | perf_event_exit_task_context(child, ctxn); | |
7533 | } | |
7534 | ||
889ff015 FW |
7535 | static void perf_free_event(struct perf_event *event, |
7536 | struct perf_event_context *ctx) | |
7537 | { | |
7538 | struct perf_event *parent = event->parent; | |
7539 | ||
7540 | if (WARN_ON_ONCE(!parent)) | |
7541 | return; | |
7542 | ||
7543 | mutex_lock(&parent->child_mutex); | |
7544 | list_del_init(&event->child_list); | |
7545 | mutex_unlock(&parent->child_mutex); | |
7546 | ||
a6fa941d | 7547 | put_event(parent); |
889ff015 | 7548 | |
8a49542c | 7549 | perf_group_detach(event); |
889ff015 FW |
7550 | list_del_event(event, ctx); |
7551 | free_event(event); | |
7552 | } | |
7553 | ||
bbbee908 PZ |
7554 | /* |
7555 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7556 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7557 | */ |
cdd6c482 | 7558 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7559 | { |
8dc85d54 | 7560 | struct perf_event_context *ctx; |
cdd6c482 | 7561 | struct perf_event *event, *tmp; |
8dc85d54 | 7562 | int ctxn; |
bbbee908 | 7563 | |
8dc85d54 PZ |
7564 | for_each_task_context_nr(ctxn) { |
7565 | ctx = task->perf_event_ctxp[ctxn]; | |
7566 | if (!ctx) | |
7567 | continue; | |
bbbee908 | 7568 | |
8dc85d54 | 7569 | mutex_lock(&ctx->mutex); |
bbbee908 | 7570 | again: |
8dc85d54 PZ |
7571 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7572 | group_entry) | |
7573 | perf_free_event(event, ctx); | |
bbbee908 | 7574 | |
8dc85d54 PZ |
7575 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7576 | group_entry) | |
7577 | perf_free_event(event, ctx); | |
bbbee908 | 7578 | |
8dc85d54 PZ |
7579 | if (!list_empty(&ctx->pinned_groups) || |
7580 | !list_empty(&ctx->flexible_groups)) | |
7581 | goto again; | |
bbbee908 | 7582 | |
8dc85d54 | 7583 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7584 | |
8dc85d54 PZ |
7585 | put_ctx(ctx); |
7586 | } | |
889ff015 FW |
7587 | } |
7588 | ||
4e231c79 PZ |
7589 | void perf_event_delayed_put(struct task_struct *task) |
7590 | { | |
7591 | int ctxn; | |
7592 | ||
7593 | for_each_task_context_nr(ctxn) | |
7594 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7595 | } | |
7596 | ||
97dee4f3 PZ |
7597 | /* |
7598 | * inherit a event from parent task to child task: | |
7599 | */ | |
7600 | static struct perf_event * | |
7601 | inherit_event(struct perf_event *parent_event, | |
7602 | struct task_struct *parent, | |
7603 | struct perf_event_context *parent_ctx, | |
7604 | struct task_struct *child, | |
7605 | struct perf_event *group_leader, | |
7606 | struct perf_event_context *child_ctx) | |
7607 | { | |
7608 | struct perf_event *child_event; | |
cee010ec | 7609 | unsigned long flags; |
97dee4f3 PZ |
7610 | |
7611 | /* | |
7612 | * Instead of creating recursive hierarchies of events, | |
7613 | * we link inherited events back to the original parent, | |
7614 | * which has a filp for sure, which we use as the reference | |
7615 | * count: | |
7616 | */ | |
7617 | if (parent_event->parent) | |
7618 | parent_event = parent_event->parent; | |
7619 | ||
7620 | child_event = perf_event_alloc(&parent_event->attr, | |
7621 | parent_event->cpu, | |
d580ff86 | 7622 | child, |
97dee4f3 | 7623 | group_leader, parent_event, |
4dc0da86 | 7624 | NULL, NULL); |
97dee4f3 PZ |
7625 | if (IS_ERR(child_event)) |
7626 | return child_event; | |
a6fa941d AV |
7627 | |
7628 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7629 | free_event(child_event); | |
7630 | return NULL; | |
7631 | } | |
7632 | ||
97dee4f3 PZ |
7633 | get_ctx(child_ctx); |
7634 | ||
7635 | /* | |
7636 | * Make the child state follow the state of the parent event, | |
7637 | * not its attr.disabled bit. We hold the parent's mutex, | |
7638 | * so we won't race with perf_event_{en, dis}able_family. | |
7639 | */ | |
7640 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7641 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7642 | else | |
7643 | child_event->state = PERF_EVENT_STATE_OFF; | |
7644 | ||
7645 | if (parent_event->attr.freq) { | |
7646 | u64 sample_period = parent_event->hw.sample_period; | |
7647 | struct hw_perf_event *hwc = &child_event->hw; | |
7648 | ||
7649 | hwc->sample_period = sample_period; | |
7650 | hwc->last_period = sample_period; | |
7651 | ||
7652 | local64_set(&hwc->period_left, sample_period); | |
7653 | } | |
7654 | ||
7655 | child_event->ctx = child_ctx; | |
7656 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7657 | child_event->overflow_handler_context |
7658 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7659 | |
614b6780 TG |
7660 | /* |
7661 | * Precalculate sample_data sizes | |
7662 | */ | |
7663 | perf_event__header_size(child_event); | |
6844c09d | 7664 | perf_event__id_header_size(child_event); |
614b6780 | 7665 | |
97dee4f3 PZ |
7666 | /* |
7667 | * Link it up in the child's context: | |
7668 | */ | |
cee010ec | 7669 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7670 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7671 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7672 | |
97dee4f3 PZ |
7673 | /* |
7674 | * Link this into the parent event's child list | |
7675 | */ | |
7676 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7677 | mutex_lock(&parent_event->child_mutex); | |
7678 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7679 | mutex_unlock(&parent_event->child_mutex); | |
7680 | ||
7681 | return child_event; | |
7682 | } | |
7683 | ||
7684 | static int inherit_group(struct perf_event *parent_event, | |
7685 | struct task_struct *parent, | |
7686 | struct perf_event_context *parent_ctx, | |
7687 | struct task_struct *child, | |
7688 | struct perf_event_context *child_ctx) | |
7689 | { | |
7690 | struct perf_event *leader; | |
7691 | struct perf_event *sub; | |
7692 | struct perf_event *child_ctr; | |
7693 | ||
7694 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7695 | child, NULL, child_ctx); | |
7696 | if (IS_ERR(leader)) | |
7697 | return PTR_ERR(leader); | |
7698 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7699 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7700 | child, leader, child_ctx); | |
7701 | if (IS_ERR(child_ctr)) | |
7702 | return PTR_ERR(child_ctr); | |
7703 | } | |
7704 | return 0; | |
889ff015 FW |
7705 | } |
7706 | ||
7707 | static int | |
7708 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7709 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7710 | struct task_struct *child, int ctxn, |
889ff015 FW |
7711 | int *inherited_all) |
7712 | { | |
7713 | int ret; | |
8dc85d54 | 7714 | struct perf_event_context *child_ctx; |
889ff015 FW |
7715 | |
7716 | if (!event->attr.inherit) { | |
7717 | *inherited_all = 0; | |
7718 | return 0; | |
bbbee908 PZ |
7719 | } |
7720 | ||
fe4b04fa | 7721 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7722 | if (!child_ctx) { |
7723 | /* | |
7724 | * This is executed from the parent task context, so | |
7725 | * inherit events that have been marked for cloning. | |
7726 | * First allocate and initialize a context for the | |
7727 | * child. | |
7728 | */ | |
bbbee908 | 7729 | |
734df5ab | 7730 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7731 | if (!child_ctx) |
7732 | return -ENOMEM; | |
bbbee908 | 7733 | |
8dc85d54 | 7734 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7735 | } |
7736 | ||
7737 | ret = inherit_group(event, parent, parent_ctx, | |
7738 | child, child_ctx); | |
7739 | ||
7740 | if (ret) | |
7741 | *inherited_all = 0; | |
7742 | ||
7743 | return ret; | |
bbbee908 PZ |
7744 | } |
7745 | ||
9b51f66d | 7746 | /* |
cdd6c482 | 7747 | * Initialize the perf_event context in task_struct |
9b51f66d | 7748 | */ |
8dc85d54 | 7749 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7750 | { |
889ff015 | 7751 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7752 | struct perf_event_context *cloned_ctx; |
7753 | struct perf_event *event; | |
9b51f66d | 7754 | struct task_struct *parent = current; |
564c2b21 | 7755 | int inherited_all = 1; |
dddd3379 | 7756 | unsigned long flags; |
6ab423e0 | 7757 | int ret = 0; |
9b51f66d | 7758 | |
8dc85d54 | 7759 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7760 | return 0; |
7761 | ||
ad3a37de | 7762 | /* |
25346b93 PM |
7763 | * If the parent's context is a clone, pin it so it won't get |
7764 | * swapped under us. | |
ad3a37de | 7765 | */ |
8dc85d54 | 7766 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
7767 | if (!parent_ctx) |
7768 | return 0; | |
25346b93 | 7769 | |
ad3a37de PM |
7770 | /* |
7771 | * No need to check if parent_ctx != NULL here; since we saw | |
7772 | * it non-NULL earlier, the only reason for it to become NULL | |
7773 | * is if we exit, and since we're currently in the middle of | |
7774 | * a fork we can't be exiting at the same time. | |
7775 | */ | |
ad3a37de | 7776 | |
9b51f66d IM |
7777 | /* |
7778 | * Lock the parent list. No need to lock the child - not PID | |
7779 | * hashed yet and not running, so nobody can access it. | |
7780 | */ | |
d859e29f | 7781 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7782 | |
7783 | /* | |
7784 | * We dont have to disable NMIs - we are only looking at | |
7785 | * the list, not manipulating it: | |
7786 | */ | |
889ff015 | 7787 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7788 | ret = inherit_task_group(event, parent, parent_ctx, |
7789 | child, ctxn, &inherited_all); | |
889ff015 FW |
7790 | if (ret) |
7791 | break; | |
7792 | } | |
b93f7978 | 7793 | |
dddd3379 TG |
7794 | /* |
7795 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7796 | * to allocations, but we need to prevent rotation because | |
7797 | * rotate_ctx() will change the list from interrupt context. | |
7798 | */ | |
7799 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7800 | parent_ctx->rotate_disable = 1; | |
7801 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7802 | ||
889ff015 | 7803 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7804 | ret = inherit_task_group(event, parent, parent_ctx, |
7805 | child, ctxn, &inherited_all); | |
889ff015 | 7806 | if (ret) |
9b51f66d | 7807 | break; |
564c2b21 PM |
7808 | } |
7809 | ||
dddd3379 TG |
7810 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7811 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7812 | |
8dc85d54 | 7813 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7814 | |
05cbaa28 | 7815 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7816 | /* |
7817 | * Mark the child context as a clone of the parent | |
7818 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7819 | * |
7820 | * Note that if the parent is a clone, the holding of | |
7821 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7822 | */ |
c5ed5145 | 7823 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7824 | if (cloned_ctx) { |
7825 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7826 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7827 | } else { |
7828 | child_ctx->parent_ctx = parent_ctx; | |
7829 | child_ctx->parent_gen = parent_ctx->generation; | |
7830 | } | |
7831 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7832 | } |
7833 | ||
c5ed5145 | 7834 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7835 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7836 | |
25346b93 | 7837 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7838 | put_ctx(parent_ctx); |
ad3a37de | 7839 | |
6ab423e0 | 7840 | return ret; |
9b51f66d IM |
7841 | } |
7842 | ||
8dc85d54 PZ |
7843 | /* |
7844 | * Initialize the perf_event context in task_struct | |
7845 | */ | |
7846 | int perf_event_init_task(struct task_struct *child) | |
7847 | { | |
7848 | int ctxn, ret; | |
7849 | ||
8550d7cb ON |
7850 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7851 | mutex_init(&child->perf_event_mutex); | |
7852 | INIT_LIST_HEAD(&child->perf_event_list); | |
7853 | ||
8dc85d54 PZ |
7854 | for_each_task_context_nr(ctxn) { |
7855 | ret = perf_event_init_context(child, ctxn); | |
7856 | if (ret) | |
7857 | return ret; | |
7858 | } | |
7859 | ||
7860 | return 0; | |
7861 | } | |
7862 | ||
220b140b PM |
7863 | static void __init perf_event_init_all_cpus(void) |
7864 | { | |
b28ab83c | 7865 | struct swevent_htable *swhash; |
220b140b | 7866 | int cpu; |
220b140b PM |
7867 | |
7868 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7869 | swhash = &per_cpu(swevent_htable, cpu); |
7870 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7871 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7872 | } |
7873 | } | |
7874 | ||
0db0628d | 7875 | static void perf_event_init_cpu(int cpu) |
0793a61d | 7876 | { |
108b02cf | 7877 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7878 | |
b28ab83c | 7879 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 7880 | swhash->online = true; |
4536e4d1 | 7881 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7882 | struct swevent_hlist *hlist; |
7883 | ||
b28ab83c PZ |
7884 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7885 | WARN_ON(!hlist); | |
7886 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7887 | } |
b28ab83c | 7888 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7889 | } |
7890 | ||
c277443c | 7891 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7892 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7893 | { |
e9d2b064 PZ |
7894 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7895 | ||
7896 | WARN_ON(!irqs_disabled()); | |
7897 | ||
7898 | list_del_init(&cpuctx->rotation_list); | |
7899 | } | |
7900 | ||
108b02cf | 7901 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7902 | { |
46ce0fe9 | 7903 | struct remove_event re = { .detach_group = false }; |
108b02cf | 7904 | struct perf_event_context *ctx = __info; |
0793a61d | 7905 | |
108b02cf | 7906 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7907 | |
e3703f8c | 7908 | rcu_read_lock(); |
46ce0fe9 PZ |
7909 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
7910 | __perf_remove_from_context(&re); | |
e3703f8c | 7911 | rcu_read_unlock(); |
0793a61d | 7912 | } |
108b02cf PZ |
7913 | |
7914 | static void perf_event_exit_cpu_context(int cpu) | |
7915 | { | |
7916 | struct perf_event_context *ctx; | |
7917 | struct pmu *pmu; | |
7918 | int idx; | |
7919 | ||
7920 | idx = srcu_read_lock(&pmus_srcu); | |
7921 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7922 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7923 | |
7924 | mutex_lock(&ctx->mutex); | |
7925 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7926 | mutex_unlock(&ctx->mutex); | |
7927 | } | |
7928 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7929 | } |
7930 | ||
cdd6c482 | 7931 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7932 | { |
b28ab83c | 7933 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7934 | |
e3703f8c PZ |
7935 | perf_event_exit_cpu_context(cpu); |
7936 | ||
b28ab83c | 7937 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 7938 | swhash->online = false; |
b28ab83c PZ |
7939 | swevent_hlist_release(swhash); |
7940 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
7941 | } |
7942 | #else | |
cdd6c482 | 7943 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7944 | #endif |
7945 | ||
c277443c PZ |
7946 | static int |
7947 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7948 | { | |
7949 | int cpu; | |
7950 | ||
7951 | for_each_online_cpu(cpu) | |
7952 | perf_event_exit_cpu(cpu); | |
7953 | ||
7954 | return NOTIFY_OK; | |
7955 | } | |
7956 | ||
7957 | /* | |
7958 | * Run the perf reboot notifier at the very last possible moment so that | |
7959 | * the generic watchdog code runs as long as possible. | |
7960 | */ | |
7961 | static struct notifier_block perf_reboot_notifier = { | |
7962 | .notifier_call = perf_reboot, | |
7963 | .priority = INT_MIN, | |
7964 | }; | |
7965 | ||
0db0628d | 7966 | static int |
0793a61d TG |
7967 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
7968 | { | |
7969 | unsigned int cpu = (long)hcpu; | |
7970 | ||
4536e4d1 | 7971 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7972 | |
7973 | case CPU_UP_PREPARE: | |
5e11637e | 7974 | case CPU_DOWN_FAILED: |
cdd6c482 | 7975 | perf_event_init_cpu(cpu); |
0793a61d TG |
7976 | break; |
7977 | ||
5e11637e | 7978 | case CPU_UP_CANCELED: |
0793a61d | 7979 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7980 | perf_event_exit_cpu(cpu); |
0793a61d | 7981 | break; |
0793a61d TG |
7982 | default: |
7983 | break; | |
7984 | } | |
7985 | ||
7986 | return NOTIFY_OK; | |
7987 | } | |
7988 | ||
cdd6c482 | 7989 | void __init perf_event_init(void) |
0793a61d | 7990 | { |
3c502e7a JW |
7991 | int ret; |
7992 | ||
2e80a82a PZ |
7993 | idr_init(&pmu_idr); |
7994 | ||
220b140b | 7995 | perf_event_init_all_cpus(); |
b0a873eb | 7996 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7997 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7998 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7999 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
8000 | perf_tp_register(); |
8001 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 8002 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
8003 | |
8004 | ret = init_hw_breakpoint(); | |
8005 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
8006 | |
8007 | /* do not patch jump label more than once per second */ | |
8008 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
8009 | |
8010 | /* | |
8011 | * Build time assertion that we keep the data_head at the intended | |
8012 | * location. IOW, validation we got the __reserved[] size right. | |
8013 | */ | |
8014 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
8015 | != 1024); | |
0793a61d | 8016 | } |
abe43400 PZ |
8017 | |
8018 | static int __init perf_event_sysfs_init(void) | |
8019 | { | |
8020 | struct pmu *pmu; | |
8021 | int ret; | |
8022 | ||
8023 | mutex_lock(&pmus_lock); | |
8024 | ||
8025 | ret = bus_register(&pmu_bus); | |
8026 | if (ret) | |
8027 | goto unlock; | |
8028 | ||
8029 | list_for_each_entry(pmu, &pmus, entry) { | |
8030 | if (!pmu->name || pmu->type < 0) | |
8031 | continue; | |
8032 | ||
8033 | ret = pmu_dev_alloc(pmu); | |
8034 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8035 | } | |
8036 | pmu_bus_running = 1; | |
8037 | ret = 0; | |
8038 | ||
8039 | unlock: | |
8040 | mutex_unlock(&pmus_lock); | |
8041 | ||
8042 | return ret; | |
8043 | } | |
8044 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8045 | |
8046 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8047 | static struct cgroup_subsys_state * |
8048 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8049 | { |
8050 | struct perf_cgroup *jc; | |
e5d1367f | 8051 | |
1b15d055 | 8052 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8053 | if (!jc) |
8054 | return ERR_PTR(-ENOMEM); | |
8055 | ||
e5d1367f SE |
8056 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8057 | if (!jc->info) { | |
8058 | kfree(jc); | |
8059 | return ERR_PTR(-ENOMEM); | |
8060 | } | |
8061 | ||
e5d1367f SE |
8062 | return &jc->css; |
8063 | } | |
8064 | ||
eb95419b | 8065 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8066 | { |
eb95419b TH |
8067 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8068 | ||
e5d1367f SE |
8069 | free_percpu(jc->info); |
8070 | kfree(jc); | |
8071 | } | |
8072 | ||
8073 | static int __perf_cgroup_move(void *info) | |
8074 | { | |
8075 | struct task_struct *task = info; | |
8076 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8077 | return 0; | |
8078 | } | |
8079 | ||
eb95419b TH |
8080 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8081 | struct cgroup_taskset *tset) | |
e5d1367f | 8082 | { |
bb9d97b6 TH |
8083 | struct task_struct *task; |
8084 | ||
924f0d9a | 8085 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 8086 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8087 | } |
8088 | ||
eb95419b TH |
8089 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8090 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8091 | struct task_struct *task) |
e5d1367f SE |
8092 | { |
8093 | /* | |
8094 | * cgroup_exit() is called in the copy_process() failure path. | |
8095 | * Ignore this case since the task hasn't ran yet, this avoids | |
8096 | * trying to poke a half freed task state from generic code. | |
8097 | */ | |
8098 | if (!(task->flags & PF_EXITING)) | |
8099 | return; | |
8100 | ||
bb9d97b6 | 8101 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8102 | } |
8103 | ||
073219e9 | 8104 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
8105 | .css_alloc = perf_cgroup_css_alloc, |
8106 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8107 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8108 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8109 | }; |
8110 | #endif /* CONFIG_CGROUP_PERF */ |