]>
Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
39bed6cb | 37 | #include <linux/cgroup.h> |
cdd6c482 | 38 | #include <linux/perf_event.h> |
6fb2915d | 39 | #include <linux/ftrace_event.h> |
3c502e7a | 40 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 41 | #include <linux/mm_types.h> |
c464c76e | 42 | #include <linux/module.h> |
f972eb63 | 43 | #include <linux/mman.h> |
b3f20785 | 44 | #include <linux/compat.h> |
0793a61d | 45 | |
76369139 FW |
46 | #include "internal.h" |
47 | ||
4e193bd4 TB |
48 | #include <asm/irq_regs.h> |
49 | ||
fadfe7be JO |
50 | static struct workqueue_struct *perf_wq; |
51 | ||
fe4b04fa | 52 | struct remote_function_call { |
e7e7ee2e IM |
53 | struct task_struct *p; |
54 | int (*func)(void *info); | |
55 | void *info; | |
56 | int ret; | |
fe4b04fa PZ |
57 | }; |
58 | ||
59 | static void remote_function(void *data) | |
60 | { | |
61 | struct remote_function_call *tfc = data; | |
62 | struct task_struct *p = tfc->p; | |
63 | ||
64 | if (p) { | |
65 | tfc->ret = -EAGAIN; | |
66 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
67 | return; | |
68 | } | |
69 | ||
70 | tfc->ret = tfc->func(tfc->info); | |
71 | } | |
72 | ||
73 | /** | |
74 | * task_function_call - call a function on the cpu on which a task runs | |
75 | * @p: the task to evaluate | |
76 | * @func: the function to be called | |
77 | * @info: the function call argument | |
78 | * | |
79 | * Calls the function @func when the task is currently running. This might | |
80 | * be on the current CPU, which just calls the function directly | |
81 | * | |
82 | * returns: @func return value, or | |
83 | * -ESRCH - when the process isn't running | |
84 | * -EAGAIN - when the process moved away | |
85 | */ | |
86 | static int | |
87 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
88 | { | |
89 | struct remote_function_call data = { | |
e7e7ee2e IM |
90 | .p = p, |
91 | .func = func, | |
92 | .info = info, | |
93 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
94 | }; |
95 | ||
96 | if (task_curr(p)) | |
97 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
98 | ||
99 | return data.ret; | |
100 | } | |
101 | ||
102 | /** | |
103 | * cpu_function_call - call a function on the cpu | |
104 | * @func: the function to be called | |
105 | * @info: the function call argument | |
106 | * | |
107 | * Calls the function @func on the remote cpu. | |
108 | * | |
109 | * returns: @func return value or -ENXIO when the cpu is offline | |
110 | */ | |
111 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
112 | { | |
113 | struct remote_function_call data = { | |
e7e7ee2e IM |
114 | .p = NULL, |
115 | .func = func, | |
116 | .info = info, | |
117 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
118 | }; |
119 | ||
120 | smp_call_function_single(cpu, remote_function, &data, 1); | |
121 | ||
122 | return data.ret; | |
123 | } | |
124 | ||
f8697762 JO |
125 | #define EVENT_OWNER_KERNEL ((void *) -1) |
126 | ||
127 | static bool is_kernel_event(struct perf_event *event) | |
128 | { | |
129 | return event->owner == EVENT_OWNER_KERNEL; | |
130 | } | |
131 | ||
e5d1367f SE |
132 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
133 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
134 | PERF_FLAG_PID_CGROUP |\ |
135 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 136 | |
bce38cd5 SE |
137 | /* |
138 | * branch priv levels that need permission checks | |
139 | */ | |
140 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
141 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
142 | PERF_SAMPLE_BRANCH_HV) | |
143 | ||
0b3fcf17 SE |
144 | enum event_type_t { |
145 | EVENT_FLEXIBLE = 0x1, | |
146 | EVENT_PINNED = 0x2, | |
147 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
148 | }; | |
149 | ||
e5d1367f SE |
150 | /* |
151 | * perf_sched_events : >0 events exist | |
152 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
153 | */ | |
c5905afb | 154 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 155 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
ba532500 | 156 | static DEFINE_PER_CPU(int, perf_sched_cb_usages); |
e5d1367f | 157 | |
cdd6c482 IM |
158 | static atomic_t nr_mmap_events __read_mostly; |
159 | static atomic_t nr_comm_events __read_mostly; | |
160 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 161 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 162 | |
108b02cf PZ |
163 | static LIST_HEAD(pmus); |
164 | static DEFINE_MUTEX(pmus_lock); | |
165 | static struct srcu_struct pmus_srcu; | |
166 | ||
0764771d | 167 | /* |
cdd6c482 | 168 | * perf event paranoia level: |
0fbdea19 IM |
169 | * -1 - not paranoid at all |
170 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 171 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 172 | * 2 - disallow kernel profiling for unpriv |
0764771d | 173 | */ |
cdd6c482 | 174 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 175 | |
20443384 FW |
176 | /* Minimum for 512 kiB + 1 user control page */ |
177 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
178 | |
179 | /* | |
cdd6c482 | 180 | * max perf event sample rate |
df58ab24 | 181 | */ |
14c63f17 DH |
182 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
183 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
184 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
185 | ||
186 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
187 | ||
188 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
189 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
190 | ||
d9494cb4 PZ |
191 | static int perf_sample_allowed_ns __read_mostly = |
192 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
193 | |
194 | void update_perf_cpu_limits(void) | |
195 | { | |
196 | u64 tmp = perf_sample_period_ns; | |
197 | ||
198 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 199 | do_div(tmp, 100); |
d9494cb4 | 200 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 201 | } |
163ec435 | 202 | |
9e630205 SE |
203 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
204 | ||
163ec435 PZ |
205 | int perf_proc_update_handler(struct ctl_table *table, int write, |
206 | void __user *buffer, size_t *lenp, | |
207 | loff_t *ppos) | |
208 | { | |
723478c8 | 209 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
210 | |
211 | if (ret || !write) | |
212 | return ret; | |
213 | ||
214 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
215 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
216 | update_perf_cpu_limits(); | |
217 | ||
218 | return 0; | |
219 | } | |
220 | ||
221 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
222 | ||
223 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
224 | void __user *buffer, size_t *lenp, | |
225 | loff_t *ppos) | |
226 | { | |
227 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
228 | ||
229 | if (ret || !write) | |
230 | return ret; | |
231 | ||
232 | update_perf_cpu_limits(); | |
163ec435 PZ |
233 | |
234 | return 0; | |
235 | } | |
1ccd1549 | 236 | |
14c63f17 DH |
237 | /* |
238 | * perf samples are done in some very critical code paths (NMIs). | |
239 | * If they take too much CPU time, the system can lock up and not | |
240 | * get any real work done. This will drop the sample rate when | |
241 | * we detect that events are taking too long. | |
242 | */ | |
243 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 244 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 245 | |
6a02ad66 | 246 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 247 | { |
6a02ad66 | 248 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 249 | u64 avg_local_sample_len; |
e5302920 | 250 | u64 local_samples_len; |
6a02ad66 | 251 | |
4a32fea9 | 252 | local_samples_len = __this_cpu_read(running_sample_length); |
6a02ad66 PZ |
253 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; |
254 | ||
255 | printk_ratelimited(KERN_WARNING | |
256 | "perf interrupt took too long (%lld > %lld), lowering " | |
257 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 258 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
259 | sysctl_perf_event_sample_rate); |
260 | } | |
261 | ||
262 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
263 | ||
264 | void perf_sample_event_took(u64 sample_len_ns) | |
265 | { | |
d9494cb4 | 266 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
267 | u64 avg_local_sample_len; |
268 | u64 local_samples_len; | |
14c63f17 | 269 | |
d9494cb4 | 270 | if (allowed_ns == 0) |
14c63f17 DH |
271 | return; |
272 | ||
273 | /* decay the counter by 1 average sample */ | |
4a32fea9 | 274 | local_samples_len = __this_cpu_read(running_sample_length); |
14c63f17 DH |
275 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; |
276 | local_samples_len += sample_len_ns; | |
4a32fea9 | 277 | __this_cpu_write(running_sample_length, local_samples_len); |
14c63f17 DH |
278 | |
279 | /* | |
280 | * note: this will be biased artifically low until we have | |
281 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
282 | * from having to maintain a count. | |
283 | */ | |
284 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
285 | ||
d9494cb4 | 286 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
287 | return; |
288 | ||
289 | if (max_samples_per_tick <= 1) | |
290 | return; | |
291 | ||
292 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
293 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
294 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
295 | ||
14c63f17 | 296 | update_perf_cpu_limits(); |
6a02ad66 | 297 | |
cd578abb PZ |
298 | if (!irq_work_queue(&perf_duration_work)) { |
299 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
300 | "kernel.perf_event_max_sample_rate to %d\n", | |
301 | avg_local_sample_len, allowed_ns >> 1, | |
302 | sysctl_perf_event_sample_rate); | |
303 | } | |
14c63f17 DH |
304 | } |
305 | ||
cdd6c482 | 306 | static atomic64_t perf_event_id; |
a96bbc16 | 307 | |
0b3fcf17 SE |
308 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
309 | enum event_type_t event_type); | |
310 | ||
311 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
312 | enum event_type_t event_type, |
313 | struct task_struct *task); | |
314 | ||
315 | static void update_context_time(struct perf_event_context *ctx); | |
316 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 317 | |
cdd6c482 | 318 | void __weak perf_event_print_debug(void) { } |
0793a61d | 319 | |
84c79910 | 320 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 321 | { |
84c79910 | 322 | return "pmu"; |
0793a61d TG |
323 | } |
324 | ||
0b3fcf17 SE |
325 | static inline u64 perf_clock(void) |
326 | { | |
327 | return local_clock(); | |
328 | } | |
329 | ||
e5d1367f SE |
330 | static inline struct perf_cpu_context * |
331 | __get_cpu_context(struct perf_event_context *ctx) | |
332 | { | |
333 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
334 | } | |
335 | ||
facc4307 PZ |
336 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
337 | struct perf_event_context *ctx) | |
338 | { | |
339 | raw_spin_lock(&cpuctx->ctx.lock); | |
340 | if (ctx) | |
341 | raw_spin_lock(&ctx->lock); | |
342 | } | |
343 | ||
344 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
345 | struct perf_event_context *ctx) | |
346 | { | |
347 | if (ctx) | |
348 | raw_spin_unlock(&ctx->lock); | |
349 | raw_spin_unlock(&cpuctx->ctx.lock); | |
350 | } | |
351 | ||
e5d1367f SE |
352 | #ifdef CONFIG_CGROUP_PERF |
353 | ||
e5d1367f SE |
354 | static inline bool |
355 | perf_cgroup_match(struct perf_event *event) | |
356 | { | |
357 | struct perf_event_context *ctx = event->ctx; | |
358 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
359 | ||
ef824fa1 TH |
360 | /* @event doesn't care about cgroup */ |
361 | if (!event->cgrp) | |
362 | return true; | |
363 | ||
364 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
365 | if (!cpuctx->cgrp) | |
366 | return false; | |
367 | ||
368 | /* | |
369 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
370 | * also enabled for all its descendant cgroups. If @cpuctx's | |
371 | * cgroup is a descendant of @event's (the test covers identity | |
372 | * case), it's a match. | |
373 | */ | |
374 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
375 | event->cgrp->css.cgroup); | |
e5d1367f SE |
376 | } |
377 | ||
e5d1367f SE |
378 | static inline void perf_detach_cgroup(struct perf_event *event) |
379 | { | |
4e2ba650 | 380 | css_put(&event->cgrp->css); |
e5d1367f SE |
381 | event->cgrp = NULL; |
382 | } | |
383 | ||
384 | static inline int is_cgroup_event(struct perf_event *event) | |
385 | { | |
386 | return event->cgrp != NULL; | |
387 | } | |
388 | ||
389 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
390 | { | |
391 | struct perf_cgroup_info *t; | |
392 | ||
393 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
394 | return t->time; | |
395 | } | |
396 | ||
397 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
398 | { | |
399 | struct perf_cgroup_info *info; | |
400 | u64 now; | |
401 | ||
402 | now = perf_clock(); | |
403 | ||
404 | info = this_cpu_ptr(cgrp->info); | |
405 | ||
406 | info->time += now - info->timestamp; | |
407 | info->timestamp = now; | |
408 | } | |
409 | ||
410 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
411 | { | |
412 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
413 | if (cgrp_out) | |
414 | __update_cgrp_time(cgrp_out); | |
415 | } | |
416 | ||
417 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
418 | { | |
3f7cce3c SE |
419 | struct perf_cgroup *cgrp; |
420 | ||
e5d1367f | 421 | /* |
3f7cce3c SE |
422 | * ensure we access cgroup data only when needed and |
423 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 424 | */ |
3f7cce3c | 425 | if (!is_cgroup_event(event)) |
e5d1367f SE |
426 | return; |
427 | ||
3f7cce3c SE |
428 | cgrp = perf_cgroup_from_task(current); |
429 | /* | |
430 | * Do not update time when cgroup is not active | |
431 | */ | |
432 | if (cgrp == event->cgrp) | |
433 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
434 | } |
435 | ||
436 | static inline void | |
3f7cce3c SE |
437 | perf_cgroup_set_timestamp(struct task_struct *task, |
438 | struct perf_event_context *ctx) | |
e5d1367f SE |
439 | { |
440 | struct perf_cgroup *cgrp; | |
441 | struct perf_cgroup_info *info; | |
442 | ||
3f7cce3c SE |
443 | /* |
444 | * ctx->lock held by caller | |
445 | * ensure we do not access cgroup data | |
446 | * unless we have the cgroup pinned (css_get) | |
447 | */ | |
448 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
449 | return; |
450 | ||
451 | cgrp = perf_cgroup_from_task(task); | |
452 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 453 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
454 | } |
455 | ||
456 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
457 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
458 | ||
459 | /* | |
460 | * reschedule events based on the cgroup constraint of task. | |
461 | * | |
462 | * mode SWOUT : schedule out everything | |
463 | * mode SWIN : schedule in based on cgroup for next | |
464 | */ | |
465 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
466 | { | |
467 | struct perf_cpu_context *cpuctx; | |
468 | struct pmu *pmu; | |
469 | unsigned long flags; | |
470 | ||
471 | /* | |
472 | * disable interrupts to avoid geting nr_cgroup | |
473 | * changes via __perf_event_disable(). Also | |
474 | * avoids preemption. | |
475 | */ | |
476 | local_irq_save(flags); | |
477 | ||
478 | /* | |
479 | * we reschedule only in the presence of cgroup | |
480 | * constrained events. | |
481 | */ | |
482 | rcu_read_lock(); | |
483 | ||
484 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 485 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
486 | if (cpuctx->unique_pmu != pmu) |
487 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 488 | |
e5d1367f SE |
489 | /* |
490 | * perf_cgroup_events says at least one | |
491 | * context on this CPU has cgroup events. | |
492 | * | |
493 | * ctx->nr_cgroups reports the number of cgroup | |
494 | * events for a context. | |
495 | */ | |
496 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
497 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
498 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
499 | |
500 | if (mode & PERF_CGROUP_SWOUT) { | |
501 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
502 | /* | |
503 | * must not be done before ctxswout due | |
504 | * to event_filter_match() in event_sched_out() | |
505 | */ | |
506 | cpuctx->cgrp = NULL; | |
507 | } | |
508 | ||
509 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 510 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
511 | /* |
512 | * set cgrp before ctxsw in to allow | |
513 | * event_filter_match() to not have to pass | |
514 | * task around | |
e5d1367f SE |
515 | */ |
516 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
517 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
518 | } | |
facc4307 PZ |
519 | perf_pmu_enable(cpuctx->ctx.pmu); |
520 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 521 | } |
e5d1367f SE |
522 | } |
523 | ||
524 | rcu_read_unlock(); | |
525 | ||
526 | local_irq_restore(flags); | |
527 | } | |
528 | ||
a8d757ef SE |
529 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
530 | struct task_struct *next) | |
e5d1367f | 531 | { |
a8d757ef SE |
532 | struct perf_cgroup *cgrp1; |
533 | struct perf_cgroup *cgrp2 = NULL; | |
534 | ||
535 | /* | |
536 | * we come here when we know perf_cgroup_events > 0 | |
537 | */ | |
538 | cgrp1 = perf_cgroup_from_task(task); | |
539 | ||
540 | /* | |
541 | * next is NULL when called from perf_event_enable_on_exec() | |
542 | * that will systematically cause a cgroup_switch() | |
543 | */ | |
544 | if (next) | |
545 | cgrp2 = perf_cgroup_from_task(next); | |
546 | ||
547 | /* | |
548 | * only schedule out current cgroup events if we know | |
549 | * that we are switching to a different cgroup. Otherwise, | |
550 | * do no touch the cgroup events. | |
551 | */ | |
552 | if (cgrp1 != cgrp2) | |
553 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
554 | } |
555 | ||
a8d757ef SE |
556 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
557 | struct task_struct *task) | |
e5d1367f | 558 | { |
a8d757ef SE |
559 | struct perf_cgroup *cgrp1; |
560 | struct perf_cgroup *cgrp2 = NULL; | |
561 | ||
562 | /* | |
563 | * we come here when we know perf_cgroup_events > 0 | |
564 | */ | |
565 | cgrp1 = perf_cgroup_from_task(task); | |
566 | ||
567 | /* prev can never be NULL */ | |
568 | cgrp2 = perf_cgroup_from_task(prev); | |
569 | ||
570 | /* | |
571 | * only need to schedule in cgroup events if we are changing | |
572 | * cgroup during ctxsw. Cgroup events were not scheduled | |
573 | * out of ctxsw out if that was not the case. | |
574 | */ | |
575 | if (cgrp1 != cgrp2) | |
576 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
577 | } |
578 | ||
579 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
580 | struct perf_event_attr *attr, | |
581 | struct perf_event *group_leader) | |
582 | { | |
583 | struct perf_cgroup *cgrp; | |
584 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
585 | struct fd f = fdget(fd); |
586 | int ret = 0; | |
e5d1367f | 587 | |
2903ff01 | 588 | if (!f.file) |
e5d1367f SE |
589 | return -EBADF; |
590 | ||
b583043e | 591 | css = css_tryget_online_from_dir(f.file->f_path.dentry, |
ec903c0c | 592 | &perf_event_cgrp_subsys); |
3db272c0 LZ |
593 | if (IS_ERR(css)) { |
594 | ret = PTR_ERR(css); | |
595 | goto out; | |
596 | } | |
e5d1367f SE |
597 | |
598 | cgrp = container_of(css, struct perf_cgroup, css); | |
599 | event->cgrp = cgrp; | |
600 | ||
601 | /* | |
602 | * all events in a group must monitor | |
603 | * the same cgroup because a task belongs | |
604 | * to only one perf cgroup at a time | |
605 | */ | |
606 | if (group_leader && group_leader->cgrp != cgrp) { | |
607 | perf_detach_cgroup(event); | |
608 | ret = -EINVAL; | |
e5d1367f | 609 | } |
3db272c0 | 610 | out: |
2903ff01 | 611 | fdput(f); |
e5d1367f SE |
612 | return ret; |
613 | } | |
614 | ||
615 | static inline void | |
616 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
617 | { | |
618 | struct perf_cgroup_info *t; | |
619 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
620 | event->shadow_ctx_time = now - t->timestamp; | |
621 | } | |
622 | ||
623 | static inline void | |
624 | perf_cgroup_defer_enabled(struct perf_event *event) | |
625 | { | |
626 | /* | |
627 | * when the current task's perf cgroup does not match | |
628 | * the event's, we need to remember to call the | |
629 | * perf_mark_enable() function the first time a task with | |
630 | * a matching perf cgroup is scheduled in. | |
631 | */ | |
632 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
633 | event->cgrp_defer_enabled = 1; | |
634 | } | |
635 | ||
636 | static inline void | |
637 | perf_cgroup_mark_enabled(struct perf_event *event, | |
638 | struct perf_event_context *ctx) | |
639 | { | |
640 | struct perf_event *sub; | |
641 | u64 tstamp = perf_event_time(event); | |
642 | ||
643 | if (!event->cgrp_defer_enabled) | |
644 | return; | |
645 | ||
646 | event->cgrp_defer_enabled = 0; | |
647 | ||
648 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
649 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
650 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
651 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
652 | sub->cgrp_defer_enabled = 0; | |
653 | } | |
654 | } | |
655 | } | |
656 | #else /* !CONFIG_CGROUP_PERF */ | |
657 | ||
658 | static inline bool | |
659 | perf_cgroup_match(struct perf_event *event) | |
660 | { | |
661 | return true; | |
662 | } | |
663 | ||
664 | static inline void perf_detach_cgroup(struct perf_event *event) | |
665 | {} | |
666 | ||
667 | static inline int is_cgroup_event(struct perf_event *event) | |
668 | { | |
669 | return 0; | |
670 | } | |
671 | ||
672 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
673 | { | |
674 | return 0; | |
675 | } | |
676 | ||
677 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
678 | { | |
679 | } | |
680 | ||
681 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
682 | { | |
683 | } | |
684 | ||
a8d757ef SE |
685 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
686 | struct task_struct *next) | |
e5d1367f SE |
687 | { |
688 | } | |
689 | ||
a8d757ef SE |
690 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
691 | struct task_struct *task) | |
e5d1367f SE |
692 | { |
693 | } | |
694 | ||
695 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
696 | struct perf_event_attr *attr, | |
697 | struct perf_event *group_leader) | |
698 | { | |
699 | return -EINVAL; | |
700 | } | |
701 | ||
702 | static inline void | |
3f7cce3c SE |
703 | perf_cgroup_set_timestamp(struct task_struct *task, |
704 | struct perf_event_context *ctx) | |
e5d1367f SE |
705 | { |
706 | } | |
707 | ||
708 | void | |
709 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
710 | { | |
711 | } | |
712 | ||
713 | static inline void | |
714 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
715 | { | |
716 | } | |
717 | ||
718 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
719 | { | |
720 | return 0; | |
721 | } | |
722 | ||
723 | static inline void | |
724 | perf_cgroup_defer_enabled(struct perf_event *event) | |
725 | { | |
726 | } | |
727 | ||
728 | static inline void | |
729 | perf_cgroup_mark_enabled(struct perf_event *event, | |
730 | struct perf_event_context *ctx) | |
731 | { | |
732 | } | |
733 | #endif | |
734 | ||
9e630205 SE |
735 | /* |
736 | * set default to be dependent on timer tick just | |
737 | * like original code | |
738 | */ | |
739 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
740 | /* | |
741 | * function must be called with interrupts disbled | |
742 | */ | |
743 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
744 | { | |
745 | struct perf_cpu_context *cpuctx; | |
746 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
747 | int rotations = 0; | |
748 | ||
749 | WARN_ON(!irqs_disabled()); | |
750 | ||
751 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
752 | ||
753 | rotations = perf_rotate_context(cpuctx); | |
754 | ||
755 | /* | |
756 | * arm timer if needed | |
757 | */ | |
758 | if (rotations) { | |
759 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
760 | ret = HRTIMER_RESTART; | |
761 | } | |
762 | ||
763 | return ret; | |
764 | } | |
765 | ||
766 | /* CPU is going down */ | |
767 | void perf_cpu_hrtimer_cancel(int cpu) | |
768 | { | |
769 | struct perf_cpu_context *cpuctx; | |
770 | struct pmu *pmu; | |
771 | unsigned long flags; | |
772 | ||
773 | if (WARN_ON(cpu != smp_processor_id())) | |
774 | return; | |
775 | ||
776 | local_irq_save(flags); | |
777 | ||
778 | rcu_read_lock(); | |
779 | ||
780 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
781 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
782 | ||
783 | if (pmu->task_ctx_nr == perf_sw_context) | |
784 | continue; | |
785 | ||
786 | hrtimer_cancel(&cpuctx->hrtimer); | |
787 | } | |
788 | ||
789 | rcu_read_unlock(); | |
790 | ||
791 | local_irq_restore(flags); | |
792 | } | |
793 | ||
794 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
795 | { | |
796 | struct hrtimer *hr = &cpuctx->hrtimer; | |
797 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 798 | int timer; |
9e630205 SE |
799 | |
800 | /* no multiplexing needed for SW PMU */ | |
801 | if (pmu->task_ctx_nr == perf_sw_context) | |
802 | return; | |
803 | ||
62b85639 SE |
804 | /* |
805 | * check default is sane, if not set then force to | |
806 | * default interval (1/tick) | |
807 | */ | |
808 | timer = pmu->hrtimer_interval_ms; | |
809 | if (timer < 1) | |
810 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
811 | ||
812 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
813 | |
814 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
815 | hr->function = perf_cpu_hrtimer_handler; | |
816 | } | |
817 | ||
818 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
819 | { | |
820 | struct hrtimer *hr = &cpuctx->hrtimer; | |
821 | struct pmu *pmu = cpuctx->ctx.pmu; | |
822 | ||
823 | /* not for SW PMU */ | |
824 | if (pmu->task_ctx_nr == perf_sw_context) | |
825 | return; | |
826 | ||
827 | if (hrtimer_active(hr)) | |
828 | return; | |
829 | ||
830 | if (!hrtimer_callback_running(hr)) | |
831 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
832 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
833 | } | |
834 | ||
33696fc0 | 835 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 836 | { |
33696fc0 PZ |
837 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
838 | if (!(*count)++) | |
839 | pmu->pmu_disable(pmu); | |
9e35ad38 | 840 | } |
9e35ad38 | 841 | |
33696fc0 | 842 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 843 | { |
33696fc0 PZ |
844 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
845 | if (!--(*count)) | |
846 | pmu->pmu_enable(pmu); | |
9e35ad38 | 847 | } |
9e35ad38 | 848 | |
2fde4f94 | 849 | static DEFINE_PER_CPU(struct list_head, active_ctx_list); |
e9d2b064 PZ |
850 | |
851 | /* | |
2fde4f94 MR |
852 | * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and |
853 | * perf_event_task_tick() are fully serialized because they're strictly cpu | |
854 | * affine and perf_event_ctx{activate,deactivate} are called with IRQs | |
855 | * disabled, while perf_event_task_tick is called from IRQ context. | |
e9d2b064 | 856 | */ |
2fde4f94 | 857 | static void perf_event_ctx_activate(struct perf_event_context *ctx) |
9e35ad38 | 858 | { |
2fde4f94 | 859 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
b5ab4cd5 | 860 | |
e9d2b064 | 861 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 862 | |
2fde4f94 MR |
863 | WARN_ON(!list_empty(&ctx->active_ctx_list)); |
864 | ||
865 | list_add(&ctx->active_ctx_list, head); | |
866 | } | |
867 | ||
868 | static void perf_event_ctx_deactivate(struct perf_event_context *ctx) | |
869 | { | |
870 | WARN_ON(!irqs_disabled()); | |
871 | ||
872 | WARN_ON(list_empty(&ctx->active_ctx_list)); | |
873 | ||
874 | list_del_init(&ctx->active_ctx_list); | |
9e35ad38 | 875 | } |
9e35ad38 | 876 | |
cdd6c482 | 877 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 878 | { |
e5289d4a | 879 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
880 | } |
881 | ||
4af57ef2 YZ |
882 | static void free_ctx(struct rcu_head *head) |
883 | { | |
884 | struct perf_event_context *ctx; | |
885 | ||
886 | ctx = container_of(head, struct perf_event_context, rcu_head); | |
887 | kfree(ctx->task_ctx_data); | |
888 | kfree(ctx); | |
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); | |
4af57ef2 | 898 | call_rcu(&ctx->rcu_head, free_ctx); |
564c2b21 | 899 | } |
a63eaf34 PM |
900 | } |
901 | ||
f63a8daa PZ |
902 | /* |
903 | * Because of perf_event::ctx migration in sys_perf_event_open::move_group and | |
904 | * perf_pmu_migrate_context() we need some magic. | |
905 | * | |
906 | * Those places that change perf_event::ctx will hold both | |
907 | * perf_event_ctx::mutex of the 'old' and 'new' ctx value. | |
908 | * | |
909 | * Lock ordering is by mutex address. There is one other site where | |
910 | * perf_event_context::mutex nests and that is put_event(). But remember that | |
911 | * that is a parent<->child context relation, and migration does not affect | |
912 | * children, therefore these two orderings should not interact. | |
913 | * | |
914 | * The change in perf_event::ctx does not affect children (as claimed above) | |
915 | * because the sys_perf_event_open() case will install a new event and break | |
916 | * the ctx parent<->child relation, and perf_pmu_migrate_context() is only | |
917 | * concerned with cpuctx and that doesn't have children. | |
918 | * | |
919 | * The places that change perf_event::ctx will issue: | |
920 | * | |
921 | * perf_remove_from_context(); | |
922 | * synchronize_rcu(); | |
923 | * perf_install_in_context(); | |
924 | * | |
925 | * to affect the change. The remove_from_context() + synchronize_rcu() should | |
926 | * quiesce the event, after which we can install it in the new location. This | |
927 | * means that only external vectors (perf_fops, prctl) can perturb the event | |
928 | * while in transit. Therefore all such accessors should also acquire | |
929 | * perf_event_context::mutex to serialize against this. | |
930 | * | |
931 | * However; because event->ctx can change while we're waiting to acquire | |
932 | * ctx->mutex we must be careful and use the below perf_event_ctx_lock() | |
933 | * function. | |
934 | * | |
935 | * Lock order: | |
936 | * task_struct::perf_event_mutex | |
937 | * perf_event_context::mutex | |
938 | * perf_event_context::lock | |
939 | * perf_event::child_mutex; | |
940 | * perf_event::mmap_mutex | |
941 | * mmap_sem | |
942 | */ | |
a83fe28e PZ |
943 | static struct perf_event_context * |
944 | perf_event_ctx_lock_nested(struct perf_event *event, int nesting) | |
f63a8daa PZ |
945 | { |
946 | struct perf_event_context *ctx; | |
947 | ||
948 | again: | |
949 | rcu_read_lock(); | |
950 | ctx = ACCESS_ONCE(event->ctx); | |
951 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
952 | rcu_read_unlock(); | |
953 | goto again; | |
954 | } | |
955 | rcu_read_unlock(); | |
956 | ||
a83fe28e | 957 | mutex_lock_nested(&ctx->mutex, nesting); |
f63a8daa PZ |
958 | if (event->ctx != ctx) { |
959 | mutex_unlock(&ctx->mutex); | |
960 | put_ctx(ctx); | |
961 | goto again; | |
962 | } | |
963 | ||
964 | return ctx; | |
965 | } | |
966 | ||
a83fe28e PZ |
967 | static inline struct perf_event_context * |
968 | perf_event_ctx_lock(struct perf_event *event) | |
969 | { | |
970 | return perf_event_ctx_lock_nested(event, 0); | |
971 | } | |
972 | ||
f63a8daa PZ |
973 | static void perf_event_ctx_unlock(struct perf_event *event, |
974 | struct perf_event_context *ctx) | |
975 | { | |
976 | mutex_unlock(&ctx->mutex); | |
977 | put_ctx(ctx); | |
978 | } | |
979 | ||
211de6eb PZ |
980 | /* |
981 | * This must be done under the ctx->lock, such as to serialize against | |
982 | * context_equiv(), therefore we cannot call put_ctx() since that might end up | |
983 | * calling scheduler related locks and ctx->lock nests inside those. | |
984 | */ | |
985 | static __must_check struct perf_event_context * | |
986 | unclone_ctx(struct perf_event_context *ctx) | |
71a851b4 | 987 | { |
211de6eb PZ |
988 | struct perf_event_context *parent_ctx = ctx->parent_ctx; |
989 | ||
990 | lockdep_assert_held(&ctx->lock); | |
991 | ||
992 | if (parent_ctx) | |
71a851b4 | 993 | ctx->parent_ctx = NULL; |
5a3126d4 | 994 | ctx->generation++; |
211de6eb PZ |
995 | |
996 | return parent_ctx; | |
71a851b4 PZ |
997 | } |
998 | ||
6844c09d ACM |
999 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
1000 | { | |
1001 | /* | |
1002 | * only top level events have the pid namespace they were created in | |
1003 | */ | |
1004 | if (event->parent) | |
1005 | event = event->parent; | |
1006 | ||
1007 | return task_tgid_nr_ns(p, event->ns); | |
1008 | } | |
1009 | ||
1010 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
1011 | { | |
1012 | /* | |
1013 | * only top level events have the pid namespace they were created in | |
1014 | */ | |
1015 | if (event->parent) | |
1016 | event = event->parent; | |
1017 | ||
1018 | return task_pid_nr_ns(p, event->ns); | |
1019 | } | |
1020 | ||
7f453c24 | 1021 | /* |
cdd6c482 | 1022 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
1023 | * to userspace. |
1024 | */ | |
cdd6c482 | 1025 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 1026 | { |
cdd6c482 | 1027 | u64 id = event->id; |
7f453c24 | 1028 | |
cdd6c482 IM |
1029 | if (event->parent) |
1030 | id = event->parent->id; | |
7f453c24 PZ |
1031 | |
1032 | return id; | |
1033 | } | |
1034 | ||
25346b93 | 1035 | /* |
cdd6c482 | 1036 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
1037 | * This has to cope with with the fact that until it is locked, |
1038 | * the context could get moved to another task. | |
1039 | */ | |
cdd6c482 | 1040 | static struct perf_event_context * |
8dc85d54 | 1041 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 1042 | { |
cdd6c482 | 1043 | struct perf_event_context *ctx; |
25346b93 | 1044 | |
9ed6060d | 1045 | retry: |
058ebd0e PZ |
1046 | /* |
1047 | * One of the few rules of preemptible RCU is that one cannot do | |
1048 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
1049 | * part of the read side critical section was preemptible -- see | |
1050 | * rcu_read_unlock_special(). | |
1051 | * | |
1052 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
1053 | * side critical section is non-preemptible. | |
1054 | */ | |
1055 | preempt_disable(); | |
1056 | rcu_read_lock(); | |
8dc85d54 | 1057 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
1058 | if (ctx) { |
1059 | /* | |
1060 | * If this context is a clone of another, it might | |
1061 | * get swapped for another underneath us by | |
cdd6c482 | 1062 | * perf_event_task_sched_out, though the |
25346b93 PM |
1063 | * rcu_read_lock() protects us from any context |
1064 | * getting freed. Lock the context and check if it | |
1065 | * got swapped before we could get the lock, and retry | |
1066 | * if so. If we locked the right context, then it | |
1067 | * can't get swapped on us any more. | |
1068 | */ | |
e625cce1 | 1069 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 1070 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 1071 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
1072 | rcu_read_unlock(); |
1073 | preempt_enable(); | |
25346b93 PM |
1074 | goto retry; |
1075 | } | |
b49a9e7e PZ |
1076 | |
1077 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 1078 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1079 | ctx = NULL; |
1080 | } | |
25346b93 PM |
1081 | } |
1082 | rcu_read_unlock(); | |
058ebd0e | 1083 | preempt_enable(); |
25346b93 PM |
1084 | return ctx; |
1085 | } | |
1086 | ||
1087 | /* | |
1088 | * Get the context for a task and increment its pin_count so it | |
1089 | * can't get swapped to another task. This also increments its | |
1090 | * reference count so that the context can't get freed. | |
1091 | */ | |
8dc85d54 PZ |
1092 | static struct perf_event_context * |
1093 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1094 | { |
cdd6c482 | 1095 | struct perf_event_context *ctx; |
25346b93 PM |
1096 | unsigned long flags; |
1097 | ||
8dc85d54 | 1098 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1099 | if (ctx) { |
1100 | ++ctx->pin_count; | |
e625cce1 | 1101 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1102 | } |
1103 | return ctx; | |
1104 | } | |
1105 | ||
cdd6c482 | 1106 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1107 | { |
1108 | unsigned long flags; | |
1109 | ||
e625cce1 | 1110 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1111 | --ctx->pin_count; |
e625cce1 | 1112 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1113 | } |
1114 | ||
f67218c3 PZ |
1115 | /* |
1116 | * Update the record of the current time in a context. | |
1117 | */ | |
1118 | static void update_context_time(struct perf_event_context *ctx) | |
1119 | { | |
1120 | u64 now = perf_clock(); | |
1121 | ||
1122 | ctx->time += now - ctx->timestamp; | |
1123 | ctx->timestamp = now; | |
1124 | } | |
1125 | ||
4158755d SE |
1126 | static u64 perf_event_time(struct perf_event *event) |
1127 | { | |
1128 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1129 | |
1130 | if (is_cgroup_event(event)) | |
1131 | return perf_cgroup_event_time(event); | |
1132 | ||
4158755d SE |
1133 | return ctx ? ctx->time : 0; |
1134 | } | |
1135 | ||
f67218c3 PZ |
1136 | /* |
1137 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1138 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1139 | */ |
1140 | static void update_event_times(struct perf_event *event) | |
1141 | { | |
1142 | struct perf_event_context *ctx = event->ctx; | |
1143 | u64 run_end; | |
1144 | ||
1145 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1146 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1147 | return; | |
e5d1367f SE |
1148 | /* |
1149 | * in cgroup mode, time_enabled represents | |
1150 | * the time the event was enabled AND active | |
1151 | * tasks were in the monitored cgroup. This is | |
1152 | * independent of the activity of the context as | |
1153 | * there may be a mix of cgroup and non-cgroup events. | |
1154 | * | |
1155 | * That is why we treat cgroup events differently | |
1156 | * here. | |
1157 | */ | |
1158 | if (is_cgroup_event(event)) | |
46cd6a7f | 1159 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1160 | else if (ctx->is_active) |
1161 | run_end = ctx->time; | |
acd1d7c1 PZ |
1162 | else |
1163 | run_end = event->tstamp_stopped; | |
1164 | ||
1165 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1166 | |
1167 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1168 | run_end = event->tstamp_stopped; | |
1169 | else | |
4158755d | 1170 | run_end = perf_event_time(event); |
f67218c3 PZ |
1171 | |
1172 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1173 | |
f67218c3 PZ |
1174 | } |
1175 | ||
96c21a46 PZ |
1176 | /* |
1177 | * Update total_time_enabled and total_time_running for all events in a group. | |
1178 | */ | |
1179 | static void update_group_times(struct perf_event *leader) | |
1180 | { | |
1181 | struct perf_event *event; | |
1182 | ||
1183 | update_event_times(leader); | |
1184 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1185 | update_event_times(event); | |
1186 | } | |
1187 | ||
889ff015 FW |
1188 | static struct list_head * |
1189 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1190 | { | |
1191 | if (event->attr.pinned) | |
1192 | return &ctx->pinned_groups; | |
1193 | else | |
1194 | return &ctx->flexible_groups; | |
1195 | } | |
1196 | ||
fccc714b | 1197 | /* |
cdd6c482 | 1198 | * Add a event from the lists for its context. |
fccc714b PZ |
1199 | * Must be called with ctx->mutex and ctx->lock held. |
1200 | */ | |
04289bb9 | 1201 | static void |
cdd6c482 | 1202 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1203 | { |
8a49542c PZ |
1204 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1205 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1206 | |
1207 | /* | |
8a49542c PZ |
1208 | * If we're a stand alone event or group leader, we go to the context |
1209 | * list, group events are kept attached to the group so that | |
1210 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1211 | */ |
8a49542c | 1212 | if (event->group_leader == event) { |
889ff015 FW |
1213 | struct list_head *list; |
1214 | ||
d6f962b5 FW |
1215 | if (is_software_event(event)) |
1216 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1217 | ||
889ff015 FW |
1218 | list = ctx_group_list(event, ctx); |
1219 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1220 | } |
592903cd | 1221 | |
08309379 | 1222 | if (is_cgroup_event(event)) |
e5d1367f | 1223 | ctx->nr_cgroups++; |
e5d1367f | 1224 | |
cdd6c482 IM |
1225 | list_add_rcu(&event->event_entry, &ctx->event_list); |
1226 | ctx->nr_events++; | |
1227 | if (event->attr.inherit_stat) | |
bfbd3381 | 1228 | ctx->nr_stat++; |
5a3126d4 PZ |
1229 | |
1230 | ctx->generation++; | |
04289bb9 IM |
1231 | } |
1232 | ||
0231bb53 JO |
1233 | /* |
1234 | * Initialize event state based on the perf_event_attr::disabled. | |
1235 | */ | |
1236 | static inline void perf_event__state_init(struct perf_event *event) | |
1237 | { | |
1238 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1239 | PERF_EVENT_STATE_INACTIVE; | |
1240 | } | |
1241 | ||
c320c7b7 ACM |
1242 | /* |
1243 | * Called at perf_event creation and when events are attached/detached from a | |
1244 | * group. | |
1245 | */ | |
1246 | static void perf_event__read_size(struct perf_event *event) | |
1247 | { | |
1248 | int entry = sizeof(u64); /* value */ | |
1249 | int size = 0; | |
1250 | int nr = 1; | |
1251 | ||
1252 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1253 | size += sizeof(u64); | |
1254 | ||
1255 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1256 | size += sizeof(u64); | |
1257 | ||
1258 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1259 | entry += sizeof(u64); | |
1260 | ||
1261 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1262 | nr += event->group_leader->nr_siblings; | |
1263 | size += sizeof(u64); | |
1264 | } | |
1265 | ||
1266 | size += entry * nr; | |
1267 | event->read_size = size; | |
1268 | } | |
1269 | ||
1270 | static void perf_event__header_size(struct perf_event *event) | |
1271 | { | |
1272 | struct perf_sample_data *data; | |
1273 | u64 sample_type = event->attr.sample_type; | |
1274 | u16 size = 0; | |
1275 | ||
1276 | perf_event__read_size(event); | |
1277 | ||
1278 | if (sample_type & PERF_SAMPLE_IP) | |
1279 | size += sizeof(data->ip); | |
1280 | ||
6844c09d ACM |
1281 | if (sample_type & PERF_SAMPLE_ADDR) |
1282 | size += sizeof(data->addr); | |
1283 | ||
1284 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1285 | size += sizeof(data->period); | |
1286 | ||
c3feedf2 AK |
1287 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1288 | size += sizeof(data->weight); | |
1289 | ||
6844c09d ACM |
1290 | if (sample_type & PERF_SAMPLE_READ) |
1291 | size += event->read_size; | |
1292 | ||
d6be9ad6 SE |
1293 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1294 | size += sizeof(data->data_src.val); | |
1295 | ||
fdfbbd07 AK |
1296 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1297 | size += sizeof(data->txn); | |
1298 | ||
6844c09d ACM |
1299 | event->header_size = size; |
1300 | } | |
1301 | ||
1302 | static void perf_event__id_header_size(struct perf_event *event) | |
1303 | { | |
1304 | struct perf_sample_data *data; | |
1305 | u64 sample_type = event->attr.sample_type; | |
1306 | u16 size = 0; | |
1307 | ||
c320c7b7 ACM |
1308 | if (sample_type & PERF_SAMPLE_TID) |
1309 | size += sizeof(data->tid_entry); | |
1310 | ||
1311 | if (sample_type & PERF_SAMPLE_TIME) | |
1312 | size += sizeof(data->time); | |
1313 | ||
ff3d527c AH |
1314 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1315 | size += sizeof(data->id); | |
1316 | ||
c320c7b7 ACM |
1317 | if (sample_type & PERF_SAMPLE_ID) |
1318 | size += sizeof(data->id); | |
1319 | ||
1320 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1321 | size += sizeof(data->stream_id); | |
1322 | ||
1323 | if (sample_type & PERF_SAMPLE_CPU) | |
1324 | size += sizeof(data->cpu_entry); | |
1325 | ||
6844c09d | 1326 | event->id_header_size = size; |
c320c7b7 ACM |
1327 | } |
1328 | ||
8a49542c PZ |
1329 | static void perf_group_attach(struct perf_event *event) |
1330 | { | |
c320c7b7 | 1331 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1332 | |
74c3337c PZ |
1333 | /* |
1334 | * We can have double attach due to group movement in perf_event_open. | |
1335 | */ | |
1336 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1337 | return; | |
1338 | ||
8a49542c PZ |
1339 | event->attach_state |= PERF_ATTACH_GROUP; |
1340 | ||
1341 | if (group_leader == event) | |
1342 | return; | |
1343 | ||
652884fe PZ |
1344 | WARN_ON_ONCE(group_leader->ctx != event->ctx); |
1345 | ||
8a49542c PZ |
1346 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
1347 | !is_software_event(event)) | |
1348 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1349 | ||
1350 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1351 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1352 | |
1353 | perf_event__header_size(group_leader); | |
1354 | ||
1355 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1356 | perf_event__header_size(pos); | |
8a49542c PZ |
1357 | } |
1358 | ||
a63eaf34 | 1359 | /* |
cdd6c482 | 1360 | * Remove a event from the lists for its context. |
fccc714b | 1361 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1362 | */ |
04289bb9 | 1363 | static void |
cdd6c482 | 1364 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1365 | { |
68cacd29 | 1366 | struct perf_cpu_context *cpuctx; |
652884fe PZ |
1367 | |
1368 | WARN_ON_ONCE(event->ctx != ctx); | |
1369 | lockdep_assert_held(&ctx->lock); | |
1370 | ||
8a49542c PZ |
1371 | /* |
1372 | * We can have double detach due to exit/hot-unplug + close. | |
1373 | */ | |
1374 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1375 | return; |
8a49542c PZ |
1376 | |
1377 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1378 | ||
68cacd29 | 1379 | if (is_cgroup_event(event)) { |
e5d1367f | 1380 | ctx->nr_cgroups--; |
68cacd29 SE |
1381 | cpuctx = __get_cpu_context(ctx); |
1382 | /* | |
1383 | * if there are no more cgroup events | |
1384 | * then cler cgrp to avoid stale pointer | |
1385 | * in update_cgrp_time_from_cpuctx() | |
1386 | */ | |
1387 | if (!ctx->nr_cgroups) | |
1388 | cpuctx->cgrp = NULL; | |
1389 | } | |
e5d1367f | 1390 | |
cdd6c482 IM |
1391 | ctx->nr_events--; |
1392 | if (event->attr.inherit_stat) | |
bfbd3381 | 1393 | ctx->nr_stat--; |
8bc20959 | 1394 | |
cdd6c482 | 1395 | list_del_rcu(&event->event_entry); |
04289bb9 | 1396 | |
8a49542c PZ |
1397 | if (event->group_leader == event) |
1398 | list_del_init(&event->group_entry); | |
5c148194 | 1399 | |
96c21a46 | 1400 | update_group_times(event); |
b2e74a26 SE |
1401 | |
1402 | /* | |
1403 | * If event was in error state, then keep it | |
1404 | * that way, otherwise bogus counts will be | |
1405 | * returned on read(). The only way to get out | |
1406 | * of error state is by explicit re-enabling | |
1407 | * of the event | |
1408 | */ | |
1409 | if (event->state > PERF_EVENT_STATE_OFF) | |
1410 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1411 | |
1412 | ctx->generation++; | |
050735b0 PZ |
1413 | } |
1414 | ||
8a49542c | 1415 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1416 | { |
1417 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1418 | struct list_head *list = NULL; |
1419 | ||
1420 | /* | |
1421 | * We can have double detach due to exit/hot-unplug + close. | |
1422 | */ | |
1423 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1424 | return; | |
1425 | ||
1426 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1427 | ||
1428 | /* | |
1429 | * If this is a sibling, remove it from its group. | |
1430 | */ | |
1431 | if (event->group_leader != event) { | |
1432 | list_del_init(&event->group_entry); | |
1433 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1434 | goto out; |
8a49542c PZ |
1435 | } |
1436 | ||
1437 | if (!list_empty(&event->group_entry)) | |
1438 | list = &event->group_entry; | |
2e2af50b | 1439 | |
04289bb9 | 1440 | /* |
cdd6c482 IM |
1441 | * If this was a group event with sibling events then |
1442 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1443 | * to whatever list we are on. |
04289bb9 | 1444 | */ |
cdd6c482 | 1445 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1446 | if (list) |
1447 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1448 | sibling->group_leader = sibling; |
d6f962b5 FW |
1449 | |
1450 | /* Inherit group flags from the previous leader */ | |
1451 | sibling->group_flags = event->group_flags; | |
652884fe PZ |
1452 | |
1453 | WARN_ON_ONCE(sibling->ctx != event->ctx); | |
04289bb9 | 1454 | } |
c320c7b7 ACM |
1455 | |
1456 | out: | |
1457 | perf_event__header_size(event->group_leader); | |
1458 | ||
1459 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1460 | perf_event__header_size(tmp); | |
04289bb9 IM |
1461 | } |
1462 | ||
fadfe7be JO |
1463 | /* |
1464 | * User event without the task. | |
1465 | */ | |
1466 | static bool is_orphaned_event(struct perf_event *event) | |
1467 | { | |
1468 | return event && !is_kernel_event(event) && !event->owner; | |
1469 | } | |
1470 | ||
1471 | /* | |
1472 | * Event has a parent but parent's task finished and it's | |
1473 | * alive only because of children holding refference. | |
1474 | */ | |
1475 | static bool is_orphaned_child(struct perf_event *event) | |
1476 | { | |
1477 | return is_orphaned_event(event->parent); | |
1478 | } | |
1479 | ||
1480 | static void orphans_remove_work(struct work_struct *work); | |
1481 | ||
1482 | static void schedule_orphans_remove(struct perf_event_context *ctx) | |
1483 | { | |
1484 | if (!ctx->task || ctx->orphans_remove_sched || !perf_wq) | |
1485 | return; | |
1486 | ||
1487 | if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) { | |
1488 | get_ctx(ctx); | |
1489 | ctx->orphans_remove_sched = true; | |
1490 | } | |
1491 | } | |
1492 | ||
1493 | static int __init perf_workqueue_init(void) | |
1494 | { | |
1495 | perf_wq = create_singlethread_workqueue("perf"); | |
1496 | WARN(!perf_wq, "failed to create perf workqueue\n"); | |
1497 | return perf_wq ? 0 : -1; | |
1498 | } | |
1499 | ||
1500 | core_initcall(perf_workqueue_init); | |
1501 | ||
fa66f07a SE |
1502 | static inline int |
1503 | event_filter_match(struct perf_event *event) | |
1504 | { | |
e5d1367f SE |
1505 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1506 | && perf_cgroup_match(event); | |
fa66f07a SE |
1507 | } |
1508 | ||
9ffcfa6f SE |
1509 | static void |
1510 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1511 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1512 | struct perf_event_context *ctx) |
3b6f9e5c | 1513 | { |
4158755d | 1514 | u64 tstamp = perf_event_time(event); |
fa66f07a | 1515 | u64 delta; |
652884fe PZ |
1516 | |
1517 | WARN_ON_ONCE(event->ctx != ctx); | |
1518 | lockdep_assert_held(&ctx->lock); | |
1519 | ||
fa66f07a SE |
1520 | /* |
1521 | * An event which could not be activated because of | |
1522 | * filter mismatch still needs to have its timings | |
1523 | * maintained, otherwise bogus information is return | |
1524 | * via read() for time_enabled, time_running: | |
1525 | */ | |
1526 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1527 | && !event_filter_match(event)) { | |
e5d1367f | 1528 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1529 | event->tstamp_running += delta; |
4158755d | 1530 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1531 | } |
1532 | ||
cdd6c482 | 1533 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1534 | return; |
3b6f9e5c | 1535 | |
44377277 AS |
1536 | perf_pmu_disable(event->pmu); |
1537 | ||
cdd6c482 IM |
1538 | event->state = PERF_EVENT_STATE_INACTIVE; |
1539 | if (event->pending_disable) { | |
1540 | event->pending_disable = 0; | |
1541 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1542 | } |
4158755d | 1543 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1544 | event->pmu->del(event, 0); |
cdd6c482 | 1545 | event->oncpu = -1; |
3b6f9e5c | 1546 | |
cdd6c482 | 1547 | if (!is_software_event(event)) |
3b6f9e5c | 1548 | cpuctx->active_oncpu--; |
2fde4f94 MR |
1549 | if (!--ctx->nr_active) |
1550 | perf_event_ctx_deactivate(ctx); | |
0f5a2601 PZ |
1551 | if (event->attr.freq && event->attr.sample_freq) |
1552 | ctx->nr_freq--; | |
cdd6c482 | 1553 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1554 | cpuctx->exclusive = 0; |
44377277 | 1555 | |
fadfe7be JO |
1556 | if (is_orphaned_child(event)) |
1557 | schedule_orphans_remove(ctx); | |
1558 | ||
44377277 | 1559 | perf_pmu_enable(event->pmu); |
3b6f9e5c PM |
1560 | } |
1561 | ||
d859e29f | 1562 | static void |
cdd6c482 | 1563 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1564 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1565 | struct perf_event_context *ctx) |
d859e29f | 1566 | { |
cdd6c482 | 1567 | struct perf_event *event; |
fa66f07a | 1568 | int state = group_event->state; |
d859e29f | 1569 | |
cdd6c482 | 1570 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1571 | |
1572 | /* | |
1573 | * Schedule out siblings (if any): | |
1574 | */ | |
cdd6c482 IM |
1575 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1576 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1577 | |
fa66f07a | 1578 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1579 | cpuctx->exclusive = 0; |
1580 | } | |
1581 | ||
46ce0fe9 PZ |
1582 | struct remove_event { |
1583 | struct perf_event *event; | |
1584 | bool detach_group; | |
1585 | }; | |
1586 | ||
0793a61d | 1587 | /* |
cdd6c482 | 1588 | * Cross CPU call to remove a performance event |
0793a61d | 1589 | * |
cdd6c482 | 1590 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1591 | * remove it from the context list. |
1592 | */ | |
fe4b04fa | 1593 | static int __perf_remove_from_context(void *info) |
0793a61d | 1594 | { |
46ce0fe9 PZ |
1595 | struct remove_event *re = info; |
1596 | struct perf_event *event = re->event; | |
cdd6c482 | 1597 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1598 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1599 | |
e625cce1 | 1600 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1601 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1602 | if (re->detach_group) |
1603 | perf_group_detach(event); | |
cdd6c482 | 1604 | list_del_event(event, ctx); |
64ce3126 PZ |
1605 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1606 | ctx->is_active = 0; | |
1607 | cpuctx->task_ctx = NULL; | |
1608 | } | |
e625cce1 | 1609 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1610 | |
1611 | return 0; | |
0793a61d TG |
1612 | } |
1613 | ||
1614 | ||
1615 | /* | |
cdd6c482 | 1616 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1617 | * |
cdd6c482 | 1618 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1619 | * call when the task is on a CPU. |
c93f7669 | 1620 | * |
cdd6c482 IM |
1621 | * If event->ctx is a cloned context, callers must make sure that |
1622 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1623 | * remains valid. This is OK when called from perf_release since |
1624 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1625 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1626 | * context has been detached from its task. |
0793a61d | 1627 | */ |
46ce0fe9 | 1628 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1629 | { |
cdd6c482 | 1630 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1631 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1632 | struct remove_event re = { |
1633 | .event = event, | |
1634 | .detach_group = detach_group, | |
1635 | }; | |
0793a61d | 1636 | |
fe4b04fa PZ |
1637 | lockdep_assert_held(&ctx->mutex); |
1638 | ||
0793a61d TG |
1639 | if (!task) { |
1640 | /* | |
226424ee MR |
1641 | * Per cpu events are removed via an smp call. The removal can |
1642 | * fail if the CPU is currently offline, but in that case we | |
1643 | * already called __perf_remove_from_context from | |
1644 | * perf_event_exit_cpu. | |
0793a61d | 1645 | */ |
46ce0fe9 | 1646 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1647 | return; |
1648 | } | |
1649 | ||
1650 | retry: | |
46ce0fe9 | 1651 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1652 | return; |
0793a61d | 1653 | |
e625cce1 | 1654 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1655 | /* |
fe4b04fa PZ |
1656 | * If we failed to find a running task, but find the context active now |
1657 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1658 | */ |
fe4b04fa | 1659 | if (ctx->is_active) { |
e625cce1 | 1660 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
1661 | /* |
1662 | * Reload the task pointer, it might have been changed by | |
1663 | * a concurrent perf_event_context_sched_out(). | |
1664 | */ | |
1665 | task = ctx->task; | |
0793a61d TG |
1666 | goto retry; |
1667 | } | |
1668 | ||
1669 | /* | |
fe4b04fa PZ |
1670 | * Since the task isn't running, its safe to remove the event, us |
1671 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1672 | */ |
46ce0fe9 PZ |
1673 | if (detach_group) |
1674 | perf_group_detach(event); | |
fe4b04fa | 1675 | list_del_event(event, ctx); |
e625cce1 | 1676 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1677 | } |
1678 | ||
d859e29f | 1679 | /* |
cdd6c482 | 1680 | * Cross CPU call to disable a performance event |
d859e29f | 1681 | */ |
500ad2d8 | 1682 | int __perf_event_disable(void *info) |
d859e29f | 1683 | { |
cdd6c482 | 1684 | struct perf_event *event = info; |
cdd6c482 | 1685 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1686 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1687 | |
1688 | /* | |
cdd6c482 IM |
1689 | * If this is a per-task event, need to check whether this |
1690 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1691 | * |
1692 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1693 | * flipping contexts around. | |
d859e29f | 1694 | */ |
665c2142 | 1695 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1696 | return -EINVAL; |
d859e29f | 1697 | |
e625cce1 | 1698 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1699 | |
1700 | /* | |
cdd6c482 | 1701 | * If the event is on, turn it off. |
d859e29f PM |
1702 | * If it is in error state, leave it in error state. |
1703 | */ | |
cdd6c482 | 1704 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1705 | update_context_time(ctx); |
e5d1367f | 1706 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1707 | update_group_times(event); |
1708 | if (event == event->group_leader) | |
1709 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1710 | else |
cdd6c482 IM |
1711 | event_sched_out(event, cpuctx, ctx); |
1712 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1713 | } |
1714 | ||
e625cce1 | 1715 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1716 | |
1717 | return 0; | |
d859e29f PM |
1718 | } |
1719 | ||
1720 | /* | |
cdd6c482 | 1721 | * Disable a event. |
c93f7669 | 1722 | * |
cdd6c482 IM |
1723 | * If event->ctx is a cloned context, callers must make sure that |
1724 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1725 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1726 | * perf_event_for_each_child or perf_event_for_each because they |
1727 | * hold the top-level event's child_mutex, so any descendant that | |
1728 | * goes to exit will block in sync_child_event. | |
1729 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1730 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1731 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1732 | */ |
f63a8daa | 1733 | static void _perf_event_disable(struct perf_event *event) |
d859e29f | 1734 | { |
cdd6c482 | 1735 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1736 | struct task_struct *task = ctx->task; |
1737 | ||
1738 | if (!task) { | |
1739 | /* | |
cdd6c482 | 1740 | * Disable the event on the cpu that it's on |
d859e29f | 1741 | */ |
fe4b04fa | 1742 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1743 | return; |
1744 | } | |
1745 | ||
9ed6060d | 1746 | retry: |
fe4b04fa PZ |
1747 | if (!task_function_call(task, __perf_event_disable, event)) |
1748 | return; | |
d859e29f | 1749 | |
e625cce1 | 1750 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1751 | /* |
cdd6c482 | 1752 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1753 | */ |
cdd6c482 | 1754 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1755 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1756 | /* |
1757 | * Reload the task pointer, it might have been changed by | |
1758 | * a concurrent perf_event_context_sched_out(). | |
1759 | */ | |
1760 | task = ctx->task; | |
d859e29f PM |
1761 | goto retry; |
1762 | } | |
1763 | ||
1764 | /* | |
1765 | * Since we have the lock this context can't be scheduled | |
1766 | * in, so we can change the state safely. | |
1767 | */ | |
cdd6c482 IM |
1768 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1769 | update_group_times(event); | |
1770 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1771 | } |
e625cce1 | 1772 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1773 | } |
f63a8daa PZ |
1774 | |
1775 | /* | |
1776 | * Strictly speaking kernel users cannot create groups and therefore this | |
1777 | * interface does not need the perf_event_ctx_lock() magic. | |
1778 | */ | |
1779 | void perf_event_disable(struct perf_event *event) | |
1780 | { | |
1781 | struct perf_event_context *ctx; | |
1782 | ||
1783 | ctx = perf_event_ctx_lock(event); | |
1784 | _perf_event_disable(event); | |
1785 | perf_event_ctx_unlock(event, ctx); | |
1786 | } | |
dcfce4a0 | 1787 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1788 | |
e5d1367f SE |
1789 | static void perf_set_shadow_time(struct perf_event *event, |
1790 | struct perf_event_context *ctx, | |
1791 | u64 tstamp) | |
1792 | { | |
1793 | /* | |
1794 | * use the correct time source for the time snapshot | |
1795 | * | |
1796 | * We could get by without this by leveraging the | |
1797 | * fact that to get to this function, the caller | |
1798 | * has most likely already called update_context_time() | |
1799 | * and update_cgrp_time_xx() and thus both timestamp | |
1800 | * are identical (or very close). Given that tstamp is, | |
1801 | * already adjusted for cgroup, we could say that: | |
1802 | * tstamp - ctx->timestamp | |
1803 | * is equivalent to | |
1804 | * tstamp - cgrp->timestamp. | |
1805 | * | |
1806 | * Then, in perf_output_read(), the calculation would | |
1807 | * work with no changes because: | |
1808 | * - event is guaranteed scheduled in | |
1809 | * - no scheduled out in between | |
1810 | * - thus the timestamp would be the same | |
1811 | * | |
1812 | * But this is a bit hairy. | |
1813 | * | |
1814 | * So instead, we have an explicit cgroup call to remain | |
1815 | * within the time time source all along. We believe it | |
1816 | * is cleaner and simpler to understand. | |
1817 | */ | |
1818 | if (is_cgroup_event(event)) | |
1819 | perf_cgroup_set_shadow_time(event, tstamp); | |
1820 | else | |
1821 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1822 | } | |
1823 | ||
4fe757dd PZ |
1824 | #define MAX_INTERRUPTS (~0ULL) |
1825 | ||
1826 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1827 | ||
235c7fc7 | 1828 | static int |
9ffcfa6f | 1829 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1830 | struct perf_cpu_context *cpuctx, |
6e37738a | 1831 | struct perf_event_context *ctx) |
235c7fc7 | 1832 | { |
4158755d | 1833 | u64 tstamp = perf_event_time(event); |
44377277 | 1834 | int ret = 0; |
4158755d | 1835 | |
63342411 PZ |
1836 | lockdep_assert_held(&ctx->lock); |
1837 | ||
cdd6c482 | 1838 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1839 | return 0; |
1840 | ||
cdd6c482 | 1841 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1842 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1843 | |
1844 | /* | |
1845 | * Unthrottle events, since we scheduled we might have missed several | |
1846 | * ticks already, also for a heavily scheduling task there is little | |
1847 | * guarantee it'll get a tick in a timely manner. | |
1848 | */ | |
1849 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1850 | perf_log_throttle(event, 1); | |
1851 | event->hw.interrupts = 0; | |
1852 | } | |
1853 | ||
235c7fc7 IM |
1854 | /* |
1855 | * The new state must be visible before we turn it on in the hardware: | |
1856 | */ | |
1857 | smp_wmb(); | |
1858 | ||
44377277 AS |
1859 | perf_pmu_disable(event->pmu); |
1860 | ||
72f669c0 SL |
1861 | event->tstamp_running += tstamp - event->tstamp_stopped; |
1862 | ||
1863 | perf_set_shadow_time(event, ctx, tstamp); | |
1864 | ||
a4eaf7f1 | 1865 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1866 | event->state = PERF_EVENT_STATE_INACTIVE; |
1867 | event->oncpu = -1; | |
44377277 AS |
1868 | ret = -EAGAIN; |
1869 | goto out; | |
235c7fc7 IM |
1870 | } |
1871 | ||
cdd6c482 | 1872 | if (!is_software_event(event)) |
3b6f9e5c | 1873 | cpuctx->active_oncpu++; |
2fde4f94 MR |
1874 | if (!ctx->nr_active++) |
1875 | perf_event_ctx_activate(ctx); | |
0f5a2601 PZ |
1876 | if (event->attr.freq && event->attr.sample_freq) |
1877 | ctx->nr_freq++; | |
235c7fc7 | 1878 | |
cdd6c482 | 1879 | if (event->attr.exclusive) |
3b6f9e5c PM |
1880 | cpuctx->exclusive = 1; |
1881 | ||
fadfe7be JO |
1882 | if (is_orphaned_child(event)) |
1883 | schedule_orphans_remove(ctx); | |
1884 | ||
44377277 AS |
1885 | out: |
1886 | perf_pmu_enable(event->pmu); | |
1887 | ||
1888 | return ret; | |
235c7fc7 IM |
1889 | } |
1890 | ||
6751b71e | 1891 | static int |
cdd6c482 | 1892 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1893 | struct perf_cpu_context *cpuctx, |
6e37738a | 1894 | struct perf_event_context *ctx) |
6751b71e | 1895 | { |
6bde9b6c | 1896 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1897 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1898 | u64 now = ctx->time; |
1899 | bool simulate = false; | |
6751b71e | 1900 | |
cdd6c482 | 1901 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1902 | return 0; |
1903 | ||
ad5133b7 | 1904 | pmu->start_txn(pmu); |
6bde9b6c | 1905 | |
9ffcfa6f | 1906 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1907 | pmu->cancel_txn(pmu); |
9e630205 | 1908 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1909 | return -EAGAIN; |
90151c35 | 1910 | } |
6751b71e PM |
1911 | |
1912 | /* | |
1913 | * Schedule in siblings as one group (if any): | |
1914 | */ | |
cdd6c482 | 1915 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1916 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1917 | partial_group = event; |
6751b71e PM |
1918 | goto group_error; |
1919 | } | |
1920 | } | |
1921 | ||
9ffcfa6f | 1922 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1923 | return 0; |
9ffcfa6f | 1924 | |
6751b71e PM |
1925 | group_error: |
1926 | /* | |
1927 | * Groups can be scheduled in as one unit only, so undo any | |
1928 | * partial group before returning: | |
d7842da4 SE |
1929 | * The events up to the failed event are scheduled out normally, |
1930 | * tstamp_stopped will be updated. | |
1931 | * | |
1932 | * The failed events and the remaining siblings need to have | |
1933 | * their timings updated as if they had gone thru event_sched_in() | |
1934 | * and event_sched_out(). This is required to get consistent timings | |
1935 | * across the group. This also takes care of the case where the group | |
1936 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1937 | * the time the event was actually stopped, such that time delta | |
1938 | * calculation in update_event_times() is correct. | |
6751b71e | 1939 | */ |
cdd6c482 IM |
1940 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1941 | if (event == partial_group) | |
d7842da4 SE |
1942 | simulate = true; |
1943 | ||
1944 | if (simulate) { | |
1945 | event->tstamp_running += now - event->tstamp_stopped; | |
1946 | event->tstamp_stopped = now; | |
1947 | } else { | |
1948 | event_sched_out(event, cpuctx, ctx); | |
1949 | } | |
6751b71e | 1950 | } |
9ffcfa6f | 1951 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1952 | |
ad5133b7 | 1953 | pmu->cancel_txn(pmu); |
90151c35 | 1954 | |
9e630205 SE |
1955 | perf_cpu_hrtimer_restart(cpuctx); |
1956 | ||
6751b71e PM |
1957 | return -EAGAIN; |
1958 | } | |
1959 | ||
3b6f9e5c | 1960 | /* |
cdd6c482 | 1961 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1962 | */ |
cdd6c482 | 1963 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1964 | struct perf_cpu_context *cpuctx, |
1965 | int can_add_hw) | |
1966 | { | |
1967 | /* | |
cdd6c482 | 1968 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1969 | */ |
d6f962b5 | 1970 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1971 | return 1; |
1972 | /* | |
1973 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1974 | * events can go on. |
3b6f9e5c PM |
1975 | */ |
1976 | if (cpuctx->exclusive) | |
1977 | return 0; | |
1978 | /* | |
1979 | * If this group is exclusive and there are already | |
cdd6c482 | 1980 | * events on the CPU, it can't go on. |
3b6f9e5c | 1981 | */ |
cdd6c482 | 1982 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1983 | return 0; |
1984 | /* | |
1985 | * Otherwise, try to add it if all previous groups were able | |
1986 | * to go on. | |
1987 | */ | |
1988 | return can_add_hw; | |
1989 | } | |
1990 | ||
cdd6c482 IM |
1991 | static void add_event_to_ctx(struct perf_event *event, |
1992 | struct perf_event_context *ctx) | |
53cfbf59 | 1993 | { |
4158755d SE |
1994 | u64 tstamp = perf_event_time(event); |
1995 | ||
cdd6c482 | 1996 | list_add_event(event, ctx); |
8a49542c | 1997 | perf_group_attach(event); |
4158755d SE |
1998 | event->tstamp_enabled = tstamp; |
1999 | event->tstamp_running = tstamp; | |
2000 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
2001 | } |
2002 | ||
2c29ef0f PZ |
2003 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
2004 | static void | |
2005 | ctx_sched_in(struct perf_event_context *ctx, | |
2006 | struct perf_cpu_context *cpuctx, | |
2007 | enum event_type_t event_type, | |
2008 | struct task_struct *task); | |
fe4b04fa | 2009 | |
dce5855b PZ |
2010 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
2011 | struct perf_event_context *ctx, | |
2012 | struct task_struct *task) | |
2013 | { | |
2014 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
2015 | if (ctx) | |
2016 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
2017 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
2018 | if (ctx) | |
2019 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
2020 | } | |
2021 | ||
0793a61d | 2022 | /* |
cdd6c482 | 2023 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
2024 | * |
2025 | * Must be called with ctx->mutex held | |
0793a61d | 2026 | */ |
fe4b04fa | 2027 | static int __perf_install_in_context(void *info) |
0793a61d | 2028 | { |
cdd6c482 IM |
2029 | struct perf_event *event = info; |
2030 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2031 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
2032 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
2033 | struct task_struct *task = current; | |
2034 | ||
b58f6b0d | 2035 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 2036 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
2037 | |
2038 | /* | |
2c29ef0f | 2039 | * If there was an active task_ctx schedule it out. |
0793a61d | 2040 | */ |
b58f6b0d | 2041 | if (task_ctx) |
2c29ef0f | 2042 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
2043 | |
2044 | /* | |
2045 | * If the context we're installing events in is not the | |
2046 | * active task_ctx, flip them. | |
2047 | */ | |
2048 | if (ctx->task && task_ctx != ctx) { | |
2049 | if (task_ctx) | |
2050 | raw_spin_unlock(&task_ctx->lock); | |
2051 | raw_spin_lock(&ctx->lock); | |
2052 | task_ctx = ctx; | |
2053 | } | |
2054 | ||
2055 | if (task_ctx) { | |
2056 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
2057 | task = task_ctx->task; |
2058 | } | |
b58f6b0d | 2059 | |
2c29ef0f | 2060 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 2061 | |
4af4998b | 2062 | update_context_time(ctx); |
e5d1367f SE |
2063 | /* |
2064 | * update cgrp time only if current cgrp | |
2065 | * matches event->cgrp. Must be done before | |
2066 | * calling add_event_to_ctx() | |
2067 | */ | |
2068 | update_cgrp_time_from_event(event); | |
0793a61d | 2069 | |
cdd6c482 | 2070 | add_event_to_ctx(event, ctx); |
0793a61d | 2071 | |
d859e29f | 2072 | /* |
2c29ef0f | 2073 | * Schedule everything back in |
d859e29f | 2074 | */ |
dce5855b | 2075 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
2076 | |
2077 | perf_pmu_enable(cpuctx->ctx.pmu); | |
2078 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
2079 | |
2080 | return 0; | |
0793a61d TG |
2081 | } |
2082 | ||
2083 | /* | |
cdd6c482 | 2084 | * Attach a performance event to a context |
0793a61d | 2085 | * |
cdd6c482 IM |
2086 | * First we add the event to the list with the hardware enable bit |
2087 | * in event->hw_config cleared. | |
0793a61d | 2088 | * |
cdd6c482 | 2089 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
2090 | * call to enable it in the task context. The task might have been |
2091 | * scheduled away, but we check this in the smp call again. | |
2092 | */ | |
2093 | static void | |
cdd6c482 IM |
2094 | perf_install_in_context(struct perf_event_context *ctx, |
2095 | struct perf_event *event, | |
0793a61d TG |
2096 | int cpu) |
2097 | { | |
2098 | struct task_struct *task = ctx->task; | |
2099 | ||
fe4b04fa PZ |
2100 | lockdep_assert_held(&ctx->mutex); |
2101 | ||
c3f00c70 | 2102 | event->ctx = ctx; |
0cda4c02 YZ |
2103 | if (event->cpu != -1) |
2104 | event->cpu = cpu; | |
c3f00c70 | 2105 | |
0793a61d TG |
2106 | if (!task) { |
2107 | /* | |
cdd6c482 | 2108 | * Per cpu events are installed via an smp call and |
af901ca1 | 2109 | * the install is always successful. |
0793a61d | 2110 | */ |
fe4b04fa | 2111 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
2112 | return; |
2113 | } | |
2114 | ||
0793a61d | 2115 | retry: |
fe4b04fa PZ |
2116 | if (!task_function_call(task, __perf_install_in_context, event)) |
2117 | return; | |
0793a61d | 2118 | |
e625cce1 | 2119 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 2120 | /* |
fe4b04fa PZ |
2121 | * If we failed to find a running task, but find the context active now |
2122 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 2123 | */ |
fe4b04fa | 2124 | if (ctx->is_active) { |
e625cce1 | 2125 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
2126 | /* |
2127 | * Reload the task pointer, it might have been changed by | |
2128 | * a concurrent perf_event_context_sched_out(). | |
2129 | */ | |
2130 | task = ctx->task; | |
0793a61d TG |
2131 | goto retry; |
2132 | } | |
2133 | ||
2134 | /* | |
fe4b04fa PZ |
2135 | * Since the task isn't running, its safe to add the event, us holding |
2136 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 2137 | */ |
fe4b04fa | 2138 | add_event_to_ctx(event, ctx); |
e625cce1 | 2139 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
2140 | } |
2141 | ||
fa289bec | 2142 | /* |
cdd6c482 | 2143 | * Put a event into inactive state and update time fields. |
fa289bec PM |
2144 | * Enabling the leader of a group effectively enables all |
2145 | * the group members that aren't explicitly disabled, so we | |
2146 | * have to update their ->tstamp_enabled also. | |
2147 | * Note: this works for group members as well as group leaders | |
2148 | * since the non-leader members' sibling_lists will be empty. | |
2149 | */ | |
1d9b482e | 2150 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 2151 | { |
cdd6c482 | 2152 | struct perf_event *sub; |
4158755d | 2153 | u64 tstamp = perf_event_time(event); |
fa289bec | 2154 | |
cdd6c482 | 2155 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 2156 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 2157 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
2158 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
2159 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 2160 | } |
fa289bec PM |
2161 | } |
2162 | ||
d859e29f | 2163 | /* |
cdd6c482 | 2164 | * Cross CPU call to enable a performance event |
d859e29f | 2165 | */ |
fe4b04fa | 2166 | static int __perf_event_enable(void *info) |
04289bb9 | 2167 | { |
cdd6c482 | 2168 | struct perf_event *event = info; |
cdd6c482 IM |
2169 | struct perf_event_context *ctx = event->ctx; |
2170 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2171 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2172 | int err; |
04289bb9 | 2173 | |
06f41796 JO |
2174 | /* |
2175 | * There's a time window between 'ctx->is_active' check | |
2176 | * in perf_event_enable function and this place having: | |
2177 | * - IRQs on | |
2178 | * - ctx->lock unlocked | |
2179 | * | |
2180 | * where the task could be killed and 'ctx' deactivated | |
2181 | * by perf_event_exit_task. | |
2182 | */ | |
2183 | if (!ctx->is_active) | |
fe4b04fa | 2184 | return -EINVAL; |
3cbed429 | 2185 | |
e625cce1 | 2186 | raw_spin_lock(&ctx->lock); |
4af4998b | 2187 | update_context_time(ctx); |
d859e29f | 2188 | |
cdd6c482 | 2189 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2190 | goto unlock; |
e5d1367f SE |
2191 | |
2192 | /* | |
2193 | * set current task's cgroup time reference point | |
2194 | */ | |
3f7cce3c | 2195 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2196 | |
1d9b482e | 2197 | __perf_event_mark_enabled(event); |
04289bb9 | 2198 | |
e5d1367f SE |
2199 | if (!event_filter_match(event)) { |
2200 | if (is_cgroup_event(event)) | |
2201 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2202 | goto unlock; |
e5d1367f | 2203 | } |
f4c4176f | 2204 | |
04289bb9 | 2205 | /* |
cdd6c482 | 2206 | * If the event is in a group and isn't the group leader, |
d859e29f | 2207 | * then don't put it on unless the group is on. |
04289bb9 | 2208 | */ |
cdd6c482 | 2209 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2210 | goto unlock; |
3b6f9e5c | 2211 | |
cdd6c482 | 2212 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2213 | err = -EEXIST; |
e758a33d | 2214 | } else { |
cdd6c482 | 2215 | if (event == leader) |
6e37738a | 2216 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2217 | else |
6e37738a | 2218 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2219 | } |
d859e29f PM |
2220 | |
2221 | if (err) { | |
2222 | /* | |
cdd6c482 | 2223 | * If this event can't go on and it's part of a |
d859e29f PM |
2224 | * group, then the whole group has to come off. |
2225 | */ | |
9e630205 | 2226 | if (leader != event) { |
d859e29f | 2227 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2228 | perf_cpu_hrtimer_restart(cpuctx); |
2229 | } | |
0d48696f | 2230 | if (leader->attr.pinned) { |
53cfbf59 | 2231 | update_group_times(leader); |
cdd6c482 | 2232 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2233 | } |
d859e29f PM |
2234 | } |
2235 | ||
9ed6060d | 2236 | unlock: |
e625cce1 | 2237 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2238 | |
2239 | return 0; | |
d859e29f PM |
2240 | } |
2241 | ||
2242 | /* | |
cdd6c482 | 2243 | * Enable a event. |
c93f7669 | 2244 | * |
cdd6c482 IM |
2245 | * If event->ctx is a cloned context, callers must make sure that |
2246 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2247 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2248 | * perf_event_for_each_child or perf_event_for_each as described |
2249 | * for perf_event_disable. | |
d859e29f | 2250 | */ |
f63a8daa | 2251 | static void _perf_event_enable(struct perf_event *event) |
d859e29f | 2252 | { |
cdd6c482 | 2253 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2254 | struct task_struct *task = ctx->task; |
2255 | ||
2256 | if (!task) { | |
2257 | /* | |
cdd6c482 | 2258 | * Enable the event on the cpu that it's on |
d859e29f | 2259 | */ |
fe4b04fa | 2260 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2261 | return; |
2262 | } | |
2263 | ||
e625cce1 | 2264 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2265 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2266 | goto out; |
2267 | ||
2268 | /* | |
cdd6c482 IM |
2269 | * If the event is in error state, clear that first. |
2270 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2271 | * know that it has gone back into error state, as distinct |
2272 | * from the task having been scheduled away before the | |
2273 | * cross-call arrived. | |
2274 | */ | |
cdd6c482 IM |
2275 | if (event->state == PERF_EVENT_STATE_ERROR) |
2276 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2277 | |
9ed6060d | 2278 | retry: |
fe4b04fa | 2279 | if (!ctx->is_active) { |
1d9b482e | 2280 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2281 | goto out; |
2282 | } | |
2283 | ||
e625cce1 | 2284 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2285 | |
2286 | if (!task_function_call(task, __perf_event_enable, event)) | |
2287 | return; | |
d859e29f | 2288 | |
e625cce1 | 2289 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2290 | |
2291 | /* | |
cdd6c482 | 2292 | * If the context is active and the event is still off, |
d859e29f PM |
2293 | * we need to retry the cross-call. |
2294 | */ | |
fe4b04fa PZ |
2295 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2296 | /* | |
2297 | * task could have been flipped by a concurrent | |
2298 | * perf_event_context_sched_out() | |
2299 | */ | |
2300 | task = ctx->task; | |
d859e29f | 2301 | goto retry; |
fe4b04fa | 2302 | } |
fa289bec | 2303 | |
9ed6060d | 2304 | out: |
e625cce1 | 2305 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2306 | } |
f63a8daa PZ |
2307 | |
2308 | /* | |
2309 | * See perf_event_disable(); | |
2310 | */ | |
2311 | void perf_event_enable(struct perf_event *event) | |
2312 | { | |
2313 | struct perf_event_context *ctx; | |
2314 | ||
2315 | ctx = perf_event_ctx_lock(event); | |
2316 | _perf_event_enable(event); | |
2317 | perf_event_ctx_unlock(event, ctx); | |
2318 | } | |
dcfce4a0 | 2319 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2320 | |
f63a8daa | 2321 | static int _perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2322 | { |
2023b359 | 2323 | /* |
cdd6c482 | 2324 | * not supported on inherited events |
2023b359 | 2325 | */ |
2e939d1d | 2326 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2327 | return -EINVAL; |
2328 | ||
cdd6c482 | 2329 | atomic_add(refresh, &event->event_limit); |
f63a8daa | 2330 | _perf_event_enable(event); |
2023b359 PZ |
2331 | |
2332 | return 0; | |
79f14641 | 2333 | } |
f63a8daa PZ |
2334 | |
2335 | /* | |
2336 | * See perf_event_disable() | |
2337 | */ | |
2338 | int perf_event_refresh(struct perf_event *event, int refresh) | |
2339 | { | |
2340 | struct perf_event_context *ctx; | |
2341 | int ret; | |
2342 | ||
2343 | ctx = perf_event_ctx_lock(event); | |
2344 | ret = _perf_event_refresh(event, refresh); | |
2345 | perf_event_ctx_unlock(event, ctx); | |
2346 | ||
2347 | return ret; | |
2348 | } | |
26ca5c11 | 2349 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2350 | |
5b0311e1 FW |
2351 | static void ctx_sched_out(struct perf_event_context *ctx, |
2352 | struct perf_cpu_context *cpuctx, | |
2353 | enum event_type_t event_type) | |
235c7fc7 | 2354 | { |
cdd6c482 | 2355 | struct perf_event *event; |
db24d33e | 2356 | int is_active = ctx->is_active; |
235c7fc7 | 2357 | |
db24d33e | 2358 | ctx->is_active &= ~event_type; |
cdd6c482 | 2359 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2360 | return; |
2361 | ||
4af4998b | 2362 | update_context_time(ctx); |
e5d1367f | 2363 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2364 | if (!ctx->nr_active) |
facc4307 | 2365 | return; |
5b0311e1 | 2366 | |
075e0b00 | 2367 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2368 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2369 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2370 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2371 | } |
889ff015 | 2372 | |
db24d33e | 2373 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2374 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2375 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2376 | } |
1b9a644f | 2377 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2378 | } |
2379 | ||
564c2b21 | 2380 | /* |
5a3126d4 PZ |
2381 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2382 | * cloned from the same version of the same context. | |
2383 | * | |
2384 | * Equivalence is measured using a generation number in the context that is | |
2385 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2386 | * and list_del_event(). | |
564c2b21 | 2387 | */ |
cdd6c482 IM |
2388 | static int context_equiv(struct perf_event_context *ctx1, |
2389 | struct perf_event_context *ctx2) | |
564c2b21 | 2390 | { |
211de6eb PZ |
2391 | lockdep_assert_held(&ctx1->lock); |
2392 | lockdep_assert_held(&ctx2->lock); | |
2393 | ||
5a3126d4 PZ |
2394 | /* Pinning disables the swap optimization */ |
2395 | if (ctx1->pin_count || ctx2->pin_count) | |
2396 | return 0; | |
2397 | ||
2398 | /* If ctx1 is the parent of ctx2 */ | |
2399 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2400 | return 1; | |
2401 | ||
2402 | /* If ctx2 is the parent of ctx1 */ | |
2403 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2404 | return 1; | |
2405 | ||
2406 | /* | |
2407 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2408 | * hierarchy, see perf_event_init_context(). | |
2409 | */ | |
2410 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2411 | ctx1->parent_gen == ctx2->parent_gen) | |
2412 | return 1; | |
2413 | ||
2414 | /* Unmatched */ | |
2415 | return 0; | |
564c2b21 PM |
2416 | } |
2417 | ||
cdd6c482 IM |
2418 | static void __perf_event_sync_stat(struct perf_event *event, |
2419 | struct perf_event *next_event) | |
bfbd3381 PZ |
2420 | { |
2421 | u64 value; | |
2422 | ||
cdd6c482 | 2423 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2424 | return; |
2425 | ||
2426 | /* | |
cdd6c482 | 2427 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2428 | * because we're in the middle of a context switch and have IRQs |
2429 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2430 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2431 | * don't need to use it. |
2432 | */ | |
cdd6c482 IM |
2433 | switch (event->state) { |
2434 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2435 | event->pmu->read(event); |
2436 | /* fall-through */ | |
bfbd3381 | 2437 | |
cdd6c482 IM |
2438 | case PERF_EVENT_STATE_INACTIVE: |
2439 | update_event_times(event); | |
bfbd3381 PZ |
2440 | break; |
2441 | ||
2442 | default: | |
2443 | break; | |
2444 | } | |
2445 | ||
2446 | /* | |
cdd6c482 | 2447 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2448 | * values when we flip the contexts. |
2449 | */ | |
e7850595 PZ |
2450 | value = local64_read(&next_event->count); |
2451 | value = local64_xchg(&event->count, value); | |
2452 | local64_set(&next_event->count, value); | |
bfbd3381 | 2453 | |
cdd6c482 IM |
2454 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2455 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2456 | |
bfbd3381 | 2457 | /* |
19d2e755 | 2458 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2459 | */ |
cdd6c482 IM |
2460 | perf_event_update_userpage(event); |
2461 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2462 | } |
2463 | ||
cdd6c482 IM |
2464 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2465 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2466 | { |
cdd6c482 | 2467 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2468 | |
2469 | if (!ctx->nr_stat) | |
2470 | return; | |
2471 | ||
02ffdbc8 PZ |
2472 | update_context_time(ctx); |
2473 | ||
cdd6c482 IM |
2474 | event = list_first_entry(&ctx->event_list, |
2475 | struct perf_event, event_entry); | |
bfbd3381 | 2476 | |
cdd6c482 IM |
2477 | next_event = list_first_entry(&next_ctx->event_list, |
2478 | struct perf_event, event_entry); | |
bfbd3381 | 2479 | |
cdd6c482 IM |
2480 | while (&event->event_entry != &ctx->event_list && |
2481 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2482 | |
cdd6c482 | 2483 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2484 | |
cdd6c482 IM |
2485 | event = list_next_entry(event, event_entry); |
2486 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2487 | } |
2488 | } | |
2489 | ||
fe4b04fa PZ |
2490 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2491 | struct task_struct *next) | |
0793a61d | 2492 | { |
8dc85d54 | 2493 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2494 | struct perf_event_context *next_ctx; |
5a3126d4 | 2495 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2496 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2497 | int do_switch = 1; |
0793a61d | 2498 | |
108b02cf PZ |
2499 | if (likely(!ctx)) |
2500 | return; | |
10989fb2 | 2501 | |
108b02cf PZ |
2502 | cpuctx = __get_cpu_context(ctx); |
2503 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2504 | return; |
2505 | ||
c93f7669 | 2506 | rcu_read_lock(); |
8dc85d54 | 2507 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2508 | if (!next_ctx) |
2509 | goto unlock; | |
2510 | ||
2511 | parent = rcu_dereference(ctx->parent_ctx); | |
2512 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2513 | ||
2514 | /* If neither context have a parent context; they cannot be clones. */ | |
802c8a61 | 2515 | if (!parent && !next_parent) |
5a3126d4 PZ |
2516 | goto unlock; |
2517 | ||
2518 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2519 | /* |
2520 | * Looks like the two contexts are clones, so we might be | |
2521 | * able to optimize the context switch. We lock both | |
2522 | * contexts and check that they are clones under the | |
2523 | * lock (including re-checking that neither has been | |
2524 | * uncloned in the meantime). It doesn't matter which | |
2525 | * order we take the locks because no other cpu could | |
2526 | * be trying to lock both of these tasks. | |
2527 | */ | |
e625cce1 TG |
2528 | raw_spin_lock(&ctx->lock); |
2529 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2530 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2531 | /* |
2532 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2533 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2534 | */ |
8dc85d54 PZ |
2535 | task->perf_event_ctxp[ctxn] = next_ctx; |
2536 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2537 | ctx->task = next; |
2538 | next_ctx->task = task; | |
5a158c3c YZ |
2539 | |
2540 | swap(ctx->task_ctx_data, next_ctx->task_ctx_data); | |
2541 | ||
c93f7669 | 2542 | do_switch = 0; |
bfbd3381 | 2543 | |
cdd6c482 | 2544 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2545 | } |
e625cce1 TG |
2546 | raw_spin_unlock(&next_ctx->lock); |
2547 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2548 | } |
5a3126d4 | 2549 | unlock: |
c93f7669 | 2550 | rcu_read_unlock(); |
564c2b21 | 2551 | |
c93f7669 | 2552 | if (do_switch) { |
facc4307 | 2553 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2554 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2555 | cpuctx->task_ctx = NULL; |
facc4307 | 2556 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2557 | } |
0793a61d TG |
2558 | } |
2559 | ||
ba532500 YZ |
2560 | void perf_sched_cb_dec(struct pmu *pmu) |
2561 | { | |
2562 | this_cpu_dec(perf_sched_cb_usages); | |
2563 | } | |
2564 | ||
2565 | void perf_sched_cb_inc(struct pmu *pmu) | |
2566 | { | |
2567 | this_cpu_inc(perf_sched_cb_usages); | |
2568 | } | |
2569 | ||
2570 | /* | |
2571 | * This function provides the context switch callback to the lower code | |
2572 | * layer. It is invoked ONLY when the context switch callback is enabled. | |
2573 | */ | |
2574 | static void perf_pmu_sched_task(struct task_struct *prev, | |
2575 | struct task_struct *next, | |
2576 | bool sched_in) | |
2577 | { | |
2578 | struct perf_cpu_context *cpuctx; | |
2579 | struct pmu *pmu; | |
2580 | unsigned long flags; | |
2581 | ||
2582 | if (prev == next) | |
2583 | return; | |
2584 | ||
2585 | local_irq_save(flags); | |
2586 | ||
2587 | rcu_read_lock(); | |
2588 | ||
2589 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2590 | if (pmu->sched_task) { | |
2591 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2592 | ||
2593 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2594 | ||
2595 | perf_pmu_disable(pmu); | |
2596 | ||
2597 | pmu->sched_task(cpuctx->task_ctx, sched_in); | |
2598 | ||
2599 | perf_pmu_enable(pmu); | |
2600 | ||
2601 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2602 | } | |
2603 | } | |
2604 | ||
2605 | rcu_read_unlock(); | |
2606 | ||
2607 | local_irq_restore(flags); | |
2608 | } | |
2609 | ||
8dc85d54 PZ |
2610 | #define for_each_task_context_nr(ctxn) \ |
2611 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2612 | ||
2613 | /* | |
2614 | * Called from scheduler to remove the events of the current task, | |
2615 | * with interrupts disabled. | |
2616 | * | |
2617 | * We stop each event and update the event value in event->count. | |
2618 | * | |
2619 | * This does not protect us against NMI, but disable() | |
2620 | * sets the disabled bit in the control field of event _before_ | |
2621 | * accessing the event control register. If a NMI hits, then it will | |
2622 | * not restart the event. | |
2623 | */ | |
ab0cce56 JO |
2624 | void __perf_event_task_sched_out(struct task_struct *task, |
2625 | struct task_struct *next) | |
8dc85d54 PZ |
2626 | { |
2627 | int ctxn; | |
2628 | ||
ba532500 YZ |
2629 | if (__this_cpu_read(perf_sched_cb_usages)) |
2630 | perf_pmu_sched_task(task, next, false); | |
2631 | ||
8dc85d54 PZ |
2632 | for_each_task_context_nr(ctxn) |
2633 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2634 | |
2635 | /* | |
2636 | * if cgroup events exist on this CPU, then we need | |
2637 | * to check if we have to switch out PMU state. | |
2638 | * cgroup event are system-wide mode only | |
2639 | */ | |
4a32fea9 | 2640 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2641 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2642 | } |
2643 | ||
04dc2dbb | 2644 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2645 | { |
108b02cf | 2646 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2647 | |
a63eaf34 PM |
2648 | if (!cpuctx->task_ctx) |
2649 | return; | |
012b84da IM |
2650 | |
2651 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2652 | return; | |
2653 | ||
04dc2dbb | 2654 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2655 | cpuctx->task_ctx = NULL; |
2656 | } | |
2657 | ||
5b0311e1 FW |
2658 | /* |
2659 | * Called with IRQs disabled | |
2660 | */ | |
2661 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2662 | enum event_type_t event_type) | |
2663 | { | |
2664 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2665 | } |
2666 | ||
235c7fc7 | 2667 | static void |
5b0311e1 | 2668 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2669 | struct perf_cpu_context *cpuctx) |
0793a61d | 2670 | { |
cdd6c482 | 2671 | struct perf_event *event; |
0793a61d | 2672 | |
889ff015 FW |
2673 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2674 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2675 | continue; |
5632ab12 | 2676 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2677 | continue; |
2678 | ||
e5d1367f SE |
2679 | /* may need to reset tstamp_enabled */ |
2680 | if (is_cgroup_event(event)) | |
2681 | perf_cgroup_mark_enabled(event, ctx); | |
2682 | ||
8c9ed8e1 | 2683 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2684 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2685 | |
2686 | /* | |
2687 | * If this pinned group hasn't been scheduled, | |
2688 | * put it in error state. | |
2689 | */ | |
cdd6c482 IM |
2690 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2691 | update_group_times(event); | |
2692 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2693 | } |
3b6f9e5c | 2694 | } |
5b0311e1 FW |
2695 | } |
2696 | ||
2697 | static void | |
2698 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2699 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2700 | { |
2701 | struct perf_event *event; | |
2702 | int can_add_hw = 1; | |
3b6f9e5c | 2703 | |
889ff015 FW |
2704 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2705 | /* Ignore events in OFF or ERROR state */ | |
2706 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2707 | continue; |
04289bb9 IM |
2708 | /* |
2709 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2710 | * of events: |
04289bb9 | 2711 | */ |
5632ab12 | 2712 | if (!event_filter_match(event)) |
0793a61d TG |
2713 | continue; |
2714 | ||
e5d1367f SE |
2715 | /* may need to reset tstamp_enabled */ |
2716 | if (is_cgroup_event(event)) | |
2717 | perf_cgroup_mark_enabled(event, ctx); | |
2718 | ||
9ed6060d | 2719 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2720 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2721 | can_add_hw = 0; |
9ed6060d | 2722 | } |
0793a61d | 2723 | } |
5b0311e1 FW |
2724 | } |
2725 | ||
2726 | static void | |
2727 | ctx_sched_in(struct perf_event_context *ctx, | |
2728 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2729 | enum event_type_t event_type, |
2730 | struct task_struct *task) | |
5b0311e1 | 2731 | { |
e5d1367f | 2732 | u64 now; |
db24d33e | 2733 | int is_active = ctx->is_active; |
e5d1367f | 2734 | |
db24d33e | 2735 | ctx->is_active |= event_type; |
5b0311e1 | 2736 | if (likely(!ctx->nr_events)) |
facc4307 | 2737 | return; |
5b0311e1 | 2738 | |
e5d1367f SE |
2739 | now = perf_clock(); |
2740 | ctx->timestamp = now; | |
3f7cce3c | 2741 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2742 | /* |
2743 | * First go through the list and put on any pinned groups | |
2744 | * in order to give them the best chance of going on. | |
2745 | */ | |
db24d33e | 2746 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2747 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2748 | |
2749 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2750 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2751 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2752 | } |
2753 | ||
329c0e01 | 2754 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2755 | enum event_type_t event_type, |
2756 | struct task_struct *task) | |
329c0e01 FW |
2757 | { |
2758 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2759 | ||
e5d1367f | 2760 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2761 | } |
2762 | ||
e5d1367f SE |
2763 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2764 | struct task_struct *task) | |
235c7fc7 | 2765 | { |
108b02cf | 2766 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2767 | |
108b02cf | 2768 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2769 | if (cpuctx->task_ctx == ctx) |
2770 | return; | |
2771 | ||
facc4307 | 2772 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2773 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2774 | /* |
2775 | * We want to keep the following priority order: | |
2776 | * cpu pinned (that don't need to move), task pinned, | |
2777 | * cpu flexible, task flexible. | |
2778 | */ | |
2779 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2780 | ||
1d5f003f GN |
2781 | if (ctx->nr_events) |
2782 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2783 | |
86b47c25 GN |
2784 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2785 | ||
facc4307 PZ |
2786 | perf_pmu_enable(ctx->pmu); |
2787 | perf_ctx_unlock(cpuctx, ctx); | |
235c7fc7 IM |
2788 | } |
2789 | ||
8dc85d54 PZ |
2790 | /* |
2791 | * Called from scheduler to add the events of the current task | |
2792 | * with interrupts disabled. | |
2793 | * | |
2794 | * We restore the event value and then enable it. | |
2795 | * | |
2796 | * This does not protect us against NMI, but enable() | |
2797 | * sets the enabled bit in the control field of event _before_ | |
2798 | * accessing the event control register. If a NMI hits, then it will | |
2799 | * keep the event running. | |
2800 | */ | |
ab0cce56 JO |
2801 | void __perf_event_task_sched_in(struct task_struct *prev, |
2802 | struct task_struct *task) | |
8dc85d54 PZ |
2803 | { |
2804 | struct perf_event_context *ctx; | |
2805 | int ctxn; | |
2806 | ||
2807 | for_each_task_context_nr(ctxn) { | |
2808 | ctx = task->perf_event_ctxp[ctxn]; | |
2809 | if (likely(!ctx)) | |
2810 | continue; | |
2811 | ||
e5d1367f | 2812 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2813 | } |
e5d1367f SE |
2814 | /* |
2815 | * if cgroup events exist on this CPU, then we need | |
2816 | * to check if we have to switch in PMU state. | |
2817 | * cgroup event are system-wide mode only | |
2818 | */ | |
4a32fea9 | 2819 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2820 | perf_cgroup_sched_in(prev, task); |
d010b332 | 2821 | |
ba532500 YZ |
2822 | if (__this_cpu_read(perf_sched_cb_usages)) |
2823 | perf_pmu_sched_task(prev, task, true); | |
235c7fc7 IM |
2824 | } |
2825 | ||
abd50713 PZ |
2826 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2827 | { | |
2828 | u64 frequency = event->attr.sample_freq; | |
2829 | u64 sec = NSEC_PER_SEC; | |
2830 | u64 divisor, dividend; | |
2831 | ||
2832 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2833 | ||
2834 | count_fls = fls64(count); | |
2835 | nsec_fls = fls64(nsec); | |
2836 | frequency_fls = fls64(frequency); | |
2837 | sec_fls = 30; | |
2838 | ||
2839 | /* | |
2840 | * We got @count in @nsec, with a target of sample_freq HZ | |
2841 | * the target period becomes: | |
2842 | * | |
2843 | * @count * 10^9 | |
2844 | * period = ------------------- | |
2845 | * @nsec * sample_freq | |
2846 | * | |
2847 | */ | |
2848 | ||
2849 | /* | |
2850 | * Reduce accuracy by one bit such that @a and @b converge | |
2851 | * to a similar magnitude. | |
2852 | */ | |
fe4b04fa | 2853 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2854 | do { \ |
2855 | if (a##_fls > b##_fls) { \ | |
2856 | a >>= 1; \ | |
2857 | a##_fls--; \ | |
2858 | } else { \ | |
2859 | b >>= 1; \ | |
2860 | b##_fls--; \ | |
2861 | } \ | |
2862 | } while (0) | |
2863 | ||
2864 | /* | |
2865 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2866 | * the other, so that finally we can do a u64/u64 division. | |
2867 | */ | |
2868 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2869 | REDUCE_FLS(nsec, frequency); | |
2870 | REDUCE_FLS(sec, count); | |
2871 | } | |
2872 | ||
2873 | if (count_fls + sec_fls > 64) { | |
2874 | divisor = nsec * frequency; | |
2875 | ||
2876 | while (count_fls + sec_fls > 64) { | |
2877 | REDUCE_FLS(count, sec); | |
2878 | divisor >>= 1; | |
2879 | } | |
2880 | ||
2881 | dividend = count * sec; | |
2882 | } else { | |
2883 | dividend = count * sec; | |
2884 | ||
2885 | while (nsec_fls + frequency_fls > 64) { | |
2886 | REDUCE_FLS(nsec, frequency); | |
2887 | dividend >>= 1; | |
2888 | } | |
2889 | ||
2890 | divisor = nsec * frequency; | |
2891 | } | |
2892 | ||
f6ab91ad PZ |
2893 | if (!divisor) |
2894 | return dividend; | |
2895 | ||
abd50713 PZ |
2896 | return div64_u64(dividend, divisor); |
2897 | } | |
2898 | ||
e050e3f0 SE |
2899 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2900 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2901 | ||
f39d47ff | 2902 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2903 | { |
cdd6c482 | 2904 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2905 | s64 period, sample_period; |
bd2b5b12 PZ |
2906 | s64 delta; |
2907 | ||
abd50713 | 2908 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2909 | |
2910 | delta = (s64)(period - hwc->sample_period); | |
2911 | delta = (delta + 7) / 8; /* low pass filter */ | |
2912 | ||
2913 | sample_period = hwc->sample_period + delta; | |
2914 | ||
2915 | if (!sample_period) | |
2916 | sample_period = 1; | |
2917 | ||
bd2b5b12 | 2918 | hwc->sample_period = sample_period; |
abd50713 | 2919 | |
e7850595 | 2920 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2921 | if (disable) |
2922 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2923 | ||
e7850595 | 2924 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2925 | |
2926 | if (disable) | |
2927 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2928 | } |
bd2b5b12 PZ |
2929 | } |
2930 | ||
e050e3f0 SE |
2931 | /* |
2932 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2933 | * events. At the same time, make sure, having freq events does not change | |
2934 | * the rate of unthrottling as that would introduce bias. | |
2935 | */ | |
2936 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2937 | int needs_unthr) | |
60db5e09 | 2938 | { |
cdd6c482 IM |
2939 | struct perf_event *event; |
2940 | struct hw_perf_event *hwc; | |
e050e3f0 | 2941 | u64 now, period = TICK_NSEC; |
abd50713 | 2942 | s64 delta; |
60db5e09 | 2943 | |
e050e3f0 SE |
2944 | /* |
2945 | * only need to iterate over all events iff: | |
2946 | * - context have events in frequency mode (needs freq adjust) | |
2947 | * - there are events to unthrottle on this cpu | |
2948 | */ | |
2949 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2950 | return; |
2951 | ||
e050e3f0 | 2952 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2953 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2954 | |
03541f8b | 2955 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2956 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2957 | continue; |
2958 | ||
5632ab12 | 2959 | if (!event_filter_match(event)) |
5d27c23d PZ |
2960 | continue; |
2961 | ||
44377277 AS |
2962 | perf_pmu_disable(event->pmu); |
2963 | ||
cdd6c482 | 2964 | hwc = &event->hw; |
6a24ed6c | 2965 | |
ae23bff1 | 2966 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2967 | hwc->interrupts = 0; |
cdd6c482 | 2968 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2969 | event->pmu->start(event, 0); |
a78ac325 PZ |
2970 | } |
2971 | ||
cdd6c482 | 2972 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2973 | goto next; |
60db5e09 | 2974 | |
e050e3f0 SE |
2975 | /* |
2976 | * stop the event and update event->count | |
2977 | */ | |
2978 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2979 | ||
e7850595 | 2980 | now = local64_read(&event->count); |
abd50713 PZ |
2981 | delta = now - hwc->freq_count_stamp; |
2982 | hwc->freq_count_stamp = now; | |
60db5e09 | 2983 | |
e050e3f0 SE |
2984 | /* |
2985 | * restart the event | |
2986 | * reload only if value has changed | |
f39d47ff SE |
2987 | * we have stopped the event so tell that |
2988 | * to perf_adjust_period() to avoid stopping it | |
2989 | * twice. | |
e050e3f0 | 2990 | */ |
abd50713 | 2991 | if (delta > 0) |
f39d47ff | 2992 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2993 | |
2994 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
2995 | next: |
2996 | perf_pmu_enable(event->pmu); | |
60db5e09 | 2997 | } |
e050e3f0 | 2998 | |
f39d47ff | 2999 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 3000 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
3001 | } |
3002 | ||
235c7fc7 | 3003 | /* |
cdd6c482 | 3004 | * Round-robin a context's events: |
235c7fc7 | 3005 | */ |
cdd6c482 | 3006 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 3007 | { |
dddd3379 TG |
3008 | /* |
3009 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
3010 | * disabled by the inheritance code. | |
3011 | */ | |
3012 | if (!ctx->rotate_disable) | |
3013 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
3014 | } |
3015 | ||
9e630205 | 3016 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 3017 | { |
8dc85d54 | 3018 | struct perf_event_context *ctx = NULL; |
2fde4f94 | 3019 | int rotate = 0; |
7fc23a53 | 3020 | |
b5ab4cd5 | 3021 | if (cpuctx->ctx.nr_events) { |
b5ab4cd5 PZ |
3022 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
3023 | rotate = 1; | |
3024 | } | |
235c7fc7 | 3025 | |
8dc85d54 | 3026 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 3027 | if (ctx && ctx->nr_events) { |
b5ab4cd5 PZ |
3028 | if (ctx->nr_events != ctx->nr_active) |
3029 | rotate = 1; | |
3030 | } | |
9717e6cd | 3031 | |
e050e3f0 | 3032 | if (!rotate) |
0f5a2601 PZ |
3033 | goto done; |
3034 | ||
facc4307 | 3035 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 3036 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 3037 | |
e050e3f0 SE |
3038 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
3039 | if (ctx) | |
3040 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 3041 | |
e050e3f0 SE |
3042 | rotate_ctx(&cpuctx->ctx); |
3043 | if (ctx) | |
3044 | rotate_ctx(ctx); | |
235c7fc7 | 3045 | |
e050e3f0 | 3046 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 3047 | |
0f5a2601 PZ |
3048 | perf_pmu_enable(cpuctx->ctx.pmu); |
3049 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 3050 | done: |
9e630205 SE |
3051 | |
3052 | return rotate; | |
e9d2b064 PZ |
3053 | } |
3054 | ||
026249ef FW |
3055 | #ifdef CONFIG_NO_HZ_FULL |
3056 | bool perf_event_can_stop_tick(void) | |
3057 | { | |
948b26b6 | 3058 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 3059 | __this_cpu_read(perf_throttled_count)) |
026249ef | 3060 | return false; |
d84153d6 FW |
3061 | else |
3062 | return true; | |
026249ef FW |
3063 | } |
3064 | #endif | |
3065 | ||
e9d2b064 PZ |
3066 | void perf_event_task_tick(void) |
3067 | { | |
2fde4f94 MR |
3068 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
3069 | struct perf_event_context *ctx, *tmp; | |
e050e3f0 | 3070 | int throttled; |
b5ab4cd5 | 3071 | |
e9d2b064 PZ |
3072 | WARN_ON(!irqs_disabled()); |
3073 | ||
e050e3f0 SE |
3074 | __this_cpu_inc(perf_throttled_seq); |
3075 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
3076 | ||
2fde4f94 | 3077 | list_for_each_entry_safe(ctx, tmp, head, active_ctx_list) |
e050e3f0 | 3078 | perf_adjust_freq_unthr_context(ctx, throttled); |
0793a61d TG |
3079 | } |
3080 | ||
889ff015 FW |
3081 | static int event_enable_on_exec(struct perf_event *event, |
3082 | struct perf_event_context *ctx) | |
3083 | { | |
3084 | if (!event->attr.enable_on_exec) | |
3085 | return 0; | |
3086 | ||
3087 | event->attr.enable_on_exec = 0; | |
3088 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
3089 | return 0; | |
3090 | ||
1d9b482e | 3091 | __perf_event_mark_enabled(event); |
889ff015 FW |
3092 | |
3093 | return 1; | |
3094 | } | |
3095 | ||
57e7986e | 3096 | /* |
cdd6c482 | 3097 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
3098 | * This expects task == current. |
3099 | */ | |
8dc85d54 | 3100 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 3101 | { |
211de6eb | 3102 | struct perf_event_context *clone_ctx = NULL; |
cdd6c482 | 3103 | struct perf_event *event; |
57e7986e PM |
3104 | unsigned long flags; |
3105 | int enabled = 0; | |
889ff015 | 3106 | int ret; |
57e7986e PM |
3107 | |
3108 | local_irq_save(flags); | |
cdd6c482 | 3109 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
3110 | goto out; |
3111 | ||
e566b76e SE |
3112 | /* |
3113 | * We must ctxsw out cgroup events to avoid conflict | |
3114 | * when invoking perf_task_event_sched_in() later on | |
3115 | * in this function. Otherwise we end up trying to | |
3116 | * ctxswin cgroup events which are already scheduled | |
3117 | * in. | |
3118 | */ | |
a8d757ef | 3119 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 3120 | |
e625cce1 | 3121 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 3122 | task_ctx_sched_out(ctx); |
57e7986e | 3123 | |
b79387ef | 3124 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
3125 | ret = event_enable_on_exec(event, ctx); |
3126 | if (ret) | |
3127 | enabled = 1; | |
57e7986e PM |
3128 | } |
3129 | ||
3130 | /* | |
cdd6c482 | 3131 | * Unclone this context if we enabled any event. |
57e7986e | 3132 | */ |
71a851b4 | 3133 | if (enabled) |
211de6eb | 3134 | clone_ctx = unclone_ctx(ctx); |
57e7986e | 3135 | |
e625cce1 | 3136 | raw_spin_unlock(&ctx->lock); |
57e7986e | 3137 | |
e566b76e SE |
3138 | /* |
3139 | * Also calls ctxswin for cgroup events, if any: | |
3140 | */ | |
e5d1367f | 3141 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 3142 | out: |
57e7986e | 3143 | local_irq_restore(flags); |
211de6eb PZ |
3144 | |
3145 | if (clone_ctx) | |
3146 | put_ctx(clone_ctx); | |
57e7986e PM |
3147 | } |
3148 | ||
e041e328 PZ |
3149 | void perf_event_exec(void) |
3150 | { | |
3151 | struct perf_event_context *ctx; | |
3152 | int ctxn; | |
3153 | ||
3154 | rcu_read_lock(); | |
3155 | for_each_task_context_nr(ctxn) { | |
3156 | ctx = current->perf_event_ctxp[ctxn]; | |
3157 | if (!ctx) | |
3158 | continue; | |
3159 | ||
3160 | perf_event_enable_on_exec(ctx); | |
3161 | } | |
3162 | rcu_read_unlock(); | |
3163 | } | |
3164 | ||
0793a61d | 3165 | /* |
cdd6c482 | 3166 | * Cross CPU call to read the hardware event |
0793a61d | 3167 | */ |
cdd6c482 | 3168 | static void __perf_event_read(void *info) |
0793a61d | 3169 | { |
cdd6c482 IM |
3170 | struct perf_event *event = info; |
3171 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 3172 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3173 | |
e1ac3614 PM |
3174 | /* |
3175 | * If this is a task context, we need to check whether it is | |
3176 | * the current task context of this cpu. If not it has been | |
3177 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3178 | * event->count would have been updated to a recent sample |
3179 | * when the event was scheduled out. | |
e1ac3614 PM |
3180 | */ |
3181 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3182 | return; | |
3183 | ||
e625cce1 | 3184 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3185 | if (ctx->is_active) { |
542e72fc | 3186 | update_context_time(ctx); |
e5d1367f SE |
3187 | update_cgrp_time_from_event(event); |
3188 | } | |
cdd6c482 | 3189 | update_event_times(event); |
542e72fc PZ |
3190 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3191 | event->pmu->read(event); | |
e625cce1 | 3192 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3193 | } |
3194 | ||
b5e58793 PZ |
3195 | static inline u64 perf_event_count(struct perf_event *event) |
3196 | { | |
eacd3ecc MF |
3197 | if (event->pmu->count) |
3198 | return event->pmu->count(event); | |
3199 | ||
3200 | return __perf_event_count(event); | |
b5e58793 PZ |
3201 | } |
3202 | ||
cdd6c482 | 3203 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3204 | { |
3205 | /* | |
cdd6c482 IM |
3206 | * If event is enabled and currently active on a CPU, update the |
3207 | * value in the event structure: | |
0793a61d | 3208 | */ |
cdd6c482 IM |
3209 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3210 | smp_call_function_single(event->oncpu, | |
3211 | __perf_event_read, event, 1); | |
3212 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3213 | struct perf_event_context *ctx = event->ctx; |
3214 | unsigned long flags; | |
3215 | ||
e625cce1 | 3216 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3217 | /* |
3218 | * may read while context is not active | |
3219 | * (e.g., thread is blocked), in that case | |
3220 | * we cannot update context time | |
3221 | */ | |
e5d1367f | 3222 | if (ctx->is_active) { |
c530ccd9 | 3223 | update_context_time(ctx); |
e5d1367f SE |
3224 | update_cgrp_time_from_event(event); |
3225 | } | |
cdd6c482 | 3226 | update_event_times(event); |
e625cce1 | 3227 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3228 | } |
3229 | ||
b5e58793 | 3230 | return perf_event_count(event); |
0793a61d TG |
3231 | } |
3232 | ||
a63eaf34 | 3233 | /* |
cdd6c482 | 3234 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3235 | */ |
eb184479 | 3236 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3237 | { |
e625cce1 | 3238 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3239 | mutex_init(&ctx->mutex); |
2fde4f94 | 3240 | INIT_LIST_HEAD(&ctx->active_ctx_list); |
889ff015 FW |
3241 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3242 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3243 | INIT_LIST_HEAD(&ctx->event_list); |
3244 | atomic_set(&ctx->refcount, 1); | |
fadfe7be | 3245 | INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work); |
eb184479 PZ |
3246 | } |
3247 | ||
3248 | static struct perf_event_context * | |
3249 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3250 | { | |
3251 | struct perf_event_context *ctx; | |
3252 | ||
3253 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3254 | if (!ctx) | |
3255 | return NULL; | |
3256 | ||
3257 | __perf_event_init_context(ctx); | |
3258 | if (task) { | |
3259 | ctx->task = task; | |
3260 | get_task_struct(task); | |
0793a61d | 3261 | } |
eb184479 PZ |
3262 | ctx->pmu = pmu; |
3263 | ||
3264 | return ctx; | |
a63eaf34 PM |
3265 | } |
3266 | ||
2ebd4ffb MH |
3267 | static struct task_struct * |
3268 | find_lively_task_by_vpid(pid_t vpid) | |
3269 | { | |
3270 | struct task_struct *task; | |
3271 | int err; | |
0793a61d TG |
3272 | |
3273 | rcu_read_lock(); | |
2ebd4ffb | 3274 | if (!vpid) |
0793a61d TG |
3275 | task = current; |
3276 | else | |
2ebd4ffb | 3277 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3278 | if (task) |
3279 | get_task_struct(task); | |
3280 | rcu_read_unlock(); | |
3281 | ||
3282 | if (!task) | |
3283 | return ERR_PTR(-ESRCH); | |
3284 | ||
0793a61d | 3285 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3286 | err = -EACCES; |
3287 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3288 | goto errout; | |
3289 | ||
2ebd4ffb MH |
3290 | return task; |
3291 | errout: | |
3292 | put_task_struct(task); | |
3293 | return ERR_PTR(err); | |
3294 | ||
3295 | } | |
3296 | ||
fe4b04fa PZ |
3297 | /* |
3298 | * Returns a matching context with refcount and pincount. | |
3299 | */ | |
108b02cf | 3300 | static struct perf_event_context * |
4af57ef2 YZ |
3301 | find_get_context(struct pmu *pmu, struct task_struct *task, |
3302 | struct perf_event *event) | |
0793a61d | 3303 | { |
211de6eb | 3304 | struct perf_event_context *ctx, *clone_ctx = NULL; |
22a4f650 | 3305 | struct perf_cpu_context *cpuctx; |
4af57ef2 | 3306 | void *task_ctx_data = NULL; |
25346b93 | 3307 | unsigned long flags; |
8dc85d54 | 3308 | int ctxn, err; |
4af57ef2 | 3309 | int cpu = event->cpu; |
0793a61d | 3310 | |
22a4ec72 | 3311 | if (!task) { |
cdd6c482 | 3312 | /* Must be root to operate on a CPU event: */ |
0764771d | 3313 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3314 | return ERR_PTR(-EACCES); |
3315 | ||
0793a61d | 3316 | /* |
cdd6c482 | 3317 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3318 | * offline CPU and activate it when the CPU comes up, but |
3319 | * that's for later. | |
3320 | */ | |
f6325e30 | 3321 | if (!cpu_online(cpu)) |
0793a61d TG |
3322 | return ERR_PTR(-ENODEV); |
3323 | ||
108b02cf | 3324 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3325 | ctx = &cpuctx->ctx; |
c93f7669 | 3326 | get_ctx(ctx); |
fe4b04fa | 3327 | ++ctx->pin_count; |
0793a61d | 3328 | |
0793a61d TG |
3329 | return ctx; |
3330 | } | |
3331 | ||
8dc85d54 PZ |
3332 | err = -EINVAL; |
3333 | ctxn = pmu->task_ctx_nr; | |
3334 | if (ctxn < 0) | |
3335 | goto errout; | |
3336 | ||
4af57ef2 YZ |
3337 | if (event->attach_state & PERF_ATTACH_TASK_DATA) { |
3338 | task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL); | |
3339 | if (!task_ctx_data) { | |
3340 | err = -ENOMEM; | |
3341 | goto errout; | |
3342 | } | |
3343 | } | |
3344 | ||
9ed6060d | 3345 | retry: |
8dc85d54 | 3346 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3347 | if (ctx) { |
211de6eb | 3348 | clone_ctx = unclone_ctx(ctx); |
fe4b04fa | 3349 | ++ctx->pin_count; |
4af57ef2 YZ |
3350 | |
3351 | if (task_ctx_data && !ctx->task_ctx_data) { | |
3352 | ctx->task_ctx_data = task_ctx_data; | |
3353 | task_ctx_data = NULL; | |
3354 | } | |
e625cce1 | 3355 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
211de6eb PZ |
3356 | |
3357 | if (clone_ctx) | |
3358 | put_ctx(clone_ctx); | |
9137fb28 | 3359 | } else { |
eb184479 | 3360 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3361 | err = -ENOMEM; |
3362 | if (!ctx) | |
3363 | goto errout; | |
eb184479 | 3364 | |
4af57ef2 YZ |
3365 | if (task_ctx_data) { |
3366 | ctx->task_ctx_data = task_ctx_data; | |
3367 | task_ctx_data = NULL; | |
3368 | } | |
3369 | ||
dbe08d82 ON |
3370 | err = 0; |
3371 | mutex_lock(&task->perf_event_mutex); | |
3372 | /* | |
3373 | * If it has already passed perf_event_exit_task(). | |
3374 | * we must see PF_EXITING, it takes this mutex too. | |
3375 | */ | |
3376 | if (task->flags & PF_EXITING) | |
3377 | err = -ESRCH; | |
3378 | else if (task->perf_event_ctxp[ctxn]) | |
3379 | err = -EAGAIN; | |
fe4b04fa | 3380 | else { |
9137fb28 | 3381 | get_ctx(ctx); |
fe4b04fa | 3382 | ++ctx->pin_count; |
dbe08d82 | 3383 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3384 | } |
dbe08d82 ON |
3385 | mutex_unlock(&task->perf_event_mutex); |
3386 | ||
3387 | if (unlikely(err)) { | |
9137fb28 | 3388 | put_ctx(ctx); |
dbe08d82 ON |
3389 | |
3390 | if (err == -EAGAIN) | |
3391 | goto retry; | |
3392 | goto errout; | |
a63eaf34 PM |
3393 | } |
3394 | } | |
3395 | ||
4af57ef2 | 3396 | kfree(task_ctx_data); |
0793a61d | 3397 | return ctx; |
c93f7669 | 3398 | |
9ed6060d | 3399 | errout: |
4af57ef2 | 3400 | kfree(task_ctx_data); |
c93f7669 | 3401 | return ERR_PTR(err); |
0793a61d TG |
3402 | } |
3403 | ||
6fb2915d LZ |
3404 | static void perf_event_free_filter(struct perf_event *event); |
3405 | ||
cdd6c482 | 3406 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3407 | { |
cdd6c482 | 3408 | struct perf_event *event; |
592903cd | 3409 | |
cdd6c482 IM |
3410 | event = container_of(head, struct perf_event, rcu_head); |
3411 | if (event->ns) | |
3412 | put_pid_ns(event->ns); | |
6fb2915d | 3413 | perf_event_free_filter(event); |
cdd6c482 | 3414 | kfree(event); |
592903cd PZ |
3415 | } |
3416 | ||
76369139 | 3417 | static void ring_buffer_put(struct ring_buffer *rb); |
b69cf536 PZ |
3418 | static void ring_buffer_attach(struct perf_event *event, |
3419 | struct ring_buffer *rb); | |
925d519a | 3420 | |
4beb31f3 | 3421 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3422 | { |
4beb31f3 FW |
3423 | if (event->parent) |
3424 | return; | |
3425 | ||
4beb31f3 FW |
3426 | if (is_cgroup_event(event)) |
3427 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3428 | } | |
925d519a | 3429 | |
4beb31f3 FW |
3430 | static void unaccount_event(struct perf_event *event) |
3431 | { | |
3432 | if (event->parent) | |
3433 | return; | |
3434 | ||
3435 | if (event->attach_state & PERF_ATTACH_TASK) | |
3436 | static_key_slow_dec_deferred(&perf_sched_events); | |
3437 | if (event->attr.mmap || event->attr.mmap_data) | |
3438 | atomic_dec(&nr_mmap_events); | |
3439 | if (event->attr.comm) | |
3440 | atomic_dec(&nr_comm_events); | |
3441 | if (event->attr.task) | |
3442 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3443 | if (event->attr.freq) |
3444 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3445 | if (is_cgroup_event(event)) |
3446 | static_key_slow_dec_deferred(&perf_sched_events); | |
3447 | if (has_branch_stack(event)) | |
3448 | static_key_slow_dec_deferred(&perf_sched_events); | |
3449 | ||
3450 | unaccount_event_cpu(event, event->cpu); | |
3451 | } | |
925d519a | 3452 | |
766d6c07 FW |
3453 | static void __free_event(struct perf_event *event) |
3454 | { | |
cdd6c482 | 3455 | if (!event->parent) { |
927c7a9e FW |
3456 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3457 | put_callchain_buffers(); | |
f344011c | 3458 | } |
9ee318a7 | 3459 | |
766d6c07 FW |
3460 | if (event->destroy) |
3461 | event->destroy(event); | |
3462 | ||
3463 | if (event->ctx) | |
3464 | put_ctx(event->ctx); | |
3465 | ||
c464c76e YZ |
3466 | if (event->pmu) |
3467 | module_put(event->pmu->module); | |
3468 | ||
766d6c07 FW |
3469 | call_rcu(&event->rcu_head, free_event_rcu); |
3470 | } | |
683ede43 PZ |
3471 | |
3472 | static void _free_event(struct perf_event *event) | |
f1600952 | 3473 | { |
e360adbe | 3474 | irq_work_sync(&event->pending); |
925d519a | 3475 | |
4beb31f3 | 3476 | unaccount_event(event); |
9ee318a7 | 3477 | |
76369139 | 3478 | if (event->rb) { |
9bb5d40c PZ |
3479 | /* |
3480 | * Can happen when we close an event with re-directed output. | |
3481 | * | |
3482 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3483 | * over us; possibly making our ring_buffer_put() the last. | |
3484 | */ | |
3485 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3486 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3487 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3488 | } |
3489 | ||
e5d1367f SE |
3490 | if (is_cgroup_event(event)) |
3491 | perf_detach_cgroup(event); | |
3492 | ||
766d6c07 | 3493 | __free_event(event); |
f1600952 PZ |
3494 | } |
3495 | ||
683ede43 PZ |
3496 | /* |
3497 | * Used to free events which have a known refcount of 1, such as in error paths | |
3498 | * where the event isn't exposed yet and inherited events. | |
3499 | */ | |
3500 | static void free_event(struct perf_event *event) | |
0793a61d | 3501 | { |
683ede43 PZ |
3502 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3503 | "unexpected event refcount: %ld; ptr=%p\n", | |
3504 | atomic_long_read(&event->refcount), event)) { | |
3505 | /* leak to avoid use-after-free */ | |
3506 | return; | |
3507 | } | |
0793a61d | 3508 | |
683ede43 | 3509 | _free_event(event); |
0793a61d TG |
3510 | } |
3511 | ||
a66a3052 | 3512 | /* |
f8697762 | 3513 | * Remove user event from the owner task. |
a66a3052 | 3514 | */ |
f8697762 | 3515 | static void perf_remove_from_owner(struct perf_event *event) |
fb0459d7 | 3516 | { |
8882135b | 3517 | struct task_struct *owner; |
fb0459d7 | 3518 | |
8882135b PZ |
3519 | rcu_read_lock(); |
3520 | owner = ACCESS_ONCE(event->owner); | |
3521 | /* | |
3522 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3523 | * !owner it means the list deletion is complete and we can indeed | |
3524 | * free this event, otherwise we need to serialize on | |
3525 | * owner->perf_event_mutex. | |
3526 | */ | |
3527 | smp_read_barrier_depends(); | |
3528 | if (owner) { | |
3529 | /* | |
3530 | * Since delayed_put_task_struct() also drops the last | |
3531 | * task reference we can safely take a new reference | |
3532 | * while holding the rcu_read_lock(). | |
3533 | */ | |
3534 | get_task_struct(owner); | |
3535 | } | |
3536 | rcu_read_unlock(); | |
3537 | ||
3538 | if (owner) { | |
f63a8daa PZ |
3539 | /* |
3540 | * If we're here through perf_event_exit_task() we're already | |
3541 | * holding ctx->mutex which would be an inversion wrt. the | |
3542 | * normal lock order. | |
3543 | * | |
3544 | * However we can safely take this lock because its the child | |
3545 | * ctx->mutex. | |
3546 | */ | |
3547 | mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); | |
3548 | ||
8882135b PZ |
3549 | /* |
3550 | * We have to re-check the event->owner field, if it is cleared | |
3551 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3552 | * ensured they're done, and we can proceed with freeing the | |
3553 | * event. | |
3554 | */ | |
3555 | if (event->owner) | |
3556 | list_del_init(&event->owner_entry); | |
3557 | mutex_unlock(&owner->perf_event_mutex); | |
3558 | put_task_struct(owner); | |
3559 | } | |
f8697762 JO |
3560 | } |
3561 | ||
3562 | /* | |
3563 | * Called when the last reference to the file is gone. | |
3564 | */ | |
3565 | static void put_event(struct perf_event *event) | |
3566 | { | |
a83fe28e | 3567 | struct perf_event_context *ctx; |
f8697762 JO |
3568 | |
3569 | if (!atomic_long_dec_and_test(&event->refcount)) | |
3570 | return; | |
3571 | ||
3572 | if (!is_kernel_event(event)) | |
3573 | perf_remove_from_owner(event); | |
8882135b | 3574 | |
683ede43 PZ |
3575 | /* |
3576 | * There are two ways this annotation is useful: | |
3577 | * | |
3578 | * 1) there is a lock recursion from perf_event_exit_task | |
3579 | * see the comment there. | |
3580 | * | |
3581 | * 2) there is a lock-inversion with mmap_sem through | |
3582 | * perf_event_read_group(), which takes faults while | |
3583 | * holding ctx->mutex, however this is called after | |
3584 | * the last filedesc died, so there is no possibility | |
3585 | * to trigger the AB-BA case. | |
3586 | */ | |
a83fe28e PZ |
3587 | ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING); |
3588 | WARN_ON_ONCE(ctx->parent_ctx); | |
683ede43 | 3589 | perf_remove_from_context(event, true); |
d415a7f1 | 3590 | perf_event_ctx_unlock(event, ctx); |
683ede43 PZ |
3591 | |
3592 | _free_event(event); | |
a6fa941d AV |
3593 | } |
3594 | ||
683ede43 PZ |
3595 | int perf_event_release_kernel(struct perf_event *event) |
3596 | { | |
3597 | put_event(event); | |
3598 | return 0; | |
3599 | } | |
3600 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3601 | ||
a6fa941d AV |
3602 | static int perf_release(struct inode *inode, struct file *file) |
3603 | { | |
3604 | put_event(file->private_data); | |
3605 | return 0; | |
fb0459d7 | 3606 | } |
fb0459d7 | 3607 | |
fadfe7be JO |
3608 | /* |
3609 | * Remove all orphanes events from the context. | |
3610 | */ | |
3611 | static void orphans_remove_work(struct work_struct *work) | |
3612 | { | |
3613 | struct perf_event_context *ctx; | |
3614 | struct perf_event *event, *tmp; | |
3615 | ||
3616 | ctx = container_of(work, struct perf_event_context, | |
3617 | orphans_remove.work); | |
3618 | ||
3619 | mutex_lock(&ctx->mutex); | |
3620 | list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) { | |
3621 | struct perf_event *parent_event = event->parent; | |
3622 | ||
3623 | if (!is_orphaned_child(event)) | |
3624 | continue; | |
3625 | ||
3626 | perf_remove_from_context(event, true); | |
3627 | ||
3628 | mutex_lock(&parent_event->child_mutex); | |
3629 | list_del_init(&event->child_list); | |
3630 | mutex_unlock(&parent_event->child_mutex); | |
3631 | ||
3632 | free_event(event); | |
3633 | put_event(parent_event); | |
3634 | } | |
3635 | ||
3636 | raw_spin_lock_irq(&ctx->lock); | |
3637 | ctx->orphans_remove_sched = false; | |
3638 | raw_spin_unlock_irq(&ctx->lock); | |
3639 | mutex_unlock(&ctx->mutex); | |
3640 | ||
3641 | put_ctx(ctx); | |
3642 | } | |
3643 | ||
59ed446f | 3644 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3645 | { |
cdd6c482 | 3646 | struct perf_event *child; |
e53c0994 PZ |
3647 | u64 total = 0; |
3648 | ||
59ed446f PZ |
3649 | *enabled = 0; |
3650 | *running = 0; | |
3651 | ||
6f10581a | 3652 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3653 | total += perf_event_read(event); |
59ed446f PZ |
3654 | *enabled += event->total_time_enabled + |
3655 | atomic64_read(&event->child_total_time_enabled); | |
3656 | *running += event->total_time_running + | |
3657 | atomic64_read(&event->child_total_time_running); | |
3658 | ||
3659 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3660 | total += perf_event_read(child); |
59ed446f PZ |
3661 | *enabled += child->total_time_enabled; |
3662 | *running += child->total_time_running; | |
3663 | } | |
6f10581a | 3664 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3665 | |
3666 | return total; | |
3667 | } | |
fb0459d7 | 3668 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3669 | |
cdd6c482 | 3670 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3671 | u64 read_format, char __user *buf) |
3672 | { | |
cdd6c482 | 3673 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a | 3674 | struct perf_event_context *ctx = leader->ctx; |
f63a8daa | 3675 | int n = 0, size = 0, ret; |
59ed446f | 3676 | u64 count, enabled, running; |
f63a8daa PZ |
3677 | u64 values[5]; |
3678 | ||
3679 | lockdep_assert_held(&ctx->mutex); | |
abf4868b | 3680 | |
59ed446f | 3681 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3682 | |
3683 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3684 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3685 | values[n++] = enabled; | |
3686 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3687 | values[n++] = running; | |
abf4868b PZ |
3688 | values[n++] = count; |
3689 | if (read_format & PERF_FORMAT_ID) | |
3690 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3691 | |
3692 | size = n * sizeof(u64); | |
3693 | ||
3694 | if (copy_to_user(buf, values, size)) | |
f63a8daa | 3695 | return -EFAULT; |
3dab77fb | 3696 | |
6f10581a | 3697 | ret = size; |
3dab77fb | 3698 | |
65abc865 | 3699 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3700 | n = 0; |
3dab77fb | 3701 | |
59ed446f | 3702 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3703 | if (read_format & PERF_FORMAT_ID) |
3704 | values[n++] = primary_event_id(sub); | |
3705 | ||
3706 | size = n * sizeof(u64); | |
3707 | ||
184d3da8 | 3708 | if (copy_to_user(buf + ret, values, size)) { |
f63a8daa | 3709 | return -EFAULT; |
6f10581a | 3710 | } |
abf4868b PZ |
3711 | |
3712 | ret += size; | |
3dab77fb PZ |
3713 | } |
3714 | ||
abf4868b | 3715 | return ret; |
3dab77fb PZ |
3716 | } |
3717 | ||
cdd6c482 | 3718 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3719 | u64 read_format, char __user *buf) |
3720 | { | |
59ed446f | 3721 | u64 enabled, running; |
3dab77fb PZ |
3722 | u64 values[4]; |
3723 | int n = 0; | |
3724 | ||
59ed446f PZ |
3725 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3726 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3727 | values[n++] = enabled; | |
3728 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3729 | values[n++] = running; | |
3dab77fb | 3730 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3731 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3732 | |
3733 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3734 | return -EFAULT; | |
3735 | ||
3736 | return n * sizeof(u64); | |
3737 | } | |
3738 | ||
dc633982 JO |
3739 | static bool is_event_hup(struct perf_event *event) |
3740 | { | |
3741 | bool no_children; | |
3742 | ||
3743 | if (event->state != PERF_EVENT_STATE_EXIT) | |
3744 | return false; | |
3745 | ||
3746 | mutex_lock(&event->child_mutex); | |
3747 | no_children = list_empty(&event->child_list); | |
3748 | mutex_unlock(&event->child_mutex); | |
3749 | return no_children; | |
3750 | } | |
3751 | ||
0793a61d | 3752 | /* |
cdd6c482 | 3753 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3754 | */ |
3755 | static ssize_t | |
cdd6c482 | 3756 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3757 | { |
cdd6c482 | 3758 | u64 read_format = event->attr.read_format; |
3dab77fb | 3759 | int ret; |
0793a61d | 3760 | |
3b6f9e5c | 3761 | /* |
cdd6c482 | 3762 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3763 | * error state (i.e. because it was pinned but it couldn't be |
3764 | * scheduled on to the CPU at some point). | |
3765 | */ | |
cdd6c482 | 3766 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3767 | return 0; |
3768 | ||
c320c7b7 | 3769 | if (count < event->read_size) |
3dab77fb PZ |
3770 | return -ENOSPC; |
3771 | ||
cdd6c482 | 3772 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3773 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3774 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3775 | else |
cdd6c482 | 3776 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3777 | |
3dab77fb | 3778 | return ret; |
0793a61d TG |
3779 | } |
3780 | ||
0793a61d TG |
3781 | static ssize_t |
3782 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3783 | { | |
cdd6c482 | 3784 | struct perf_event *event = file->private_data; |
f63a8daa PZ |
3785 | struct perf_event_context *ctx; |
3786 | int ret; | |
0793a61d | 3787 | |
f63a8daa PZ |
3788 | ctx = perf_event_ctx_lock(event); |
3789 | ret = perf_read_hw(event, buf, count); | |
3790 | perf_event_ctx_unlock(event, ctx); | |
3791 | ||
3792 | return ret; | |
0793a61d TG |
3793 | } |
3794 | ||
3795 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3796 | { | |
cdd6c482 | 3797 | struct perf_event *event = file->private_data; |
76369139 | 3798 | struct ring_buffer *rb; |
61b67684 | 3799 | unsigned int events = POLLHUP; |
c7138f37 | 3800 | |
e708d7ad | 3801 | poll_wait(file, &event->waitq, wait); |
179033b3 | 3802 | |
dc633982 | 3803 | if (is_event_hup(event)) |
179033b3 | 3804 | return events; |
c7138f37 | 3805 | |
10c6db11 | 3806 | /* |
9bb5d40c PZ |
3807 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3808 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3809 | */ |
3810 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3811 | rb = event->rb; |
3812 | if (rb) | |
76369139 | 3813 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3814 | mutex_unlock(&event->mmap_mutex); |
0793a61d TG |
3815 | return events; |
3816 | } | |
3817 | ||
f63a8daa | 3818 | static void _perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3819 | { |
cdd6c482 | 3820 | (void)perf_event_read(event); |
e7850595 | 3821 | local64_set(&event->count, 0); |
cdd6c482 | 3822 | perf_event_update_userpage(event); |
3df5edad PZ |
3823 | } |
3824 | ||
c93f7669 | 3825 | /* |
cdd6c482 IM |
3826 | * Holding the top-level event's child_mutex means that any |
3827 | * descendant process that has inherited this event will block | |
3828 | * in sync_child_event if it goes to exit, thus satisfying the | |
3829 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3830 | */ |
cdd6c482 IM |
3831 | static void perf_event_for_each_child(struct perf_event *event, |
3832 | void (*func)(struct perf_event *)) | |
3df5edad | 3833 | { |
cdd6c482 | 3834 | struct perf_event *child; |
3df5edad | 3835 | |
cdd6c482 | 3836 | WARN_ON_ONCE(event->ctx->parent_ctx); |
f63a8daa | 3837 | |
cdd6c482 IM |
3838 | mutex_lock(&event->child_mutex); |
3839 | func(event); | |
3840 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3841 | func(child); |
cdd6c482 | 3842 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3843 | } |
3844 | ||
cdd6c482 IM |
3845 | static void perf_event_for_each(struct perf_event *event, |
3846 | void (*func)(struct perf_event *)) | |
3df5edad | 3847 | { |
cdd6c482 IM |
3848 | struct perf_event_context *ctx = event->ctx; |
3849 | struct perf_event *sibling; | |
3df5edad | 3850 | |
f63a8daa PZ |
3851 | lockdep_assert_held(&ctx->mutex); |
3852 | ||
cdd6c482 | 3853 | event = event->group_leader; |
75f937f2 | 3854 | |
cdd6c482 | 3855 | perf_event_for_each_child(event, func); |
cdd6c482 | 3856 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3857 | perf_event_for_each_child(sibling, func); |
6de6a7b9 PZ |
3858 | } |
3859 | ||
cdd6c482 | 3860 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3861 | { |
cdd6c482 | 3862 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3863 | int ret = 0, active; |
08247e31 PZ |
3864 | u64 value; |
3865 | ||
6c7e550f | 3866 | if (!is_sampling_event(event)) |
08247e31 PZ |
3867 | return -EINVAL; |
3868 | ||
ad0cf347 | 3869 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3870 | return -EFAULT; |
3871 | ||
3872 | if (!value) | |
3873 | return -EINVAL; | |
3874 | ||
e625cce1 | 3875 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3876 | if (event->attr.freq) { |
3877 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3878 | ret = -EINVAL; |
3879 | goto unlock; | |
3880 | } | |
3881 | ||
cdd6c482 | 3882 | event->attr.sample_freq = value; |
08247e31 | 3883 | } else { |
cdd6c482 IM |
3884 | event->attr.sample_period = value; |
3885 | event->hw.sample_period = value; | |
08247e31 | 3886 | } |
bad7192b PZ |
3887 | |
3888 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
3889 | if (active) { | |
3890 | perf_pmu_disable(ctx->pmu); | |
3891 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3892 | } | |
3893 | ||
3894 | local64_set(&event->hw.period_left, 0); | |
3895 | ||
3896 | if (active) { | |
3897 | event->pmu->start(event, PERF_EF_RELOAD); | |
3898 | perf_pmu_enable(ctx->pmu); | |
3899 | } | |
3900 | ||
08247e31 | 3901 | unlock: |
e625cce1 | 3902 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3903 | |
3904 | return ret; | |
3905 | } | |
3906 | ||
ac9721f3 PZ |
3907 | static const struct file_operations perf_fops; |
3908 | ||
2903ff01 | 3909 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3910 | { |
2903ff01 AV |
3911 | struct fd f = fdget(fd); |
3912 | if (!f.file) | |
3913 | return -EBADF; | |
ac9721f3 | 3914 | |
2903ff01 AV |
3915 | if (f.file->f_op != &perf_fops) { |
3916 | fdput(f); | |
3917 | return -EBADF; | |
ac9721f3 | 3918 | } |
2903ff01 AV |
3919 | *p = f; |
3920 | return 0; | |
ac9721f3 PZ |
3921 | } |
3922 | ||
3923 | static int perf_event_set_output(struct perf_event *event, | |
3924 | struct perf_event *output_event); | |
6fb2915d | 3925 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3926 | |
f63a8daa | 3927 | static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg) |
d859e29f | 3928 | { |
cdd6c482 | 3929 | void (*func)(struct perf_event *); |
3df5edad | 3930 | u32 flags = arg; |
d859e29f PM |
3931 | |
3932 | switch (cmd) { | |
cdd6c482 | 3933 | case PERF_EVENT_IOC_ENABLE: |
f63a8daa | 3934 | func = _perf_event_enable; |
d859e29f | 3935 | break; |
cdd6c482 | 3936 | case PERF_EVENT_IOC_DISABLE: |
f63a8daa | 3937 | func = _perf_event_disable; |
79f14641 | 3938 | break; |
cdd6c482 | 3939 | case PERF_EVENT_IOC_RESET: |
f63a8daa | 3940 | func = _perf_event_reset; |
6de6a7b9 | 3941 | break; |
3df5edad | 3942 | |
cdd6c482 | 3943 | case PERF_EVENT_IOC_REFRESH: |
f63a8daa | 3944 | return _perf_event_refresh(event, arg); |
08247e31 | 3945 | |
cdd6c482 IM |
3946 | case PERF_EVENT_IOC_PERIOD: |
3947 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3948 | |
cf4957f1 JO |
3949 | case PERF_EVENT_IOC_ID: |
3950 | { | |
3951 | u64 id = primary_event_id(event); | |
3952 | ||
3953 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3954 | return -EFAULT; | |
3955 | return 0; | |
3956 | } | |
3957 | ||
cdd6c482 | 3958 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3959 | { |
ac9721f3 | 3960 | int ret; |
ac9721f3 | 3961 | if (arg != -1) { |
2903ff01 AV |
3962 | struct perf_event *output_event; |
3963 | struct fd output; | |
3964 | ret = perf_fget_light(arg, &output); | |
3965 | if (ret) | |
3966 | return ret; | |
3967 | output_event = output.file->private_data; | |
3968 | ret = perf_event_set_output(event, output_event); | |
3969 | fdput(output); | |
3970 | } else { | |
3971 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3972 | } |
ac9721f3 PZ |
3973 | return ret; |
3974 | } | |
a4be7c27 | 3975 | |
6fb2915d LZ |
3976 | case PERF_EVENT_IOC_SET_FILTER: |
3977 | return perf_event_set_filter(event, (void __user *)arg); | |
3978 | ||
d859e29f | 3979 | default: |
3df5edad | 3980 | return -ENOTTY; |
d859e29f | 3981 | } |
3df5edad PZ |
3982 | |
3983 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3984 | perf_event_for_each(event, func); |
3df5edad | 3985 | else |
cdd6c482 | 3986 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3987 | |
3988 | return 0; | |
d859e29f PM |
3989 | } |
3990 | ||
f63a8daa PZ |
3991 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3992 | { | |
3993 | struct perf_event *event = file->private_data; | |
3994 | struct perf_event_context *ctx; | |
3995 | long ret; | |
3996 | ||
3997 | ctx = perf_event_ctx_lock(event); | |
3998 | ret = _perf_ioctl(event, cmd, arg); | |
3999 | perf_event_ctx_unlock(event, ctx); | |
4000 | ||
4001 | return ret; | |
4002 | } | |
4003 | ||
b3f20785 PM |
4004 | #ifdef CONFIG_COMPAT |
4005 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, | |
4006 | unsigned long arg) | |
4007 | { | |
4008 | switch (_IOC_NR(cmd)) { | |
4009 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): | |
4010 | case _IOC_NR(PERF_EVENT_IOC_ID): | |
4011 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ | |
4012 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { | |
4013 | cmd &= ~IOCSIZE_MASK; | |
4014 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; | |
4015 | } | |
4016 | break; | |
4017 | } | |
4018 | return perf_ioctl(file, cmd, arg); | |
4019 | } | |
4020 | #else | |
4021 | # define perf_compat_ioctl NULL | |
4022 | #endif | |
4023 | ||
cdd6c482 | 4024 | int perf_event_task_enable(void) |
771d7cde | 4025 | { |
f63a8daa | 4026 | struct perf_event_context *ctx; |
cdd6c482 | 4027 | struct perf_event *event; |
771d7cde | 4028 | |
cdd6c482 | 4029 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4030 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4031 | ctx = perf_event_ctx_lock(event); | |
4032 | perf_event_for_each_child(event, _perf_event_enable); | |
4033 | perf_event_ctx_unlock(event, ctx); | |
4034 | } | |
cdd6c482 | 4035 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4036 | |
4037 | return 0; | |
4038 | } | |
4039 | ||
cdd6c482 | 4040 | int perf_event_task_disable(void) |
771d7cde | 4041 | { |
f63a8daa | 4042 | struct perf_event_context *ctx; |
cdd6c482 | 4043 | struct perf_event *event; |
771d7cde | 4044 | |
cdd6c482 | 4045 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4046 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4047 | ctx = perf_event_ctx_lock(event); | |
4048 | perf_event_for_each_child(event, _perf_event_disable); | |
4049 | perf_event_ctx_unlock(event, ctx); | |
4050 | } | |
cdd6c482 | 4051 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4052 | |
4053 | return 0; | |
4054 | } | |
4055 | ||
cdd6c482 | 4056 | static int perf_event_index(struct perf_event *event) |
194002b2 | 4057 | { |
a4eaf7f1 PZ |
4058 | if (event->hw.state & PERF_HES_STOPPED) |
4059 | return 0; | |
4060 | ||
cdd6c482 | 4061 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
4062 | return 0; |
4063 | ||
35edc2a5 | 4064 | return event->pmu->event_idx(event); |
194002b2 PZ |
4065 | } |
4066 | ||
c4794295 | 4067 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 4068 | u64 *now, |
7f310a5d EM |
4069 | u64 *enabled, |
4070 | u64 *running) | |
c4794295 | 4071 | { |
e3f3541c | 4072 | u64 ctx_time; |
c4794295 | 4073 | |
e3f3541c PZ |
4074 | *now = perf_clock(); |
4075 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
4076 | *enabled = ctx_time - event->tstamp_enabled; |
4077 | *running = ctx_time - event->tstamp_running; | |
4078 | } | |
4079 | ||
fa731587 PZ |
4080 | static void perf_event_init_userpage(struct perf_event *event) |
4081 | { | |
4082 | struct perf_event_mmap_page *userpg; | |
4083 | struct ring_buffer *rb; | |
4084 | ||
4085 | rcu_read_lock(); | |
4086 | rb = rcu_dereference(event->rb); | |
4087 | if (!rb) | |
4088 | goto unlock; | |
4089 | ||
4090 | userpg = rb->user_page; | |
4091 | ||
4092 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
4093 | userpg->cap_bit0_is_deprecated = 1; | |
4094 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
4095 | ||
4096 | unlock: | |
4097 | rcu_read_unlock(); | |
4098 | } | |
4099 | ||
c1317ec2 AL |
4100 | void __weak arch_perf_update_userpage( |
4101 | struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now) | |
e3f3541c PZ |
4102 | { |
4103 | } | |
4104 | ||
38ff667b PZ |
4105 | /* |
4106 | * Callers need to ensure there can be no nesting of this function, otherwise | |
4107 | * the seqlock logic goes bad. We can not serialize this because the arch | |
4108 | * code calls this from NMI context. | |
4109 | */ | |
cdd6c482 | 4110 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 4111 | { |
cdd6c482 | 4112 | struct perf_event_mmap_page *userpg; |
76369139 | 4113 | struct ring_buffer *rb; |
e3f3541c | 4114 | u64 enabled, running, now; |
38ff667b PZ |
4115 | |
4116 | rcu_read_lock(); | |
5ec4c599 PZ |
4117 | rb = rcu_dereference(event->rb); |
4118 | if (!rb) | |
4119 | goto unlock; | |
4120 | ||
0d641208 EM |
4121 | /* |
4122 | * compute total_time_enabled, total_time_running | |
4123 | * based on snapshot values taken when the event | |
4124 | * was last scheduled in. | |
4125 | * | |
4126 | * we cannot simply called update_context_time() | |
4127 | * because of locking issue as we can be called in | |
4128 | * NMI context | |
4129 | */ | |
e3f3541c | 4130 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 4131 | |
76369139 | 4132 | userpg = rb->user_page; |
7b732a75 PZ |
4133 | /* |
4134 | * Disable preemption so as to not let the corresponding user-space | |
4135 | * spin too long if we get preempted. | |
4136 | */ | |
4137 | preempt_disable(); | |
37d81828 | 4138 | ++userpg->lock; |
92f22a38 | 4139 | barrier(); |
cdd6c482 | 4140 | userpg->index = perf_event_index(event); |
b5e58793 | 4141 | userpg->offset = perf_event_count(event); |
365a4038 | 4142 | if (userpg->index) |
e7850595 | 4143 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 4144 | |
0d641208 | 4145 | userpg->time_enabled = enabled + |
cdd6c482 | 4146 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 4147 | |
0d641208 | 4148 | userpg->time_running = running + |
cdd6c482 | 4149 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 4150 | |
c1317ec2 | 4151 | arch_perf_update_userpage(event, userpg, now); |
e3f3541c | 4152 | |
92f22a38 | 4153 | barrier(); |
37d81828 | 4154 | ++userpg->lock; |
7b732a75 | 4155 | preempt_enable(); |
38ff667b | 4156 | unlock: |
7b732a75 | 4157 | rcu_read_unlock(); |
37d81828 PM |
4158 | } |
4159 | ||
906010b2 PZ |
4160 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
4161 | { | |
4162 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 4163 | struct ring_buffer *rb; |
906010b2 PZ |
4164 | int ret = VM_FAULT_SIGBUS; |
4165 | ||
4166 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
4167 | if (vmf->pgoff == 0) | |
4168 | ret = 0; | |
4169 | return ret; | |
4170 | } | |
4171 | ||
4172 | rcu_read_lock(); | |
76369139 FW |
4173 | rb = rcu_dereference(event->rb); |
4174 | if (!rb) | |
906010b2 PZ |
4175 | goto unlock; |
4176 | ||
4177 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
4178 | goto unlock; | |
4179 | ||
76369139 | 4180 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
4181 | if (!vmf->page) |
4182 | goto unlock; | |
4183 | ||
4184 | get_page(vmf->page); | |
4185 | vmf->page->mapping = vma->vm_file->f_mapping; | |
4186 | vmf->page->index = vmf->pgoff; | |
4187 | ||
4188 | ret = 0; | |
4189 | unlock: | |
4190 | rcu_read_unlock(); | |
4191 | ||
4192 | return ret; | |
4193 | } | |
4194 | ||
10c6db11 PZ |
4195 | static void ring_buffer_attach(struct perf_event *event, |
4196 | struct ring_buffer *rb) | |
4197 | { | |
b69cf536 | 4198 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
4199 | unsigned long flags; |
4200 | ||
b69cf536 PZ |
4201 | if (event->rb) { |
4202 | /* | |
4203 | * Should be impossible, we set this when removing | |
4204 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
4205 | */ | |
4206 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 4207 | |
b69cf536 PZ |
4208 | old_rb = event->rb; |
4209 | event->rcu_batches = get_state_synchronize_rcu(); | |
4210 | event->rcu_pending = 1; | |
10c6db11 | 4211 | |
b69cf536 PZ |
4212 | spin_lock_irqsave(&old_rb->event_lock, flags); |
4213 | list_del_rcu(&event->rb_entry); | |
4214 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
4215 | } | |
10c6db11 | 4216 | |
b69cf536 PZ |
4217 | if (event->rcu_pending && rb) { |
4218 | cond_synchronize_rcu(event->rcu_batches); | |
4219 | event->rcu_pending = 0; | |
4220 | } | |
10c6db11 | 4221 | |
b69cf536 PZ |
4222 | if (rb) { |
4223 | spin_lock_irqsave(&rb->event_lock, flags); | |
4224 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
4225 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
4226 | } | |
4227 | ||
4228 | rcu_assign_pointer(event->rb, rb); | |
4229 | ||
4230 | if (old_rb) { | |
4231 | ring_buffer_put(old_rb); | |
4232 | /* | |
4233 | * Since we detached before setting the new rb, so that we | |
4234 | * could attach the new rb, we could have missed a wakeup. | |
4235 | * Provide it now. | |
4236 | */ | |
4237 | wake_up_all(&event->waitq); | |
4238 | } | |
10c6db11 PZ |
4239 | } |
4240 | ||
4241 | static void ring_buffer_wakeup(struct perf_event *event) | |
4242 | { | |
4243 | struct ring_buffer *rb; | |
4244 | ||
4245 | rcu_read_lock(); | |
4246 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
4247 | if (rb) { |
4248 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
4249 | wake_up_all(&event->waitq); | |
4250 | } | |
10c6db11 PZ |
4251 | rcu_read_unlock(); |
4252 | } | |
4253 | ||
76369139 | 4254 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 4255 | { |
76369139 | 4256 | struct ring_buffer *rb; |
906010b2 | 4257 | |
76369139 FW |
4258 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
4259 | rb_free(rb); | |
7b732a75 PZ |
4260 | } |
4261 | ||
76369139 | 4262 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 4263 | { |
76369139 | 4264 | struct ring_buffer *rb; |
7b732a75 | 4265 | |
ac9721f3 | 4266 | rcu_read_lock(); |
76369139 FW |
4267 | rb = rcu_dereference(event->rb); |
4268 | if (rb) { | |
4269 | if (!atomic_inc_not_zero(&rb->refcount)) | |
4270 | rb = NULL; | |
ac9721f3 PZ |
4271 | } |
4272 | rcu_read_unlock(); | |
4273 | ||
76369139 | 4274 | return rb; |
ac9721f3 PZ |
4275 | } |
4276 | ||
76369139 | 4277 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 4278 | { |
76369139 | 4279 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 4280 | return; |
7b732a75 | 4281 | |
9bb5d40c | 4282 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 4283 | |
76369139 | 4284 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
4285 | } |
4286 | ||
4287 | static void perf_mmap_open(struct vm_area_struct *vma) | |
4288 | { | |
cdd6c482 | 4289 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4290 | |
cdd6c482 | 4291 | atomic_inc(&event->mmap_count); |
9bb5d40c | 4292 | atomic_inc(&event->rb->mmap_count); |
1e0fb9ec AL |
4293 | |
4294 | if (event->pmu->event_mapped) | |
4295 | event->pmu->event_mapped(event); | |
7b732a75 PZ |
4296 | } |
4297 | ||
9bb5d40c PZ |
4298 | /* |
4299 | * A buffer can be mmap()ed multiple times; either directly through the same | |
4300 | * event, or through other events by use of perf_event_set_output(). | |
4301 | * | |
4302 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
4303 | * the buffer here, where we still have a VM context. This means we need | |
4304 | * to detach all events redirecting to us. | |
4305 | */ | |
7b732a75 PZ |
4306 | static void perf_mmap_close(struct vm_area_struct *vma) |
4307 | { | |
cdd6c482 | 4308 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4309 | |
b69cf536 | 4310 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
4311 | struct user_struct *mmap_user = rb->mmap_user; |
4312 | int mmap_locked = rb->mmap_locked; | |
4313 | unsigned long size = perf_data_size(rb); | |
789f90fc | 4314 | |
1e0fb9ec AL |
4315 | if (event->pmu->event_unmapped) |
4316 | event->pmu->event_unmapped(event); | |
4317 | ||
9bb5d40c PZ |
4318 | atomic_dec(&rb->mmap_count); |
4319 | ||
4320 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4321 | goto out_put; |
9bb5d40c | 4322 | |
b69cf536 | 4323 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4324 | mutex_unlock(&event->mmap_mutex); |
4325 | ||
4326 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4327 | if (atomic_read(&rb->mmap_count)) |
4328 | goto out_put; | |
ac9721f3 | 4329 | |
9bb5d40c PZ |
4330 | /* |
4331 | * No other mmap()s, detach from all other events that might redirect | |
4332 | * into the now unreachable buffer. Somewhat complicated by the | |
4333 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4334 | */ | |
4335 | again: | |
4336 | rcu_read_lock(); | |
4337 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4338 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4339 | /* | |
4340 | * This event is en-route to free_event() which will | |
4341 | * detach it and remove it from the list. | |
4342 | */ | |
4343 | continue; | |
4344 | } | |
4345 | rcu_read_unlock(); | |
789f90fc | 4346 | |
9bb5d40c PZ |
4347 | mutex_lock(&event->mmap_mutex); |
4348 | /* | |
4349 | * Check we didn't race with perf_event_set_output() which can | |
4350 | * swizzle the rb from under us while we were waiting to | |
4351 | * acquire mmap_mutex. | |
4352 | * | |
4353 | * If we find a different rb; ignore this event, a next | |
4354 | * iteration will no longer find it on the list. We have to | |
4355 | * still restart the iteration to make sure we're not now | |
4356 | * iterating the wrong list. | |
4357 | */ | |
b69cf536 PZ |
4358 | if (event->rb == rb) |
4359 | ring_buffer_attach(event, NULL); | |
4360 | ||
cdd6c482 | 4361 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4362 | put_event(event); |
ac9721f3 | 4363 | |
9bb5d40c PZ |
4364 | /* |
4365 | * Restart the iteration; either we're on the wrong list or | |
4366 | * destroyed its integrity by doing a deletion. | |
4367 | */ | |
4368 | goto again; | |
7b732a75 | 4369 | } |
9bb5d40c PZ |
4370 | rcu_read_unlock(); |
4371 | ||
4372 | /* | |
4373 | * It could be there's still a few 0-ref events on the list; they'll | |
4374 | * get cleaned up by free_event() -- they'll also still have their | |
4375 | * ref on the rb and will free it whenever they are done with it. | |
4376 | * | |
4377 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4378 | * undo the VM accounting. | |
4379 | */ | |
4380 | ||
4381 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4382 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4383 | free_uid(mmap_user); | |
4384 | ||
b69cf536 | 4385 | out_put: |
9bb5d40c | 4386 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4387 | } |
4388 | ||
f0f37e2f | 4389 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
4390 | .open = perf_mmap_open, |
4391 | .close = perf_mmap_close, | |
4392 | .fault = perf_mmap_fault, | |
4393 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4394 | }; |
4395 | ||
4396 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4397 | { | |
cdd6c482 | 4398 | struct perf_event *event = file->private_data; |
22a4f650 | 4399 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4400 | struct user_struct *user = current_user(); |
22a4f650 | 4401 | unsigned long locked, lock_limit; |
76369139 | 4402 | struct ring_buffer *rb; |
7b732a75 PZ |
4403 | unsigned long vma_size; |
4404 | unsigned long nr_pages; | |
789f90fc | 4405 | long user_extra, extra; |
d57e34fd | 4406 | int ret = 0, flags = 0; |
37d81828 | 4407 | |
c7920614 PZ |
4408 | /* |
4409 | * Don't allow mmap() of inherited per-task counters. This would | |
4410 | * create a performance issue due to all children writing to the | |
76369139 | 4411 | * same rb. |
c7920614 PZ |
4412 | */ |
4413 | if (event->cpu == -1 && event->attr.inherit) | |
4414 | return -EINVAL; | |
4415 | ||
43a21ea8 | 4416 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4417 | return -EINVAL; |
7b732a75 PZ |
4418 | |
4419 | vma_size = vma->vm_end - vma->vm_start; | |
4420 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4421 | ||
7730d865 | 4422 | /* |
76369139 | 4423 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4424 | * can do bitmasks instead of modulo. |
4425 | */ | |
2ed11312 | 4426 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) |
37d81828 PM |
4427 | return -EINVAL; |
4428 | ||
7b732a75 | 4429 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4430 | return -EINVAL; |
4431 | ||
7b732a75 PZ |
4432 | if (vma->vm_pgoff != 0) |
4433 | return -EINVAL; | |
37d81828 | 4434 | |
cdd6c482 | 4435 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4436 | again: |
cdd6c482 | 4437 | mutex_lock(&event->mmap_mutex); |
76369139 | 4438 | if (event->rb) { |
9bb5d40c | 4439 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4440 | ret = -EINVAL; |
9bb5d40c PZ |
4441 | goto unlock; |
4442 | } | |
4443 | ||
4444 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4445 | /* | |
4446 | * Raced against perf_mmap_close() through | |
4447 | * perf_event_set_output(). Try again, hope for better | |
4448 | * luck. | |
4449 | */ | |
4450 | mutex_unlock(&event->mmap_mutex); | |
4451 | goto again; | |
4452 | } | |
4453 | ||
ebb3c4c4 PZ |
4454 | goto unlock; |
4455 | } | |
4456 | ||
789f90fc | 4457 | user_extra = nr_pages + 1; |
cdd6c482 | 4458 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4459 | |
4460 | /* | |
4461 | * Increase the limit linearly with more CPUs: | |
4462 | */ | |
4463 | user_lock_limit *= num_online_cpus(); | |
4464 | ||
789f90fc | 4465 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4466 | |
789f90fc PZ |
4467 | extra = 0; |
4468 | if (user_locked > user_lock_limit) | |
4469 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4470 | |
78d7d407 | 4471 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4472 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4473 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4474 | |
459ec28a IM |
4475 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4476 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4477 | ret = -EPERM; |
4478 | goto unlock; | |
4479 | } | |
7b732a75 | 4480 | |
76369139 | 4481 | WARN_ON(event->rb); |
906010b2 | 4482 | |
d57e34fd | 4483 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4484 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4485 | |
4ec8363d VW |
4486 | rb = rb_alloc(nr_pages, |
4487 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4488 | event->cpu, flags); | |
4489 | ||
76369139 | 4490 | if (!rb) { |
ac9721f3 | 4491 | ret = -ENOMEM; |
ebb3c4c4 | 4492 | goto unlock; |
ac9721f3 | 4493 | } |
26cb63ad | 4494 | |
9bb5d40c | 4495 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4496 | rb->mmap_locked = extra; |
4497 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4498 | |
ac9721f3 | 4499 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4500 | vma->vm_mm->pinned_vm += extra; |
4501 | ||
9bb5d40c | 4502 | ring_buffer_attach(event, rb); |
ac9721f3 | 4503 | |
fa731587 | 4504 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4505 | perf_event_update_userpage(event); |
4506 | ||
ebb3c4c4 | 4507 | unlock: |
ac9721f3 PZ |
4508 | if (!ret) |
4509 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4510 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4511 | |
9bb5d40c PZ |
4512 | /* |
4513 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4514 | * vma. | |
4515 | */ | |
26cb63ad | 4516 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4517 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 | 4518 | |
1e0fb9ec AL |
4519 | if (event->pmu->event_mapped) |
4520 | event->pmu->event_mapped(event); | |
4521 | ||
7b732a75 | 4522 | return ret; |
37d81828 PM |
4523 | } |
4524 | ||
3c446b3d PZ |
4525 | static int perf_fasync(int fd, struct file *filp, int on) |
4526 | { | |
496ad9aa | 4527 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4528 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4529 | int retval; |
4530 | ||
4531 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4532 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4533 | mutex_unlock(&inode->i_mutex); |
4534 | ||
4535 | if (retval < 0) | |
4536 | return retval; | |
4537 | ||
4538 | return 0; | |
4539 | } | |
4540 | ||
0793a61d | 4541 | static const struct file_operations perf_fops = { |
3326c1ce | 4542 | .llseek = no_llseek, |
0793a61d TG |
4543 | .release = perf_release, |
4544 | .read = perf_read, | |
4545 | .poll = perf_poll, | |
d859e29f | 4546 | .unlocked_ioctl = perf_ioctl, |
b3f20785 | 4547 | .compat_ioctl = perf_compat_ioctl, |
37d81828 | 4548 | .mmap = perf_mmap, |
3c446b3d | 4549 | .fasync = perf_fasync, |
0793a61d TG |
4550 | }; |
4551 | ||
925d519a | 4552 | /* |
cdd6c482 | 4553 | * Perf event wakeup |
925d519a PZ |
4554 | * |
4555 | * If there's data, ensure we set the poll() state and publish everything | |
4556 | * to user-space before waking everybody up. | |
4557 | */ | |
4558 | ||
cdd6c482 | 4559 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4560 | { |
10c6db11 | 4561 | ring_buffer_wakeup(event); |
4c9e2542 | 4562 | |
cdd6c482 IM |
4563 | if (event->pending_kill) { |
4564 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4565 | event->pending_kill = 0; | |
4c9e2542 | 4566 | } |
925d519a PZ |
4567 | } |
4568 | ||
e360adbe | 4569 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4570 | { |
cdd6c482 IM |
4571 | struct perf_event *event = container_of(entry, |
4572 | struct perf_event, pending); | |
d525211f PZ |
4573 | int rctx; |
4574 | ||
4575 | rctx = perf_swevent_get_recursion_context(); | |
4576 | /* | |
4577 | * If we 'fail' here, that's OK, it means recursion is already disabled | |
4578 | * and we won't recurse 'further'. | |
4579 | */ | |
79f14641 | 4580 | |
cdd6c482 IM |
4581 | if (event->pending_disable) { |
4582 | event->pending_disable = 0; | |
4583 | __perf_event_disable(event); | |
79f14641 PZ |
4584 | } |
4585 | ||
cdd6c482 IM |
4586 | if (event->pending_wakeup) { |
4587 | event->pending_wakeup = 0; | |
4588 | perf_event_wakeup(event); | |
79f14641 | 4589 | } |
d525211f PZ |
4590 | |
4591 | if (rctx >= 0) | |
4592 | perf_swevent_put_recursion_context(rctx); | |
79f14641 PZ |
4593 | } |
4594 | ||
39447b38 ZY |
4595 | /* |
4596 | * We assume there is only KVM supporting the callbacks. | |
4597 | * Later on, we might change it to a list if there is | |
4598 | * another virtualization implementation supporting the callbacks. | |
4599 | */ | |
4600 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4601 | ||
4602 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4603 | { | |
4604 | perf_guest_cbs = cbs; | |
4605 | return 0; | |
4606 | } | |
4607 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4608 | ||
4609 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4610 | { | |
4611 | perf_guest_cbs = NULL; | |
4612 | return 0; | |
4613 | } | |
4614 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4615 | ||
4018994f JO |
4616 | static void |
4617 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4618 | struct pt_regs *regs, u64 mask) | |
4619 | { | |
4620 | int bit; | |
4621 | ||
4622 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4623 | sizeof(mask) * BITS_PER_BYTE) { | |
4624 | u64 val; | |
4625 | ||
4626 | val = perf_reg_value(regs, bit); | |
4627 | perf_output_put(handle, val); | |
4628 | } | |
4629 | } | |
4630 | ||
60e2364e | 4631 | static void perf_sample_regs_user(struct perf_regs *regs_user, |
88a7c26a AL |
4632 | struct pt_regs *regs, |
4633 | struct pt_regs *regs_user_copy) | |
4018994f | 4634 | { |
88a7c26a AL |
4635 | if (user_mode(regs)) { |
4636 | regs_user->abi = perf_reg_abi(current); | |
2565711f | 4637 | regs_user->regs = regs; |
88a7c26a AL |
4638 | } else if (current->mm) { |
4639 | perf_get_regs_user(regs_user, regs, regs_user_copy); | |
2565711f PZ |
4640 | } else { |
4641 | regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; | |
4642 | regs_user->regs = NULL; | |
4018994f JO |
4643 | } |
4644 | } | |
4645 | ||
60e2364e SE |
4646 | static void perf_sample_regs_intr(struct perf_regs *regs_intr, |
4647 | struct pt_regs *regs) | |
4648 | { | |
4649 | regs_intr->regs = regs; | |
4650 | regs_intr->abi = perf_reg_abi(current); | |
4651 | } | |
4652 | ||
4653 | ||
c5ebcedb JO |
4654 | /* |
4655 | * Get remaining task size from user stack pointer. | |
4656 | * | |
4657 | * It'd be better to take stack vma map and limit this more | |
4658 | * precisly, but there's no way to get it safely under interrupt, | |
4659 | * so using TASK_SIZE as limit. | |
4660 | */ | |
4661 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4662 | { | |
4663 | unsigned long addr = perf_user_stack_pointer(regs); | |
4664 | ||
4665 | if (!addr || addr >= TASK_SIZE) | |
4666 | return 0; | |
4667 | ||
4668 | return TASK_SIZE - addr; | |
4669 | } | |
4670 | ||
4671 | static u16 | |
4672 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4673 | struct pt_regs *regs) | |
4674 | { | |
4675 | u64 task_size; | |
4676 | ||
4677 | /* No regs, no stack pointer, no dump. */ | |
4678 | if (!regs) | |
4679 | return 0; | |
4680 | ||
4681 | /* | |
4682 | * Check if we fit in with the requested stack size into the: | |
4683 | * - TASK_SIZE | |
4684 | * If we don't, we limit the size to the TASK_SIZE. | |
4685 | * | |
4686 | * - remaining sample size | |
4687 | * If we don't, we customize the stack size to | |
4688 | * fit in to the remaining sample size. | |
4689 | */ | |
4690 | ||
4691 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4692 | stack_size = min(stack_size, (u16) task_size); | |
4693 | ||
4694 | /* Current header size plus static size and dynamic size. */ | |
4695 | header_size += 2 * sizeof(u64); | |
4696 | ||
4697 | /* Do we fit in with the current stack dump size? */ | |
4698 | if ((u16) (header_size + stack_size) < header_size) { | |
4699 | /* | |
4700 | * If we overflow the maximum size for the sample, | |
4701 | * we customize the stack dump size to fit in. | |
4702 | */ | |
4703 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4704 | stack_size = round_up(stack_size, sizeof(u64)); | |
4705 | } | |
4706 | ||
4707 | return stack_size; | |
4708 | } | |
4709 | ||
4710 | static void | |
4711 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4712 | struct pt_regs *regs) | |
4713 | { | |
4714 | /* Case of a kernel thread, nothing to dump */ | |
4715 | if (!regs) { | |
4716 | u64 size = 0; | |
4717 | perf_output_put(handle, size); | |
4718 | } else { | |
4719 | unsigned long sp; | |
4720 | unsigned int rem; | |
4721 | u64 dyn_size; | |
4722 | ||
4723 | /* | |
4724 | * We dump: | |
4725 | * static size | |
4726 | * - the size requested by user or the best one we can fit | |
4727 | * in to the sample max size | |
4728 | * data | |
4729 | * - user stack dump data | |
4730 | * dynamic size | |
4731 | * - the actual dumped size | |
4732 | */ | |
4733 | ||
4734 | /* Static size. */ | |
4735 | perf_output_put(handle, dump_size); | |
4736 | ||
4737 | /* Data. */ | |
4738 | sp = perf_user_stack_pointer(regs); | |
4739 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4740 | dyn_size = dump_size - rem; | |
4741 | ||
4742 | perf_output_skip(handle, rem); | |
4743 | ||
4744 | /* Dynamic size. */ | |
4745 | perf_output_put(handle, dyn_size); | |
4746 | } | |
4747 | } | |
4748 | ||
c980d109 ACM |
4749 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4750 | struct perf_sample_data *data, | |
4751 | struct perf_event *event) | |
6844c09d ACM |
4752 | { |
4753 | u64 sample_type = event->attr.sample_type; | |
4754 | ||
4755 | data->type = sample_type; | |
4756 | header->size += event->id_header_size; | |
4757 | ||
4758 | if (sample_type & PERF_SAMPLE_TID) { | |
4759 | /* namespace issues */ | |
4760 | data->tid_entry.pid = perf_event_pid(event, current); | |
4761 | data->tid_entry.tid = perf_event_tid(event, current); | |
4762 | } | |
4763 | ||
4764 | if (sample_type & PERF_SAMPLE_TIME) | |
4765 | data->time = perf_clock(); | |
4766 | ||
ff3d527c | 4767 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4768 | data->id = primary_event_id(event); |
4769 | ||
4770 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4771 | data->stream_id = event->id; | |
4772 | ||
4773 | if (sample_type & PERF_SAMPLE_CPU) { | |
4774 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4775 | data->cpu_entry.reserved = 0; | |
4776 | } | |
4777 | } | |
4778 | ||
76369139 FW |
4779 | void perf_event_header__init_id(struct perf_event_header *header, |
4780 | struct perf_sample_data *data, | |
4781 | struct perf_event *event) | |
c980d109 ACM |
4782 | { |
4783 | if (event->attr.sample_id_all) | |
4784 | __perf_event_header__init_id(header, data, event); | |
4785 | } | |
4786 | ||
4787 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4788 | struct perf_sample_data *data) | |
4789 | { | |
4790 | u64 sample_type = data->type; | |
4791 | ||
4792 | if (sample_type & PERF_SAMPLE_TID) | |
4793 | perf_output_put(handle, data->tid_entry); | |
4794 | ||
4795 | if (sample_type & PERF_SAMPLE_TIME) | |
4796 | perf_output_put(handle, data->time); | |
4797 | ||
4798 | if (sample_type & PERF_SAMPLE_ID) | |
4799 | perf_output_put(handle, data->id); | |
4800 | ||
4801 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4802 | perf_output_put(handle, data->stream_id); | |
4803 | ||
4804 | if (sample_type & PERF_SAMPLE_CPU) | |
4805 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4806 | |
4807 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4808 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4809 | } |
4810 | ||
76369139 FW |
4811 | void perf_event__output_id_sample(struct perf_event *event, |
4812 | struct perf_output_handle *handle, | |
4813 | struct perf_sample_data *sample) | |
c980d109 ACM |
4814 | { |
4815 | if (event->attr.sample_id_all) | |
4816 | __perf_event__output_id_sample(handle, sample); | |
4817 | } | |
4818 | ||
3dab77fb | 4819 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4820 | struct perf_event *event, |
4821 | u64 enabled, u64 running) | |
3dab77fb | 4822 | { |
cdd6c482 | 4823 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4824 | u64 values[4]; |
4825 | int n = 0; | |
4826 | ||
b5e58793 | 4827 | values[n++] = perf_event_count(event); |
3dab77fb | 4828 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4829 | values[n++] = enabled + |
cdd6c482 | 4830 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4831 | } |
4832 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4833 | values[n++] = running + |
cdd6c482 | 4834 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4835 | } |
4836 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4837 | values[n++] = primary_event_id(event); |
3dab77fb | 4838 | |
76369139 | 4839 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4840 | } |
4841 | ||
4842 | /* | |
cdd6c482 | 4843 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4844 | */ |
4845 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4846 | struct perf_event *event, |
4847 | u64 enabled, u64 running) | |
3dab77fb | 4848 | { |
cdd6c482 IM |
4849 | struct perf_event *leader = event->group_leader, *sub; |
4850 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4851 | u64 values[5]; |
4852 | int n = 0; | |
4853 | ||
4854 | values[n++] = 1 + leader->nr_siblings; | |
4855 | ||
4856 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4857 | values[n++] = enabled; |
3dab77fb PZ |
4858 | |
4859 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4860 | values[n++] = running; |
3dab77fb | 4861 | |
cdd6c482 | 4862 | if (leader != event) |
3dab77fb PZ |
4863 | leader->pmu->read(leader); |
4864 | ||
b5e58793 | 4865 | values[n++] = perf_event_count(leader); |
3dab77fb | 4866 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4867 | values[n++] = primary_event_id(leader); |
3dab77fb | 4868 | |
76369139 | 4869 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4870 | |
65abc865 | 4871 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4872 | n = 0; |
4873 | ||
6f5ab001 JO |
4874 | if ((sub != event) && |
4875 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4876 | sub->pmu->read(sub); |
4877 | ||
b5e58793 | 4878 | values[n++] = perf_event_count(sub); |
3dab77fb | 4879 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4880 | values[n++] = primary_event_id(sub); |
3dab77fb | 4881 | |
76369139 | 4882 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4883 | } |
4884 | } | |
4885 | ||
eed01528 SE |
4886 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4887 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4888 | ||
3dab77fb | 4889 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4890 | struct perf_event *event) |
3dab77fb | 4891 | { |
e3f3541c | 4892 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4893 | u64 read_format = event->attr.read_format; |
4894 | ||
4895 | /* | |
4896 | * compute total_time_enabled, total_time_running | |
4897 | * based on snapshot values taken when the event | |
4898 | * was last scheduled in. | |
4899 | * | |
4900 | * we cannot simply called update_context_time() | |
4901 | * because of locking issue as we are called in | |
4902 | * NMI context | |
4903 | */ | |
c4794295 | 4904 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4905 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4906 | |
cdd6c482 | 4907 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4908 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4909 | else |
eed01528 | 4910 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4911 | } |
4912 | ||
5622f295 MM |
4913 | void perf_output_sample(struct perf_output_handle *handle, |
4914 | struct perf_event_header *header, | |
4915 | struct perf_sample_data *data, | |
cdd6c482 | 4916 | struct perf_event *event) |
5622f295 MM |
4917 | { |
4918 | u64 sample_type = data->type; | |
4919 | ||
4920 | perf_output_put(handle, *header); | |
4921 | ||
ff3d527c AH |
4922 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4923 | perf_output_put(handle, data->id); | |
4924 | ||
5622f295 MM |
4925 | if (sample_type & PERF_SAMPLE_IP) |
4926 | perf_output_put(handle, data->ip); | |
4927 | ||
4928 | if (sample_type & PERF_SAMPLE_TID) | |
4929 | perf_output_put(handle, data->tid_entry); | |
4930 | ||
4931 | if (sample_type & PERF_SAMPLE_TIME) | |
4932 | perf_output_put(handle, data->time); | |
4933 | ||
4934 | if (sample_type & PERF_SAMPLE_ADDR) | |
4935 | perf_output_put(handle, data->addr); | |
4936 | ||
4937 | if (sample_type & PERF_SAMPLE_ID) | |
4938 | perf_output_put(handle, data->id); | |
4939 | ||
4940 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4941 | perf_output_put(handle, data->stream_id); | |
4942 | ||
4943 | if (sample_type & PERF_SAMPLE_CPU) | |
4944 | perf_output_put(handle, data->cpu_entry); | |
4945 | ||
4946 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4947 | perf_output_put(handle, data->period); | |
4948 | ||
4949 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4950 | perf_output_read(handle, event); |
5622f295 MM |
4951 | |
4952 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4953 | if (data->callchain) { | |
4954 | int size = 1; | |
4955 | ||
4956 | if (data->callchain) | |
4957 | size += data->callchain->nr; | |
4958 | ||
4959 | size *= sizeof(u64); | |
4960 | ||
76369139 | 4961 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4962 | } else { |
4963 | u64 nr = 0; | |
4964 | perf_output_put(handle, nr); | |
4965 | } | |
4966 | } | |
4967 | ||
4968 | if (sample_type & PERF_SAMPLE_RAW) { | |
4969 | if (data->raw) { | |
4970 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4971 | __output_copy(handle, data->raw->data, |
4972 | data->raw->size); | |
5622f295 MM |
4973 | } else { |
4974 | struct { | |
4975 | u32 size; | |
4976 | u32 data; | |
4977 | } raw = { | |
4978 | .size = sizeof(u32), | |
4979 | .data = 0, | |
4980 | }; | |
4981 | perf_output_put(handle, raw); | |
4982 | } | |
4983 | } | |
a7ac67ea | 4984 | |
bce38cd5 SE |
4985 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4986 | if (data->br_stack) { | |
4987 | size_t size; | |
4988 | ||
4989 | size = data->br_stack->nr | |
4990 | * sizeof(struct perf_branch_entry); | |
4991 | ||
4992 | perf_output_put(handle, data->br_stack->nr); | |
4993 | perf_output_copy(handle, data->br_stack->entries, size); | |
4994 | } else { | |
4995 | /* | |
4996 | * we always store at least the value of nr | |
4997 | */ | |
4998 | u64 nr = 0; | |
4999 | perf_output_put(handle, nr); | |
5000 | } | |
5001 | } | |
4018994f JO |
5002 | |
5003 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
5004 | u64 abi = data->regs_user.abi; | |
5005 | ||
5006 | /* | |
5007 | * If there are no regs to dump, notice it through | |
5008 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5009 | */ | |
5010 | perf_output_put(handle, abi); | |
5011 | ||
5012 | if (abi) { | |
5013 | u64 mask = event->attr.sample_regs_user; | |
5014 | perf_output_sample_regs(handle, | |
5015 | data->regs_user.regs, | |
5016 | mask); | |
5017 | } | |
5018 | } | |
c5ebcedb | 5019 | |
a5cdd40c | 5020 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
5021 | perf_output_sample_ustack(handle, |
5022 | data->stack_user_size, | |
5023 | data->regs_user.regs); | |
a5cdd40c | 5024 | } |
c3feedf2 AK |
5025 | |
5026 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
5027 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
5028 | |
5029 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
5030 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 5031 | |
fdfbbd07 AK |
5032 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
5033 | perf_output_put(handle, data->txn); | |
5034 | ||
60e2364e SE |
5035 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
5036 | u64 abi = data->regs_intr.abi; | |
5037 | /* | |
5038 | * If there are no regs to dump, notice it through | |
5039 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5040 | */ | |
5041 | perf_output_put(handle, abi); | |
5042 | ||
5043 | if (abi) { | |
5044 | u64 mask = event->attr.sample_regs_intr; | |
5045 | ||
5046 | perf_output_sample_regs(handle, | |
5047 | data->regs_intr.regs, | |
5048 | mask); | |
5049 | } | |
5050 | } | |
5051 | ||
a5cdd40c PZ |
5052 | if (!event->attr.watermark) { |
5053 | int wakeup_events = event->attr.wakeup_events; | |
5054 | ||
5055 | if (wakeup_events) { | |
5056 | struct ring_buffer *rb = handle->rb; | |
5057 | int events = local_inc_return(&rb->events); | |
5058 | ||
5059 | if (events >= wakeup_events) { | |
5060 | local_sub(wakeup_events, &rb->events); | |
5061 | local_inc(&rb->wakeup); | |
5062 | } | |
5063 | } | |
5064 | } | |
5622f295 MM |
5065 | } |
5066 | ||
5067 | void perf_prepare_sample(struct perf_event_header *header, | |
5068 | struct perf_sample_data *data, | |
cdd6c482 | 5069 | struct perf_event *event, |
5622f295 | 5070 | struct pt_regs *regs) |
7b732a75 | 5071 | { |
cdd6c482 | 5072 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 5073 | |
cdd6c482 | 5074 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 5075 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
5076 | |
5077 | header->misc = 0; | |
5078 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 5079 | |
c980d109 | 5080 | __perf_event_header__init_id(header, data, event); |
6844c09d | 5081 | |
c320c7b7 | 5082 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
5083 | data->ip = perf_instruction_pointer(regs); |
5084 | ||
b23f3325 | 5085 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 5086 | int size = 1; |
394ee076 | 5087 | |
e6dab5ff | 5088 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
5089 | |
5090 | if (data->callchain) | |
5091 | size += data->callchain->nr; | |
5092 | ||
5093 | header->size += size * sizeof(u64); | |
394ee076 PZ |
5094 | } |
5095 | ||
3a43ce68 | 5096 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
5097 | int size = sizeof(u32); |
5098 | ||
5099 | if (data->raw) | |
5100 | size += data->raw->size; | |
5101 | else | |
5102 | size += sizeof(u32); | |
5103 | ||
5104 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 5105 | header->size += size; |
7f453c24 | 5106 | } |
bce38cd5 SE |
5107 | |
5108 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
5109 | int size = sizeof(u64); /* nr */ | |
5110 | if (data->br_stack) { | |
5111 | size += data->br_stack->nr | |
5112 | * sizeof(struct perf_branch_entry); | |
5113 | } | |
5114 | header->size += size; | |
5115 | } | |
4018994f | 5116 | |
2565711f | 5117 | if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER)) |
88a7c26a AL |
5118 | perf_sample_regs_user(&data->regs_user, regs, |
5119 | &data->regs_user_copy); | |
2565711f | 5120 | |
4018994f JO |
5121 | if (sample_type & PERF_SAMPLE_REGS_USER) { |
5122 | /* regs dump ABI info */ | |
5123 | int size = sizeof(u64); | |
5124 | ||
4018994f JO |
5125 | if (data->regs_user.regs) { |
5126 | u64 mask = event->attr.sample_regs_user; | |
5127 | size += hweight64(mask) * sizeof(u64); | |
5128 | } | |
5129 | ||
5130 | header->size += size; | |
5131 | } | |
c5ebcedb JO |
5132 | |
5133 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
5134 | /* | |
5135 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
5136 | * processed as the last one or have additional check added | |
5137 | * in case new sample type is added, because we could eat | |
5138 | * up the rest of the sample size. | |
5139 | */ | |
c5ebcedb JO |
5140 | u16 stack_size = event->attr.sample_stack_user; |
5141 | u16 size = sizeof(u64); | |
5142 | ||
c5ebcedb | 5143 | stack_size = perf_sample_ustack_size(stack_size, header->size, |
2565711f | 5144 | data->regs_user.regs); |
c5ebcedb JO |
5145 | |
5146 | /* | |
5147 | * If there is something to dump, add space for the dump | |
5148 | * itself and for the field that tells the dynamic size, | |
5149 | * which is how many have been actually dumped. | |
5150 | */ | |
5151 | if (stack_size) | |
5152 | size += sizeof(u64) + stack_size; | |
5153 | ||
5154 | data->stack_user_size = stack_size; | |
5155 | header->size += size; | |
5156 | } | |
60e2364e SE |
5157 | |
5158 | if (sample_type & PERF_SAMPLE_REGS_INTR) { | |
5159 | /* regs dump ABI info */ | |
5160 | int size = sizeof(u64); | |
5161 | ||
5162 | perf_sample_regs_intr(&data->regs_intr, regs); | |
5163 | ||
5164 | if (data->regs_intr.regs) { | |
5165 | u64 mask = event->attr.sample_regs_intr; | |
5166 | ||
5167 | size += hweight64(mask) * sizeof(u64); | |
5168 | } | |
5169 | ||
5170 | header->size += size; | |
5171 | } | |
5622f295 | 5172 | } |
7f453c24 | 5173 | |
a8b0ca17 | 5174 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
5175 | struct perf_sample_data *data, |
5176 | struct pt_regs *regs) | |
5177 | { | |
5178 | struct perf_output_handle handle; | |
5179 | struct perf_event_header header; | |
689802b2 | 5180 | |
927c7a9e FW |
5181 | /* protect the callchain buffers */ |
5182 | rcu_read_lock(); | |
5183 | ||
cdd6c482 | 5184 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 5185 | |
a7ac67ea | 5186 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 5187 | goto exit; |
0322cd6e | 5188 | |
cdd6c482 | 5189 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 5190 | |
8a057d84 | 5191 | perf_output_end(&handle); |
927c7a9e FW |
5192 | |
5193 | exit: | |
5194 | rcu_read_unlock(); | |
0322cd6e PZ |
5195 | } |
5196 | ||
38b200d6 | 5197 | /* |
cdd6c482 | 5198 | * read event_id |
38b200d6 PZ |
5199 | */ |
5200 | ||
5201 | struct perf_read_event { | |
5202 | struct perf_event_header header; | |
5203 | ||
5204 | u32 pid; | |
5205 | u32 tid; | |
38b200d6 PZ |
5206 | }; |
5207 | ||
5208 | static void | |
cdd6c482 | 5209 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
5210 | struct task_struct *task) |
5211 | { | |
5212 | struct perf_output_handle handle; | |
c980d109 | 5213 | struct perf_sample_data sample; |
dfc65094 | 5214 | struct perf_read_event read_event = { |
38b200d6 | 5215 | .header = { |
cdd6c482 | 5216 | .type = PERF_RECORD_READ, |
38b200d6 | 5217 | .misc = 0, |
c320c7b7 | 5218 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 5219 | }, |
cdd6c482 IM |
5220 | .pid = perf_event_pid(event, task), |
5221 | .tid = perf_event_tid(event, task), | |
38b200d6 | 5222 | }; |
3dab77fb | 5223 | int ret; |
38b200d6 | 5224 | |
c980d109 | 5225 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 5226 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
5227 | if (ret) |
5228 | return; | |
5229 | ||
dfc65094 | 5230 | perf_output_put(&handle, read_event); |
cdd6c482 | 5231 | perf_output_read(&handle, event); |
c980d109 | 5232 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 5233 | |
38b200d6 PZ |
5234 | perf_output_end(&handle); |
5235 | } | |
5236 | ||
52d857a8 JO |
5237 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
5238 | ||
5239 | static void | |
5240 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
5241 | perf_event_aux_output_cb output, |
5242 | void *data) | |
5243 | { | |
5244 | struct perf_event *event; | |
5245 | ||
5246 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5247 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
5248 | continue; | |
5249 | if (!event_filter_match(event)) | |
5250 | continue; | |
67516844 | 5251 | output(event, data); |
52d857a8 JO |
5252 | } |
5253 | } | |
5254 | ||
5255 | static void | |
67516844 | 5256 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
5257 | struct perf_event_context *task_ctx) |
5258 | { | |
5259 | struct perf_cpu_context *cpuctx; | |
5260 | struct perf_event_context *ctx; | |
5261 | struct pmu *pmu; | |
5262 | int ctxn; | |
5263 | ||
5264 | rcu_read_lock(); | |
5265 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
5266 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
5267 | if (cpuctx->unique_pmu != pmu) | |
5268 | goto next; | |
67516844 | 5269 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
5270 | if (task_ctx) |
5271 | goto next; | |
5272 | ctxn = pmu->task_ctx_nr; | |
5273 | if (ctxn < 0) | |
5274 | goto next; | |
5275 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
5276 | if (ctx) | |
67516844 | 5277 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
5278 | next: |
5279 | put_cpu_ptr(pmu->pmu_cpu_context); | |
5280 | } | |
5281 | ||
5282 | if (task_ctx) { | |
5283 | preempt_disable(); | |
67516844 | 5284 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
5285 | preempt_enable(); |
5286 | } | |
5287 | rcu_read_unlock(); | |
5288 | } | |
5289 | ||
60313ebe | 5290 | /* |
9f498cc5 PZ |
5291 | * task tracking -- fork/exit |
5292 | * | |
13d7a241 | 5293 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
5294 | */ |
5295 | ||
9f498cc5 | 5296 | struct perf_task_event { |
3a80b4a3 | 5297 | struct task_struct *task; |
cdd6c482 | 5298 | struct perf_event_context *task_ctx; |
60313ebe PZ |
5299 | |
5300 | struct { | |
5301 | struct perf_event_header header; | |
5302 | ||
5303 | u32 pid; | |
5304 | u32 ppid; | |
9f498cc5 PZ |
5305 | u32 tid; |
5306 | u32 ptid; | |
393b2ad8 | 5307 | u64 time; |
cdd6c482 | 5308 | } event_id; |
60313ebe PZ |
5309 | }; |
5310 | ||
67516844 JO |
5311 | static int perf_event_task_match(struct perf_event *event) |
5312 | { | |
13d7a241 SE |
5313 | return event->attr.comm || event->attr.mmap || |
5314 | event->attr.mmap2 || event->attr.mmap_data || | |
5315 | event->attr.task; | |
67516844 JO |
5316 | } |
5317 | ||
cdd6c482 | 5318 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 5319 | void *data) |
60313ebe | 5320 | { |
52d857a8 | 5321 | struct perf_task_event *task_event = data; |
60313ebe | 5322 | struct perf_output_handle handle; |
c980d109 | 5323 | struct perf_sample_data sample; |
9f498cc5 | 5324 | struct task_struct *task = task_event->task; |
c980d109 | 5325 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 5326 | |
67516844 JO |
5327 | if (!perf_event_task_match(event)) |
5328 | return; | |
5329 | ||
c980d109 | 5330 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 5331 | |
c980d109 | 5332 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 5333 | task_event->event_id.header.size); |
ef60777c | 5334 | if (ret) |
c980d109 | 5335 | goto out; |
60313ebe | 5336 | |
cdd6c482 IM |
5337 | task_event->event_id.pid = perf_event_pid(event, task); |
5338 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 5339 | |
cdd6c482 IM |
5340 | task_event->event_id.tid = perf_event_tid(event, task); |
5341 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 5342 | |
cdd6c482 | 5343 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 5344 | |
c980d109 ACM |
5345 | perf_event__output_id_sample(event, &handle, &sample); |
5346 | ||
60313ebe | 5347 | perf_output_end(&handle); |
c980d109 ACM |
5348 | out: |
5349 | task_event->event_id.header.size = size; | |
60313ebe PZ |
5350 | } |
5351 | ||
cdd6c482 IM |
5352 | static void perf_event_task(struct task_struct *task, |
5353 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 5354 | int new) |
60313ebe | 5355 | { |
9f498cc5 | 5356 | struct perf_task_event task_event; |
60313ebe | 5357 | |
cdd6c482 IM |
5358 | if (!atomic_read(&nr_comm_events) && |
5359 | !atomic_read(&nr_mmap_events) && | |
5360 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
5361 | return; |
5362 | ||
9f498cc5 | 5363 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
5364 | .task = task, |
5365 | .task_ctx = task_ctx, | |
cdd6c482 | 5366 | .event_id = { |
60313ebe | 5367 | .header = { |
cdd6c482 | 5368 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 5369 | .misc = 0, |
cdd6c482 | 5370 | .size = sizeof(task_event.event_id), |
60313ebe | 5371 | }, |
573402db PZ |
5372 | /* .pid */ |
5373 | /* .ppid */ | |
9f498cc5 PZ |
5374 | /* .tid */ |
5375 | /* .ptid */ | |
6f93d0a7 | 5376 | .time = perf_clock(), |
60313ebe PZ |
5377 | }, |
5378 | }; | |
5379 | ||
67516844 | 5380 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5381 | &task_event, |
5382 | task_ctx); | |
9f498cc5 PZ |
5383 | } |
5384 | ||
cdd6c482 | 5385 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5386 | { |
cdd6c482 | 5387 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5388 | } |
5389 | ||
8d1b2d93 PZ |
5390 | /* |
5391 | * comm tracking | |
5392 | */ | |
5393 | ||
5394 | struct perf_comm_event { | |
22a4f650 IM |
5395 | struct task_struct *task; |
5396 | char *comm; | |
8d1b2d93 PZ |
5397 | int comm_size; |
5398 | ||
5399 | struct { | |
5400 | struct perf_event_header header; | |
5401 | ||
5402 | u32 pid; | |
5403 | u32 tid; | |
cdd6c482 | 5404 | } event_id; |
8d1b2d93 PZ |
5405 | }; |
5406 | ||
67516844 JO |
5407 | static int perf_event_comm_match(struct perf_event *event) |
5408 | { | |
5409 | return event->attr.comm; | |
5410 | } | |
5411 | ||
cdd6c482 | 5412 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5413 | void *data) |
8d1b2d93 | 5414 | { |
52d857a8 | 5415 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5416 | struct perf_output_handle handle; |
c980d109 | 5417 | struct perf_sample_data sample; |
cdd6c482 | 5418 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5419 | int ret; |
5420 | ||
67516844 JO |
5421 | if (!perf_event_comm_match(event)) |
5422 | return; | |
5423 | ||
c980d109 ACM |
5424 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5425 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5426 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5427 | |
5428 | if (ret) | |
c980d109 | 5429 | goto out; |
8d1b2d93 | 5430 | |
cdd6c482 IM |
5431 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5432 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5433 | |
cdd6c482 | 5434 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5435 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5436 | comm_event->comm_size); |
c980d109 ACM |
5437 | |
5438 | perf_event__output_id_sample(event, &handle, &sample); | |
5439 | ||
8d1b2d93 | 5440 | perf_output_end(&handle); |
c980d109 ACM |
5441 | out: |
5442 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5443 | } |
5444 | ||
cdd6c482 | 5445 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5446 | { |
413ee3b4 | 5447 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5448 | unsigned int size; |
8d1b2d93 | 5449 | |
413ee3b4 | 5450 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5451 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5452 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5453 | |
5454 | comm_event->comm = comm; | |
5455 | comm_event->comm_size = size; | |
5456 | ||
cdd6c482 | 5457 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5458 | |
67516844 | 5459 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5460 | comm_event, |
5461 | NULL); | |
8d1b2d93 PZ |
5462 | } |
5463 | ||
82b89778 | 5464 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5465 | { |
9ee318a7 PZ |
5466 | struct perf_comm_event comm_event; |
5467 | ||
cdd6c482 | 5468 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5469 | return; |
a63eaf34 | 5470 | |
9ee318a7 | 5471 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5472 | .task = task, |
573402db PZ |
5473 | /* .comm */ |
5474 | /* .comm_size */ | |
cdd6c482 | 5475 | .event_id = { |
573402db | 5476 | .header = { |
cdd6c482 | 5477 | .type = PERF_RECORD_COMM, |
82b89778 | 5478 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5479 | /* .size */ |
5480 | }, | |
5481 | /* .pid */ | |
5482 | /* .tid */ | |
8d1b2d93 PZ |
5483 | }, |
5484 | }; | |
5485 | ||
cdd6c482 | 5486 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5487 | } |
5488 | ||
0a4a9391 PZ |
5489 | /* |
5490 | * mmap tracking | |
5491 | */ | |
5492 | ||
5493 | struct perf_mmap_event { | |
089dd79d PZ |
5494 | struct vm_area_struct *vma; |
5495 | ||
5496 | const char *file_name; | |
5497 | int file_size; | |
13d7a241 SE |
5498 | int maj, min; |
5499 | u64 ino; | |
5500 | u64 ino_generation; | |
f972eb63 | 5501 | u32 prot, flags; |
0a4a9391 PZ |
5502 | |
5503 | struct { | |
5504 | struct perf_event_header header; | |
5505 | ||
5506 | u32 pid; | |
5507 | u32 tid; | |
5508 | u64 start; | |
5509 | u64 len; | |
5510 | u64 pgoff; | |
cdd6c482 | 5511 | } event_id; |
0a4a9391 PZ |
5512 | }; |
5513 | ||
67516844 JO |
5514 | static int perf_event_mmap_match(struct perf_event *event, |
5515 | void *data) | |
5516 | { | |
5517 | struct perf_mmap_event *mmap_event = data; | |
5518 | struct vm_area_struct *vma = mmap_event->vma; | |
5519 | int executable = vma->vm_flags & VM_EXEC; | |
5520 | ||
5521 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5522 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5523 | } |
5524 | ||
cdd6c482 | 5525 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5526 | void *data) |
0a4a9391 | 5527 | { |
52d857a8 | 5528 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5529 | struct perf_output_handle handle; |
c980d109 | 5530 | struct perf_sample_data sample; |
cdd6c482 | 5531 | int size = mmap_event->event_id.header.size; |
c980d109 | 5532 | int ret; |
0a4a9391 | 5533 | |
67516844 JO |
5534 | if (!perf_event_mmap_match(event, data)) |
5535 | return; | |
5536 | ||
13d7a241 SE |
5537 | if (event->attr.mmap2) { |
5538 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5539 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5540 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5541 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5542 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
f972eb63 PZ |
5543 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
5544 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); | |
13d7a241 SE |
5545 | } |
5546 | ||
c980d109 ACM |
5547 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5548 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5549 | mmap_event->event_id.header.size); |
0a4a9391 | 5550 | if (ret) |
c980d109 | 5551 | goto out; |
0a4a9391 | 5552 | |
cdd6c482 IM |
5553 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5554 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5555 | |
cdd6c482 | 5556 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5557 | |
5558 | if (event->attr.mmap2) { | |
5559 | perf_output_put(&handle, mmap_event->maj); | |
5560 | perf_output_put(&handle, mmap_event->min); | |
5561 | perf_output_put(&handle, mmap_event->ino); | |
5562 | perf_output_put(&handle, mmap_event->ino_generation); | |
f972eb63 PZ |
5563 | perf_output_put(&handle, mmap_event->prot); |
5564 | perf_output_put(&handle, mmap_event->flags); | |
13d7a241 SE |
5565 | } |
5566 | ||
76369139 | 5567 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5568 | mmap_event->file_size); |
c980d109 ACM |
5569 | |
5570 | perf_event__output_id_sample(event, &handle, &sample); | |
5571 | ||
78d613eb | 5572 | perf_output_end(&handle); |
c980d109 ACM |
5573 | out: |
5574 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5575 | } |
5576 | ||
cdd6c482 | 5577 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5578 | { |
089dd79d PZ |
5579 | struct vm_area_struct *vma = mmap_event->vma; |
5580 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5581 | int maj = 0, min = 0; |
5582 | u64 ino = 0, gen = 0; | |
f972eb63 | 5583 | u32 prot = 0, flags = 0; |
0a4a9391 PZ |
5584 | unsigned int size; |
5585 | char tmp[16]; | |
5586 | char *buf = NULL; | |
2c42cfbf | 5587 | char *name; |
413ee3b4 | 5588 | |
0a4a9391 | 5589 | if (file) { |
13d7a241 SE |
5590 | struct inode *inode; |
5591 | dev_t dev; | |
3ea2f2b9 | 5592 | |
2c42cfbf | 5593 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5594 | if (!buf) { |
c7e548b4 ON |
5595 | name = "//enomem"; |
5596 | goto cpy_name; | |
0a4a9391 | 5597 | } |
413ee3b4 | 5598 | /* |
3ea2f2b9 | 5599 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5600 | * need to add enough zero bytes after the string to handle |
5601 | * the 64bit alignment we do later. | |
5602 | */ | |
3ea2f2b9 | 5603 | name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5604 | if (IS_ERR(name)) { |
c7e548b4 ON |
5605 | name = "//toolong"; |
5606 | goto cpy_name; | |
0a4a9391 | 5607 | } |
13d7a241 SE |
5608 | inode = file_inode(vma->vm_file); |
5609 | dev = inode->i_sb->s_dev; | |
5610 | ino = inode->i_ino; | |
5611 | gen = inode->i_generation; | |
5612 | maj = MAJOR(dev); | |
5613 | min = MINOR(dev); | |
f972eb63 PZ |
5614 | |
5615 | if (vma->vm_flags & VM_READ) | |
5616 | prot |= PROT_READ; | |
5617 | if (vma->vm_flags & VM_WRITE) | |
5618 | prot |= PROT_WRITE; | |
5619 | if (vma->vm_flags & VM_EXEC) | |
5620 | prot |= PROT_EXEC; | |
5621 | ||
5622 | if (vma->vm_flags & VM_MAYSHARE) | |
5623 | flags = MAP_SHARED; | |
5624 | else | |
5625 | flags = MAP_PRIVATE; | |
5626 | ||
5627 | if (vma->vm_flags & VM_DENYWRITE) | |
5628 | flags |= MAP_DENYWRITE; | |
5629 | if (vma->vm_flags & VM_MAYEXEC) | |
5630 | flags |= MAP_EXECUTABLE; | |
5631 | if (vma->vm_flags & VM_LOCKED) | |
5632 | flags |= MAP_LOCKED; | |
5633 | if (vma->vm_flags & VM_HUGETLB) | |
5634 | flags |= MAP_HUGETLB; | |
5635 | ||
c7e548b4 | 5636 | goto got_name; |
0a4a9391 | 5637 | } else { |
fbe26abe JO |
5638 | if (vma->vm_ops && vma->vm_ops->name) { |
5639 | name = (char *) vma->vm_ops->name(vma); | |
5640 | if (name) | |
5641 | goto cpy_name; | |
5642 | } | |
5643 | ||
2c42cfbf | 5644 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5645 | if (name) |
5646 | goto cpy_name; | |
089dd79d | 5647 | |
32c5fb7e | 5648 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5649 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5650 | name = "[heap]"; |
5651 | goto cpy_name; | |
32c5fb7e ON |
5652 | } |
5653 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5654 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5655 | name = "[stack]"; |
5656 | goto cpy_name; | |
089dd79d PZ |
5657 | } |
5658 | ||
c7e548b4 ON |
5659 | name = "//anon"; |
5660 | goto cpy_name; | |
0a4a9391 PZ |
5661 | } |
5662 | ||
c7e548b4 ON |
5663 | cpy_name: |
5664 | strlcpy(tmp, name, sizeof(tmp)); | |
5665 | name = tmp; | |
0a4a9391 | 5666 | got_name: |
2c42cfbf PZ |
5667 | /* |
5668 | * Since our buffer works in 8 byte units we need to align our string | |
5669 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5670 | * zero'd out to avoid leaking random bits to userspace. | |
5671 | */ | |
5672 | size = strlen(name)+1; | |
5673 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5674 | name[size++] = '\0'; | |
0a4a9391 PZ |
5675 | |
5676 | mmap_event->file_name = name; | |
5677 | mmap_event->file_size = size; | |
13d7a241 SE |
5678 | mmap_event->maj = maj; |
5679 | mmap_event->min = min; | |
5680 | mmap_event->ino = ino; | |
5681 | mmap_event->ino_generation = gen; | |
f972eb63 PZ |
5682 | mmap_event->prot = prot; |
5683 | mmap_event->flags = flags; | |
0a4a9391 | 5684 | |
2fe85427 SE |
5685 | if (!(vma->vm_flags & VM_EXEC)) |
5686 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5687 | ||
cdd6c482 | 5688 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5689 | |
67516844 | 5690 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5691 | mmap_event, |
5692 | NULL); | |
665c2142 | 5693 | |
0a4a9391 PZ |
5694 | kfree(buf); |
5695 | } | |
5696 | ||
3af9e859 | 5697 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5698 | { |
9ee318a7 PZ |
5699 | struct perf_mmap_event mmap_event; |
5700 | ||
cdd6c482 | 5701 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5702 | return; |
5703 | ||
5704 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5705 | .vma = vma, |
573402db PZ |
5706 | /* .file_name */ |
5707 | /* .file_size */ | |
cdd6c482 | 5708 | .event_id = { |
573402db | 5709 | .header = { |
cdd6c482 | 5710 | .type = PERF_RECORD_MMAP, |
39447b38 | 5711 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5712 | /* .size */ |
5713 | }, | |
5714 | /* .pid */ | |
5715 | /* .tid */ | |
089dd79d PZ |
5716 | .start = vma->vm_start, |
5717 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5718 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5719 | }, |
13d7a241 SE |
5720 | /* .maj (attr_mmap2 only) */ |
5721 | /* .min (attr_mmap2 only) */ | |
5722 | /* .ino (attr_mmap2 only) */ | |
5723 | /* .ino_generation (attr_mmap2 only) */ | |
f972eb63 PZ |
5724 | /* .prot (attr_mmap2 only) */ |
5725 | /* .flags (attr_mmap2 only) */ | |
0a4a9391 PZ |
5726 | }; |
5727 | ||
cdd6c482 | 5728 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5729 | } |
5730 | ||
a78ac325 PZ |
5731 | /* |
5732 | * IRQ throttle logging | |
5733 | */ | |
5734 | ||
cdd6c482 | 5735 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5736 | { |
5737 | struct perf_output_handle handle; | |
c980d109 | 5738 | struct perf_sample_data sample; |
a78ac325 PZ |
5739 | int ret; |
5740 | ||
5741 | struct { | |
5742 | struct perf_event_header header; | |
5743 | u64 time; | |
cca3f454 | 5744 | u64 id; |
7f453c24 | 5745 | u64 stream_id; |
a78ac325 PZ |
5746 | } throttle_event = { |
5747 | .header = { | |
cdd6c482 | 5748 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5749 | .misc = 0, |
5750 | .size = sizeof(throttle_event), | |
5751 | }, | |
def0a9b2 | 5752 | .time = perf_clock(), |
cdd6c482 IM |
5753 | .id = primary_event_id(event), |
5754 | .stream_id = event->id, | |
a78ac325 PZ |
5755 | }; |
5756 | ||
966ee4d6 | 5757 | if (enable) |
cdd6c482 | 5758 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5759 | |
c980d109 ACM |
5760 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5761 | ||
5762 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5763 | throttle_event.header.size); |
a78ac325 PZ |
5764 | if (ret) |
5765 | return; | |
5766 | ||
5767 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5768 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5769 | perf_output_end(&handle); |
5770 | } | |
5771 | ||
f6c7d5fe | 5772 | /* |
cdd6c482 | 5773 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5774 | */ |
5775 | ||
a8b0ca17 | 5776 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5777 | int throttle, struct perf_sample_data *data, |
5778 | struct pt_regs *regs) | |
f6c7d5fe | 5779 | { |
cdd6c482 IM |
5780 | int events = atomic_read(&event->event_limit); |
5781 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5782 | u64 seq; |
79f14641 PZ |
5783 | int ret = 0; |
5784 | ||
96398826 PZ |
5785 | /* |
5786 | * Non-sampling counters might still use the PMI to fold short | |
5787 | * hardware counters, ignore those. | |
5788 | */ | |
5789 | if (unlikely(!is_sampling_event(event))) | |
5790 | return 0; | |
5791 | ||
e050e3f0 SE |
5792 | seq = __this_cpu_read(perf_throttled_seq); |
5793 | if (seq != hwc->interrupts_seq) { | |
5794 | hwc->interrupts_seq = seq; | |
5795 | hwc->interrupts = 1; | |
5796 | } else { | |
5797 | hwc->interrupts++; | |
5798 | if (unlikely(throttle | |
5799 | && hwc->interrupts >= max_samples_per_tick)) { | |
5800 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5801 | hwc->interrupts = MAX_INTERRUPTS; |
5802 | perf_log_throttle(event, 0); | |
d84153d6 | 5803 | tick_nohz_full_kick(); |
a78ac325 PZ |
5804 | ret = 1; |
5805 | } | |
e050e3f0 | 5806 | } |
60db5e09 | 5807 | |
cdd6c482 | 5808 | if (event->attr.freq) { |
def0a9b2 | 5809 | u64 now = perf_clock(); |
abd50713 | 5810 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5811 | |
abd50713 | 5812 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5813 | |
abd50713 | 5814 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5815 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5816 | } |
5817 | ||
2023b359 PZ |
5818 | /* |
5819 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5820 | * events |
2023b359 PZ |
5821 | */ |
5822 | ||
cdd6c482 IM |
5823 | event->pending_kill = POLL_IN; |
5824 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5825 | ret = 1; |
cdd6c482 | 5826 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5827 | event->pending_disable = 1; |
5828 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5829 | } |
5830 | ||
453f19ee | 5831 | if (event->overflow_handler) |
a8b0ca17 | 5832 | event->overflow_handler(event, data, regs); |
453f19ee | 5833 | else |
a8b0ca17 | 5834 | perf_event_output(event, data, regs); |
453f19ee | 5835 | |
f506b3dc | 5836 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5837 | event->pending_wakeup = 1; |
5838 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5839 | } |
5840 | ||
79f14641 | 5841 | return ret; |
f6c7d5fe PZ |
5842 | } |
5843 | ||
a8b0ca17 | 5844 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5845 | struct perf_sample_data *data, |
5846 | struct pt_regs *regs) | |
850bc73f | 5847 | { |
a8b0ca17 | 5848 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5849 | } |
5850 | ||
15dbf27c | 5851 | /* |
cdd6c482 | 5852 | * Generic software event infrastructure |
15dbf27c PZ |
5853 | */ |
5854 | ||
b28ab83c PZ |
5855 | struct swevent_htable { |
5856 | struct swevent_hlist *swevent_hlist; | |
5857 | struct mutex hlist_mutex; | |
5858 | int hlist_refcount; | |
5859 | ||
5860 | /* Recursion avoidance in each contexts */ | |
5861 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
5862 | |
5863 | /* Keeps track of cpu being initialized/exited */ | |
5864 | bool online; | |
b28ab83c PZ |
5865 | }; |
5866 | ||
5867 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5868 | ||
7b4b6658 | 5869 | /* |
cdd6c482 IM |
5870 | * We directly increment event->count and keep a second value in |
5871 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5872 | * is kept in the range [-sample_period, 0] so that we can use the |
5873 | * sign as trigger. | |
5874 | */ | |
5875 | ||
ab573844 | 5876 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5877 | { |
cdd6c482 | 5878 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5879 | u64 period = hwc->last_period; |
5880 | u64 nr, offset; | |
5881 | s64 old, val; | |
5882 | ||
5883 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5884 | |
5885 | again: | |
e7850595 | 5886 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5887 | if (val < 0) |
5888 | return 0; | |
15dbf27c | 5889 | |
7b4b6658 PZ |
5890 | nr = div64_u64(period + val, period); |
5891 | offset = nr * period; | |
5892 | val -= offset; | |
e7850595 | 5893 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5894 | goto again; |
15dbf27c | 5895 | |
7b4b6658 | 5896 | return nr; |
15dbf27c PZ |
5897 | } |
5898 | ||
0cff784a | 5899 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5900 | struct perf_sample_data *data, |
5622f295 | 5901 | struct pt_regs *regs) |
15dbf27c | 5902 | { |
cdd6c482 | 5903 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5904 | int throttle = 0; |
15dbf27c | 5905 | |
0cff784a PZ |
5906 | if (!overflow) |
5907 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5908 | |
7b4b6658 PZ |
5909 | if (hwc->interrupts == MAX_INTERRUPTS) |
5910 | return; | |
15dbf27c | 5911 | |
7b4b6658 | 5912 | for (; overflow; overflow--) { |
a8b0ca17 | 5913 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5914 | data, regs)) { |
7b4b6658 PZ |
5915 | /* |
5916 | * We inhibit the overflow from happening when | |
5917 | * hwc->interrupts == MAX_INTERRUPTS. | |
5918 | */ | |
5919 | break; | |
5920 | } | |
cf450a73 | 5921 | throttle = 1; |
7b4b6658 | 5922 | } |
15dbf27c PZ |
5923 | } |
5924 | ||
a4eaf7f1 | 5925 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5926 | struct perf_sample_data *data, |
5622f295 | 5927 | struct pt_regs *regs) |
7b4b6658 | 5928 | { |
cdd6c482 | 5929 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5930 | |
e7850595 | 5931 | local64_add(nr, &event->count); |
d6d020e9 | 5932 | |
0cff784a PZ |
5933 | if (!regs) |
5934 | return; | |
5935 | ||
6c7e550f | 5936 | if (!is_sampling_event(event)) |
7b4b6658 | 5937 | return; |
d6d020e9 | 5938 | |
5d81e5cf AV |
5939 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5940 | data->period = nr; | |
5941 | return perf_swevent_overflow(event, 1, data, regs); | |
5942 | } else | |
5943 | data->period = event->hw.last_period; | |
5944 | ||
0cff784a | 5945 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5946 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5947 | |
e7850595 | 5948 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5949 | return; |
df1a132b | 5950 | |
a8b0ca17 | 5951 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5952 | } |
5953 | ||
f5ffe02e FW |
5954 | static int perf_exclude_event(struct perf_event *event, |
5955 | struct pt_regs *regs) | |
5956 | { | |
a4eaf7f1 | 5957 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5958 | return 1; |
a4eaf7f1 | 5959 | |
f5ffe02e FW |
5960 | if (regs) { |
5961 | if (event->attr.exclude_user && user_mode(regs)) | |
5962 | return 1; | |
5963 | ||
5964 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5965 | return 1; | |
5966 | } | |
5967 | ||
5968 | return 0; | |
5969 | } | |
5970 | ||
cdd6c482 | 5971 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5972 | enum perf_type_id type, |
6fb2915d LZ |
5973 | u32 event_id, |
5974 | struct perf_sample_data *data, | |
5975 | struct pt_regs *regs) | |
15dbf27c | 5976 | { |
cdd6c482 | 5977 | if (event->attr.type != type) |
a21ca2ca | 5978 | return 0; |
f5ffe02e | 5979 | |
cdd6c482 | 5980 | if (event->attr.config != event_id) |
15dbf27c PZ |
5981 | return 0; |
5982 | ||
f5ffe02e FW |
5983 | if (perf_exclude_event(event, regs)) |
5984 | return 0; | |
15dbf27c PZ |
5985 | |
5986 | return 1; | |
5987 | } | |
5988 | ||
76e1d904 FW |
5989 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5990 | { | |
5991 | u64 val = event_id | (type << 32); | |
5992 | ||
5993 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5994 | } | |
5995 | ||
49f135ed FW |
5996 | static inline struct hlist_head * |
5997 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5998 | { |
49f135ed FW |
5999 | u64 hash = swevent_hash(type, event_id); |
6000 | ||
6001 | return &hlist->heads[hash]; | |
6002 | } | |
76e1d904 | 6003 | |
49f135ed FW |
6004 | /* For the read side: events when they trigger */ |
6005 | static inline struct hlist_head * | |
b28ab83c | 6006 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
6007 | { |
6008 | struct swevent_hlist *hlist; | |
76e1d904 | 6009 | |
b28ab83c | 6010 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
6011 | if (!hlist) |
6012 | return NULL; | |
6013 | ||
49f135ed FW |
6014 | return __find_swevent_head(hlist, type, event_id); |
6015 | } | |
6016 | ||
6017 | /* For the event head insertion and removal in the hlist */ | |
6018 | static inline struct hlist_head * | |
b28ab83c | 6019 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
6020 | { |
6021 | struct swevent_hlist *hlist; | |
6022 | u32 event_id = event->attr.config; | |
6023 | u64 type = event->attr.type; | |
6024 | ||
6025 | /* | |
6026 | * Event scheduling is always serialized against hlist allocation | |
6027 | * and release. Which makes the protected version suitable here. | |
6028 | * The context lock guarantees that. | |
6029 | */ | |
b28ab83c | 6030 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
6031 | lockdep_is_held(&event->ctx->lock)); |
6032 | if (!hlist) | |
6033 | return NULL; | |
6034 | ||
6035 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
6036 | } |
6037 | ||
6038 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 6039 | u64 nr, |
76e1d904 FW |
6040 | struct perf_sample_data *data, |
6041 | struct pt_regs *regs) | |
15dbf27c | 6042 | { |
4a32fea9 | 6043 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6044 | struct perf_event *event; |
76e1d904 | 6045 | struct hlist_head *head; |
15dbf27c | 6046 | |
76e1d904 | 6047 | rcu_read_lock(); |
b28ab83c | 6048 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
6049 | if (!head) |
6050 | goto end; | |
6051 | ||
b67bfe0d | 6052 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 6053 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 6054 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 6055 | } |
76e1d904 FW |
6056 | end: |
6057 | rcu_read_unlock(); | |
15dbf27c PZ |
6058 | } |
6059 | ||
86038c5e PZI |
6060 | DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]); |
6061 | ||
4ed7c92d | 6062 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 6063 | { |
4a32fea9 | 6064 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
96f6d444 | 6065 | |
b28ab83c | 6066 | return get_recursion_context(swhash->recursion); |
96f6d444 | 6067 | } |
645e8cc0 | 6068 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 6069 | |
fa9f90be | 6070 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 6071 | { |
4a32fea9 | 6072 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
927c7a9e | 6073 | |
b28ab83c | 6074 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 6075 | } |
15dbf27c | 6076 | |
86038c5e | 6077 | void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 6078 | { |
a4234bfc | 6079 | struct perf_sample_data data; |
4ed7c92d | 6080 | |
86038c5e | 6081 | if (WARN_ON_ONCE(!regs)) |
4ed7c92d | 6082 | return; |
a4234bfc | 6083 | |
fd0d000b | 6084 | perf_sample_data_init(&data, addr, 0); |
a8b0ca17 | 6085 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
86038c5e PZI |
6086 | } |
6087 | ||
6088 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) | |
6089 | { | |
6090 | int rctx; | |
6091 | ||
6092 | preempt_disable_notrace(); | |
6093 | rctx = perf_swevent_get_recursion_context(); | |
6094 | if (unlikely(rctx < 0)) | |
6095 | goto fail; | |
6096 | ||
6097 | ___perf_sw_event(event_id, nr, regs, addr); | |
4ed7c92d PZ |
6098 | |
6099 | perf_swevent_put_recursion_context(rctx); | |
86038c5e | 6100 | fail: |
1c024eca | 6101 | preempt_enable_notrace(); |
b8e83514 PZ |
6102 | } |
6103 | ||
cdd6c482 | 6104 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 6105 | { |
15dbf27c PZ |
6106 | } |
6107 | ||
a4eaf7f1 | 6108 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 6109 | { |
4a32fea9 | 6110 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6111 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
6112 | struct hlist_head *head; |
6113 | ||
6c7e550f | 6114 | if (is_sampling_event(event)) { |
7b4b6658 | 6115 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 6116 | perf_swevent_set_period(event); |
7b4b6658 | 6117 | } |
76e1d904 | 6118 | |
a4eaf7f1 PZ |
6119 | hwc->state = !(flags & PERF_EF_START); |
6120 | ||
b28ab83c | 6121 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
6122 | if (!head) { |
6123 | /* | |
6124 | * We can race with cpu hotplug code. Do not | |
6125 | * WARN if the cpu just got unplugged. | |
6126 | */ | |
6127 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 6128 | return -EINVAL; |
39af6b16 | 6129 | } |
76e1d904 FW |
6130 | |
6131 | hlist_add_head_rcu(&event->hlist_entry, head); | |
6a694a60 | 6132 | perf_event_update_userpage(event); |
76e1d904 | 6133 | |
15dbf27c PZ |
6134 | return 0; |
6135 | } | |
6136 | ||
a4eaf7f1 | 6137 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 6138 | { |
76e1d904 | 6139 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
6140 | } |
6141 | ||
a4eaf7f1 | 6142 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 6143 | { |
a4eaf7f1 | 6144 | event->hw.state = 0; |
d6d020e9 | 6145 | } |
aa9c4c0f | 6146 | |
a4eaf7f1 | 6147 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 6148 | { |
a4eaf7f1 | 6149 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
6150 | } |
6151 | ||
49f135ed FW |
6152 | /* Deref the hlist from the update side */ |
6153 | static inline struct swevent_hlist * | |
b28ab83c | 6154 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 6155 | { |
b28ab83c PZ |
6156 | return rcu_dereference_protected(swhash->swevent_hlist, |
6157 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
6158 | } |
6159 | ||
b28ab83c | 6160 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 6161 | { |
b28ab83c | 6162 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 6163 | |
49f135ed | 6164 | if (!hlist) |
76e1d904 FW |
6165 | return; |
6166 | ||
70691d4a | 6167 | RCU_INIT_POINTER(swhash->swevent_hlist, NULL); |
fa4bbc4c | 6168 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
6169 | } |
6170 | ||
6171 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
6172 | { | |
b28ab83c | 6173 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 6174 | |
b28ab83c | 6175 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6176 | |
b28ab83c PZ |
6177 | if (!--swhash->hlist_refcount) |
6178 | swevent_hlist_release(swhash); | |
76e1d904 | 6179 | |
b28ab83c | 6180 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6181 | } |
6182 | ||
6183 | static void swevent_hlist_put(struct perf_event *event) | |
6184 | { | |
6185 | int cpu; | |
6186 | ||
76e1d904 FW |
6187 | for_each_possible_cpu(cpu) |
6188 | swevent_hlist_put_cpu(event, cpu); | |
6189 | } | |
6190 | ||
6191 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
6192 | { | |
b28ab83c | 6193 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
6194 | int err = 0; |
6195 | ||
b28ab83c | 6196 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6197 | |
b28ab83c | 6198 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
6199 | struct swevent_hlist *hlist; |
6200 | ||
6201 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
6202 | if (!hlist) { | |
6203 | err = -ENOMEM; | |
6204 | goto exit; | |
6205 | } | |
b28ab83c | 6206 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 6207 | } |
b28ab83c | 6208 | swhash->hlist_refcount++; |
9ed6060d | 6209 | exit: |
b28ab83c | 6210 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6211 | |
6212 | return err; | |
6213 | } | |
6214 | ||
6215 | static int swevent_hlist_get(struct perf_event *event) | |
6216 | { | |
6217 | int err; | |
6218 | int cpu, failed_cpu; | |
6219 | ||
76e1d904 FW |
6220 | get_online_cpus(); |
6221 | for_each_possible_cpu(cpu) { | |
6222 | err = swevent_hlist_get_cpu(event, cpu); | |
6223 | if (err) { | |
6224 | failed_cpu = cpu; | |
6225 | goto fail; | |
6226 | } | |
6227 | } | |
6228 | put_online_cpus(); | |
6229 | ||
6230 | return 0; | |
9ed6060d | 6231 | fail: |
76e1d904 FW |
6232 | for_each_possible_cpu(cpu) { |
6233 | if (cpu == failed_cpu) | |
6234 | break; | |
6235 | swevent_hlist_put_cpu(event, cpu); | |
6236 | } | |
6237 | ||
6238 | put_online_cpus(); | |
6239 | return err; | |
6240 | } | |
6241 | ||
c5905afb | 6242 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 6243 | |
b0a873eb PZ |
6244 | static void sw_perf_event_destroy(struct perf_event *event) |
6245 | { | |
6246 | u64 event_id = event->attr.config; | |
95476b64 | 6247 | |
b0a873eb PZ |
6248 | WARN_ON(event->parent); |
6249 | ||
c5905afb | 6250 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6251 | swevent_hlist_put(event); |
6252 | } | |
6253 | ||
6254 | static int perf_swevent_init(struct perf_event *event) | |
6255 | { | |
8176cced | 6256 | u64 event_id = event->attr.config; |
b0a873eb PZ |
6257 | |
6258 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6259 | return -ENOENT; | |
6260 | ||
2481c5fa SE |
6261 | /* |
6262 | * no branch sampling for software events | |
6263 | */ | |
6264 | if (has_branch_stack(event)) | |
6265 | return -EOPNOTSUPP; | |
6266 | ||
b0a873eb PZ |
6267 | switch (event_id) { |
6268 | case PERF_COUNT_SW_CPU_CLOCK: | |
6269 | case PERF_COUNT_SW_TASK_CLOCK: | |
6270 | return -ENOENT; | |
6271 | ||
6272 | default: | |
6273 | break; | |
6274 | } | |
6275 | ||
ce677831 | 6276 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
6277 | return -ENOENT; |
6278 | ||
6279 | if (!event->parent) { | |
6280 | int err; | |
6281 | ||
6282 | err = swevent_hlist_get(event); | |
6283 | if (err) | |
6284 | return err; | |
6285 | ||
c5905afb | 6286 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6287 | event->destroy = sw_perf_event_destroy; |
6288 | } | |
6289 | ||
6290 | return 0; | |
6291 | } | |
6292 | ||
6293 | static struct pmu perf_swevent = { | |
89a1e187 | 6294 | .task_ctx_nr = perf_sw_context, |
95476b64 | 6295 | |
b0a873eb | 6296 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
6297 | .add = perf_swevent_add, |
6298 | .del = perf_swevent_del, | |
6299 | .start = perf_swevent_start, | |
6300 | .stop = perf_swevent_stop, | |
1c024eca | 6301 | .read = perf_swevent_read, |
1c024eca PZ |
6302 | }; |
6303 | ||
b0a873eb PZ |
6304 | #ifdef CONFIG_EVENT_TRACING |
6305 | ||
1c024eca PZ |
6306 | static int perf_tp_filter_match(struct perf_event *event, |
6307 | struct perf_sample_data *data) | |
6308 | { | |
6309 | void *record = data->raw->data; | |
6310 | ||
6311 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
6312 | return 1; | |
6313 | return 0; | |
6314 | } | |
6315 | ||
6316 | static int perf_tp_event_match(struct perf_event *event, | |
6317 | struct perf_sample_data *data, | |
6318 | struct pt_regs *regs) | |
6319 | { | |
a0f7d0f7 FW |
6320 | if (event->hw.state & PERF_HES_STOPPED) |
6321 | return 0; | |
580d607c PZ |
6322 | /* |
6323 | * All tracepoints are from kernel-space. | |
6324 | */ | |
6325 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
6326 | return 0; |
6327 | ||
6328 | if (!perf_tp_filter_match(event, data)) | |
6329 | return 0; | |
6330 | ||
6331 | return 1; | |
6332 | } | |
6333 | ||
6334 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
6335 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
6336 | struct task_struct *task) | |
95476b64 FW |
6337 | { |
6338 | struct perf_sample_data data; | |
1c024eca | 6339 | struct perf_event *event; |
1c024eca | 6340 | |
95476b64 FW |
6341 | struct perf_raw_record raw = { |
6342 | .size = entry_size, | |
6343 | .data = record, | |
6344 | }; | |
6345 | ||
fd0d000b | 6346 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
6347 | data.raw = &raw; |
6348 | ||
b67bfe0d | 6349 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 6350 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 6351 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 6352 | } |
ecc55f84 | 6353 | |
e6dab5ff AV |
6354 | /* |
6355 | * If we got specified a target task, also iterate its context and | |
6356 | * deliver this event there too. | |
6357 | */ | |
6358 | if (task && task != current) { | |
6359 | struct perf_event_context *ctx; | |
6360 | struct trace_entry *entry = record; | |
6361 | ||
6362 | rcu_read_lock(); | |
6363 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
6364 | if (!ctx) | |
6365 | goto unlock; | |
6366 | ||
6367 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
6368 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6369 | continue; | |
6370 | if (event->attr.config != entry->type) | |
6371 | continue; | |
6372 | if (perf_tp_event_match(event, &data, regs)) | |
6373 | perf_swevent_event(event, count, &data, regs); | |
6374 | } | |
6375 | unlock: | |
6376 | rcu_read_unlock(); | |
6377 | } | |
6378 | ||
ecc55f84 | 6379 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
6380 | } |
6381 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
6382 | ||
cdd6c482 | 6383 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 6384 | { |
1c024eca | 6385 | perf_trace_destroy(event); |
e077df4f PZ |
6386 | } |
6387 | ||
b0a873eb | 6388 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 6389 | { |
76e1d904 FW |
6390 | int err; |
6391 | ||
b0a873eb PZ |
6392 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
6393 | return -ENOENT; | |
6394 | ||
2481c5fa SE |
6395 | /* |
6396 | * no branch sampling for tracepoint events | |
6397 | */ | |
6398 | if (has_branch_stack(event)) | |
6399 | return -EOPNOTSUPP; | |
6400 | ||
1c024eca PZ |
6401 | err = perf_trace_init(event); |
6402 | if (err) | |
b0a873eb | 6403 | return err; |
e077df4f | 6404 | |
cdd6c482 | 6405 | event->destroy = tp_perf_event_destroy; |
e077df4f | 6406 | |
b0a873eb PZ |
6407 | return 0; |
6408 | } | |
6409 | ||
6410 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
6411 | .task_ctx_nr = perf_sw_context, |
6412 | ||
b0a873eb | 6413 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6414 | .add = perf_trace_add, |
6415 | .del = perf_trace_del, | |
6416 | .start = perf_swevent_start, | |
6417 | .stop = perf_swevent_stop, | |
b0a873eb | 6418 | .read = perf_swevent_read, |
b0a873eb PZ |
6419 | }; |
6420 | ||
6421 | static inline void perf_tp_register(void) | |
6422 | { | |
2e80a82a | 6423 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6424 | } |
6fb2915d LZ |
6425 | |
6426 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6427 | { | |
6428 | char *filter_str; | |
6429 | int ret; | |
6430 | ||
6431 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6432 | return -EINVAL; | |
6433 | ||
6434 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6435 | if (IS_ERR(filter_str)) | |
6436 | return PTR_ERR(filter_str); | |
6437 | ||
6438 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6439 | ||
6440 | kfree(filter_str); | |
6441 | return ret; | |
6442 | } | |
6443 | ||
6444 | static void perf_event_free_filter(struct perf_event *event) | |
6445 | { | |
6446 | ftrace_profile_free_filter(event); | |
6447 | } | |
6448 | ||
e077df4f | 6449 | #else |
6fb2915d | 6450 | |
b0a873eb | 6451 | static inline void perf_tp_register(void) |
e077df4f | 6452 | { |
e077df4f | 6453 | } |
6fb2915d LZ |
6454 | |
6455 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6456 | { | |
6457 | return -ENOENT; | |
6458 | } | |
6459 | ||
6460 | static void perf_event_free_filter(struct perf_event *event) | |
6461 | { | |
6462 | } | |
6463 | ||
07b139c8 | 6464 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6465 | |
24f1e32c | 6466 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6467 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6468 | { |
f5ffe02e FW |
6469 | struct perf_sample_data sample; |
6470 | struct pt_regs *regs = data; | |
6471 | ||
fd0d000b | 6472 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6473 | |
a4eaf7f1 | 6474 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6475 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6476 | } |
6477 | #endif | |
6478 | ||
b0a873eb PZ |
6479 | /* |
6480 | * hrtimer based swevent callback | |
6481 | */ | |
f29ac756 | 6482 | |
b0a873eb | 6483 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6484 | { |
b0a873eb PZ |
6485 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6486 | struct perf_sample_data data; | |
6487 | struct pt_regs *regs; | |
6488 | struct perf_event *event; | |
6489 | u64 period; | |
f29ac756 | 6490 | |
b0a873eb | 6491 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6492 | |
6493 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6494 | return HRTIMER_NORESTART; | |
6495 | ||
b0a873eb | 6496 | event->pmu->read(event); |
f344011c | 6497 | |
fd0d000b | 6498 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6499 | regs = get_irq_regs(); |
6500 | ||
6501 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6502 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6503 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6504 | ret = HRTIMER_NORESTART; |
6505 | } | |
24f1e32c | 6506 | |
b0a873eb PZ |
6507 | period = max_t(u64, 10000, event->hw.sample_period); |
6508 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6509 | |
b0a873eb | 6510 | return ret; |
f29ac756 PZ |
6511 | } |
6512 | ||
b0a873eb | 6513 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6514 | { |
b0a873eb | 6515 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6516 | s64 period; |
6517 | ||
6518 | if (!is_sampling_event(event)) | |
6519 | return; | |
f5ffe02e | 6520 | |
5d508e82 FBH |
6521 | period = local64_read(&hwc->period_left); |
6522 | if (period) { | |
6523 | if (period < 0) | |
6524 | period = 10000; | |
fa407f35 | 6525 | |
5d508e82 FBH |
6526 | local64_set(&hwc->period_left, 0); |
6527 | } else { | |
6528 | period = max_t(u64, 10000, hwc->sample_period); | |
6529 | } | |
6530 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6531 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6532 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6533 | } |
b0a873eb PZ |
6534 | |
6535 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6536 | { |
b0a873eb PZ |
6537 | struct hw_perf_event *hwc = &event->hw; |
6538 | ||
6c7e550f | 6539 | if (is_sampling_event(event)) { |
b0a873eb | 6540 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6541 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6542 | |
6543 | hrtimer_cancel(&hwc->hrtimer); | |
6544 | } | |
24f1e32c FW |
6545 | } |
6546 | ||
ba3dd36c PZ |
6547 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6548 | { | |
6549 | struct hw_perf_event *hwc = &event->hw; | |
6550 | ||
6551 | if (!is_sampling_event(event)) | |
6552 | return; | |
6553 | ||
6554 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6555 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6556 | ||
6557 | /* | |
6558 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6559 | * mapping and avoid the whole period adjust feedback stuff. | |
6560 | */ | |
6561 | if (event->attr.freq) { | |
6562 | long freq = event->attr.sample_freq; | |
6563 | ||
6564 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6565 | hwc->sample_period = event->attr.sample_period; | |
6566 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6567 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6568 | event->attr.freq = 0; |
6569 | } | |
6570 | } | |
6571 | ||
b0a873eb PZ |
6572 | /* |
6573 | * Software event: cpu wall time clock | |
6574 | */ | |
6575 | ||
6576 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6577 | { |
b0a873eb PZ |
6578 | s64 prev; |
6579 | u64 now; | |
6580 | ||
a4eaf7f1 | 6581 | now = local_clock(); |
b0a873eb PZ |
6582 | prev = local64_xchg(&event->hw.prev_count, now); |
6583 | local64_add(now - prev, &event->count); | |
24f1e32c | 6584 | } |
24f1e32c | 6585 | |
a4eaf7f1 | 6586 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6587 | { |
a4eaf7f1 | 6588 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6589 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6590 | } |
6591 | ||
a4eaf7f1 | 6592 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6593 | { |
b0a873eb PZ |
6594 | perf_swevent_cancel_hrtimer(event); |
6595 | cpu_clock_event_update(event); | |
6596 | } | |
f29ac756 | 6597 | |
a4eaf7f1 PZ |
6598 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6599 | { | |
6600 | if (flags & PERF_EF_START) | |
6601 | cpu_clock_event_start(event, flags); | |
6a694a60 | 6602 | perf_event_update_userpage(event); |
a4eaf7f1 PZ |
6603 | |
6604 | return 0; | |
6605 | } | |
6606 | ||
6607 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6608 | { | |
6609 | cpu_clock_event_stop(event, flags); | |
6610 | } | |
6611 | ||
b0a873eb PZ |
6612 | static void cpu_clock_event_read(struct perf_event *event) |
6613 | { | |
6614 | cpu_clock_event_update(event); | |
6615 | } | |
f344011c | 6616 | |
b0a873eb PZ |
6617 | static int cpu_clock_event_init(struct perf_event *event) |
6618 | { | |
6619 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6620 | return -ENOENT; | |
6621 | ||
6622 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6623 | return -ENOENT; | |
6624 | ||
2481c5fa SE |
6625 | /* |
6626 | * no branch sampling for software events | |
6627 | */ | |
6628 | if (has_branch_stack(event)) | |
6629 | return -EOPNOTSUPP; | |
6630 | ||
ba3dd36c PZ |
6631 | perf_swevent_init_hrtimer(event); |
6632 | ||
b0a873eb | 6633 | return 0; |
f29ac756 PZ |
6634 | } |
6635 | ||
b0a873eb | 6636 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6637 | .task_ctx_nr = perf_sw_context, |
6638 | ||
b0a873eb | 6639 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6640 | .add = cpu_clock_event_add, |
6641 | .del = cpu_clock_event_del, | |
6642 | .start = cpu_clock_event_start, | |
6643 | .stop = cpu_clock_event_stop, | |
b0a873eb PZ |
6644 | .read = cpu_clock_event_read, |
6645 | }; | |
6646 | ||
6647 | /* | |
6648 | * Software event: task time clock | |
6649 | */ | |
6650 | ||
6651 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6652 | { |
b0a873eb PZ |
6653 | u64 prev; |
6654 | s64 delta; | |
5c92d124 | 6655 | |
b0a873eb PZ |
6656 | prev = local64_xchg(&event->hw.prev_count, now); |
6657 | delta = now - prev; | |
6658 | local64_add(delta, &event->count); | |
6659 | } | |
5c92d124 | 6660 | |
a4eaf7f1 | 6661 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6662 | { |
a4eaf7f1 | 6663 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6664 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6665 | } |
6666 | ||
a4eaf7f1 | 6667 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6668 | { |
6669 | perf_swevent_cancel_hrtimer(event); | |
6670 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6671 | } |
6672 | ||
6673 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6674 | { | |
6675 | if (flags & PERF_EF_START) | |
6676 | task_clock_event_start(event, flags); | |
6a694a60 | 6677 | perf_event_update_userpage(event); |
b0a873eb | 6678 | |
a4eaf7f1 PZ |
6679 | return 0; |
6680 | } | |
6681 | ||
6682 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6683 | { | |
6684 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6685 | } |
6686 | ||
6687 | static void task_clock_event_read(struct perf_event *event) | |
6688 | { | |
768a06e2 PZ |
6689 | u64 now = perf_clock(); |
6690 | u64 delta = now - event->ctx->timestamp; | |
6691 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6692 | |
6693 | task_clock_event_update(event, time); | |
6694 | } | |
6695 | ||
6696 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6697 | { |
b0a873eb PZ |
6698 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6699 | return -ENOENT; | |
6700 | ||
6701 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6702 | return -ENOENT; | |
6703 | ||
2481c5fa SE |
6704 | /* |
6705 | * no branch sampling for software events | |
6706 | */ | |
6707 | if (has_branch_stack(event)) | |
6708 | return -EOPNOTSUPP; | |
6709 | ||
ba3dd36c PZ |
6710 | perf_swevent_init_hrtimer(event); |
6711 | ||
b0a873eb | 6712 | return 0; |
6fb2915d LZ |
6713 | } |
6714 | ||
b0a873eb | 6715 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6716 | .task_ctx_nr = perf_sw_context, |
6717 | ||
b0a873eb | 6718 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6719 | .add = task_clock_event_add, |
6720 | .del = task_clock_event_del, | |
6721 | .start = task_clock_event_start, | |
6722 | .stop = task_clock_event_stop, | |
b0a873eb PZ |
6723 | .read = task_clock_event_read, |
6724 | }; | |
6fb2915d | 6725 | |
ad5133b7 | 6726 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6727 | { |
e077df4f | 6728 | } |
6fb2915d | 6729 | |
ad5133b7 | 6730 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6731 | { |
ad5133b7 | 6732 | return 0; |
6fb2915d LZ |
6733 | } |
6734 | ||
ad5133b7 | 6735 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6736 | { |
ad5133b7 | 6737 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6738 | } |
6739 | ||
ad5133b7 PZ |
6740 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6741 | { | |
6742 | perf_pmu_enable(pmu); | |
6743 | return 0; | |
6744 | } | |
e077df4f | 6745 | |
ad5133b7 | 6746 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6747 | { |
ad5133b7 | 6748 | perf_pmu_enable(pmu); |
24f1e32c FW |
6749 | } |
6750 | ||
35edc2a5 PZ |
6751 | static int perf_event_idx_default(struct perf_event *event) |
6752 | { | |
c719f560 | 6753 | return 0; |
35edc2a5 PZ |
6754 | } |
6755 | ||
8dc85d54 PZ |
6756 | /* |
6757 | * Ensures all contexts with the same task_ctx_nr have the same | |
6758 | * pmu_cpu_context too. | |
6759 | */ | |
9e317041 | 6760 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 6761 | { |
8dc85d54 | 6762 | struct pmu *pmu; |
b326e956 | 6763 | |
8dc85d54 PZ |
6764 | if (ctxn < 0) |
6765 | return NULL; | |
24f1e32c | 6766 | |
8dc85d54 PZ |
6767 | list_for_each_entry(pmu, &pmus, entry) { |
6768 | if (pmu->task_ctx_nr == ctxn) | |
6769 | return pmu->pmu_cpu_context; | |
6770 | } | |
24f1e32c | 6771 | |
8dc85d54 | 6772 | return NULL; |
24f1e32c FW |
6773 | } |
6774 | ||
51676957 | 6775 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6776 | { |
51676957 PZ |
6777 | int cpu; |
6778 | ||
6779 | for_each_possible_cpu(cpu) { | |
6780 | struct perf_cpu_context *cpuctx; | |
6781 | ||
6782 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6783 | ||
3f1f3320 PZ |
6784 | if (cpuctx->unique_pmu == old_pmu) |
6785 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6786 | } |
6787 | } | |
6788 | ||
6789 | static void free_pmu_context(struct pmu *pmu) | |
6790 | { | |
6791 | struct pmu *i; | |
f5ffe02e | 6792 | |
8dc85d54 | 6793 | mutex_lock(&pmus_lock); |
0475f9ea | 6794 | /* |
8dc85d54 | 6795 | * Like a real lame refcount. |
0475f9ea | 6796 | */ |
51676957 PZ |
6797 | list_for_each_entry(i, &pmus, entry) { |
6798 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6799 | update_pmu_context(i, pmu); | |
8dc85d54 | 6800 | goto out; |
51676957 | 6801 | } |
8dc85d54 | 6802 | } |
d6d020e9 | 6803 | |
51676957 | 6804 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6805 | out: |
6806 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6807 | } |
2e80a82a | 6808 | static struct idr pmu_idr; |
d6d020e9 | 6809 | |
abe43400 PZ |
6810 | static ssize_t |
6811 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6812 | { | |
6813 | struct pmu *pmu = dev_get_drvdata(dev); | |
6814 | ||
6815 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6816 | } | |
90826ca7 | 6817 | static DEVICE_ATTR_RO(type); |
abe43400 | 6818 | |
62b85639 SE |
6819 | static ssize_t |
6820 | perf_event_mux_interval_ms_show(struct device *dev, | |
6821 | struct device_attribute *attr, | |
6822 | char *page) | |
6823 | { | |
6824 | struct pmu *pmu = dev_get_drvdata(dev); | |
6825 | ||
6826 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6827 | } | |
6828 | ||
6829 | static ssize_t | |
6830 | perf_event_mux_interval_ms_store(struct device *dev, | |
6831 | struct device_attribute *attr, | |
6832 | const char *buf, size_t count) | |
6833 | { | |
6834 | struct pmu *pmu = dev_get_drvdata(dev); | |
6835 | int timer, cpu, ret; | |
6836 | ||
6837 | ret = kstrtoint(buf, 0, &timer); | |
6838 | if (ret) | |
6839 | return ret; | |
6840 | ||
6841 | if (timer < 1) | |
6842 | return -EINVAL; | |
6843 | ||
6844 | /* same value, noting to do */ | |
6845 | if (timer == pmu->hrtimer_interval_ms) | |
6846 | return count; | |
6847 | ||
6848 | pmu->hrtimer_interval_ms = timer; | |
6849 | ||
6850 | /* update all cpuctx for this PMU */ | |
6851 | for_each_possible_cpu(cpu) { | |
6852 | struct perf_cpu_context *cpuctx; | |
6853 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6854 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6855 | ||
6856 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6857 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6858 | } | |
6859 | ||
6860 | return count; | |
6861 | } | |
90826ca7 | 6862 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 6863 | |
90826ca7 GKH |
6864 | static struct attribute *pmu_dev_attrs[] = { |
6865 | &dev_attr_type.attr, | |
6866 | &dev_attr_perf_event_mux_interval_ms.attr, | |
6867 | NULL, | |
abe43400 | 6868 | }; |
90826ca7 | 6869 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
6870 | |
6871 | static int pmu_bus_running; | |
6872 | static struct bus_type pmu_bus = { | |
6873 | .name = "event_source", | |
90826ca7 | 6874 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
6875 | }; |
6876 | ||
6877 | static void pmu_dev_release(struct device *dev) | |
6878 | { | |
6879 | kfree(dev); | |
6880 | } | |
6881 | ||
6882 | static int pmu_dev_alloc(struct pmu *pmu) | |
6883 | { | |
6884 | int ret = -ENOMEM; | |
6885 | ||
6886 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6887 | if (!pmu->dev) | |
6888 | goto out; | |
6889 | ||
0c9d42ed | 6890 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6891 | device_initialize(pmu->dev); |
6892 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6893 | if (ret) | |
6894 | goto free_dev; | |
6895 | ||
6896 | dev_set_drvdata(pmu->dev, pmu); | |
6897 | pmu->dev->bus = &pmu_bus; | |
6898 | pmu->dev->release = pmu_dev_release; | |
6899 | ret = device_add(pmu->dev); | |
6900 | if (ret) | |
6901 | goto free_dev; | |
6902 | ||
6903 | out: | |
6904 | return ret; | |
6905 | ||
6906 | free_dev: | |
6907 | put_device(pmu->dev); | |
6908 | goto out; | |
6909 | } | |
6910 | ||
547e9fd7 | 6911 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6912 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6913 | |
03d8e80b | 6914 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6915 | { |
108b02cf | 6916 | int cpu, ret; |
24f1e32c | 6917 | |
b0a873eb | 6918 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6919 | ret = -ENOMEM; |
6920 | pmu->pmu_disable_count = alloc_percpu(int); | |
6921 | if (!pmu->pmu_disable_count) | |
6922 | goto unlock; | |
f29ac756 | 6923 | |
2e80a82a PZ |
6924 | pmu->type = -1; |
6925 | if (!name) | |
6926 | goto skip_type; | |
6927 | pmu->name = name; | |
6928 | ||
6929 | if (type < 0) { | |
0e9c3be2 TH |
6930 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6931 | if (type < 0) { | |
6932 | ret = type; | |
2e80a82a PZ |
6933 | goto free_pdc; |
6934 | } | |
6935 | } | |
6936 | pmu->type = type; | |
6937 | ||
abe43400 PZ |
6938 | if (pmu_bus_running) { |
6939 | ret = pmu_dev_alloc(pmu); | |
6940 | if (ret) | |
6941 | goto free_idr; | |
6942 | } | |
6943 | ||
2e80a82a | 6944 | skip_type: |
8dc85d54 PZ |
6945 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6946 | if (pmu->pmu_cpu_context) | |
6947 | goto got_cpu_context; | |
f29ac756 | 6948 | |
c4814202 | 6949 | ret = -ENOMEM; |
108b02cf PZ |
6950 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6951 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6952 | goto free_dev; |
f344011c | 6953 | |
108b02cf PZ |
6954 | for_each_possible_cpu(cpu) { |
6955 | struct perf_cpu_context *cpuctx; | |
6956 | ||
6957 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6958 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6959 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6960 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
108b02cf | 6961 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6962 | |
6963 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6964 | ||
3f1f3320 | 6965 | cpuctx->unique_pmu = pmu; |
108b02cf | 6966 | } |
76e1d904 | 6967 | |
8dc85d54 | 6968 | got_cpu_context: |
ad5133b7 PZ |
6969 | if (!pmu->start_txn) { |
6970 | if (pmu->pmu_enable) { | |
6971 | /* | |
6972 | * If we have pmu_enable/pmu_disable calls, install | |
6973 | * transaction stubs that use that to try and batch | |
6974 | * hardware accesses. | |
6975 | */ | |
6976 | pmu->start_txn = perf_pmu_start_txn; | |
6977 | pmu->commit_txn = perf_pmu_commit_txn; | |
6978 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6979 | } else { | |
6980 | pmu->start_txn = perf_pmu_nop_void; | |
6981 | pmu->commit_txn = perf_pmu_nop_int; | |
6982 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6983 | } |
5c92d124 | 6984 | } |
15dbf27c | 6985 | |
ad5133b7 PZ |
6986 | if (!pmu->pmu_enable) { |
6987 | pmu->pmu_enable = perf_pmu_nop_void; | |
6988 | pmu->pmu_disable = perf_pmu_nop_void; | |
6989 | } | |
6990 | ||
35edc2a5 PZ |
6991 | if (!pmu->event_idx) |
6992 | pmu->event_idx = perf_event_idx_default; | |
6993 | ||
b0a873eb | 6994 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6995 | ret = 0; |
6996 | unlock: | |
b0a873eb PZ |
6997 | mutex_unlock(&pmus_lock); |
6998 | ||
33696fc0 | 6999 | return ret; |
108b02cf | 7000 | |
abe43400 PZ |
7001 | free_dev: |
7002 | device_del(pmu->dev); | |
7003 | put_device(pmu->dev); | |
7004 | ||
2e80a82a PZ |
7005 | free_idr: |
7006 | if (pmu->type >= PERF_TYPE_MAX) | |
7007 | idr_remove(&pmu_idr, pmu->type); | |
7008 | ||
108b02cf PZ |
7009 | free_pdc: |
7010 | free_percpu(pmu->pmu_disable_count); | |
7011 | goto unlock; | |
f29ac756 | 7012 | } |
c464c76e | 7013 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 7014 | |
b0a873eb | 7015 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 7016 | { |
b0a873eb PZ |
7017 | mutex_lock(&pmus_lock); |
7018 | list_del_rcu(&pmu->entry); | |
7019 | mutex_unlock(&pmus_lock); | |
5c92d124 | 7020 | |
0475f9ea | 7021 | /* |
cde8e884 PZ |
7022 | * We dereference the pmu list under both SRCU and regular RCU, so |
7023 | * synchronize against both of those. | |
0475f9ea | 7024 | */ |
b0a873eb | 7025 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 7026 | synchronize_rcu(); |
d6d020e9 | 7027 | |
33696fc0 | 7028 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
7029 | if (pmu->type >= PERF_TYPE_MAX) |
7030 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
7031 | device_del(pmu->dev); |
7032 | put_device(pmu->dev); | |
51676957 | 7033 | free_pmu_context(pmu); |
b0a873eb | 7034 | } |
c464c76e | 7035 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 7036 | |
cc34b98b MR |
7037 | static int perf_try_init_event(struct pmu *pmu, struct perf_event *event) |
7038 | { | |
7039 | int ret; | |
7040 | ||
7041 | if (!try_module_get(pmu->module)) | |
7042 | return -ENODEV; | |
7043 | event->pmu = pmu; | |
7044 | ret = pmu->event_init(event); | |
7045 | if (ret) | |
7046 | module_put(pmu->module); | |
7047 | ||
7048 | return ret; | |
7049 | } | |
7050 | ||
b0a873eb PZ |
7051 | struct pmu *perf_init_event(struct perf_event *event) |
7052 | { | |
7053 | struct pmu *pmu = NULL; | |
7054 | int idx; | |
940c5b29 | 7055 | int ret; |
b0a873eb PZ |
7056 | |
7057 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
7058 | |
7059 | rcu_read_lock(); | |
7060 | pmu = idr_find(&pmu_idr, event->attr.type); | |
7061 | rcu_read_unlock(); | |
940c5b29 | 7062 | if (pmu) { |
cc34b98b | 7063 | ret = perf_try_init_event(pmu, event); |
940c5b29 LM |
7064 | if (ret) |
7065 | pmu = ERR_PTR(ret); | |
2e80a82a | 7066 | goto unlock; |
940c5b29 | 7067 | } |
2e80a82a | 7068 | |
b0a873eb | 7069 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
cc34b98b | 7070 | ret = perf_try_init_event(pmu, event); |
b0a873eb | 7071 | if (!ret) |
e5f4d339 | 7072 | goto unlock; |
76e1d904 | 7073 | |
b0a873eb PZ |
7074 | if (ret != -ENOENT) { |
7075 | pmu = ERR_PTR(ret); | |
e5f4d339 | 7076 | goto unlock; |
f344011c | 7077 | } |
5c92d124 | 7078 | } |
e5f4d339 PZ |
7079 | pmu = ERR_PTR(-ENOENT); |
7080 | unlock: | |
b0a873eb | 7081 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 7082 | |
4aeb0b42 | 7083 | return pmu; |
5c92d124 IM |
7084 | } |
7085 | ||
4beb31f3 FW |
7086 | static void account_event_cpu(struct perf_event *event, int cpu) |
7087 | { | |
7088 | if (event->parent) | |
7089 | return; | |
7090 | ||
4beb31f3 FW |
7091 | if (is_cgroup_event(event)) |
7092 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
7093 | } | |
7094 | ||
766d6c07 FW |
7095 | static void account_event(struct perf_event *event) |
7096 | { | |
4beb31f3 FW |
7097 | if (event->parent) |
7098 | return; | |
7099 | ||
766d6c07 FW |
7100 | if (event->attach_state & PERF_ATTACH_TASK) |
7101 | static_key_slow_inc(&perf_sched_events.key); | |
7102 | if (event->attr.mmap || event->attr.mmap_data) | |
7103 | atomic_inc(&nr_mmap_events); | |
7104 | if (event->attr.comm) | |
7105 | atomic_inc(&nr_comm_events); | |
7106 | if (event->attr.task) | |
7107 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
7108 | if (event->attr.freq) { |
7109 | if (atomic_inc_return(&nr_freq_events) == 1) | |
7110 | tick_nohz_full_kick_all(); | |
7111 | } | |
4beb31f3 | 7112 | if (has_branch_stack(event)) |
766d6c07 | 7113 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 7114 | if (is_cgroup_event(event)) |
766d6c07 | 7115 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
7116 | |
7117 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
7118 | } |
7119 | ||
0793a61d | 7120 | /* |
cdd6c482 | 7121 | * Allocate and initialize a event structure |
0793a61d | 7122 | */ |
cdd6c482 | 7123 | static struct perf_event * |
c3f00c70 | 7124 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
7125 | struct task_struct *task, |
7126 | struct perf_event *group_leader, | |
7127 | struct perf_event *parent_event, | |
4dc0da86 | 7128 | perf_overflow_handler_t overflow_handler, |
79dff51e | 7129 | void *context, int cgroup_fd) |
0793a61d | 7130 | { |
51b0fe39 | 7131 | struct pmu *pmu; |
cdd6c482 IM |
7132 | struct perf_event *event; |
7133 | struct hw_perf_event *hwc; | |
90983b16 | 7134 | long err = -EINVAL; |
0793a61d | 7135 | |
66832eb4 ON |
7136 | if ((unsigned)cpu >= nr_cpu_ids) { |
7137 | if (!task || cpu != -1) | |
7138 | return ERR_PTR(-EINVAL); | |
7139 | } | |
7140 | ||
c3f00c70 | 7141 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 7142 | if (!event) |
d5d2bc0d | 7143 | return ERR_PTR(-ENOMEM); |
0793a61d | 7144 | |
04289bb9 | 7145 | /* |
cdd6c482 | 7146 | * Single events are their own group leaders, with an |
04289bb9 IM |
7147 | * empty sibling list: |
7148 | */ | |
7149 | if (!group_leader) | |
cdd6c482 | 7150 | group_leader = event; |
04289bb9 | 7151 | |
cdd6c482 IM |
7152 | mutex_init(&event->child_mutex); |
7153 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 7154 | |
cdd6c482 IM |
7155 | INIT_LIST_HEAD(&event->group_entry); |
7156 | INIT_LIST_HEAD(&event->event_entry); | |
7157 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 7158 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 7159 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
7160 | INIT_HLIST_NODE(&event->hlist_entry); |
7161 | ||
10c6db11 | 7162 | |
cdd6c482 | 7163 | init_waitqueue_head(&event->waitq); |
e360adbe | 7164 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 7165 | |
cdd6c482 | 7166 | mutex_init(&event->mmap_mutex); |
7b732a75 | 7167 | |
a6fa941d | 7168 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
7169 | event->cpu = cpu; |
7170 | event->attr = *attr; | |
7171 | event->group_leader = group_leader; | |
7172 | event->pmu = NULL; | |
cdd6c482 | 7173 | event->oncpu = -1; |
a96bbc16 | 7174 | |
cdd6c482 | 7175 | event->parent = parent_event; |
b84fbc9f | 7176 | |
17cf22c3 | 7177 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 7178 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 7179 | |
cdd6c482 | 7180 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 7181 | |
d580ff86 PZ |
7182 | if (task) { |
7183 | event->attach_state = PERF_ATTACH_TASK; | |
d580ff86 | 7184 | /* |
50f16a8b PZ |
7185 | * XXX pmu::event_init needs to know what task to account to |
7186 | * and we cannot use the ctx information because we need the | |
7187 | * pmu before we get a ctx. | |
d580ff86 | 7188 | */ |
50f16a8b | 7189 | event->hw.target = task; |
d580ff86 PZ |
7190 | } |
7191 | ||
4dc0da86 | 7192 | if (!overflow_handler && parent_event) { |
b326e956 | 7193 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
7194 | context = parent_event->overflow_handler_context; |
7195 | } | |
66832eb4 | 7196 | |
b326e956 | 7197 | event->overflow_handler = overflow_handler; |
4dc0da86 | 7198 | event->overflow_handler_context = context; |
97eaf530 | 7199 | |
0231bb53 | 7200 | perf_event__state_init(event); |
a86ed508 | 7201 | |
4aeb0b42 | 7202 | pmu = NULL; |
b8e83514 | 7203 | |
cdd6c482 | 7204 | hwc = &event->hw; |
bd2b5b12 | 7205 | hwc->sample_period = attr->sample_period; |
0d48696f | 7206 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 7207 | hwc->sample_period = 1; |
eced1dfc | 7208 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 7209 | |
e7850595 | 7210 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 7211 | |
2023b359 | 7212 | /* |
cdd6c482 | 7213 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 7214 | */ |
3dab77fb | 7215 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 7216 | goto err_ns; |
a46a2300 YZ |
7217 | |
7218 | if (!has_branch_stack(event)) | |
7219 | event->attr.branch_sample_type = 0; | |
2023b359 | 7220 | |
79dff51e MF |
7221 | if (cgroup_fd != -1) { |
7222 | err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader); | |
7223 | if (err) | |
7224 | goto err_ns; | |
7225 | } | |
7226 | ||
b0a873eb | 7227 | pmu = perf_init_event(event); |
4aeb0b42 | 7228 | if (!pmu) |
90983b16 FW |
7229 | goto err_ns; |
7230 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 7231 | err = PTR_ERR(pmu); |
90983b16 | 7232 | goto err_ns; |
621a01ea | 7233 | } |
d5d2bc0d | 7234 | |
cdd6c482 | 7235 | if (!event->parent) { |
927c7a9e FW |
7236 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
7237 | err = get_callchain_buffers(); | |
90983b16 FW |
7238 | if (err) |
7239 | goto err_pmu; | |
d010b332 | 7240 | } |
f344011c | 7241 | } |
9ee318a7 | 7242 | |
cdd6c482 | 7243 | return event; |
90983b16 FW |
7244 | |
7245 | err_pmu: | |
7246 | if (event->destroy) | |
7247 | event->destroy(event); | |
c464c76e | 7248 | module_put(pmu->module); |
90983b16 | 7249 | err_ns: |
79dff51e MF |
7250 | if (is_cgroup_event(event)) |
7251 | perf_detach_cgroup(event); | |
90983b16 FW |
7252 | if (event->ns) |
7253 | put_pid_ns(event->ns); | |
7254 | kfree(event); | |
7255 | ||
7256 | return ERR_PTR(err); | |
0793a61d TG |
7257 | } |
7258 | ||
cdd6c482 IM |
7259 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
7260 | struct perf_event_attr *attr) | |
974802ea | 7261 | { |
974802ea | 7262 | u32 size; |
cdf8073d | 7263 | int ret; |
974802ea PZ |
7264 | |
7265 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
7266 | return -EFAULT; | |
7267 | ||
7268 | /* | |
7269 | * zero the full structure, so that a short copy will be nice. | |
7270 | */ | |
7271 | memset(attr, 0, sizeof(*attr)); | |
7272 | ||
7273 | ret = get_user(size, &uattr->size); | |
7274 | if (ret) | |
7275 | return ret; | |
7276 | ||
7277 | if (size > PAGE_SIZE) /* silly large */ | |
7278 | goto err_size; | |
7279 | ||
7280 | if (!size) /* abi compat */ | |
7281 | size = PERF_ATTR_SIZE_VER0; | |
7282 | ||
7283 | if (size < PERF_ATTR_SIZE_VER0) | |
7284 | goto err_size; | |
7285 | ||
7286 | /* | |
7287 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
7288 | * ensure all the unknown bits are 0 - i.e. new |
7289 | * user-space does not rely on any kernel feature | |
7290 | * extensions we dont know about yet. | |
974802ea PZ |
7291 | */ |
7292 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
7293 | unsigned char __user *addr; |
7294 | unsigned char __user *end; | |
7295 | unsigned char val; | |
974802ea | 7296 | |
cdf8073d IS |
7297 | addr = (void __user *)uattr + sizeof(*attr); |
7298 | end = (void __user *)uattr + size; | |
974802ea | 7299 | |
cdf8073d | 7300 | for (; addr < end; addr++) { |
974802ea PZ |
7301 | ret = get_user(val, addr); |
7302 | if (ret) | |
7303 | return ret; | |
7304 | if (val) | |
7305 | goto err_size; | |
7306 | } | |
b3e62e35 | 7307 | size = sizeof(*attr); |
974802ea PZ |
7308 | } |
7309 | ||
7310 | ret = copy_from_user(attr, uattr, size); | |
7311 | if (ret) | |
7312 | return -EFAULT; | |
7313 | ||
cd757645 | 7314 | if (attr->__reserved_1) |
974802ea PZ |
7315 | return -EINVAL; |
7316 | ||
7317 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
7318 | return -EINVAL; | |
7319 | ||
7320 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
7321 | return -EINVAL; | |
7322 | ||
bce38cd5 SE |
7323 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
7324 | u64 mask = attr->branch_sample_type; | |
7325 | ||
7326 | /* only using defined bits */ | |
7327 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
7328 | return -EINVAL; | |
7329 | ||
7330 | /* at least one branch bit must be set */ | |
7331 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
7332 | return -EINVAL; | |
7333 | ||
bce38cd5 SE |
7334 | /* propagate priv level, when not set for branch */ |
7335 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
7336 | ||
7337 | /* exclude_kernel checked on syscall entry */ | |
7338 | if (!attr->exclude_kernel) | |
7339 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
7340 | ||
7341 | if (!attr->exclude_user) | |
7342 | mask |= PERF_SAMPLE_BRANCH_USER; | |
7343 | ||
7344 | if (!attr->exclude_hv) | |
7345 | mask |= PERF_SAMPLE_BRANCH_HV; | |
7346 | /* | |
7347 | * adjust user setting (for HW filter setup) | |
7348 | */ | |
7349 | attr->branch_sample_type = mask; | |
7350 | } | |
e712209a SE |
7351 | /* privileged levels capture (kernel, hv): check permissions */ |
7352 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
7353 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
7354 | return -EACCES; | |
bce38cd5 | 7355 | } |
4018994f | 7356 | |
c5ebcedb | 7357 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 7358 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
7359 | if (ret) |
7360 | return ret; | |
7361 | } | |
7362 | ||
7363 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
7364 | if (!arch_perf_have_user_stack_dump()) | |
7365 | return -ENOSYS; | |
7366 | ||
7367 | /* | |
7368 | * We have __u32 type for the size, but so far | |
7369 | * we can only use __u16 as maximum due to the | |
7370 | * __u16 sample size limit. | |
7371 | */ | |
7372 | if (attr->sample_stack_user >= USHRT_MAX) | |
7373 | ret = -EINVAL; | |
7374 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
7375 | ret = -EINVAL; | |
7376 | } | |
4018994f | 7377 | |
60e2364e SE |
7378 | if (attr->sample_type & PERF_SAMPLE_REGS_INTR) |
7379 | ret = perf_reg_validate(attr->sample_regs_intr); | |
974802ea PZ |
7380 | out: |
7381 | return ret; | |
7382 | ||
7383 | err_size: | |
7384 | put_user(sizeof(*attr), &uattr->size); | |
7385 | ret = -E2BIG; | |
7386 | goto out; | |
7387 | } | |
7388 | ||
ac9721f3 PZ |
7389 | static int |
7390 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 7391 | { |
b69cf536 | 7392 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
7393 | int ret = -EINVAL; |
7394 | ||
ac9721f3 | 7395 | if (!output_event) |
a4be7c27 PZ |
7396 | goto set; |
7397 | ||
ac9721f3 PZ |
7398 | /* don't allow circular references */ |
7399 | if (event == output_event) | |
a4be7c27 PZ |
7400 | goto out; |
7401 | ||
0f139300 PZ |
7402 | /* |
7403 | * Don't allow cross-cpu buffers | |
7404 | */ | |
7405 | if (output_event->cpu != event->cpu) | |
7406 | goto out; | |
7407 | ||
7408 | /* | |
76369139 | 7409 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
7410 | */ |
7411 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
7412 | goto out; | |
7413 | ||
a4be7c27 | 7414 | set: |
cdd6c482 | 7415 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7416 | /* Can't redirect output if we've got an active mmap() */ |
7417 | if (atomic_read(&event->mmap_count)) | |
7418 | goto unlock; | |
a4be7c27 | 7419 | |
ac9721f3 | 7420 | if (output_event) { |
76369139 FW |
7421 | /* get the rb we want to redirect to */ |
7422 | rb = ring_buffer_get(output_event); | |
7423 | if (!rb) | |
ac9721f3 | 7424 | goto unlock; |
a4be7c27 PZ |
7425 | } |
7426 | ||
b69cf536 | 7427 | ring_buffer_attach(event, rb); |
9bb5d40c | 7428 | |
a4be7c27 | 7429 | ret = 0; |
ac9721f3 PZ |
7430 | unlock: |
7431 | mutex_unlock(&event->mmap_mutex); | |
7432 | ||
a4be7c27 | 7433 | out: |
a4be7c27 PZ |
7434 | return ret; |
7435 | } | |
7436 | ||
f63a8daa PZ |
7437 | static void mutex_lock_double(struct mutex *a, struct mutex *b) |
7438 | { | |
7439 | if (b < a) | |
7440 | swap(a, b); | |
7441 | ||
7442 | mutex_lock(a); | |
7443 | mutex_lock_nested(b, SINGLE_DEPTH_NESTING); | |
7444 | } | |
7445 | ||
0793a61d | 7446 | /** |
cdd6c482 | 7447 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7448 | * |
cdd6c482 | 7449 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7450 | * @pid: target pid |
9f66a381 | 7451 | * @cpu: target cpu |
cdd6c482 | 7452 | * @group_fd: group leader event fd |
0793a61d | 7453 | */ |
cdd6c482 IM |
7454 | SYSCALL_DEFINE5(perf_event_open, |
7455 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7456 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7457 | { |
b04243ef PZ |
7458 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7459 | struct perf_event *event, *sibling; | |
cdd6c482 | 7460 | struct perf_event_attr attr; |
f63a8daa | 7461 | struct perf_event_context *ctx, *uninitialized_var(gctx); |
cdd6c482 | 7462 | struct file *event_file = NULL; |
2903ff01 | 7463 | struct fd group = {NULL, 0}; |
38a81da2 | 7464 | struct task_struct *task = NULL; |
89a1e187 | 7465 | struct pmu *pmu; |
ea635c64 | 7466 | int event_fd; |
b04243ef | 7467 | int move_group = 0; |
dc86cabe | 7468 | int err; |
a21b0b35 | 7469 | int f_flags = O_RDWR; |
79dff51e | 7470 | int cgroup_fd = -1; |
0793a61d | 7471 | |
2743a5b0 | 7472 | /* for future expandability... */ |
e5d1367f | 7473 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
7474 | return -EINVAL; |
7475 | ||
dc86cabe IM |
7476 | err = perf_copy_attr(attr_uptr, &attr); |
7477 | if (err) | |
7478 | return err; | |
eab656ae | 7479 | |
0764771d PZ |
7480 | if (!attr.exclude_kernel) { |
7481 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
7482 | return -EACCES; | |
7483 | } | |
7484 | ||
df58ab24 | 7485 | if (attr.freq) { |
cdd6c482 | 7486 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 7487 | return -EINVAL; |
0819b2e3 PZ |
7488 | } else { |
7489 | if (attr.sample_period & (1ULL << 63)) | |
7490 | return -EINVAL; | |
df58ab24 PZ |
7491 | } |
7492 | ||
e5d1367f SE |
7493 | /* |
7494 | * In cgroup mode, the pid argument is used to pass the fd | |
7495 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
7496 | * designates the cpu on which to monitor threads from that | |
7497 | * cgroup. | |
7498 | */ | |
7499 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
7500 | return -EINVAL; | |
7501 | ||
a21b0b35 YD |
7502 | if (flags & PERF_FLAG_FD_CLOEXEC) |
7503 | f_flags |= O_CLOEXEC; | |
7504 | ||
7505 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
7506 | if (event_fd < 0) |
7507 | return event_fd; | |
7508 | ||
ac9721f3 | 7509 | if (group_fd != -1) { |
2903ff01 AV |
7510 | err = perf_fget_light(group_fd, &group); |
7511 | if (err) | |
d14b12d7 | 7512 | goto err_fd; |
2903ff01 | 7513 | group_leader = group.file->private_data; |
ac9721f3 PZ |
7514 | if (flags & PERF_FLAG_FD_OUTPUT) |
7515 | output_event = group_leader; | |
7516 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
7517 | group_leader = NULL; | |
7518 | } | |
7519 | ||
e5d1367f | 7520 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
7521 | task = find_lively_task_by_vpid(pid); |
7522 | if (IS_ERR(task)) { | |
7523 | err = PTR_ERR(task); | |
7524 | goto err_group_fd; | |
7525 | } | |
7526 | } | |
7527 | ||
1f4ee503 PZ |
7528 | if (task && group_leader && |
7529 | group_leader->attr.inherit != attr.inherit) { | |
7530 | err = -EINVAL; | |
7531 | goto err_task; | |
7532 | } | |
7533 | ||
fbfc623f YZ |
7534 | get_online_cpus(); |
7535 | ||
79dff51e MF |
7536 | if (flags & PERF_FLAG_PID_CGROUP) |
7537 | cgroup_fd = pid; | |
7538 | ||
4dc0da86 | 7539 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
79dff51e | 7540 | NULL, NULL, cgroup_fd); |
d14b12d7 SE |
7541 | if (IS_ERR(event)) { |
7542 | err = PTR_ERR(event); | |
1f4ee503 | 7543 | goto err_cpus; |
d14b12d7 SE |
7544 | } |
7545 | ||
53b25335 VW |
7546 | if (is_sampling_event(event)) { |
7547 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
7548 | err = -ENOTSUPP; | |
7549 | goto err_alloc; | |
7550 | } | |
7551 | } | |
7552 | ||
766d6c07 FW |
7553 | account_event(event); |
7554 | ||
89a1e187 PZ |
7555 | /* |
7556 | * Special case software events and allow them to be part of | |
7557 | * any hardware group. | |
7558 | */ | |
7559 | pmu = event->pmu; | |
b04243ef PZ |
7560 | |
7561 | if (group_leader && | |
7562 | (is_software_event(event) != is_software_event(group_leader))) { | |
7563 | if (is_software_event(event)) { | |
7564 | /* | |
7565 | * If event and group_leader are not both a software | |
7566 | * event, and event is, then group leader is not. | |
7567 | * | |
7568 | * Allow the addition of software events to !software | |
7569 | * groups, this is safe because software events never | |
7570 | * fail to schedule. | |
7571 | */ | |
7572 | pmu = group_leader->pmu; | |
7573 | } else if (is_software_event(group_leader) && | |
7574 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7575 | /* | |
7576 | * In case the group is a pure software group, and we | |
7577 | * try to add a hardware event, move the whole group to | |
7578 | * the hardware context. | |
7579 | */ | |
7580 | move_group = 1; | |
7581 | } | |
7582 | } | |
89a1e187 PZ |
7583 | |
7584 | /* | |
7585 | * Get the target context (task or percpu): | |
7586 | */ | |
4af57ef2 | 7587 | ctx = find_get_context(pmu, task, event); |
89a1e187 PZ |
7588 | if (IS_ERR(ctx)) { |
7589 | err = PTR_ERR(ctx); | |
c6be5a5c | 7590 | goto err_alloc; |
89a1e187 PZ |
7591 | } |
7592 | ||
fd1edb3a PZ |
7593 | if (task) { |
7594 | put_task_struct(task); | |
7595 | task = NULL; | |
7596 | } | |
7597 | ||
ccff286d | 7598 | /* |
cdd6c482 | 7599 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7600 | */ |
ac9721f3 | 7601 | if (group_leader) { |
dc86cabe | 7602 | err = -EINVAL; |
04289bb9 | 7603 | |
04289bb9 | 7604 | /* |
ccff286d IM |
7605 | * Do not allow a recursive hierarchy (this new sibling |
7606 | * becoming part of another group-sibling): | |
7607 | */ | |
7608 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7609 | goto err_context; |
ccff286d IM |
7610 | /* |
7611 | * Do not allow to attach to a group in a different | |
7612 | * task or CPU context: | |
04289bb9 | 7613 | */ |
b04243ef | 7614 | if (move_group) { |
c3c87e77 PZ |
7615 | /* |
7616 | * Make sure we're both on the same task, or both | |
7617 | * per-cpu events. | |
7618 | */ | |
7619 | if (group_leader->ctx->task != ctx->task) | |
7620 | goto err_context; | |
7621 | ||
7622 | /* | |
7623 | * Make sure we're both events for the same CPU; | |
7624 | * grouping events for different CPUs is broken; since | |
7625 | * you can never concurrently schedule them anyhow. | |
7626 | */ | |
7627 | if (group_leader->cpu != event->cpu) | |
b04243ef PZ |
7628 | goto err_context; |
7629 | } else { | |
7630 | if (group_leader->ctx != ctx) | |
7631 | goto err_context; | |
7632 | } | |
7633 | ||
3b6f9e5c PM |
7634 | /* |
7635 | * Only a group leader can be exclusive or pinned | |
7636 | */ | |
0d48696f | 7637 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7638 | goto err_context; |
ac9721f3 PZ |
7639 | } |
7640 | ||
7641 | if (output_event) { | |
7642 | err = perf_event_set_output(event, output_event); | |
7643 | if (err) | |
c3f00c70 | 7644 | goto err_context; |
ac9721f3 | 7645 | } |
0793a61d | 7646 | |
a21b0b35 YD |
7647 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
7648 | f_flags); | |
ea635c64 AV |
7649 | if (IS_ERR(event_file)) { |
7650 | err = PTR_ERR(event_file); | |
c3f00c70 | 7651 | goto err_context; |
ea635c64 | 7652 | } |
9b51f66d | 7653 | |
b04243ef | 7654 | if (move_group) { |
f63a8daa PZ |
7655 | gctx = group_leader->ctx; |
7656 | ||
7657 | /* | |
7658 | * See perf_event_ctx_lock() for comments on the details | |
7659 | * of swizzling perf_event::ctx. | |
7660 | */ | |
7661 | mutex_lock_double(&gctx->mutex, &ctx->mutex); | |
b04243ef | 7662 | |
46ce0fe9 | 7663 | perf_remove_from_context(group_leader, false); |
0231bb53 | 7664 | |
b04243ef PZ |
7665 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7666 | group_entry) { | |
46ce0fe9 | 7667 | perf_remove_from_context(sibling, false); |
b04243ef PZ |
7668 | put_ctx(gctx); |
7669 | } | |
f63a8daa PZ |
7670 | } else { |
7671 | mutex_lock(&ctx->mutex); | |
ea635c64 | 7672 | } |
9b51f66d | 7673 | |
ad3a37de | 7674 | WARN_ON_ONCE(ctx->parent_ctx); |
b04243ef PZ |
7675 | |
7676 | if (move_group) { | |
f63a8daa PZ |
7677 | /* |
7678 | * Wait for everybody to stop referencing the events through | |
7679 | * the old lists, before installing it on new lists. | |
7680 | */ | |
0cda4c02 | 7681 | synchronize_rcu(); |
f63a8daa | 7682 | |
8f95b435 PZI |
7683 | /* |
7684 | * Install the group siblings before the group leader. | |
7685 | * | |
7686 | * Because a group leader will try and install the entire group | |
7687 | * (through the sibling list, which is still in-tact), we can | |
7688 | * end up with siblings installed in the wrong context. | |
7689 | * | |
7690 | * By installing siblings first we NO-OP because they're not | |
7691 | * reachable through the group lists. | |
7692 | */ | |
b04243ef PZ |
7693 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7694 | group_entry) { | |
8f95b435 | 7695 | perf_event__state_init(sibling); |
9fc81d87 | 7696 | perf_install_in_context(ctx, sibling, sibling->cpu); |
b04243ef PZ |
7697 | get_ctx(ctx); |
7698 | } | |
8f95b435 PZI |
7699 | |
7700 | /* | |
7701 | * Removing from the context ends up with disabled | |
7702 | * event. What we want here is event in the initial | |
7703 | * startup state, ready to be add into new context. | |
7704 | */ | |
7705 | perf_event__state_init(group_leader); | |
7706 | perf_install_in_context(ctx, group_leader, group_leader->cpu); | |
7707 | get_ctx(ctx); | |
b04243ef PZ |
7708 | } |
7709 | ||
e2d37cd2 | 7710 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 7711 | perf_unpin_context(ctx); |
f63a8daa PZ |
7712 | |
7713 | if (move_group) { | |
7714 | mutex_unlock(&gctx->mutex); | |
7715 | put_ctx(gctx); | |
7716 | } | |
d859e29f | 7717 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7718 | |
fbfc623f YZ |
7719 | put_online_cpus(); |
7720 | ||
cdd6c482 | 7721 | event->owner = current; |
8882135b | 7722 | |
cdd6c482 IM |
7723 | mutex_lock(¤t->perf_event_mutex); |
7724 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7725 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7726 | |
c320c7b7 ACM |
7727 | /* |
7728 | * Precalculate sample_data sizes | |
7729 | */ | |
7730 | perf_event__header_size(event); | |
6844c09d | 7731 | perf_event__id_header_size(event); |
c320c7b7 | 7732 | |
8a49542c PZ |
7733 | /* |
7734 | * Drop the reference on the group_event after placing the | |
7735 | * new event on the sibling_list. This ensures destruction | |
7736 | * of the group leader will find the pointer to itself in | |
7737 | * perf_group_detach(). | |
7738 | */ | |
2903ff01 | 7739 | fdput(group); |
ea635c64 AV |
7740 | fd_install(event_fd, event_file); |
7741 | return event_fd; | |
0793a61d | 7742 | |
c3f00c70 | 7743 | err_context: |
fe4b04fa | 7744 | perf_unpin_context(ctx); |
ea635c64 | 7745 | put_ctx(ctx); |
c6be5a5c | 7746 | err_alloc: |
ea635c64 | 7747 | free_event(event); |
1f4ee503 | 7748 | err_cpus: |
fbfc623f | 7749 | put_online_cpus(); |
1f4ee503 | 7750 | err_task: |
e7d0bc04 PZ |
7751 | if (task) |
7752 | put_task_struct(task); | |
89a1e187 | 7753 | err_group_fd: |
2903ff01 | 7754 | fdput(group); |
ea635c64 AV |
7755 | err_fd: |
7756 | put_unused_fd(event_fd); | |
dc86cabe | 7757 | return err; |
0793a61d TG |
7758 | } |
7759 | ||
fb0459d7 AV |
7760 | /** |
7761 | * perf_event_create_kernel_counter | |
7762 | * | |
7763 | * @attr: attributes of the counter to create | |
7764 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7765 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7766 | */ |
7767 | struct perf_event * | |
7768 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7769 | struct task_struct *task, |
4dc0da86 AK |
7770 | perf_overflow_handler_t overflow_handler, |
7771 | void *context) | |
fb0459d7 | 7772 | { |
fb0459d7 | 7773 | struct perf_event_context *ctx; |
c3f00c70 | 7774 | struct perf_event *event; |
fb0459d7 | 7775 | int err; |
d859e29f | 7776 | |
fb0459d7 AV |
7777 | /* |
7778 | * Get the target context (task or percpu): | |
7779 | */ | |
d859e29f | 7780 | |
4dc0da86 | 7781 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
79dff51e | 7782 | overflow_handler, context, -1); |
c3f00c70 PZ |
7783 | if (IS_ERR(event)) { |
7784 | err = PTR_ERR(event); | |
7785 | goto err; | |
7786 | } | |
d859e29f | 7787 | |
f8697762 JO |
7788 | /* Mark owner so we could distinguish it from user events. */ |
7789 | event->owner = EVENT_OWNER_KERNEL; | |
7790 | ||
766d6c07 FW |
7791 | account_event(event); |
7792 | ||
4af57ef2 | 7793 | ctx = find_get_context(event->pmu, task, event); |
c6567f64 FW |
7794 | if (IS_ERR(ctx)) { |
7795 | err = PTR_ERR(ctx); | |
c3f00c70 | 7796 | goto err_free; |
d859e29f | 7797 | } |
fb0459d7 | 7798 | |
fb0459d7 AV |
7799 | WARN_ON_ONCE(ctx->parent_ctx); |
7800 | mutex_lock(&ctx->mutex); | |
7801 | perf_install_in_context(ctx, event, cpu); | |
fe4b04fa | 7802 | perf_unpin_context(ctx); |
fb0459d7 AV |
7803 | mutex_unlock(&ctx->mutex); |
7804 | ||
fb0459d7 AV |
7805 | return event; |
7806 | ||
c3f00c70 PZ |
7807 | err_free: |
7808 | free_event(event); | |
7809 | err: | |
c6567f64 | 7810 | return ERR_PTR(err); |
9b51f66d | 7811 | } |
fb0459d7 | 7812 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7813 | |
0cda4c02 YZ |
7814 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7815 | { | |
7816 | struct perf_event_context *src_ctx; | |
7817 | struct perf_event_context *dst_ctx; | |
7818 | struct perf_event *event, *tmp; | |
7819 | LIST_HEAD(events); | |
7820 | ||
7821 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7822 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7823 | ||
f63a8daa PZ |
7824 | /* |
7825 | * See perf_event_ctx_lock() for comments on the details | |
7826 | * of swizzling perf_event::ctx. | |
7827 | */ | |
7828 | mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); | |
0cda4c02 YZ |
7829 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, |
7830 | event_entry) { | |
46ce0fe9 | 7831 | perf_remove_from_context(event, false); |
9a545de0 | 7832 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7833 | put_ctx(src_ctx); |
9886167d | 7834 | list_add(&event->migrate_entry, &events); |
0cda4c02 | 7835 | } |
0cda4c02 | 7836 | |
8f95b435 PZI |
7837 | /* |
7838 | * Wait for the events to quiesce before re-instating them. | |
7839 | */ | |
0cda4c02 YZ |
7840 | synchronize_rcu(); |
7841 | ||
8f95b435 PZI |
7842 | /* |
7843 | * Re-instate events in 2 passes. | |
7844 | * | |
7845 | * Skip over group leaders and only install siblings on this first | |
7846 | * pass, siblings will not get enabled without a leader, however a | |
7847 | * leader will enable its siblings, even if those are still on the old | |
7848 | * context. | |
7849 | */ | |
7850 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { | |
7851 | if (event->group_leader == event) | |
7852 | continue; | |
7853 | ||
7854 | list_del(&event->migrate_entry); | |
7855 | if (event->state >= PERF_EVENT_STATE_OFF) | |
7856 | event->state = PERF_EVENT_STATE_INACTIVE; | |
7857 | account_event_cpu(event, dst_cpu); | |
7858 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
7859 | get_ctx(dst_ctx); | |
7860 | } | |
7861 | ||
7862 | /* | |
7863 | * Once all the siblings are setup properly, install the group leaders | |
7864 | * to make it go. | |
7865 | */ | |
9886167d PZ |
7866 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7867 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7868 | if (event->state >= PERF_EVENT_STATE_OFF) |
7869 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7870 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7871 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7872 | get_ctx(dst_ctx); | |
7873 | } | |
7874 | mutex_unlock(&dst_ctx->mutex); | |
f63a8daa | 7875 | mutex_unlock(&src_ctx->mutex); |
0cda4c02 YZ |
7876 | } |
7877 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7878 | ||
cdd6c482 | 7879 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7880 | struct task_struct *child) |
d859e29f | 7881 | { |
cdd6c482 | 7882 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7883 | u64 child_val; |
d859e29f | 7884 | |
cdd6c482 IM |
7885 | if (child_event->attr.inherit_stat) |
7886 | perf_event_read_event(child_event, child); | |
38b200d6 | 7887 | |
b5e58793 | 7888 | child_val = perf_event_count(child_event); |
d859e29f PM |
7889 | |
7890 | /* | |
7891 | * Add back the child's count to the parent's count: | |
7892 | */ | |
a6e6dea6 | 7893 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7894 | atomic64_add(child_event->total_time_enabled, |
7895 | &parent_event->child_total_time_enabled); | |
7896 | atomic64_add(child_event->total_time_running, | |
7897 | &parent_event->child_total_time_running); | |
d859e29f PM |
7898 | |
7899 | /* | |
cdd6c482 | 7900 | * Remove this event from the parent's list |
d859e29f | 7901 | */ |
cdd6c482 IM |
7902 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7903 | mutex_lock(&parent_event->child_mutex); | |
7904 | list_del_init(&child_event->child_list); | |
7905 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f | 7906 | |
dc633982 JO |
7907 | /* |
7908 | * Make sure user/parent get notified, that we just | |
7909 | * lost one event. | |
7910 | */ | |
7911 | perf_event_wakeup(parent_event); | |
7912 | ||
d859e29f | 7913 | /* |
cdd6c482 | 7914 | * Release the parent event, if this was the last |
d859e29f PM |
7915 | * reference to it. |
7916 | */ | |
a6fa941d | 7917 | put_event(parent_event); |
d859e29f PM |
7918 | } |
7919 | ||
9b51f66d | 7920 | static void |
cdd6c482 IM |
7921 | __perf_event_exit_task(struct perf_event *child_event, |
7922 | struct perf_event_context *child_ctx, | |
38b200d6 | 7923 | struct task_struct *child) |
9b51f66d | 7924 | { |
1903d50c PZ |
7925 | /* |
7926 | * Do not destroy the 'original' grouping; because of the context | |
7927 | * switch optimization the original events could've ended up in a | |
7928 | * random child task. | |
7929 | * | |
7930 | * If we were to destroy the original group, all group related | |
7931 | * operations would cease to function properly after this random | |
7932 | * child dies. | |
7933 | * | |
7934 | * Do destroy all inherited groups, we don't care about those | |
7935 | * and being thorough is better. | |
7936 | */ | |
7937 | perf_remove_from_context(child_event, !!child_event->parent); | |
0cc0c027 | 7938 | |
9b51f66d | 7939 | /* |
38b435b1 | 7940 | * It can happen that the parent exits first, and has events |
9b51f66d | 7941 | * that are still around due to the child reference. These |
38b435b1 | 7942 | * events need to be zapped. |
9b51f66d | 7943 | */ |
38b435b1 | 7944 | if (child_event->parent) { |
cdd6c482 IM |
7945 | sync_child_event(child_event, child); |
7946 | free_event(child_event); | |
179033b3 JO |
7947 | } else { |
7948 | child_event->state = PERF_EVENT_STATE_EXIT; | |
7949 | perf_event_wakeup(child_event); | |
4bcf349a | 7950 | } |
9b51f66d IM |
7951 | } |
7952 | ||
8dc85d54 | 7953 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7954 | { |
ebf905fc | 7955 | struct perf_event *child_event, *next; |
211de6eb | 7956 | struct perf_event_context *child_ctx, *clone_ctx = NULL; |
a63eaf34 | 7957 | unsigned long flags; |
9b51f66d | 7958 | |
8dc85d54 | 7959 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7960 | perf_event_task(child, NULL, 0); |
9b51f66d | 7961 | return; |
9f498cc5 | 7962 | } |
9b51f66d | 7963 | |
a63eaf34 | 7964 | local_irq_save(flags); |
ad3a37de PM |
7965 | /* |
7966 | * We can't reschedule here because interrupts are disabled, | |
7967 | * and either child is current or it is a task that can't be | |
7968 | * scheduled, so we are now safe from rescheduling changing | |
7969 | * our context. | |
7970 | */ | |
806839b2 | 7971 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7972 | |
7973 | /* | |
7974 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7975 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7976 | * incremented the context's refcount before we do put_ctx below. |
7977 | */ | |
e625cce1 | 7978 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7979 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7980 | child->perf_event_ctxp[ctxn] = NULL; |
4a1c0f26 | 7981 | |
71a851b4 PZ |
7982 | /* |
7983 | * If this context is a clone; unclone it so it can't get | |
7984 | * swapped to another process while we're removing all | |
cdd6c482 | 7985 | * the events from it. |
71a851b4 | 7986 | */ |
211de6eb | 7987 | clone_ctx = unclone_ctx(child_ctx); |
5e942bb3 | 7988 | update_context_time(child_ctx); |
e625cce1 | 7989 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 | 7990 | |
211de6eb PZ |
7991 | if (clone_ctx) |
7992 | put_ctx(clone_ctx); | |
4a1c0f26 | 7993 | |
9f498cc5 | 7994 | /* |
cdd6c482 IM |
7995 | * Report the task dead after unscheduling the events so that we |
7996 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7997 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7998 | */ |
cdd6c482 | 7999 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 8000 | |
66fff224 PZ |
8001 | /* |
8002 | * We can recurse on the same lock type through: | |
8003 | * | |
cdd6c482 IM |
8004 | * __perf_event_exit_task() |
8005 | * sync_child_event() | |
a6fa941d AV |
8006 | * put_event() |
8007 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
8008 | * |
8009 | * But since its the parent context it won't be the same instance. | |
8010 | */ | |
a0507c84 | 8011 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 8012 | |
ebf905fc | 8013 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 8014 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 8015 | |
a63eaf34 PM |
8016 | mutex_unlock(&child_ctx->mutex); |
8017 | ||
8018 | put_ctx(child_ctx); | |
9b51f66d IM |
8019 | } |
8020 | ||
8dc85d54 PZ |
8021 | /* |
8022 | * When a child task exits, feed back event values to parent events. | |
8023 | */ | |
8024 | void perf_event_exit_task(struct task_struct *child) | |
8025 | { | |
8882135b | 8026 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
8027 | int ctxn; |
8028 | ||
8882135b PZ |
8029 | mutex_lock(&child->perf_event_mutex); |
8030 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
8031 | owner_entry) { | |
8032 | list_del_init(&event->owner_entry); | |
8033 | ||
8034 | /* | |
8035 | * Ensure the list deletion is visible before we clear | |
8036 | * the owner, closes a race against perf_release() where | |
8037 | * we need to serialize on the owner->perf_event_mutex. | |
8038 | */ | |
8039 | smp_wmb(); | |
8040 | event->owner = NULL; | |
8041 | } | |
8042 | mutex_unlock(&child->perf_event_mutex); | |
8043 | ||
8dc85d54 PZ |
8044 | for_each_task_context_nr(ctxn) |
8045 | perf_event_exit_task_context(child, ctxn); | |
8046 | } | |
8047 | ||
889ff015 FW |
8048 | static void perf_free_event(struct perf_event *event, |
8049 | struct perf_event_context *ctx) | |
8050 | { | |
8051 | struct perf_event *parent = event->parent; | |
8052 | ||
8053 | if (WARN_ON_ONCE(!parent)) | |
8054 | return; | |
8055 | ||
8056 | mutex_lock(&parent->child_mutex); | |
8057 | list_del_init(&event->child_list); | |
8058 | mutex_unlock(&parent->child_mutex); | |
8059 | ||
a6fa941d | 8060 | put_event(parent); |
889ff015 | 8061 | |
652884fe | 8062 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 8063 | perf_group_detach(event); |
889ff015 | 8064 | list_del_event(event, ctx); |
652884fe | 8065 | raw_spin_unlock_irq(&ctx->lock); |
889ff015 FW |
8066 | free_event(event); |
8067 | } | |
8068 | ||
bbbee908 | 8069 | /* |
652884fe | 8070 | * Free an unexposed, unused context as created by inheritance by |
8dc85d54 | 8071 | * perf_event_init_task below, used by fork() in case of fail. |
652884fe PZ |
8072 | * |
8073 | * Not all locks are strictly required, but take them anyway to be nice and | |
8074 | * help out with the lockdep assertions. | |
bbbee908 | 8075 | */ |
cdd6c482 | 8076 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 8077 | { |
8dc85d54 | 8078 | struct perf_event_context *ctx; |
cdd6c482 | 8079 | struct perf_event *event, *tmp; |
8dc85d54 | 8080 | int ctxn; |
bbbee908 | 8081 | |
8dc85d54 PZ |
8082 | for_each_task_context_nr(ctxn) { |
8083 | ctx = task->perf_event_ctxp[ctxn]; | |
8084 | if (!ctx) | |
8085 | continue; | |
bbbee908 | 8086 | |
8dc85d54 | 8087 | mutex_lock(&ctx->mutex); |
bbbee908 | 8088 | again: |
8dc85d54 PZ |
8089 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
8090 | group_entry) | |
8091 | perf_free_event(event, ctx); | |
bbbee908 | 8092 | |
8dc85d54 PZ |
8093 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
8094 | group_entry) | |
8095 | perf_free_event(event, ctx); | |
bbbee908 | 8096 | |
8dc85d54 PZ |
8097 | if (!list_empty(&ctx->pinned_groups) || |
8098 | !list_empty(&ctx->flexible_groups)) | |
8099 | goto again; | |
bbbee908 | 8100 | |
8dc85d54 | 8101 | mutex_unlock(&ctx->mutex); |
bbbee908 | 8102 | |
8dc85d54 PZ |
8103 | put_ctx(ctx); |
8104 | } | |
889ff015 FW |
8105 | } |
8106 | ||
4e231c79 PZ |
8107 | void perf_event_delayed_put(struct task_struct *task) |
8108 | { | |
8109 | int ctxn; | |
8110 | ||
8111 | for_each_task_context_nr(ctxn) | |
8112 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
8113 | } | |
8114 | ||
97dee4f3 PZ |
8115 | /* |
8116 | * inherit a event from parent task to child task: | |
8117 | */ | |
8118 | static struct perf_event * | |
8119 | inherit_event(struct perf_event *parent_event, | |
8120 | struct task_struct *parent, | |
8121 | struct perf_event_context *parent_ctx, | |
8122 | struct task_struct *child, | |
8123 | struct perf_event *group_leader, | |
8124 | struct perf_event_context *child_ctx) | |
8125 | { | |
1929def9 | 8126 | enum perf_event_active_state parent_state = parent_event->state; |
97dee4f3 | 8127 | struct perf_event *child_event; |
cee010ec | 8128 | unsigned long flags; |
97dee4f3 PZ |
8129 | |
8130 | /* | |
8131 | * Instead of creating recursive hierarchies of events, | |
8132 | * we link inherited events back to the original parent, | |
8133 | * which has a filp for sure, which we use as the reference | |
8134 | * count: | |
8135 | */ | |
8136 | if (parent_event->parent) | |
8137 | parent_event = parent_event->parent; | |
8138 | ||
8139 | child_event = perf_event_alloc(&parent_event->attr, | |
8140 | parent_event->cpu, | |
d580ff86 | 8141 | child, |
97dee4f3 | 8142 | group_leader, parent_event, |
79dff51e | 8143 | NULL, NULL, -1); |
97dee4f3 PZ |
8144 | if (IS_ERR(child_event)) |
8145 | return child_event; | |
a6fa941d | 8146 | |
fadfe7be JO |
8147 | if (is_orphaned_event(parent_event) || |
8148 | !atomic_long_inc_not_zero(&parent_event->refcount)) { | |
a6fa941d AV |
8149 | free_event(child_event); |
8150 | return NULL; | |
8151 | } | |
8152 | ||
97dee4f3 PZ |
8153 | get_ctx(child_ctx); |
8154 | ||
8155 | /* | |
8156 | * Make the child state follow the state of the parent event, | |
8157 | * not its attr.disabled bit. We hold the parent's mutex, | |
8158 | * so we won't race with perf_event_{en, dis}able_family. | |
8159 | */ | |
1929def9 | 8160 | if (parent_state >= PERF_EVENT_STATE_INACTIVE) |
97dee4f3 PZ |
8161 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
8162 | else | |
8163 | child_event->state = PERF_EVENT_STATE_OFF; | |
8164 | ||
8165 | if (parent_event->attr.freq) { | |
8166 | u64 sample_period = parent_event->hw.sample_period; | |
8167 | struct hw_perf_event *hwc = &child_event->hw; | |
8168 | ||
8169 | hwc->sample_period = sample_period; | |
8170 | hwc->last_period = sample_period; | |
8171 | ||
8172 | local64_set(&hwc->period_left, sample_period); | |
8173 | } | |
8174 | ||
8175 | child_event->ctx = child_ctx; | |
8176 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
8177 | child_event->overflow_handler_context |
8178 | = parent_event->overflow_handler_context; | |
97dee4f3 | 8179 | |
614b6780 TG |
8180 | /* |
8181 | * Precalculate sample_data sizes | |
8182 | */ | |
8183 | perf_event__header_size(child_event); | |
6844c09d | 8184 | perf_event__id_header_size(child_event); |
614b6780 | 8185 | |
97dee4f3 PZ |
8186 | /* |
8187 | * Link it up in the child's context: | |
8188 | */ | |
cee010ec | 8189 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 8190 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 8191 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 8192 | |
97dee4f3 PZ |
8193 | /* |
8194 | * Link this into the parent event's child list | |
8195 | */ | |
8196 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
8197 | mutex_lock(&parent_event->child_mutex); | |
8198 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
8199 | mutex_unlock(&parent_event->child_mutex); | |
8200 | ||
8201 | return child_event; | |
8202 | } | |
8203 | ||
8204 | static int inherit_group(struct perf_event *parent_event, | |
8205 | struct task_struct *parent, | |
8206 | struct perf_event_context *parent_ctx, | |
8207 | struct task_struct *child, | |
8208 | struct perf_event_context *child_ctx) | |
8209 | { | |
8210 | struct perf_event *leader; | |
8211 | struct perf_event *sub; | |
8212 | struct perf_event *child_ctr; | |
8213 | ||
8214 | leader = inherit_event(parent_event, parent, parent_ctx, | |
8215 | child, NULL, child_ctx); | |
8216 | if (IS_ERR(leader)) | |
8217 | return PTR_ERR(leader); | |
8218 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
8219 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
8220 | child, leader, child_ctx); | |
8221 | if (IS_ERR(child_ctr)) | |
8222 | return PTR_ERR(child_ctr); | |
8223 | } | |
8224 | return 0; | |
889ff015 FW |
8225 | } |
8226 | ||
8227 | static int | |
8228 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
8229 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 8230 | struct task_struct *child, int ctxn, |
889ff015 FW |
8231 | int *inherited_all) |
8232 | { | |
8233 | int ret; | |
8dc85d54 | 8234 | struct perf_event_context *child_ctx; |
889ff015 FW |
8235 | |
8236 | if (!event->attr.inherit) { | |
8237 | *inherited_all = 0; | |
8238 | return 0; | |
bbbee908 PZ |
8239 | } |
8240 | ||
fe4b04fa | 8241 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
8242 | if (!child_ctx) { |
8243 | /* | |
8244 | * This is executed from the parent task context, so | |
8245 | * inherit events that have been marked for cloning. | |
8246 | * First allocate and initialize a context for the | |
8247 | * child. | |
8248 | */ | |
bbbee908 | 8249 | |
734df5ab | 8250 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
8251 | if (!child_ctx) |
8252 | return -ENOMEM; | |
bbbee908 | 8253 | |
8dc85d54 | 8254 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
8255 | } |
8256 | ||
8257 | ret = inherit_group(event, parent, parent_ctx, | |
8258 | child, child_ctx); | |
8259 | ||
8260 | if (ret) | |
8261 | *inherited_all = 0; | |
8262 | ||
8263 | return ret; | |
bbbee908 PZ |
8264 | } |
8265 | ||
9b51f66d | 8266 | /* |
cdd6c482 | 8267 | * Initialize the perf_event context in task_struct |
9b51f66d | 8268 | */ |
985c8dcb | 8269 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 8270 | { |
889ff015 | 8271 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
8272 | struct perf_event_context *cloned_ctx; |
8273 | struct perf_event *event; | |
9b51f66d | 8274 | struct task_struct *parent = current; |
564c2b21 | 8275 | int inherited_all = 1; |
dddd3379 | 8276 | unsigned long flags; |
6ab423e0 | 8277 | int ret = 0; |
9b51f66d | 8278 | |
8dc85d54 | 8279 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
8280 | return 0; |
8281 | ||
ad3a37de | 8282 | /* |
25346b93 PM |
8283 | * If the parent's context is a clone, pin it so it won't get |
8284 | * swapped under us. | |
ad3a37de | 8285 | */ |
8dc85d54 | 8286 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
8287 | if (!parent_ctx) |
8288 | return 0; | |
25346b93 | 8289 | |
ad3a37de PM |
8290 | /* |
8291 | * No need to check if parent_ctx != NULL here; since we saw | |
8292 | * it non-NULL earlier, the only reason for it to become NULL | |
8293 | * is if we exit, and since we're currently in the middle of | |
8294 | * a fork we can't be exiting at the same time. | |
8295 | */ | |
ad3a37de | 8296 | |
9b51f66d IM |
8297 | /* |
8298 | * Lock the parent list. No need to lock the child - not PID | |
8299 | * hashed yet and not running, so nobody can access it. | |
8300 | */ | |
d859e29f | 8301 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
8302 | |
8303 | /* | |
8304 | * We dont have to disable NMIs - we are only looking at | |
8305 | * the list, not manipulating it: | |
8306 | */ | |
889ff015 | 8307 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
8308 | ret = inherit_task_group(event, parent, parent_ctx, |
8309 | child, ctxn, &inherited_all); | |
889ff015 FW |
8310 | if (ret) |
8311 | break; | |
8312 | } | |
b93f7978 | 8313 | |
dddd3379 TG |
8314 | /* |
8315 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
8316 | * to allocations, but we need to prevent rotation because | |
8317 | * rotate_ctx() will change the list from interrupt context. | |
8318 | */ | |
8319 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
8320 | parent_ctx->rotate_disable = 1; | |
8321 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
8322 | ||
889ff015 | 8323 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
8324 | ret = inherit_task_group(event, parent, parent_ctx, |
8325 | child, ctxn, &inherited_all); | |
889ff015 | 8326 | if (ret) |
9b51f66d | 8327 | break; |
564c2b21 PM |
8328 | } |
8329 | ||
dddd3379 TG |
8330 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
8331 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 8332 | |
8dc85d54 | 8333 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 8334 | |
05cbaa28 | 8335 | if (child_ctx && inherited_all) { |
564c2b21 PM |
8336 | /* |
8337 | * Mark the child context as a clone of the parent | |
8338 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
8339 | * |
8340 | * Note that if the parent is a clone, the holding of | |
8341 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 8342 | */ |
c5ed5145 | 8343 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
8344 | if (cloned_ctx) { |
8345 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 8346 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
8347 | } else { |
8348 | child_ctx->parent_ctx = parent_ctx; | |
8349 | child_ctx->parent_gen = parent_ctx->generation; | |
8350 | } | |
8351 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
8352 | } |
8353 | ||
c5ed5145 | 8354 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 8355 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 8356 | |
25346b93 | 8357 | perf_unpin_context(parent_ctx); |
fe4b04fa | 8358 | put_ctx(parent_ctx); |
ad3a37de | 8359 | |
6ab423e0 | 8360 | return ret; |
9b51f66d IM |
8361 | } |
8362 | ||
8dc85d54 PZ |
8363 | /* |
8364 | * Initialize the perf_event context in task_struct | |
8365 | */ | |
8366 | int perf_event_init_task(struct task_struct *child) | |
8367 | { | |
8368 | int ctxn, ret; | |
8369 | ||
8550d7cb ON |
8370 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
8371 | mutex_init(&child->perf_event_mutex); | |
8372 | INIT_LIST_HEAD(&child->perf_event_list); | |
8373 | ||
8dc85d54 PZ |
8374 | for_each_task_context_nr(ctxn) { |
8375 | ret = perf_event_init_context(child, ctxn); | |
6c72e350 PZ |
8376 | if (ret) { |
8377 | perf_event_free_task(child); | |
8dc85d54 | 8378 | return ret; |
6c72e350 | 8379 | } |
8dc85d54 PZ |
8380 | } |
8381 | ||
8382 | return 0; | |
8383 | } | |
8384 | ||
220b140b PM |
8385 | static void __init perf_event_init_all_cpus(void) |
8386 | { | |
b28ab83c | 8387 | struct swevent_htable *swhash; |
220b140b | 8388 | int cpu; |
220b140b PM |
8389 | |
8390 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
8391 | swhash = &per_cpu(swevent_htable, cpu); |
8392 | mutex_init(&swhash->hlist_mutex); | |
2fde4f94 | 8393 | INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu)); |
220b140b PM |
8394 | } |
8395 | } | |
8396 | ||
0db0628d | 8397 | static void perf_event_init_cpu(int cpu) |
0793a61d | 8398 | { |
108b02cf | 8399 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 8400 | |
b28ab83c | 8401 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8402 | swhash->online = true; |
4536e4d1 | 8403 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
8404 | struct swevent_hlist *hlist; |
8405 | ||
b28ab83c PZ |
8406 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
8407 | WARN_ON(!hlist); | |
8408 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 8409 | } |
b28ab83c | 8410 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
8411 | } |
8412 | ||
c277443c | 8413 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
108b02cf | 8414 | static void __perf_event_exit_context(void *__info) |
0793a61d | 8415 | { |
226424ee | 8416 | struct remove_event re = { .detach_group = true }; |
108b02cf | 8417 | struct perf_event_context *ctx = __info; |
0793a61d | 8418 | |
e3703f8c | 8419 | rcu_read_lock(); |
46ce0fe9 PZ |
8420 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
8421 | __perf_remove_from_context(&re); | |
e3703f8c | 8422 | rcu_read_unlock(); |
0793a61d | 8423 | } |
108b02cf PZ |
8424 | |
8425 | static void perf_event_exit_cpu_context(int cpu) | |
8426 | { | |
8427 | struct perf_event_context *ctx; | |
8428 | struct pmu *pmu; | |
8429 | int idx; | |
8430 | ||
8431 | idx = srcu_read_lock(&pmus_srcu); | |
8432 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 8433 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
8434 | |
8435 | mutex_lock(&ctx->mutex); | |
8436 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
8437 | mutex_unlock(&ctx->mutex); | |
8438 | } | |
8439 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
8440 | } |
8441 | ||
cdd6c482 | 8442 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 8443 | { |
b28ab83c | 8444 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 8445 | |
e3703f8c PZ |
8446 | perf_event_exit_cpu_context(cpu); |
8447 | ||
b28ab83c | 8448 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8449 | swhash->online = false; |
b28ab83c PZ |
8450 | swevent_hlist_release(swhash); |
8451 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
8452 | } |
8453 | #else | |
cdd6c482 | 8454 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
8455 | #endif |
8456 | ||
c277443c PZ |
8457 | static int |
8458 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
8459 | { | |
8460 | int cpu; | |
8461 | ||
8462 | for_each_online_cpu(cpu) | |
8463 | perf_event_exit_cpu(cpu); | |
8464 | ||
8465 | return NOTIFY_OK; | |
8466 | } | |
8467 | ||
8468 | /* | |
8469 | * Run the perf reboot notifier at the very last possible moment so that | |
8470 | * the generic watchdog code runs as long as possible. | |
8471 | */ | |
8472 | static struct notifier_block perf_reboot_notifier = { | |
8473 | .notifier_call = perf_reboot, | |
8474 | .priority = INT_MIN, | |
8475 | }; | |
8476 | ||
0db0628d | 8477 | static int |
0793a61d TG |
8478 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
8479 | { | |
8480 | unsigned int cpu = (long)hcpu; | |
8481 | ||
4536e4d1 | 8482 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
8483 | |
8484 | case CPU_UP_PREPARE: | |
5e11637e | 8485 | case CPU_DOWN_FAILED: |
cdd6c482 | 8486 | perf_event_init_cpu(cpu); |
0793a61d TG |
8487 | break; |
8488 | ||
5e11637e | 8489 | case CPU_UP_CANCELED: |
0793a61d | 8490 | case CPU_DOWN_PREPARE: |
cdd6c482 | 8491 | perf_event_exit_cpu(cpu); |
0793a61d | 8492 | break; |
0793a61d TG |
8493 | default: |
8494 | break; | |
8495 | } | |
8496 | ||
8497 | return NOTIFY_OK; | |
8498 | } | |
8499 | ||
cdd6c482 | 8500 | void __init perf_event_init(void) |
0793a61d | 8501 | { |
3c502e7a JW |
8502 | int ret; |
8503 | ||
2e80a82a PZ |
8504 | idr_init(&pmu_idr); |
8505 | ||
220b140b | 8506 | perf_event_init_all_cpus(); |
b0a873eb | 8507 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
8508 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
8509 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
8510 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
8511 | perf_tp_register(); |
8512 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 8513 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
8514 | |
8515 | ret = init_hw_breakpoint(); | |
8516 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
8517 | |
8518 | /* do not patch jump label more than once per second */ | |
8519 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
8520 | |
8521 | /* | |
8522 | * Build time assertion that we keep the data_head at the intended | |
8523 | * location. IOW, validation we got the __reserved[] size right. | |
8524 | */ | |
8525 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
8526 | != 1024); | |
0793a61d | 8527 | } |
abe43400 | 8528 | |
fd979c01 CS |
8529 | ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr, |
8530 | char *page) | |
8531 | { | |
8532 | struct perf_pmu_events_attr *pmu_attr = | |
8533 | container_of(attr, struct perf_pmu_events_attr, attr); | |
8534 | ||
8535 | if (pmu_attr->event_str) | |
8536 | return sprintf(page, "%s\n", pmu_attr->event_str); | |
8537 | ||
8538 | return 0; | |
8539 | } | |
8540 | ||
abe43400 PZ |
8541 | static int __init perf_event_sysfs_init(void) |
8542 | { | |
8543 | struct pmu *pmu; | |
8544 | int ret; | |
8545 | ||
8546 | mutex_lock(&pmus_lock); | |
8547 | ||
8548 | ret = bus_register(&pmu_bus); | |
8549 | if (ret) | |
8550 | goto unlock; | |
8551 | ||
8552 | list_for_each_entry(pmu, &pmus, entry) { | |
8553 | if (!pmu->name || pmu->type < 0) | |
8554 | continue; | |
8555 | ||
8556 | ret = pmu_dev_alloc(pmu); | |
8557 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
8558 | } | |
8559 | pmu_bus_running = 1; | |
8560 | ret = 0; | |
8561 | ||
8562 | unlock: | |
8563 | mutex_unlock(&pmus_lock); | |
8564 | ||
8565 | return ret; | |
8566 | } | |
8567 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
8568 | |
8569 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
8570 | static struct cgroup_subsys_state * |
8571 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
8572 | { |
8573 | struct perf_cgroup *jc; | |
e5d1367f | 8574 | |
1b15d055 | 8575 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
8576 | if (!jc) |
8577 | return ERR_PTR(-ENOMEM); | |
8578 | ||
e5d1367f SE |
8579 | jc->info = alloc_percpu(struct perf_cgroup_info); |
8580 | if (!jc->info) { | |
8581 | kfree(jc); | |
8582 | return ERR_PTR(-ENOMEM); | |
8583 | } | |
8584 | ||
e5d1367f SE |
8585 | return &jc->css; |
8586 | } | |
8587 | ||
eb95419b | 8588 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 8589 | { |
eb95419b TH |
8590 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
8591 | ||
e5d1367f SE |
8592 | free_percpu(jc->info); |
8593 | kfree(jc); | |
8594 | } | |
8595 | ||
8596 | static int __perf_cgroup_move(void *info) | |
8597 | { | |
8598 | struct task_struct *task = info; | |
8599 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
8600 | return 0; | |
8601 | } | |
8602 | ||
eb95419b TH |
8603 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
8604 | struct cgroup_taskset *tset) | |
e5d1367f | 8605 | { |
bb9d97b6 TH |
8606 | struct task_struct *task; |
8607 | ||
924f0d9a | 8608 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 8609 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8610 | } |
8611 | ||
eb95419b TH |
8612 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
8613 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 8614 | struct task_struct *task) |
e5d1367f SE |
8615 | { |
8616 | /* | |
8617 | * cgroup_exit() is called in the copy_process() failure path. | |
8618 | * Ignore this case since the task hasn't ran yet, this avoids | |
8619 | * trying to poke a half freed task state from generic code. | |
8620 | */ | |
8621 | if (!(task->flags & PF_EXITING)) | |
8622 | return; | |
8623 | ||
bb9d97b6 | 8624 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
8625 | } |
8626 | ||
073219e9 | 8627 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
8628 | .css_alloc = perf_cgroup_css_alloc, |
8629 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 8630 | .exit = perf_cgroup_exit, |
bb9d97b6 | 8631 | .attach = perf_cgroup_attach, |
e5d1367f SE |
8632 | }; |
8633 | #endif /* CONFIG_CGROUP_PERF */ |