Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
39bed6cb | 37 | #include <linux/cgroup.h> |
cdd6c482 | 38 | #include <linux/perf_event.h> |
af658dca | 39 | #include <linux/trace_events.h> |
3c502e7a | 40 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 41 | #include <linux/mm_types.h> |
c464c76e | 42 | #include <linux/module.h> |
f972eb63 | 43 | #include <linux/mman.h> |
b3f20785 | 44 | #include <linux/compat.h> |
2541517c AS |
45 | #include <linux/bpf.h> |
46 | #include <linux/filter.h> | |
0793a61d | 47 | |
76369139 FW |
48 | #include "internal.h" |
49 | ||
4e193bd4 TB |
50 | #include <asm/irq_regs.h> |
51 | ||
fadfe7be JO |
52 | static struct workqueue_struct *perf_wq; |
53 | ||
272325c4 PZ |
54 | typedef int (*remote_function_f)(void *); |
55 | ||
fe4b04fa | 56 | struct remote_function_call { |
e7e7ee2e | 57 | struct task_struct *p; |
272325c4 | 58 | remote_function_f func; |
e7e7ee2e IM |
59 | void *info; |
60 | int ret; | |
fe4b04fa PZ |
61 | }; |
62 | ||
63 | static void remote_function(void *data) | |
64 | { | |
65 | struct remote_function_call *tfc = data; | |
66 | struct task_struct *p = tfc->p; | |
67 | ||
68 | if (p) { | |
69 | tfc->ret = -EAGAIN; | |
70 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
71 | return; | |
72 | } | |
73 | ||
74 | tfc->ret = tfc->func(tfc->info); | |
75 | } | |
76 | ||
77 | /** | |
78 | * task_function_call - call a function on the cpu on which a task runs | |
79 | * @p: the task to evaluate | |
80 | * @func: the function to be called | |
81 | * @info: the function call argument | |
82 | * | |
83 | * Calls the function @func when the task is currently running. This might | |
84 | * be on the current CPU, which just calls the function directly | |
85 | * | |
86 | * returns: @func return value, or | |
87 | * -ESRCH - when the process isn't running | |
88 | * -EAGAIN - when the process moved away | |
89 | */ | |
90 | static int | |
272325c4 | 91 | task_function_call(struct task_struct *p, remote_function_f func, void *info) |
fe4b04fa PZ |
92 | { |
93 | struct remote_function_call data = { | |
e7e7ee2e IM |
94 | .p = p, |
95 | .func = func, | |
96 | .info = info, | |
97 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
98 | }; |
99 | ||
100 | if (task_curr(p)) | |
101 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
102 | ||
103 | return data.ret; | |
104 | } | |
105 | ||
106 | /** | |
107 | * cpu_function_call - call a function on the cpu | |
108 | * @func: the function to be called | |
109 | * @info: the function call argument | |
110 | * | |
111 | * Calls the function @func on the remote cpu. | |
112 | * | |
113 | * returns: @func return value or -ENXIO when the cpu is offline | |
114 | */ | |
272325c4 | 115 | static int cpu_function_call(int cpu, remote_function_f func, void *info) |
fe4b04fa PZ |
116 | { |
117 | struct remote_function_call data = { | |
e7e7ee2e IM |
118 | .p = NULL, |
119 | .func = func, | |
120 | .info = info, | |
121 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
122 | }; |
123 | ||
124 | smp_call_function_single(cpu, remote_function, &data, 1); | |
125 | ||
126 | return data.ret; | |
127 | } | |
128 | ||
f8697762 JO |
129 | #define EVENT_OWNER_KERNEL ((void *) -1) |
130 | ||
131 | static bool is_kernel_event(struct perf_event *event) | |
132 | { | |
133 | return event->owner == EVENT_OWNER_KERNEL; | |
134 | } | |
135 | ||
e5d1367f SE |
136 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
137 | PERF_FLAG_FD_OUTPUT |\ | |
a21b0b35 YD |
138 | PERF_FLAG_PID_CGROUP |\ |
139 | PERF_FLAG_FD_CLOEXEC) | |
e5d1367f | 140 | |
bce38cd5 SE |
141 | /* |
142 | * branch priv levels that need permission checks | |
143 | */ | |
144 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
145 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
146 | PERF_SAMPLE_BRANCH_HV) | |
147 | ||
0b3fcf17 SE |
148 | enum event_type_t { |
149 | EVENT_FLEXIBLE = 0x1, | |
150 | EVENT_PINNED = 0x2, | |
151 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
152 | }; | |
153 | ||
e5d1367f SE |
154 | /* |
155 | * perf_sched_events : >0 events exist | |
156 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
157 | */ | |
c5905afb | 158 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 159 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
ba532500 | 160 | static DEFINE_PER_CPU(int, perf_sched_cb_usages); |
e5d1367f | 161 | |
cdd6c482 IM |
162 | static atomic_t nr_mmap_events __read_mostly; |
163 | static atomic_t nr_comm_events __read_mostly; | |
164 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 165 | static atomic_t nr_freq_events __read_mostly; |
45ac1403 | 166 | static atomic_t nr_switch_events __read_mostly; |
9ee318a7 | 167 | |
108b02cf PZ |
168 | static LIST_HEAD(pmus); |
169 | static DEFINE_MUTEX(pmus_lock); | |
170 | static struct srcu_struct pmus_srcu; | |
171 | ||
0764771d | 172 | /* |
cdd6c482 | 173 | * perf event paranoia level: |
0fbdea19 IM |
174 | * -1 - not paranoid at all |
175 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 176 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 177 | * 2 - disallow kernel profiling for unpriv |
0764771d | 178 | */ |
cdd6c482 | 179 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 180 | |
20443384 FW |
181 | /* Minimum for 512 kiB + 1 user control page */ |
182 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
183 | |
184 | /* | |
cdd6c482 | 185 | * max perf event sample rate |
df58ab24 | 186 | */ |
14c63f17 DH |
187 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
188 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
189 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
190 | ||
191 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
192 | ||
193 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
194 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
195 | ||
d9494cb4 PZ |
196 | static int perf_sample_allowed_ns __read_mostly = |
197 | DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100; | |
14c63f17 DH |
198 | |
199 | void update_perf_cpu_limits(void) | |
200 | { | |
201 | u64 tmp = perf_sample_period_ns; | |
202 | ||
203 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 204 | do_div(tmp, 100); |
d9494cb4 | 205 | ACCESS_ONCE(perf_sample_allowed_ns) = tmp; |
14c63f17 | 206 | } |
163ec435 | 207 | |
9e630205 SE |
208 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
209 | ||
163ec435 PZ |
210 | int perf_proc_update_handler(struct ctl_table *table, int write, |
211 | void __user *buffer, size_t *lenp, | |
212 | loff_t *ppos) | |
213 | { | |
723478c8 | 214 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
163ec435 PZ |
215 | |
216 | if (ret || !write) | |
217 | return ret; | |
218 | ||
219 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
220 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
221 | update_perf_cpu_limits(); | |
222 | ||
223 | return 0; | |
224 | } | |
225 | ||
226 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
227 | ||
228 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
229 | void __user *buffer, size_t *lenp, | |
230 | loff_t *ppos) | |
231 | { | |
232 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
233 | ||
234 | if (ret || !write) | |
235 | return ret; | |
236 | ||
237 | update_perf_cpu_limits(); | |
163ec435 PZ |
238 | |
239 | return 0; | |
240 | } | |
1ccd1549 | 241 | |
14c63f17 DH |
242 | /* |
243 | * perf samples are done in some very critical code paths (NMIs). | |
244 | * If they take too much CPU time, the system can lock up and not | |
245 | * get any real work done. This will drop the sample rate when | |
246 | * we detect that events are taking too long. | |
247 | */ | |
248 | #define NR_ACCUMULATED_SAMPLES 128 | |
d9494cb4 | 249 | static DEFINE_PER_CPU(u64, running_sample_length); |
14c63f17 | 250 | |
6a02ad66 | 251 | static void perf_duration_warn(struct irq_work *w) |
14c63f17 | 252 | { |
6a02ad66 | 253 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
14c63f17 | 254 | u64 avg_local_sample_len; |
e5302920 | 255 | u64 local_samples_len; |
6a02ad66 | 256 | |
4a32fea9 | 257 | local_samples_len = __this_cpu_read(running_sample_length); |
6a02ad66 PZ |
258 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; |
259 | ||
260 | printk_ratelimited(KERN_WARNING | |
261 | "perf interrupt took too long (%lld > %lld), lowering " | |
262 | "kernel.perf_event_max_sample_rate to %d\n", | |
cd578abb | 263 | avg_local_sample_len, allowed_ns >> 1, |
6a02ad66 PZ |
264 | sysctl_perf_event_sample_rate); |
265 | } | |
266 | ||
267 | static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn); | |
268 | ||
269 | void perf_sample_event_took(u64 sample_len_ns) | |
270 | { | |
d9494cb4 | 271 | u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns); |
6a02ad66 PZ |
272 | u64 avg_local_sample_len; |
273 | u64 local_samples_len; | |
14c63f17 | 274 | |
d9494cb4 | 275 | if (allowed_ns == 0) |
14c63f17 DH |
276 | return; |
277 | ||
278 | /* decay the counter by 1 average sample */ | |
4a32fea9 | 279 | local_samples_len = __this_cpu_read(running_sample_length); |
14c63f17 DH |
280 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; |
281 | local_samples_len += sample_len_ns; | |
4a32fea9 | 282 | __this_cpu_write(running_sample_length, local_samples_len); |
14c63f17 DH |
283 | |
284 | /* | |
285 | * note: this will be biased artifically low until we have | |
286 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
287 | * from having to maintain a count. | |
288 | */ | |
289 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
290 | ||
d9494cb4 | 291 | if (avg_local_sample_len <= allowed_ns) |
14c63f17 DH |
292 | return; |
293 | ||
294 | if (max_samples_per_tick <= 1) | |
295 | return; | |
296 | ||
297 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
298 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
299 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
300 | ||
14c63f17 | 301 | update_perf_cpu_limits(); |
6a02ad66 | 302 | |
cd578abb PZ |
303 | if (!irq_work_queue(&perf_duration_work)) { |
304 | early_printk("perf interrupt took too long (%lld > %lld), lowering " | |
305 | "kernel.perf_event_max_sample_rate to %d\n", | |
306 | avg_local_sample_len, allowed_ns >> 1, | |
307 | sysctl_perf_event_sample_rate); | |
308 | } | |
14c63f17 DH |
309 | } |
310 | ||
cdd6c482 | 311 | static atomic64_t perf_event_id; |
a96bbc16 | 312 | |
0b3fcf17 SE |
313 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
314 | enum event_type_t event_type); | |
315 | ||
316 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
317 | enum event_type_t event_type, |
318 | struct task_struct *task); | |
319 | ||
320 | static void update_context_time(struct perf_event_context *ctx); | |
321 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 322 | |
cdd6c482 | 323 | void __weak perf_event_print_debug(void) { } |
0793a61d | 324 | |
84c79910 | 325 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 326 | { |
84c79910 | 327 | return "pmu"; |
0793a61d TG |
328 | } |
329 | ||
0b3fcf17 SE |
330 | static inline u64 perf_clock(void) |
331 | { | |
332 | return local_clock(); | |
333 | } | |
334 | ||
34f43927 PZ |
335 | static inline u64 perf_event_clock(struct perf_event *event) |
336 | { | |
337 | return event->clock(); | |
338 | } | |
339 | ||
e5d1367f SE |
340 | static inline struct perf_cpu_context * |
341 | __get_cpu_context(struct perf_event_context *ctx) | |
342 | { | |
343 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
344 | } | |
345 | ||
facc4307 PZ |
346 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
347 | struct perf_event_context *ctx) | |
348 | { | |
349 | raw_spin_lock(&cpuctx->ctx.lock); | |
350 | if (ctx) | |
351 | raw_spin_lock(&ctx->lock); | |
352 | } | |
353 | ||
354 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
355 | struct perf_event_context *ctx) | |
356 | { | |
357 | if (ctx) | |
358 | raw_spin_unlock(&ctx->lock); | |
359 | raw_spin_unlock(&cpuctx->ctx.lock); | |
360 | } | |
361 | ||
e5d1367f SE |
362 | #ifdef CONFIG_CGROUP_PERF |
363 | ||
e5d1367f SE |
364 | static inline bool |
365 | perf_cgroup_match(struct perf_event *event) | |
366 | { | |
367 | struct perf_event_context *ctx = event->ctx; | |
368 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
369 | ||
ef824fa1 TH |
370 | /* @event doesn't care about cgroup */ |
371 | if (!event->cgrp) | |
372 | return true; | |
373 | ||
374 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
375 | if (!cpuctx->cgrp) | |
376 | return false; | |
377 | ||
378 | /* | |
379 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
380 | * also enabled for all its descendant cgroups. If @cpuctx's | |
381 | * cgroup is a descendant of @event's (the test covers identity | |
382 | * case), it's a match. | |
383 | */ | |
384 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
385 | event->cgrp->css.cgroup); | |
e5d1367f SE |
386 | } |
387 | ||
e5d1367f SE |
388 | static inline void perf_detach_cgroup(struct perf_event *event) |
389 | { | |
4e2ba650 | 390 | css_put(&event->cgrp->css); |
e5d1367f SE |
391 | event->cgrp = NULL; |
392 | } | |
393 | ||
394 | static inline int is_cgroup_event(struct perf_event *event) | |
395 | { | |
396 | return event->cgrp != NULL; | |
397 | } | |
398 | ||
399 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
400 | { | |
401 | struct perf_cgroup_info *t; | |
402 | ||
403 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
404 | return t->time; | |
405 | } | |
406 | ||
407 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
408 | { | |
409 | struct perf_cgroup_info *info; | |
410 | u64 now; | |
411 | ||
412 | now = perf_clock(); | |
413 | ||
414 | info = this_cpu_ptr(cgrp->info); | |
415 | ||
416 | info->time += now - info->timestamp; | |
417 | info->timestamp = now; | |
418 | } | |
419 | ||
420 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
421 | { | |
422 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
423 | if (cgrp_out) | |
424 | __update_cgrp_time(cgrp_out); | |
425 | } | |
426 | ||
427 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
428 | { | |
3f7cce3c SE |
429 | struct perf_cgroup *cgrp; |
430 | ||
e5d1367f | 431 | /* |
3f7cce3c SE |
432 | * ensure we access cgroup data only when needed and |
433 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 434 | */ |
3f7cce3c | 435 | if (!is_cgroup_event(event)) |
e5d1367f SE |
436 | return; |
437 | ||
3f7cce3c SE |
438 | cgrp = perf_cgroup_from_task(current); |
439 | /* | |
440 | * Do not update time when cgroup is not active | |
441 | */ | |
442 | if (cgrp == event->cgrp) | |
443 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
444 | } |
445 | ||
446 | static inline void | |
3f7cce3c SE |
447 | perf_cgroup_set_timestamp(struct task_struct *task, |
448 | struct perf_event_context *ctx) | |
e5d1367f SE |
449 | { |
450 | struct perf_cgroup *cgrp; | |
451 | struct perf_cgroup_info *info; | |
452 | ||
3f7cce3c SE |
453 | /* |
454 | * ctx->lock held by caller | |
455 | * ensure we do not access cgroup data | |
456 | * unless we have the cgroup pinned (css_get) | |
457 | */ | |
458 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
459 | return; |
460 | ||
461 | cgrp = perf_cgroup_from_task(task); | |
462 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 463 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
464 | } |
465 | ||
466 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
467 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
468 | ||
469 | /* | |
470 | * reschedule events based on the cgroup constraint of task. | |
471 | * | |
472 | * mode SWOUT : schedule out everything | |
473 | * mode SWIN : schedule in based on cgroup for next | |
474 | */ | |
475 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
476 | { | |
477 | struct perf_cpu_context *cpuctx; | |
478 | struct pmu *pmu; | |
479 | unsigned long flags; | |
480 | ||
481 | /* | |
482 | * disable interrupts to avoid geting nr_cgroup | |
483 | * changes via __perf_event_disable(). Also | |
484 | * avoids preemption. | |
485 | */ | |
486 | local_irq_save(flags); | |
487 | ||
488 | /* | |
489 | * we reschedule only in the presence of cgroup | |
490 | * constrained events. | |
491 | */ | |
492 | rcu_read_lock(); | |
493 | ||
494 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 495 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
496 | if (cpuctx->unique_pmu != pmu) |
497 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 498 | |
e5d1367f SE |
499 | /* |
500 | * perf_cgroup_events says at least one | |
501 | * context on this CPU has cgroup events. | |
502 | * | |
503 | * ctx->nr_cgroups reports the number of cgroup | |
504 | * events for a context. | |
505 | */ | |
506 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
507 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
508 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
509 | |
510 | if (mode & PERF_CGROUP_SWOUT) { | |
511 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
512 | /* | |
513 | * must not be done before ctxswout due | |
514 | * to event_filter_match() in event_sched_out() | |
515 | */ | |
516 | cpuctx->cgrp = NULL; | |
517 | } | |
518 | ||
519 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 520 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
521 | /* |
522 | * set cgrp before ctxsw in to allow | |
523 | * event_filter_match() to not have to pass | |
524 | * task around | |
e5d1367f SE |
525 | */ |
526 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
527 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
528 | } | |
facc4307 PZ |
529 | perf_pmu_enable(cpuctx->ctx.pmu); |
530 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 531 | } |
e5d1367f SE |
532 | } |
533 | ||
534 | rcu_read_unlock(); | |
535 | ||
536 | local_irq_restore(flags); | |
537 | } | |
538 | ||
a8d757ef SE |
539 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
540 | struct task_struct *next) | |
e5d1367f | 541 | { |
a8d757ef SE |
542 | struct perf_cgroup *cgrp1; |
543 | struct perf_cgroup *cgrp2 = NULL; | |
544 | ||
545 | /* | |
546 | * we come here when we know perf_cgroup_events > 0 | |
547 | */ | |
548 | cgrp1 = perf_cgroup_from_task(task); | |
549 | ||
550 | /* | |
551 | * next is NULL when called from perf_event_enable_on_exec() | |
552 | * that will systematically cause a cgroup_switch() | |
553 | */ | |
554 | if (next) | |
555 | cgrp2 = perf_cgroup_from_task(next); | |
556 | ||
557 | /* | |
558 | * only schedule out current cgroup events if we know | |
559 | * that we are switching to a different cgroup. Otherwise, | |
560 | * do no touch the cgroup events. | |
561 | */ | |
562 | if (cgrp1 != cgrp2) | |
563 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
564 | } |
565 | ||
a8d757ef SE |
566 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
567 | struct task_struct *task) | |
e5d1367f | 568 | { |
a8d757ef SE |
569 | struct perf_cgroup *cgrp1; |
570 | struct perf_cgroup *cgrp2 = NULL; | |
571 | ||
572 | /* | |
573 | * we come here when we know perf_cgroup_events > 0 | |
574 | */ | |
575 | cgrp1 = perf_cgroup_from_task(task); | |
576 | ||
577 | /* prev can never be NULL */ | |
578 | cgrp2 = perf_cgroup_from_task(prev); | |
579 | ||
580 | /* | |
581 | * only need to schedule in cgroup events if we are changing | |
582 | * cgroup during ctxsw. Cgroup events were not scheduled | |
583 | * out of ctxsw out if that was not the case. | |
584 | */ | |
585 | if (cgrp1 != cgrp2) | |
586 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
587 | } |
588 | ||
589 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
590 | struct perf_event_attr *attr, | |
591 | struct perf_event *group_leader) | |
592 | { | |
593 | struct perf_cgroup *cgrp; | |
594 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
595 | struct fd f = fdget(fd); |
596 | int ret = 0; | |
e5d1367f | 597 | |
2903ff01 | 598 | if (!f.file) |
e5d1367f SE |
599 | return -EBADF; |
600 | ||
b583043e | 601 | css = css_tryget_online_from_dir(f.file->f_path.dentry, |
ec903c0c | 602 | &perf_event_cgrp_subsys); |
3db272c0 LZ |
603 | if (IS_ERR(css)) { |
604 | ret = PTR_ERR(css); | |
605 | goto out; | |
606 | } | |
e5d1367f SE |
607 | |
608 | cgrp = container_of(css, struct perf_cgroup, css); | |
609 | event->cgrp = cgrp; | |
610 | ||
611 | /* | |
612 | * all events in a group must monitor | |
613 | * the same cgroup because a task belongs | |
614 | * to only one perf cgroup at a time | |
615 | */ | |
616 | if (group_leader && group_leader->cgrp != cgrp) { | |
617 | perf_detach_cgroup(event); | |
618 | ret = -EINVAL; | |
e5d1367f | 619 | } |
3db272c0 | 620 | out: |
2903ff01 | 621 | fdput(f); |
e5d1367f SE |
622 | return ret; |
623 | } | |
624 | ||
625 | static inline void | |
626 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
627 | { | |
628 | struct perf_cgroup_info *t; | |
629 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
630 | event->shadow_ctx_time = now - t->timestamp; | |
631 | } | |
632 | ||
633 | static inline void | |
634 | perf_cgroup_defer_enabled(struct perf_event *event) | |
635 | { | |
636 | /* | |
637 | * when the current task's perf cgroup does not match | |
638 | * the event's, we need to remember to call the | |
639 | * perf_mark_enable() function the first time a task with | |
640 | * a matching perf cgroup is scheduled in. | |
641 | */ | |
642 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
643 | event->cgrp_defer_enabled = 1; | |
644 | } | |
645 | ||
646 | static inline void | |
647 | perf_cgroup_mark_enabled(struct perf_event *event, | |
648 | struct perf_event_context *ctx) | |
649 | { | |
650 | struct perf_event *sub; | |
651 | u64 tstamp = perf_event_time(event); | |
652 | ||
653 | if (!event->cgrp_defer_enabled) | |
654 | return; | |
655 | ||
656 | event->cgrp_defer_enabled = 0; | |
657 | ||
658 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
659 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
660 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
661 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
662 | sub->cgrp_defer_enabled = 0; | |
663 | } | |
664 | } | |
665 | } | |
666 | #else /* !CONFIG_CGROUP_PERF */ | |
667 | ||
668 | static inline bool | |
669 | perf_cgroup_match(struct perf_event *event) | |
670 | { | |
671 | return true; | |
672 | } | |
673 | ||
674 | static inline void perf_detach_cgroup(struct perf_event *event) | |
675 | {} | |
676 | ||
677 | static inline int is_cgroup_event(struct perf_event *event) | |
678 | { | |
679 | return 0; | |
680 | } | |
681 | ||
682 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
683 | { | |
684 | return 0; | |
685 | } | |
686 | ||
687 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
688 | { | |
689 | } | |
690 | ||
691 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
692 | { | |
693 | } | |
694 | ||
a8d757ef SE |
695 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
696 | struct task_struct *next) | |
e5d1367f SE |
697 | { |
698 | } | |
699 | ||
a8d757ef SE |
700 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
701 | struct task_struct *task) | |
e5d1367f SE |
702 | { |
703 | } | |
704 | ||
705 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
706 | struct perf_event_attr *attr, | |
707 | struct perf_event *group_leader) | |
708 | { | |
709 | return -EINVAL; | |
710 | } | |
711 | ||
712 | static inline void | |
3f7cce3c SE |
713 | perf_cgroup_set_timestamp(struct task_struct *task, |
714 | struct perf_event_context *ctx) | |
e5d1367f SE |
715 | { |
716 | } | |
717 | ||
718 | void | |
719 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
720 | { | |
721 | } | |
722 | ||
723 | static inline void | |
724 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
725 | { | |
726 | } | |
727 | ||
728 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
729 | { | |
730 | return 0; | |
731 | } | |
732 | ||
733 | static inline void | |
734 | perf_cgroup_defer_enabled(struct perf_event *event) | |
735 | { | |
736 | } | |
737 | ||
738 | static inline void | |
739 | perf_cgroup_mark_enabled(struct perf_event *event, | |
740 | struct perf_event_context *ctx) | |
741 | { | |
742 | } | |
743 | #endif | |
744 | ||
9e630205 SE |
745 | /* |
746 | * set default to be dependent on timer tick just | |
747 | * like original code | |
748 | */ | |
749 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
750 | /* | |
751 | * function must be called with interrupts disbled | |
752 | */ | |
272325c4 | 753 | static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr) |
9e630205 SE |
754 | { |
755 | struct perf_cpu_context *cpuctx; | |
9e630205 SE |
756 | int rotations = 0; |
757 | ||
758 | WARN_ON(!irqs_disabled()); | |
759 | ||
760 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
9e630205 SE |
761 | rotations = perf_rotate_context(cpuctx); |
762 | ||
4cfafd30 PZ |
763 | raw_spin_lock(&cpuctx->hrtimer_lock); |
764 | if (rotations) | |
9e630205 | 765 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); |
4cfafd30 PZ |
766 | else |
767 | cpuctx->hrtimer_active = 0; | |
768 | raw_spin_unlock(&cpuctx->hrtimer_lock); | |
9e630205 | 769 | |
4cfafd30 | 770 | return rotations ? HRTIMER_RESTART : HRTIMER_NORESTART; |
9e630205 SE |
771 | } |
772 | ||
272325c4 | 773 | static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) |
9e630205 | 774 | { |
272325c4 | 775 | struct hrtimer *timer = &cpuctx->hrtimer; |
9e630205 | 776 | struct pmu *pmu = cpuctx->ctx.pmu; |
272325c4 | 777 | u64 interval; |
9e630205 SE |
778 | |
779 | /* no multiplexing needed for SW PMU */ | |
780 | if (pmu->task_ctx_nr == perf_sw_context) | |
781 | return; | |
782 | ||
62b85639 SE |
783 | /* |
784 | * check default is sane, if not set then force to | |
785 | * default interval (1/tick) | |
786 | */ | |
272325c4 PZ |
787 | interval = pmu->hrtimer_interval_ms; |
788 | if (interval < 1) | |
789 | interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
62b85639 | 790 | |
272325c4 | 791 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval); |
9e630205 | 792 | |
4cfafd30 PZ |
793 | raw_spin_lock_init(&cpuctx->hrtimer_lock); |
794 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); | |
272325c4 | 795 | timer->function = perf_mux_hrtimer_handler; |
9e630205 SE |
796 | } |
797 | ||
272325c4 | 798 | static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx) |
9e630205 | 799 | { |
272325c4 | 800 | struct hrtimer *timer = &cpuctx->hrtimer; |
9e630205 | 801 | struct pmu *pmu = cpuctx->ctx.pmu; |
4cfafd30 | 802 | unsigned long flags; |
9e630205 SE |
803 | |
804 | /* not for SW PMU */ | |
805 | if (pmu->task_ctx_nr == perf_sw_context) | |
272325c4 | 806 | return 0; |
9e630205 | 807 | |
4cfafd30 PZ |
808 | raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags); |
809 | if (!cpuctx->hrtimer_active) { | |
810 | cpuctx->hrtimer_active = 1; | |
811 | hrtimer_forward_now(timer, cpuctx->hrtimer_interval); | |
812 | hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED); | |
813 | } | |
814 | raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags); | |
9e630205 | 815 | |
272325c4 | 816 | return 0; |
9e630205 SE |
817 | } |
818 | ||
33696fc0 | 819 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 820 | { |
33696fc0 PZ |
821 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
822 | if (!(*count)++) | |
823 | pmu->pmu_disable(pmu); | |
9e35ad38 | 824 | } |
9e35ad38 | 825 | |
33696fc0 | 826 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 827 | { |
33696fc0 PZ |
828 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
829 | if (!--(*count)) | |
830 | pmu->pmu_enable(pmu); | |
9e35ad38 | 831 | } |
9e35ad38 | 832 | |
2fde4f94 | 833 | static DEFINE_PER_CPU(struct list_head, active_ctx_list); |
e9d2b064 PZ |
834 | |
835 | /* | |
2fde4f94 MR |
836 | * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and |
837 | * perf_event_task_tick() are fully serialized because they're strictly cpu | |
838 | * affine and perf_event_ctx{activate,deactivate} are called with IRQs | |
839 | * disabled, while perf_event_task_tick is called from IRQ context. | |
e9d2b064 | 840 | */ |
2fde4f94 | 841 | static void perf_event_ctx_activate(struct perf_event_context *ctx) |
9e35ad38 | 842 | { |
2fde4f94 | 843 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
b5ab4cd5 | 844 | |
e9d2b064 | 845 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 846 | |
2fde4f94 MR |
847 | WARN_ON(!list_empty(&ctx->active_ctx_list)); |
848 | ||
849 | list_add(&ctx->active_ctx_list, head); | |
850 | } | |
851 | ||
852 | static void perf_event_ctx_deactivate(struct perf_event_context *ctx) | |
853 | { | |
854 | WARN_ON(!irqs_disabled()); | |
855 | ||
856 | WARN_ON(list_empty(&ctx->active_ctx_list)); | |
857 | ||
858 | list_del_init(&ctx->active_ctx_list); | |
9e35ad38 | 859 | } |
9e35ad38 | 860 | |
cdd6c482 | 861 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 862 | { |
e5289d4a | 863 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
864 | } |
865 | ||
4af57ef2 YZ |
866 | static void free_ctx(struct rcu_head *head) |
867 | { | |
868 | struct perf_event_context *ctx; | |
869 | ||
870 | ctx = container_of(head, struct perf_event_context, rcu_head); | |
871 | kfree(ctx->task_ctx_data); | |
872 | kfree(ctx); | |
873 | } | |
874 | ||
cdd6c482 | 875 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 876 | { |
564c2b21 PM |
877 | if (atomic_dec_and_test(&ctx->refcount)) { |
878 | if (ctx->parent_ctx) | |
879 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
880 | if (ctx->task) |
881 | put_task_struct(ctx->task); | |
4af57ef2 | 882 | call_rcu(&ctx->rcu_head, free_ctx); |
564c2b21 | 883 | } |
a63eaf34 PM |
884 | } |
885 | ||
f63a8daa PZ |
886 | /* |
887 | * Because of perf_event::ctx migration in sys_perf_event_open::move_group and | |
888 | * perf_pmu_migrate_context() we need some magic. | |
889 | * | |
890 | * Those places that change perf_event::ctx will hold both | |
891 | * perf_event_ctx::mutex of the 'old' and 'new' ctx value. | |
892 | * | |
8b10c5e2 PZ |
893 | * Lock ordering is by mutex address. There are two other sites where |
894 | * perf_event_context::mutex nests and those are: | |
895 | * | |
896 | * - perf_event_exit_task_context() [ child , 0 ] | |
897 | * __perf_event_exit_task() | |
898 | * sync_child_event() | |
899 | * put_event() [ parent, 1 ] | |
900 | * | |
901 | * - perf_event_init_context() [ parent, 0 ] | |
902 | * inherit_task_group() | |
903 | * inherit_group() | |
904 | * inherit_event() | |
905 | * perf_event_alloc() | |
906 | * perf_init_event() | |
907 | * perf_try_init_event() [ child , 1 ] | |
908 | * | |
909 | * While it appears there is an obvious deadlock here -- the parent and child | |
910 | * nesting levels are inverted between the two. This is in fact safe because | |
911 | * life-time rules separate them. That is an exiting task cannot fork, and a | |
912 | * spawning task cannot (yet) exit. | |
913 | * | |
914 | * But remember that that these are parent<->child context relations, and | |
915 | * migration does not affect children, therefore these two orderings should not | |
916 | * interact. | |
f63a8daa PZ |
917 | * |
918 | * The change in perf_event::ctx does not affect children (as claimed above) | |
919 | * because the sys_perf_event_open() case will install a new event and break | |
920 | * the ctx parent<->child relation, and perf_pmu_migrate_context() is only | |
921 | * concerned with cpuctx and that doesn't have children. | |
922 | * | |
923 | * The places that change perf_event::ctx will issue: | |
924 | * | |
925 | * perf_remove_from_context(); | |
926 | * synchronize_rcu(); | |
927 | * perf_install_in_context(); | |
928 | * | |
929 | * to affect the change. The remove_from_context() + synchronize_rcu() should | |
930 | * quiesce the event, after which we can install it in the new location. This | |
931 | * means that only external vectors (perf_fops, prctl) can perturb the event | |
932 | * while in transit. Therefore all such accessors should also acquire | |
933 | * perf_event_context::mutex to serialize against this. | |
934 | * | |
935 | * However; because event->ctx can change while we're waiting to acquire | |
936 | * ctx->mutex we must be careful and use the below perf_event_ctx_lock() | |
937 | * function. | |
938 | * | |
939 | * Lock order: | |
940 | * task_struct::perf_event_mutex | |
941 | * perf_event_context::mutex | |
942 | * perf_event_context::lock | |
943 | * perf_event::child_mutex; | |
944 | * perf_event::mmap_mutex | |
945 | * mmap_sem | |
946 | */ | |
a83fe28e PZ |
947 | static struct perf_event_context * |
948 | perf_event_ctx_lock_nested(struct perf_event *event, int nesting) | |
f63a8daa PZ |
949 | { |
950 | struct perf_event_context *ctx; | |
951 | ||
952 | again: | |
953 | rcu_read_lock(); | |
954 | ctx = ACCESS_ONCE(event->ctx); | |
955 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
956 | rcu_read_unlock(); | |
957 | goto again; | |
958 | } | |
959 | rcu_read_unlock(); | |
960 | ||
a83fe28e | 961 | mutex_lock_nested(&ctx->mutex, nesting); |
f63a8daa PZ |
962 | if (event->ctx != ctx) { |
963 | mutex_unlock(&ctx->mutex); | |
964 | put_ctx(ctx); | |
965 | goto again; | |
966 | } | |
967 | ||
968 | return ctx; | |
969 | } | |
970 | ||
a83fe28e PZ |
971 | static inline struct perf_event_context * |
972 | perf_event_ctx_lock(struct perf_event *event) | |
973 | { | |
974 | return perf_event_ctx_lock_nested(event, 0); | |
975 | } | |
976 | ||
f63a8daa PZ |
977 | static void perf_event_ctx_unlock(struct perf_event *event, |
978 | struct perf_event_context *ctx) | |
979 | { | |
980 | mutex_unlock(&ctx->mutex); | |
981 | put_ctx(ctx); | |
982 | } | |
983 | ||
211de6eb PZ |
984 | /* |
985 | * This must be done under the ctx->lock, such as to serialize against | |
986 | * context_equiv(), therefore we cannot call put_ctx() since that might end up | |
987 | * calling scheduler related locks and ctx->lock nests inside those. | |
988 | */ | |
989 | static __must_check struct perf_event_context * | |
990 | unclone_ctx(struct perf_event_context *ctx) | |
71a851b4 | 991 | { |
211de6eb PZ |
992 | struct perf_event_context *parent_ctx = ctx->parent_ctx; |
993 | ||
994 | lockdep_assert_held(&ctx->lock); | |
995 | ||
996 | if (parent_ctx) | |
71a851b4 | 997 | ctx->parent_ctx = NULL; |
5a3126d4 | 998 | ctx->generation++; |
211de6eb PZ |
999 | |
1000 | return parent_ctx; | |
71a851b4 PZ |
1001 | } |
1002 | ||
6844c09d ACM |
1003 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
1004 | { | |
1005 | /* | |
1006 | * only top level events have the pid namespace they were created in | |
1007 | */ | |
1008 | if (event->parent) | |
1009 | event = event->parent; | |
1010 | ||
1011 | return task_tgid_nr_ns(p, event->ns); | |
1012 | } | |
1013 | ||
1014 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
1015 | { | |
1016 | /* | |
1017 | * only top level events have the pid namespace they were created in | |
1018 | */ | |
1019 | if (event->parent) | |
1020 | event = event->parent; | |
1021 | ||
1022 | return task_pid_nr_ns(p, event->ns); | |
1023 | } | |
1024 | ||
7f453c24 | 1025 | /* |
cdd6c482 | 1026 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
1027 | * to userspace. |
1028 | */ | |
cdd6c482 | 1029 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 1030 | { |
cdd6c482 | 1031 | u64 id = event->id; |
7f453c24 | 1032 | |
cdd6c482 IM |
1033 | if (event->parent) |
1034 | id = event->parent->id; | |
7f453c24 PZ |
1035 | |
1036 | return id; | |
1037 | } | |
1038 | ||
25346b93 | 1039 | /* |
cdd6c482 | 1040 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
1041 | * This has to cope with with the fact that until it is locked, |
1042 | * the context could get moved to another task. | |
1043 | */ | |
cdd6c482 | 1044 | static struct perf_event_context * |
8dc85d54 | 1045 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 1046 | { |
cdd6c482 | 1047 | struct perf_event_context *ctx; |
25346b93 | 1048 | |
9ed6060d | 1049 | retry: |
058ebd0e PZ |
1050 | /* |
1051 | * One of the few rules of preemptible RCU is that one cannot do | |
1052 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
1053 | * part of the read side critical section was preemptible -- see | |
1054 | * rcu_read_unlock_special(). | |
1055 | * | |
1056 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
1057 | * side critical section is non-preemptible. | |
1058 | */ | |
1059 | preempt_disable(); | |
1060 | rcu_read_lock(); | |
8dc85d54 | 1061 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
1062 | if (ctx) { |
1063 | /* | |
1064 | * If this context is a clone of another, it might | |
1065 | * get swapped for another underneath us by | |
cdd6c482 | 1066 | * perf_event_task_sched_out, though the |
25346b93 PM |
1067 | * rcu_read_lock() protects us from any context |
1068 | * getting freed. Lock the context and check if it | |
1069 | * got swapped before we could get the lock, and retry | |
1070 | * if so. If we locked the right context, then it | |
1071 | * can't get swapped on us any more. | |
1072 | */ | |
e625cce1 | 1073 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 1074 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 1075 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
1076 | rcu_read_unlock(); |
1077 | preempt_enable(); | |
25346b93 PM |
1078 | goto retry; |
1079 | } | |
b49a9e7e PZ |
1080 | |
1081 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 1082 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
1083 | ctx = NULL; |
1084 | } | |
25346b93 PM |
1085 | } |
1086 | rcu_read_unlock(); | |
058ebd0e | 1087 | preempt_enable(); |
25346b93 PM |
1088 | return ctx; |
1089 | } | |
1090 | ||
1091 | /* | |
1092 | * Get the context for a task and increment its pin_count so it | |
1093 | * can't get swapped to another task. This also increments its | |
1094 | * reference count so that the context can't get freed. | |
1095 | */ | |
8dc85d54 PZ |
1096 | static struct perf_event_context * |
1097 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 1098 | { |
cdd6c482 | 1099 | struct perf_event_context *ctx; |
25346b93 PM |
1100 | unsigned long flags; |
1101 | ||
8dc85d54 | 1102 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1103 | if (ctx) { |
1104 | ++ctx->pin_count; | |
e625cce1 | 1105 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1106 | } |
1107 | return ctx; | |
1108 | } | |
1109 | ||
cdd6c482 | 1110 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1111 | { |
1112 | unsigned long flags; | |
1113 | ||
e625cce1 | 1114 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1115 | --ctx->pin_count; |
e625cce1 | 1116 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1117 | } |
1118 | ||
f67218c3 PZ |
1119 | /* |
1120 | * Update the record of the current time in a context. | |
1121 | */ | |
1122 | static void update_context_time(struct perf_event_context *ctx) | |
1123 | { | |
1124 | u64 now = perf_clock(); | |
1125 | ||
1126 | ctx->time += now - ctx->timestamp; | |
1127 | ctx->timestamp = now; | |
1128 | } | |
1129 | ||
4158755d SE |
1130 | static u64 perf_event_time(struct perf_event *event) |
1131 | { | |
1132 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1133 | |
1134 | if (is_cgroup_event(event)) | |
1135 | return perf_cgroup_event_time(event); | |
1136 | ||
4158755d SE |
1137 | return ctx ? ctx->time : 0; |
1138 | } | |
1139 | ||
f67218c3 PZ |
1140 | /* |
1141 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1142 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1143 | */ |
1144 | static void update_event_times(struct perf_event *event) | |
1145 | { | |
1146 | struct perf_event_context *ctx = event->ctx; | |
1147 | u64 run_end; | |
1148 | ||
1149 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1150 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1151 | return; | |
e5d1367f SE |
1152 | /* |
1153 | * in cgroup mode, time_enabled represents | |
1154 | * the time the event was enabled AND active | |
1155 | * tasks were in the monitored cgroup. This is | |
1156 | * independent of the activity of the context as | |
1157 | * there may be a mix of cgroup and non-cgroup events. | |
1158 | * | |
1159 | * That is why we treat cgroup events differently | |
1160 | * here. | |
1161 | */ | |
1162 | if (is_cgroup_event(event)) | |
46cd6a7f | 1163 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1164 | else if (ctx->is_active) |
1165 | run_end = ctx->time; | |
acd1d7c1 PZ |
1166 | else |
1167 | run_end = event->tstamp_stopped; | |
1168 | ||
1169 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1170 | |
1171 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1172 | run_end = event->tstamp_stopped; | |
1173 | else | |
4158755d | 1174 | run_end = perf_event_time(event); |
f67218c3 PZ |
1175 | |
1176 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1177 | |
f67218c3 PZ |
1178 | } |
1179 | ||
96c21a46 PZ |
1180 | /* |
1181 | * Update total_time_enabled and total_time_running for all events in a group. | |
1182 | */ | |
1183 | static void update_group_times(struct perf_event *leader) | |
1184 | { | |
1185 | struct perf_event *event; | |
1186 | ||
1187 | update_event_times(leader); | |
1188 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1189 | update_event_times(event); | |
1190 | } | |
1191 | ||
889ff015 FW |
1192 | static struct list_head * |
1193 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1194 | { | |
1195 | if (event->attr.pinned) | |
1196 | return &ctx->pinned_groups; | |
1197 | else | |
1198 | return &ctx->flexible_groups; | |
1199 | } | |
1200 | ||
fccc714b | 1201 | /* |
cdd6c482 | 1202 | * Add a event from the lists for its context. |
fccc714b PZ |
1203 | * Must be called with ctx->mutex and ctx->lock held. |
1204 | */ | |
04289bb9 | 1205 | static void |
cdd6c482 | 1206 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1207 | { |
8a49542c PZ |
1208 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1209 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1210 | |
1211 | /* | |
8a49542c PZ |
1212 | * If we're a stand alone event or group leader, we go to the context |
1213 | * list, group events are kept attached to the group so that | |
1214 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1215 | */ |
8a49542c | 1216 | if (event->group_leader == event) { |
889ff015 FW |
1217 | struct list_head *list; |
1218 | ||
d6f962b5 FW |
1219 | if (is_software_event(event)) |
1220 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1221 | ||
889ff015 FW |
1222 | list = ctx_group_list(event, ctx); |
1223 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1224 | } |
592903cd | 1225 | |
08309379 | 1226 | if (is_cgroup_event(event)) |
e5d1367f | 1227 | ctx->nr_cgroups++; |
e5d1367f | 1228 | |
cdd6c482 IM |
1229 | list_add_rcu(&event->event_entry, &ctx->event_list); |
1230 | ctx->nr_events++; | |
1231 | if (event->attr.inherit_stat) | |
bfbd3381 | 1232 | ctx->nr_stat++; |
5a3126d4 PZ |
1233 | |
1234 | ctx->generation++; | |
04289bb9 IM |
1235 | } |
1236 | ||
0231bb53 JO |
1237 | /* |
1238 | * Initialize event state based on the perf_event_attr::disabled. | |
1239 | */ | |
1240 | static inline void perf_event__state_init(struct perf_event *event) | |
1241 | { | |
1242 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1243 | PERF_EVENT_STATE_INACTIVE; | |
1244 | } | |
1245 | ||
c320c7b7 ACM |
1246 | /* |
1247 | * Called at perf_event creation and when events are attached/detached from a | |
1248 | * group. | |
1249 | */ | |
1250 | static void perf_event__read_size(struct perf_event *event) | |
1251 | { | |
1252 | int entry = sizeof(u64); /* value */ | |
1253 | int size = 0; | |
1254 | int nr = 1; | |
1255 | ||
1256 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1257 | size += sizeof(u64); | |
1258 | ||
1259 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1260 | size += sizeof(u64); | |
1261 | ||
1262 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1263 | entry += sizeof(u64); | |
1264 | ||
1265 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1266 | nr += event->group_leader->nr_siblings; | |
1267 | size += sizeof(u64); | |
1268 | } | |
1269 | ||
1270 | size += entry * nr; | |
1271 | event->read_size = size; | |
1272 | } | |
1273 | ||
1274 | static void perf_event__header_size(struct perf_event *event) | |
1275 | { | |
1276 | struct perf_sample_data *data; | |
1277 | u64 sample_type = event->attr.sample_type; | |
1278 | u16 size = 0; | |
1279 | ||
1280 | perf_event__read_size(event); | |
1281 | ||
1282 | if (sample_type & PERF_SAMPLE_IP) | |
1283 | size += sizeof(data->ip); | |
1284 | ||
6844c09d ACM |
1285 | if (sample_type & PERF_SAMPLE_ADDR) |
1286 | size += sizeof(data->addr); | |
1287 | ||
1288 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1289 | size += sizeof(data->period); | |
1290 | ||
c3feedf2 AK |
1291 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1292 | size += sizeof(data->weight); | |
1293 | ||
6844c09d ACM |
1294 | if (sample_type & PERF_SAMPLE_READ) |
1295 | size += event->read_size; | |
1296 | ||
d6be9ad6 SE |
1297 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1298 | size += sizeof(data->data_src.val); | |
1299 | ||
fdfbbd07 AK |
1300 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
1301 | size += sizeof(data->txn); | |
1302 | ||
6844c09d ACM |
1303 | event->header_size = size; |
1304 | } | |
1305 | ||
1306 | static void perf_event__id_header_size(struct perf_event *event) | |
1307 | { | |
1308 | struct perf_sample_data *data; | |
1309 | u64 sample_type = event->attr.sample_type; | |
1310 | u16 size = 0; | |
1311 | ||
c320c7b7 ACM |
1312 | if (sample_type & PERF_SAMPLE_TID) |
1313 | size += sizeof(data->tid_entry); | |
1314 | ||
1315 | if (sample_type & PERF_SAMPLE_TIME) | |
1316 | size += sizeof(data->time); | |
1317 | ||
ff3d527c AH |
1318 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1319 | size += sizeof(data->id); | |
1320 | ||
c320c7b7 ACM |
1321 | if (sample_type & PERF_SAMPLE_ID) |
1322 | size += sizeof(data->id); | |
1323 | ||
1324 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1325 | size += sizeof(data->stream_id); | |
1326 | ||
1327 | if (sample_type & PERF_SAMPLE_CPU) | |
1328 | size += sizeof(data->cpu_entry); | |
1329 | ||
6844c09d | 1330 | event->id_header_size = size; |
c320c7b7 ACM |
1331 | } |
1332 | ||
8a49542c PZ |
1333 | static void perf_group_attach(struct perf_event *event) |
1334 | { | |
c320c7b7 | 1335 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1336 | |
74c3337c PZ |
1337 | /* |
1338 | * We can have double attach due to group movement in perf_event_open. | |
1339 | */ | |
1340 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1341 | return; | |
1342 | ||
8a49542c PZ |
1343 | event->attach_state |= PERF_ATTACH_GROUP; |
1344 | ||
1345 | if (group_leader == event) | |
1346 | return; | |
1347 | ||
652884fe PZ |
1348 | WARN_ON_ONCE(group_leader->ctx != event->ctx); |
1349 | ||
8a49542c PZ |
1350 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && |
1351 | !is_software_event(event)) | |
1352 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1353 | ||
1354 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1355 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1356 | |
1357 | perf_event__header_size(group_leader); | |
1358 | ||
1359 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1360 | perf_event__header_size(pos); | |
8a49542c PZ |
1361 | } |
1362 | ||
a63eaf34 | 1363 | /* |
cdd6c482 | 1364 | * Remove a event from the lists for its context. |
fccc714b | 1365 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1366 | */ |
04289bb9 | 1367 | static void |
cdd6c482 | 1368 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1369 | { |
68cacd29 | 1370 | struct perf_cpu_context *cpuctx; |
652884fe PZ |
1371 | |
1372 | WARN_ON_ONCE(event->ctx != ctx); | |
1373 | lockdep_assert_held(&ctx->lock); | |
1374 | ||
8a49542c PZ |
1375 | /* |
1376 | * We can have double detach due to exit/hot-unplug + close. | |
1377 | */ | |
1378 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1379 | return; |
8a49542c PZ |
1380 | |
1381 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1382 | ||
68cacd29 | 1383 | if (is_cgroup_event(event)) { |
e5d1367f | 1384 | ctx->nr_cgroups--; |
68cacd29 SE |
1385 | cpuctx = __get_cpu_context(ctx); |
1386 | /* | |
1387 | * if there are no more cgroup events | |
1388 | * then cler cgrp to avoid stale pointer | |
1389 | * in update_cgrp_time_from_cpuctx() | |
1390 | */ | |
1391 | if (!ctx->nr_cgroups) | |
1392 | cpuctx->cgrp = NULL; | |
1393 | } | |
e5d1367f | 1394 | |
cdd6c482 IM |
1395 | ctx->nr_events--; |
1396 | if (event->attr.inherit_stat) | |
bfbd3381 | 1397 | ctx->nr_stat--; |
8bc20959 | 1398 | |
cdd6c482 | 1399 | list_del_rcu(&event->event_entry); |
04289bb9 | 1400 | |
8a49542c PZ |
1401 | if (event->group_leader == event) |
1402 | list_del_init(&event->group_entry); | |
5c148194 | 1403 | |
96c21a46 | 1404 | update_group_times(event); |
b2e74a26 SE |
1405 | |
1406 | /* | |
1407 | * If event was in error state, then keep it | |
1408 | * that way, otherwise bogus counts will be | |
1409 | * returned on read(). The only way to get out | |
1410 | * of error state is by explicit re-enabling | |
1411 | * of the event | |
1412 | */ | |
1413 | if (event->state > PERF_EVENT_STATE_OFF) | |
1414 | event->state = PERF_EVENT_STATE_OFF; | |
5a3126d4 PZ |
1415 | |
1416 | ctx->generation++; | |
050735b0 PZ |
1417 | } |
1418 | ||
8a49542c | 1419 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1420 | { |
1421 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1422 | struct list_head *list = NULL; |
1423 | ||
1424 | /* | |
1425 | * We can have double detach due to exit/hot-unplug + close. | |
1426 | */ | |
1427 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1428 | return; | |
1429 | ||
1430 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1431 | ||
1432 | /* | |
1433 | * If this is a sibling, remove it from its group. | |
1434 | */ | |
1435 | if (event->group_leader != event) { | |
1436 | list_del_init(&event->group_entry); | |
1437 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1438 | goto out; |
8a49542c PZ |
1439 | } |
1440 | ||
1441 | if (!list_empty(&event->group_entry)) | |
1442 | list = &event->group_entry; | |
2e2af50b | 1443 | |
04289bb9 | 1444 | /* |
cdd6c482 IM |
1445 | * If this was a group event with sibling events then |
1446 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1447 | * to whatever list we are on. |
04289bb9 | 1448 | */ |
cdd6c482 | 1449 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1450 | if (list) |
1451 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1452 | sibling->group_leader = sibling; |
d6f962b5 FW |
1453 | |
1454 | /* Inherit group flags from the previous leader */ | |
1455 | sibling->group_flags = event->group_flags; | |
652884fe PZ |
1456 | |
1457 | WARN_ON_ONCE(sibling->ctx != event->ctx); | |
04289bb9 | 1458 | } |
c320c7b7 ACM |
1459 | |
1460 | out: | |
1461 | perf_event__header_size(event->group_leader); | |
1462 | ||
1463 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1464 | perf_event__header_size(tmp); | |
04289bb9 IM |
1465 | } |
1466 | ||
fadfe7be JO |
1467 | /* |
1468 | * User event without the task. | |
1469 | */ | |
1470 | static bool is_orphaned_event(struct perf_event *event) | |
1471 | { | |
1472 | return event && !is_kernel_event(event) && !event->owner; | |
1473 | } | |
1474 | ||
1475 | /* | |
1476 | * Event has a parent but parent's task finished and it's | |
1477 | * alive only because of children holding refference. | |
1478 | */ | |
1479 | static bool is_orphaned_child(struct perf_event *event) | |
1480 | { | |
1481 | return is_orphaned_event(event->parent); | |
1482 | } | |
1483 | ||
1484 | static void orphans_remove_work(struct work_struct *work); | |
1485 | ||
1486 | static void schedule_orphans_remove(struct perf_event_context *ctx) | |
1487 | { | |
1488 | if (!ctx->task || ctx->orphans_remove_sched || !perf_wq) | |
1489 | return; | |
1490 | ||
1491 | if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) { | |
1492 | get_ctx(ctx); | |
1493 | ctx->orphans_remove_sched = true; | |
1494 | } | |
1495 | } | |
1496 | ||
1497 | static int __init perf_workqueue_init(void) | |
1498 | { | |
1499 | perf_wq = create_singlethread_workqueue("perf"); | |
1500 | WARN(!perf_wq, "failed to create perf workqueue\n"); | |
1501 | return perf_wq ? 0 : -1; | |
1502 | } | |
1503 | ||
1504 | core_initcall(perf_workqueue_init); | |
1505 | ||
66eb579e MR |
1506 | static inline int pmu_filter_match(struct perf_event *event) |
1507 | { | |
1508 | struct pmu *pmu = event->pmu; | |
1509 | return pmu->filter_match ? pmu->filter_match(event) : 1; | |
1510 | } | |
1511 | ||
fa66f07a SE |
1512 | static inline int |
1513 | event_filter_match(struct perf_event *event) | |
1514 | { | |
e5d1367f | 1515 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
66eb579e | 1516 | && perf_cgroup_match(event) && pmu_filter_match(event); |
fa66f07a SE |
1517 | } |
1518 | ||
9ffcfa6f SE |
1519 | static void |
1520 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1521 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1522 | struct perf_event_context *ctx) |
3b6f9e5c | 1523 | { |
4158755d | 1524 | u64 tstamp = perf_event_time(event); |
fa66f07a | 1525 | u64 delta; |
652884fe PZ |
1526 | |
1527 | WARN_ON_ONCE(event->ctx != ctx); | |
1528 | lockdep_assert_held(&ctx->lock); | |
1529 | ||
fa66f07a SE |
1530 | /* |
1531 | * An event which could not be activated because of | |
1532 | * filter mismatch still needs to have its timings | |
1533 | * maintained, otherwise bogus information is return | |
1534 | * via read() for time_enabled, time_running: | |
1535 | */ | |
1536 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1537 | && !event_filter_match(event)) { | |
e5d1367f | 1538 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1539 | event->tstamp_running += delta; |
4158755d | 1540 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1541 | } |
1542 | ||
cdd6c482 | 1543 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1544 | return; |
3b6f9e5c | 1545 | |
44377277 AS |
1546 | perf_pmu_disable(event->pmu); |
1547 | ||
cdd6c482 IM |
1548 | event->state = PERF_EVENT_STATE_INACTIVE; |
1549 | if (event->pending_disable) { | |
1550 | event->pending_disable = 0; | |
1551 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1552 | } |
4158755d | 1553 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1554 | event->pmu->del(event, 0); |
cdd6c482 | 1555 | event->oncpu = -1; |
3b6f9e5c | 1556 | |
cdd6c482 | 1557 | if (!is_software_event(event)) |
3b6f9e5c | 1558 | cpuctx->active_oncpu--; |
2fde4f94 MR |
1559 | if (!--ctx->nr_active) |
1560 | perf_event_ctx_deactivate(ctx); | |
0f5a2601 PZ |
1561 | if (event->attr.freq && event->attr.sample_freq) |
1562 | ctx->nr_freq--; | |
cdd6c482 | 1563 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c | 1564 | cpuctx->exclusive = 0; |
44377277 | 1565 | |
fadfe7be JO |
1566 | if (is_orphaned_child(event)) |
1567 | schedule_orphans_remove(ctx); | |
1568 | ||
44377277 | 1569 | perf_pmu_enable(event->pmu); |
3b6f9e5c PM |
1570 | } |
1571 | ||
d859e29f | 1572 | static void |
cdd6c482 | 1573 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1574 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1575 | struct perf_event_context *ctx) |
d859e29f | 1576 | { |
cdd6c482 | 1577 | struct perf_event *event; |
fa66f07a | 1578 | int state = group_event->state; |
d859e29f | 1579 | |
cdd6c482 | 1580 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1581 | |
1582 | /* | |
1583 | * Schedule out siblings (if any): | |
1584 | */ | |
cdd6c482 IM |
1585 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1586 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1587 | |
fa66f07a | 1588 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1589 | cpuctx->exclusive = 0; |
1590 | } | |
1591 | ||
46ce0fe9 PZ |
1592 | struct remove_event { |
1593 | struct perf_event *event; | |
1594 | bool detach_group; | |
1595 | }; | |
1596 | ||
0793a61d | 1597 | /* |
cdd6c482 | 1598 | * Cross CPU call to remove a performance event |
0793a61d | 1599 | * |
cdd6c482 | 1600 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1601 | * remove it from the context list. |
1602 | */ | |
fe4b04fa | 1603 | static int __perf_remove_from_context(void *info) |
0793a61d | 1604 | { |
46ce0fe9 PZ |
1605 | struct remove_event *re = info; |
1606 | struct perf_event *event = re->event; | |
cdd6c482 | 1607 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1608 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1609 | |
e625cce1 | 1610 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1611 | event_sched_out(event, cpuctx, ctx); |
46ce0fe9 PZ |
1612 | if (re->detach_group) |
1613 | perf_group_detach(event); | |
cdd6c482 | 1614 | list_del_event(event, ctx); |
64ce3126 PZ |
1615 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1616 | ctx->is_active = 0; | |
1617 | cpuctx->task_ctx = NULL; | |
1618 | } | |
e625cce1 | 1619 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1620 | |
1621 | return 0; | |
0793a61d TG |
1622 | } |
1623 | ||
1624 | ||
1625 | /* | |
cdd6c482 | 1626 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1627 | * |
cdd6c482 | 1628 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1629 | * call when the task is on a CPU. |
c93f7669 | 1630 | * |
cdd6c482 IM |
1631 | * If event->ctx is a cloned context, callers must make sure that |
1632 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1633 | * remains valid. This is OK when called from perf_release since |
1634 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1635 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1636 | * context has been detached from its task. |
0793a61d | 1637 | */ |
46ce0fe9 | 1638 | static void perf_remove_from_context(struct perf_event *event, bool detach_group) |
0793a61d | 1639 | { |
cdd6c482 | 1640 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 1641 | struct task_struct *task = ctx->task; |
46ce0fe9 PZ |
1642 | struct remove_event re = { |
1643 | .event = event, | |
1644 | .detach_group = detach_group, | |
1645 | }; | |
0793a61d | 1646 | |
fe4b04fa PZ |
1647 | lockdep_assert_held(&ctx->mutex); |
1648 | ||
0793a61d TG |
1649 | if (!task) { |
1650 | /* | |
226424ee MR |
1651 | * Per cpu events are removed via an smp call. The removal can |
1652 | * fail if the CPU is currently offline, but in that case we | |
1653 | * already called __perf_remove_from_context from | |
1654 | * perf_event_exit_cpu. | |
0793a61d | 1655 | */ |
46ce0fe9 | 1656 | cpu_function_call(event->cpu, __perf_remove_from_context, &re); |
0793a61d TG |
1657 | return; |
1658 | } | |
1659 | ||
1660 | retry: | |
46ce0fe9 | 1661 | if (!task_function_call(task, __perf_remove_from_context, &re)) |
fe4b04fa | 1662 | return; |
0793a61d | 1663 | |
e625cce1 | 1664 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1665 | /* |
fe4b04fa PZ |
1666 | * If we failed to find a running task, but find the context active now |
1667 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1668 | */ |
fe4b04fa | 1669 | if (ctx->is_active) { |
e625cce1 | 1670 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
1671 | /* |
1672 | * Reload the task pointer, it might have been changed by | |
1673 | * a concurrent perf_event_context_sched_out(). | |
1674 | */ | |
1675 | task = ctx->task; | |
0793a61d TG |
1676 | goto retry; |
1677 | } | |
1678 | ||
1679 | /* | |
fe4b04fa PZ |
1680 | * Since the task isn't running, its safe to remove the event, us |
1681 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1682 | */ |
46ce0fe9 PZ |
1683 | if (detach_group) |
1684 | perf_group_detach(event); | |
fe4b04fa | 1685 | list_del_event(event, ctx); |
e625cce1 | 1686 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1687 | } |
1688 | ||
d859e29f | 1689 | /* |
cdd6c482 | 1690 | * Cross CPU call to disable a performance event |
d859e29f | 1691 | */ |
500ad2d8 | 1692 | int __perf_event_disable(void *info) |
d859e29f | 1693 | { |
cdd6c482 | 1694 | struct perf_event *event = info; |
cdd6c482 | 1695 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1696 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1697 | |
1698 | /* | |
cdd6c482 IM |
1699 | * If this is a per-task event, need to check whether this |
1700 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1701 | * |
1702 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1703 | * flipping contexts around. | |
d859e29f | 1704 | */ |
665c2142 | 1705 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1706 | return -EINVAL; |
d859e29f | 1707 | |
e625cce1 | 1708 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1709 | |
1710 | /* | |
cdd6c482 | 1711 | * If the event is on, turn it off. |
d859e29f PM |
1712 | * If it is in error state, leave it in error state. |
1713 | */ | |
cdd6c482 | 1714 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1715 | update_context_time(ctx); |
e5d1367f | 1716 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1717 | update_group_times(event); |
1718 | if (event == event->group_leader) | |
1719 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1720 | else |
cdd6c482 IM |
1721 | event_sched_out(event, cpuctx, ctx); |
1722 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1723 | } |
1724 | ||
e625cce1 | 1725 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1726 | |
1727 | return 0; | |
d859e29f PM |
1728 | } |
1729 | ||
1730 | /* | |
cdd6c482 | 1731 | * Disable a event. |
c93f7669 | 1732 | * |
cdd6c482 IM |
1733 | * If event->ctx is a cloned context, callers must make sure that |
1734 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1735 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1736 | * perf_event_for_each_child or perf_event_for_each because they |
1737 | * hold the top-level event's child_mutex, so any descendant that | |
1738 | * goes to exit will block in sync_child_event. | |
1739 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1740 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1741 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1742 | */ |
f63a8daa | 1743 | static void _perf_event_disable(struct perf_event *event) |
d859e29f | 1744 | { |
cdd6c482 | 1745 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1746 | struct task_struct *task = ctx->task; |
1747 | ||
1748 | if (!task) { | |
1749 | /* | |
cdd6c482 | 1750 | * Disable the event on the cpu that it's on |
d859e29f | 1751 | */ |
fe4b04fa | 1752 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1753 | return; |
1754 | } | |
1755 | ||
9ed6060d | 1756 | retry: |
fe4b04fa PZ |
1757 | if (!task_function_call(task, __perf_event_disable, event)) |
1758 | return; | |
d859e29f | 1759 | |
e625cce1 | 1760 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1761 | /* |
cdd6c482 | 1762 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1763 | */ |
cdd6c482 | 1764 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1765 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1766 | /* |
1767 | * Reload the task pointer, it might have been changed by | |
1768 | * a concurrent perf_event_context_sched_out(). | |
1769 | */ | |
1770 | task = ctx->task; | |
d859e29f PM |
1771 | goto retry; |
1772 | } | |
1773 | ||
1774 | /* | |
1775 | * Since we have the lock this context can't be scheduled | |
1776 | * in, so we can change the state safely. | |
1777 | */ | |
cdd6c482 IM |
1778 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1779 | update_group_times(event); | |
1780 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1781 | } |
e625cce1 | 1782 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1783 | } |
f63a8daa PZ |
1784 | |
1785 | /* | |
1786 | * Strictly speaking kernel users cannot create groups and therefore this | |
1787 | * interface does not need the perf_event_ctx_lock() magic. | |
1788 | */ | |
1789 | void perf_event_disable(struct perf_event *event) | |
1790 | { | |
1791 | struct perf_event_context *ctx; | |
1792 | ||
1793 | ctx = perf_event_ctx_lock(event); | |
1794 | _perf_event_disable(event); | |
1795 | perf_event_ctx_unlock(event, ctx); | |
1796 | } | |
dcfce4a0 | 1797 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1798 | |
e5d1367f SE |
1799 | static void perf_set_shadow_time(struct perf_event *event, |
1800 | struct perf_event_context *ctx, | |
1801 | u64 tstamp) | |
1802 | { | |
1803 | /* | |
1804 | * use the correct time source for the time snapshot | |
1805 | * | |
1806 | * We could get by without this by leveraging the | |
1807 | * fact that to get to this function, the caller | |
1808 | * has most likely already called update_context_time() | |
1809 | * and update_cgrp_time_xx() and thus both timestamp | |
1810 | * are identical (or very close). Given that tstamp is, | |
1811 | * already adjusted for cgroup, we could say that: | |
1812 | * tstamp - ctx->timestamp | |
1813 | * is equivalent to | |
1814 | * tstamp - cgrp->timestamp. | |
1815 | * | |
1816 | * Then, in perf_output_read(), the calculation would | |
1817 | * work with no changes because: | |
1818 | * - event is guaranteed scheduled in | |
1819 | * - no scheduled out in between | |
1820 | * - thus the timestamp would be the same | |
1821 | * | |
1822 | * But this is a bit hairy. | |
1823 | * | |
1824 | * So instead, we have an explicit cgroup call to remain | |
1825 | * within the time time source all along. We believe it | |
1826 | * is cleaner and simpler to understand. | |
1827 | */ | |
1828 | if (is_cgroup_event(event)) | |
1829 | perf_cgroup_set_shadow_time(event, tstamp); | |
1830 | else | |
1831 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1832 | } | |
1833 | ||
4fe757dd PZ |
1834 | #define MAX_INTERRUPTS (~0ULL) |
1835 | ||
1836 | static void perf_log_throttle(struct perf_event *event, int enable); | |
ec0d7729 | 1837 | static void perf_log_itrace_start(struct perf_event *event); |
4fe757dd | 1838 | |
235c7fc7 | 1839 | static int |
9ffcfa6f | 1840 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1841 | struct perf_cpu_context *cpuctx, |
6e37738a | 1842 | struct perf_event_context *ctx) |
235c7fc7 | 1843 | { |
4158755d | 1844 | u64 tstamp = perf_event_time(event); |
44377277 | 1845 | int ret = 0; |
4158755d | 1846 | |
63342411 PZ |
1847 | lockdep_assert_held(&ctx->lock); |
1848 | ||
cdd6c482 | 1849 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1850 | return 0; |
1851 | ||
cdd6c482 | 1852 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1853 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1854 | |
1855 | /* | |
1856 | * Unthrottle events, since we scheduled we might have missed several | |
1857 | * ticks already, also for a heavily scheduling task there is little | |
1858 | * guarantee it'll get a tick in a timely manner. | |
1859 | */ | |
1860 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1861 | perf_log_throttle(event, 1); | |
1862 | event->hw.interrupts = 0; | |
1863 | } | |
1864 | ||
235c7fc7 IM |
1865 | /* |
1866 | * The new state must be visible before we turn it on in the hardware: | |
1867 | */ | |
1868 | smp_wmb(); | |
1869 | ||
44377277 AS |
1870 | perf_pmu_disable(event->pmu); |
1871 | ||
72f669c0 SL |
1872 | perf_set_shadow_time(event, ctx, tstamp); |
1873 | ||
ec0d7729 AS |
1874 | perf_log_itrace_start(event); |
1875 | ||
a4eaf7f1 | 1876 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1877 | event->state = PERF_EVENT_STATE_INACTIVE; |
1878 | event->oncpu = -1; | |
44377277 AS |
1879 | ret = -EAGAIN; |
1880 | goto out; | |
235c7fc7 IM |
1881 | } |
1882 | ||
00a2916f PZ |
1883 | event->tstamp_running += tstamp - event->tstamp_stopped; |
1884 | ||
cdd6c482 | 1885 | if (!is_software_event(event)) |
3b6f9e5c | 1886 | cpuctx->active_oncpu++; |
2fde4f94 MR |
1887 | if (!ctx->nr_active++) |
1888 | perf_event_ctx_activate(ctx); | |
0f5a2601 PZ |
1889 | if (event->attr.freq && event->attr.sample_freq) |
1890 | ctx->nr_freq++; | |
235c7fc7 | 1891 | |
cdd6c482 | 1892 | if (event->attr.exclusive) |
3b6f9e5c PM |
1893 | cpuctx->exclusive = 1; |
1894 | ||
fadfe7be JO |
1895 | if (is_orphaned_child(event)) |
1896 | schedule_orphans_remove(ctx); | |
1897 | ||
44377277 AS |
1898 | out: |
1899 | perf_pmu_enable(event->pmu); | |
1900 | ||
1901 | return ret; | |
235c7fc7 IM |
1902 | } |
1903 | ||
6751b71e | 1904 | static int |
cdd6c482 | 1905 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1906 | struct perf_cpu_context *cpuctx, |
6e37738a | 1907 | struct perf_event_context *ctx) |
6751b71e | 1908 | { |
6bde9b6c | 1909 | struct perf_event *event, *partial_group = NULL; |
4a234593 | 1910 | struct pmu *pmu = ctx->pmu; |
d7842da4 SE |
1911 | u64 now = ctx->time; |
1912 | bool simulate = false; | |
6751b71e | 1913 | |
cdd6c482 | 1914 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1915 | return 0; |
1916 | ||
ad5133b7 | 1917 | pmu->start_txn(pmu); |
6bde9b6c | 1918 | |
9ffcfa6f | 1919 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1920 | pmu->cancel_txn(pmu); |
272325c4 | 1921 | perf_mux_hrtimer_restart(cpuctx); |
6751b71e | 1922 | return -EAGAIN; |
90151c35 | 1923 | } |
6751b71e PM |
1924 | |
1925 | /* | |
1926 | * Schedule in siblings as one group (if any): | |
1927 | */ | |
cdd6c482 | 1928 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1929 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1930 | partial_group = event; |
6751b71e PM |
1931 | goto group_error; |
1932 | } | |
1933 | } | |
1934 | ||
9ffcfa6f | 1935 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1936 | return 0; |
9ffcfa6f | 1937 | |
6751b71e PM |
1938 | group_error: |
1939 | /* | |
1940 | * Groups can be scheduled in as one unit only, so undo any | |
1941 | * partial group before returning: | |
d7842da4 SE |
1942 | * The events up to the failed event are scheduled out normally, |
1943 | * tstamp_stopped will be updated. | |
1944 | * | |
1945 | * The failed events and the remaining siblings need to have | |
1946 | * their timings updated as if they had gone thru event_sched_in() | |
1947 | * and event_sched_out(). This is required to get consistent timings | |
1948 | * across the group. This also takes care of the case where the group | |
1949 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1950 | * the time the event was actually stopped, such that time delta | |
1951 | * calculation in update_event_times() is correct. | |
6751b71e | 1952 | */ |
cdd6c482 IM |
1953 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1954 | if (event == partial_group) | |
d7842da4 SE |
1955 | simulate = true; |
1956 | ||
1957 | if (simulate) { | |
1958 | event->tstamp_running += now - event->tstamp_stopped; | |
1959 | event->tstamp_stopped = now; | |
1960 | } else { | |
1961 | event_sched_out(event, cpuctx, ctx); | |
1962 | } | |
6751b71e | 1963 | } |
9ffcfa6f | 1964 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1965 | |
ad5133b7 | 1966 | pmu->cancel_txn(pmu); |
90151c35 | 1967 | |
272325c4 | 1968 | perf_mux_hrtimer_restart(cpuctx); |
9e630205 | 1969 | |
6751b71e PM |
1970 | return -EAGAIN; |
1971 | } | |
1972 | ||
3b6f9e5c | 1973 | /* |
cdd6c482 | 1974 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1975 | */ |
cdd6c482 | 1976 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1977 | struct perf_cpu_context *cpuctx, |
1978 | int can_add_hw) | |
1979 | { | |
1980 | /* | |
cdd6c482 | 1981 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1982 | */ |
d6f962b5 | 1983 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1984 | return 1; |
1985 | /* | |
1986 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1987 | * events can go on. |
3b6f9e5c PM |
1988 | */ |
1989 | if (cpuctx->exclusive) | |
1990 | return 0; | |
1991 | /* | |
1992 | * If this group is exclusive and there are already | |
cdd6c482 | 1993 | * events on the CPU, it can't go on. |
3b6f9e5c | 1994 | */ |
cdd6c482 | 1995 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1996 | return 0; |
1997 | /* | |
1998 | * Otherwise, try to add it if all previous groups were able | |
1999 | * to go on. | |
2000 | */ | |
2001 | return can_add_hw; | |
2002 | } | |
2003 | ||
cdd6c482 IM |
2004 | static void add_event_to_ctx(struct perf_event *event, |
2005 | struct perf_event_context *ctx) | |
53cfbf59 | 2006 | { |
4158755d SE |
2007 | u64 tstamp = perf_event_time(event); |
2008 | ||
cdd6c482 | 2009 | list_add_event(event, ctx); |
8a49542c | 2010 | perf_group_attach(event); |
4158755d SE |
2011 | event->tstamp_enabled = tstamp; |
2012 | event->tstamp_running = tstamp; | |
2013 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
2014 | } |
2015 | ||
2c29ef0f PZ |
2016 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
2017 | static void | |
2018 | ctx_sched_in(struct perf_event_context *ctx, | |
2019 | struct perf_cpu_context *cpuctx, | |
2020 | enum event_type_t event_type, | |
2021 | struct task_struct *task); | |
fe4b04fa | 2022 | |
dce5855b PZ |
2023 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
2024 | struct perf_event_context *ctx, | |
2025 | struct task_struct *task) | |
2026 | { | |
2027 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
2028 | if (ctx) | |
2029 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
2030 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
2031 | if (ctx) | |
2032 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
2033 | } | |
2034 | ||
0793a61d | 2035 | /* |
cdd6c482 | 2036 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
2037 | * |
2038 | * Must be called with ctx->mutex held | |
0793a61d | 2039 | */ |
fe4b04fa | 2040 | static int __perf_install_in_context(void *info) |
0793a61d | 2041 | { |
cdd6c482 IM |
2042 | struct perf_event *event = info; |
2043 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2044 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
2045 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
2046 | struct task_struct *task = current; | |
2047 | ||
b58f6b0d | 2048 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 2049 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
2050 | |
2051 | /* | |
2c29ef0f | 2052 | * If there was an active task_ctx schedule it out. |
0793a61d | 2053 | */ |
b58f6b0d | 2054 | if (task_ctx) |
2c29ef0f | 2055 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
2056 | |
2057 | /* | |
2058 | * If the context we're installing events in is not the | |
2059 | * active task_ctx, flip them. | |
2060 | */ | |
2061 | if (ctx->task && task_ctx != ctx) { | |
2062 | if (task_ctx) | |
2063 | raw_spin_unlock(&task_ctx->lock); | |
2064 | raw_spin_lock(&ctx->lock); | |
2065 | task_ctx = ctx; | |
2066 | } | |
2067 | ||
2068 | if (task_ctx) { | |
2069 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
2070 | task = task_ctx->task; |
2071 | } | |
b58f6b0d | 2072 | |
2c29ef0f | 2073 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 2074 | |
4af4998b | 2075 | update_context_time(ctx); |
e5d1367f SE |
2076 | /* |
2077 | * update cgrp time only if current cgrp | |
2078 | * matches event->cgrp. Must be done before | |
2079 | * calling add_event_to_ctx() | |
2080 | */ | |
2081 | update_cgrp_time_from_event(event); | |
0793a61d | 2082 | |
cdd6c482 | 2083 | add_event_to_ctx(event, ctx); |
0793a61d | 2084 | |
d859e29f | 2085 | /* |
2c29ef0f | 2086 | * Schedule everything back in |
d859e29f | 2087 | */ |
dce5855b | 2088 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
2089 | |
2090 | perf_pmu_enable(cpuctx->ctx.pmu); | |
2091 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
2092 | |
2093 | return 0; | |
0793a61d TG |
2094 | } |
2095 | ||
2096 | /* | |
cdd6c482 | 2097 | * Attach a performance event to a context |
0793a61d | 2098 | * |
cdd6c482 IM |
2099 | * First we add the event to the list with the hardware enable bit |
2100 | * in event->hw_config cleared. | |
0793a61d | 2101 | * |
cdd6c482 | 2102 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
2103 | * call to enable it in the task context. The task might have been |
2104 | * scheduled away, but we check this in the smp call again. | |
2105 | */ | |
2106 | static void | |
cdd6c482 IM |
2107 | perf_install_in_context(struct perf_event_context *ctx, |
2108 | struct perf_event *event, | |
0793a61d TG |
2109 | int cpu) |
2110 | { | |
2111 | struct task_struct *task = ctx->task; | |
2112 | ||
fe4b04fa PZ |
2113 | lockdep_assert_held(&ctx->mutex); |
2114 | ||
c3f00c70 | 2115 | event->ctx = ctx; |
0cda4c02 YZ |
2116 | if (event->cpu != -1) |
2117 | event->cpu = cpu; | |
c3f00c70 | 2118 | |
0793a61d TG |
2119 | if (!task) { |
2120 | /* | |
cdd6c482 | 2121 | * Per cpu events are installed via an smp call and |
af901ca1 | 2122 | * the install is always successful. |
0793a61d | 2123 | */ |
fe4b04fa | 2124 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
2125 | return; |
2126 | } | |
2127 | ||
0793a61d | 2128 | retry: |
fe4b04fa PZ |
2129 | if (!task_function_call(task, __perf_install_in_context, event)) |
2130 | return; | |
0793a61d | 2131 | |
e625cce1 | 2132 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 2133 | /* |
fe4b04fa PZ |
2134 | * If we failed to find a running task, but find the context active now |
2135 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 2136 | */ |
fe4b04fa | 2137 | if (ctx->is_active) { |
e625cce1 | 2138 | raw_spin_unlock_irq(&ctx->lock); |
3577af70 CW |
2139 | /* |
2140 | * Reload the task pointer, it might have been changed by | |
2141 | * a concurrent perf_event_context_sched_out(). | |
2142 | */ | |
2143 | task = ctx->task; | |
0793a61d TG |
2144 | goto retry; |
2145 | } | |
2146 | ||
2147 | /* | |
fe4b04fa PZ |
2148 | * Since the task isn't running, its safe to add the event, us holding |
2149 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 2150 | */ |
fe4b04fa | 2151 | add_event_to_ctx(event, ctx); |
e625cce1 | 2152 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
2153 | } |
2154 | ||
fa289bec | 2155 | /* |
cdd6c482 | 2156 | * Put a event into inactive state and update time fields. |
fa289bec PM |
2157 | * Enabling the leader of a group effectively enables all |
2158 | * the group members that aren't explicitly disabled, so we | |
2159 | * have to update their ->tstamp_enabled also. | |
2160 | * Note: this works for group members as well as group leaders | |
2161 | * since the non-leader members' sibling_lists will be empty. | |
2162 | */ | |
1d9b482e | 2163 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 2164 | { |
cdd6c482 | 2165 | struct perf_event *sub; |
4158755d | 2166 | u64 tstamp = perf_event_time(event); |
fa289bec | 2167 | |
cdd6c482 | 2168 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 2169 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 2170 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
2171 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
2172 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 2173 | } |
fa289bec PM |
2174 | } |
2175 | ||
d859e29f | 2176 | /* |
cdd6c482 | 2177 | * Cross CPU call to enable a performance event |
d859e29f | 2178 | */ |
fe4b04fa | 2179 | static int __perf_event_enable(void *info) |
04289bb9 | 2180 | { |
cdd6c482 | 2181 | struct perf_event *event = info; |
cdd6c482 IM |
2182 | struct perf_event_context *ctx = event->ctx; |
2183 | struct perf_event *leader = event->group_leader; | |
108b02cf | 2184 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 2185 | int err; |
04289bb9 | 2186 | |
06f41796 JO |
2187 | /* |
2188 | * There's a time window between 'ctx->is_active' check | |
2189 | * in perf_event_enable function and this place having: | |
2190 | * - IRQs on | |
2191 | * - ctx->lock unlocked | |
2192 | * | |
2193 | * where the task could be killed and 'ctx' deactivated | |
2194 | * by perf_event_exit_task. | |
2195 | */ | |
2196 | if (!ctx->is_active) | |
fe4b04fa | 2197 | return -EINVAL; |
3cbed429 | 2198 | |
e625cce1 | 2199 | raw_spin_lock(&ctx->lock); |
4af4998b | 2200 | update_context_time(ctx); |
d859e29f | 2201 | |
cdd6c482 | 2202 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 2203 | goto unlock; |
e5d1367f SE |
2204 | |
2205 | /* | |
2206 | * set current task's cgroup time reference point | |
2207 | */ | |
3f7cce3c | 2208 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 2209 | |
1d9b482e | 2210 | __perf_event_mark_enabled(event); |
04289bb9 | 2211 | |
e5d1367f SE |
2212 | if (!event_filter_match(event)) { |
2213 | if (is_cgroup_event(event)) | |
2214 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 2215 | goto unlock; |
e5d1367f | 2216 | } |
f4c4176f | 2217 | |
04289bb9 | 2218 | /* |
cdd6c482 | 2219 | * If the event is in a group and isn't the group leader, |
d859e29f | 2220 | * then don't put it on unless the group is on. |
04289bb9 | 2221 | */ |
cdd6c482 | 2222 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2223 | goto unlock; |
3b6f9e5c | 2224 | |
cdd6c482 | 2225 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2226 | err = -EEXIST; |
e758a33d | 2227 | } else { |
cdd6c482 | 2228 | if (event == leader) |
6e37738a | 2229 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2230 | else |
6e37738a | 2231 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2232 | } |
d859e29f PM |
2233 | |
2234 | if (err) { | |
2235 | /* | |
cdd6c482 | 2236 | * If this event can't go on and it's part of a |
d859e29f PM |
2237 | * group, then the whole group has to come off. |
2238 | */ | |
9e630205 | 2239 | if (leader != event) { |
d859e29f | 2240 | group_sched_out(leader, cpuctx, ctx); |
272325c4 | 2241 | perf_mux_hrtimer_restart(cpuctx); |
9e630205 | 2242 | } |
0d48696f | 2243 | if (leader->attr.pinned) { |
53cfbf59 | 2244 | update_group_times(leader); |
cdd6c482 | 2245 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2246 | } |
d859e29f PM |
2247 | } |
2248 | ||
9ed6060d | 2249 | unlock: |
e625cce1 | 2250 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2251 | |
2252 | return 0; | |
d859e29f PM |
2253 | } |
2254 | ||
2255 | /* | |
cdd6c482 | 2256 | * Enable a event. |
c93f7669 | 2257 | * |
cdd6c482 IM |
2258 | * If event->ctx is a cloned context, callers must make sure that |
2259 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2260 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2261 | * perf_event_for_each_child or perf_event_for_each as described |
2262 | * for perf_event_disable. | |
d859e29f | 2263 | */ |
f63a8daa | 2264 | static void _perf_event_enable(struct perf_event *event) |
d859e29f | 2265 | { |
cdd6c482 | 2266 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2267 | struct task_struct *task = ctx->task; |
2268 | ||
2269 | if (!task) { | |
2270 | /* | |
cdd6c482 | 2271 | * Enable the event on the cpu that it's on |
d859e29f | 2272 | */ |
fe4b04fa | 2273 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2274 | return; |
2275 | } | |
2276 | ||
e625cce1 | 2277 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2278 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2279 | goto out; |
2280 | ||
2281 | /* | |
cdd6c482 IM |
2282 | * If the event is in error state, clear that first. |
2283 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2284 | * know that it has gone back into error state, as distinct |
2285 | * from the task having been scheduled away before the | |
2286 | * cross-call arrived. | |
2287 | */ | |
cdd6c482 IM |
2288 | if (event->state == PERF_EVENT_STATE_ERROR) |
2289 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2290 | |
9ed6060d | 2291 | retry: |
fe4b04fa | 2292 | if (!ctx->is_active) { |
1d9b482e | 2293 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2294 | goto out; |
2295 | } | |
2296 | ||
e625cce1 | 2297 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2298 | |
2299 | if (!task_function_call(task, __perf_event_enable, event)) | |
2300 | return; | |
d859e29f | 2301 | |
e625cce1 | 2302 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2303 | |
2304 | /* | |
cdd6c482 | 2305 | * If the context is active and the event is still off, |
d859e29f PM |
2306 | * we need to retry the cross-call. |
2307 | */ | |
fe4b04fa PZ |
2308 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2309 | /* | |
2310 | * task could have been flipped by a concurrent | |
2311 | * perf_event_context_sched_out() | |
2312 | */ | |
2313 | task = ctx->task; | |
d859e29f | 2314 | goto retry; |
fe4b04fa | 2315 | } |
fa289bec | 2316 | |
9ed6060d | 2317 | out: |
e625cce1 | 2318 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2319 | } |
f63a8daa PZ |
2320 | |
2321 | /* | |
2322 | * See perf_event_disable(); | |
2323 | */ | |
2324 | void perf_event_enable(struct perf_event *event) | |
2325 | { | |
2326 | struct perf_event_context *ctx; | |
2327 | ||
2328 | ctx = perf_event_ctx_lock(event); | |
2329 | _perf_event_enable(event); | |
2330 | perf_event_ctx_unlock(event, ctx); | |
2331 | } | |
dcfce4a0 | 2332 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2333 | |
f63a8daa | 2334 | static int _perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2335 | { |
2023b359 | 2336 | /* |
cdd6c482 | 2337 | * not supported on inherited events |
2023b359 | 2338 | */ |
2e939d1d | 2339 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2340 | return -EINVAL; |
2341 | ||
cdd6c482 | 2342 | atomic_add(refresh, &event->event_limit); |
f63a8daa | 2343 | _perf_event_enable(event); |
2023b359 PZ |
2344 | |
2345 | return 0; | |
79f14641 | 2346 | } |
f63a8daa PZ |
2347 | |
2348 | /* | |
2349 | * See perf_event_disable() | |
2350 | */ | |
2351 | int perf_event_refresh(struct perf_event *event, int refresh) | |
2352 | { | |
2353 | struct perf_event_context *ctx; | |
2354 | int ret; | |
2355 | ||
2356 | ctx = perf_event_ctx_lock(event); | |
2357 | ret = _perf_event_refresh(event, refresh); | |
2358 | perf_event_ctx_unlock(event, ctx); | |
2359 | ||
2360 | return ret; | |
2361 | } | |
26ca5c11 | 2362 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2363 | |
5b0311e1 FW |
2364 | static void ctx_sched_out(struct perf_event_context *ctx, |
2365 | struct perf_cpu_context *cpuctx, | |
2366 | enum event_type_t event_type) | |
235c7fc7 | 2367 | { |
cdd6c482 | 2368 | struct perf_event *event; |
db24d33e | 2369 | int is_active = ctx->is_active; |
235c7fc7 | 2370 | |
db24d33e | 2371 | ctx->is_active &= ~event_type; |
cdd6c482 | 2372 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2373 | return; |
2374 | ||
4af4998b | 2375 | update_context_time(ctx); |
e5d1367f | 2376 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2377 | if (!ctx->nr_active) |
facc4307 | 2378 | return; |
5b0311e1 | 2379 | |
075e0b00 | 2380 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2381 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2382 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2383 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2384 | } |
889ff015 | 2385 | |
db24d33e | 2386 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2387 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2388 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2389 | } |
1b9a644f | 2390 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2391 | } |
2392 | ||
564c2b21 | 2393 | /* |
5a3126d4 PZ |
2394 | * Test whether two contexts are equivalent, i.e. whether they have both been |
2395 | * cloned from the same version of the same context. | |
2396 | * | |
2397 | * Equivalence is measured using a generation number in the context that is | |
2398 | * incremented on each modification to it; see unclone_ctx(), list_add_event() | |
2399 | * and list_del_event(). | |
564c2b21 | 2400 | */ |
cdd6c482 IM |
2401 | static int context_equiv(struct perf_event_context *ctx1, |
2402 | struct perf_event_context *ctx2) | |
564c2b21 | 2403 | { |
211de6eb PZ |
2404 | lockdep_assert_held(&ctx1->lock); |
2405 | lockdep_assert_held(&ctx2->lock); | |
2406 | ||
5a3126d4 PZ |
2407 | /* Pinning disables the swap optimization */ |
2408 | if (ctx1->pin_count || ctx2->pin_count) | |
2409 | return 0; | |
2410 | ||
2411 | /* If ctx1 is the parent of ctx2 */ | |
2412 | if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen) | |
2413 | return 1; | |
2414 | ||
2415 | /* If ctx2 is the parent of ctx1 */ | |
2416 | if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation) | |
2417 | return 1; | |
2418 | ||
2419 | /* | |
2420 | * If ctx1 and ctx2 have the same parent; we flatten the parent | |
2421 | * hierarchy, see perf_event_init_context(). | |
2422 | */ | |
2423 | if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx && | |
2424 | ctx1->parent_gen == ctx2->parent_gen) | |
2425 | return 1; | |
2426 | ||
2427 | /* Unmatched */ | |
2428 | return 0; | |
564c2b21 PM |
2429 | } |
2430 | ||
cdd6c482 IM |
2431 | static void __perf_event_sync_stat(struct perf_event *event, |
2432 | struct perf_event *next_event) | |
bfbd3381 PZ |
2433 | { |
2434 | u64 value; | |
2435 | ||
cdd6c482 | 2436 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2437 | return; |
2438 | ||
2439 | /* | |
cdd6c482 | 2440 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2441 | * because we're in the middle of a context switch and have IRQs |
2442 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2443 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2444 | * don't need to use it. |
2445 | */ | |
cdd6c482 IM |
2446 | switch (event->state) { |
2447 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2448 | event->pmu->read(event); |
2449 | /* fall-through */ | |
bfbd3381 | 2450 | |
cdd6c482 IM |
2451 | case PERF_EVENT_STATE_INACTIVE: |
2452 | update_event_times(event); | |
bfbd3381 PZ |
2453 | break; |
2454 | ||
2455 | default: | |
2456 | break; | |
2457 | } | |
2458 | ||
2459 | /* | |
cdd6c482 | 2460 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2461 | * values when we flip the contexts. |
2462 | */ | |
e7850595 PZ |
2463 | value = local64_read(&next_event->count); |
2464 | value = local64_xchg(&event->count, value); | |
2465 | local64_set(&next_event->count, value); | |
bfbd3381 | 2466 | |
cdd6c482 IM |
2467 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2468 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2469 | |
bfbd3381 | 2470 | /* |
19d2e755 | 2471 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2472 | */ |
cdd6c482 IM |
2473 | perf_event_update_userpage(event); |
2474 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2475 | } |
2476 | ||
cdd6c482 IM |
2477 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2478 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2479 | { |
cdd6c482 | 2480 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2481 | |
2482 | if (!ctx->nr_stat) | |
2483 | return; | |
2484 | ||
02ffdbc8 PZ |
2485 | update_context_time(ctx); |
2486 | ||
cdd6c482 IM |
2487 | event = list_first_entry(&ctx->event_list, |
2488 | struct perf_event, event_entry); | |
bfbd3381 | 2489 | |
cdd6c482 IM |
2490 | next_event = list_first_entry(&next_ctx->event_list, |
2491 | struct perf_event, event_entry); | |
bfbd3381 | 2492 | |
cdd6c482 IM |
2493 | while (&event->event_entry != &ctx->event_list && |
2494 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2495 | |
cdd6c482 | 2496 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2497 | |
cdd6c482 IM |
2498 | event = list_next_entry(event, event_entry); |
2499 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2500 | } |
2501 | } | |
2502 | ||
fe4b04fa PZ |
2503 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2504 | struct task_struct *next) | |
0793a61d | 2505 | { |
8dc85d54 | 2506 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 | 2507 | struct perf_event_context *next_ctx; |
5a3126d4 | 2508 | struct perf_event_context *parent, *next_parent; |
108b02cf | 2509 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2510 | int do_switch = 1; |
0793a61d | 2511 | |
108b02cf PZ |
2512 | if (likely(!ctx)) |
2513 | return; | |
10989fb2 | 2514 | |
108b02cf PZ |
2515 | cpuctx = __get_cpu_context(ctx); |
2516 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2517 | return; |
2518 | ||
c93f7669 | 2519 | rcu_read_lock(); |
8dc85d54 | 2520 | next_ctx = next->perf_event_ctxp[ctxn]; |
5a3126d4 PZ |
2521 | if (!next_ctx) |
2522 | goto unlock; | |
2523 | ||
2524 | parent = rcu_dereference(ctx->parent_ctx); | |
2525 | next_parent = rcu_dereference(next_ctx->parent_ctx); | |
2526 | ||
2527 | /* If neither context have a parent context; they cannot be clones. */ | |
802c8a61 | 2528 | if (!parent && !next_parent) |
5a3126d4 PZ |
2529 | goto unlock; |
2530 | ||
2531 | if (next_parent == ctx || next_ctx == parent || next_parent == parent) { | |
c93f7669 PM |
2532 | /* |
2533 | * Looks like the two contexts are clones, so we might be | |
2534 | * able to optimize the context switch. We lock both | |
2535 | * contexts and check that they are clones under the | |
2536 | * lock (including re-checking that neither has been | |
2537 | * uncloned in the meantime). It doesn't matter which | |
2538 | * order we take the locks because no other cpu could | |
2539 | * be trying to lock both of these tasks. | |
2540 | */ | |
e625cce1 TG |
2541 | raw_spin_lock(&ctx->lock); |
2542 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2543 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2544 | /* |
2545 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2546 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2547 | */ |
8dc85d54 PZ |
2548 | task->perf_event_ctxp[ctxn] = next_ctx; |
2549 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2550 | ctx->task = next; |
2551 | next_ctx->task = task; | |
5a158c3c YZ |
2552 | |
2553 | swap(ctx->task_ctx_data, next_ctx->task_ctx_data); | |
2554 | ||
c93f7669 | 2555 | do_switch = 0; |
bfbd3381 | 2556 | |
cdd6c482 | 2557 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2558 | } |
e625cce1 TG |
2559 | raw_spin_unlock(&next_ctx->lock); |
2560 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2561 | } |
5a3126d4 | 2562 | unlock: |
c93f7669 | 2563 | rcu_read_unlock(); |
564c2b21 | 2564 | |
c93f7669 | 2565 | if (do_switch) { |
facc4307 | 2566 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2567 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2568 | cpuctx->task_ctx = NULL; |
facc4307 | 2569 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2570 | } |
0793a61d TG |
2571 | } |
2572 | ||
ba532500 YZ |
2573 | void perf_sched_cb_dec(struct pmu *pmu) |
2574 | { | |
2575 | this_cpu_dec(perf_sched_cb_usages); | |
2576 | } | |
2577 | ||
2578 | void perf_sched_cb_inc(struct pmu *pmu) | |
2579 | { | |
2580 | this_cpu_inc(perf_sched_cb_usages); | |
2581 | } | |
2582 | ||
2583 | /* | |
2584 | * This function provides the context switch callback to the lower code | |
2585 | * layer. It is invoked ONLY when the context switch callback is enabled. | |
2586 | */ | |
2587 | static void perf_pmu_sched_task(struct task_struct *prev, | |
2588 | struct task_struct *next, | |
2589 | bool sched_in) | |
2590 | { | |
2591 | struct perf_cpu_context *cpuctx; | |
2592 | struct pmu *pmu; | |
2593 | unsigned long flags; | |
2594 | ||
2595 | if (prev == next) | |
2596 | return; | |
2597 | ||
2598 | local_irq_save(flags); | |
2599 | ||
2600 | rcu_read_lock(); | |
2601 | ||
2602 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2603 | if (pmu->sched_task) { | |
2604 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2605 | ||
2606 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2607 | ||
2608 | perf_pmu_disable(pmu); | |
2609 | ||
2610 | pmu->sched_task(cpuctx->task_ctx, sched_in); | |
2611 | ||
2612 | perf_pmu_enable(pmu); | |
2613 | ||
2614 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2615 | } | |
2616 | } | |
2617 | ||
2618 | rcu_read_unlock(); | |
2619 | ||
2620 | local_irq_restore(flags); | |
2621 | } | |
2622 | ||
45ac1403 AH |
2623 | static void perf_event_switch(struct task_struct *task, |
2624 | struct task_struct *next_prev, bool sched_in); | |
2625 | ||
8dc85d54 PZ |
2626 | #define for_each_task_context_nr(ctxn) \ |
2627 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2628 | ||
2629 | /* | |
2630 | * Called from scheduler to remove the events of the current task, | |
2631 | * with interrupts disabled. | |
2632 | * | |
2633 | * We stop each event and update the event value in event->count. | |
2634 | * | |
2635 | * This does not protect us against NMI, but disable() | |
2636 | * sets the disabled bit in the control field of event _before_ | |
2637 | * accessing the event control register. If a NMI hits, then it will | |
2638 | * not restart the event. | |
2639 | */ | |
ab0cce56 JO |
2640 | void __perf_event_task_sched_out(struct task_struct *task, |
2641 | struct task_struct *next) | |
8dc85d54 PZ |
2642 | { |
2643 | int ctxn; | |
2644 | ||
ba532500 YZ |
2645 | if (__this_cpu_read(perf_sched_cb_usages)) |
2646 | perf_pmu_sched_task(task, next, false); | |
2647 | ||
45ac1403 AH |
2648 | if (atomic_read(&nr_switch_events)) |
2649 | perf_event_switch(task, next, false); | |
2650 | ||
8dc85d54 PZ |
2651 | for_each_task_context_nr(ctxn) |
2652 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2653 | |
2654 | /* | |
2655 | * if cgroup events exist on this CPU, then we need | |
2656 | * to check if we have to switch out PMU state. | |
2657 | * cgroup event are system-wide mode only | |
2658 | */ | |
4a32fea9 | 2659 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2660 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2661 | } |
2662 | ||
04dc2dbb | 2663 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2664 | { |
108b02cf | 2665 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2666 | |
a63eaf34 PM |
2667 | if (!cpuctx->task_ctx) |
2668 | return; | |
012b84da IM |
2669 | |
2670 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2671 | return; | |
2672 | ||
04dc2dbb | 2673 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2674 | cpuctx->task_ctx = NULL; |
2675 | } | |
2676 | ||
5b0311e1 FW |
2677 | /* |
2678 | * Called with IRQs disabled | |
2679 | */ | |
2680 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2681 | enum event_type_t event_type) | |
2682 | { | |
2683 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2684 | } |
2685 | ||
235c7fc7 | 2686 | static void |
5b0311e1 | 2687 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2688 | struct perf_cpu_context *cpuctx) |
0793a61d | 2689 | { |
cdd6c482 | 2690 | struct perf_event *event; |
0793a61d | 2691 | |
889ff015 FW |
2692 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2693 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2694 | continue; |
5632ab12 | 2695 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2696 | continue; |
2697 | ||
e5d1367f SE |
2698 | /* may need to reset tstamp_enabled */ |
2699 | if (is_cgroup_event(event)) | |
2700 | perf_cgroup_mark_enabled(event, ctx); | |
2701 | ||
8c9ed8e1 | 2702 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2703 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2704 | |
2705 | /* | |
2706 | * If this pinned group hasn't been scheduled, | |
2707 | * put it in error state. | |
2708 | */ | |
cdd6c482 IM |
2709 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2710 | update_group_times(event); | |
2711 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2712 | } |
3b6f9e5c | 2713 | } |
5b0311e1 FW |
2714 | } |
2715 | ||
2716 | static void | |
2717 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2718 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2719 | { |
2720 | struct perf_event *event; | |
2721 | int can_add_hw = 1; | |
3b6f9e5c | 2722 | |
889ff015 FW |
2723 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2724 | /* Ignore events in OFF or ERROR state */ | |
2725 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2726 | continue; |
04289bb9 IM |
2727 | /* |
2728 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2729 | * of events: |
04289bb9 | 2730 | */ |
5632ab12 | 2731 | if (!event_filter_match(event)) |
0793a61d TG |
2732 | continue; |
2733 | ||
e5d1367f SE |
2734 | /* may need to reset tstamp_enabled */ |
2735 | if (is_cgroup_event(event)) | |
2736 | perf_cgroup_mark_enabled(event, ctx); | |
2737 | ||
9ed6060d | 2738 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2739 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2740 | can_add_hw = 0; |
9ed6060d | 2741 | } |
0793a61d | 2742 | } |
5b0311e1 FW |
2743 | } |
2744 | ||
2745 | static void | |
2746 | ctx_sched_in(struct perf_event_context *ctx, | |
2747 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2748 | enum event_type_t event_type, |
2749 | struct task_struct *task) | |
5b0311e1 | 2750 | { |
e5d1367f | 2751 | u64 now; |
db24d33e | 2752 | int is_active = ctx->is_active; |
e5d1367f | 2753 | |
db24d33e | 2754 | ctx->is_active |= event_type; |
5b0311e1 | 2755 | if (likely(!ctx->nr_events)) |
facc4307 | 2756 | return; |
5b0311e1 | 2757 | |
e5d1367f SE |
2758 | now = perf_clock(); |
2759 | ctx->timestamp = now; | |
3f7cce3c | 2760 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2761 | /* |
2762 | * First go through the list and put on any pinned groups | |
2763 | * in order to give them the best chance of going on. | |
2764 | */ | |
db24d33e | 2765 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2766 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2767 | |
2768 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2769 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2770 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2771 | } |
2772 | ||
329c0e01 | 2773 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2774 | enum event_type_t event_type, |
2775 | struct task_struct *task) | |
329c0e01 FW |
2776 | { |
2777 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2778 | ||
e5d1367f | 2779 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2780 | } |
2781 | ||
e5d1367f SE |
2782 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2783 | struct task_struct *task) | |
235c7fc7 | 2784 | { |
108b02cf | 2785 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2786 | |
108b02cf | 2787 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2788 | if (cpuctx->task_ctx == ctx) |
2789 | return; | |
2790 | ||
facc4307 | 2791 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2792 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2793 | /* |
2794 | * We want to keep the following priority order: | |
2795 | * cpu pinned (that don't need to move), task pinned, | |
2796 | * cpu flexible, task flexible. | |
2797 | */ | |
2798 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2799 | ||
1d5f003f GN |
2800 | if (ctx->nr_events) |
2801 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2802 | |
86b47c25 GN |
2803 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2804 | ||
facc4307 PZ |
2805 | perf_pmu_enable(ctx->pmu); |
2806 | perf_ctx_unlock(cpuctx, ctx); | |
235c7fc7 IM |
2807 | } |
2808 | ||
8dc85d54 PZ |
2809 | /* |
2810 | * Called from scheduler to add the events of the current task | |
2811 | * with interrupts disabled. | |
2812 | * | |
2813 | * We restore the event value and then enable it. | |
2814 | * | |
2815 | * This does not protect us against NMI, but enable() | |
2816 | * sets the enabled bit in the control field of event _before_ | |
2817 | * accessing the event control register. If a NMI hits, then it will | |
2818 | * keep the event running. | |
2819 | */ | |
ab0cce56 JO |
2820 | void __perf_event_task_sched_in(struct task_struct *prev, |
2821 | struct task_struct *task) | |
8dc85d54 PZ |
2822 | { |
2823 | struct perf_event_context *ctx; | |
2824 | int ctxn; | |
2825 | ||
2826 | for_each_task_context_nr(ctxn) { | |
2827 | ctx = task->perf_event_ctxp[ctxn]; | |
2828 | if (likely(!ctx)) | |
2829 | continue; | |
2830 | ||
e5d1367f | 2831 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2832 | } |
e5d1367f SE |
2833 | /* |
2834 | * if cgroup events exist on this CPU, then we need | |
2835 | * to check if we have to switch in PMU state. | |
2836 | * cgroup event are system-wide mode only | |
2837 | */ | |
4a32fea9 | 2838 | if (atomic_read(this_cpu_ptr(&perf_cgroup_events))) |
a8d757ef | 2839 | perf_cgroup_sched_in(prev, task); |
d010b332 | 2840 | |
45ac1403 AH |
2841 | if (atomic_read(&nr_switch_events)) |
2842 | perf_event_switch(task, prev, true); | |
2843 | ||
ba532500 YZ |
2844 | if (__this_cpu_read(perf_sched_cb_usages)) |
2845 | perf_pmu_sched_task(prev, task, true); | |
235c7fc7 IM |
2846 | } |
2847 | ||
abd50713 PZ |
2848 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2849 | { | |
2850 | u64 frequency = event->attr.sample_freq; | |
2851 | u64 sec = NSEC_PER_SEC; | |
2852 | u64 divisor, dividend; | |
2853 | ||
2854 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2855 | ||
2856 | count_fls = fls64(count); | |
2857 | nsec_fls = fls64(nsec); | |
2858 | frequency_fls = fls64(frequency); | |
2859 | sec_fls = 30; | |
2860 | ||
2861 | /* | |
2862 | * We got @count in @nsec, with a target of sample_freq HZ | |
2863 | * the target period becomes: | |
2864 | * | |
2865 | * @count * 10^9 | |
2866 | * period = ------------------- | |
2867 | * @nsec * sample_freq | |
2868 | * | |
2869 | */ | |
2870 | ||
2871 | /* | |
2872 | * Reduce accuracy by one bit such that @a and @b converge | |
2873 | * to a similar magnitude. | |
2874 | */ | |
fe4b04fa | 2875 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2876 | do { \ |
2877 | if (a##_fls > b##_fls) { \ | |
2878 | a >>= 1; \ | |
2879 | a##_fls--; \ | |
2880 | } else { \ | |
2881 | b >>= 1; \ | |
2882 | b##_fls--; \ | |
2883 | } \ | |
2884 | } while (0) | |
2885 | ||
2886 | /* | |
2887 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2888 | * the other, so that finally we can do a u64/u64 division. | |
2889 | */ | |
2890 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2891 | REDUCE_FLS(nsec, frequency); | |
2892 | REDUCE_FLS(sec, count); | |
2893 | } | |
2894 | ||
2895 | if (count_fls + sec_fls > 64) { | |
2896 | divisor = nsec * frequency; | |
2897 | ||
2898 | while (count_fls + sec_fls > 64) { | |
2899 | REDUCE_FLS(count, sec); | |
2900 | divisor >>= 1; | |
2901 | } | |
2902 | ||
2903 | dividend = count * sec; | |
2904 | } else { | |
2905 | dividend = count * sec; | |
2906 | ||
2907 | while (nsec_fls + frequency_fls > 64) { | |
2908 | REDUCE_FLS(nsec, frequency); | |
2909 | dividend >>= 1; | |
2910 | } | |
2911 | ||
2912 | divisor = nsec * frequency; | |
2913 | } | |
2914 | ||
f6ab91ad PZ |
2915 | if (!divisor) |
2916 | return dividend; | |
2917 | ||
abd50713 PZ |
2918 | return div64_u64(dividend, divisor); |
2919 | } | |
2920 | ||
e050e3f0 SE |
2921 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2922 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2923 | ||
f39d47ff | 2924 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2925 | { |
cdd6c482 | 2926 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2927 | s64 period, sample_period; |
bd2b5b12 PZ |
2928 | s64 delta; |
2929 | ||
abd50713 | 2930 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2931 | |
2932 | delta = (s64)(period - hwc->sample_period); | |
2933 | delta = (delta + 7) / 8; /* low pass filter */ | |
2934 | ||
2935 | sample_period = hwc->sample_period + delta; | |
2936 | ||
2937 | if (!sample_period) | |
2938 | sample_period = 1; | |
2939 | ||
bd2b5b12 | 2940 | hwc->sample_period = sample_period; |
abd50713 | 2941 | |
e7850595 | 2942 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2943 | if (disable) |
2944 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2945 | ||
e7850595 | 2946 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2947 | |
2948 | if (disable) | |
2949 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2950 | } |
bd2b5b12 PZ |
2951 | } |
2952 | ||
e050e3f0 SE |
2953 | /* |
2954 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2955 | * events. At the same time, make sure, having freq events does not change | |
2956 | * the rate of unthrottling as that would introduce bias. | |
2957 | */ | |
2958 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2959 | int needs_unthr) | |
60db5e09 | 2960 | { |
cdd6c482 IM |
2961 | struct perf_event *event; |
2962 | struct hw_perf_event *hwc; | |
e050e3f0 | 2963 | u64 now, period = TICK_NSEC; |
abd50713 | 2964 | s64 delta; |
60db5e09 | 2965 | |
e050e3f0 SE |
2966 | /* |
2967 | * only need to iterate over all events iff: | |
2968 | * - context have events in frequency mode (needs freq adjust) | |
2969 | * - there are events to unthrottle on this cpu | |
2970 | */ | |
2971 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2972 | return; |
2973 | ||
e050e3f0 | 2974 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2975 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2976 | |
03541f8b | 2977 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2978 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2979 | continue; |
2980 | ||
5632ab12 | 2981 | if (!event_filter_match(event)) |
5d27c23d PZ |
2982 | continue; |
2983 | ||
44377277 AS |
2984 | perf_pmu_disable(event->pmu); |
2985 | ||
cdd6c482 | 2986 | hwc = &event->hw; |
6a24ed6c | 2987 | |
ae23bff1 | 2988 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2989 | hwc->interrupts = 0; |
cdd6c482 | 2990 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2991 | event->pmu->start(event, 0); |
a78ac325 PZ |
2992 | } |
2993 | ||
cdd6c482 | 2994 | if (!event->attr.freq || !event->attr.sample_freq) |
44377277 | 2995 | goto next; |
60db5e09 | 2996 | |
e050e3f0 SE |
2997 | /* |
2998 | * stop the event and update event->count | |
2999 | */ | |
3000 | event->pmu->stop(event, PERF_EF_UPDATE); | |
3001 | ||
e7850595 | 3002 | now = local64_read(&event->count); |
abd50713 PZ |
3003 | delta = now - hwc->freq_count_stamp; |
3004 | hwc->freq_count_stamp = now; | |
60db5e09 | 3005 | |
e050e3f0 SE |
3006 | /* |
3007 | * restart the event | |
3008 | * reload only if value has changed | |
f39d47ff SE |
3009 | * we have stopped the event so tell that |
3010 | * to perf_adjust_period() to avoid stopping it | |
3011 | * twice. | |
e050e3f0 | 3012 | */ |
abd50713 | 3013 | if (delta > 0) |
f39d47ff | 3014 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
3015 | |
3016 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
44377277 AS |
3017 | next: |
3018 | perf_pmu_enable(event->pmu); | |
60db5e09 | 3019 | } |
e050e3f0 | 3020 | |
f39d47ff | 3021 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 3022 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
3023 | } |
3024 | ||
235c7fc7 | 3025 | /* |
cdd6c482 | 3026 | * Round-robin a context's events: |
235c7fc7 | 3027 | */ |
cdd6c482 | 3028 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 3029 | { |
dddd3379 TG |
3030 | /* |
3031 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
3032 | * disabled by the inheritance code. | |
3033 | */ | |
3034 | if (!ctx->rotate_disable) | |
3035 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
3036 | } |
3037 | ||
9e630205 | 3038 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 3039 | { |
8dc85d54 | 3040 | struct perf_event_context *ctx = NULL; |
2fde4f94 | 3041 | int rotate = 0; |
7fc23a53 | 3042 | |
b5ab4cd5 | 3043 | if (cpuctx->ctx.nr_events) { |
b5ab4cd5 PZ |
3044 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
3045 | rotate = 1; | |
3046 | } | |
235c7fc7 | 3047 | |
8dc85d54 | 3048 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 3049 | if (ctx && ctx->nr_events) { |
b5ab4cd5 PZ |
3050 | if (ctx->nr_events != ctx->nr_active) |
3051 | rotate = 1; | |
3052 | } | |
9717e6cd | 3053 | |
e050e3f0 | 3054 | if (!rotate) |
0f5a2601 PZ |
3055 | goto done; |
3056 | ||
facc4307 | 3057 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 3058 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 3059 | |
e050e3f0 SE |
3060 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
3061 | if (ctx) | |
3062 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 3063 | |
e050e3f0 SE |
3064 | rotate_ctx(&cpuctx->ctx); |
3065 | if (ctx) | |
3066 | rotate_ctx(ctx); | |
235c7fc7 | 3067 | |
e050e3f0 | 3068 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 3069 | |
0f5a2601 PZ |
3070 | perf_pmu_enable(cpuctx->ctx.pmu); |
3071 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 3072 | done: |
9e630205 SE |
3073 | |
3074 | return rotate; | |
e9d2b064 PZ |
3075 | } |
3076 | ||
026249ef FW |
3077 | #ifdef CONFIG_NO_HZ_FULL |
3078 | bool perf_event_can_stop_tick(void) | |
3079 | { | |
948b26b6 | 3080 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 3081 | __this_cpu_read(perf_throttled_count)) |
026249ef | 3082 | return false; |
d84153d6 FW |
3083 | else |
3084 | return true; | |
026249ef FW |
3085 | } |
3086 | #endif | |
3087 | ||
e9d2b064 PZ |
3088 | void perf_event_task_tick(void) |
3089 | { | |
2fde4f94 MR |
3090 | struct list_head *head = this_cpu_ptr(&active_ctx_list); |
3091 | struct perf_event_context *ctx, *tmp; | |
e050e3f0 | 3092 | int throttled; |
b5ab4cd5 | 3093 | |
e9d2b064 PZ |
3094 | WARN_ON(!irqs_disabled()); |
3095 | ||
e050e3f0 SE |
3096 | __this_cpu_inc(perf_throttled_seq); |
3097 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
3098 | ||
2fde4f94 | 3099 | list_for_each_entry_safe(ctx, tmp, head, active_ctx_list) |
e050e3f0 | 3100 | perf_adjust_freq_unthr_context(ctx, throttled); |
0793a61d TG |
3101 | } |
3102 | ||
889ff015 FW |
3103 | static int event_enable_on_exec(struct perf_event *event, |
3104 | struct perf_event_context *ctx) | |
3105 | { | |
3106 | if (!event->attr.enable_on_exec) | |
3107 | return 0; | |
3108 | ||
3109 | event->attr.enable_on_exec = 0; | |
3110 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
3111 | return 0; | |
3112 | ||
1d9b482e | 3113 | __perf_event_mark_enabled(event); |
889ff015 FW |
3114 | |
3115 | return 1; | |
3116 | } | |
3117 | ||
57e7986e | 3118 | /* |
cdd6c482 | 3119 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
3120 | * This expects task == current. |
3121 | */ | |
8dc85d54 | 3122 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 3123 | { |
211de6eb | 3124 | struct perf_event_context *clone_ctx = NULL; |
cdd6c482 | 3125 | struct perf_event *event; |
57e7986e PM |
3126 | unsigned long flags; |
3127 | int enabled = 0; | |
889ff015 | 3128 | int ret; |
57e7986e PM |
3129 | |
3130 | local_irq_save(flags); | |
cdd6c482 | 3131 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
3132 | goto out; |
3133 | ||
e566b76e SE |
3134 | /* |
3135 | * We must ctxsw out cgroup events to avoid conflict | |
3136 | * when invoking perf_task_event_sched_in() later on | |
3137 | * in this function. Otherwise we end up trying to | |
3138 | * ctxswin cgroup events which are already scheduled | |
3139 | * in. | |
3140 | */ | |
a8d757ef | 3141 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 3142 | |
e625cce1 | 3143 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 3144 | task_ctx_sched_out(ctx); |
57e7986e | 3145 | |
b79387ef | 3146 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
3147 | ret = event_enable_on_exec(event, ctx); |
3148 | if (ret) | |
3149 | enabled = 1; | |
57e7986e PM |
3150 | } |
3151 | ||
3152 | /* | |
cdd6c482 | 3153 | * Unclone this context if we enabled any event. |
57e7986e | 3154 | */ |
71a851b4 | 3155 | if (enabled) |
211de6eb | 3156 | clone_ctx = unclone_ctx(ctx); |
57e7986e | 3157 | |
e625cce1 | 3158 | raw_spin_unlock(&ctx->lock); |
57e7986e | 3159 | |
e566b76e SE |
3160 | /* |
3161 | * Also calls ctxswin for cgroup events, if any: | |
3162 | */ | |
e5d1367f | 3163 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 3164 | out: |
57e7986e | 3165 | local_irq_restore(flags); |
211de6eb PZ |
3166 | |
3167 | if (clone_ctx) | |
3168 | put_ctx(clone_ctx); | |
57e7986e PM |
3169 | } |
3170 | ||
e041e328 PZ |
3171 | void perf_event_exec(void) |
3172 | { | |
3173 | struct perf_event_context *ctx; | |
3174 | int ctxn; | |
3175 | ||
3176 | rcu_read_lock(); | |
3177 | for_each_task_context_nr(ctxn) { | |
3178 | ctx = current->perf_event_ctxp[ctxn]; | |
3179 | if (!ctx) | |
3180 | continue; | |
3181 | ||
3182 | perf_event_enable_on_exec(ctx); | |
3183 | } | |
3184 | rcu_read_unlock(); | |
3185 | } | |
3186 | ||
0793a61d | 3187 | /* |
cdd6c482 | 3188 | * Cross CPU call to read the hardware event |
0793a61d | 3189 | */ |
cdd6c482 | 3190 | static void __perf_event_read(void *info) |
0793a61d | 3191 | { |
cdd6c482 IM |
3192 | struct perf_event *event = info; |
3193 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 3194 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 3195 | |
e1ac3614 PM |
3196 | /* |
3197 | * If this is a task context, we need to check whether it is | |
3198 | * the current task context of this cpu. If not it has been | |
3199 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
3200 | * event->count would have been updated to a recent sample |
3201 | * when the event was scheduled out. | |
e1ac3614 PM |
3202 | */ |
3203 | if (ctx->task && cpuctx->task_ctx != ctx) | |
3204 | return; | |
3205 | ||
e625cce1 | 3206 | raw_spin_lock(&ctx->lock); |
e5d1367f | 3207 | if (ctx->is_active) { |
542e72fc | 3208 | update_context_time(ctx); |
e5d1367f SE |
3209 | update_cgrp_time_from_event(event); |
3210 | } | |
cdd6c482 | 3211 | update_event_times(event); |
542e72fc PZ |
3212 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
3213 | event->pmu->read(event); | |
e625cce1 | 3214 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
3215 | } |
3216 | ||
b5e58793 PZ |
3217 | static inline u64 perf_event_count(struct perf_event *event) |
3218 | { | |
eacd3ecc MF |
3219 | if (event->pmu->count) |
3220 | return event->pmu->count(event); | |
3221 | ||
3222 | return __perf_event_count(event); | |
b5e58793 PZ |
3223 | } |
3224 | ||
cdd6c482 | 3225 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
3226 | { |
3227 | /* | |
cdd6c482 IM |
3228 | * If event is enabled and currently active on a CPU, update the |
3229 | * value in the event structure: | |
0793a61d | 3230 | */ |
cdd6c482 IM |
3231 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
3232 | smp_call_function_single(event->oncpu, | |
3233 | __perf_event_read, event, 1); | |
3234 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
3235 | struct perf_event_context *ctx = event->ctx; |
3236 | unsigned long flags; | |
3237 | ||
e625cce1 | 3238 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
3239 | /* |
3240 | * may read while context is not active | |
3241 | * (e.g., thread is blocked), in that case | |
3242 | * we cannot update context time | |
3243 | */ | |
e5d1367f | 3244 | if (ctx->is_active) { |
c530ccd9 | 3245 | update_context_time(ctx); |
e5d1367f SE |
3246 | update_cgrp_time_from_event(event); |
3247 | } | |
cdd6c482 | 3248 | update_event_times(event); |
e625cce1 | 3249 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
3250 | } |
3251 | ||
b5e58793 | 3252 | return perf_event_count(event); |
0793a61d TG |
3253 | } |
3254 | ||
a63eaf34 | 3255 | /* |
cdd6c482 | 3256 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 3257 | */ |
eb184479 | 3258 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 3259 | { |
e625cce1 | 3260 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 3261 | mutex_init(&ctx->mutex); |
2fde4f94 | 3262 | INIT_LIST_HEAD(&ctx->active_ctx_list); |
889ff015 FW |
3263 | INIT_LIST_HEAD(&ctx->pinned_groups); |
3264 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
3265 | INIT_LIST_HEAD(&ctx->event_list); |
3266 | atomic_set(&ctx->refcount, 1); | |
fadfe7be | 3267 | INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work); |
eb184479 PZ |
3268 | } |
3269 | ||
3270 | static struct perf_event_context * | |
3271 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
3272 | { | |
3273 | struct perf_event_context *ctx; | |
3274 | ||
3275 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
3276 | if (!ctx) | |
3277 | return NULL; | |
3278 | ||
3279 | __perf_event_init_context(ctx); | |
3280 | if (task) { | |
3281 | ctx->task = task; | |
3282 | get_task_struct(task); | |
0793a61d | 3283 | } |
eb184479 PZ |
3284 | ctx->pmu = pmu; |
3285 | ||
3286 | return ctx; | |
a63eaf34 PM |
3287 | } |
3288 | ||
2ebd4ffb MH |
3289 | static struct task_struct * |
3290 | find_lively_task_by_vpid(pid_t vpid) | |
3291 | { | |
3292 | struct task_struct *task; | |
3293 | int err; | |
0793a61d TG |
3294 | |
3295 | rcu_read_lock(); | |
2ebd4ffb | 3296 | if (!vpid) |
0793a61d TG |
3297 | task = current; |
3298 | else | |
2ebd4ffb | 3299 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3300 | if (task) |
3301 | get_task_struct(task); | |
3302 | rcu_read_unlock(); | |
3303 | ||
3304 | if (!task) | |
3305 | return ERR_PTR(-ESRCH); | |
3306 | ||
0793a61d | 3307 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3308 | err = -EACCES; |
3309 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3310 | goto errout; | |
3311 | ||
2ebd4ffb MH |
3312 | return task; |
3313 | errout: | |
3314 | put_task_struct(task); | |
3315 | return ERR_PTR(err); | |
3316 | ||
3317 | } | |
3318 | ||
fe4b04fa PZ |
3319 | /* |
3320 | * Returns a matching context with refcount and pincount. | |
3321 | */ | |
108b02cf | 3322 | static struct perf_event_context * |
4af57ef2 YZ |
3323 | find_get_context(struct pmu *pmu, struct task_struct *task, |
3324 | struct perf_event *event) | |
0793a61d | 3325 | { |
211de6eb | 3326 | struct perf_event_context *ctx, *clone_ctx = NULL; |
22a4f650 | 3327 | struct perf_cpu_context *cpuctx; |
4af57ef2 | 3328 | void *task_ctx_data = NULL; |
25346b93 | 3329 | unsigned long flags; |
8dc85d54 | 3330 | int ctxn, err; |
4af57ef2 | 3331 | int cpu = event->cpu; |
0793a61d | 3332 | |
22a4ec72 | 3333 | if (!task) { |
cdd6c482 | 3334 | /* Must be root to operate on a CPU event: */ |
0764771d | 3335 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3336 | return ERR_PTR(-EACCES); |
3337 | ||
0793a61d | 3338 | /* |
cdd6c482 | 3339 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3340 | * offline CPU and activate it when the CPU comes up, but |
3341 | * that's for later. | |
3342 | */ | |
f6325e30 | 3343 | if (!cpu_online(cpu)) |
0793a61d TG |
3344 | return ERR_PTR(-ENODEV); |
3345 | ||
108b02cf | 3346 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3347 | ctx = &cpuctx->ctx; |
c93f7669 | 3348 | get_ctx(ctx); |
fe4b04fa | 3349 | ++ctx->pin_count; |
0793a61d | 3350 | |
0793a61d TG |
3351 | return ctx; |
3352 | } | |
3353 | ||
8dc85d54 PZ |
3354 | err = -EINVAL; |
3355 | ctxn = pmu->task_ctx_nr; | |
3356 | if (ctxn < 0) | |
3357 | goto errout; | |
3358 | ||
4af57ef2 YZ |
3359 | if (event->attach_state & PERF_ATTACH_TASK_DATA) { |
3360 | task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL); | |
3361 | if (!task_ctx_data) { | |
3362 | err = -ENOMEM; | |
3363 | goto errout; | |
3364 | } | |
3365 | } | |
3366 | ||
9ed6060d | 3367 | retry: |
8dc85d54 | 3368 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3369 | if (ctx) { |
211de6eb | 3370 | clone_ctx = unclone_ctx(ctx); |
fe4b04fa | 3371 | ++ctx->pin_count; |
4af57ef2 YZ |
3372 | |
3373 | if (task_ctx_data && !ctx->task_ctx_data) { | |
3374 | ctx->task_ctx_data = task_ctx_data; | |
3375 | task_ctx_data = NULL; | |
3376 | } | |
e625cce1 | 3377 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
211de6eb PZ |
3378 | |
3379 | if (clone_ctx) | |
3380 | put_ctx(clone_ctx); | |
9137fb28 | 3381 | } else { |
eb184479 | 3382 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3383 | err = -ENOMEM; |
3384 | if (!ctx) | |
3385 | goto errout; | |
eb184479 | 3386 | |
4af57ef2 YZ |
3387 | if (task_ctx_data) { |
3388 | ctx->task_ctx_data = task_ctx_data; | |
3389 | task_ctx_data = NULL; | |
3390 | } | |
3391 | ||
dbe08d82 ON |
3392 | err = 0; |
3393 | mutex_lock(&task->perf_event_mutex); | |
3394 | /* | |
3395 | * If it has already passed perf_event_exit_task(). | |
3396 | * we must see PF_EXITING, it takes this mutex too. | |
3397 | */ | |
3398 | if (task->flags & PF_EXITING) | |
3399 | err = -ESRCH; | |
3400 | else if (task->perf_event_ctxp[ctxn]) | |
3401 | err = -EAGAIN; | |
fe4b04fa | 3402 | else { |
9137fb28 | 3403 | get_ctx(ctx); |
fe4b04fa | 3404 | ++ctx->pin_count; |
dbe08d82 | 3405 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3406 | } |
dbe08d82 ON |
3407 | mutex_unlock(&task->perf_event_mutex); |
3408 | ||
3409 | if (unlikely(err)) { | |
9137fb28 | 3410 | put_ctx(ctx); |
dbe08d82 ON |
3411 | |
3412 | if (err == -EAGAIN) | |
3413 | goto retry; | |
3414 | goto errout; | |
a63eaf34 PM |
3415 | } |
3416 | } | |
3417 | ||
4af57ef2 | 3418 | kfree(task_ctx_data); |
0793a61d | 3419 | return ctx; |
c93f7669 | 3420 | |
9ed6060d | 3421 | errout: |
4af57ef2 | 3422 | kfree(task_ctx_data); |
c93f7669 | 3423 | return ERR_PTR(err); |
0793a61d TG |
3424 | } |
3425 | ||
6fb2915d | 3426 | static void perf_event_free_filter(struct perf_event *event); |
2541517c | 3427 | static void perf_event_free_bpf_prog(struct perf_event *event); |
6fb2915d | 3428 | |
cdd6c482 | 3429 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3430 | { |
cdd6c482 | 3431 | struct perf_event *event; |
592903cd | 3432 | |
cdd6c482 IM |
3433 | event = container_of(head, struct perf_event, rcu_head); |
3434 | if (event->ns) | |
3435 | put_pid_ns(event->ns); | |
6fb2915d | 3436 | perf_event_free_filter(event); |
cdd6c482 | 3437 | kfree(event); |
592903cd PZ |
3438 | } |
3439 | ||
b69cf536 PZ |
3440 | static void ring_buffer_attach(struct perf_event *event, |
3441 | struct ring_buffer *rb); | |
925d519a | 3442 | |
4beb31f3 | 3443 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3444 | { |
4beb31f3 FW |
3445 | if (event->parent) |
3446 | return; | |
3447 | ||
4beb31f3 FW |
3448 | if (is_cgroup_event(event)) |
3449 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3450 | } | |
925d519a | 3451 | |
4beb31f3 FW |
3452 | static void unaccount_event(struct perf_event *event) |
3453 | { | |
3454 | if (event->parent) | |
3455 | return; | |
3456 | ||
3457 | if (event->attach_state & PERF_ATTACH_TASK) | |
3458 | static_key_slow_dec_deferred(&perf_sched_events); | |
3459 | if (event->attr.mmap || event->attr.mmap_data) | |
3460 | atomic_dec(&nr_mmap_events); | |
3461 | if (event->attr.comm) | |
3462 | atomic_dec(&nr_comm_events); | |
3463 | if (event->attr.task) | |
3464 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3465 | if (event->attr.freq) |
3466 | atomic_dec(&nr_freq_events); | |
45ac1403 AH |
3467 | if (event->attr.context_switch) { |
3468 | static_key_slow_dec_deferred(&perf_sched_events); | |
3469 | atomic_dec(&nr_switch_events); | |
3470 | } | |
4beb31f3 FW |
3471 | if (is_cgroup_event(event)) |
3472 | static_key_slow_dec_deferred(&perf_sched_events); | |
3473 | if (has_branch_stack(event)) | |
3474 | static_key_slow_dec_deferred(&perf_sched_events); | |
3475 | ||
3476 | unaccount_event_cpu(event, event->cpu); | |
3477 | } | |
925d519a | 3478 | |
bed5b25a AS |
3479 | /* |
3480 | * The following implement mutual exclusion of events on "exclusive" pmus | |
3481 | * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled | |
3482 | * at a time, so we disallow creating events that might conflict, namely: | |
3483 | * | |
3484 | * 1) cpu-wide events in the presence of per-task events, | |
3485 | * 2) per-task events in the presence of cpu-wide events, | |
3486 | * 3) two matching events on the same context. | |
3487 | * | |
3488 | * The former two cases are handled in the allocation path (perf_event_alloc(), | |
3489 | * __free_event()), the latter -- before the first perf_install_in_context(). | |
3490 | */ | |
3491 | static int exclusive_event_init(struct perf_event *event) | |
3492 | { | |
3493 | struct pmu *pmu = event->pmu; | |
3494 | ||
3495 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) | |
3496 | return 0; | |
3497 | ||
3498 | /* | |
3499 | * Prevent co-existence of per-task and cpu-wide events on the | |
3500 | * same exclusive pmu. | |
3501 | * | |
3502 | * Negative pmu::exclusive_cnt means there are cpu-wide | |
3503 | * events on this "exclusive" pmu, positive means there are | |
3504 | * per-task events. | |
3505 | * | |
3506 | * Since this is called in perf_event_alloc() path, event::ctx | |
3507 | * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK | |
3508 | * to mean "per-task event", because unlike other attach states it | |
3509 | * never gets cleared. | |
3510 | */ | |
3511 | if (event->attach_state & PERF_ATTACH_TASK) { | |
3512 | if (!atomic_inc_unless_negative(&pmu->exclusive_cnt)) | |
3513 | return -EBUSY; | |
3514 | } else { | |
3515 | if (!atomic_dec_unless_positive(&pmu->exclusive_cnt)) | |
3516 | return -EBUSY; | |
3517 | } | |
3518 | ||
3519 | return 0; | |
3520 | } | |
3521 | ||
3522 | static void exclusive_event_destroy(struct perf_event *event) | |
3523 | { | |
3524 | struct pmu *pmu = event->pmu; | |
3525 | ||
3526 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) | |
3527 | return; | |
3528 | ||
3529 | /* see comment in exclusive_event_init() */ | |
3530 | if (event->attach_state & PERF_ATTACH_TASK) | |
3531 | atomic_dec(&pmu->exclusive_cnt); | |
3532 | else | |
3533 | atomic_inc(&pmu->exclusive_cnt); | |
3534 | } | |
3535 | ||
3536 | static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2) | |
3537 | { | |
3538 | if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && | |
3539 | (e1->cpu == e2->cpu || | |
3540 | e1->cpu == -1 || | |
3541 | e2->cpu == -1)) | |
3542 | return true; | |
3543 | return false; | |
3544 | } | |
3545 | ||
3546 | /* Called under the same ctx::mutex as perf_install_in_context() */ | |
3547 | static bool exclusive_event_installable(struct perf_event *event, | |
3548 | struct perf_event_context *ctx) | |
3549 | { | |
3550 | struct perf_event *iter_event; | |
3551 | struct pmu *pmu = event->pmu; | |
3552 | ||
3553 | if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE)) | |
3554 | return true; | |
3555 | ||
3556 | list_for_each_entry(iter_event, &ctx->event_list, event_entry) { | |
3557 | if (exclusive_event_match(iter_event, event)) | |
3558 | return false; | |
3559 | } | |
3560 | ||
3561 | return true; | |
3562 | } | |
3563 | ||
766d6c07 FW |
3564 | static void __free_event(struct perf_event *event) |
3565 | { | |
cdd6c482 | 3566 | if (!event->parent) { |
927c7a9e FW |
3567 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3568 | put_callchain_buffers(); | |
f344011c | 3569 | } |
9ee318a7 | 3570 | |
dead9f29 AS |
3571 | perf_event_free_bpf_prog(event); |
3572 | ||
766d6c07 FW |
3573 | if (event->destroy) |
3574 | event->destroy(event); | |
3575 | ||
3576 | if (event->ctx) | |
3577 | put_ctx(event->ctx); | |
3578 | ||
bed5b25a AS |
3579 | if (event->pmu) { |
3580 | exclusive_event_destroy(event); | |
c464c76e | 3581 | module_put(event->pmu->module); |
bed5b25a | 3582 | } |
c464c76e | 3583 | |
766d6c07 FW |
3584 | call_rcu(&event->rcu_head, free_event_rcu); |
3585 | } | |
683ede43 PZ |
3586 | |
3587 | static void _free_event(struct perf_event *event) | |
f1600952 | 3588 | { |
e360adbe | 3589 | irq_work_sync(&event->pending); |
925d519a | 3590 | |
4beb31f3 | 3591 | unaccount_event(event); |
9ee318a7 | 3592 | |
76369139 | 3593 | if (event->rb) { |
9bb5d40c PZ |
3594 | /* |
3595 | * Can happen when we close an event with re-directed output. | |
3596 | * | |
3597 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3598 | * over us; possibly making our ring_buffer_put() the last. | |
3599 | */ | |
3600 | mutex_lock(&event->mmap_mutex); | |
b69cf536 | 3601 | ring_buffer_attach(event, NULL); |
9bb5d40c | 3602 | mutex_unlock(&event->mmap_mutex); |
a4be7c27 PZ |
3603 | } |
3604 | ||
e5d1367f SE |
3605 | if (is_cgroup_event(event)) |
3606 | perf_detach_cgroup(event); | |
3607 | ||
766d6c07 | 3608 | __free_event(event); |
f1600952 PZ |
3609 | } |
3610 | ||
683ede43 PZ |
3611 | /* |
3612 | * Used to free events which have a known refcount of 1, such as in error paths | |
3613 | * where the event isn't exposed yet and inherited events. | |
3614 | */ | |
3615 | static void free_event(struct perf_event *event) | |
0793a61d | 3616 | { |
683ede43 PZ |
3617 | if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1, |
3618 | "unexpected event refcount: %ld; ptr=%p\n", | |
3619 | atomic_long_read(&event->refcount), event)) { | |
3620 | /* leak to avoid use-after-free */ | |
3621 | return; | |
3622 | } | |
0793a61d | 3623 | |
683ede43 | 3624 | _free_event(event); |
0793a61d TG |
3625 | } |
3626 | ||
a66a3052 | 3627 | /* |
f8697762 | 3628 | * Remove user event from the owner task. |
a66a3052 | 3629 | */ |
f8697762 | 3630 | static void perf_remove_from_owner(struct perf_event *event) |
fb0459d7 | 3631 | { |
8882135b | 3632 | struct task_struct *owner; |
fb0459d7 | 3633 | |
8882135b PZ |
3634 | rcu_read_lock(); |
3635 | owner = ACCESS_ONCE(event->owner); | |
3636 | /* | |
3637 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3638 | * !owner it means the list deletion is complete and we can indeed | |
3639 | * free this event, otherwise we need to serialize on | |
3640 | * owner->perf_event_mutex. | |
3641 | */ | |
3642 | smp_read_barrier_depends(); | |
3643 | if (owner) { | |
3644 | /* | |
3645 | * Since delayed_put_task_struct() also drops the last | |
3646 | * task reference we can safely take a new reference | |
3647 | * while holding the rcu_read_lock(). | |
3648 | */ | |
3649 | get_task_struct(owner); | |
3650 | } | |
3651 | rcu_read_unlock(); | |
3652 | ||
3653 | if (owner) { | |
f63a8daa PZ |
3654 | /* |
3655 | * If we're here through perf_event_exit_task() we're already | |
3656 | * holding ctx->mutex which would be an inversion wrt. the | |
3657 | * normal lock order. | |
3658 | * | |
3659 | * However we can safely take this lock because its the child | |
3660 | * ctx->mutex. | |
3661 | */ | |
3662 | mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING); | |
3663 | ||
8882135b PZ |
3664 | /* |
3665 | * We have to re-check the event->owner field, if it is cleared | |
3666 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3667 | * ensured they're done, and we can proceed with freeing the | |
3668 | * event. | |
3669 | */ | |
3670 | if (event->owner) | |
3671 | list_del_init(&event->owner_entry); | |
3672 | mutex_unlock(&owner->perf_event_mutex); | |
3673 | put_task_struct(owner); | |
3674 | } | |
f8697762 JO |
3675 | } |
3676 | ||
f8697762 JO |
3677 | static void put_event(struct perf_event *event) |
3678 | { | |
a83fe28e | 3679 | struct perf_event_context *ctx; |
f8697762 JO |
3680 | |
3681 | if (!atomic_long_dec_and_test(&event->refcount)) | |
3682 | return; | |
3683 | ||
3684 | if (!is_kernel_event(event)) | |
3685 | perf_remove_from_owner(event); | |
8882135b | 3686 | |
683ede43 PZ |
3687 | /* |
3688 | * There are two ways this annotation is useful: | |
3689 | * | |
3690 | * 1) there is a lock recursion from perf_event_exit_task | |
3691 | * see the comment there. | |
3692 | * | |
3693 | * 2) there is a lock-inversion with mmap_sem through | |
3694 | * perf_event_read_group(), which takes faults while | |
3695 | * holding ctx->mutex, however this is called after | |
3696 | * the last filedesc died, so there is no possibility | |
3697 | * to trigger the AB-BA case. | |
3698 | */ | |
a83fe28e PZ |
3699 | ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING); |
3700 | WARN_ON_ONCE(ctx->parent_ctx); | |
683ede43 | 3701 | perf_remove_from_context(event, true); |
d415a7f1 | 3702 | perf_event_ctx_unlock(event, ctx); |
683ede43 PZ |
3703 | |
3704 | _free_event(event); | |
a6fa941d AV |
3705 | } |
3706 | ||
683ede43 PZ |
3707 | int perf_event_release_kernel(struct perf_event *event) |
3708 | { | |
3709 | put_event(event); | |
3710 | return 0; | |
3711 | } | |
3712 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); | |
3713 | ||
8b10c5e2 PZ |
3714 | /* |
3715 | * Called when the last reference to the file is gone. | |
3716 | */ | |
a6fa941d AV |
3717 | static int perf_release(struct inode *inode, struct file *file) |
3718 | { | |
3719 | put_event(file->private_data); | |
3720 | return 0; | |
fb0459d7 | 3721 | } |
fb0459d7 | 3722 | |
fadfe7be JO |
3723 | /* |
3724 | * Remove all orphanes events from the context. | |
3725 | */ | |
3726 | static void orphans_remove_work(struct work_struct *work) | |
3727 | { | |
3728 | struct perf_event_context *ctx; | |
3729 | struct perf_event *event, *tmp; | |
3730 | ||
3731 | ctx = container_of(work, struct perf_event_context, | |
3732 | orphans_remove.work); | |
3733 | ||
3734 | mutex_lock(&ctx->mutex); | |
3735 | list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) { | |
3736 | struct perf_event *parent_event = event->parent; | |
3737 | ||
3738 | if (!is_orphaned_child(event)) | |
3739 | continue; | |
3740 | ||
3741 | perf_remove_from_context(event, true); | |
3742 | ||
3743 | mutex_lock(&parent_event->child_mutex); | |
3744 | list_del_init(&event->child_list); | |
3745 | mutex_unlock(&parent_event->child_mutex); | |
3746 | ||
3747 | free_event(event); | |
3748 | put_event(parent_event); | |
3749 | } | |
3750 | ||
3751 | raw_spin_lock_irq(&ctx->lock); | |
3752 | ctx->orphans_remove_sched = false; | |
3753 | raw_spin_unlock_irq(&ctx->lock); | |
3754 | mutex_unlock(&ctx->mutex); | |
3755 | ||
3756 | put_ctx(ctx); | |
3757 | } | |
3758 | ||
59ed446f | 3759 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3760 | { |
cdd6c482 | 3761 | struct perf_event *child; |
e53c0994 PZ |
3762 | u64 total = 0; |
3763 | ||
59ed446f PZ |
3764 | *enabled = 0; |
3765 | *running = 0; | |
3766 | ||
6f10581a | 3767 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3768 | total += perf_event_read(event); |
59ed446f PZ |
3769 | *enabled += event->total_time_enabled + |
3770 | atomic64_read(&event->child_total_time_enabled); | |
3771 | *running += event->total_time_running + | |
3772 | atomic64_read(&event->child_total_time_running); | |
3773 | ||
3774 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3775 | total += perf_event_read(child); |
59ed446f PZ |
3776 | *enabled += child->total_time_enabled; |
3777 | *running += child->total_time_running; | |
3778 | } | |
6f10581a | 3779 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3780 | |
3781 | return total; | |
3782 | } | |
fb0459d7 | 3783 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3784 | |
cdd6c482 | 3785 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3786 | u64 read_format, char __user *buf) |
3787 | { | |
cdd6c482 | 3788 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a | 3789 | struct perf_event_context *ctx = leader->ctx; |
f63a8daa | 3790 | int n = 0, size = 0, ret; |
59ed446f | 3791 | u64 count, enabled, running; |
f63a8daa PZ |
3792 | u64 values[5]; |
3793 | ||
3794 | lockdep_assert_held(&ctx->mutex); | |
abf4868b | 3795 | |
59ed446f | 3796 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3797 | |
3798 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3799 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3800 | values[n++] = enabled; | |
3801 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3802 | values[n++] = running; | |
abf4868b PZ |
3803 | values[n++] = count; |
3804 | if (read_format & PERF_FORMAT_ID) | |
3805 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3806 | |
3807 | size = n * sizeof(u64); | |
3808 | ||
3809 | if (copy_to_user(buf, values, size)) | |
f63a8daa | 3810 | return -EFAULT; |
3dab77fb | 3811 | |
6f10581a | 3812 | ret = size; |
3dab77fb | 3813 | |
65abc865 | 3814 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3815 | n = 0; |
3dab77fb | 3816 | |
59ed446f | 3817 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3818 | if (read_format & PERF_FORMAT_ID) |
3819 | values[n++] = primary_event_id(sub); | |
3820 | ||
3821 | size = n * sizeof(u64); | |
3822 | ||
184d3da8 | 3823 | if (copy_to_user(buf + ret, values, size)) { |
f63a8daa | 3824 | return -EFAULT; |
6f10581a | 3825 | } |
abf4868b PZ |
3826 | |
3827 | ret += size; | |
3dab77fb PZ |
3828 | } |
3829 | ||
abf4868b | 3830 | return ret; |
3dab77fb PZ |
3831 | } |
3832 | ||
cdd6c482 | 3833 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3834 | u64 read_format, char __user *buf) |
3835 | { | |
59ed446f | 3836 | u64 enabled, running; |
3dab77fb PZ |
3837 | u64 values[4]; |
3838 | int n = 0; | |
3839 | ||
59ed446f PZ |
3840 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3841 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3842 | values[n++] = enabled; | |
3843 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3844 | values[n++] = running; | |
3dab77fb | 3845 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3846 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3847 | |
3848 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3849 | return -EFAULT; | |
3850 | ||
3851 | return n * sizeof(u64); | |
3852 | } | |
3853 | ||
dc633982 JO |
3854 | static bool is_event_hup(struct perf_event *event) |
3855 | { | |
3856 | bool no_children; | |
3857 | ||
3858 | if (event->state != PERF_EVENT_STATE_EXIT) | |
3859 | return false; | |
3860 | ||
3861 | mutex_lock(&event->child_mutex); | |
3862 | no_children = list_empty(&event->child_list); | |
3863 | mutex_unlock(&event->child_mutex); | |
3864 | return no_children; | |
3865 | } | |
3866 | ||
0793a61d | 3867 | /* |
cdd6c482 | 3868 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3869 | */ |
3870 | static ssize_t | |
cdd6c482 | 3871 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3872 | { |
cdd6c482 | 3873 | u64 read_format = event->attr.read_format; |
3dab77fb | 3874 | int ret; |
0793a61d | 3875 | |
3b6f9e5c | 3876 | /* |
cdd6c482 | 3877 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3878 | * error state (i.e. because it was pinned but it couldn't be |
3879 | * scheduled on to the CPU at some point). | |
3880 | */ | |
cdd6c482 | 3881 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3882 | return 0; |
3883 | ||
c320c7b7 | 3884 | if (count < event->read_size) |
3dab77fb PZ |
3885 | return -ENOSPC; |
3886 | ||
cdd6c482 | 3887 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3888 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3889 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3890 | else |
cdd6c482 | 3891 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3892 | |
3dab77fb | 3893 | return ret; |
0793a61d TG |
3894 | } |
3895 | ||
0793a61d TG |
3896 | static ssize_t |
3897 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3898 | { | |
cdd6c482 | 3899 | struct perf_event *event = file->private_data; |
f63a8daa PZ |
3900 | struct perf_event_context *ctx; |
3901 | int ret; | |
0793a61d | 3902 | |
f63a8daa PZ |
3903 | ctx = perf_event_ctx_lock(event); |
3904 | ret = perf_read_hw(event, buf, count); | |
3905 | perf_event_ctx_unlock(event, ctx); | |
3906 | ||
3907 | return ret; | |
0793a61d TG |
3908 | } |
3909 | ||
3910 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3911 | { | |
cdd6c482 | 3912 | struct perf_event *event = file->private_data; |
76369139 | 3913 | struct ring_buffer *rb; |
61b67684 | 3914 | unsigned int events = POLLHUP; |
c7138f37 | 3915 | |
e708d7ad | 3916 | poll_wait(file, &event->waitq, wait); |
179033b3 | 3917 | |
dc633982 | 3918 | if (is_event_hup(event)) |
179033b3 | 3919 | return events; |
c7138f37 | 3920 | |
10c6db11 | 3921 | /* |
9bb5d40c PZ |
3922 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3923 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3924 | */ |
3925 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3926 | rb = event->rb; |
3927 | if (rb) | |
76369139 | 3928 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 | 3929 | mutex_unlock(&event->mmap_mutex); |
0793a61d TG |
3930 | return events; |
3931 | } | |
3932 | ||
f63a8daa | 3933 | static void _perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3934 | { |
cdd6c482 | 3935 | (void)perf_event_read(event); |
e7850595 | 3936 | local64_set(&event->count, 0); |
cdd6c482 | 3937 | perf_event_update_userpage(event); |
3df5edad PZ |
3938 | } |
3939 | ||
c93f7669 | 3940 | /* |
cdd6c482 IM |
3941 | * Holding the top-level event's child_mutex means that any |
3942 | * descendant process that has inherited this event will block | |
3943 | * in sync_child_event if it goes to exit, thus satisfying the | |
3944 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3945 | */ |
cdd6c482 IM |
3946 | static void perf_event_for_each_child(struct perf_event *event, |
3947 | void (*func)(struct perf_event *)) | |
3df5edad | 3948 | { |
cdd6c482 | 3949 | struct perf_event *child; |
3df5edad | 3950 | |
cdd6c482 | 3951 | WARN_ON_ONCE(event->ctx->parent_ctx); |
f63a8daa | 3952 | |
cdd6c482 IM |
3953 | mutex_lock(&event->child_mutex); |
3954 | func(event); | |
3955 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3956 | func(child); |
cdd6c482 | 3957 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3958 | } |
3959 | ||
cdd6c482 IM |
3960 | static void perf_event_for_each(struct perf_event *event, |
3961 | void (*func)(struct perf_event *)) | |
3df5edad | 3962 | { |
cdd6c482 IM |
3963 | struct perf_event_context *ctx = event->ctx; |
3964 | struct perf_event *sibling; | |
3df5edad | 3965 | |
f63a8daa PZ |
3966 | lockdep_assert_held(&ctx->mutex); |
3967 | ||
cdd6c482 | 3968 | event = event->group_leader; |
75f937f2 | 3969 | |
cdd6c482 | 3970 | perf_event_for_each_child(event, func); |
cdd6c482 | 3971 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3972 | perf_event_for_each_child(sibling, func); |
6de6a7b9 PZ |
3973 | } |
3974 | ||
cdd6c482 | 3975 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3976 | { |
cdd6c482 | 3977 | struct perf_event_context *ctx = event->ctx; |
bad7192b | 3978 | int ret = 0, active; |
08247e31 PZ |
3979 | u64 value; |
3980 | ||
6c7e550f | 3981 | if (!is_sampling_event(event)) |
08247e31 PZ |
3982 | return -EINVAL; |
3983 | ||
ad0cf347 | 3984 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3985 | return -EFAULT; |
3986 | ||
3987 | if (!value) | |
3988 | return -EINVAL; | |
3989 | ||
e625cce1 | 3990 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3991 | if (event->attr.freq) { |
3992 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3993 | ret = -EINVAL; |
3994 | goto unlock; | |
3995 | } | |
3996 | ||
cdd6c482 | 3997 | event->attr.sample_freq = value; |
08247e31 | 3998 | } else { |
cdd6c482 IM |
3999 | event->attr.sample_period = value; |
4000 | event->hw.sample_period = value; | |
08247e31 | 4001 | } |
bad7192b PZ |
4002 | |
4003 | active = (event->state == PERF_EVENT_STATE_ACTIVE); | |
4004 | if (active) { | |
4005 | perf_pmu_disable(ctx->pmu); | |
4006 | event->pmu->stop(event, PERF_EF_UPDATE); | |
4007 | } | |
4008 | ||
4009 | local64_set(&event->hw.period_left, 0); | |
4010 | ||
4011 | if (active) { | |
4012 | event->pmu->start(event, PERF_EF_RELOAD); | |
4013 | perf_pmu_enable(ctx->pmu); | |
4014 | } | |
4015 | ||
08247e31 | 4016 | unlock: |
e625cce1 | 4017 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
4018 | |
4019 | return ret; | |
4020 | } | |
4021 | ||
ac9721f3 PZ |
4022 | static const struct file_operations perf_fops; |
4023 | ||
2903ff01 | 4024 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 4025 | { |
2903ff01 AV |
4026 | struct fd f = fdget(fd); |
4027 | if (!f.file) | |
4028 | return -EBADF; | |
ac9721f3 | 4029 | |
2903ff01 AV |
4030 | if (f.file->f_op != &perf_fops) { |
4031 | fdput(f); | |
4032 | return -EBADF; | |
ac9721f3 | 4033 | } |
2903ff01 AV |
4034 | *p = f; |
4035 | return 0; | |
ac9721f3 PZ |
4036 | } |
4037 | ||
4038 | static int perf_event_set_output(struct perf_event *event, | |
4039 | struct perf_event *output_event); | |
6fb2915d | 4040 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
2541517c | 4041 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd); |
a4be7c27 | 4042 | |
f63a8daa | 4043 | static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg) |
d859e29f | 4044 | { |
cdd6c482 | 4045 | void (*func)(struct perf_event *); |
3df5edad | 4046 | u32 flags = arg; |
d859e29f PM |
4047 | |
4048 | switch (cmd) { | |
cdd6c482 | 4049 | case PERF_EVENT_IOC_ENABLE: |
f63a8daa | 4050 | func = _perf_event_enable; |
d859e29f | 4051 | break; |
cdd6c482 | 4052 | case PERF_EVENT_IOC_DISABLE: |
f63a8daa | 4053 | func = _perf_event_disable; |
79f14641 | 4054 | break; |
cdd6c482 | 4055 | case PERF_EVENT_IOC_RESET: |
f63a8daa | 4056 | func = _perf_event_reset; |
6de6a7b9 | 4057 | break; |
3df5edad | 4058 | |
cdd6c482 | 4059 | case PERF_EVENT_IOC_REFRESH: |
f63a8daa | 4060 | return _perf_event_refresh(event, arg); |
08247e31 | 4061 | |
cdd6c482 IM |
4062 | case PERF_EVENT_IOC_PERIOD: |
4063 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 4064 | |
cf4957f1 JO |
4065 | case PERF_EVENT_IOC_ID: |
4066 | { | |
4067 | u64 id = primary_event_id(event); | |
4068 | ||
4069 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
4070 | return -EFAULT; | |
4071 | return 0; | |
4072 | } | |
4073 | ||
cdd6c482 | 4074 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 4075 | { |
ac9721f3 | 4076 | int ret; |
ac9721f3 | 4077 | if (arg != -1) { |
2903ff01 AV |
4078 | struct perf_event *output_event; |
4079 | struct fd output; | |
4080 | ret = perf_fget_light(arg, &output); | |
4081 | if (ret) | |
4082 | return ret; | |
4083 | output_event = output.file->private_data; | |
4084 | ret = perf_event_set_output(event, output_event); | |
4085 | fdput(output); | |
4086 | } else { | |
4087 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 4088 | } |
ac9721f3 PZ |
4089 | return ret; |
4090 | } | |
a4be7c27 | 4091 | |
6fb2915d LZ |
4092 | case PERF_EVENT_IOC_SET_FILTER: |
4093 | return perf_event_set_filter(event, (void __user *)arg); | |
4094 | ||
2541517c AS |
4095 | case PERF_EVENT_IOC_SET_BPF: |
4096 | return perf_event_set_bpf_prog(event, arg); | |
4097 | ||
d859e29f | 4098 | default: |
3df5edad | 4099 | return -ENOTTY; |
d859e29f | 4100 | } |
3df5edad PZ |
4101 | |
4102 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 4103 | perf_event_for_each(event, func); |
3df5edad | 4104 | else |
cdd6c482 | 4105 | perf_event_for_each_child(event, func); |
3df5edad PZ |
4106 | |
4107 | return 0; | |
d859e29f PM |
4108 | } |
4109 | ||
f63a8daa PZ |
4110 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
4111 | { | |
4112 | struct perf_event *event = file->private_data; | |
4113 | struct perf_event_context *ctx; | |
4114 | long ret; | |
4115 | ||
4116 | ctx = perf_event_ctx_lock(event); | |
4117 | ret = _perf_ioctl(event, cmd, arg); | |
4118 | perf_event_ctx_unlock(event, ctx); | |
4119 | ||
4120 | return ret; | |
4121 | } | |
4122 | ||
b3f20785 PM |
4123 | #ifdef CONFIG_COMPAT |
4124 | static long perf_compat_ioctl(struct file *file, unsigned int cmd, | |
4125 | unsigned long arg) | |
4126 | { | |
4127 | switch (_IOC_NR(cmd)) { | |
4128 | case _IOC_NR(PERF_EVENT_IOC_SET_FILTER): | |
4129 | case _IOC_NR(PERF_EVENT_IOC_ID): | |
4130 | /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */ | |
4131 | if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) { | |
4132 | cmd &= ~IOCSIZE_MASK; | |
4133 | cmd |= sizeof(void *) << IOCSIZE_SHIFT; | |
4134 | } | |
4135 | break; | |
4136 | } | |
4137 | return perf_ioctl(file, cmd, arg); | |
4138 | } | |
4139 | #else | |
4140 | # define perf_compat_ioctl NULL | |
4141 | #endif | |
4142 | ||
cdd6c482 | 4143 | int perf_event_task_enable(void) |
771d7cde | 4144 | { |
f63a8daa | 4145 | struct perf_event_context *ctx; |
cdd6c482 | 4146 | struct perf_event *event; |
771d7cde | 4147 | |
cdd6c482 | 4148 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4149 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4150 | ctx = perf_event_ctx_lock(event); | |
4151 | perf_event_for_each_child(event, _perf_event_enable); | |
4152 | perf_event_ctx_unlock(event, ctx); | |
4153 | } | |
cdd6c482 | 4154 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4155 | |
4156 | return 0; | |
4157 | } | |
4158 | ||
cdd6c482 | 4159 | int perf_event_task_disable(void) |
771d7cde | 4160 | { |
f63a8daa | 4161 | struct perf_event_context *ctx; |
cdd6c482 | 4162 | struct perf_event *event; |
771d7cde | 4163 | |
cdd6c482 | 4164 | mutex_lock(¤t->perf_event_mutex); |
f63a8daa PZ |
4165 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) { |
4166 | ctx = perf_event_ctx_lock(event); | |
4167 | perf_event_for_each_child(event, _perf_event_disable); | |
4168 | perf_event_ctx_unlock(event, ctx); | |
4169 | } | |
cdd6c482 | 4170 | mutex_unlock(¤t->perf_event_mutex); |
771d7cde PZ |
4171 | |
4172 | return 0; | |
4173 | } | |
4174 | ||
cdd6c482 | 4175 | static int perf_event_index(struct perf_event *event) |
194002b2 | 4176 | { |
a4eaf7f1 PZ |
4177 | if (event->hw.state & PERF_HES_STOPPED) |
4178 | return 0; | |
4179 | ||
cdd6c482 | 4180 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
4181 | return 0; |
4182 | ||
35edc2a5 | 4183 | return event->pmu->event_idx(event); |
194002b2 PZ |
4184 | } |
4185 | ||
c4794295 | 4186 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 4187 | u64 *now, |
7f310a5d EM |
4188 | u64 *enabled, |
4189 | u64 *running) | |
c4794295 | 4190 | { |
e3f3541c | 4191 | u64 ctx_time; |
c4794295 | 4192 | |
e3f3541c PZ |
4193 | *now = perf_clock(); |
4194 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
4195 | *enabled = ctx_time - event->tstamp_enabled; |
4196 | *running = ctx_time - event->tstamp_running; | |
4197 | } | |
4198 | ||
fa731587 PZ |
4199 | static void perf_event_init_userpage(struct perf_event *event) |
4200 | { | |
4201 | struct perf_event_mmap_page *userpg; | |
4202 | struct ring_buffer *rb; | |
4203 | ||
4204 | rcu_read_lock(); | |
4205 | rb = rcu_dereference(event->rb); | |
4206 | if (!rb) | |
4207 | goto unlock; | |
4208 | ||
4209 | userpg = rb->user_page; | |
4210 | ||
4211 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
4212 | userpg->cap_bit0_is_deprecated = 1; | |
4213 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
e8c6deac AS |
4214 | userpg->data_offset = PAGE_SIZE; |
4215 | userpg->data_size = perf_data_size(rb); | |
fa731587 PZ |
4216 | |
4217 | unlock: | |
4218 | rcu_read_unlock(); | |
4219 | } | |
4220 | ||
c1317ec2 AL |
4221 | void __weak arch_perf_update_userpage( |
4222 | struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now) | |
e3f3541c PZ |
4223 | { |
4224 | } | |
4225 | ||
38ff667b PZ |
4226 | /* |
4227 | * Callers need to ensure there can be no nesting of this function, otherwise | |
4228 | * the seqlock logic goes bad. We can not serialize this because the arch | |
4229 | * code calls this from NMI context. | |
4230 | */ | |
cdd6c482 | 4231 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 4232 | { |
cdd6c482 | 4233 | struct perf_event_mmap_page *userpg; |
76369139 | 4234 | struct ring_buffer *rb; |
e3f3541c | 4235 | u64 enabled, running, now; |
38ff667b PZ |
4236 | |
4237 | rcu_read_lock(); | |
5ec4c599 PZ |
4238 | rb = rcu_dereference(event->rb); |
4239 | if (!rb) | |
4240 | goto unlock; | |
4241 | ||
0d641208 EM |
4242 | /* |
4243 | * compute total_time_enabled, total_time_running | |
4244 | * based on snapshot values taken when the event | |
4245 | * was last scheduled in. | |
4246 | * | |
4247 | * we cannot simply called update_context_time() | |
4248 | * because of locking issue as we can be called in | |
4249 | * NMI context | |
4250 | */ | |
e3f3541c | 4251 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 4252 | |
76369139 | 4253 | userpg = rb->user_page; |
7b732a75 PZ |
4254 | /* |
4255 | * Disable preemption so as to not let the corresponding user-space | |
4256 | * spin too long if we get preempted. | |
4257 | */ | |
4258 | preempt_disable(); | |
37d81828 | 4259 | ++userpg->lock; |
92f22a38 | 4260 | barrier(); |
cdd6c482 | 4261 | userpg->index = perf_event_index(event); |
b5e58793 | 4262 | userpg->offset = perf_event_count(event); |
365a4038 | 4263 | if (userpg->index) |
e7850595 | 4264 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 4265 | |
0d641208 | 4266 | userpg->time_enabled = enabled + |
cdd6c482 | 4267 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 4268 | |
0d641208 | 4269 | userpg->time_running = running + |
cdd6c482 | 4270 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 4271 | |
c1317ec2 | 4272 | arch_perf_update_userpage(event, userpg, now); |
e3f3541c | 4273 | |
92f22a38 | 4274 | barrier(); |
37d81828 | 4275 | ++userpg->lock; |
7b732a75 | 4276 | preempt_enable(); |
38ff667b | 4277 | unlock: |
7b732a75 | 4278 | rcu_read_unlock(); |
37d81828 PM |
4279 | } |
4280 | ||
906010b2 PZ |
4281 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
4282 | { | |
4283 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 4284 | struct ring_buffer *rb; |
906010b2 PZ |
4285 | int ret = VM_FAULT_SIGBUS; |
4286 | ||
4287 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
4288 | if (vmf->pgoff == 0) | |
4289 | ret = 0; | |
4290 | return ret; | |
4291 | } | |
4292 | ||
4293 | rcu_read_lock(); | |
76369139 FW |
4294 | rb = rcu_dereference(event->rb); |
4295 | if (!rb) | |
906010b2 PZ |
4296 | goto unlock; |
4297 | ||
4298 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
4299 | goto unlock; | |
4300 | ||
76369139 | 4301 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
4302 | if (!vmf->page) |
4303 | goto unlock; | |
4304 | ||
4305 | get_page(vmf->page); | |
4306 | vmf->page->mapping = vma->vm_file->f_mapping; | |
4307 | vmf->page->index = vmf->pgoff; | |
4308 | ||
4309 | ret = 0; | |
4310 | unlock: | |
4311 | rcu_read_unlock(); | |
4312 | ||
4313 | return ret; | |
4314 | } | |
4315 | ||
10c6db11 PZ |
4316 | static void ring_buffer_attach(struct perf_event *event, |
4317 | struct ring_buffer *rb) | |
4318 | { | |
b69cf536 | 4319 | struct ring_buffer *old_rb = NULL; |
10c6db11 PZ |
4320 | unsigned long flags; |
4321 | ||
b69cf536 PZ |
4322 | if (event->rb) { |
4323 | /* | |
4324 | * Should be impossible, we set this when removing | |
4325 | * event->rb_entry and wait/clear when adding event->rb_entry. | |
4326 | */ | |
4327 | WARN_ON_ONCE(event->rcu_pending); | |
10c6db11 | 4328 | |
b69cf536 | 4329 | old_rb = event->rb; |
b69cf536 PZ |
4330 | spin_lock_irqsave(&old_rb->event_lock, flags); |
4331 | list_del_rcu(&event->rb_entry); | |
4332 | spin_unlock_irqrestore(&old_rb->event_lock, flags); | |
10c6db11 | 4333 | |
2f993cf0 ON |
4334 | event->rcu_batches = get_state_synchronize_rcu(); |
4335 | event->rcu_pending = 1; | |
b69cf536 | 4336 | } |
10c6db11 | 4337 | |
b69cf536 | 4338 | if (rb) { |
2f993cf0 ON |
4339 | if (event->rcu_pending) { |
4340 | cond_synchronize_rcu(event->rcu_batches); | |
4341 | event->rcu_pending = 0; | |
4342 | } | |
4343 | ||
b69cf536 PZ |
4344 | spin_lock_irqsave(&rb->event_lock, flags); |
4345 | list_add_rcu(&event->rb_entry, &rb->event_list); | |
4346 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
4347 | } | |
4348 | ||
4349 | rcu_assign_pointer(event->rb, rb); | |
4350 | ||
4351 | if (old_rb) { | |
4352 | ring_buffer_put(old_rb); | |
4353 | /* | |
4354 | * Since we detached before setting the new rb, so that we | |
4355 | * could attach the new rb, we could have missed a wakeup. | |
4356 | * Provide it now. | |
4357 | */ | |
4358 | wake_up_all(&event->waitq); | |
4359 | } | |
10c6db11 PZ |
4360 | } |
4361 | ||
4362 | static void ring_buffer_wakeup(struct perf_event *event) | |
4363 | { | |
4364 | struct ring_buffer *rb; | |
4365 | ||
4366 | rcu_read_lock(); | |
4367 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
4368 | if (rb) { |
4369 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
4370 | wake_up_all(&event->waitq); | |
4371 | } | |
10c6db11 PZ |
4372 | rcu_read_unlock(); |
4373 | } | |
4374 | ||
fdc26706 | 4375 | struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 4376 | { |
76369139 | 4377 | struct ring_buffer *rb; |
7b732a75 | 4378 | |
ac9721f3 | 4379 | rcu_read_lock(); |
76369139 FW |
4380 | rb = rcu_dereference(event->rb); |
4381 | if (rb) { | |
4382 | if (!atomic_inc_not_zero(&rb->refcount)) | |
4383 | rb = NULL; | |
ac9721f3 PZ |
4384 | } |
4385 | rcu_read_unlock(); | |
4386 | ||
76369139 | 4387 | return rb; |
ac9721f3 PZ |
4388 | } |
4389 | ||
fdc26706 | 4390 | void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 4391 | { |
76369139 | 4392 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 4393 | return; |
7b732a75 | 4394 | |
9bb5d40c | 4395 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 4396 | |
76369139 | 4397 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
4398 | } |
4399 | ||
4400 | static void perf_mmap_open(struct vm_area_struct *vma) | |
4401 | { | |
cdd6c482 | 4402 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4403 | |
cdd6c482 | 4404 | atomic_inc(&event->mmap_count); |
9bb5d40c | 4405 | atomic_inc(&event->rb->mmap_count); |
1e0fb9ec | 4406 | |
45bfb2e5 PZ |
4407 | if (vma->vm_pgoff) |
4408 | atomic_inc(&event->rb->aux_mmap_count); | |
4409 | ||
1e0fb9ec AL |
4410 | if (event->pmu->event_mapped) |
4411 | event->pmu->event_mapped(event); | |
7b732a75 PZ |
4412 | } |
4413 | ||
9bb5d40c PZ |
4414 | /* |
4415 | * A buffer can be mmap()ed multiple times; either directly through the same | |
4416 | * event, or through other events by use of perf_event_set_output(). | |
4417 | * | |
4418 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
4419 | * the buffer here, where we still have a VM context. This means we need | |
4420 | * to detach all events redirecting to us. | |
4421 | */ | |
7b732a75 PZ |
4422 | static void perf_mmap_close(struct vm_area_struct *vma) |
4423 | { | |
cdd6c482 | 4424 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 4425 | |
b69cf536 | 4426 | struct ring_buffer *rb = ring_buffer_get(event); |
9bb5d40c PZ |
4427 | struct user_struct *mmap_user = rb->mmap_user; |
4428 | int mmap_locked = rb->mmap_locked; | |
4429 | unsigned long size = perf_data_size(rb); | |
789f90fc | 4430 | |
1e0fb9ec AL |
4431 | if (event->pmu->event_unmapped) |
4432 | event->pmu->event_unmapped(event); | |
4433 | ||
45bfb2e5 PZ |
4434 | /* |
4435 | * rb->aux_mmap_count will always drop before rb->mmap_count and | |
4436 | * event->mmap_count, so it is ok to use event->mmap_mutex to | |
4437 | * serialize with perf_mmap here. | |
4438 | */ | |
4439 | if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff && | |
4440 | atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) { | |
4441 | atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm); | |
4442 | vma->vm_mm->pinned_vm -= rb->aux_mmap_locked; | |
4443 | ||
4444 | rb_free_aux(rb); | |
4445 | mutex_unlock(&event->mmap_mutex); | |
4446 | } | |
4447 | ||
9bb5d40c PZ |
4448 | atomic_dec(&rb->mmap_count); |
4449 | ||
4450 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
b69cf536 | 4451 | goto out_put; |
9bb5d40c | 4452 | |
b69cf536 | 4453 | ring_buffer_attach(event, NULL); |
9bb5d40c PZ |
4454 | mutex_unlock(&event->mmap_mutex); |
4455 | ||
4456 | /* If there's still other mmap()s of this buffer, we're done. */ | |
b69cf536 PZ |
4457 | if (atomic_read(&rb->mmap_count)) |
4458 | goto out_put; | |
ac9721f3 | 4459 | |
9bb5d40c PZ |
4460 | /* |
4461 | * No other mmap()s, detach from all other events that might redirect | |
4462 | * into the now unreachable buffer. Somewhat complicated by the | |
4463 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
4464 | */ | |
4465 | again: | |
4466 | rcu_read_lock(); | |
4467 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
4468 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
4469 | /* | |
4470 | * This event is en-route to free_event() which will | |
4471 | * detach it and remove it from the list. | |
4472 | */ | |
4473 | continue; | |
4474 | } | |
4475 | rcu_read_unlock(); | |
789f90fc | 4476 | |
9bb5d40c PZ |
4477 | mutex_lock(&event->mmap_mutex); |
4478 | /* | |
4479 | * Check we didn't race with perf_event_set_output() which can | |
4480 | * swizzle the rb from under us while we were waiting to | |
4481 | * acquire mmap_mutex. | |
4482 | * | |
4483 | * If we find a different rb; ignore this event, a next | |
4484 | * iteration will no longer find it on the list. We have to | |
4485 | * still restart the iteration to make sure we're not now | |
4486 | * iterating the wrong list. | |
4487 | */ | |
b69cf536 PZ |
4488 | if (event->rb == rb) |
4489 | ring_buffer_attach(event, NULL); | |
4490 | ||
cdd6c482 | 4491 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 4492 | put_event(event); |
ac9721f3 | 4493 | |
9bb5d40c PZ |
4494 | /* |
4495 | * Restart the iteration; either we're on the wrong list or | |
4496 | * destroyed its integrity by doing a deletion. | |
4497 | */ | |
4498 | goto again; | |
7b732a75 | 4499 | } |
9bb5d40c PZ |
4500 | rcu_read_unlock(); |
4501 | ||
4502 | /* | |
4503 | * It could be there's still a few 0-ref events on the list; they'll | |
4504 | * get cleaned up by free_event() -- they'll also still have their | |
4505 | * ref on the rb and will free it whenever they are done with it. | |
4506 | * | |
4507 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
4508 | * undo the VM accounting. | |
4509 | */ | |
4510 | ||
4511 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
4512 | vma->vm_mm->pinned_vm -= mmap_locked; | |
4513 | free_uid(mmap_user); | |
4514 | ||
b69cf536 | 4515 | out_put: |
9bb5d40c | 4516 | ring_buffer_put(rb); /* could be last */ |
37d81828 PM |
4517 | } |
4518 | ||
f0f37e2f | 4519 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 | 4520 | .open = perf_mmap_open, |
45bfb2e5 | 4521 | .close = perf_mmap_close, /* non mergable */ |
43a21ea8 PZ |
4522 | .fault = perf_mmap_fault, |
4523 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
4524 | }; |
4525 | ||
4526 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
4527 | { | |
cdd6c482 | 4528 | struct perf_event *event = file->private_data; |
22a4f650 | 4529 | unsigned long user_locked, user_lock_limit; |
789f90fc | 4530 | struct user_struct *user = current_user(); |
22a4f650 | 4531 | unsigned long locked, lock_limit; |
45bfb2e5 | 4532 | struct ring_buffer *rb = NULL; |
7b732a75 PZ |
4533 | unsigned long vma_size; |
4534 | unsigned long nr_pages; | |
45bfb2e5 | 4535 | long user_extra = 0, extra = 0; |
d57e34fd | 4536 | int ret = 0, flags = 0; |
37d81828 | 4537 | |
c7920614 PZ |
4538 | /* |
4539 | * Don't allow mmap() of inherited per-task counters. This would | |
4540 | * create a performance issue due to all children writing to the | |
76369139 | 4541 | * same rb. |
c7920614 PZ |
4542 | */ |
4543 | if (event->cpu == -1 && event->attr.inherit) | |
4544 | return -EINVAL; | |
4545 | ||
43a21ea8 | 4546 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 4547 | return -EINVAL; |
7b732a75 PZ |
4548 | |
4549 | vma_size = vma->vm_end - vma->vm_start; | |
45bfb2e5 PZ |
4550 | |
4551 | if (vma->vm_pgoff == 0) { | |
4552 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
4553 | } else { | |
4554 | /* | |
4555 | * AUX area mapping: if rb->aux_nr_pages != 0, it's already | |
4556 | * mapped, all subsequent mappings should have the same size | |
4557 | * and offset. Must be above the normal perf buffer. | |
4558 | */ | |
4559 | u64 aux_offset, aux_size; | |
4560 | ||
4561 | if (!event->rb) | |
4562 | return -EINVAL; | |
4563 | ||
4564 | nr_pages = vma_size / PAGE_SIZE; | |
4565 | ||
4566 | mutex_lock(&event->mmap_mutex); | |
4567 | ret = -EINVAL; | |
4568 | ||
4569 | rb = event->rb; | |
4570 | if (!rb) | |
4571 | goto aux_unlock; | |
4572 | ||
4573 | aux_offset = ACCESS_ONCE(rb->user_page->aux_offset); | |
4574 | aux_size = ACCESS_ONCE(rb->user_page->aux_size); | |
4575 | ||
4576 | if (aux_offset < perf_data_size(rb) + PAGE_SIZE) | |
4577 | goto aux_unlock; | |
4578 | ||
4579 | if (aux_offset != vma->vm_pgoff << PAGE_SHIFT) | |
4580 | goto aux_unlock; | |
4581 | ||
4582 | /* already mapped with a different offset */ | |
4583 | if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff) | |
4584 | goto aux_unlock; | |
4585 | ||
4586 | if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE) | |
4587 | goto aux_unlock; | |
4588 | ||
4589 | /* already mapped with a different size */ | |
4590 | if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages) | |
4591 | goto aux_unlock; | |
4592 | ||
4593 | if (!is_power_of_2(nr_pages)) | |
4594 | goto aux_unlock; | |
4595 | ||
4596 | if (!atomic_inc_not_zero(&rb->mmap_count)) | |
4597 | goto aux_unlock; | |
4598 | ||
4599 | if (rb_has_aux(rb)) { | |
4600 | atomic_inc(&rb->aux_mmap_count); | |
4601 | ret = 0; | |
4602 | goto unlock; | |
4603 | } | |
4604 | ||
4605 | atomic_set(&rb->aux_mmap_count, 1); | |
4606 | user_extra = nr_pages; | |
4607 | ||
4608 | goto accounting; | |
4609 | } | |
7b732a75 | 4610 | |
7730d865 | 4611 | /* |
76369139 | 4612 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
4613 | * can do bitmasks instead of modulo. |
4614 | */ | |
2ed11312 | 4615 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) |
37d81828 PM |
4616 | return -EINVAL; |
4617 | ||
7b732a75 | 4618 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
4619 | return -EINVAL; |
4620 | ||
cdd6c482 | 4621 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 4622 | again: |
cdd6c482 | 4623 | mutex_lock(&event->mmap_mutex); |
76369139 | 4624 | if (event->rb) { |
9bb5d40c | 4625 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4626 | ret = -EINVAL; |
9bb5d40c PZ |
4627 | goto unlock; |
4628 | } | |
4629 | ||
4630 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4631 | /* | |
4632 | * Raced against perf_mmap_close() through | |
4633 | * perf_event_set_output(). Try again, hope for better | |
4634 | * luck. | |
4635 | */ | |
4636 | mutex_unlock(&event->mmap_mutex); | |
4637 | goto again; | |
4638 | } | |
4639 | ||
ebb3c4c4 PZ |
4640 | goto unlock; |
4641 | } | |
4642 | ||
789f90fc | 4643 | user_extra = nr_pages + 1; |
45bfb2e5 PZ |
4644 | |
4645 | accounting: | |
cdd6c482 | 4646 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4647 | |
4648 | /* | |
4649 | * Increase the limit linearly with more CPUs: | |
4650 | */ | |
4651 | user_lock_limit *= num_online_cpus(); | |
4652 | ||
789f90fc | 4653 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4654 | |
789f90fc PZ |
4655 | if (user_locked > user_lock_limit) |
4656 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4657 | |
78d7d407 | 4658 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4659 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4660 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4661 | |
459ec28a IM |
4662 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4663 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4664 | ret = -EPERM; |
4665 | goto unlock; | |
4666 | } | |
7b732a75 | 4667 | |
45bfb2e5 | 4668 | WARN_ON(!rb && event->rb); |
906010b2 | 4669 | |
d57e34fd | 4670 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4671 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4672 | |
76369139 | 4673 | if (!rb) { |
45bfb2e5 PZ |
4674 | rb = rb_alloc(nr_pages, |
4675 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4676 | event->cpu, flags); | |
26cb63ad | 4677 | |
45bfb2e5 PZ |
4678 | if (!rb) { |
4679 | ret = -ENOMEM; | |
4680 | goto unlock; | |
4681 | } | |
43a21ea8 | 4682 | |
45bfb2e5 PZ |
4683 | atomic_set(&rb->mmap_count, 1); |
4684 | rb->mmap_user = get_current_user(); | |
4685 | rb->mmap_locked = extra; | |
26cb63ad | 4686 | |
45bfb2e5 | 4687 | ring_buffer_attach(event, rb); |
ac9721f3 | 4688 | |
45bfb2e5 PZ |
4689 | perf_event_init_userpage(event); |
4690 | perf_event_update_userpage(event); | |
4691 | } else { | |
1a594131 AS |
4692 | ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages, |
4693 | event->attr.aux_watermark, flags); | |
45bfb2e5 PZ |
4694 | if (!ret) |
4695 | rb->aux_mmap_locked = extra; | |
4696 | } | |
9a0f05cb | 4697 | |
ebb3c4c4 | 4698 | unlock: |
45bfb2e5 PZ |
4699 | if (!ret) { |
4700 | atomic_long_add(user_extra, &user->locked_vm); | |
4701 | vma->vm_mm->pinned_vm += extra; | |
4702 | ||
ac9721f3 | 4703 | atomic_inc(&event->mmap_count); |
45bfb2e5 PZ |
4704 | } else if (rb) { |
4705 | atomic_dec(&rb->mmap_count); | |
4706 | } | |
4707 | aux_unlock: | |
cdd6c482 | 4708 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4709 | |
9bb5d40c PZ |
4710 | /* |
4711 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4712 | * vma. | |
4713 | */ | |
26cb63ad | 4714 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4715 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 | 4716 | |
1e0fb9ec AL |
4717 | if (event->pmu->event_mapped) |
4718 | event->pmu->event_mapped(event); | |
4719 | ||
7b732a75 | 4720 | return ret; |
37d81828 PM |
4721 | } |
4722 | ||
3c446b3d PZ |
4723 | static int perf_fasync(int fd, struct file *filp, int on) |
4724 | { | |
496ad9aa | 4725 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4726 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4727 | int retval; |
4728 | ||
4729 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4730 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4731 | mutex_unlock(&inode->i_mutex); |
4732 | ||
4733 | if (retval < 0) | |
4734 | return retval; | |
4735 | ||
4736 | return 0; | |
4737 | } | |
4738 | ||
0793a61d | 4739 | static const struct file_operations perf_fops = { |
3326c1ce | 4740 | .llseek = no_llseek, |
0793a61d TG |
4741 | .release = perf_release, |
4742 | .read = perf_read, | |
4743 | .poll = perf_poll, | |
d859e29f | 4744 | .unlocked_ioctl = perf_ioctl, |
b3f20785 | 4745 | .compat_ioctl = perf_compat_ioctl, |
37d81828 | 4746 | .mmap = perf_mmap, |
3c446b3d | 4747 | .fasync = perf_fasync, |
0793a61d TG |
4748 | }; |
4749 | ||
925d519a | 4750 | /* |
cdd6c482 | 4751 | * Perf event wakeup |
925d519a PZ |
4752 | * |
4753 | * If there's data, ensure we set the poll() state and publish everything | |
4754 | * to user-space before waking everybody up. | |
4755 | */ | |
4756 | ||
cdd6c482 | 4757 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4758 | { |
10c6db11 | 4759 | ring_buffer_wakeup(event); |
4c9e2542 | 4760 | |
cdd6c482 IM |
4761 | if (event->pending_kill) { |
4762 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4763 | event->pending_kill = 0; | |
4c9e2542 | 4764 | } |
925d519a PZ |
4765 | } |
4766 | ||
e360adbe | 4767 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4768 | { |
cdd6c482 IM |
4769 | struct perf_event *event = container_of(entry, |
4770 | struct perf_event, pending); | |
d525211f PZ |
4771 | int rctx; |
4772 | ||
4773 | rctx = perf_swevent_get_recursion_context(); | |
4774 | /* | |
4775 | * If we 'fail' here, that's OK, it means recursion is already disabled | |
4776 | * and we won't recurse 'further'. | |
4777 | */ | |
79f14641 | 4778 | |
cdd6c482 IM |
4779 | if (event->pending_disable) { |
4780 | event->pending_disable = 0; | |
4781 | __perf_event_disable(event); | |
79f14641 PZ |
4782 | } |
4783 | ||
cdd6c482 IM |
4784 | if (event->pending_wakeup) { |
4785 | event->pending_wakeup = 0; | |
4786 | perf_event_wakeup(event); | |
79f14641 | 4787 | } |
d525211f PZ |
4788 | |
4789 | if (rctx >= 0) | |
4790 | perf_swevent_put_recursion_context(rctx); | |
79f14641 PZ |
4791 | } |
4792 | ||
39447b38 ZY |
4793 | /* |
4794 | * We assume there is only KVM supporting the callbacks. | |
4795 | * Later on, we might change it to a list if there is | |
4796 | * another virtualization implementation supporting the callbacks. | |
4797 | */ | |
4798 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4799 | ||
4800 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4801 | { | |
4802 | perf_guest_cbs = cbs; | |
4803 | return 0; | |
4804 | } | |
4805 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4806 | ||
4807 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4808 | { | |
4809 | perf_guest_cbs = NULL; | |
4810 | return 0; | |
4811 | } | |
4812 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4813 | ||
4018994f JO |
4814 | static void |
4815 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4816 | struct pt_regs *regs, u64 mask) | |
4817 | { | |
4818 | int bit; | |
4819 | ||
4820 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4821 | sizeof(mask) * BITS_PER_BYTE) { | |
4822 | u64 val; | |
4823 | ||
4824 | val = perf_reg_value(regs, bit); | |
4825 | perf_output_put(handle, val); | |
4826 | } | |
4827 | } | |
4828 | ||
60e2364e | 4829 | static void perf_sample_regs_user(struct perf_regs *regs_user, |
88a7c26a AL |
4830 | struct pt_regs *regs, |
4831 | struct pt_regs *regs_user_copy) | |
4018994f | 4832 | { |
88a7c26a AL |
4833 | if (user_mode(regs)) { |
4834 | regs_user->abi = perf_reg_abi(current); | |
2565711f | 4835 | regs_user->regs = regs; |
88a7c26a AL |
4836 | } else if (current->mm) { |
4837 | perf_get_regs_user(regs_user, regs, regs_user_copy); | |
2565711f PZ |
4838 | } else { |
4839 | regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE; | |
4840 | regs_user->regs = NULL; | |
4018994f JO |
4841 | } |
4842 | } | |
4843 | ||
60e2364e SE |
4844 | static void perf_sample_regs_intr(struct perf_regs *regs_intr, |
4845 | struct pt_regs *regs) | |
4846 | { | |
4847 | regs_intr->regs = regs; | |
4848 | regs_intr->abi = perf_reg_abi(current); | |
4849 | } | |
4850 | ||
4851 | ||
c5ebcedb JO |
4852 | /* |
4853 | * Get remaining task size from user stack pointer. | |
4854 | * | |
4855 | * It'd be better to take stack vma map and limit this more | |
4856 | * precisly, but there's no way to get it safely under interrupt, | |
4857 | * so using TASK_SIZE as limit. | |
4858 | */ | |
4859 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4860 | { | |
4861 | unsigned long addr = perf_user_stack_pointer(regs); | |
4862 | ||
4863 | if (!addr || addr >= TASK_SIZE) | |
4864 | return 0; | |
4865 | ||
4866 | return TASK_SIZE - addr; | |
4867 | } | |
4868 | ||
4869 | static u16 | |
4870 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4871 | struct pt_regs *regs) | |
4872 | { | |
4873 | u64 task_size; | |
4874 | ||
4875 | /* No regs, no stack pointer, no dump. */ | |
4876 | if (!regs) | |
4877 | return 0; | |
4878 | ||
4879 | /* | |
4880 | * Check if we fit in with the requested stack size into the: | |
4881 | * - TASK_SIZE | |
4882 | * If we don't, we limit the size to the TASK_SIZE. | |
4883 | * | |
4884 | * - remaining sample size | |
4885 | * If we don't, we customize the stack size to | |
4886 | * fit in to the remaining sample size. | |
4887 | */ | |
4888 | ||
4889 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4890 | stack_size = min(stack_size, (u16) task_size); | |
4891 | ||
4892 | /* Current header size plus static size and dynamic size. */ | |
4893 | header_size += 2 * sizeof(u64); | |
4894 | ||
4895 | /* Do we fit in with the current stack dump size? */ | |
4896 | if ((u16) (header_size + stack_size) < header_size) { | |
4897 | /* | |
4898 | * If we overflow the maximum size for the sample, | |
4899 | * we customize the stack dump size to fit in. | |
4900 | */ | |
4901 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4902 | stack_size = round_up(stack_size, sizeof(u64)); | |
4903 | } | |
4904 | ||
4905 | return stack_size; | |
4906 | } | |
4907 | ||
4908 | static void | |
4909 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4910 | struct pt_regs *regs) | |
4911 | { | |
4912 | /* Case of a kernel thread, nothing to dump */ | |
4913 | if (!regs) { | |
4914 | u64 size = 0; | |
4915 | perf_output_put(handle, size); | |
4916 | } else { | |
4917 | unsigned long sp; | |
4918 | unsigned int rem; | |
4919 | u64 dyn_size; | |
4920 | ||
4921 | /* | |
4922 | * We dump: | |
4923 | * static size | |
4924 | * - the size requested by user or the best one we can fit | |
4925 | * in to the sample max size | |
4926 | * data | |
4927 | * - user stack dump data | |
4928 | * dynamic size | |
4929 | * - the actual dumped size | |
4930 | */ | |
4931 | ||
4932 | /* Static size. */ | |
4933 | perf_output_put(handle, dump_size); | |
4934 | ||
4935 | /* Data. */ | |
4936 | sp = perf_user_stack_pointer(regs); | |
4937 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4938 | dyn_size = dump_size - rem; | |
4939 | ||
4940 | perf_output_skip(handle, rem); | |
4941 | ||
4942 | /* Dynamic size. */ | |
4943 | perf_output_put(handle, dyn_size); | |
4944 | } | |
4945 | } | |
4946 | ||
c980d109 ACM |
4947 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4948 | struct perf_sample_data *data, | |
4949 | struct perf_event *event) | |
6844c09d ACM |
4950 | { |
4951 | u64 sample_type = event->attr.sample_type; | |
4952 | ||
4953 | data->type = sample_type; | |
4954 | header->size += event->id_header_size; | |
4955 | ||
4956 | if (sample_type & PERF_SAMPLE_TID) { | |
4957 | /* namespace issues */ | |
4958 | data->tid_entry.pid = perf_event_pid(event, current); | |
4959 | data->tid_entry.tid = perf_event_tid(event, current); | |
4960 | } | |
4961 | ||
4962 | if (sample_type & PERF_SAMPLE_TIME) | |
34f43927 | 4963 | data->time = perf_event_clock(event); |
6844c09d | 4964 | |
ff3d527c | 4965 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4966 | data->id = primary_event_id(event); |
4967 | ||
4968 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4969 | data->stream_id = event->id; | |
4970 | ||
4971 | if (sample_type & PERF_SAMPLE_CPU) { | |
4972 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4973 | data->cpu_entry.reserved = 0; | |
4974 | } | |
4975 | } | |
4976 | ||
76369139 FW |
4977 | void perf_event_header__init_id(struct perf_event_header *header, |
4978 | struct perf_sample_data *data, | |
4979 | struct perf_event *event) | |
c980d109 ACM |
4980 | { |
4981 | if (event->attr.sample_id_all) | |
4982 | __perf_event_header__init_id(header, data, event); | |
4983 | } | |
4984 | ||
4985 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4986 | struct perf_sample_data *data) | |
4987 | { | |
4988 | u64 sample_type = data->type; | |
4989 | ||
4990 | if (sample_type & PERF_SAMPLE_TID) | |
4991 | perf_output_put(handle, data->tid_entry); | |
4992 | ||
4993 | if (sample_type & PERF_SAMPLE_TIME) | |
4994 | perf_output_put(handle, data->time); | |
4995 | ||
4996 | if (sample_type & PERF_SAMPLE_ID) | |
4997 | perf_output_put(handle, data->id); | |
4998 | ||
4999 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
5000 | perf_output_put(handle, data->stream_id); | |
5001 | ||
5002 | if (sample_type & PERF_SAMPLE_CPU) | |
5003 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
5004 | |
5005 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
5006 | perf_output_put(handle, data->id); | |
c980d109 ACM |
5007 | } |
5008 | ||
76369139 FW |
5009 | void perf_event__output_id_sample(struct perf_event *event, |
5010 | struct perf_output_handle *handle, | |
5011 | struct perf_sample_data *sample) | |
c980d109 ACM |
5012 | { |
5013 | if (event->attr.sample_id_all) | |
5014 | __perf_event__output_id_sample(handle, sample); | |
5015 | } | |
5016 | ||
3dab77fb | 5017 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
5018 | struct perf_event *event, |
5019 | u64 enabled, u64 running) | |
3dab77fb | 5020 | { |
cdd6c482 | 5021 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
5022 | u64 values[4]; |
5023 | int n = 0; | |
5024 | ||
b5e58793 | 5025 | values[n++] = perf_event_count(event); |
3dab77fb | 5026 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 5027 | values[n++] = enabled + |
cdd6c482 | 5028 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
5029 | } |
5030 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 5031 | values[n++] = running + |
cdd6c482 | 5032 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
5033 | } |
5034 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 5035 | values[n++] = primary_event_id(event); |
3dab77fb | 5036 | |
76369139 | 5037 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
5038 | } |
5039 | ||
5040 | /* | |
cdd6c482 | 5041 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
5042 | */ |
5043 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
5044 | struct perf_event *event, |
5045 | u64 enabled, u64 running) | |
3dab77fb | 5046 | { |
cdd6c482 IM |
5047 | struct perf_event *leader = event->group_leader, *sub; |
5048 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
5049 | u64 values[5]; |
5050 | int n = 0; | |
5051 | ||
5052 | values[n++] = 1 + leader->nr_siblings; | |
5053 | ||
5054 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 5055 | values[n++] = enabled; |
3dab77fb PZ |
5056 | |
5057 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 5058 | values[n++] = running; |
3dab77fb | 5059 | |
cdd6c482 | 5060 | if (leader != event) |
3dab77fb PZ |
5061 | leader->pmu->read(leader); |
5062 | ||
b5e58793 | 5063 | values[n++] = perf_event_count(leader); |
3dab77fb | 5064 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 5065 | values[n++] = primary_event_id(leader); |
3dab77fb | 5066 | |
76369139 | 5067 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 5068 | |
65abc865 | 5069 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
5070 | n = 0; |
5071 | ||
6f5ab001 JO |
5072 | if ((sub != event) && |
5073 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
5074 | sub->pmu->read(sub); |
5075 | ||
b5e58793 | 5076 | values[n++] = perf_event_count(sub); |
3dab77fb | 5077 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 5078 | values[n++] = primary_event_id(sub); |
3dab77fb | 5079 | |
76369139 | 5080 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
5081 | } |
5082 | } | |
5083 | ||
eed01528 SE |
5084 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
5085 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
5086 | ||
3dab77fb | 5087 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 5088 | struct perf_event *event) |
3dab77fb | 5089 | { |
e3f3541c | 5090 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
5091 | u64 read_format = event->attr.read_format; |
5092 | ||
5093 | /* | |
5094 | * compute total_time_enabled, total_time_running | |
5095 | * based on snapshot values taken when the event | |
5096 | * was last scheduled in. | |
5097 | * | |
5098 | * we cannot simply called update_context_time() | |
5099 | * because of locking issue as we are called in | |
5100 | * NMI context | |
5101 | */ | |
c4794295 | 5102 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 5103 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 5104 | |
cdd6c482 | 5105 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 5106 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 5107 | else |
eed01528 | 5108 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
5109 | } |
5110 | ||
5622f295 MM |
5111 | void perf_output_sample(struct perf_output_handle *handle, |
5112 | struct perf_event_header *header, | |
5113 | struct perf_sample_data *data, | |
cdd6c482 | 5114 | struct perf_event *event) |
5622f295 MM |
5115 | { |
5116 | u64 sample_type = data->type; | |
5117 | ||
5118 | perf_output_put(handle, *header); | |
5119 | ||
ff3d527c AH |
5120 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
5121 | perf_output_put(handle, data->id); | |
5122 | ||
5622f295 MM |
5123 | if (sample_type & PERF_SAMPLE_IP) |
5124 | perf_output_put(handle, data->ip); | |
5125 | ||
5126 | if (sample_type & PERF_SAMPLE_TID) | |
5127 | perf_output_put(handle, data->tid_entry); | |
5128 | ||
5129 | if (sample_type & PERF_SAMPLE_TIME) | |
5130 | perf_output_put(handle, data->time); | |
5131 | ||
5132 | if (sample_type & PERF_SAMPLE_ADDR) | |
5133 | perf_output_put(handle, data->addr); | |
5134 | ||
5135 | if (sample_type & PERF_SAMPLE_ID) | |
5136 | perf_output_put(handle, data->id); | |
5137 | ||
5138 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
5139 | perf_output_put(handle, data->stream_id); | |
5140 | ||
5141 | if (sample_type & PERF_SAMPLE_CPU) | |
5142 | perf_output_put(handle, data->cpu_entry); | |
5143 | ||
5144 | if (sample_type & PERF_SAMPLE_PERIOD) | |
5145 | perf_output_put(handle, data->period); | |
5146 | ||
5147 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 5148 | perf_output_read(handle, event); |
5622f295 MM |
5149 | |
5150 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
5151 | if (data->callchain) { | |
5152 | int size = 1; | |
5153 | ||
5154 | if (data->callchain) | |
5155 | size += data->callchain->nr; | |
5156 | ||
5157 | size *= sizeof(u64); | |
5158 | ||
76369139 | 5159 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
5160 | } else { |
5161 | u64 nr = 0; | |
5162 | perf_output_put(handle, nr); | |
5163 | } | |
5164 | } | |
5165 | ||
5166 | if (sample_type & PERF_SAMPLE_RAW) { | |
5167 | if (data->raw) { | |
5168 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
5169 | __output_copy(handle, data->raw->data, |
5170 | data->raw->size); | |
5622f295 MM |
5171 | } else { |
5172 | struct { | |
5173 | u32 size; | |
5174 | u32 data; | |
5175 | } raw = { | |
5176 | .size = sizeof(u32), | |
5177 | .data = 0, | |
5178 | }; | |
5179 | perf_output_put(handle, raw); | |
5180 | } | |
5181 | } | |
a7ac67ea | 5182 | |
bce38cd5 SE |
5183 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
5184 | if (data->br_stack) { | |
5185 | size_t size; | |
5186 | ||
5187 | size = data->br_stack->nr | |
5188 | * sizeof(struct perf_branch_entry); | |
5189 | ||
5190 | perf_output_put(handle, data->br_stack->nr); | |
5191 | perf_output_copy(handle, data->br_stack->entries, size); | |
5192 | } else { | |
5193 | /* | |
5194 | * we always store at least the value of nr | |
5195 | */ | |
5196 | u64 nr = 0; | |
5197 | perf_output_put(handle, nr); | |
5198 | } | |
5199 | } | |
4018994f JO |
5200 | |
5201 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
5202 | u64 abi = data->regs_user.abi; | |
5203 | ||
5204 | /* | |
5205 | * If there are no regs to dump, notice it through | |
5206 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5207 | */ | |
5208 | perf_output_put(handle, abi); | |
5209 | ||
5210 | if (abi) { | |
5211 | u64 mask = event->attr.sample_regs_user; | |
5212 | perf_output_sample_regs(handle, | |
5213 | data->regs_user.regs, | |
5214 | mask); | |
5215 | } | |
5216 | } | |
c5ebcedb | 5217 | |
a5cdd40c | 5218 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
5219 | perf_output_sample_ustack(handle, |
5220 | data->stack_user_size, | |
5221 | data->regs_user.regs); | |
a5cdd40c | 5222 | } |
c3feedf2 AK |
5223 | |
5224 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
5225 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
5226 | |
5227 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
5228 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c | 5229 | |
fdfbbd07 AK |
5230 | if (sample_type & PERF_SAMPLE_TRANSACTION) |
5231 | perf_output_put(handle, data->txn); | |
5232 | ||
60e2364e SE |
5233 | if (sample_type & PERF_SAMPLE_REGS_INTR) { |
5234 | u64 abi = data->regs_intr.abi; | |
5235 | /* | |
5236 | * If there are no regs to dump, notice it through | |
5237 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
5238 | */ | |
5239 | perf_output_put(handle, abi); | |
5240 | ||
5241 | if (abi) { | |
5242 | u64 mask = event->attr.sample_regs_intr; | |
5243 | ||
5244 | perf_output_sample_regs(handle, | |
5245 | data->regs_intr.regs, | |
5246 | mask); | |
5247 | } | |
5248 | } | |
5249 | ||
a5cdd40c PZ |
5250 | if (!event->attr.watermark) { |
5251 | int wakeup_events = event->attr.wakeup_events; | |
5252 | ||
5253 | if (wakeup_events) { | |
5254 | struct ring_buffer *rb = handle->rb; | |
5255 | int events = local_inc_return(&rb->events); | |
5256 | ||
5257 | if (events >= wakeup_events) { | |
5258 | local_sub(wakeup_events, &rb->events); | |
5259 | local_inc(&rb->wakeup); | |
5260 | } | |
5261 | } | |
5262 | } | |
5622f295 MM |
5263 | } |
5264 | ||
5265 | void perf_prepare_sample(struct perf_event_header *header, | |
5266 | struct perf_sample_data *data, | |
cdd6c482 | 5267 | struct perf_event *event, |
5622f295 | 5268 | struct pt_regs *regs) |
7b732a75 | 5269 | { |
cdd6c482 | 5270 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 5271 | |
cdd6c482 | 5272 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 5273 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
5274 | |
5275 | header->misc = 0; | |
5276 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 5277 | |
c980d109 | 5278 | __perf_event_header__init_id(header, data, event); |
6844c09d | 5279 | |
c320c7b7 | 5280 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
5281 | data->ip = perf_instruction_pointer(regs); |
5282 | ||
b23f3325 | 5283 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 5284 | int size = 1; |
394ee076 | 5285 | |
e6dab5ff | 5286 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
5287 | |
5288 | if (data->callchain) | |
5289 | size += data->callchain->nr; | |
5290 | ||
5291 | header->size += size * sizeof(u64); | |
394ee076 PZ |
5292 | } |
5293 | ||
3a43ce68 | 5294 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
5295 | int size = sizeof(u32); |
5296 | ||
5297 | if (data->raw) | |
5298 | size += data->raw->size; | |
5299 | else | |
5300 | size += sizeof(u32); | |
5301 | ||
5302 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 5303 | header->size += size; |
7f453c24 | 5304 | } |
bce38cd5 SE |
5305 | |
5306 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
5307 | int size = sizeof(u64); /* nr */ | |
5308 | if (data->br_stack) { | |
5309 | size += data->br_stack->nr | |
5310 | * sizeof(struct perf_branch_entry); | |
5311 | } | |
5312 | header->size += size; | |
5313 | } | |
4018994f | 5314 | |
2565711f | 5315 | if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER)) |
88a7c26a AL |
5316 | perf_sample_regs_user(&data->regs_user, regs, |
5317 | &data->regs_user_copy); | |
2565711f | 5318 | |
4018994f JO |
5319 | if (sample_type & PERF_SAMPLE_REGS_USER) { |
5320 | /* regs dump ABI info */ | |
5321 | int size = sizeof(u64); | |
5322 | ||
4018994f JO |
5323 | if (data->regs_user.regs) { |
5324 | u64 mask = event->attr.sample_regs_user; | |
5325 | size += hweight64(mask) * sizeof(u64); | |
5326 | } | |
5327 | ||
5328 | header->size += size; | |
5329 | } | |
c5ebcedb JO |
5330 | |
5331 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
5332 | /* | |
5333 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
5334 | * processed as the last one or have additional check added | |
5335 | * in case new sample type is added, because we could eat | |
5336 | * up the rest of the sample size. | |
5337 | */ | |
c5ebcedb JO |
5338 | u16 stack_size = event->attr.sample_stack_user; |
5339 | u16 size = sizeof(u64); | |
5340 | ||
c5ebcedb | 5341 | stack_size = perf_sample_ustack_size(stack_size, header->size, |
2565711f | 5342 | data->regs_user.regs); |
c5ebcedb JO |
5343 | |
5344 | /* | |
5345 | * If there is something to dump, add space for the dump | |
5346 | * itself and for the field that tells the dynamic size, | |
5347 | * which is how many have been actually dumped. | |
5348 | */ | |
5349 | if (stack_size) | |
5350 | size += sizeof(u64) + stack_size; | |
5351 | ||
5352 | data->stack_user_size = stack_size; | |
5353 | header->size += size; | |
5354 | } | |
60e2364e SE |
5355 | |
5356 | if (sample_type & PERF_SAMPLE_REGS_INTR) { | |
5357 | /* regs dump ABI info */ | |
5358 | int size = sizeof(u64); | |
5359 | ||
5360 | perf_sample_regs_intr(&data->regs_intr, regs); | |
5361 | ||
5362 | if (data->regs_intr.regs) { | |
5363 | u64 mask = event->attr.sample_regs_intr; | |
5364 | ||
5365 | size += hweight64(mask) * sizeof(u64); | |
5366 | } | |
5367 | ||
5368 | header->size += size; | |
5369 | } | |
5622f295 | 5370 | } |
7f453c24 | 5371 | |
21509084 YZ |
5372 | void perf_event_output(struct perf_event *event, |
5373 | struct perf_sample_data *data, | |
5374 | struct pt_regs *regs) | |
5622f295 MM |
5375 | { |
5376 | struct perf_output_handle handle; | |
5377 | struct perf_event_header header; | |
689802b2 | 5378 | |
927c7a9e FW |
5379 | /* protect the callchain buffers */ |
5380 | rcu_read_lock(); | |
5381 | ||
cdd6c482 | 5382 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 5383 | |
a7ac67ea | 5384 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 5385 | goto exit; |
0322cd6e | 5386 | |
cdd6c482 | 5387 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 5388 | |
8a057d84 | 5389 | perf_output_end(&handle); |
927c7a9e FW |
5390 | |
5391 | exit: | |
5392 | rcu_read_unlock(); | |
0322cd6e PZ |
5393 | } |
5394 | ||
38b200d6 | 5395 | /* |
cdd6c482 | 5396 | * read event_id |
38b200d6 PZ |
5397 | */ |
5398 | ||
5399 | struct perf_read_event { | |
5400 | struct perf_event_header header; | |
5401 | ||
5402 | u32 pid; | |
5403 | u32 tid; | |
38b200d6 PZ |
5404 | }; |
5405 | ||
5406 | static void | |
cdd6c482 | 5407 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
5408 | struct task_struct *task) |
5409 | { | |
5410 | struct perf_output_handle handle; | |
c980d109 | 5411 | struct perf_sample_data sample; |
dfc65094 | 5412 | struct perf_read_event read_event = { |
38b200d6 | 5413 | .header = { |
cdd6c482 | 5414 | .type = PERF_RECORD_READ, |
38b200d6 | 5415 | .misc = 0, |
c320c7b7 | 5416 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 5417 | }, |
cdd6c482 IM |
5418 | .pid = perf_event_pid(event, task), |
5419 | .tid = perf_event_tid(event, task), | |
38b200d6 | 5420 | }; |
3dab77fb | 5421 | int ret; |
38b200d6 | 5422 | |
c980d109 | 5423 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 5424 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
5425 | if (ret) |
5426 | return; | |
5427 | ||
dfc65094 | 5428 | perf_output_put(&handle, read_event); |
cdd6c482 | 5429 | perf_output_read(&handle, event); |
c980d109 | 5430 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 5431 | |
38b200d6 PZ |
5432 | perf_output_end(&handle); |
5433 | } | |
5434 | ||
52d857a8 JO |
5435 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
5436 | ||
5437 | static void | |
5438 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
5439 | perf_event_aux_output_cb output, |
5440 | void *data) | |
5441 | { | |
5442 | struct perf_event *event; | |
5443 | ||
5444 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5445 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
5446 | continue; | |
5447 | if (!event_filter_match(event)) | |
5448 | continue; | |
67516844 | 5449 | output(event, data); |
52d857a8 JO |
5450 | } |
5451 | } | |
5452 | ||
5453 | static void | |
67516844 | 5454 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
5455 | struct perf_event_context *task_ctx) |
5456 | { | |
5457 | struct perf_cpu_context *cpuctx; | |
5458 | struct perf_event_context *ctx; | |
5459 | struct pmu *pmu; | |
5460 | int ctxn; | |
5461 | ||
5462 | rcu_read_lock(); | |
5463 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
5464 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
5465 | if (cpuctx->unique_pmu != pmu) | |
5466 | goto next; | |
67516844 | 5467 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
5468 | if (task_ctx) |
5469 | goto next; | |
5470 | ctxn = pmu->task_ctx_nr; | |
5471 | if (ctxn < 0) | |
5472 | goto next; | |
5473 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
5474 | if (ctx) | |
67516844 | 5475 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
5476 | next: |
5477 | put_cpu_ptr(pmu->pmu_cpu_context); | |
5478 | } | |
5479 | ||
5480 | if (task_ctx) { | |
5481 | preempt_disable(); | |
67516844 | 5482 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
5483 | preempt_enable(); |
5484 | } | |
5485 | rcu_read_unlock(); | |
5486 | } | |
5487 | ||
60313ebe | 5488 | /* |
9f498cc5 PZ |
5489 | * task tracking -- fork/exit |
5490 | * | |
13d7a241 | 5491 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
5492 | */ |
5493 | ||
9f498cc5 | 5494 | struct perf_task_event { |
3a80b4a3 | 5495 | struct task_struct *task; |
cdd6c482 | 5496 | struct perf_event_context *task_ctx; |
60313ebe PZ |
5497 | |
5498 | struct { | |
5499 | struct perf_event_header header; | |
5500 | ||
5501 | u32 pid; | |
5502 | u32 ppid; | |
9f498cc5 PZ |
5503 | u32 tid; |
5504 | u32 ptid; | |
393b2ad8 | 5505 | u64 time; |
cdd6c482 | 5506 | } event_id; |
60313ebe PZ |
5507 | }; |
5508 | ||
67516844 JO |
5509 | static int perf_event_task_match(struct perf_event *event) |
5510 | { | |
13d7a241 SE |
5511 | return event->attr.comm || event->attr.mmap || |
5512 | event->attr.mmap2 || event->attr.mmap_data || | |
5513 | event->attr.task; | |
67516844 JO |
5514 | } |
5515 | ||
cdd6c482 | 5516 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 5517 | void *data) |
60313ebe | 5518 | { |
52d857a8 | 5519 | struct perf_task_event *task_event = data; |
60313ebe | 5520 | struct perf_output_handle handle; |
c980d109 | 5521 | struct perf_sample_data sample; |
9f498cc5 | 5522 | struct task_struct *task = task_event->task; |
c980d109 | 5523 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 5524 | |
67516844 JO |
5525 | if (!perf_event_task_match(event)) |
5526 | return; | |
5527 | ||
c980d109 | 5528 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 5529 | |
c980d109 | 5530 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 5531 | task_event->event_id.header.size); |
ef60777c | 5532 | if (ret) |
c980d109 | 5533 | goto out; |
60313ebe | 5534 | |
cdd6c482 IM |
5535 | task_event->event_id.pid = perf_event_pid(event, task); |
5536 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 5537 | |
cdd6c482 IM |
5538 | task_event->event_id.tid = perf_event_tid(event, task); |
5539 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 5540 | |
34f43927 PZ |
5541 | task_event->event_id.time = perf_event_clock(event); |
5542 | ||
cdd6c482 | 5543 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 5544 | |
c980d109 ACM |
5545 | perf_event__output_id_sample(event, &handle, &sample); |
5546 | ||
60313ebe | 5547 | perf_output_end(&handle); |
c980d109 ACM |
5548 | out: |
5549 | task_event->event_id.header.size = size; | |
60313ebe PZ |
5550 | } |
5551 | ||
cdd6c482 IM |
5552 | static void perf_event_task(struct task_struct *task, |
5553 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 5554 | int new) |
60313ebe | 5555 | { |
9f498cc5 | 5556 | struct perf_task_event task_event; |
60313ebe | 5557 | |
cdd6c482 IM |
5558 | if (!atomic_read(&nr_comm_events) && |
5559 | !atomic_read(&nr_mmap_events) && | |
5560 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
5561 | return; |
5562 | ||
9f498cc5 | 5563 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
5564 | .task = task, |
5565 | .task_ctx = task_ctx, | |
cdd6c482 | 5566 | .event_id = { |
60313ebe | 5567 | .header = { |
cdd6c482 | 5568 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 5569 | .misc = 0, |
cdd6c482 | 5570 | .size = sizeof(task_event.event_id), |
60313ebe | 5571 | }, |
573402db PZ |
5572 | /* .pid */ |
5573 | /* .ppid */ | |
9f498cc5 PZ |
5574 | /* .tid */ |
5575 | /* .ptid */ | |
34f43927 | 5576 | /* .time */ |
60313ebe PZ |
5577 | }, |
5578 | }; | |
5579 | ||
67516844 | 5580 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
5581 | &task_event, |
5582 | task_ctx); | |
9f498cc5 PZ |
5583 | } |
5584 | ||
cdd6c482 | 5585 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 5586 | { |
cdd6c482 | 5587 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
5588 | } |
5589 | ||
8d1b2d93 PZ |
5590 | /* |
5591 | * comm tracking | |
5592 | */ | |
5593 | ||
5594 | struct perf_comm_event { | |
22a4f650 IM |
5595 | struct task_struct *task; |
5596 | char *comm; | |
8d1b2d93 PZ |
5597 | int comm_size; |
5598 | ||
5599 | struct { | |
5600 | struct perf_event_header header; | |
5601 | ||
5602 | u32 pid; | |
5603 | u32 tid; | |
cdd6c482 | 5604 | } event_id; |
8d1b2d93 PZ |
5605 | }; |
5606 | ||
67516844 JO |
5607 | static int perf_event_comm_match(struct perf_event *event) |
5608 | { | |
5609 | return event->attr.comm; | |
5610 | } | |
5611 | ||
cdd6c482 | 5612 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 5613 | void *data) |
8d1b2d93 | 5614 | { |
52d857a8 | 5615 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 5616 | struct perf_output_handle handle; |
c980d109 | 5617 | struct perf_sample_data sample; |
cdd6c482 | 5618 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
5619 | int ret; |
5620 | ||
67516844 JO |
5621 | if (!perf_event_comm_match(event)) |
5622 | return; | |
5623 | ||
c980d109 ACM |
5624 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
5625 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5626 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
5627 | |
5628 | if (ret) | |
c980d109 | 5629 | goto out; |
8d1b2d93 | 5630 | |
cdd6c482 IM |
5631 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
5632 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 5633 | |
cdd6c482 | 5634 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 5635 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 5636 | comm_event->comm_size); |
c980d109 ACM |
5637 | |
5638 | perf_event__output_id_sample(event, &handle, &sample); | |
5639 | ||
8d1b2d93 | 5640 | perf_output_end(&handle); |
c980d109 ACM |
5641 | out: |
5642 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
5643 | } |
5644 | ||
cdd6c482 | 5645 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 5646 | { |
413ee3b4 | 5647 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 5648 | unsigned int size; |
8d1b2d93 | 5649 | |
413ee3b4 | 5650 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 5651 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 5652 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
5653 | |
5654 | comm_event->comm = comm; | |
5655 | comm_event->comm_size = size; | |
5656 | ||
cdd6c482 | 5657 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 5658 | |
67516844 | 5659 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
5660 | comm_event, |
5661 | NULL); | |
8d1b2d93 PZ |
5662 | } |
5663 | ||
82b89778 | 5664 | void perf_event_comm(struct task_struct *task, bool exec) |
8d1b2d93 | 5665 | { |
9ee318a7 PZ |
5666 | struct perf_comm_event comm_event; |
5667 | ||
cdd6c482 | 5668 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 5669 | return; |
a63eaf34 | 5670 | |
9ee318a7 | 5671 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 5672 | .task = task, |
573402db PZ |
5673 | /* .comm */ |
5674 | /* .comm_size */ | |
cdd6c482 | 5675 | .event_id = { |
573402db | 5676 | .header = { |
cdd6c482 | 5677 | .type = PERF_RECORD_COMM, |
82b89778 | 5678 | .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0, |
573402db PZ |
5679 | /* .size */ |
5680 | }, | |
5681 | /* .pid */ | |
5682 | /* .tid */ | |
8d1b2d93 PZ |
5683 | }, |
5684 | }; | |
5685 | ||
cdd6c482 | 5686 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5687 | } |
5688 | ||
0a4a9391 PZ |
5689 | /* |
5690 | * mmap tracking | |
5691 | */ | |
5692 | ||
5693 | struct perf_mmap_event { | |
089dd79d PZ |
5694 | struct vm_area_struct *vma; |
5695 | ||
5696 | const char *file_name; | |
5697 | int file_size; | |
13d7a241 SE |
5698 | int maj, min; |
5699 | u64 ino; | |
5700 | u64 ino_generation; | |
f972eb63 | 5701 | u32 prot, flags; |
0a4a9391 PZ |
5702 | |
5703 | struct { | |
5704 | struct perf_event_header header; | |
5705 | ||
5706 | u32 pid; | |
5707 | u32 tid; | |
5708 | u64 start; | |
5709 | u64 len; | |
5710 | u64 pgoff; | |
cdd6c482 | 5711 | } event_id; |
0a4a9391 PZ |
5712 | }; |
5713 | ||
67516844 JO |
5714 | static int perf_event_mmap_match(struct perf_event *event, |
5715 | void *data) | |
5716 | { | |
5717 | struct perf_mmap_event *mmap_event = data; | |
5718 | struct vm_area_struct *vma = mmap_event->vma; | |
5719 | int executable = vma->vm_flags & VM_EXEC; | |
5720 | ||
5721 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5722 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5723 | } |
5724 | ||
cdd6c482 | 5725 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5726 | void *data) |
0a4a9391 | 5727 | { |
52d857a8 | 5728 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5729 | struct perf_output_handle handle; |
c980d109 | 5730 | struct perf_sample_data sample; |
cdd6c482 | 5731 | int size = mmap_event->event_id.header.size; |
c980d109 | 5732 | int ret; |
0a4a9391 | 5733 | |
67516844 JO |
5734 | if (!perf_event_mmap_match(event, data)) |
5735 | return; | |
5736 | ||
13d7a241 SE |
5737 | if (event->attr.mmap2) { |
5738 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5739 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5740 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5741 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5742 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
f972eb63 PZ |
5743 | mmap_event->event_id.header.size += sizeof(mmap_event->prot); |
5744 | mmap_event->event_id.header.size += sizeof(mmap_event->flags); | |
13d7a241 SE |
5745 | } |
5746 | ||
c980d109 ACM |
5747 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5748 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5749 | mmap_event->event_id.header.size); |
0a4a9391 | 5750 | if (ret) |
c980d109 | 5751 | goto out; |
0a4a9391 | 5752 | |
cdd6c482 IM |
5753 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5754 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5755 | |
cdd6c482 | 5756 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5757 | |
5758 | if (event->attr.mmap2) { | |
5759 | perf_output_put(&handle, mmap_event->maj); | |
5760 | perf_output_put(&handle, mmap_event->min); | |
5761 | perf_output_put(&handle, mmap_event->ino); | |
5762 | perf_output_put(&handle, mmap_event->ino_generation); | |
f972eb63 PZ |
5763 | perf_output_put(&handle, mmap_event->prot); |
5764 | perf_output_put(&handle, mmap_event->flags); | |
13d7a241 SE |
5765 | } |
5766 | ||
76369139 | 5767 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5768 | mmap_event->file_size); |
c980d109 ACM |
5769 | |
5770 | perf_event__output_id_sample(event, &handle, &sample); | |
5771 | ||
78d613eb | 5772 | perf_output_end(&handle); |
c980d109 ACM |
5773 | out: |
5774 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5775 | } |
5776 | ||
cdd6c482 | 5777 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5778 | { |
089dd79d PZ |
5779 | struct vm_area_struct *vma = mmap_event->vma; |
5780 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5781 | int maj = 0, min = 0; |
5782 | u64 ino = 0, gen = 0; | |
f972eb63 | 5783 | u32 prot = 0, flags = 0; |
0a4a9391 PZ |
5784 | unsigned int size; |
5785 | char tmp[16]; | |
5786 | char *buf = NULL; | |
2c42cfbf | 5787 | char *name; |
413ee3b4 | 5788 | |
0a4a9391 | 5789 | if (file) { |
13d7a241 SE |
5790 | struct inode *inode; |
5791 | dev_t dev; | |
3ea2f2b9 | 5792 | |
2c42cfbf | 5793 | buf = kmalloc(PATH_MAX, GFP_KERNEL); |
0a4a9391 | 5794 | if (!buf) { |
c7e548b4 ON |
5795 | name = "//enomem"; |
5796 | goto cpy_name; | |
0a4a9391 | 5797 | } |
413ee3b4 | 5798 | /* |
3ea2f2b9 | 5799 | * d_path() works from the end of the rb backwards, so we |
413ee3b4 AB |
5800 | * need to add enough zero bytes after the string to handle |
5801 | * the 64bit alignment we do later. | |
5802 | */ | |
9bf39ab2 | 5803 | name = file_path(file, buf, PATH_MAX - sizeof(u64)); |
0a4a9391 | 5804 | if (IS_ERR(name)) { |
c7e548b4 ON |
5805 | name = "//toolong"; |
5806 | goto cpy_name; | |
0a4a9391 | 5807 | } |
13d7a241 SE |
5808 | inode = file_inode(vma->vm_file); |
5809 | dev = inode->i_sb->s_dev; | |
5810 | ino = inode->i_ino; | |
5811 | gen = inode->i_generation; | |
5812 | maj = MAJOR(dev); | |
5813 | min = MINOR(dev); | |
f972eb63 PZ |
5814 | |
5815 | if (vma->vm_flags & VM_READ) | |
5816 | prot |= PROT_READ; | |
5817 | if (vma->vm_flags & VM_WRITE) | |
5818 | prot |= PROT_WRITE; | |
5819 | if (vma->vm_flags & VM_EXEC) | |
5820 | prot |= PROT_EXEC; | |
5821 | ||
5822 | if (vma->vm_flags & VM_MAYSHARE) | |
5823 | flags = MAP_SHARED; | |
5824 | else | |
5825 | flags = MAP_PRIVATE; | |
5826 | ||
5827 | if (vma->vm_flags & VM_DENYWRITE) | |
5828 | flags |= MAP_DENYWRITE; | |
5829 | if (vma->vm_flags & VM_MAYEXEC) | |
5830 | flags |= MAP_EXECUTABLE; | |
5831 | if (vma->vm_flags & VM_LOCKED) | |
5832 | flags |= MAP_LOCKED; | |
5833 | if (vma->vm_flags & VM_HUGETLB) | |
5834 | flags |= MAP_HUGETLB; | |
5835 | ||
c7e548b4 | 5836 | goto got_name; |
0a4a9391 | 5837 | } else { |
fbe26abe JO |
5838 | if (vma->vm_ops && vma->vm_ops->name) { |
5839 | name = (char *) vma->vm_ops->name(vma); | |
5840 | if (name) | |
5841 | goto cpy_name; | |
5842 | } | |
5843 | ||
2c42cfbf | 5844 | name = (char *)arch_vma_name(vma); |
c7e548b4 ON |
5845 | if (name) |
5846 | goto cpy_name; | |
089dd79d | 5847 | |
32c5fb7e | 5848 | if (vma->vm_start <= vma->vm_mm->start_brk && |
3af9e859 | 5849 | vma->vm_end >= vma->vm_mm->brk) { |
c7e548b4 ON |
5850 | name = "[heap]"; |
5851 | goto cpy_name; | |
32c5fb7e ON |
5852 | } |
5853 | if (vma->vm_start <= vma->vm_mm->start_stack && | |
3af9e859 | 5854 | vma->vm_end >= vma->vm_mm->start_stack) { |
c7e548b4 ON |
5855 | name = "[stack]"; |
5856 | goto cpy_name; | |
089dd79d PZ |
5857 | } |
5858 | ||
c7e548b4 ON |
5859 | name = "//anon"; |
5860 | goto cpy_name; | |
0a4a9391 PZ |
5861 | } |
5862 | ||
c7e548b4 ON |
5863 | cpy_name: |
5864 | strlcpy(tmp, name, sizeof(tmp)); | |
5865 | name = tmp; | |
0a4a9391 | 5866 | got_name: |
2c42cfbf PZ |
5867 | /* |
5868 | * Since our buffer works in 8 byte units we need to align our string | |
5869 | * size to a multiple of 8. However, we must guarantee the tail end is | |
5870 | * zero'd out to avoid leaking random bits to userspace. | |
5871 | */ | |
5872 | size = strlen(name)+1; | |
5873 | while (!IS_ALIGNED(size, sizeof(u64))) | |
5874 | name[size++] = '\0'; | |
0a4a9391 PZ |
5875 | |
5876 | mmap_event->file_name = name; | |
5877 | mmap_event->file_size = size; | |
13d7a241 SE |
5878 | mmap_event->maj = maj; |
5879 | mmap_event->min = min; | |
5880 | mmap_event->ino = ino; | |
5881 | mmap_event->ino_generation = gen; | |
f972eb63 PZ |
5882 | mmap_event->prot = prot; |
5883 | mmap_event->flags = flags; | |
0a4a9391 | 5884 | |
2fe85427 SE |
5885 | if (!(vma->vm_flags & VM_EXEC)) |
5886 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5887 | ||
cdd6c482 | 5888 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5889 | |
67516844 | 5890 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5891 | mmap_event, |
5892 | NULL); | |
665c2142 | 5893 | |
0a4a9391 PZ |
5894 | kfree(buf); |
5895 | } | |
5896 | ||
3af9e859 | 5897 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5898 | { |
9ee318a7 PZ |
5899 | struct perf_mmap_event mmap_event; |
5900 | ||
cdd6c482 | 5901 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5902 | return; |
5903 | ||
5904 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5905 | .vma = vma, |
573402db PZ |
5906 | /* .file_name */ |
5907 | /* .file_size */ | |
cdd6c482 | 5908 | .event_id = { |
573402db | 5909 | .header = { |
cdd6c482 | 5910 | .type = PERF_RECORD_MMAP, |
39447b38 | 5911 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5912 | /* .size */ |
5913 | }, | |
5914 | /* .pid */ | |
5915 | /* .tid */ | |
089dd79d PZ |
5916 | .start = vma->vm_start, |
5917 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5918 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5919 | }, |
13d7a241 SE |
5920 | /* .maj (attr_mmap2 only) */ |
5921 | /* .min (attr_mmap2 only) */ | |
5922 | /* .ino (attr_mmap2 only) */ | |
5923 | /* .ino_generation (attr_mmap2 only) */ | |
f972eb63 PZ |
5924 | /* .prot (attr_mmap2 only) */ |
5925 | /* .flags (attr_mmap2 only) */ | |
0a4a9391 PZ |
5926 | }; |
5927 | ||
cdd6c482 | 5928 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5929 | } |
5930 | ||
68db7e98 AS |
5931 | void perf_event_aux_event(struct perf_event *event, unsigned long head, |
5932 | unsigned long size, u64 flags) | |
5933 | { | |
5934 | struct perf_output_handle handle; | |
5935 | struct perf_sample_data sample; | |
5936 | struct perf_aux_event { | |
5937 | struct perf_event_header header; | |
5938 | u64 offset; | |
5939 | u64 size; | |
5940 | u64 flags; | |
5941 | } rec = { | |
5942 | .header = { | |
5943 | .type = PERF_RECORD_AUX, | |
5944 | .misc = 0, | |
5945 | .size = sizeof(rec), | |
5946 | }, | |
5947 | .offset = head, | |
5948 | .size = size, | |
5949 | .flags = flags, | |
5950 | }; | |
5951 | int ret; | |
5952 | ||
5953 | perf_event_header__init_id(&rec.header, &sample, event); | |
5954 | ret = perf_output_begin(&handle, event, rec.header.size); | |
5955 | ||
5956 | if (ret) | |
5957 | return; | |
5958 | ||
5959 | perf_output_put(&handle, rec); | |
5960 | perf_event__output_id_sample(event, &handle, &sample); | |
5961 | ||
5962 | perf_output_end(&handle); | |
5963 | } | |
5964 | ||
f38b0dbb KL |
5965 | /* |
5966 | * Lost/dropped samples logging | |
5967 | */ | |
5968 | void perf_log_lost_samples(struct perf_event *event, u64 lost) | |
5969 | { | |
5970 | struct perf_output_handle handle; | |
5971 | struct perf_sample_data sample; | |
5972 | int ret; | |
5973 | ||
5974 | struct { | |
5975 | struct perf_event_header header; | |
5976 | u64 lost; | |
5977 | } lost_samples_event = { | |
5978 | .header = { | |
5979 | .type = PERF_RECORD_LOST_SAMPLES, | |
5980 | .misc = 0, | |
5981 | .size = sizeof(lost_samples_event), | |
5982 | }, | |
5983 | .lost = lost, | |
5984 | }; | |
5985 | ||
5986 | perf_event_header__init_id(&lost_samples_event.header, &sample, event); | |
5987 | ||
5988 | ret = perf_output_begin(&handle, event, | |
5989 | lost_samples_event.header.size); | |
5990 | if (ret) | |
5991 | return; | |
5992 | ||
5993 | perf_output_put(&handle, lost_samples_event); | |
5994 | perf_event__output_id_sample(event, &handle, &sample); | |
5995 | perf_output_end(&handle); | |
5996 | } | |
5997 | ||
45ac1403 AH |
5998 | /* |
5999 | * context_switch tracking | |
6000 | */ | |
6001 | ||
6002 | struct perf_switch_event { | |
6003 | struct task_struct *task; | |
6004 | struct task_struct *next_prev; | |
6005 | ||
6006 | struct { | |
6007 | struct perf_event_header header; | |
6008 | u32 next_prev_pid; | |
6009 | u32 next_prev_tid; | |
6010 | } event_id; | |
6011 | }; | |
6012 | ||
6013 | static int perf_event_switch_match(struct perf_event *event) | |
6014 | { | |
6015 | return event->attr.context_switch; | |
6016 | } | |
6017 | ||
6018 | static void perf_event_switch_output(struct perf_event *event, void *data) | |
6019 | { | |
6020 | struct perf_switch_event *se = data; | |
6021 | struct perf_output_handle handle; | |
6022 | struct perf_sample_data sample; | |
6023 | int ret; | |
6024 | ||
6025 | if (!perf_event_switch_match(event)) | |
6026 | return; | |
6027 | ||
6028 | /* Only CPU-wide events are allowed to see next/prev pid/tid */ | |
6029 | if (event->ctx->task) { | |
6030 | se->event_id.header.type = PERF_RECORD_SWITCH; | |
6031 | se->event_id.header.size = sizeof(se->event_id.header); | |
6032 | } else { | |
6033 | se->event_id.header.type = PERF_RECORD_SWITCH_CPU_WIDE; | |
6034 | se->event_id.header.size = sizeof(se->event_id); | |
6035 | se->event_id.next_prev_pid = | |
6036 | perf_event_pid(event, se->next_prev); | |
6037 | se->event_id.next_prev_tid = | |
6038 | perf_event_tid(event, se->next_prev); | |
6039 | } | |
6040 | ||
6041 | perf_event_header__init_id(&se->event_id.header, &sample, event); | |
6042 | ||
6043 | ret = perf_output_begin(&handle, event, se->event_id.header.size); | |
6044 | if (ret) | |
6045 | return; | |
6046 | ||
6047 | if (event->ctx->task) | |
6048 | perf_output_put(&handle, se->event_id.header); | |
6049 | else | |
6050 | perf_output_put(&handle, se->event_id); | |
6051 | ||
6052 | perf_event__output_id_sample(event, &handle, &sample); | |
6053 | ||
6054 | perf_output_end(&handle); | |
6055 | } | |
6056 | ||
6057 | static void perf_event_switch(struct task_struct *task, | |
6058 | struct task_struct *next_prev, bool sched_in) | |
6059 | { | |
6060 | struct perf_switch_event switch_event; | |
6061 | ||
6062 | /* N.B. caller checks nr_switch_events != 0 */ | |
6063 | ||
6064 | switch_event = (struct perf_switch_event){ | |
6065 | .task = task, | |
6066 | .next_prev = next_prev, | |
6067 | .event_id = { | |
6068 | .header = { | |
6069 | /* .type */ | |
6070 | .misc = sched_in ? 0 : PERF_RECORD_MISC_SWITCH_OUT, | |
6071 | /* .size */ | |
6072 | }, | |
6073 | /* .next_prev_pid */ | |
6074 | /* .next_prev_tid */ | |
6075 | }, | |
6076 | }; | |
6077 | ||
6078 | perf_event_aux(perf_event_switch_output, | |
6079 | &switch_event, | |
6080 | NULL); | |
6081 | } | |
6082 | ||
a78ac325 PZ |
6083 | /* |
6084 | * IRQ throttle logging | |
6085 | */ | |
6086 | ||
cdd6c482 | 6087 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
6088 | { |
6089 | struct perf_output_handle handle; | |
c980d109 | 6090 | struct perf_sample_data sample; |
a78ac325 PZ |
6091 | int ret; |
6092 | ||
6093 | struct { | |
6094 | struct perf_event_header header; | |
6095 | u64 time; | |
cca3f454 | 6096 | u64 id; |
7f453c24 | 6097 | u64 stream_id; |
a78ac325 PZ |
6098 | } throttle_event = { |
6099 | .header = { | |
cdd6c482 | 6100 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
6101 | .misc = 0, |
6102 | .size = sizeof(throttle_event), | |
6103 | }, | |
34f43927 | 6104 | .time = perf_event_clock(event), |
cdd6c482 IM |
6105 | .id = primary_event_id(event), |
6106 | .stream_id = event->id, | |
a78ac325 PZ |
6107 | }; |
6108 | ||
966ee4d6 | 6109 | if (enable) |
cdd6c482 | 6110 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 6111 | |
c980d109 ACM |
6112 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
6113 | ||
6114 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 6115 | throttle_event.header.size); |
a78ac325 PZ |
6116 | if (ret) |
6117 | return; | |
6118 | ||
6119 | perf_output_put(&handle, throttle_event); | |
c980d109 | 6120 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
6121 | perf_output_end(&handle); |
6122 | } | |
6123 | ||
ec0d7729 AS |
6124 | static void perf_log_itrace_start(struct perf_event *event) |
6125 | { | |
6126 | struct perf_output_handle handle; | |
6127 | struct perf_sample_data sample; | |
6128 | struct perf_aux_event { | |
6129 | struct perf_event_header header; | |
6130 | u32 pid; | |
6131 | u32 tid; | |
6132 | } rec; | |
6133 | int ret; | |
6134 | ||
6135 | if (event->parent) | |
6136 | event = event->parent; | |
6137 | ||
6138 | if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) || | |
6139 | event->hw.itrace_started) | |
6140 | return; | |
6141 | ||
ec0d7729 AS |
6142 | rec.header.type = PERF_RECORD_ITRACE_START; |
6143 | rec.header.misc = 0; | |
6144 | rec.header.size = sizeof(rec); | |
6145 | rec.pid = perf_event_pid(event, current); | |
6146 | rec.tid = perf_event_tid(event, current); | |
6147 | ||
6148 | perf_event_header__init_id(&rec.header, &sample, event); | |
6149 | ret = perf_output_begin(&handle, event, rec.header.size); | |
6150 | ||
6151 | if (ret) | |
6152 | return; | |
6153 | ||
6154 | perf_output_put(&handle, rec); | |
6155 | perf_event__output_id_sample(event, &handle, &sample); | |
6156 | ||
6157 | perf_output_end(&handle); | |
6158 | } | |
6159 | ||
f6c7d5fe | 6160 | /* |
cdd6c482 | 6161 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
6162 | */ |
6163 | ||
a8b0ca17 | 6164 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
6165 | int throttle, struct perf_sample_data *data, |
6166 | struct pt_regs *regs) | |
f6c7d5fe | 6167 | { |
cdd6c482 IM |
6168 | int events = atomic_read(&event->event_limit); |
6169 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 6170 | u64 seq; |
79f14641 PZ |
6171 | int ret = 0; |
6172 | ||
96398826 PZ |
6173 | /* |
6174 | * Non-sampling counters might still use the PMI to fold short | |
6175 | * hardware counters, ignore those. | |
6176 | */ | |
6177 | if (unlikely(!is_sampling_event(event))) | |
6178 | return 0; | |
6179 | ||
e050e3f0 SE |
6180 | seq = __this_cpu_read(perf_throttled_seq); |
6181 | if (seq != hwc->interrupts_seq) { | |
6182 | hwc->interrupts_seq = seq; | |
6183 | hwc->interrupts = 1; | |
6184 | } else { | |
6185 | hwc->interrupts++; | |
6186 | if (unlikely(throttle | |
6187 | && hwc->interrupts >= max_samples_per_tick)) { | |
6188 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
6189 | hwc->interrupts = MAX_INTERRUPTS; |
6190 | perf_log_throttle(event, 0); | |
d84153d6 | 6191 | tick_nohz_full_kick(); |
a78ac325 PZ |
6192 | ret = 1; |
6193 | } | |
e050e3f0 | 6194 | } |
60db5e09 | 6195 | |
cdd6c482 | 6196 | if (event->attr.freq) { |
def0a9b2 | 6197 | u64 now = perf_clock(); |
abd50713 | 6198 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 6199 | |
abd50713 | 6200 | hwc->freq_time_stamp = now; |
bd2b5b12 | 6201 | |
abd50713 | 6202 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 6203 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
6204 | } |
6205 | ||
2023b359 PZ |
6206 | /* |
6207 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 6208 | * events |
2023b359 PZ |
6209 | */ |
6210 | ||
cdd6c482 IM |
6211 | event->pending_kill = POLL_IN; |
6212 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 6213 | ret = 1; |
cdd6c482 | 6214 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
6215 | event->pending_disable = 1; |
6216 | irq_work_queue(&event->pending); | |
79f14641 PZ |
6217 | } |
6218 | ||
453f19ee | 6219 | if (event->overflow_handler) |
a8b0ca17 | 6220 | event->overflow_handler(event, data, regs); |
453f19ee | 6221 | else |
a8b0ca17 | 6222 | perf_event_output(event, data, regs); |
453f19ee | 6223 | |
f506b3dc | 6224 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
6225 | event->pending_wakeup = 1; |
6226 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
6227 | } |
6228 | ||
79f14641 | 6229 | return ret; |
f6c7d5fe PZ |
6230 | } |
6231 | ||
a8b0ca17 | 6232 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
6233 | struct perf_sample_data *data, |
6234 | struct pt_regs *regs) | |
850bc73f | 6235 | { |
a8b0ca17 | 6236 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
6237 | } |
6238 | ||
15dbf27c | 6239 | /* |
cdd6c482 | 6240 | * Generic software event infrastructure |
15dbf27c PZ |
6241 | */ |
6242 | ||
b28ab83c PZ |
6243 | struct swevent_htable { |
6244 | struct swevent_hlist *swevent_hlist; | |
6245 | struct mutex hlist_mutex; | |
6246 | int hlist_refcount; | |
6247 | ||
6248 | /* Recursion avoidance in each contexts */ | |
6249 | int recursion[PERF_NR_CONTEXTS]; | |
39af6b16 JO |
6250 | |
6251 | /* Keeps track of cpu being initialized/exited */ | |
6252 | bool online; | |
b28ab83c PZ |
6253 | }; |
6254 | ||
6255 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
6256 | ||
7b4b6658 | 6257 | /* |
cdd6c482 IM |
6258 | * We directly increment event->count and keep a second value in |
6259 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
6260 | * is kept in the range [-sample_period, 0] so that we can use the |
6261 | * sign as trigger. | |
6262 | */ | |
6263 | ||
ab573844 | 6264 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 6265 | { |
cdd6c482 | 6266 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
6267 | u64 period = hwc->last_period; |
6268 | u64 nr, offset; | |
6269 | s64 old, val; | |
6270 | ||
6271 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
6272 | |
6273 | again: | |
e7850595 | 6274 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
6275 | if (val < 0) |
6276 | return 0; | |
15dbf27c | 6277 | |
7b4b6658 PZ |
6278 | nr = div64_u64(period + val, period); |
6279 | offset = nr * period; | |
6280 | val -= offset; | |
e7850595 | 6281 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 6282 | goto again; |
15dbf27c | 6283 | |
7b4b6658 | 6284 | return nr; |
15dbf27c PZ |
6285 | } |
6286 | ||
0cff784a | 6287 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 6288 | struct perf_sample_data *data, |
5622f295 | 6289 | struct pt_regs *regs) |
15dbf27c | 6290 | { |
cdd6c482 | 6291 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 6292 | int throttle = 0; |
15dbf27c | 6293 | |
0cff784a PZ |
6294 | if (!overflow) |
6295 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 6296 | |
7b4b6658 PZ |
6297 | if (hwc->interrupts == MAX_INTERRUPTS) |
6298 | return; | |
15dbf27c | 6299 | |
7b4b6658 | 6300 | for (; overflow; overflow--) { |
a8b0ca17 | 6301 | if (__perf_event_overflow(event, throttle, |
5622f295 | 6302 | data, regs)) { |
7b4b6658 PZ |
6303 | /* |
6304 | * We inhibit the overflow from happening when | |
6305 | * hwc->interrupts == MAX_INTERRUPTS. | |
6306 | */ | |
6307 | break; | |
6308 | } | |
cf450a73 | 6309 | throttle = 1; |
7b4b6658 | 6310 | } |
15dbf27c PZ |
6311 | } |
6312 | ||
a4eaf7f1 | 6313 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 6314 | struct perf_sample_data *data, |
5622f295 | 6315 | struct pt_regs *regs) |
7b4b6658 | 6316 | { |
cdd6c482 | 6317 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 6318 | |
e7850595 | 6319 | local64_add(nr, &event->count); |
d6d020e9 | 6320 | |
0cff784a PZ |
6321 | if (!regs) |
6322 | return; | |
6323 | ||
6c7e550f | 6324 | if (!is_sampling_event(event)) |
7b4b6658 | 6325 | return; |
d6d020e9 | 6326 | |
5d81e5cf AV |
6327 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
6328 | data->period = nr; | |
6329 | return perf_swevent_overflow(event, 1, data, regs); | |
6330 | } else | |
6331 | data->period = event->hw.last_period; | |
6332 | ||
0cff784a | 6333 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 6334 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 6335 | |
e7850595 | 6336 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 6337 | return; |
df1a132b | 6338 | |
a8b0ca17 | 6339 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
6340 | } |
6341 | ||
f5ffe02e FW |
6342 | static int perf_exclude_event(struct perf_event *event, |
6343 | struct pt_regs *regs) | |
6344 | { | |
a4eaf7f1 | 6345 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 6346 | return 1; |
a4eaf7f1 | 6347 | |
f5ffe02e FW |
6348 | if (regs) { |
6349 | if (event->attr.exclude_user && user_mode(regs)) | |
6350 | return 1; | |
6351 | ||
6352 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
6353 | return 1; | |
6354 | } | |
6355 | ||
6356 | return 0; | |
6357 | } | |
6358 | ||
cdd6c482 | 6359 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 6360 | enum perf_type_id type, |
6fb2915d LZ |
6361 | u32 event_id, |
6362 | struct perf_sample_data *data, | |
6363 | struct pt_regs *regs) | |
15dbf27c | 6364 | { |
cdd6c482 | 6365 | if (event->attr.type != type) |
a21ca2ca | 6366 | return 0; |
f5ffe02e | 6367 | |
cdd6c482 | 6368 | if (event->attr.config != event_id) |
15dbf27c PZ |
6369 | return 0; |
6370 | ||
f5ffe02e FW |
6371 | if (perf_exclude_event(event, regs)) |
6372 | return 0; | |
15dbf27c PZ |
6373 | |
6374 | return 1; | |
6375 | } | |
6376 | ||
76e1d904 FW |
6377 | static inline u64 swevent_hash(u64 type, u32 event_id) |
6378 | { | |
6379 | u64 val = event_id | (type << 32); | |
6380 | ||
6381 | return hash_64(val, SWEVENT_HLIST_BITS); | |
6382 | } | |
6383 | ||
49f135ed FW |
6384 | static inline struct hlist_head * |
6385 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 6386 | { |
49f135ed FW |
6387 | u64 hash = swevent_hash(type, event_id); |
6388 | ||
6389 | return &hlist->heads[hash]; | |
6390 | } | |
76e1d904 | 6391 | |
49f135ed FW |
6392 | /* For the read side: events when they trigger */ |
6393 | static inline struct hlist_head * | |
b28ab83c | 6394 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
6395 | { |
6396 | struct swevent_hlist *hlist; | |
76e1d904 | 6397 | |
b28ab83c | 6398 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
6399 | if (!hlist) |
6400 | return NULL; | |
6401 | ||
49f135ed FW |
6402 | return __find_swevent_head(hlist, type, event_id); |
6403 | } | |
6404 | ||
6405 | /* For the event head insertion and removal in the hlist */ | |
6406 | static inline struct hlist_head * | |
b28ab83c | 6407 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
6408 | { |
6409 | struct swevent_hlist *hlist; | |
6410 | u32 event_id = event->attr.config; | |
6411 | u64 type = event->attr.type; | |
6412 | ||
6413 | /* | |
6414 | * Event scheduling is always serialized against hlist allocation | |
6415 | * and release. Which makes the protected version suitable here. | |
6416 | * The context lock guarantees that. | |
6417 | */ | |
b28ab83c | 6418 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
6419 | lockdep_is_held(&event->ctx->lock)); |
6420 | if (!hlist) | |
6421 | return NULL; | |
6422 | ||
6423 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
6424 | } |
6425 | ||
6426 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 6427 | u64 nr, |
76e1d904 FW |
6428 | struct perf_sample_data *data, |
6429 | struct pt_regs *regs) | |
15dbf27c | 6430 | { |
4a32fea9 | 6431 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6432 | struct perf_event *event; |
76e1d904 | 6433 | struct hlist_head *head; |
15dbf27c | 6434 | |
76e1d904 | 6435 | rcu_read_lock(); |
b28ab83c | 6436 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
6437 | if (!head) |
6438 | goto end; | |
6439 | ||
b67bfe0d | 6440 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 6441 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 6442 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 6443 | } |
76e1d904 FW |
6444 | end: |
6445 | rcu_read_unlock(); | |
15dbf27c PZ |
6446 | } |
6447 | ||
86038c5e PZI |
6448 | DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]); |
6449 | ||
4ed7c92d | 6450 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 6451 | { |
4a32fea9 | 6452 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
96f6d444 | 6453 | |
b28ab83c | 6454 | return get_recursion_context(swhash->recursion); |
96f6d444 | 6455 | } |
645e8cc0 | 6456 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 6457 | |
fa9f90be | 6458 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 6459 | { |
4a32fea9 | 6460 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
927c7a9e | 6461 | |
b28ab83c | 6462 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 6463 | } |
15dbf27c | 6464 | |
86038c5e | 6465 | void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 6466 | { |
a4234bfc | 6467 | struct perf_sample_data data; |
4ed7c92d | 6468 | |
86038c5e | 6469 | if (WARN_ON_ONCE(!regs)) |
4ed7c92d | 6470 | return; |
a4234bfc | 6471 | |
fd0d000b | 6472 | perf_sample_data_init(&data, addr, 0); |
a8b0ca17 | 6473 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
86038c5e PZI |
6474 | } |
6475 | ||
6476 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) | |
6477 | { | |
6478 | int rctx; | |
6479 | ||
6480 | preempt_disable_notrace(); | |
6481 | rctx = perf_swevent_get_recursion_context(); | |
6482 | if (unlikely(rctx < 0)) | |
6483 | goto fail; | |
6484 | ||
6485 | ___perf_sw_event(event_id, nr, regs, addr); | |
4ed7c92d PZ |
6486 | |
6487 | perf_swevent_put_recursion_context(rctx); | |
86038c5e | 6488 | fail: |
1c024eca | 6489 | preempt_enable_notrace(); |
b8e83514 PZ |
6490 | } |
6491 | ||
cdd6c482 | 6492 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 6493 | { |
15dbf27c PZ |
6494 | } |
6495 | ||
a4eaf7f1 | 6496 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 6497 | { |
4a32fea9 | 6498 | struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable); |
cdd6c482 | 6499 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
6500 | struct hlist_head *head; |
6501 | ||
6c7e550f | 6502 | if (is_sampling_event(event)) { |
7b4b6658 | 6503 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 6504 | perf_swevent_set_period(event); |
7b4b6658 | 6505 | } |
76e1d904 | 6506 | |
a4eaf7f1 PZ |
6507 | hwc->state = !(flags & PERF_EF_START); |
6508 | ||
b28ab83c | 6509 | head = find_swevent_head(swhash, event); |
39af6b16 JO |
6510 | if (!head) { |
6511 | /* | |
6512 | * We can race with cpu hotplug code. Do not | |
6513 | * WARN if the cpu just got unplugged. | |
6514 | */ | |
6515 | WARN_ON_ONCE(swhash->online); | |
76e1d904 | 6516 | return -EINVAL; |
39af6b16 | 6517 | } |
76e1d904 FW |
6518 | |
6519 | hlist_add_head_rcu(&event->hlist_entry, head); | |
6a694a60 | 6520 | perf_event_update_userpage(event); |
76e1d904 | 6521 | |
15dbf27c PZ |
6522 | return 0; |
6523 | } | |
6524 | ||
a4eaf7f1 | 6525 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 6526 | { |
76e1d904 | 6527 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
6528 | } |
6529 | ||
a4eaf7f1 | 6530 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 6531 | { |
a4eaf7f1 | 6532 | event->hw.state = 0; |
d6d020e9 | 6533 | } |
aa9c4c0f | 6534 | |
a4eaf7f1 | 6535 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 6536 | { |
a4eaf7f1 | 6537 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
6538 | } |
6539 | ||
49f135ed FW |
6540 | /* Deref the hlist from the update side */ |
6541 | static inline struct swevent_hlist * | |
b28ab83c | 6542 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 6543 | { |
b28ab83c PZ |
6544 | return rcu_dereference_protected(swhash->swevent_hlist, |
6545 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
6546 | } |
6547 | ||
b28ab83c | 6548 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 6549 | { |
b28ab83c | 6550 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 6551 | |
49f135ed | 6552 | if (!hlist) |
76e1d904 FW |
6553 | return; |
6554 | ||
70691d4a | 6555 | RCU_INIT_POINTER(swhash->swevent_hlist, NULL); |
fa4bbc4c | 6556 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
6557 | } |
6558 | ||
6559 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
6560 | { | |
b28ab83c | 6561 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 6562 | |
b28ab83c | 6563 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6564 | |
b28ab83c PZ |
6565 | if (!--swhash->hlist_refcount) |
6566 | swevent_hlist_release(swhash); | |
76e1d904 | 6567 | |
b28ab83c | 6568 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6569 | } |
6570 | ||
6571 | static void swevent_hlist_put(struct perf_event *event) | |
6572 | { | |
6573 | int cpu; | |
6574 | ||
76e1d904 FW |
6575 | for_each_possible_cpu(cpu) |
6576 | swevent_hlist_put_cpu(event, cpu); | |
6577 | } | |
6578 | ||
6579 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
6580 | { | |
b28ab83c | 6581 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
6582 | int err = 0; |
6583 | ||
b28ab83c | 6584 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 6585 | |
b28ab83c | 6586 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
6587 | struct swevent_hlist *hlist; |
6588 | ||
6589 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
6590 | if (!hlist) { | |
6591 | err = -ENOMEM; | |
6592 | goto exit; | |
6593 | } | |
b28ab83c | 6594 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 6595 | } |
b28ab83c | 6596 | swhash->hlist_refcount++; |
9ed6060d | 6597 | exit: |
b28ab83c | 6598 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
6599 | |
6600 | return err; | |
6601 | } | |
6602 | ||
6603 | static int swevent_hlist_get(struct perf_event *event) | |
6604 | { | |
6605 | int err; | |
6606 | int cpu, failed_cpu; | |
6607 | ||
76e1d904 FW |
6608 | get_online_cpus(); |
6609 | for_each_possible_cpu(cpu) { | |
6610 | err = swevent_hlist_get_cpu(event, cpu); | |
6611 | if (err) { | |
6612 | failed_cpu = cpu; | |
6613 | goto fail; | |
6614 | } | |
6615 | } | |
6616 | put_online_cpus(); | |
6617 | ||
6618 | return 0; | |
9ed6060d | 6619 | fail: |
76e1d904 FW |
6620 | for_each_possible_cpu(cpu) { |
6621 | if (cpu == failed_cpu) | |
6622 | break; | |
6623 | swevent_hlist_put_cpu(event, cpu); | |
6624 | } | |
6625 | ||
6626 | put_online_cpus(); | |
6627 | return err; | |
6628 | } | |
6629 | ||
c5905afb | 6630 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 6631 | |
b0a873eb PZ |
6632 | static void sw_perf_event_destroy(struct perf_event *event) |
6633 | { | |
6634 | u64 event_id = event->attr.config; | |
95476b64 | 6635 | |
b0a873eb PZ |
6636 | WARN_ON(event->parent); |
6637 | ||
c5905afb | 6638 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6639 | swevent_hlist_put(event); |
6640 | } | |
6641 | ||
6642 | static int perf_swevent_init(struct perf_event *event) | |
6643 | { | |
8176cced | 6644 | u64 event_id = event->attr.config; |
b0a873eb PZ |
6645 | |
6646 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6647 | return -ENOENT; | |
6648 | ||
2481c5fa SE |
6649 | /* |
6650 | * no branch sampling for software events | |
6651 | */ | |
6652 | if (has_branch_stack(event)) | |
6653 | return -EOPNOTSUPP; | |
6654 | ||
b0a873eb PZ |
6655 | switch (event_id) { |
6656 | case PERF_COUNT_SW_CPU_CLOCK: | |
6657 | case PERF_COUNT_SW_TASK_CLOCK: | |
6658 | return -ENOENT; | |
6659 | ||
6660 | default: | |
6661 | break; | |
6662 | } | |
6663 | ||
ce677831 | 6664 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
6665 | return -ENOENT; |
6666 | ||
6667 | if (!event->parent) { | |
6668 | int err; | |
6669 | ||
6670 | err = swevent_hlist_get(event); | |
6671 | if (err) | |
6672 | return err; | |
6673 | ||
c5905afb | 6674 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
6675 | event->destroy = sw_perf_event_destroy; |
6676 | } | |
6677 | ||
6678 | return 0; | |
6679 | } | |
6680 | ||
6681 | static struct pmu perf_swevent = { | |
89a1e187 | 6682 | .task_ctx_nr = perf_sw_context, |
95476b64 | 6683 | |
34f43927 PZ |
6684 | .capabilities = PERF_PMU_CAP_NO_NMI, |
6685 | ||
b0a873eb | 6686 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
6687 | .add = perf_swevent_add, |
6688 | .del = perf_swevent_del, | |
6689 | .start = perf_swevent_start, | |
6690 | .stop = perf_swevent_stop, | |
1c024eca | 6691 | .read = perf_swevent_read, |
1c024eca PZ |
6692 | }; |
6693 | ||
b0a873eb PZ |
6694 | #ifdef CONFIG_EVENT_TRACING |
6695 | ||
1c024eca PZ |
6696 | static int perf_tp_filter_match(struct perf_event *event, |
6697 | struct perf_sample_data *data) | |
6698 | { | |
6699 | void *record = data->raw->data; | |
6700 | ||
6701 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
6702 | return 1; | |
6703 | return 0; | |
6704 | } | |
6705 | ||
6706 | static int perf_tp_event_match(struct perf_event *event, | |
6707 | struct perf_sample_data *data, | |
6708 | struct pt_regs *regs) | |
6709 | { | |
a0f7d0f7 FW |
6710 | if (event->hw.state & PERF_HES_STOPPED) |
6711 | return 0; | |
580d607c PZ |
6712 | /* |
6713 | * All tracepoints are from kernel-space. | |
6714 | */ | |
6715 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
6716 | return 0; |
6717 | ||
6718 | if (!perf_tp_filter_match(event, data)) | |
6719 | return 0; | |
6720 | ||
6721 | return 1; | |
6722 | } | |
6723 | ||
6724 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
6725 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
6726 | struct task_struct *task) | |
95476b64 FW |
6727 | { |
6728 | struct perf_sample_data data; | |
1c024eca | 6729 | struct perf_event *event; |
1c024eca | 6730 | |
95476b64 FW |
6731 | struct perf_raw_record raw = { |
6732 | .size = entry_size, | |
6733 | .data = record, | |
6734 | }; | |
6735 | ||
fd0d000b | 6736 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
6737 | data.raw = &raw; |
6738 | ||
b67bfe0d | 6739 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 6740 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 6741 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 6742 | } |
ecc55f84 | 6743 | |
e6dab5ff AV |
6744 | /* |
6745 | * If we got specified a target task, also iterate its context and | |
6746 | * deliver this event there too. | |
6747 | */ | |
6748 | if (task && task != current) { | |
6749 | struct perf_event_context *ctx; | |
6750 | struct trace_entry *entry = record; | |
6751 | ||
6752 | rcu_read_lock(); | |
6753 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
6754 | if (!ctx) | |
6755 | goto unlock; | |
6756 | ||
6757 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
6758 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6759 | continue; | |
6760 | if (event->attr.config != entry->type) | |
6761 | continue; | |
6762 | if (perf_tp_event_match(event, &data, regs)) | |
6763 | perf_swevent_event(event, count, &data, regs); | |
6764 | } | |
6765 | unlock: | |
6766 | rcu_read_unlock(); | |
6767 | } | |
6768 | ||
ecc55f84 | 6769 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
6770 | } |
6771 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
6772 | ||
cdd6c482 | 6773 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 6774 | { |
1c024eca | 6775 | perf_trace_destroy(event); |
e077df4f PZ |
6776 | } |
6777 | ||
b0a873eb | 6778 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 6779 | { |
76e1d904 FW |
6780 | int err; |
6781 | ||
b0a873eb PZ |
6782 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
6783 | return -ENOENT; | |
6784 | ||
2481c5fa SE |
6785 | /* |
6786 | * no branch sampling for tracepoint events | |
6787 | */ | |
6788 | if (has_branch_stack(event)) | |
6789 | return -EOPNOTSUPP; | |
6790 | ||
1c024eca PZ |
6791 | err = perf_trace_init(event); |
6792 | if (err) | |
b0a873eb | 6793 | return err; |
e077df4f | 6794 | |
cdd6c482 | 6795 | event->destroy = tp_perf_event_destroy; |
e077df4f | 6796 | |
b0a873eb PZ |
6797 | return 0; |
6798 | } | |
6799 | ||
6800 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
6801 | .task_ctx_nr = perf_sw_context, |
6802 | ||
b0a873eb | 6803 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
6804 | .add = perf_trace_add, |
6805 | .del = perf_trace_del, | |
6806 | .start = perf_swevent_start, | |
6807 | .stop = perf_swevent_stop, | |
b0a873eb | 6808 | .read = perf_swevent_read, |
b0a873eb PZ |
6809 | }; |
6810 | ||
6811 | static inline void perf_tp_register(void) | |
6812 | { | |
2e80a82a | 6813 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 6814 | } |
6fb2915d LZ |
6815 | |
6816 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6817 | { | |
6818 | char *filter_str; | |
6819 | int ret; | |
6820 | ||
6821 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6822 | return -EINVAL; | |
6823 | ||
6824 | filter_str = strndup_user(arg, PAGE_SIZE); | |
6825 | if (IS_ERR(filter_str)) | |
6826 | return PTR_ERR(filter_str); | |
6827 | ||
6828 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
6829 | ||
6830 | kfree(filter_str); | |
6831 | return ret; | |
6832 | } | |
6833 | ||
6834 | static void perf_event_free_filter(struct perf_event *event) | |
6835 | { | |
6836 | ftrace_profile_free_filter(event); | |
6837 | } | |
6838 | ||
2541517c AS |
6839 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd) |
6840 | { | |
6841 | struct bpf_prog *prog; | |
6842 | ||
6843 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
6844 | return -EINVAL; | |
6845 | ||
6846 | if (event->tp_event->prog) | |
6847 | return -EEXIST; | |
6848 | ||
6849 | if (!(event->tp_event->flags & TRACE_EVENT_FL_KPROBE)) | |
6850 | /* bpf programs can only be attached to kprobes */ | |
6851 | return -EINVAL; | |
6852 | ||
6853 | prog = bpf_prog_get(prog_fd); | |
6854 | if (IS_ERR(prog)) | |
6855 | return PTR_ERR(prog); | |
6856 | ||
6c373ca8 | 6857 | if (prog->type != BPF_PROG_TYPE_KPROBE) { |
2541517c AS |
6858 | /* valid fd, but invalid bpf program type */ |
6859 | bpf_prog_put(prog); | |
6860 | return -EINVAL; | |
6861 | } | |
6862 | ||
6863 | event->tp_event->prog = prog; | |
6864 | ||
6865 | return 0; | |
6866 | } | |
6867 | ||
6868 | static void perf_event_free_bpf_prog(struct perf_event *event) | |
6869 | { | |
6870 | struct bpf_prog *prog; | |
6871 | ||
6872 | if (!event->tp_event) | |
6873 | return; | |
6874 | ||
6875 | prog = event->tp_event->prog; | |
6876 | if (prog) { | |
6877 | event->tp_event->prog = NULL; | |
6878 | bpf_prog_put(prog); | |
6879 | } | |
6880 | } | |
6881 | ||
e077df4f | 6882 | #else |
6fb2915d | 6883 | |
b0a873eb | 6884 | static inline void perf_tp_register(void) |
e077df4f | 6885 | { |
e077df4f | 6886 | } |
6fb2915d LZ |
6887 | |
6888 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
6889 | { | |
6890 | return -ENOENT; | |
6891 | } | |
6892 | ||
6893 | static void perf_event_free_filter(struct perf_event *event) | |
6894 | { | |
6895 | } | |
6896 | ||
2541517c AS |
6897 | static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd) |
6898 | { | |
6899 | return -ENOENT; | |
6900 | } | |
6901 | ||
6902 | static void perf_event_free_bpf_prog(struct perf_event *event) | |
6903 | { | |
6904 | } | |
07b139c8 | 6905 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 6906 | |
24f1e32c | 6907 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 6908 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 6909 | { |
f5ffe02e FW |
6910 | struct perf_sample_data sample; |
6911 | struct pt_regs *regs = data; | |
6912 | ||
fd0d000b | 6913 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 6914 | |
a4eaf7f1 | 6915 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 6916 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
6917 | } |
6918 | #endif | |
6919 | ||
b0a873eb PZ |
6920 | /* |
6921 | * hrtimer based swevent callback | |
6922 | */ | |
f29ac756 | 6923 | |
b0a873eb | 6924 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 6925 | { |
b0a873eb PZ |
6926 | enum hrtimer_restart ret = HRTIMER_RESTART; |
6927 | struct perf_sample_data data; | |
6928 | struct pt_regs *regs; | |
6929 | struct perf_event *event; | |
6930 | u64 period; | |
f29ac756 | 6931 | |
b0a873eb | 6932 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
6933 | |
6934 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
6935 | return HRTIMER_NORESTART; | |
6936 | ||
b0a873eb | 6937 | event->pmu->read(event); |
f344011c | 6938 | |
fd0d000b | 6939 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
6940 | regs = get_irq_regs(); |
6941 | ||
6942 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 6943 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 6944 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
6945 | ret = HRTIMER_NORESTART; |
6946 | } | |
24f1e32c | 6947 | |
b0a873eb PZ |
6948 | period = max_t(u64, 10000, event->hw.sample_period); |
6949 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 6950 | |
b0a873eb | 6951 | return ret; |
f29ac756 PZ |
6952 | } |
6953 | ||
b0a873eb | 6954 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 6955 | { |
b0a873eb | 6956 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
6957 | s64 period; |
6958 | ||
6959 | if (!is_sampling_event(event)) | |
6960 | return; | |
f5ffe02e | 6961 | |
5d508e82 FBH |
6962 | period = local64_read(&hwc->period_left); |
6963 | if (period) { | |
6964 | if (period < 0) | |
6965 | period = 10000; | |
fa407f35 | 6966 | |
5d508e82 FBH |
6967 | local64_set(&hwc->period_left, 0); |
6968 | } else { | |
6969 | period = max_t(u64, 10000, hwc->sample_period); | |
6970 | } | |
3497d206 TG |
6971 | hrtimer_start(&hwc->hrtimer, ns_to_ktime(period), |
6972 | HRTIMER_MODE_REL_PINNED); | |
24f1e32c | 6973 | } |
b0a873eb PZ |
6974 | |
6975 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6976 | { |
b0a873eb PZ |
6977 | struct hw_perf_event *hwc = &event->hw; |
6978 | ||
6c7e550f | 6979 | if (is_sampling_event(event)) { |
b0a873eb | 6980 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6981 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6982 | |
6983 | hrtimer_cancel(&hwc->hrtimer); | |
6984 | } | |
24f1e32c FW |
6985 | } |
6986 | ||
ba3dd36c PZ |
6987 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6988 | { | |
6989 | struct hw_perf_event *hwc = &event->hw; | |
6990 | ||
6991 | if (!is_sampling_event(event)) | |
6992 | return; | |
6993 | ||
6994 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6995 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6996 | ||
6997 | /* | |
6998 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6999 | * mapping and avoid the whole period adjust feedback stuff. | |
7000 | */ | |
7001 | if (event->attr.freq) { | |
7002 | long freq = event->attr.sample_freq; | |
7003 | ||
7004 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
7005 | hwc->sample_period = event->attr.sample_period; | |
7006 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 7007 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
7008 | event->attr.freq = 0; |
7009 | } | |
7010 | } | |
7011 | ||
b0a873eb PZ |
7012 | /* |
7013 | * Software event: cpu wall time clock | |
7014 | */ | |
7015 | ||
7016 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 7017 | { |
b0a873eb PZ |
7018 | s64 prev; |
7019 | u64 now; | |
7020 | ||
a4eaf7f1 | 7021 | now = local_clock(); |
b0a873eb PZ |
7022 | prev = local64_xchg(&event->hw.prev_count, now); |
7023 | local64_add(now - prev, &event->count); | |
24f1e32c | 7024 | } |
24f1e32c | 7025 | |
a4eaf7f1 | 7026 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 7027 | { |
a4eaf7f1 | 7028 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 7029 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
7030 | } |
7031 | ||
a4eaf7f1 | 7032 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 7033 | { |
b0a873eb PZ |
7034 | perf_swevent_cancel_hrtimer(event); |
7035 | cpu_clock_event_update(event); | |
7036 | } | |
f29ac756 | 7037 | |
a4eaf7f1 PZ |
7038 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
7039 | { | |
7040 | if (flags & PERF_EF_START) | |
7041 | cpu_clock_event_start(event, flags); | |
6a694a60 | 7042 | perf_event_update_userpage(event); |
a4eaf7f1 PZ |
7043 | |
7044 | return 0; | |
7045 | } | |
7046 | ||
7047 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
7048 | { | |
7049 | cpu_clock_event_stop(event, flags); | |
7050 | } | |
7051 | ||
b0a873eb PZ |
7052 | static void cpu_clock_event_read(struct perf_event *event) |
7053 | { | |
7054 | cpu_clock_event_update(event); | |
7055 | } | |
f344011c | 7056 | |
b0a873eb PZ |
7057 | static int cpu_clock_event_init(struct perf_event *event) |
7058 | { | |
7059 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
7060 | return -ENOENT; | |
7061 | ||
7062 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
7063 | return -ENOENT; | |
7064 | ||
2481c5fa SE |
7065 | /* |
7066 | * no branch sampling for software events | |
7067 | */ | |
7068 | if (has_branch_stack(event)) | |
7069 | return -EOPNOTSUPP; | |
7070 | ||
ba3dd36c PZ |
7071 | perf_swevent_init_hrtimer(event); |
7072 | ||
b0a873eb | 7073 | return 0; |
f29ac756 PZ |
7074 | } |
7075 | ||
b0a873eb | 7076 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
7077 | .task_ctx_nr = perf_sw_context, |
7078 | ||
34f43927 PZ |
7079 | .capabilities = PERF_PMU_CAP_NO_NMI, |
7080 | ||
b0a873eb | 7081 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
7082 | .add = cpu_clock_event_add, |
7083 | .del = cpu_clock_event_del, | |
7084 | .start = cpu_clock_event_start, | |
7085 | .stop = cpu_clock_event_stop, | |
b0a873eb PZ |
7086 | .read = cpu_clock_event_read, |
7087 | }; | |
7088 | ||
7089 | /* | |
7090 | * Software event: task time clock | |
7091 | */ | |
7092 | ||
7093 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 7094 | { |
b0a873eb PZ |
7095 | u64 prev; |
7096 | s64 delta; | |
5c92d124 | 7097 | |
b0a873eb PZ |
7098 | prev = local64_xchg(&event->hw.prev_count, now); |
7099 | delta = now - prev; | |
7100 | local64_add(delta, &event->count); | |
7101 | } | |
5c92d124 | 7102 | |
a4eaf7f1 | 7103 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 7104 | { |
a4eaf7f1 | 7105 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 7106 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
7107 | } |
7108 | ||
a4eaf7f1 | 7109 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
7110 | { |
7111 | perf_swevent_cancel_hrtimer(event); | |
7112 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
7113 | } |
7114 | ||
7115 | static int task_clock_event_add(struct perf_event *event, int flags) | |
7116 | { | |
7117 | if (flags & PERF_EF_START) | |
7118 | task_clock_event_start(event, flags); | |
6a694a60 | 7119 | perf_event_update_userpage(event); |
b0a873eb | 7120 | |
a4eaf7f1 PZ |
7121 | return 0; |
7122 | } | |
7123 | ||
7124 | static void task_clock_event_del(struct perf_event *event, int flags) | |
7125 | { | |
7126 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
7127 | } |
7128 | ||
7129 | static void task_clock_event_read(struct perf_event *event) | |
7130 | { | |
768a06e2 PZ |
7131 | u64 now = perf_clock(); |
7132 | u64 delta = now - event->ctx->timestamp; | |
7133 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
7134 | |
7135 | task_clock_event_update(event, time); | |
7136 | } | |
7137 | ||
7138 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 7139 | { |
b0a873eb PZ |
7140 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
7141 | return -ENOENT; | |
7142 | ||
7143 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
7144 | return -ENOENT; | |
7145 | ||
2481c5fa SE |
7146 | /* |
7147 | * no branch sampling for software events | |
7148 | */ | |
7149 | if (has_branch_stack(event)) | |
7150 | return -EOPNOTSUPP; | |
7151 | ||
ba3dd36c PZ |
7152 | perf_swevent_init_hrtimer(event); |
7153 | ||
b0a873eb | 7154 | return 0; |
6fb2915d LZ |
7155 | } |
7156 | ||
b0a873eb | 7157 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
7158 | .task_ctx_nr = perf_sw_context, |
7159 | ||
34f43927 PZ |
7160 | .capabilities = PERF_PMU_CAP_NO_NMI, |
7161 | ||
b0a873eb | 7162 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
7163 | .add = task_clock_event_add, |
7164 | .del = task_clock_event_del, | |
7165 | .start = task_clock_event_start, | |
7166 | .stop = task_clock_event_stop, | |
b0a873eb PZ |
7167 | .read = task_clock_event_read, |
7168 | }; | |
6fb2915d | 7169 | |
ad5133b7 | 7170 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 7171 | { |
e077df4f | 7172 | } |
6fb2915d | 7173 | |
ad5133b7 | 7174 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 7175 | { |
ad5133b7 | 7176 | return 0; |
6fb2915d LZ |
7177 | } |
7178 | ||
ad5133b7 | 7179 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 7180 | { |
ad5133b7 | 7181 | perf_pmu_disable(pmu); |
6fb2915d LZ |
7182 | } |
7183 | ||
ad5133b7 PZ |
7184 | static int perf_pmu_commit_txn(struct pmu *pmu) |
7185 | { | |
7186 | perf_pmu_enable(pmu); | |
7187 | return 0; | |
7188 | } | |
e077df4f | 7189 | |
ad5133b7 | 7190 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 7191 | { |
ad5133b7 | 7192 | perf_pmu_enable(pmu); |
24f1e32c FW |
7193 | } |
7194 | ||
35edc2a5 PZ |
7195 | static int perf_event_idx_default(struct perf_event *event) |
7196 | { | |
c719f560 | 7197 | return 0; |
35edc2a5 PZ |
7198 | } |
7199 | ||
8dc85d54 PZ |
7200 | /* |
7201 | * Ensures all contexts with the same task_ctx_nr have the same | |
7202 | * pmu_cpu_context too. | |
7203 | */ | |
9e317041 | 7204 | static struct perf_cpu_context __percpu *find_pmu_context(int ctxn) |
24f1e32c | 7205 | { |
8dc85d54 | 7206 | struct pmu *pmu; |
b326e956 | 7207 | |
8dc85d54 PZ |
7208 | if (ctxn < 0) |
7209 | return NULL; | |
24f1e32c | 7210 | |
8dc85d54 PZ |
7211 | list_for_each_entry(pmu, &pmus, entry) { |
7212 | if (pmu->task_ctx_nr == ctxn) | |
7213 | return pmu->pmu_cpu_context; | |
7214 | } | |
24f1e32c | 7215 | |
8dc85d54 | 7216 | return NULL; |
24f1e32c FW |
7217 | } |
7218 | ||
51676957 | 7219 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 7220 | { |
51676957 PZ |
7221 | int cpu; |
7222 | ||
7223 | for_each_possible_cpu(cpu) { | |
7224 | struct perf_cpu_context *cpuctx; | |
7225 | ||
7226 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
7227 | ||
3f1f3320 PZ |
7228 | if (cpuctx->unique_pmu == old_pmu) |
7229 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
7230 | } |
7231 | } | |
7232 | ||
7233 | static void free_pmu_context(struct pmu *pmu) | |
7234 | { | |
7235 | struct pmu *i; | |
f5ffe02e | 7236 | |
8dc85d54 | 7237 | mutex_lock(&pmus_lock); |
0475f9ea | 7238 | /* |
8dc85d54 | 7239 | * Like a real lame refcount. |
0475f9ea | 7240 | */ |
51676957 PZ |
7241 | list_for_each_entry(i, &pmus, entry) { |
7242 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
7243 | update_pmu_context(i, pmu); | |
8dc85d54 | 7244 | goto out; |
51676957 | 7245 | } |
8dc85d54 | 7246 | } |
d6d020e9 | 7247 | |
51676957 | 7248 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
7249 | out: |
7250 | mutex_unlock(&pmus_lock); | |
24f1e32c | 7251 | } |
2e80a82a | 7252 | static struct idr pmu_idr; |
d6d020e9 | 7253 | |
abe43400 PZ |
7254 | static ssize_t |
7255 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
7256 | { | |
7257 | struct pmu *pmu = dev_get_drvdata(dev); | |
7258 | ||
7259 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
7260 | } | |
90826ca7 | 7261 | static DEVICE_ATTR_RO(type); |
abe43400 | 7262 | |
62b85639 SE |
7263 | static ssize_t |
7264 | perf_event_mux_interval_ms_show(struct device *dev, | |
7265 | struct device_attribute *attr, | |
7266 | char *page) | |
7267 | { | |
7268 | struct pmu *pmu = dev_get_drvdata(dev); | |
7269 | ||
7270 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
7271 | } | |
7272 | ||
272325c4 PZ |
7273 | static DEFINE_MUTEX(mux_interval_mutex); |
7274 | ||
62b85639 SE |
7275 | static ssize_t |
7276 | perf_event_mux_interval_ms_store(struct device *dev, | |
7277 | struct device_attribute *attr, | |
7278 | const char *buf, size_t count) | |
7279 | { | |
7280 | struct pmu *pmu = dev_get_drvdata(dev); | |
7281 | int timer, cpu, ret; | |
7282 | ||
7283 | ret = kstrtoint(buf, 0, &timer); | |
7284 | if (ret) | |
7285 | return ret; | |
7286 | ||
7287 | if (timer < 1) | |
7288 | return -EINVAL; | |
7289 | ||
7290 | /* same value, noting to do */ | |
7291 | if (timer == pmu->hrtimer_interval_ms) | |
7292 | return count; | |
7293 | ||
272325c4 | 7294 | mutex_lock(&mux_interval_mutex); |
62b85639 SE |
7295 | pmu->hrtimer_interval_ms = timer; |
7296 | ||
7297 | /* update all cpuctx for this PMU */ | |
272325c4 PZ |
7298 | get_online_cpus(); |
7299 | for_each_online_cpu(cpu) { | |
62b85639 SE |
7300 | struct perf_cpu_context *cpuctx; |
7301 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
7302 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
7303 | ||
272325c4 PZ |
7304 | cpu_function_call(cpu, |
7305 | (remote_function_f)perf_mux_hrtimer_restart, cpuctx); | |
62b85639 | 7306 | } |
272325c4 PZ |
7307 | put_online_cpus(); |
7308 | mutex_unlock(&mux_interval_mutex); | |
62b85639 SE |
7309 | |
7310 | return count; | |
7311 | } | |
90826ca7 | 7312 | static DEVICE_ATTR_RW(perf_event_mux_interval_ms); |
62b85639 | 7313 | |
90826ca7 GKH |
7314 | static struct attribute *pmu_dev_attrs[] = { |
7315 | &dev_attr_type.attr, | |
7316 | &dev_attr_perf_event_mux_interval_ms.attr, | |
7317 | NULL, | |
abe43400 | 7318 | }; |
90826ca7 | 7319 | ATTRIBUTE_GROUPS(pmu_dev); |
abe43400 PZ |
7320 | |
7321 | static int pmu_bus_running; | |
7322 | static struct bus_type pmu_bus = { | |
7323 | .name = "event_source", | |
90826ca7 | 7324 | .dev_groups = pmu_dev_groups, |
abe43400 PZ |
7325 | }; |
7326 | ||
7327 | static void pmu_dev_release(struct device *dev) | |
7328 | { | |
7329 | kfree(dev); | |
7330 | } | |
7331 | ||
7332 | static int pmu_dev_alloc(struct pmu *pmu) | |
7333 | { | |
7334 | int ret = -ENOMEM; | |
7335 | ||
7336 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
7337 | if (!pmu->dev) | |
7338 | goto out; | |
7339 | ||
0c9d42ed | 7340 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
7341 | device_initialize(pmu->dev); |
7342 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
7343 | if (ret) | |
7344 | goto free_dev; | |
7345 | ||
7346 | dev_set_drvdata(pmu->dev, pmu); | |
7347 | pmu->dev->bus = &pmu_bus; | |
7348 | pmu->dev->release = pmu_dev_release; | |
7349 | ret = device_add(pmu->dev); | |
7350 | if (ret) | |
7351 | goto free_dev; | |
7352 | ||
7353 | out: | |
7354 | return ret; | |
7355 | ||
7356 | free_dev: | |
7357 | put_device(pmu->dev); | |
7358 | goto out; | |
7359 | } | |
7360 | ||
547e9fd7 | 7361 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 7362 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 7363 | |
03d8e80b | 7364 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 7365 | { |
108b02cf | 7366 | int cpu, ret; |
24f1e32c | 7367 | |
b0a873eb | 7368 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
7369 | ret = -ENOMEM; |
7370 | pmu->pmu_disable_count = alloc_percpu(int); | |
7371 | if (!pmu->pmu_disable_count) | |
7372 | goto unlock; | |
f29ac756 | 7373 | |
2e80a82a PZ |
7374 | pmu->type = -1; |
7375 | if (!name) | |
7376 | goto skip_type; | |
7377 | pmu->name = name; | |
7378 | ||
7379 | if (type < 0) { | |
0e9c3be2 TH |
7380 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
7381 | if (type < 0) { | |
7382 | ret = type; | |
2e80a82a PZ |
7383 | goto free_pdc; |
7384 | } | |
7385 | } | |
7386 | pmu->type = type; | |
7387 | ||
abe43400 PZ |
7388 | if (pmu_bus_running) { |
7389 | ret = pmu_dev_alloc(pmu); | |
7390 | if (ret) | |
7391 | goto free_idr; | |
7392 | } | |
7393 | ||
2e80a82a | 7394 | skip_type: |
8dc85d54 PZ |
7395 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
7396 | if (pmu->pmu_cpu_context) | |
7397 | goto got_cpu_context; | |
f29ac756 | 7398 | |
c4814202 | 7399 | ret = -ENOMEM; |
108b02cf PZ |
7400 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
7401 | if (!pmu->pmu_cpu_context) | |
abe43400 | 7402 | goto free_dev; |
f344011c | 7403 | |
108b02cf PZ |
7404 | for_each_possible_cpu(cpu) { |
7405 | struct perf_cpu_context *cpuctx; | |
7406 | ||
7407 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 7408 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 7409 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 7410 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
108b02cf | 7411 | cpuctx->ctx.pmu = pmu; |
9e630205 | 7412 | |
272325c4 | 7413 | __perf_mux_hrtimer_init(cpuctx, cpu); |
9e630205 | 7414 | |
3f1f3320 | 7415 | cpuctx->unique_pmu = pmu; |
108b02cf | 7416 | } |
76e1d904 | 7417 | |
8dc85d54 | 7418 | got_cpu_context: |
ad5133b7 PZ |
7419 | if (!pmu->start_txn) { |
7420 | if (pmu->pmu_enable) { | |
7421 | /* | |
7422 | * If we have pmu_enable/pmu_disable calls, install | |
7423 | * transaction stubs that use that to try and batch | |
7424 | * hardware accesses. | |
7425 | */ | |
7426 | pmu->start_txn = perf_pmu_start_txn; | |
7427 | pmu->commit_txn = perf_pmu_commit_txn; | |
7428 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
7429 | } else { | |
7430 | pmu->start_txn = perf_pmu_nop_void; | |
7431 | pmu->commit_txn = perf_pmu_nop_int; | |
7432 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 7433 | } |
5c92d124 | 7434 | } |
15dbf27c | 7435 | |
ad5133b7 PZ |
7436 | if (!pmu->pmu_enable) { |
7437 | pmu->pmu_enable = perf_pmu_nop_void; | |
7438 | pmu->pmu_disable = perf_pmu_nop_void; | |
7439 | } | |
7440 | ||
35edc2a5 PZ |
7441 | if (!pmu->event_idx) |
7442 | pmu->event_idx = perf_event_idx_default; | |
7443 | ||
b0a873eb | 7444 | list_add_rcu(&pmu->entry, &pmus); |
bed5b25a | 7445 | atomic_set(&pmu->exclusive_cnt, 0); |
33696fc0 PZ |
7446 | ret = 0; |
7447 | unlock: | |
b0a873eb PZ |
7448 | mutex_unlock(&pmus_lock); |
7449 | ||
33696fc0 | 7450 | return ret; |
108b02cf | 7451 | |
abe43400 PZ |
7452 | free_dev: |
7453 | device_del(pmu->dev); | |
7454 | put_device(pmu->dev); | |
7455 | ||
2e80a82a PZ |
7456 | free_idr: |
7457 | if (pmu->type >= PERF_TYPE_MAX) | |
7458 | idr_remove(&pmu_idr, pmu->type); | |
7459 | ||
108b02cf PZ |
7460 | free_pdc: |
7461 | free_percpu(pmu->pmu_disable_count); | |
7462 | goto unlock; | |
f29ac756 | 7463 | } |
c464c76e | 7464 | EXPORT_SYMBOL_GPL(perf_pmu_register); |
f29ac756 | 7465 | |
b0a873eb | 7466 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 7467 | { |
b0a873eb PZ |
7468 | mutex_lock(&pmus_lock); |
7469 | list_del_rcu(&pmu->entry); | |
7470 | mutex_unlock(&pmus_lock); | |
5c92d124 | 7471 | |
0475f9ea | 7472 | /* |
cde8e884 PZ |
7473 | * We dereference the pmu list under both SRCU and regular RCU, so |
7474 | * synchronize against both of those. | |
0475f9ea | 7475 | */ |
b0a873eb | 7476 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 7477 | synchronize_rcu(); |
d6d020e9 | 7478 | |
33696fc0 | 7479 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
7480 | if (pmu->type >= PERF_TYPE_MAX) |
7481 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
7482 | device_del(pmu->dev); |
7483 | put_device(pmu->dev); | |
51676957 | 7484 | free_pmu_context(pmu); |
b0a873eb | 7485 | } |
c464c76e | 7486 | EXPORT_SYMBOL_GPL(perf_pmu_unregister); |
d6d020e9 | 7487 | |
cc34b98b MR |
7488 | static int perf_try_init_event(struct pmu *pmu, struct perf_event *event) |
7489 | { | |
ccd41c86 | 7490 | struct perf_event_context *ctx = NULL; |
cc34b98b MR |
7491 | int ret; |
7492 | ||
7493 | if (!try_module_get(pmu->module)) | |
7494 | return -ENODEV; | |
ccd41c86 PZ |
7495 | |
7496 | if (event->group_leader != event) { | |
8b10c5e2 PZ |
7497 | /* |
7498 | * This ctx->mutex can nest when we're called through | |
7499 | * inheritance. See the perf_event_ctx_lock_nested() comment. | |
7500 | */ | |
7501 | ctx = perf_event_ctx_lock_nested(event->group_leader, | |
7502 | SINGLE_DEPTH_NESTING); | |
ccd41c86 PZ |
7503 | BUG_ON(!ctx); |
7504 | } | |
7505 | ||
cc34b98b MR |
7506 | event->pmu = pmu; |
7507 | ret = pmu->event_init(event); | |
ccd41c86 PZ |
7508 | |
7509 | if (ctx) | |
7510 | perf_event_ctx_unlock(event->group_leader, ctx); | |
7511 | ||
cc34b98b MR |
7512 | if (ret) |
7513 | module_put(pmu->module); | |
7514 | ||
7515 | return ret; | |
7516 | } | |
7517 | ||
b0a873eb PZ |
7518 | struct pmu *perf_init_event(struct perf_event *event) |
7519 | { | |
7520 | struct pmu *pmu = NULL; | |
7521 | int idx; | |
940c5b29 | 7522 | int ret; |
b0a873eb PZ |
7523 | |
7524 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
7525 | |
7526 | rcu_read_lock(); | |
7527 | pmu = idr_find(&pmu_idr, event->attr.type); | |
7528 | rcu_read_unlock(); | |
940c5b29 | 7529 | if (pmu) { |
cc34b98b | 7530 | ret = perf_try_init_event(pmu, event); |
940c5b29 LM |
7531 | if (ret) |
7532 | pmu = ERR_PTR(ret); | |
2e80a82a | 7533 | goto unlock; |
940c5b29 | 7534 | } |
2e80a82a | 7535 | |
b0a873eb | 7536 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
cc34b98b | 7537 | ret = perf_try_init_event(pmu, event); |
b0a873eb | 7538 | if (!ret) |
e5f4d339 | 7539 | goto unlock; |
76e1d904 | 7540 | |
b0a873eb PZ |
7541 | if (ret != -ENOENT) { |
7542 | pmu = ERR_PTR(ret); | |
e5f4d339 | 7543 | goto unlock; |
f344011c | 7544 | } |
5c92d124 | 7545 | } |
e5f4d339 PZ |
7546 | pmu = ERR_PTR(-ENOENT); |
7547 | unlock: | |
b0a873eb | 7548 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 7549 | |
4aeb0b42 | 7550 | return pmu; |
5c92d124 IM |
7551 | } |
7552 | ||
4beb31f3 FW |
7553 | static void account_event_cpu(struct perf_event *event, int cpu) |
7554 | { | |
7555 | if (event->parent) | |
7556 | return; | |
7557 | ||
4beb31f3 FW |
7558 | if (is_cgroup_event(event)) |
7559 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
7560 | } | |
7561 | ||
766d6c07 FW |
7562 | static void account_event(struct perf_event *event) |
7563 | { | |
4beb31f3 FW |
7564 | if (event->parent) |
7565 | return; | |
7566 | ||
766d6c07 FW |
7567 | if (event->attach_state & PERF_ATTACH_TASK) |
7568 | static_key_slow_inc(&perf_sched_events.key); | |
7569 | if (event->attr.mmap || event->attr.mmap_data) | |
7570 | atomic_inc(&nr_mmap_events); | |
7571 | if (event->attr.comm) | |
7572 | atomic_inc(&nr_comm_events); | |
7573 | if (event->attr.task) | |
7574 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
7575 | if (event->attr.freq) { |
7576 | if (atomic_inc_return(&nr_freq_events) == 1) | |
7577 | tick_nohz_full_kick_all(); | |
7578 | } | |
45ac1403 AH |
7579 | if (event->attr.context_switch) { |
7580 | atomic_inc(&nr_switch_events); | |
7581 | static_key_slow_inc(&perf_sched_events.key); | |
7582 | } | |
4beb31f3 | 7583 | if (has_branch_stack(event)) |
766d6c07 | 7584 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 7585 | if (is_cgroup_event(event)) |
766d6c07 | 7586 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
7587 | |
7588 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
7589 | } |
7590 | ||
0793a61d | 7591 | /* |
cdd6c482 | 7592 | * Allocate and initialize a event structure |
0793a61d | 7593 | */ |
cdd6c482 | 7594 | static struct perf_event * |
c3f00c70 | 7595 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
7596 | struct task_struct *task, |
7597 | struct perf_event *group_leader, | |
7598 | struct perf_event *parent_event, | |
4dc0da86 | 7599 | perf_overflow_handler_t overflow_handler, |
79dff51e | 7600 | void *context, int cgroup_fd) |
0793a61d | 7601 | { |
51b0fe39 | 7602 | struct pmu *pmu; |
cdd6c482 IM |
7603 | struct perf_event *event; |
7604 | struct hw_perf_event *hwc; | |
90983b16 | 7605 | long err = -EINVAL; |
0793a61d | 7606 | |
66832eb4 ON |
7607 | if ((unsigned)cpu >= nr_cpu_ids) { |
7608 | if (!task || cpu != -1) | |
7609 | return ERR_PTR(-EINVAL); | |
7610 | } | |
7611 | ||
c3f00c70 | 7612 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 7613 | if (!event) |
d5d2bc0d | 7614 | return ERR_PTR(-ENOMEM); |
0793a61d | 7615 | |
04289bb9 | 7616 | /* |
cdd6c482 | 7617 | * Single events are their own group leaders, with an |
04289bb9 IM |
7618 | * empty sibling list: |
7619 | */ | |
7620 | if (!group_leader) | |
cdd6c482 | 7621 | group_leader = event; |
04289bb9 | 7622 | |
cdd6c482 IM |
7623 | mutex_init(&event->child_mutex); |
7624 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 7625 | |
cdd6c482 IM |
7626 | INIT_LIST_HEAD(&event->group_entry); |
7627 | INIT_LIST_HEAD(&event->event_entry); | |
7628 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 | 7629 | INIT_LIST_HEAD(&event->rb_entry); |
71ad88ef | 7630 | INIT_LIST_HEAD(&event->active_entry); |
f3ae75de SE |
7631 | INIT_HLIST_NODE(&event->hlist_entry); |
7632 | ||
10c6db11 | 7633 | |
cdd6c482 | 7634 | init_waitqueue_head(&event->waitq); |
e360adbe | 7635 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 7636 | |
cdd6c482 | 7637 | mutex_init(&event->mmap_mutex); |
7b732a75 | 7638 | |
a6fa941d | 7639 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
7640 | event->cpu = cpu; |
7641 | event->attr = *attr; | |
7642 | event->group_leader = group_leader; | |
7643 | event->pmu = NULL; | |
cdd6c482 | 7644 | event->oncpu = -1; |
a96bbc16 | 7645 | |
cdd6c482 | 7646 | event->parent = parent_event; |
b84fbc9f | 7647 | |
17cf22c3 | 7648 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 7649 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 7650 | |
cdd6c482 | 7651 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 7652 | |
d580ff86 PZ |
7653 | if (task) { |
7654 | event->attach_state = PERF_ATTACH_TASK; | |
d580ff86 | 7655 | /* |
50f16a8b PZ |
7656 | * XXX pmu::event_init needs to know what task to account to |
7657 | * and we cannot use the ctx information because we need the | |
7658 | * pmu before we get a ctx. | |
d580ff86 | 7659 | */ |
50f16a8b | 7660 | event->hw.target = task; |
d580ff86 PZ |
7661 | } |
7662 | ||
34f43927 PZ |
7663 | event->clock = &local_clock; |
7664 | if (parent_event) | |
7665 | event->clock = parent_event->clock; | |
7666 | ||
4dc0da86 | 7667 | if (!overflow_handler && parent_event) { |
b326e956 | 7668 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
7669 | context = parent_event->overflow_handler_context; |
7670 | } | |
66832eb4 | 7671 | |
b326e956 | 7672 | event->overflow_handler = overflow_handler; |
4dc0da86 | 7673 | event->overflow_handler_context = context; |
97eaf530 | 7674 | |
0231bb53 | 7675 | perf_event__state_init(event); |
a86ed508 | 7676 | |
4aeb0b42 | 7677 | pmu = NULL; |
b8e83514 | 7678 | |
cdd6c482 | 7679 | hwc = &event->hw; |
bd2b5b12 | 7680 | hwc->sample_period = attr->sample_period; |
0d48696f | 7681 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 7682 | hwc->sample_period = 1; |
eced1dfc | 7683 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 7684 | |
e7850595 | 7685 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 7686 | |
2023b359 | 7687 | /* |
cdd6c482 | 7688 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 7689 | */ |
3dab77fb | 7690 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 7691 | goto err_ns; |
a46a2300 YZ |
7692 | |
7693 | if (!has_branch_stack(event)) | |
7694 | event->attr.branch_sample_type = 0; | |
2023b359 | 7695 | |
79dff51e MF |
7696 | if (cgroup_fd != -1) { |
7697 | err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader); | |
7698 | if (err) | |
7699 | goto err_ns; | |
7700 | } | |
7701 | ||
b0a873eb | 7702 | pmu = perf_init_event(event); |
4aeb0b42 | 7703 | if (!pmu) |
90983b16 FW |
7704 | goto err_ns; |
7705 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 7706 | err = PTR_ERR(pmu); |
90983b16 | 7707 | goto err_ns; |
621a01ea | 7708 | } |
d5d2bc0d | 7709 | |
bed5b25a AS |
7710 | err = exclusive_event_init(event); |
7711 | if (err) | |
7712 | goto err_pmu; | |
7713 | ||
cdd6c482 | 7714 | if (!event->parent) { |
927c7a9e FW |
7715 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
7716 | err = get_callchain_buffers(); | |
90983b16 | 7717 | if (err) |
bed5b25a | 7718 | goto err_per_task; |
d010b332 | 7719 | } |
f344011c | 7720 | } |
9ee318a7 | 7721 | |
cdd6c482 | 7722 | return event; |
90983b16 | 7723 | |
bed5b25a AS |
7724 | err_per_task: |
7725 | exclusive_event_destroy(event); | |
7726 | ||
90983b16 FW |
7727 | err_pmu: |
7728 | if (event->destroy) | |
7729 | event->destroy(event); | |
c464c76e | 7730 | module_put(pmu->module); |
90983b16 | 7731 | err_ns: |
79dff51e MF |
7732 | if (is_cgroup_event(event)) |
7733 | perf_detach_cgroup(event); | |
90983b16 FW |
7734 | if (event->ns) |
7735 | put_pid_ns(event->ns); | |
7736 | kfree(event); | |
7737 | ||
7738 | return ERR_PTR(err); | |
0793a61d TG |
7739 | } |
7740 | ||
cdd6c482 IM |
7741 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
7742 | struct perf_event_attr *attr) | |
974802ea | 7743 | { |
974802ea | 7744 | u32 size; |
cdf8073d | 7745 | int ret; |
974802ea PZ |
7746 | |
7747 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
7748 | return -EFAULT; | |
7749 | ||
7750 | /* | |
7751 | * zero the full structure, so that a short copy will be nice. | |
7752 | */ | |
7753 | memset(attr, 0, sizeof(*attr)); | |
7754 | ||
7755 | ret = get_user(size, &uattr->size); | |
7756 | if (ret) | |
7757 | return ret; | |
7758 | ||
7759 | if (size > PAGE_SIZE) /* silly large */ | |
7760 | goto err_size; | |
7761 | ||
7762 | if (!size) /* abi compat */ | |
7763 | size = PERF_ATTR_SIZE_VER0; | |
7764 | ||
7765 | if (size < PERF_ATTR_SIZE_VER0) | |
7766 | goto err_size; | |
7767 | ||
7768 | /* | |
7769 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
7770 | * ensure all the unknown bits are 0 - i.e. new |
7771 | * user-space does not rely on any kernel feature | |
7772 | * extensions we dont know about yet. | |
974802ea PZ |
7773 | */ |
7774 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
7775 | unsigned char __user *addr; |
7776 | unsigned char __user *end; | |
7777 | unsigned char val; | |
974802ea | 7778 | |
cdf8073d IS |
7779 | addr = (void __user *)uattr + sizeof(*attr); |
7780 | end = (void __user *)uattr + size; | |
974802ea | 7781 | |
cdf8073d | 7782 | for (; addr < end; addr++) { |
974802ea PZ |
7783 | ret = get_user(val, addr); |
7784 | if (ret) | |
7785 | return ret; | |
7786 | if (val) | |
7787 | goto err_size; | |
7788 | } | |
b3e62e35 | 7789 | size = sizeof(*attr); |
974802ea PZ |
7790 | } |
7791 | ||
7792 | ret = copy_from_user(attr, uattr, size); | |
7793 | if (ret) | |
7794 | return -EFAULT; | |
7795 | ||
cd757645 | 7796 | if (attr->__reserved_1) |
974802ea PZ |
7797 | return -EINVAL; |
7798 | ||
7799 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
7800 | return -EINVAL; | |
7801 | ||
7802 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
7803 | return -EINVAL; | |
7804 | ||
bce38cd5 SE |
7805 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
7806 | u64 mask = attr->branch_sample_type; | |
7807 | ||
7808 | /* only using defined bits */ | |
7809 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
7810 | return -EINVAL; | |
7811 | ||
7812 | /* at least one branch bit must be set */ | |
7813 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
7814 | return -EINVAL; | |
7815 | ||
bce38cd5 SE |
7816 | /* propagate priv level, when not set for branch */ |
7817 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
7818 | ||
7819 | /* exclude_kernel checked on syscall entry */ | |
7820 | if (!attr->exclude_kernel) | |
7821 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
7822 | ||
7823 | if (!attr->exclude_user) | |
7824 | mask |= PERF_SAMPLE_BRANCH_USER; | |
7825 | ||
7826 | if (!attr->exclude_hv) | |
7827 | mask |= PERF_SAMPLE_BRANCH_HV; | |
7828 | /* | |
7829 | * adjust user setting (for HW filter setup) | |
7830 | */ | |
7831 | attr->branch_sample_type = mask; | |
7832 | } | |
e712209a SE |
7833 | /* privileged levels capture (kernel, hv): check permissions */ |
7834 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
7835 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
7836 | return -EACCES; | |
bce38cd5 | 7837 | } |
4018994f | 7838 | |
c5ebcedb | 7839 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 7840 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
7841 | if (ret) |
7842 | return ret; | |
7843 | } | |
7844 | ||
7845 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
7846 | if (!arch_perf_have_user_stack_dump()) | |
7847 | return -ENOSYS; | |
7848 | ||
7849 | /* | |
7850 | * We have __u32 type for the size, but so far | |
7851 | * we can only use __u16 as maximum due to the | |
7852 | * __u16 sample size limit. | |
7853 | */ | |
7854 | if (attr->sample_stack_user >= USHRT_MAX) | |
7855 | ret = -EINVAL; | |
7856 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
7857 | ret = -EINVAL; | |
7858 | } | |
4018994f | 7859 | |
60e2364e SE |
7860 | if (attr->sample_type & PERF_SAMPLE_REGS_INTR) |
7861 | ret = perf_reg_validate(attr->sample_regs_intr); | |
974802ea PZ |
7862 | out: |
7863 | return ret; | |
7864 | ||
7865 | err_size: | |
7866 | put_user(sizeof(*attr), &uattr->size); | |
7867 | ret = -E2BIG; | |
7868 | goto out; | |
7869 | } | |
7870 | ||
ac9721f3 PZ |
7871 | static int |
7872 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 7873 | { |
b69cf536 | 7874 | struct ring_buffer *rb = NULL; |
a4be7c27 PZ |
7875 | int ret = -EINVAL; |
7876 | ||
ac9721f3 | 7877 | if (!output_event) |
a4be7c27 PZ |
7878 | goto set; |
7879 | ||
ac9721f3 PZ |
7880 | /* don't allow circular references */ |
7881 | if (event == output_event) | |
a4be7c27 PZ |
7882 | goto out; |
7883 | ||
0f139300 PZ |
7884 | /* |
7885 | * Don't allow cross-cpu buffers | |
7886 | */ | |
7887 | if (output_event->cpu != event->cpu) | |
7888 | goto out; | |
7889 | ||
7890 | /* | |
76369139 | 7891 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
7892 | */ |
7893 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
7894 | goto out; | |
7895 | ||
34f43927 PZ |
7896 | /* |
7897 | * Mixing clocks in the same buffer is trouble you don't need. | |
7898 | */ | |
7899 | if (output_event->clock != event->clock) | |
7900 | goto out; | |
7901 | ||
45bfb2e5 PZ |
7902 | /* |
7903 | * If both events generate aux data, they must be on the same PMU | |
7904 | */ | |
7905 | if (has_aux(event) && has_aux(output_event) && | |
7906 | event->pmu != output_event->pmu) | |
7907 | goto out; | |
7908 | ||
a4be7c27 | 7909 | set: |
cdd6c482 | 7910 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
7911 | /* Can't redirect output if we've got an active mmap() */ |
7912 | if (atomic_read(&event->mmap_count)) | |
7913 | goto unlock; | |
a4be7c27 | 7914 | |
ac9721f3 | 7915 | if (output_event) { |
76369139 FW |
7916 | /* get the rb we want to redirect to */ |
7917 | rb = ring_buffer_get(output_event); | |
7918 | if (!rb) | |
ac9721f3 | 7919 | goto unlock; |
a4be7c27 PZ |
7920 | } |
7921 | ||
b69cf536 | 7922 | ring_buffer_attach(event, rb); |
9bb5d40c | 7923 | |
a4be7c27 | 7924 | ret = 0; |
ac9721f3 PZ |
7925 | unlock: |
7926 | mutex_unlock(&event->mmap_mutex); | |
7927 | ||
a4be7c27 | 7928 | out: |
a4be7c27 PZ |
7929 | return ret; |
7930 | } | |
7931 | ||
f63a8daa PZ |
7932 | static void mutex_lock_double(struct mutex *a, struct mutex *b) |
7933 | { | |
7934 | if (b < a) | |
7935 | swap(a, b); | |
7936 | ||
7937 | mutex_lock(a); | |
7938 | mutex_lock_nested(b, SINGLE_DEPTH_NESTING); | |
7939 | } | |
7940 | ||
34f43927 PZ |
7941 | static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id) |
7942 | { | |
7943 | bool nmi_safe = false; | |
7944 | ||
7945 | switch (clk_id) { | |
7946 | case CLOCK_MONOTONIC: | |
7947 | event->clock = &ktime_get_mono_fast_ns; | |
7948 | nmi_safe = true; | |
7949 | break; | |
7950 | ||
7951 | case CLOCK_MONOTONIC_RAW: | |
7952 | event->clock = &ktime_get_raw_fast_ns; | |
7953 | nmi_safe = true; | |
7954 | break; | |
7955 | ||
7956 | case CLOCK_REALTIME: | |
7957 | event->clock = &ktime_get_real_ns; | |
7958 | break; | |
7959 | ||
7960 | case CLOCK_BOOTTIME: | |
7961 | event->clock = &ktime_get_boot_ns; | |
7962 | break; | |
7963 | ||
7964 | case CLOCK_TAI: | |
7965 | event->clock = &ktime_get_tai_ns; | |
7966 | break; | |
7967 | ||
7968 | default: | |
7969 | return -EINVAL; | |
7970 | } | |
7971 | ||
7972 | if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI)) | |
7973 | return -EINVAL; | |
7974 | ||
7975 | return 0; | |
7976 | } | |
7977 | ||
0793a61d | 7978 | /** |
cdd6c482 | 7979 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 7980 | * |
cdd6c482 | 7981 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 7982 | * @pid: target pid |
9f66a381 | 7983 | * @cpu: target cpu |
cdd6c482 | 7984 | * @group_fd: group leader event fd |
0793a61d | 7985 | */ |
cdd6c482 IM |
7986 | SYSCALL_DEFINE5(perf_event_open, |
7987 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 7988 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 7989 | { |
b04243ef PZ |
7990 | struct perf_event *group_leader = NULL, *output_event = NULL; |
7991 | struct perf_event *event, *sibling; | |
cdd6c482 | 7992 | struct perf_event_attr attr; |
f63a8daa | 7993 | struct perf_event_context *ctx, *uninitialized_var(gctx); |
cdd6c482 | 7994 | struct file *event_file = NULL; |
2903ff01 | 7995 | struct fd group = {NULL, 0}; |
38a81da2 | 7996 | struct task_struct *task = NULL; |
89a1e187 | 7997 | struct pmu *pmu; |
ea635c64 | 7998 | int event_fd; |
b04243ef | 7999 | int move_group = 0; |
dc86cabe | 8000 | int err; |
a21b0b35 | 8001 | int f_flags = O_RDWR; |
79dff51e | 8002 | int cgroup_fd = -1; |
0793a61d | 8003 | |
2743a5b0 | 8004 | /* for future expandability... */ |
e5d1367f | 8005 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
8006 | return -EINVAL; |
8007 | ||
dc86cabe IM |
8008 | err = perf_copy_attr(attr_uptr, &attr); |
8009 | if (err) | |
8010 | return err; | |
eab656ae | 8011 | |
0764771d PZ |
8012 | if (!attr.exclude_kernel) { |
8013 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
8014 | return -EACCES; | |
8015 | } | |
8016 | ||
df58ab24 | 8017 | if (attr.freq) { |
cdd6c482 | 8018 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 | 8019 | return -EINVAL; |
0819b2e3 PZ |
8020 | } else { |
8021 | if (attr.sample_period & (1ULL << 63)) | |
8022 | return -EINVAL; | |
df58ab24 PZ |
8023 | } |
8024 | ||
e5d1367f SE |
8025 | /* |
8026 | * In cgroup mode, the pid argument is used to pass the fd | |
8027 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
8028 | * designates the cpu on which to monitor threads from that | |
8029 | * cgroup. | |
8030 | */ | |
8031 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
8032 | return -EINVAL; | |
8033 | ||
a21b0b35 YD |
8034 | if (flags & PERF_FLAG_FD_CLOEXEC) |
8035 | f_flags |= O_CLOEXEC; | |
8036 | ||
8037 | event_fd = get_unused_fd_flags(f_flags); | |
ea635c64 AV |
8038 | if (event_fd < 0) |
8039 | return event_fd; | |
8040 | ||
ac9721f3 | 8041 | if (group_fd != -1) { |
2903ff01 AV |
8042 | err = perf_fget_light(group_fd, &group); |
8043 | if (err) | |
d14b12d7 | 8044 | goto err_fd; |
2903ff01 | 8045 | group_leader = group.file->private_data; |
ac9721f3 PZ |
8046 | if (flags & PERF_FLAG_FD_OUTPUT) |
8047 | output_event = group_leader; | |
8048 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
8049 | group_leader = NULL; | |
8050 | } | |
8051 | ||
e5d1367f | 8052 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
8053 | task = find_lively_task_by_vpid(pid); |
8054 | if (IS_ERR(task)) { | |
8055 | err = PTR_ERR(task); | |
8056 | goto err_group_fd; | |
8057 | } | |
8058 | } | |
8059 | ||
1f4ee503 PZ |
8060 | if (task && group_leader && |
8061 | group_leader->attr.inherit != attr.inherit) { | |
8062 | err = -EINVAL; | |
8063 | goto err_task; | |
8064 | } | |
8065 | ||
fbfc623f YZ |
8066 | get_online_cpus(); |
8067 | ||
79dff51e MF |
8068 | if (flags & PERF_FLAG_PID_CGROUP) |
8069 | cgroup_fd = pid; | |
8070 | ||
4dc0da86 | 8071 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
79dff51e | 8072 | NULL, NULL, cgroup_fd); |
d14b12d7 SE |
8073 | if (IS_ERR(event)) { |
8074 | err = PTR_ERR(event); | |
1f4ee503 | 8075 | goto err_cpus; |
d14b12d7 SE |
8076 | } |
8077 | ||
53b25335 VW |
8078 | if (is_sampling_event(event)) { |
8079 | if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) { | |
8080 | err = -ENOTSUPP; | |
8081 | goto err_alloc; | |
8082 | } | |
8083 | } | |
8084 | ||
766d6c07 FW |
8085 | account_event(event); |
8086 | ||
89a1e187 PZ |
8087 | /* |
8088 | * Special case software events and allow them to be part of | |
8089 | * any hardware group. | |
8090 | */ | |
8091 | pmu = event->pmu; | |
b04243ef | 8092 | |
34f43927 PZ |
8093 | if (attr.use_clockid) { |
8094 | err = perf_event_set_clock(event, attr.clockid); | |
8095 | if (err) | |
8096 | goto err_alloc; | |
8097 | } | |
8098 | ||
b04243ef PZ |
8099 | if (group_leader && |
8100 | (is_software_event(event) != is_software_event(group_leader))) { | |
8101 | if (is_software_event(event)) { | |
8102 | /* | |
8103 | * If event and group_leader are not both a software | |
8104 | * event, and event is, then group leader is not. | |
8105 | * | |
8106 | * Allow the addition of software events to !software | |
8107 | * groups, this is safe because software events never | |
8108 | * fail to schedule. | |
8109 | */ | |
8110 | pmu = group_leader->pmu; | |
8111 | } else if (is_software_event(group_leader) && | |
8112 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
8113 | /* | |
8114 | * In case the group is a pure software group, and we | |
8115 | * try to add a hardware event, move the whole group to | |
8116 | * the hardware context. | |
8117 | */ | |
8118 | move_group = 1; | |
8119 | } | |
8120 | } | |
89a1e187 PZ |
8121 | |
8122 | /* | |
8123 | * Get the target context (task or percpu): | |
8124 | */ | |
4af57ef2 | 8125 | ctx = find_get_context(pmu, task, event); |
89a1e187 PZ |
8126 | if (IS_ERR(ctx)) { |
8127 | err = PTR_ERR(ctx); | |
c6be5a5c | 8128 | goto err_alloc; |
89a1e187 PZ |
8129 | } |
8130 | ||
bed5b25a AS |
8131 | if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) { |
8132 | err = -EBUSY; | |
8133 | goto err_context; | |
8134 | } | |
8135 | ||
fd1edb3a PZ |
8136 | if (task) { |
8137 | put_task_struct(task); | |
8138 | task = NULL; | |
8139 | } | |
8140 | ||
ccff286d | 8141 | /* |
cdd6c482 | 8142 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 8143 | */ |
ac9721f3 | 8144 | if (group_leader) { |
dc86cabe | 8145 | err = -EINVAL; |
04289bb9 | 8146 | |
04289bb9 | 8147 | /* |
ccff286d IM |
8148 | * Do not allow a recursive hierarchy (this new sibling |
8149 | * becoming part of another group-sibling): | |
8150 | */ | |
8151 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 8152 | goto err_context; |
34f43927 PZ |
8153 | |
8154 | /* All events in a group should have the same clock */ | |
8155 | if (group_leader->clock != event->clock) | |
8156 | goto err_context; | |
8157 | ||
ccff286d IM |
8158 | /* |
8159 | * Do not allow to attach to a group in a different | |
8160 | * task or CPU context: | |
04289bb9 | 8161 | */ |
b04243ef | 8162 | if (move_group) { |
c3c87e77 PZ |
8163 | /* |
8164 | * Make sure we're both on the same task, or both | |
8165 | * per-cpu events. | |
8166 | */ | |
8167 | if (group_leader->ctx->task != ctx->task) | |
8168 | goto err_context; | |
8169 | ||
8170 | /* | |
8171 | * Make sure we're both events for the same CPU; | |
8172 | * grouping events for different CPUs is broken; since | |
8173 | * you can never concurrently schedule them anyhow. | |
8174 | */ | |
8175 | if (group_leader->cpu != event->cpu) | |
b04243ef PZ |
8176 | goto err_context; |
8177 | } else { | |
8178 | if (group_leader->ctx != ctx) | |
8179 | goto err_context; | |
8180 | } | |
8181 | ||
3b6f9e5c PM |
8182 | /* |
8183 | * Only a group leader can be exclusive or pinned | |
8184 | */ | |
0d48696f | 8185 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 8186 | goto err_context; |
ac9721f3 PZ |
8187 | } |
8188 | ||
8189 | if (output_event) { | |
8190 | err = perf_event_set_output(event, output_event); | |
8191 | if (err) | |
c3f00c70 | 8192 | goto err_context; |
ac9721f3 | 8193 | } |
0793a61d | 8194 | |
a21b0b35 YD |
8195 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, |
8196 | f_flags); | |
ea635c64 AV |
8197 | if (IS_ERR(event_file)) { |
8198 | err = PTR_ERR(event_file); | |
c3f00c70 | 8199 | goto err_context; |
ea635c64 | 8200 | } |
9b51f66d | 8201 | |
b04243ef | 8202 | if (move_group) { |
f63a8daa PZ |
8203 | gctx = group_leader->ctx; |
8204 | ||
8205 | /* | |
8206 | * See perf_event_ctx_lock() for comments on the details | |
8207 | * of swizzling perf_event::ctx. | |
8208 | */ | |
8209 | mutex_lock_double(&gctx->mutex, &ctx->mutex); | |
b04243ef | 8210 | |
46ce0fe9 | 8211 | perf_remove_from_context(group_leader, false); |
0231bb53 | 8212 | |
b04243ef PZ |
8213 | list_for_each_entry(sibling, &group_leader->sibling_list, |
8214 | group_entry) { | |
46ce0fe9 | 8215 | perf_remove_from_context(sibling, false); |
b04243ef PZ |
8216 | put_ctx(gctx); |
8217 | } | |
f63a8daa PZ |
8218 | } else { |
8219 | mutex_lock(&ctx->mutex); | |
ea635c64 | 8220 | } |
9b51f66d | 8221 | |
ad3a37de | 8222 | WARN_ON_ONCE(ctx->parent_ctx); |
b04243ef PZ |
8223 | |
8224 | if (move_group) { | |
f63a8daa PZ |
8225 | /* |
8226 | * Wait for everybody to stop referencing the events through | |
8227 | * the old lists, before installing it on new lists. | |
8228 | */ | |
0cda4c02 | 8229 | synchronize_rcu(); |
f63a8daa | 8230 | |
8f95b435 PZI |
8231 | /* |
8232 | * Install the group siblings before the group leader. | |
8233 | * | |
8234 | * Because a group leader will try and install the entire group | |
8235 | * (through the sibling list, which is still in-tact), we can | |
8236 | * end up with siblings installed in the wrong context. | |
8237 | * | |
8238 | * By installing siblings first we NO-OP because they're not | |
8239 | * reachable through the group lists. | |
8240 | */ | |
b04243ef PZ |
8241 | list_for_each_entry(sibling, &group_leader->sibling_list, |
8242 | group_entry) { | |
8f95b435 | 8243 | perf_event__state_init(sibling); |
9fc81d87 | 8244 | perf_install_in_context(ctx, sibling, sibling->cpu); |
b04243ef PZ |
8245 | get_ctx(ctx); |
8246 | } | |
8f95b435 PZI |
8247 | |
8248 | /* | |
8249 | * Removing from the context ends up with disabled | |
8250 | * event. What we want here is event in the initial | |
8251 | * startup state, ready to be add into new context. | |
8252 | */ | |
8253 | perf_event__state_init(group_leader); | |
8254 | perf_install_in_context(ctx, group_leader, group_leader->cpu); | |
8255 | get_ctx(ctx); | |
b04243ef PZ |
8256 | } |
8257 | ||
bed5b25a AS |
8258 | if (!exclusive_event_installable(event, ctx)) { |
8259 | err = -EBUSY; | |
8260 | mutex_unlock(&ctx->mutex); | |
8261 | fput(event_file); | |
8262 | goto err_context; | |
8263 | } | |
8264 | ||
e2d37cd2 | 8265 | perf_install_in_context(ctx, event, event->cpu); |
fe4b04fa | 8266 | perf_unpin_context(ctx); |
f63a8daa PZ |
8267 | |
8268 | if (move_group) { | |
8269 | mutex_unlock(&gctx->mutex); | |
8270 | put_ctx(gctx); | |
8271 | } | |
d859e29f | 8272 | mutex_unlock(&ctx->mutex); |
9b51f66d | 8273 | |
fbfc623f YZ |
8274 | put_online_cpus(); |
8275 | ||
cdd6c482 | 8276 | event->owner = current; |
8882135b | 8277 | |
cdd6c482 IM |
8278 | mutex_lock(¤t->perf_event_mutex); |
8279 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
8280 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 8281 | |
c320c7b7 ACM |
8282 | /* |
8283 | * Precalculate sample_data sizes | |
8284 | */ | |
8285 | perf_event__header_size(event); | |
6844c09d | 8286 | perf_event__id_header_size(event); |
c320c7b7 | 8287 | |
8a49542c PZ |
8288 | /* |
8289 | * Drop the reference on the group_event after placing the | |
8290 | * new event on the sibling_list. This ensures destruction | |
8291 | * of the group leader will find the pointer to itself in | |
8292 | * perf_group_detach(). | |
8293 | */ | |
2903ff01 | 8294 | fdput(group); |
ea635c64 AV |
8295 | fd_install(event_fd, event_file); |
8296 | return event_fd; | |
0793a61d | 8297 | |
c3f00c70 | 8298 | err_context: |
fe4b04fa | 8299 | perf_unpin_context(ctx); |
ea635c64 | 8300 | put_ctx(ctx); |
c6be5a5c | 8301 | err_alloc: |
ea635c64 | 8302 | free_event(event); |
1f4ee503 | 8303 | err_cpus: |
fbfc623f | 8304 | put_online_cpus(); |
1f4ee503 | 8305 | err_task: |
e7d0bc04 PZ |
8306 | if (task) |
8307 | put_task_struct(task); | |
89a1e187 | 8308 | err_group_fd: |
2903ff01 | 8309 | fdput(group); |
ea635c64 AV |
8310 | err_fd: |
8311 | put_unused_fd(event_fd); | |
dc86cabe | 8312 | return err; |
0793a61d TG |
8313 | } |
8314 | ||
fb0459d7 AV |
8315 | /** |
8316 | * perf_event_create_kernel_counter | |
8317 | * | |
8318 | * @attr: attributes of the counter to create | |
8319 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 8320 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
8321 | */ |
8322 | struct perf_event * | |
8323 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 8324 | struct task_struct *task, |
4dc0da86 AK |
8325 | perf_overflow_handler_t overflow_handler, |
8326 | void *context) | |
fb0459d7 | 8327 | { |
fb0459d7 | 8328 | struct perf_event_context *ctx; |
c3f00c70 | 8329 | struct perf_event *event; |
fb0459d7 | 8330 | int err; |
d859e29f | 8331 | |
fb0459d7 AV |
8332 | /* |
8333 | * Get the target context (task or percpu): | |
8334 | */ | |
d859e29f | 8335 | |
4dc0da86 | 8336 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
79dff51e | 8337 | overflow_handler, context, -1); |
c3f00c70 PZ |
8338 | if (IS_ERR(event)) { |
8339 | err = PTR_ERR(event); | |
8340 | goto err; | |
8341 | } | |
d859e29f | 8342 | |
f8697762 JO |
8343 | /* Mark owner so we could distinguish it from user events. */ |
8344 | event->owner = EVENT_OWNER_KERNEL; | |
8345 | ||
766d6c07 FW |
8346 | account_event(event); |
8347 | ||
4af57ef2 | 8348 | ctx = find_get_context(event->pmu, task, event); |
c6567f64 FW |
8349 | if (IS_ERR(ctx)) { |
8350 | err = PTR_ERR(ctx); | |
c3f00c70 | 8351 | goto err_free; |
d859e29f | 8352 | } |
fb0459d7 | 8353 | |
fb0459d7 AV |
8354 | WARN_ON_ONCE(ctx->parent_ctx); |
8355 | mutex_lock(&ctx->mutex); | |
bed5b25a AS |
8356 | if (!exclusive_event_installable(event, ctx)) { |
8357 | mutex_unlock(&ctx->mutex); | |
8358 | perf_unpin_context(ctx); | |
8359 | put_ctx(ctx); | |
8360 | err = -EBUSY; | |
8361 | goto err_free; | |
8362 | } | |
8363 | ||
fb0459d7 | 8364 | perf_install_in_context(ctx, event, cpu); |
fe4b04fa | 8365 | perf_unpin_context(ctx); |
fb0459d7 AV |
8366 | mutex_unlock(&ctx->mutex); |
8367 | ||
fb0459d7 AV |
8368 | return event; |
8369 | ||
c3f00c70 PZ |
8370 | err_free: |
8371 | free_event(event); | |
8372 | err: | |
c6567f64 | 8373 | return ERR_PTR(err); |
9b51f66d | 8374 | } |
fb0459d7 | 8375 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 8376 | |
0cda4c02 YZ |
8377 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
8378 | { | |
8379 | struct perf_event_context *src_ctx; | |
8380 | struct perf_event_context *dst_ctx; | |
8381 | struct perf_event *event, *tmp; | |
8382 | LIST_HEAD(events); | |
8383 | ||
8384 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
8385 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
8386 | ||
f63a8daa PZ |
8387 | /* |
8388 | * See perf_event_ctx_lock() for comments on the details | |
8389 | * of swizzling perf_event::ctx. | |
8390 | */ | |
8391 | mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex); | |
0cda4c02 YZ |
8392 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, |
8393 | event_entry) { | |
46ce0fe9 | 8394 | perf_remove_from_context(event, false); |
9a545de0 | 8395 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 8396 | put_ctx(src_ctx); |
9886167d | 8397 | list_add(&event->migrate_entry, &events); |
0cda4c02 | 8398 | } |
0cda4c02 | 8399 | |
8f95b435 PZI |
8400 | /* |
8401 | * Wait for the events to quiesce before re-instating them. | |
8402 | */ | |
0cda4c02 YZ |
8403 | synchronize_rcu(); |
8404 | ||
8f95b435 PZI |
8405 | /* |
8406 | * Re-instate events in 2 passes. | |
8407 | * | |
8408 | * Skip over group leaders and only install siblings on this first | |
8409 | * pass, siblings will not get enabled without a leader, however a | |
8410 | * leader will enable its siblings, even if those are still on the old | |
8411 | * context. | |
8412 | */ | |
8413 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { | |
8414 | if (event->group_leader == event) | |
8415 | continue; | |
8416 | ||
8417 | list_del(&event->migrate_entry); | |
8418 | if (event->state >= PERF_EVENT_STATE_OFF) | |
8419 | event->state = PERF_EVENT_STATE_INACTIVE; | |
8420 | account_event_cpu(event, dst_cpu); | |
8421 | perf_install_in_context(dst_ctx, event, dst_cpu); | |
8422 | get_ctx(dst_ctx); | |
8423 | } | |
8424 | ||
8425 | /* | |
8426 | * Once all the siblings are setup properly, install the group leaders | |
8427 | * to make it go. | |
8428 | */ | |
9886167d PZ |
8429 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
8430 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
8431 | if (event->state >= PERF_EVENT_STATE_OFF) |
8432 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 8433 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
8434 | perf_install_in_context(dst_ctx, event, dst_cpu); |
8435 | get_ctx(dst_ctx); | |
8436 | } | |
8437 | mutex_unlock(&dst_ctx->mutex); | |
f63a8daa | 8438 | mutex_unlock(&src_ctx->mutex); |
0cda4c02 YZ |
8439 | } |
8440 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
8441 | ||
cdd6c482 | 8442 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 8443 | struct task_struct *child) |
d859e29f | 8444 | { |
cdd6c482 | 8445 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 8446 | u64 child_val; |
d859e29f | 8447 | |
cdd6c482 IM |
8448 | if (child_event->attr.inherit_stat) |
8449 | perf_event_read_event(child_event, child); | |
38b200d6 | 8450 | |
b5e58793 | 8451 | child_val = perf_event_count(child_event); |
d859e29f PM |
8452 | |
8453 | /* | |
8454 | * Add back the child's count to the parent's count: | |
8455 | */ | |
a6e6dea6 | 8456 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
8457 | atomic64_add(child_event->total_time_enabled, |
8458 | &parent_event->child_total_time_enabled); | |
8459 | atomic64_add(child_event->total_time_running, | |
8460 | &parent_event->child_total_time_running); | |
d859e29f PM |
8461 | |
8462 | /* | |
cdd6c482 | 8463 | * Remove this event from the parent's list |
d859e29f | 8464 | */ |
cdd6c482 IM |
8465 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
8466 | mutex_lock(&parent_event->child_mutex); | |
8467 | list_del_init(&child_event->child_list); | |
8468 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f | 8469 | |
dc633982 JO |
8470 | /* |
8471 | * Make sure user/parent get notified, that we just | |
8472 | * lost one event. | |
8473 | */ | |
8474 | perf_event_wakeup(parent_event); | |
8475 | ||
d859e29f | 8476 | /* |
cdd6c482 | 8477 | * Release the parent event, if this was the last |
d859e29f PM |
8478 | * reference to it. |
8479 | */ | |
a6fa941d | 8480 | put_event(parent_event); |
d859e29f PM |
8481 | } |
8482 | ||
9b51f66d | 8483 | static void |
cdd6c482 IM |
8484 | __perf_event_exit_task(struct perf_event *child_event, |
8485 | struct perf_event_context *child_ctx, | |
38b200d6 | 8486 | struct task_struct *child) |
9b51f66d | 8487 | { |
1903d50c PZ |
8488 | /* |
8489 | * Do not destroy the 'original' grouping; because of the context | |
8490 | * switch optimization the original events could've ended up in a | |
8491 | * random child task. | |
8492 | * | |
8493 | * If we were to destroy the original group, all group related | |
8494 | * operations would cease to function properly after this random | |
8495 | * child dies. | |
8496 | * | |
8497 | * Do destroy all inherited groups, we don't care about those | |
8498 | * and being thorough is better. | |
8499 | */ | |
8500 | perf_remove_from_context(child_event, !!child_event->parent); | |
0cc0c027 | 8501 | |
9b51f66d | 8502 | /* |
38b435b1 | 8503 | * It can happen that the parent exits first, and has events |
9b51f66d | 8504 | * that are still around due to the child reference. These |
38b435b1 | 8505 | * events need to be zapped. |
9b51f66d | 8506 | */ |
38b435b1 | 8507 | if (child_event->parent) { |
cdd6c482 IM |
8508 | sync_child_event(child_event, child); |
8509 | free_event(child_event); | |
179033b3 JO |
8510 | } else { |
8511 | child_event->state = PERF_EVENT_STATE_EXIT; | |
8512 | perf_event_wakeup(child_event); | |
4bcf349a | 8513 | } |
9b51f66d IM |
8514 | } |
8515 | ||
8dc85d54 | 8516 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 8517 | { |
ebf905fc | 8518 | struct perf_event *child_event, *next; |
211de6eb | 8519 | struct perf_event_context *child_ctx, *clone_ctx = NULL; |
a63eaf34 | 8520 | unsigned long flags; |
9b51f66d | 8521 | |
8dc85d54 | 8522 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 8523 | perf_event_task(child, NULL, 0); |
9b51f66d | 8524 | return; |
9f498cc5 | 8525 | } |
9b51f66d | 8526 | |
a63eaf34 | 8527 | local_irq_save(flags); |
ad3a37de PM |
8528 | /* |
8529 | * We can't reschedule here because interrupts are disabled, | |
8530 | * and either child is current or it is a task that can't be | |
8531 | * scheduled, so we are now safe from rescheduling changing | |
8532 | * our context. | |
8533 | */ | |
806839b2 | 8534 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
8535 | |
8536 | /* | |
8537 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 8538 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
8539 | * incremented the context's refcount before we do put_ctx below. |
8540 | */ | |
e625cce1 | 8541 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 8542 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 8543 | child->perf_event_ctxp[ctxn] = NULL; |
4a1c0f26 | 8544 | |
71a851b4 PZ |
8545 | /* |
8546 | * If this context is a clone; unclone it so it can't get | |
8547 | * swapped to another process while we're removing all | |
cdd6c482 | 8548 | * the events from it. |
71a851b4 | 8549 | */ |
211de6eb | 8550 | clone_ctx = unclone_ctx(child_ctx); |
5e942bb3 | 8551 | update_context_time(child_ctx); |
e625cce1 | 8552 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 | 8553 | |
211de6eb PZ |
8554 | if (clone_ctx) |
8555 | put_ctx(clone_ctx); | |
4a1c0f26 | 8556 | |
9f498cc5 | 8557 | /* |
cdd6c482 IM |
8558 | * Report the task dead after unscheduling the events so that we |
8559 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
8560 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 8561 | */ |
cdd6c482 | 8562 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 8563 | |
66fff224 PZ |
8564 | /* |
8565 | * We can recurse on the same lock type through: | |
8566 | * | |
cdd6c482 IM |
8567 | * __perf_event_exit_task() |
8568 | * sync_child_event() | |
a6fa941d AV |
8569 | * put_event() |
8570 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
8571 | * |
8572 | * But since its the parent context it won't be the same instance. | |
8573 | */ | |
a0507c84 | 8574 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 8575 | |
ebf905fc | 8576 | list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry) |
cdd6c482 | 8577 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 | 8578 | |
a63eaf34 PM |
8579 | mutex_unlock(&child_ctx->mutex); |
8580 | ||
8581 | put_ctx(child_ctx); | |
9b51f66d IM |
8582 | } |
8583 | ||
8dc85d54 PZ |
8584 | /* |
8585 | * When a child task exits, feed back event values to parent events. | |
8586 | */ | |
8587 | void perf_event_exit_task(struct task_struct *child) | |
8588 | { | |
8882135b | 8589 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
8590 | int ctxn; |
8591 | ||
8882135b PZ |
8592 | mutex_lock(&child->perf_event_mutex); |
8593 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
8594 | owner_entry) { | |
8595 | list_del_init(&event->owner_entry); | |
8596 | ||
8597 | /* | |
8598 | * Ensure the list deletion is visible before we clear | |
8599 | * the owner, closes a race against perf_release() where | |
8600 | * we need to serialize on the owner->perf_event_mutex. | |
8601 | */ | |
8602 | smp_wmb(); | |
8603 | event->owner = NULL; | |
8604 | } | |
8605 | mutex_unlock(&child->perf_event_mutex); | |
8606 | ||
8dc85d54 PZ |
8607 | for_each_task_context_nr(ctxn) |
8608 | perf_event_exit_task_context(child, ctxn); | |
8609 | } | |
8610 | ||
889ff015 FW |
8611 | static void perf_free_event(struct perf_event *event, |
8612 | struct perf_event_context *ctx) | |
8613 | { | |
8614 | struct perf_event *parent = event->parent; | |
8615 | ||
8616 | if (WARN_ON_ONCE(!parent)) | |
8617 | return; | |
8618 | ||
8619 | mutex_lock(&parent->child_mutex); | |
8620 | list_del_init(&event->child_list); | |
8621 | mutex_unlock(&parent->child_mutex); | |
8622 | ||
a6fa941d | 8623 | put_event(parent); |
889ff015 | 8624 | |
652884fe | 8625 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 8626 | perf_group_detach(event); |
889ff015 | 8627 | list_del_event(event, ctx); |
652884fe | 8628 | raw_spin_unlock_irq(&ctx->lock); |
889ff015 FW |
8629 | free_event(event); |
8630 | } | |
8631 | ||
bbbee908 | 8632 | /* |
652884fe | 8633 | * Free an unexposed, unused context as created by inheritance by |
8dc85d54 | 8634 | * perf_event_init_task below, used by fork() in case of fail. |
652884fe PZ |
8635 | * |
8636 | * Not all locks are strictly required, but take them anyway to be nice and | |
8637 | * help out with the lockdep assertions. | |
bbbee908 | 8638 | */ |
cdd6c482 | 8639 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 8640 | { |
8dc85d54 | 8641 | struct perf_event_context *ctx; |
cdd6c482 | 8642 | struct perf_event *event, *tmp; |
8dc85d54 | 8643 | int ctxn; |
bbbee908 | 8644 | |
8dc85d54 PZ |
8645 | for_each_task_context_nr(ctxn) { |
8646 | ctx = task->perf_event_ctxp[ctxn]; | |
8647 | if (!ctx) | |
8648 | continue; | |
bbbee908 | 8649 | |
8dc85d54 | 8650 | mutex_lock(&ctx->mutex); |
bbbee908 | 8651 | again: |
8dc85d54 PZ |
8652 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
8653 | group_entry) | |
8654 | perf_free_event(event, ctx); | |
bbbee908 | 8655 | |
8dc85d54 PZ |
8656 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
8657 | group_entry) | |
8658 | perf_free_event(event, ctx); | |
bbbee908 | 8659 | |
8dc85d54 PZ |
8660 | if (!list_empty(&ctx->pinned_groups) || |
8661 | !list_empty(&ctx->flexible_groups)) | |
8662 | goto again; | |
bbbee908 | 8663 | |
8dc85d54 | 8664 | mutex_unlock(&ctx->mutex); |
bbbee908 | 8665 | |
8dc85d54 PZ |
8666 | put_ctx(ctx); |
8667 | } | |
889ff015 FW |
8668 | } |
8669 | ||
4e231c79 PZ |
8670 | void perf_event_delayed_put(struct task_struct *task) |
8671 | { | |
8672 | int ctxn; | |
8673 | ||
8674 | for_each_task_context_nr(ctxn) | |
8675 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
8676 | } | |
8677 | ||
97dee4f3 PZ |
8678 | /* |
8679 | * inherit a event from parent task to child task: | |
8680 | */ | |
8681 | static struct perf_event * | |
8682 | inherit_event(struct perf_event *parent_event, | |
8683 | struct task_struct *parent, | |
8684 | struct perf_event_context *parent_ctx, | |
8685 | struct task_struct *child, | |
8686 | struct perf_event *group_leader, | |
8687 | struct perf_event_context *child_ctx) | |
8688 | { | |
1929def9 | 8689 | enum perf_event_active_state parent_state = parent_event->state; |
97dee4f3 | 8690 | struct perf_event *child_event; |
cee010ec | 8691 | unsigned long flags; |
97dee4f3 PZ |
8692 | |
8693 | /* | |
8694 | * Instead of creating recursive hierarchies of events, | |
8695 | * we link inherited events back to the original parent, | |
8696 | * which has a filp for sure, which we use as the reference | |
8697 | * count: | |
8698 | */ | |
8699 | if (parent_event->parent) | |
8700 | parent_event = parent_event->parent; | |
8701 | ||
8702 | child_event = perf_event_alloc(&parent_event->attr, | |
8703 | parent_event->cpu, | |
d580ff86 | 8704 | child, |
97dee4f3 | 8705 | group_leader, parent_event, |
79dff51e | 8706 | NULL, NULL, -1); |
97dee4f3 PZ |
8707 | if (IS_ERR(child_event)) |
8708 | return child_event; | |
a6fa941d | 8709 | |
fadfe7be JO |
8710 | if (is_orphaned_event(parent_event) || |
8711 | !atomic_long_inc_not_zero(&parent_event->refcount)) { | |
a6fa941d AV |
8712 | free_event(child_event); |
8713 | return NULL; | |
8714 | } | |
8715 | ||
97dee4f3 PZ |
8716 | get_ctx(child_ctx); |
8717 | ||
8718 | /* | |
8719 | * Make the child state follow the state of the parent event, | |
8720 | * not its attr.disabled bit. We hold the parent's mutex, | |
8721 | * so we won't race with perf_event_{en, dis}able_family. | |
8722 | */ | |
1929def9 | 8723 | if (parent_state >= PERF_EVENT_STATE_INACTIVE) |
97dee4f3 PZ |
8724 | child_event->state = PERF_EVENT_STATE_INACTIVE; |
8725 | else | |
8726 | child_event->state = PERF_EVENT_STATE_OFF; | |
8727 | ||
8728 | if (parent_event->attr.freq) { | |
8729 | u64 sample_period = parent_event->hw.sample_period; | |
8730 | struct hw_perf_event *hwc = &child_event->hw; | |
8731 | ||
8732 | hwc->sample_period = sample_period; | |
8733 | hwc->last_period = sample_period; | |
8734 | ||
8735 | local64_set(&hwc->period_left, sample_period); | |
8736 | } | |
8737 | ||
8738 | child_event->ctx = child_ctx; | |
8739 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
8740 | child_event->overflow_handler_context |
8741 | = parent_event->overflow_handler_context; | |
97dee4f3 | 8742 | |
614b6780 TG |
8743 | /* |
8744 | * Precalculate sample_data sizes | |
8745 | */ | |
8746 | perf_event__header_size(child_event); | |
6844c09d | 8747 | perf_event__id_header_size(child_event); |
614b6780 | 8748 | |
97dee4f3 PZ |
8749 | /* |
8750 | * Link it up in the child's context: | |
8751 | */ | |
cee010ec | 8752 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 8753 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 8754 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 8755 | |
97dee4f3 PZ |
8756 | /* |
8757 | * Link this into the parent event's child list | |
8758 | */ | |
8759 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
8760 | mutex_lock(&parent_event->child_mutex); | |
8761 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
8762 | mutex_unlock(&parent_event->child_mutex); | |
8763 | ||
8764 | return child_event; | |
8765 | } | |
8766 | ||
8767 | static int inherit_group(struct perf_event *parent_event, | |
8768 | struct task_struct *parent, | |
8769 | struct perf_event_context *parent_ctx, | |
8770 | struct task_struct *child, | |
8771 | struct perf_event_context *child_ctx) | |
8772 | { | |
8773 | struct perf_event *leader; | |
8774 | struct perf_event *sub; | |
8775 | struct perf_event *child_ctr; | |
8776 | ||
8777 | leader = inherit_event(parent_event, parent, parent_ctx, | |
8778 | child, NULL, child_ctx); | |
8779 | if (IS_ERR(leader)) | |
8780 | return PTR_ERR(leader); | |
8781 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
8782 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
8783 | child, leader, child_ctx); | |
8784 | if (IS_ERR(child_ctr)) | |
8785 | return PTR_ERR(child_ctr); | |
8786 | } | |
8787 | return 0; | |
889ff015 FW |
8788 | } |
8789 | ||
8790 | static int | |
8791 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
8792 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 8793 | struct task_struct *child, int ctxn, |
889ff015 FW |
8794 | int *inherited_all) |
8795 | { | |
8796 | int ret; | |
8dc85d54 | 8797 | struct perf_event_context *child_ctx; |
889ff015 FW |
8798 | |
8799 | if (!event->attr.inherit) { | |
8800 | *inherited_all = 0; | |
8801 | return 0; | |
bbbee908 PZ |
8802 | } |
8803 | ||
fe4b04fa | 8804 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
8805 | if (!child_ctx) { |
8806 | /* | |
8807 | * This is executed from the parent task context, so | |
8808 | * inherit events that have been marked for cloning. | |
8809 | * First allocate and initialize a context for the | |
8810 | * child. | |
8811 | */ | |
bbbee908 | 8812 | |
734df5ab | 8813 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
8814 | if (!child_ctx) |
8815 | return -ENOMEM; | |
bbbee908 | 8816 | |
8dc85d54 | 8817 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
8818 | } |
8819 | ||
8820 | ret = inherit_group(event, parent, parent_ctx, | |
8821 | child, child_ctx); | |
8822 | ||
8823 | if (ret) | |
8824 | *inherited_all = 0; | |
8825 | ||
8826 | return ret; | |
bbbee908 PZ |
8827 | } |
8828 | ||
9b51f66d | 8829 | /* |
cdd6c482 | 8830 | * Initialize the perf_event context in task_struct |
9b51f66d | 8831 | */ |
985c8dcb | 8832 | static int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 8833 | { |
889ff015 | 8834 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
8835 | struct perf_event_context *cloned_ctx; |
8836 | struct perf_event *event; | |
9b51f66d | 8837 | struct task_struct *parent = current; |
564c2b21 | 8838 | int inherited_all = 1; |
dddd3379 | 8839 | unsigned long flags; |
6ab423e0 | 8840 | int ret = 0; |
9b51f66d | 8841 | |
8dc85d54 | 8842 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
8843 | return 0; |
8844 | ||
ad3a37de | 8845 | /* |
25346b93 PM |
8846 | * If the parent's context is a clone, pin it so it won't get |
8847 | * swapped under us. | |
ad3a37de | 8848 | */ |
8dc85d54 | 8849 | parent_ctx = perf_pin_task_context(parent, ctxn); |
ffb4ef21 PZ |
8850 | if (!parent_ctx) |
8851 | return 0; | |
25346b93 | 8852 | |
ad3a37de PM |
8853 | /* |
8854 | * No need to check if parent_ctx != NULL here; since we saw | |
8855 | * it non-NULL earlier, the only reason for it to become NULL | |
8856 | * is if we exit, and since we're currently in the middle of | |
8857 | * a fork we can't be exiting at the same time. | |
8858 | */ | |
ad3a37de | 8859 | |
9b51f66d IM |
8860 | /* |
8861 | * Lock the parent list. No need to lock the child - not PID | |
8862 | * hashed yet and not running, so nobody can access it. | |
8863 | */ | |
d859e29f | 8864 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
8865 | |
8866 | /* | |
8867 | * We dont have to disable NMIs - we are only looking at | |
8868 | * the list, not manipulating it: | |
8869 | */ | |
889ff015 | 8870 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
8871 | ret = inherit_task_group(event, parent, parent_ctx, |
8872 | child, ctxn, &inherited_all); | |
889ff015 FW |
8873 | if (ret) |
8874 | break; | |
8875 | } | |
b93f7978 | 8876 | |
dddd3379 TG |
8877 | /* |
8878 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
8879 | * to allocations, but we need to prevent rotation because | |
8880 | * rotate_ctx() will change the list from interrupt context. | |
8881 | */ | |
8882 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
8883 | parent_ctx->rotate_disable = 1; | |
8884 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
8885 | ||
889ff015 | 8886 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
8887 | ret = inherit_task_group(event, parent, parent_ctx, |
8888 | child, ctxn, &inherited_all); | |
889ff015 | 8889 | if (ret) |
9b51f66d | 8890 | break; |
564c2b21 PM |
8891 | } |
8892 | ||
dddd3379 TG |
8893 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
8894 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 8895 | |
8dc85d54 | 8896 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 8897 | |
05cbaa28 | 8898 | if (child_ctx && inherited_all) { |
564c2b21 PM |
8899 | /* |
8900 | * Mark the child context as a clone of the parent | |
8901 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
8902 | * |
8903 | * Note that if the parent is a clone, the holding of | |
8904 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 8905 | */ |
c5ed5145 | 8906 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
8907 | if (cloned_ctx) { |
8908 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 8909 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
8910 | } else { |
8911 | child_ctx->parent_ctx = parent_ctx; | |
8912 | child_ctx->parent_gen = parent_ctx->generation; | |
8913 | } | |
8914 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
8915 | } |
8916 | ||
c5ed5145 | 8917 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 8918 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 8919 | |
25346b93 | 8920 | perf_unpin_context(parent_ctx); |
fe4b04fa | 8921 | put_ctx(parent_ctx); |
ad3a37de | 8922 | |
6ab423e0 | 8923 | return ret; |
9b51f66d IM |
8924 | } |
8925 | ||
8dc85d54 PZ |
8926 | /* |
8927 | * Initialize the perf_event context in task_struct | |
8928 | */ | |
8929 | int perf_event_init_task(struct task_struct *child) | |
8930 | { | |
8931 | int ctxn, ret; | |
8932 | ||
8550d7cb ON |
8933 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
8934 | mutex_init(&child->perf_event_mutex); | |
8935 | INIT_LIST_HEAD(&child->perf_event_list); | |
8936 | ||
8dc85d54 PZ |
8937 | for_each_task_context_nr(ctxn) { |
8938 | ret = perf_event_init_context(child, ctxn); | |
6c72e350 PZ |
8939 | if (ret) { |
8940 | perf_event_free_task(child); | |
8dc85d54 | 8941 | return ret; |
6c72e350 | 8942 | } |
8dc85d54 PZ |
8943 | } |
8944 | ||
8945 | return 0; | |
8946 | } | |
8947 | ||
220b140b PM |
8948 | static void __init perf_event_init_all_cpus(void) |
8949 | { | |
b28ab83c | 8950 | struct swevent_htable *swhash; |
220b140b | 8951 | int cpu; |
220b140b PM |
8952 | |
8953 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
8954 | swhash = &per_cpu(swevent_htable, cpu); |
8955 | mutex_init(&swhash->hlist_mutex); | |
2fde4f94 | 8956 | INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu)); |
220b140b PM |
8957 | } |
8958 | } | |
8959 | ||
0db0628d | 8960 | static void perf_event_init_cpu(int cpu) |
0793a61d | 8961 | { |
108b02cf | 8962 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 8963 | |
b28ab83c | 8964 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 8965 | swhash->online = true; |
4536e4d1 | 8966 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
8967 | struct swevent_hlist *hlist; |
8968 | ||
b28ab83c PZ |
8969 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
8970 | WARN_ON(!hlist); | |
8971 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 8972 | } |
b28ab83c | 8973 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
8974 | } |
8975 | ||
c277443c | 8976 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
108b02cf | 8977 | static void __perf_event_exit_context(void *__info) |
0793a61d | 8978 | { |
226424ee | 8979 | struct remove_event re = { .detach_group = true }; |
108b02cf | 8980 | struct perf_event_context *ctx = __info; |
0793a61d | 8981 | |
e3703f8c | 8982 | rcu_read_lock(); |
46ce0fe9 PZ |
8983 | list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry) |
8984 | __perf_remove_from_context(&re); | |
e3703f8c | 8985 | rcu_read_unlock(); |
0793a61d | 8986 | } |
108b02cf PZ |
8987 | |
8988 | static void perf_event_exit_cpu_context(int cpu) | |
8989 | { | |
8990 | struct perf_event_context *ctx; | |
8991 | struct pmu *pmu; | |
8992 | int idx; | |
8993 | ||
8994 | idx = srcu_read_lock(&pmus_srcu); | |
8995 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 8996 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
8997 | |
8998 | mutex_lock(&ctx->mutex); | |
8999 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
9000 | mutex_unlock(&ctx->mutex); | |
9001 | } | |
9002 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
9003 | } |
9004 | ||
cdd6c482 | 9005 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 9006 | { |
b28ab83c | 9007 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 9008 | |
e3703f8c PZ |
9009 | perf_event_exit_cpu_context(cpu); |
9010 | ||
b28ab83c | 9011 | mutex_lock(&swhash->hlist_mutex); |
39af6b16 | 9012 | swhash->online = false; |
b28ab83c PZ |
9013 | swevent_hlist_release(swhash); |
9014 | mutex_unlock(&swhash->hlist_mutex); | |
0793a61d TG |
9015 | } |
9016 | #else | |
cdd6c482 | 9017 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
9018 | #endif |
9019 | ||
c277443c PZ |
9020 | static int |
9021 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
9022 | { | |
9023 | int cpu; | |
9024 | ||
9025 | for_each_online_cpu(cpu) | |
9026 | perf_event_exit_cpu(cpu); | |
9027 | ||
9028 | return NOTIFY_OK; | |
9029 | } | |
9030 | ||
9031 | /* | |
9032 | * Run the perf reboot notifier at the very last possible moment so that | |
9033 | * the generic watchdog code runs as long as possible. | |
9034 | */ | |
9035 | static struct notifier_block perf_reboot_notifier = { | |
9036 | .notifier_call = perf_reboot, | |
9037 | .priority = INT_MIN, | |
9038 | }; | |
9039 | ||
0db0628d | 9040 | static int |
0793a61d TG |
9041 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
9042 | { | |
9043 | unsigned int cpu = (long)hcpu; | |
9044 | ||
4536e4d1 | 9045 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
9046 | |
9047 | case CPU_UP_PREPARE: | |
5e11637e | 9048 | case CPU_DOWN_FAILED: |
cdd6c482 | 9049 | perf_event_init_cpu(cpu); |
0793a61d TG |
9050 | break; |
9051 | ||
5e11637e | 9052 | case CPU_UP_CANCELED: |
0793a61d | 9053 | case CPU_DOWN_PREPARE: |
cdd6c482 | 9054 | perf_event_exit_cpu(cpu); |
0793a61d | 9055 | break; |
0793a61d TG |
9056 | default: |
9057 | break; | |
9058 | } | |
9059 | ||
9060 | return NOTIFY_OK; | |
9061 | } | |
9062 | ||
cdd6c482 | 9063 | void __init perf_event_init(void) |
0793a61d | 9064 | { |
3c502e7a JW |
9065 | int ret; |
9066 | ||
2e80a82a PZ |
9067 | idr_init(&pmu_idr); |
9068 | ||
220b140b | 9069 | perf_event_init_all_cpus(); |
b0a873eb | 9070 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
9071 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
9072 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
9073 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
9074 | perf_tp_register(); |
9075 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 9076 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
9077 | |
9078 | ret = init_hw_breakpoint(); | |
9079 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
9080 | |
9081 | /* do not patch jump label more than once per second */ | |
9082 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
9083 | |
9084 | /* | |
9085 | * Build time assertion that we keep the data_head at the intended | |
9086 | * location. IOW, validation we got the __reserved[] size right. | |
9087 | */ | |
9088 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
9089 | != 1024); | |
0793a61d | 9090 | } |
abe43400 | 9091 | |
fd979c01 CS |
9092 | ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr, |
9093 | char *page) | |
9094 | { | |
9095 | struct perf_pmu_events_attr *pmu_attr = | |
9096 | container_of(attr, struct perf_pmu_events_attr, attr); | |
9097 | ||
9098 | if (pmu_attr->event_str) | |
9099 | return sprintf(page, "%s\n", pmu_attr->event_str); | |
9100 | ||
9101 | return 0; | |
9102 | } | |
9103 | ||
abe43400 PZ |
9104 | static int __init perf_event_sysfs_init(void) |
9105 | { | |
9106 | struct pmu *pmu; | |
9107 | int ret; | |
9108 | ||
9109 | mutex_lock(&pmus_lock); | |
9110 | ||
9111 | ret = bus_register(&pmu_bus); | |
9112 | if (ret) | |
9113 | goto unlock; | |
9114 | ||
9115 | list_for_each_entry(pmu, &pmus, entry) { | |
9116 | if (!pmu->name || pmu->type < 0) | |
9117 | continue; | |
9118 | ||
9119 | ret = pmu_dev_alloc(pmu); | |
9120 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
9121 | } | |
9122 | pmu_bus_running = 1; | |
9123 | ret = 0; | |
9124 | ||
9125 | unlock: | |
9126 | mutex_unlock(&pmus_lock); | |
9127 | ||
9128 | return ret; | |
9129 | } | |
9130 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
9131 | |
9132 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
9133 | static struct cgroup_subsys_state * |
9134 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
9135 | { |
9136 | struct perf_cgroup *jc; | |
e5d1367f | 9137 | |
1b15d055 | 9138 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
9139 | if (!jc) |
9140 | return ERR_PTR(-ENOMEM); | |
9141 | ||
e5d1367f SE |
9142 | jc->info = alloc_percpu(struct perf_cgroup_info); |
9143 | if (!jc->info) { | |
9144 | kfree(jc); | |
9145 | return ERR_PTR(-ENOMEM); | |
9146 | } | |
9147 | ||
e5d1367f SE |
9148 | return &jc->css; |
9149 | } | |
9150 | ||
eb95419b | 9151 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 9152 | { |
eb95419b TH |
9153 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
9154 | ||
e5d1367f SE |
9155 | free_percpu(jc->info); |
9156 | kfree(jc); | |
9157 | } | |
9158 | ||
9159 | static int __perf_cgroup_move(void *info) | |
9160 | { | |
9161 | struct task_struct *task = info; | |
9162 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
9163 | return 0; | |
9164 | } | |
9165 | ||
eb95419b TH |
9166 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
9167 | struct cgroup_taskset *tset) | |
e5d1367f | 9168 | { |
bb9d97b6 TH |
9169 | struct task_struct *task; |
9170 | ||
924f0d9a | 9171 | cgroup_taskset_for_each(task, tset) |
bb9d97b6 | 9172 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
9173 | } |
9174 | ||
eb95419b TH |
9175 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
9176 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 9177 | struct task_struct *task) |
e5d1367f SE |
9178 | { |
9179 | /* | |
9180 | * cgroup_exit() is called in the copy_process() failure path. | |
9181 | * Ignore this case since the task hasn't ran yet, this avoids | |
9182 | * trying to poke a half freed task state from generic code. | |
9183 | */ | |
9184 | if (!(task->flags & PF_EXITING)) | |
9185 | return; | |
9186 | ||
bb9d97b6 | 9187 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
9188 | } |
9189 | ||
073219e9 | 9190 | struct cgroup_subsys perf_event_cgrp_subsys = { |
92fb9748 TH |
9191 | .css_alloc = perf_cgroup_css_alloc, |
9192 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 9193 | .exit = perf_cgroup_exit, |
bb9d97b6 | 9194 | .attach = perf_cgroup_attach, |
e5d1367f SE |
9195 | }; |
9196 | #endif /* CONFIG_CGROUP_PERF */ |