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