core.c 187.4 KB
Newer Older
T
Thomas Gleixner 已提交
1
/*
I
Ingo Molnar 已提交
2
 * Performance events core code:
T
Thomas Gleixner 已提交
3
 *
4
 *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 6
 *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
 *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
A
Al Viro 已提交
7
 *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8
 *
I
Ingo Molnar 已提交
9
 * For licensing details see kernel-base/COPYING
T
Thomas Gleixner 已提交
10 11 12
 */

#include <linux/fs.h>
13
#include <linux/mm.h>
T
Thomas Gleixner 已提交
14 15
#include <linux/cpu.h>
#include <linux/smp.h>
P
Peter Zijlstra 已提交
16
#include <linux/idr.h>
17
#include <linux/file.h>
T
Thomas Gleixner 已提交
18
#include <linux/poll.h>
19
#include <linux/slab.h>
20
#include <linux/hash.h>
21
#include <linux/tick.h>
T
Thomas Gleixner 已提交
22
#include <linux/sysfs.h>
23
#include <linux/dcache.h>
T
Thomas Gleixner 已提交
24
#include <linux/percpu.h>
25
#include <linux/ptrace.h>
P
Peter Zijlstra 已提交
26
#include <linux/reboot.h>
27
#include <linux/vmstat.h>
P
Peter Zijlstra 已提交
28
#include <linux/device.h>
29
#include <linux/export.h>
30
#include <linux/vmalloc.h>
31 32
#include <linux/hardirq.h>
#include <linux/rculist.h>
T
Thomas Gleixner 已提交
33 34 35
#include <linux/uaccess.h>
#include <linux/syscalls.h>
#include <linux/anon_inodes.h>
I
Ingo Molnar 已提交
36
#include <linux/kernel_stat.h>
37
#include <linux/perf_event.h>
L
Li Zefan 已提交
38
#include <linux/ftrace_event.h>
39
#include <linux/hw_breakpoint.h>
40
#include <linux/mm_types.h>
41
#include <linux/cgroup.h>
42
#include <linux/module.h>
T
Thomas Gleixner 已提交
43

44 45
#include "internal.h"

46 47
#include <asm/irq_regs.h>

48
struct remote_function_call {
49 50 51 52
	struct task_struct	*p;
	int			(*func)(void *info);
	void			*info;
	int			ret;
53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
};

static void remote_function(void *data)
{
	struct remote_function_call *tfc = data;
	struct task_struct *p = tfc->p;

	if (p) {
		tfc->ret = -EAGAIN;
		if (task_cpu(p) != smp_processor_id() || !task_curr(p))
			return;
	}

	tfc->ret = tfc->func(tfc->info);
}

/**
 * task_function_call - call a function on the cpu on which a task runs
 * @p:		the task to evaluate
 * @func:	the function to be called
 * @info:	the function call argument
 *
 * Calls the function @func when the task is currently running. This might
 * be on the current CPU, which just calls the function directly
 *
 * returns: @func return value, or
 *	    -ESRCH  - when the process isn't running
 *	    -EAGAIN - when the process moved away
 */
static int
task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
{
	struct remote_function_call data = {
86 87 88 89
		.p	= p,
		.func	= func,
		.info	= info,
		.ret	= -ESRCH, /* No such (running) process */
90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
	};

	if (task_curr(p))
		smp_call_function_single(task_cpu(p), remote_function, &data, 1);

	return data.ret;
}

/**
 * cpu_function_call - call a function on the cpu
 * @func:	the function to be called
 * @info:	the function call argument
 *
 * Calls the function @func on the remote cpu.
 *
 * returns: @func return value or -ENXIO when the cpu is offline
 */
static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
{
	struct remote_function_call data = {
110 111 112 113
		.p	= NULL,
		.func	= func,
		.info	= info,
		.ret	= -ENXIO, /* No such CPU */
114 115 116 117 118 119 120
	};

	smp_call_function_single(cpu, remote_function, &data, 1);

	return data.ret;
}

S
Stephane Eranian 已提交
121 122
#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
		       PERF_FLAG_FD_OUTPUT  |\
123 124
		       PERF_FLAG_PID_CGROUP |\
		       PERF_FLAG_FD_CLOEXEC)
S
Stephane Eranian 已提交
125

126 127 128 129 130 131 132
/*
 * branch priv levels that need permission checks
 */
#define PERF_SAMPLE_BRANCH_PERM_PLM \
	(PERF_SAMPLE_BRANCH_KERNEL |\
	 PERF_SAMPLE_BRANCH_HV)

133 134 135 136 137 138
enum event_type_t {
	EVENT_FLEXIBLE = 0x1,
	EVENT_PINNED = 0x2,
	EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
};

S
Stephane Eranian 已提交
139 140 141 142
/*
 * perf_sched_events : >0 events exist
 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
 */
143
struct static_key_deferred perf_sched_events __read_mostly;
S
Stephane Eranian 已提交
144
static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
145
static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
S
Stephane Eranian 已提交
146

147 148 149
static atomic_t nr_mmap_events __read_mostly;
static atomic_t nr_comm_events __read_mostly;
static atomic_t nr_task_events __read_mostly;
150
static atomic_t nr_freq_events __read_mostly;
151

P
Peter Zijlstra 已提交
152 153 154 155
static LIST_HEAD(pmus);
static DEFINE_MUTEX(pmus_lock);
static struct srcu_struct pmus_srcu;

156
/*
157
 * perf event paranoia level:
158 159
 *  -1 - not paranoid at all
 *   0 - disallow raw tracepoint access for unpriv
160
 *   1 - disallow cpu events for unpriv
161
 *   2 - disallow kernel profiling for unpriv
162
 */
163
int sysctl_perf_event_paranoid __read_mostly = 1;
164

165 166
/* Minimum for 512 kiB + 1 user control page */
int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
167 168

/*
169
 * max perf event sample rate
170
 */
171 172 173 174 175 176 177 178 179
#define DEFAULT_MAX_SAMPLE_RATE		100000
#define DEFAULT_SAMPLE_PERIOD_NS	(NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
#define DEFAULT_CPU_TIME_MAX_PERCENT	25

int sysctl_perf_event_sample_rate __read_mostly	= DEFAULT_MAX_SAMPLE_RATE;

static int max_samples_per_tick __read_mostly	= DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
static int perf_sample_period_ns __read_mostly	= DEFAULT_SAMPLE_PERIOD_NS;

P
Peter Zijlstra 已提交
180 181
static int perf_sample_allowed_ns __read_mostly =
	DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
182 183 184 185 186 187

void update_perf_cpu_limits(void)
{
	u64 tmp = perf_sample_period_ns;

	tmp *= sysctl_perf_cpu_time_max_percent;
188
	do_div(tmp, 100);
P
Peter Zijlstra 已提交
189
	ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
190
}
P
Peter Zijlstra 已提交
191

192 193
static int perf_rotate_context(struct perf_cpu_context *cpuctx);

P
Peter Zijlstra 已提交
194 195 196 197
int perf_proc_update_handler(struct ctl_table *table, int write,
		void __user *buffer, size_t *lenp,
		loff_t *ppos)
{
198
	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
P
Peter Zijlstra 已提交
199 200 201 202 203

	if (ret || !write)
		return ret;

	max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
	perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
	update_perf_cpu_limits();

	return 0;
}

int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;

int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
				void __user *buffer, size_t *lenp,
				loff_t *ppos)
{
	int ret = proc_dointvec(table, write, buffer, lenp, ppos);

	if (ret || !write)
		return ret;

	update_perf_cpu_limits();
P
Peter Zijlstra 已提交
222 223 224

	return 0;
}
225

226 227 228 229 230 231 232
/*
 * perf samples are done in some very critical code paths (NMIs).
 * If they take too much CPU time, the system can lock up and not
 * get any real work done.  This will drop the sample rate when
 * we detect that events are taking too long.
 */
#define NR_ACCUMULATED_SAMPLES 128
P
Peter Zijlstra 已提交
233
static DEFINE_PER_CPU(u64, running_sample_length);
234

235
static void perf_duration_warn(struct irq_work *w)
236
{
237
	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
238
	u64 avg_local_sample_len;
239
	u64 local_samples_len;
240 241 242 243 244 245 246

	local_samples_len = __get_cpu_var(running_sample_length);
	avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;

	printk_ratelimited(KERN_WARNING
			"perf interrupt took too long (%lld > %lld), lowering "
			"kernel.perf_event_max_sample_rate to %d\n",
247
			avg_local_sample_len, allowed_ns >> 1,
248 249 250 251 252 253 254
			sysctl_perf_event_sample_rate);
}

static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);

void perf_sample_event_took(u64 sample_len_ns)
{
P
Peter Zijlstra 已提交
255
	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
256 257
	u64 avg_local_sample_len;
	u64 local_samples_len;
258

P
Peter Zijlstra 已提交
259
	if (allowed_ns == 0)
260 261 262 263 264 265 266 267 268 269 270 271 272 273 274
		return;

	/* decay the counter by 1 average sample */
	local_samples_len = __get_cpu_var(running_sample_length);
	local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
	local_samples_len += sample_len_ns;
	__get_cpu_var(running_sample_length) = local_samples_len;

	/*
	 * note: this will be biased artifically low until we have
	 * seen NR_ACCUMULATED_SAMPLES.  Doing it this way keeps us
	 * from having to maintain a count.
	 */
	avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;

P
Peter Zijlstra 已提交
275
	if (avg_local_sample_len <= allowed_ns)
276 277 278 279 280 281 282 283 284 285
		return;

	if (max_samples_per_tick <= 1)
		return;

	max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
	sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
	perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;

	update_perf_cpu_limits();
286

287 288 289 290 291 292
	if (!irq_work_queue(&perf_duration_work)) {
		early_printk("perf interrupt took too long (%lld > %lld), lowering "
			     "kernel.perf_event_max_sample_rate to %d\n",
			     avg_local_sample_len, allowed_ns >> 1,
			     sysctl_perf_event_sample_rate);
	}
293 294
}

295
static atomic64_t perf_event_id;
296

297 298 299 300
static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
			      enum event_type_t event_type);

static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
301 302 303 304 305
			     enum event_type_t event_type,
			     struct task_struct *task);

static void update_context_time(struct perf_event_context *ctx);
static u64 perf_event_time(struct perf_event *event);
306

307
void __weak perf_event_print_debug(void)	{ }
T
Thomas Gleixner 已提交
308

309
extern __weak const char *perf_pmu_name(void)
T
Thomas Gleixner 已提交
310
{
311
	return "pmu";
T
Thomas Gleixner 已提交
312 313
}

314 315 316 317 318
static inline u64 perf_clock(void)
{
	return local_clock();
}

S
Stephane Eranian 已提交
319 320 321 322 323 324
static inline struct perf_cpu_context *
__get_cpu_context(struct perf_event_context *ctx)
{
	return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
}

325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340
static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
			  struct perf_event_context *ctx)
{
	raw_spin_lock(&cpuctx->ctx.lock);
	if (ctx)
		raw_spin_lock(&ctx->lock);
}

static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
			    struct perf_event_context *ctx)
{
	if (ctx)
		raw_spin_unlock(&ctx->lock);
	raw_spin_unlock(&cpuctx->ctx.lock);
}

S
Stephane Eranian 已提交
341 342
#ifdef CONFIG_CGROUP_PERF

343 344 345 346 347 348 349 350 351 352 353
/*
 * perf_cgroup_info keeps track of time_enabled for a cgroup.
 * This is a per-cpu dynamically allocated data structure.
 */
struct perf_cgroup_info {
	u64				time;
	u64				timestamp;
};

struct perf_cgroup {
	struct cgroup_subsys_state	css;
354
	struct perf_cgroup_info	__percpu *info;
355 356
};

357 358 359 360 361
/*
 * Must ensure cgroup is pinned (css_get) before calling
 * this function. In other words, we cannot call this function
 * if there is no cgroup event for the current CPU context.
 */
S
Stephane Eranian 已提交
362 363 364
static inline struct perf_cgroup *
perf_cgroup_from_task(struct task_struct *task)
{
365
	return container_of(task_css(task, perf_event_cgrp_id),
366
			    struct perf_cgroup, css);
S
Stephane Eranian 已提交
367 368 369 370 371 372 373 374
}

static inline bool
perf_cgroup_match(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);

375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390
	/* @event doesn't care about cgroup */
	if (!event->cgrp)
		return true;

	/* wants specific cgroup scope but @cpuctx isn't associated with any */
	if (!cpuctx->cgrp)
		return false;

	/*
	 * Cgroup scoping is recursive.  An event enabled for a cgroup is
	 * also enabled for all its descendant cgroups.  If @cpuctx's
	 * cgroup is a descendant of @event's (the test covers identity
	 * case), it's a match.
	 */
	return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
				    event->cgrp->css.cgroup);
S
Stephane Eranian 已提交
391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438
}

static inline void perf_put_cgroup(struct perf_event *event)
{
	css_put(&event->cgrp->css);
}

static inline void perf_detach_cgroup(struct perf_event *event)
{
	perf_put_cgroup(event);
	event->cgrp = NULL;
}

static inline int is_cgroup_event(struct perf_event *event)
{
	return event->cgrp != NULL;
}

static inline u64 perf_cgroup_event_time(struct perf_event *event)
{
	struct perf_cgroup_info *t;

	t = per_cpu_ptr(event->cgrp->info, event->cpu);
	return t->time;
}

static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
{
	struct perf_cgroup_info *info;
	u64 now;

	now = perf_clock();

	info = this_cpu_ptr(cgrp->info);

	info->time += now - info->timestamp;
	info->timestamp = now;
}

static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
{
	struct perf_cgroup *cgrp_out = cpuctx->cgrp;
	if (cgrp_out)
		__update_cgrp_time(cgrp_out);
}

static inline void update_cgrp_time_from_event(struct perf_event *event)
{
439 440
	struct perf_cgroup *cgrp;

S
Stephane Eranian 已提交
441
	/*
442 443
	 * ensure we access cgroup data only when needed and
	 * when we know the cgroup is pinned (css_get)
S
Stephane Eranian 已提交
444
	 */
445
	if (!is_cgroup_event(event))
S
Stephane Eranian 已提交
446 447
		return;

448 449 450 451 452 453
	cgrp = perf_cgroup_from_task(current);
	/*
	 * Do not update time when cgroup is not active
	 */
	if (cgrp == event->cgrp)
		__update_cgrp_time(event->cgrp);
S
Stephane Eranian 已提交
454 455 456
}

static inline void
457 458
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
459 460 461 462
{
	struct perf_cgroup *cgrp;
	struct perf_cgroup_info *info;

463 464 465 466 467 468
	/*
	 * ctx->lock held by caller
	 * ensure we do not access cgroup data
	 * unless we have the cgroup pinned (css_get)
	 */
	if (!task || !ctx->nr_cgroups)
S
Stephane Eranian 已提交
469 470 471 472
		return;

	cgrp = perf_cgroup_from_task(task);
	info = this_cpu_ptr(cgrp->info);
473
	info->timestamp = ctx->timestamp;
S
Stephane Eranian 已提交
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505
}

#define PERF_CGROUP_SWOUT	0x1 /* cgroup switch out every event */
#define PERF_CGROUP_SWIN	0x2 /* cgroup switch in events based on task */

/*
 * reschedule events based on the cgroup constraint of task.
 *
 * mode SWOUT : schedule out everything
 * mode SWIN : schedule in based on cgroup for next
 */
void perf_cgroup_switch(struct task_struct *task, int mode)
{
	struct perf_cpu_context *cpuctx;
	struct pmu *pmu;
	unsigned long flags;

	/*
	 * disable interrupts to avoid geting nr_cgroup
	 * changes via __perf_event_disable(). Also
	 * avoids preemption.
	 */
	local_irq_save(flags);

	/*
	 * we reschedule only in the presence of cgroup
	 * constrained events.
	 */
	rcu_read_lock();

	list_for_each_entry_rcu(pmu, &pmus, entry) {
		cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
506 507
		if (cpuctx->unique_pmu != pmu)
			continue; /* ensure we process each cpuctx once */
S
Stephane Eranian 已提交
508 509 510 511 512 513 514 515 516

		/*
		 * perf_cgroup_events says at least one
		 * context on this CPU has cgroup events.
		 *
		 * ctx->nr_cgroups reports the number of cgroup
		 * events for a context.
		 */
		if (cpuctx->ctx.nr_cgroups > 0) {
517 518
			perf_ctx_lock(cpuctx, cpuctx->task_ctx);
			perf_pmu_disable(cpuctx->ctx.pmu);
S
Stephane Eranian 已提交
519 520 521 522 523 524 525 526 527 528 529

			if (mode & PERF_CGROUP_SWOUT) {
				cpu_ctx_sched_out(cpuctx, EVENT_ALL);
				/*
				 * must not be done before ctxswout due
				 * to event_filter_match() in event_sched_out()
				 */
				cpuctx->cgrp = NULL;
			}

			if (mode & PERF_CGROUP_SWIN) {
530
				WARN_ON_ONCE(cpuctx->cgrp);
531 532 533 534
				/*
				 * set cgrp before ctxsw in to allow
				 * event_filter_match() to not have to pass
				 * task around
S
Stephane Eranian 已提交
535 536 537 538
				 */
				cpuctx->cgrp = perf_cgroup_from_task(task);
				cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
			}
539 540
			perf_pmu_enable(cpuctx->ctx.pmu);
			perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
S
Stephane Eranian 已提交
541 542 543 544 545 546 547 548
		}
	}

	rcu_read_unlock();

	local_irq_restore(flags);
}

549 550
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
551
{
552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573
	struct perf_cgroup *cgrp1;
	struct perf_cgroup *cgrp2 = NULL;

	/*
	 * we come here when we know perf_cgroup_events > 0
	 */
	cgrp1 = perf_cgroup_from_task(task);

	/*
	 * next is NULL when called from perf_event_enable_on_exec()
	 * that will systematically cause a cgroup_switch()
	 */
	if (next)
		cgrp2 = perf_cgroup_from_task(next);

	/*
	 * only schedule out current cgroup events if we know
	 * that we are switching to a different cgroup. Otherwise,
	 * do no touch the cgroup events.
	 */
	if (cgrp1 != cgrp2)
		perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
S
Stephane Eranian 已提交
574 575
}

576 577
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
578
{
579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596
	struct perf_cgroup *cgrp1;
	struct perf_cgroup *cgrp2 = NULL;

	/*
	 * we come here when we know perf_cgroup_events > 0
	 */
	cgrp1 = perf_cgroup_from_task(task);

	/* prev can never be NULL */
	cgrp2 = perf_cgroup_from_task(prev);

	/*
	 * only need to schedule in cgroup events if we are changing
	 * cgroup during ctxsw. Cgroup events were not scheduled
	 * out of ctxsw out if that was not the case.
	 */
	if (cgrp1 != cgrp2)
		perf_cgroup_switch(task, PERF_CGROUP_SWIN);
S
Stephane Eranian 已提交
597 598 599 600 601 602 603 604
}

static inline int perf_cgroup_connect(int fd, struct perf_event *event,
				      struct perf_event_attr *attr,
				      struct perf_event *group_leader)
{
	struct perf_cgroup *cgrp;
	struct cgroup_subsys_state *css;
605 606
	struct fd f = fdget(fd);
	int ret = 0;
S
Stephane Eranian 已提交
607

608
	if (!f.file)
S
Stephane Eranian 已提交
609 610
		return -EBADF;

611
	css = css_tryget_from_dir(f.file->f_dentry, &perf_event_cgrp_subsys);
612 613 614 615
	if (IS_ERR(css)) {
		ret = PTR_ERR(css);
		goto out;
	}
S
Stephane Eranian 已提交
616 617 618 619 620 621 622 623 624 625 626 627 628

	cgrp = container_of(css, struct perf_cgroup, css);
	event->cgrp = cgrp;

	/*
	 * all events in a group must monitor
	 * the same cgroup because a task belongs
	 * to only one perf cgroup at a time
	 */
	if (group_leader && group_leader->cgrp != cgrp) {
		perf_detach_cgroup(event);
		ret = -EINVAL;
	}
629
out:
630
	fdput(f);
S
Stephane Eranian 已提交
631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703
	return ret;
}

static inline void
perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
{
	struct perf_cgroup_info *t;
	t = per_cpu_ptr(event->cgrp->info, event->cpu);
	event->shadow_ctx_time = now - t->timestamp;
}

static inline void
perf_cgroup_defer_enabled(struct perf_event *event)
{
	/*
	 * when the current task's perf cgroup does not match
	 * the event's, we need to remember to call the
	 * perf_mark_enable() function the first time a task with
	 * a matching perf cgroup is scheduled in.
	 */
	if (is_cgroup_event(event) && !perf_cgroup_match(event))
		event->cgrp_defer_enabled = 1;
}

static inline void
perf_cgroup_mark_enabled(struct perf_event *event,
			 struct perf_event_context *ctx)
{
	struct perf_event *sub;
	u64 tstamp = perf_event_time(event);

	if (!event->cgrp_defer_enabled)
		return;

	event->cgrp_defer_enabled = 0;

	event->tstamp_enabled = tstamp - event->total_time_enabled;
	list_for_each_entry(sub, &event->sibling_list, group_entry) {
		if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
			sub->tstamp_enabled = tstamp - sub->total_time_enabled;
			sub->cgrp_defer_enabled = 0;
		}
	}
}
#else /* !CONFIG_CGROUP_PERF */

static inline bool
perf_cgroup_match(struct perf_event *event)
{
	return true;
}

static inline void perf_detach_cgroup(struct perf_event *event)
{}

static inline int is_cgroup_event(struct perf_event *event)
{
	return 0;
}

static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
{
	return 0;
}

static inline void update_cgrp_time_from_event(struct perf_event *event)
{
}

static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
{
}

704 705
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
706 707 708
{
}

709 710
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
711 712 713 714 715 716 717 718 719 720 721
{
}

static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
				      struct perf_event_attr *attr,
				      struct perf_event *group_leader)
{
	return -EINVAL;
}

static inline void
722 723
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753
{
}

void
perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
{
}

static inline void
perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
{
}

static inline u64 perf_cgroup_event_time(struct perf_event *event)
{
	return 0;
}

static inline void
perf_cgroup_defer_enabled(struct perf_event *event)
{
}

static inline void
perf_cgroup_mark_enabled(struct perf_event *event,
			 struct perf_event_context *ctx)
{
}
#endif

754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816
/*
 * set default to be dependent on timer tick just
 * like original code
 */
#define PERF_CPU_HRTIMER (1000 / HZ)
/*
 * function must be called with interrupts disbled
 */
static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
{
	struct perf_cpu_context *cpuctx;
	enum hrtimer_restart ret = HRTIMER_NORESTART;
	int rotations = 0;

	WARN_ON(!irqs_disabled());

	cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);

	rotations = perf_rotate_context(cpuctx);

	/*
	 * arm timer if needed
	 */
	if (rotations) {
		hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
		ret = HRTIMER_RESTART;
	}

	return ret;
}

/* CPU is going down */
void perf_cpu_hrtimer_cancel(int cpu)
{
	struct perf_cpu_context *cpuctx;
	struct pmu *pmu;
	unsigned long flags;

	if (WARN_ON(cpu != smp_processor_id()))
		return;

	local_irq_save(flags);

	rcu_read_lock();

	list_for_each_entry_rcu(pmu, &pmus, entry) {
		cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);

		if (pmu->task_ctx_nr == perf_sw_context)
			continue;

		hrtimer_cancel(&cpuctx->hrtimer);
	}

	rcu_read_unlock();

	local_irq_restore(flags);
}

static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
{
	struct hrtimer *hr = &cpuctx->hrtimer;
	struct pmu *pmu = cpuctx->ctx.pmu;
817
	int timer;
818 819 820 821 822

	/* no multiplexing needed for SW PMU */
	if (pmu->task_ctx_nr == perf_sw_context)
		return;

823 824 825 826 827 828 829 830 831
	/*
	 * check default is sane, if not set then force to
	 * default interval (1/tick)
	 */
	timer = pmu->hrtimer_interval_ms;
	if (timer < 1)
		timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;

	cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853

	hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
	hr->function = perf_cpu_hrtimer_handler;
}

static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
{
	struct hrtimer *hr = &cpuctx->hrtimer;
	struct pmu *pmu = cpuctx->ctx.pmu;

	/* not for SW PMU */
	if (pmu->task_ctx_nr == perf_sw_context)
		return;

	if (hrtimer_active(hr))
		return;

	if (!hrtimer_callback_running(hr))
		__hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
					 0, HRTIMER_MODE_REL_PINNED, 0);
}

P
Peter Zijlstra 已提交
854
void perf_pmu_disable(struct pmu *pmu)
855
{
P
Peter Zijlstra 已提交
856 857 858
	int *count = this_cpu_ptr(pmu->pmu_disable_count);
	if (!(*count)++)
		pmu->pmu_disable(pmu);
859 860
}

P
Peter Zijlstra 已提交
861
void perf_pmu_enable(struct pmu *pmu)
862
{
P
Peter Zijlstra 已提交
863 864 865
	int *count = this_cpu_ptr(pmu->pmu_disable_count);
	if (!--(*count))
		pmu->pmu_enable(pmu);
866 867
}

868 869 870 871 872 873 874
static DEFINE_PER_CPU(struct list_head, rotation_list);

/*
 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
 * because they're strictly cpu affine and rotate_start is called with IRQs
 * disabled, while rotate_context is called from IRQ context.
 */
P
Peter Zijlstra 已提交
875
static void perf_pmu_rotate_start(struct pmu *pmu)
876
{
P
Peter Zijlstra 已提交
877
	struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
878
	struct list_head *head = &__get_cpu_var(rotation_list);
879

880
	WARN_ON(!irqs_disabled());
881

882
	if (list_empty(&cpuctx->rotation_list))
883
		list_add(&cpuctx->rotation_list, head);
884 885
}

886
static void get_ctx(struct perf_event_context *ctx)
887
{
888
	WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
889 890
}

891
static void put_ctx(struct perf_event_context *ctx)
892
{
893 894 895
	if (atomic_dec_and_test(&ctx->refcount)) {
		if (ctx->parent_ctx)
			put_ctx(ctx->parent_ctx);
896 897
		if (ctx->task)
			put_task_struct(ctx->task);
898
		kfree_rcu(ctx, rcu_head);
899
	}
900 901
}

902
static void unclone_ctx(struct perf_event_context *ctx)
903 904 905 906 907
{
	if (ctx->parent_ctx) {
		put_ctx(ctx->parent_ctx);
		ctx->parent_ctx = NULL;
	}
908
	ctx->generation++;
909 910
}

911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932
static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
{
	/*
	 * only top level events have the pid namespace they were created in
	 */
	if (event->parent)
		event = event->parent;

	return task_tgid_nr_ns(p, event->ns);
}

static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
{
	/*
	 * only top level events have the pid namespace they were created in
	 */
	if (event->parent)
		event = event->parent;

	return task_pid_nr_ns(p, event->ns);
}

933
/*
934
 * If we inherit events we want to return the parent event id
935 936
 * to userspace.
 */
937
static u64 primary_event_id(struct perf_event *event)
938
{
939
	u64 id = event->id;
940

941 942
	if (event->parent)
		id = event->parent->id;
943 944 945 946

	return id;
}

947
/*
948
 * Get the perf_event_context for a task and lock it.
949 950 951
 * This has to cope with with the fact that until it is locked,
 * the context could get moved to another task.
 */
952
static struct perf_event_context *
P
Peter Zijlstra 已提交
953
perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
954
{
955
	struct perf_event_context *ctx;
956

P
Peter Zijlstra 已提交
957
retry:
958 959 960 961 962 963 964 965 966 967 968
	/*
	 * One of the few rules of preemptible RCU is that one cannot do
	 * rcu_read_unlock() while holding a scheduler (or nested) lock when
	 * part of the read side critical section was preemptible -- see
	 * rcu_read_unlock_special().
	 *
	 * Since ctx->lock nests under rq->lock we must ensure the entire read
	 * side critical section is non-preemptible.
	 */
	preempt_disable();
	rcu_read_lock();
P
Peter Zijlstra 已提交
969
	ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
970 971 972 973
	if (ctx) {
		/*
		 * If this context is a clone of another, it might
		 * get swapped for another underneath us by
974
		 * perf_event_task_sched_out, though the
975 976 977 978 979 980
		 * rcu_read_lock() protects us from any context
		 * getting freed.  Lock the context and check if it
		 * got swapped before we could get the lock, and retry
		 * if so.  If we locked the right context, then it
		 * can't get swapped on us any more.
		 */
981
		raw_spin_lock_irqsave(&ctx->lock, *flags);
P
Peter Zijlstra 已提交
982
		if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
983
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
984 985
			rcu_read_unlock();
			preempt_enable();
986 987
			goto retry;
		}
988 989

		if (!atomic_inc_not_zero(&ctx->refcount)) {
990
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
991 992
			ctx = NULL;
		}
993 994
	}
	rcu_read_unlock();
995
	preempt_enable();
996 997 998 999 1000 1001 1002 1003
	return ctx;
}

/*
 * Get the context for a task and increment its pin_count so it
 * can't get swapped to another task.  This also increments its
 * reference count so that the context can't get freed.
 */
P
Peter Zijlstra 已提交
1004 1005
static struct perf_event_context *
perf_pin_task_context(struct task_struct *task, int ctxn)
1006
{
1007
	struct perf_event_context *ctx;
1008 1009
	unsigned long flags;

P
Peter Zijlstra 已提交
1010
	ctx = perf_lock_task_context(task, ctxn, &flags);
1011 1012
	if (ctx) {
		++ctx->pin_count;
1013
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
1014 1015 1016 1017
	}
	return ctx;
}

1018
static void perf_unpin_context(struct perf_event_context *ctx)
1019 1020 1021
{
	unsigned long flags;

1022
	raw_spin_lock_irqsave(&ctx->lock, flags);
1023
	--ctx->pin_count;
1024
	raw_spin_unlock_irqrestore(&ctx->lock, flags);
1025 1026
}

1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037
/*
 * Update the record of the current time in a context.
 */
static void update_context_time(struct perf_event_context *ctx)
{
	u64 now = perf_clock();

	ctx->time += now - ctx->timestamp;
	ctx->timestamp = now;
}

1038 1039 1040
static u64 perf_event_time(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;
S
Stephane Eranian 已提交
1041 1042 1043 1044

	if (is_cgroup_event(event))
		return perf_cgroup_event_time(event);

1045 1046 1047
	return ctx ? ctx->time : 0;
}

1048 1049
/*
 * Update the total_time_enabled and total_time_running fields for a event.
1050
 * The caller of this function needs to hold the ctx->lock.
1051 1052 1053 1054 1055 1056 1057 1058 1059
 */
static void update_event_times(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;
	u64 run_end;

	if (event->state < PERF_EVENT_STATE_INACTIVE ||
	    event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
		return;
S
Stephane Eranian 已提交
1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070
	/*
	 * in cgroup mode, time_enabled represents
	 * the time the event was enabled AND active
	 * tasks were in the monitored cgroup. This is
	 * independent of the activity of the context as
	 * there may be a mix of cgroup and non-cgroup events.
	 *
	 * That is why we treat cgroup events differently
	 * here.
	 */
	if (is_cgroup_event(event))
1071
		run_end = perf_cgroup_event_time(event);
S
Stephane Eranian 已提交
1072 1073
	else if (ctx->is_active)
		run_end = ctx->time;
1074 1075 1076 1077
	else
		run_end = event->tstamp_stopped;

	event->total_time_enabled = run_end - event->tstamp_enabled;
1078 1079 1080 1081

	if (event->state == PERF_EVENT_STATE_INACTIVE)
		run_end = event->tstamp_stopped;
	else
1082
		run_end = perf_event_time(event);
1083 1084

	event->total_time_running = run_end - event->tstamp_running;
S
Stephane Eranian 已提交
1085

1086 1087
}

1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099
/*
 * Update total_time_enabled and total_time_running for all events in a group.
 */
static void update_group_times(struct perf_event *leader)
{
	struct perf_event *event;

	update_event_times(leader);
	list_for_each_entry(event, &leader->sibling_list, group_entry)
		update_event_times(event);
}

1100 1101 1102 1103 1104 1105 1106 1107 1108
static struct list_head *
ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
{
	if (event->attr.pinned)
		return &ctx->pinned_groups;
	else
		return &ctx->flexible_groups;
}

1109
/*
1110
 * Add a event from the lists for its context.
1111 1112
 * Must be called with ctx->mutex and ctx->lock held.
 */
1113
static void
1114
list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1115
{
1116 1117
	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
	event->attach_state |= PERF_ATTACH_CONTEXT;
1118 1119

	/*
1120 1121 1122
	 * If we're a stand alone event or group leader, we go to the context
	 * list, group events are kept attached to the group so that
	 * perf_group_detach can, at all times, locate all siblings.
1123
	 */
1124
	if (event->group_leader == event) {
1125 1126
		struct list_head *list;

1127 1128 1129
		if (is_software_event(event))
			event->group_flags |= PERF_GROUP_SOFTWARE;

1130 1131
		list = ctx_group_list(event, ctx);
		list_add_tail(&event->group_entry, list);
P
Peter Zijlstra 已提交
1132
	}
P
Peter Zijlstra 已提交
1133

1134
	if (is_cgroup_event(event))
S
Stephane Eranian 已提交
1135 1136
		ctx->nr_cgroups++;

1137 1138 1139
	if (has_branch_stack(event))
		ctx->nr_branch_stack++;

1140
	list_add_rcu(&event->event_entry, &ctx->event_list);
1141
	if (!ctx->nr_events)
P
Peter Zijlstra 已提交
1142
		perf_pmu_rotate_start(ctx->pmu);
1143 1144
	ctx->nr_events++;
	if (event->attr.inherit_stat)
1145
		ctx->nr_stat++;
1146 1147

	ctx->generation++;
1148 1149
}

J
Jiri Olsa 已提交
1150 1151 1152 1153 1154 1155 1156 1157 1158
/*
 * Initialize event state based on the perf_event_attr::disabled.
 */
static inline void perf_event__state_init(struct perf_event *event)
{
	event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
					      PERF_EVENT_STATE_INACTIVE;
}

1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197
/*
 * Called at perf_event creation and when events are attached/detached from a
 * group.
 */
static void perf_event__read_size(struct perf_event *event)
{
	int entry = sizeof(u64); /* value */
	int size = 0;
	int nr = 1;

	if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
		size += sizeof(u64);

	if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
		size += sizeof(u64);

	if (event->attr.read_format & PERF_FORMAT_ID)
		entry += sizeof(u64);

	if (event->attr.read_format & PERF_FORMAT_GROUP) {
		nr += event->group_leader->nr_siblings;
		size += sizeof(u64);
	}

	size += entry * nr;
	event->read_size = size;
}

static void perf_event__header_size(struct perf_event *event)
{
	struct perf_sample_data *data;
	u64 sample_type = event->attr.sample_type;
	u16 size = 0;

	perf_event__read_size(event);

	if (sample_type & PERF_SAMPLE_IP)
		size += sizeof(data->ip);

1198 1199 1200 1201 1202 1203
	if (sample_type & PERF_SAMPLE_ADDR)
		size += sizeof(data->addr);

	if (sample_type & PERF_SAMPLE_PERIOD)
		size += sizeof(data->period);

A
Andi Kleen 已提交
1204 1205 1206
	if (sample_type & PERF_SAMPLE_WEIGHT)
		size += sizeof(data->weight);

1207 1208 1209
	if (sample_type & PERF_SAMPLE_READ)
		size += event->read_size;

1210 1211 1212
	if (sample_type & PERF_SAMPLE_DATA_SRC)
		size += sizeof(data->data_src.val);

A
Andi Kleen 已提交
1213 1214 1215
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		size += sizeof(data->txn);

1216 1217 1218 1219 1220 1221 1222 1223 1224
	event->header_size = size;
}

static void perf_event__id_header_size(struct perf_event *event)
{
	struct perf_sample_data *data;
	u64 sample_type = event->attr.sample_type;
	u16 size = 0;

1225 1226 1227 1228 1229 1230
	if (sample_type & PERF_SAMPLE_TID)
		size += sizeof(data->tid_entry);

	if (sample_type & PERF_SAMPLE_TIME)
		size += sizeof(data->time);

1231 1232 1233
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		size += sizeof(data->id);

1234 1235 1236 1237 1238 1239 1240 1241 1242
	if (sample_type & PERF_SAMPLE_ID)
		size += sizeof(data->id);

	if (sample_type & PERF_SAMPLE_STREAM_ID)
		size += sizeof(data->stream_id);

	if (sample_type & PERF_SAMPLE_CPU)
		size += sizeof(data->cpu_entry);

1243
	event->id_header_size = size;
1244 1245
}

1246 1247
static void perf_group_attach(struct perf_event *event)
{
1248
	struct perf_event *group_leader = event->group_leader, *pos;
1249

P
Peter Zijlstra 已提交
1250 1251 1252 1253 1254 1255
	/*
	 * We can have double attach due to group movement in perf_event_open.
	 */
	if (event->attach_state & PERF_ATTACH_GROUP)
		return;

1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266
	event->attach_state |= PERF_ATTACH_GROUP;

	if (group_leader == event)
		return;

	if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
			!is_software_event(event))
		group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;

	list_add_tail(&event->group_entry, &group_leader->sibling_list);
	group_leader->nr_siblings++;
1267 1268 1269 1270 1271

	perf_event__header_size(group_leader);

	list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
		perf_event__header_size(pos);
1272 1273
}

1274
/*
1275
 * Remove a event from the lists for its context.
1276
 * Must be called with ctx->mutex and ctx->lock held.
1277
 */
1278
static void
1279
list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1280
{
1281
	struct perf_cpu_context *cpuctx;
1282 1283 1284 1285
	/*
	 * We can have double detach due to exit/hot-unplug + close.
	 */
	if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1286
		return;
1287 1288 1289

	event->attach_state &= ~PERF_ATTACH_CONTEXT;

1290
	if (is_cgroup_event(event)) {
S
Stephane Eranian 已提交
1291
		ctx->nr_cgroups--;
1292 1293 1294 1295 1296 1297 1298 1299 1300
		cpuctx = __get_cpu_context(ctx);
		/*
		 * if there are no more cgroup events
		 * then cler cgrp to avoid stale pointer
		 * in update_cgrp_time_from_cpuctx()
		 */
		if (!ctx->nr_cgroups)
			cpuctx->cgrp = NULL;
	}
S
Stephane Eranian 已提交
1301

1302 1303 1304
	if (has_branch_stack(event))
		ctx->nr_branch_stack--;

1305 1306
	ctx->nr_events--;
	if (event->attr.inherit_stat)
1307
		ctx->nr_stat--;
1308

1309
	list_del_rcu(&event->event_entry);
1310

1311 1312
	if (event->group_leader == event)
		list_del_init(&event->group_entry);
P
Peter Zijlstra 已提交
1313

1314
	update_group_times(event);
1315 1316 1317 1318 1319 1320 1321 1322 1323 1324

	/*
	 * If event was in error state, then keep it
	 * that way, otherwise bogus counts will be
	 * returned on read(). The only way to get out
	 * of error state is by explicit re-enabling
	 * of the event
	 */
	if (event->state > PERF_EVENT_STATE_OFF)
		event->state = PERF_EVENT_STATE_OFF;
1325 1326

	ctx->generation++;
1327 1328
}

1329
static void perf_group_detach(struct perf_event *event)
1330 1331
{
	struct perf_event *sibling, *tmp;
1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347
	struct list_head *list = NULL;

	/*
	 * We can have double detach due to exit/hot-unplug + close.
	 */
	if (!(event->attach_state & PERF_ATTACH_GROUP))
		return;

	event->attach_state &= ~PERF_ATTACH_GROUP;

	/*
	 * If this is a sibling, remove it from its group.
	 */
	if (event->group_leader != event) {
		list_del_init(&event->group_entry);
		event->group_leader->nr_siblings--;
1348
		goto out;
1349 1350 1351 1352
	}

	if (!list_empty(&event->group_entry))
		list = &event->group_entry;
1353

1354
	/*
1355 1356
	 * If this was a group event with sibling events then
	 * upgrade the siblings to singleton events by adding them
1357
	 * to whatever list we are on.
1358
	 */
1359
	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1360 1361
		if (list)
			list_move_tail(&sibling->group_entry, list);
1362
		sibling->group_leader = sibling;
1363 1364 1365

		/* Inherit group flags from the previous leader */
		sibling->group_flags = event->group_flags;
1366
	}
1367 1368 1369 1370 1371 1372

out:
	perf_event__header_size(event->group_leader);

	list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
		perf_event__header_size(tmp);
1373 1374
}

1375 1376 1377
static inline int
event_filter_match(struct perf_event *event)
{
S
Stephane Eranian 已提交
1378 1379
	return (event->cpu == -1 || event->cpu == smp_processor_id())
	    && perf_cgroup_match(event);
1380 1381
}

1382 1383
static void
event_sched_out(struct perf_event *event,
1384
		  struct perf_cpu_context *cpuctx,
1385
		  struct perf_event_context *ctx)
1386
{
1387
	u64 tstamp = perf_event_time(event);
1388 1389 1390 1391 1392 1393 1394 1395 1396
	u64 delta;
	/*
	 * An event which could not be activated because of
	 * filter mismatch still needs to have its timings
	 * maintained, otherwise bogus information is return
	 * via read() for time_enabled, time_running:
	 */
	if (event->state == PERF_EVENT_STATE_INACTIVE
	    && !event_filter_match(event)) {
S
Stephane Eranian 已提交
1397
		delta = tstamp - event->tstamp_stopped;
1398
		event->tstamp_running += delta;
1399
		event->tstamp_stopped = tstamp;
1400 1401
	}

1402
	if (event->state != PERF_EVENT_STATE_ACTIVE)
1403
		return;
1404

1405 1406
	perf_pmu_disable(event->pmu);

1407 1408 1409 1410
	event->state = PERF_EVENT_STATE_INACTIVE;
	if (event->pending_disable) {
		event->pending_disable = 0;
		event->state = PERF_EVENT_STATE_OFF;
1411
	}
1412
	event->tstamp_stopped = tstamp;
P
Peter Zijlstra 已提交
1413
	event->pmu->del(event, 0);
1414
	event->oncpu = -1;
1415

1416
	if (!is_software_event(event))
1417 1418
		cpuctx->active_oncpu--;
	ctx->nr_active--;
1419 1420
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq--;
1421
	if (event->attr.exclusive || !cpuctx->active_oncpu)
1422
		cpuctx->exclusive = 0;
1423 1424

	perf_pmu_enable(event->pmu);
1425 1426
}

1427
static void
1428
group_sched_out(struct perf_event *group_event,
1429
		struct perf_cpu_context *cpuctx,
1430
		struct perf_event_context *ctx)
1431
{
1432
	struct perf_event *event;
1433
	int state = group_event->state;
1434

1435
	event_sched_out(group_event, cpuctx, ctx);
1436 1437 1438 1439

	/*
	 * Schedule out siblings (if any):
	 */
1440 1441
	list_for_each_entry(event, &group_event->sibling_list, group_entry)
		event_sched_out(event, cpuctx, ctx);
1442

1443
	if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1444 1445 1446
		cpuctx->exclusive = 0;
}

1447 1448 1449 1450 1451
struct remove_event {
	struct perf_event *event;
	bool detach_group;
};

T
Thomas Gleixner 已提交
1452
/*
1453
 * Cross CPU call to remove a performance event
T
Thomas Gleixner 已提交
1454
 *
1455
 * We disable the event on the hardware level first. After that we
T
Thomas Gleixner 已提交
1456 1457
 * remove it from the context list.
 */
1458
static int __perf_remove_from_context(void *info)
T
Thomas Gleixner 已提交
1459
{
1460 1461
	struct remove_event *re = info;
	struct perf_event *event = re->event;
1462
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1463
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
T
Thomas Gleixner 已提交
1464

1465
	raw_spin_lock(&ctx->lock);
1466
	event_sched_out(event, cpuctx, ctx);
1467 1468
	if (re->detach_group)
		perf_group_detach(event);
1469
	list_del_event(event, ctx);
1470 1471 1472 1473
	if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
		ctx->is_active = 0;
		cpuctx->task_ctx = NULL;
	}
1474
	raw_spin_unlock(&ctx->lock);
1475 1476

	return 0;
T
Thomas Gleixner 已提交
1477 1478 1479 1480
}


/*
1481
 * Remove the event from a task's (or a CPU's) list of events.
T
Thomas Gleixner 已提交
1482
 *
1483
 * CPU events are removed with a smp call. For task events we only
T
Thomas Gleixner 已提交
1484
 * call when the task is on a CPU.
1485
 *
1486 1487
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1488 1489
 * remains valid.  This is OK when called from perf_release since
 * that only calls us on the top-level context, which can't be a clone.
1490
 * When called from perf_event_exit_task, it's OK because the
1491
 * context has been detached from its task.
T
Thomas Gleixner 已提交
1492
 */
1493
static void perf_remove_from_context(struct perf_event *event, bool detach_group)
T
Thomas Gleixner 已提交
1494
{
1495
	struct perf_event_context *ctx = event->ctx;
T
Thomas Gleixner 已提交
1496
	struct task_struct *task = ctx->task;
1497 1498 1499 1500
	struct remove_event re = {
		.event = event,
		.detach_group = detach_group,
	};
T
Thomas Gleixner 已提交
1501

1502 1503
	lockdep_assert_held(&ctx->mutex);

T
Thomas Gleixner 已提交
1504 1505
	if (!task) {
		/*
1506
		 * Per cpu events are removed via an smp call and
1507
		 * the removal is always successful.
T
Thomas Gleixner 已提交
1508
		 */
1509
		cpu_function_call(event->cpu, __perf_remove_from_context, &re);
T
Thomas Gleixner 已提交
1510 1511 1512 1513
		return;
	}

retry:
1514
	if (!task_function_call(task, __perf_remove_from_context, &re))
1515
		return;
T
Thomas Gleixner 已提交
1516

1517
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1518
	/*
1519 1520
	 * If we failed to find a running task, but find the context active now
	 * that we've acquired the ctx->lock, retry.
T
Thomas Gleixner 已提交
1521
	 */
1522
	if (ctx->is_active) {
1523
		raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1524 1525 1526 1527
		goto retry;
	}

	/*
1528 1529
	 * Since the task isn't running, its safe to remove the event, us
	 * holding the ctx->lock ensures the task won't get scheduled in.
T
Thomas Gleixner 已提交
1530
	 */
1531 1532
	if (detach_group)
		perf_group_detach(event);
1533
	list_del_event(event, ctx);
1534
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1535 1536
}

1537
/*
1538
 * Cross CPU call to disable a performance event
1539
 */
1540
int __perf_event_disable(void *info)
1541
{
1542 1543
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1544
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1545 1546

	/*
1547 1548
	 * If this is a per-task event, need to check whether this
	 * event's task is the current task on this cpu.
1549 1550 1551
	 *
	 * Can trigger due to concurrent perf_event_context_sched_out()
	 * flipping contexts around.
1552
	 */
1553
	if (ctx->task && cpuctx->task_ctx != ctx)
1554
		return -EINVAL;
1555

1556
	raw_spin_lock(&ctx->lock);
1557 1558

	/*
1559
	 * If the event is on, turn it off.
1560 1561
	 * If it is in error state, leave it in error state.
	 */
1562
	if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1563
		update_context_time(ctx);
S
Stephane Eranian 已提交
1564
		update_cgrp_time_from_event(event);
1565 1566 1567
		update_group_times(event);
		if (event == event->group_leader)
			group_sched_out(event, cpuctx, ctx);
1568
		else
1569 1570
			event_sched_out(event, cpuctx, ctx);
		event->state = PERF_EVENT_STATE_OFF;
1571 1572
	}

1573
	raw_spin_unlock(&ctx->lock);
1574 1575

	return 0;
1576 1577 1578
}

/*
1579
 * Disable a event.
1580
 *
1581 1582
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1583
 * remains valid.  This condition is satisifed when called through
1584 1585 1586 1587
 * perf_event_for_each_child or perf_event_for_each because they
 * hold the top-level event's child_mutex, so any descendant that
 * goes to exit will block in sync_child_event.
 * When called from perf_pending_event it's OK because event->ctx
1588
 * is the current context on this CPU and preemption is disabled,
1589
 * hence we can't get into perf_event_task_sched_out for this context.
1590
 */
1591
void perf_event_disable(struct perf_event *event)
1592
{
1593
	struct perf_event_context *ctx = event->ctx;
1594 1595 1596 1597
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
1598
		 * Disable the event on the cpu that it's on
1599
		 */
1600
		cpu_function_call(event->cpu, __perf_event_disable, event);
1601 1602 1603
		return;
	}

P
Peter Zijlstra 已提交
1604
retry:
1605 1606
	if (!task_function_call(task, __perf_event_disable, event))
		return;
1607

1608
	raw_spin_lock_irq(&ctx->lock);
1609
	/*
1610
	 * If the event is still active, we need to retry the cross-call.
1611
	 */
1612
	if (event->state == PERF_EVENT_STATE_ACTIVE) {
1613
		raw_spin_unlock_irq(&ctx->lock);
1614 1615 1616 1617 1618
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
1619 1620 1621 1622 1623 1624 1625
		goto retry;
	}

	/*
	 * Since we have the lock this context can't be scheduled
	 * in, so we can change the state safely.
	 */
1626 1627 1628
	if (event->state == PERF_EVENT_STATE_INACTIVE) {
		update_group_times(event);
		event->state = PERF_EVENT_STATE_OFF;
1629
	}
1630
	raw_spin_unlock_irq(&ctx->lock);
1631
}
1632
EXPORT_SYMBOL_GPL(perf_event_disable);
1633

S
Stephane Eranian 已提交
1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668
static void perf_set_shadow_time(struct perf_event *event,
				 struct perf_event_context *ctx,
				 u64 tstamp)
{
	/*
	 * use the correct time source for the time snapshot
	 *
	 * We could get by without this by leveraging the
	 * fact that to get to this function, the caller
	 * has most likely already called update_context_time()
	 * and update_cgrp_time_xx() and thus both timestamp
	 * are identical (or very close). Given that tstamp is,
	 * already adjusted for cgroup, we could say that:
	 *    tstamp - ctx->timestamp
	 * is equivalent to
	 *    tstamp - cgrp->timestamp.
	 *
	 * Then, in perf_output_read(), the calculation would
	 * work with no changes because:
	 * - event is guaranteed scheduled in
	 * - no scheduled out in between
	 * - thus the timestamp would be the same
	 *
	 * But this is a bit hairy.
	 *
	 * So instead, we have an explicit cgroup call to remain
	 * within the time time source all along. We believe it
	 * is cleaner and simpler to understand.
	 */
	if (is_cgroup_event(event))
		perf_cgroup_set_shadow_time(event, tstamp);
	else
		event->shadow_ctx_time = tstamp - ctx->timestamp;
}

P
Peter Zijlstra 已提交
1669 1670 1671 1672
#define MAX_INTERRUPTS (~0ULL)

static void perf_log_throttle(struct perf_event *event, int enable);

1673
static int
1674
event_sched_in(struct perf_event *event,
1675
		 struct perf_cpu_context *cpuctx,
1676
		 struct perf_event_context *ctx)
1677
{
1678
	u64 tstamp = perf_event_time(event);
1679
	int ret = 0;
1680

1681 1682
	lockdep_assert_held(&ctx->lock);

1683
	if (event->state <= PERF_EVENT_STATE_OFF)
1684 1685
		return 0;

1686
	event->state = PERF_EVENT_STATE_ACTIVE;
1687
	event->oncpu = smp_processor_id();
P
Peter Zijlstra 已提交
1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698

	/*
	 * Unthrottle events, since we scheduled we might have missed several
	 * ticks already, also for a heavily scheduling task there is little
	 * guarantee it'll get a tick in a timely manner.
	 */
	if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
		perf_log_throttle(event, 1);
		event->hw.interrupts = 0;
	}

1699 1700 1701 1702 1703
	/*
	 * The new state must be visible before we turn it on in the hardware:
	 */
	smp_wmb();

1704 1705
	perf_pmu_disable(event->pmu);

P
Peter Zijlstra 已提交
1706
	if (event->pmu->add(event, PERF_EF_START)) {
1707 1708
		event->state = PERF_EVENT_STATE_INACTIVE;
		event->oncpu = -1;
1709 1710
		ret = -EAGAIN;
		goto out;
1711 1712
	}

1713
	event->tstamp_running += tstamp - event->tstamp_stopped;
1714

S
Stephane Eranian 已提交
1715
	perf_set_shadow_time(event, ctx, tstamp);
1716

1717
	if (!is_software_event(event))
1718
		cpuctx->active_oncpu++;
1719
	ctx->nr_active++;
1720 1721
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq++;
1722

1723
	if (event->attr.exclusive)
1724 1725
		cpuctx->exclusive = 1;

1726 1727 1728 1729
out:
	perf_pmu_enable(event->pmu);

	return ret;
1730 1731
}

1732
static int
1733
group_sched_in(struct perf_event *group_event,
1734
	       struct perf_cpu_context *cpuctx,
1735
	       struct perf_event_context *ctx)
1736
{
1737
	struct perf_event *event, *partial_group = NULL;
P
Peter Zijlstra 已提交
1738
	struct pmu *pmu = ctx->pmu;
1739 1740
	u64 now = ctx->time;
	bool simulate = false;
1741

1742
	if (group_event->state == PERF_EVENT_STATE_OFF)
1743 1744
		return 0;

P
Peter Zijlstra 已提交
1745
	pmu->start_txn(pmu);
1746

1747
	if (event_sched_in(group_event, cpuctx, ctx)) {
P
Peter Zijlstra 已提交
1748
		pmu->cancel_txn(pmu);
1749
		perf_cpu_hrtimer_restart(cpuctx);
1750
		return -EAGAIN;
1751
	}
1752 1753 1754 1755

	/*
	 * Schedule in siblings as one group (if any):
	 */
1756
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1757
		if (event_sched_in(event, cpuctx, ctx)) {
1758
			partial_group = event;
1759 1760 1761 1762
			goto group_error;
		}
	}

1763
	if (!pmu->commit_txn(pmu))
1764
		return 0;
1765

1766 1767 1768 1769
group_error:
	/*
	 * Groups can be scheduled in as one unit only, so undo any
	 * partial group before returning:
1770 1771 1772 1773 1774 1775 1776 1777 1778 1779
	 * The events up to the failed event are scheduled out normally,
	 * tstamp_stopped will be updated.
	 *
	 * The failed events and the remaining siblings need to have
	 * their timings updated as if they had gone thru event_sched_in()
	 * and event_sched_out(). This is required to get consistent timings
	 * across the group. This also takes care of the case where the group
	 * could never be scheduled by ensuring tstamp_stopped is set to mark
	 * the time the event was actually stopped, such that time delta
	 * calculation in update_event_times() is correct.
1780
	 */
1781 1782
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
		if (event == partial_group)
1783 1784 1785 1786 1787 1788 1789 1790
			simulate = true;

		if (simulate) {
			event->tstamp_running += now - event->tstamp_stopped;
			event->tstamp_stopped = now;
		} else {
			event_sched_out(event, cpuctx, ctx);
		}
1791
	}
1792
	event_sched_out(group_event, cpuctx, ctx);
1793

P
Peter Zijlstra 已提交
1794
	pmu->cancel_txn(pmu);
1795

1796 1797
	perf_cpu_hrtimer_restart(cpuctx);

1798 1799 1800
	return -EAGAIN;
}

1801
/*
1802
 * Work out whether we can put this event group on the CPU now.
1803
 */
1804
static int group_can_go_on(struct perf_event *event,
1805 1806 1807 1808
			   struct perf_cpu_context *cpuctx,
			   int can_add_hw)
{
	/*
1809
	 * Groups consisting entirely of software events can always go on.
1810
	 */
1811
	if (event->group_flags & PERF_GROUP_SOFTWARE)
1812 1813 1814
		return 1;
	/*
	 * If an exclusive group is already on, no other hardware
1815
	 * events can go on.
1816 1817 1818 1819 1820
	 */
	if (cpuctx->exclusive)
		return 0;
	/*
	 * If this group is exclusive and there are already
1821
	 * events on the CPU, it can't go on.
1822
	 */
1823
	if (event->attr.exclusive && cpuctx->active_oncpu)
1824 1825 1826 1827 1828 1829 1830 1831
		return 0;
	/*
	 * Otherwise, try to add it if all previous groups were able
	 * to go on.
	 */
	return can_add_hw;
}

1832 1833
static void add_event_to_ctx(struct perf_event *event,
			       struct perf_event_context *ctx)
1834
{
1835 1836
	u64 tstamp = perf_event_time(event);

1837
	list_add_event(event, ctx);
1838
	perf_group_attach(event);
1839 1840 1841
	event->tstamp_enabled = tstamp;
	event->tstamp_running = tstamp;
	event->tstamp_stopped = tstamp;
1842 1843
}

1844 1845 1846 1847 1848 1849
static void task_ctx_sched_out(struct perf_event_context *ctx);
static void
ctx_sched_in(struct perf_event_context *ctx,
	     struct perf_cpu_context *cpuctx,
	     enum event_type_t event_type,
	     struct task_struct *task);
1850

1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862
static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
				struct perf_event_context *ctx,
				struct task_struct *task)
{
	cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
	if (ctx)
		ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
	cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
	if (ctx)
		ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
}

T
Thomas Gleixner 已提交
1863
/*
1864
 * Cross CPU call to install and enable a performance event
1865 1866
 *
 * Must be called with ctx->mutex held
T
Thomas Gleixner 已提交
1867
 */
1868
static int  __perf_install_in_context(void *info)
T
Thomas Gleixner 已提交
1869
{
1870 1871
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1872
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1873 1874 1875
	struct perf_event_context *task_ctx = cpuctx->task_ctx;
	struct task_struct *task = current;

1876
	perf_ctx_lock(cpuctx, task_ctx);
1877
	perf_pmu_disable(cpuctx->ctx.pmu);
T
Thomas Gleixner 已提交
1878 1879

	/*
1880
	 * If there was an active task_ctx schedule it out.
T
Thomas Gleixner 已提交
1881
	 */
1882
	if (task_ctx)
1883
		task_ctx_sched_out(task_ctx);
1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897

	/*
	 * If the context we're installing events in is not the
	 * active task_ctx, flip them.
	 */
	if (ctx->task && task_ctx != ctx) {
		if (task_ctx)
			raw_spin_unlock(&task_ctx->lock);
		raw_spin_lock(&ctx->lock);
		task_ctx = ctx;
	}

	if (task_ctx) {
		cpuctx->task_ctx = task_ctx;
1898 1899
		task = task_ctx->task;
	}
1900

1901
	cpu_ctx_sched_out(cpuctx, EVENT_ALL);
T
Thomas Gleixner 已提交
1902

1903
	update_context_time(ctx);
S
Stephane Eranian 已提交
1904 1905 1906 1907 1908 1909
	/*
	 * update cgrp time only if current cgrp
	 * matches event->cgrp. Must be done before
	 * calling add_event_to_ctx()
	 */
	update_cgrp_time_from_event(event);
T
Thomas Gleixner 已提交
1910

1911
	add_event_to_ctx(event, ctx);
T
Thomas Gleixner 已提交
1912

1913
	/*
1914
	 * Schedule everything back in
1915
	 */
1916
	perf_event_sched_in(cpuctx, task_ctx, task);
1917 1918 1919

	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, task_ctx);
1920 1921

	return 0;
T
Thomas Gleixner 已提交
1922 1923 1924
}

/*
1925
 * Attach a performance event to a context
T
Thomas Gleixner 已提交
1926
 *
1927 1928
 * First we add the event to the list with the hardware enable bit
 * in event->hw_config cleared.
T
Thomas Gleixner 已提交
1929
 *
1930
 * If the event is attached to a task which is on a CPU we use a smp
T
Thomas Gleixner 已提交
1931 1932 1933 1934
 * call to enable it in the task context. The task might have been
 * scheduled away, but we check this in the smp call again.
 */
static void
1935 1936
perf_install_in_context(struct perf_event_context *ctx,
			struct perf_event *event,
T
Thomas Gleixner 已提交
1937 1938 1939 1940
			int cpu)
{
	struct task_struct *task = ctx->task;

1941 1942
	lockdep_assert_held(&ctx->mutex);

1943
	event->ctx = ctx;
1944 1945
	if (event->cpu != -1)
		event->cpu = cpu;
1946

T
Thomas Gleixner 已提交
1947 1948
	if (!task) {
		/*
1949
		 * Per cpu events are installed via an smp call and
1950
		 * the install is always successful.
T
Thomas Gleixner 已提交
1951
		 */
1952
		cpu_function_call(cpu, __perf_install_in_context, event);
T
Thomas Gleixner 已提交
1953 1954 1955 1956
		return;
	}

retry:
1957 1958
	if (!task_function_call(task, __perf_install_in_context, event))
		return;
T
Thomas Gleixner 已提交
1959

1960
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1961
	/*
1962 1963
	 * If we failed to find a running task, but find the context active now
	 * that we've acquired the ctx->lock, retry.
T
Thomas Gleixner 已提交
1964
	 */
1965
	if (ctx->is_active) {
1966
		raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1967 1968 1969 1970
		goto retry;
	}

	/*
1971 1972
	 * Since the task isn't running, its safe to add the event, us holding
	 * the ctx->lock ensures the task won't get scheduled in.
T
Thomas Gleixner 已提交
1973
	 */
1974
	add_event_to_ctx(event, ctx);
1975
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1976 1977
}

1978
/*
1979
 * Put a event into inactive state and update time fields.
1980 1981 1982 1983 1984 1985
 * Enabling the leader of a group effectively enables all
 * the group members that aren't explicitly disabled, so we
 * have to update their ->tstamp_enabled also.
 * Note: this works for group members as well as group leaders
 * since the non-leader members' sibling_lists will be empty.
 */
1986
static void __perf_event_mark_enabled(struct perf_event *event)
1987
{
1988
	struct perf_event *sub;
1989
	u64 tstamp = perf_event_time(event);
1990

1991
	event->state = PERF_EVENT_STATE_INACTIVE;
1992
	event->tstamp_enabled = tstamp - event->total_time_enabled;
P
Peter Zijlstra 已提交
1993
	list_for_each_entry(sub, &event->sibling_list, group_entry) {
1994 1995
		if (sub->state >= PERF_EVENT_STATE_INACTIVE)
			sub->tstamp_enabled = tstamp - sub->total_time_enabled;
P
Peter Zijlstra 已提交
1996
	}
1997 1998
}

1999
/*
2000
 * Cross CPU call to enable a performance event
2001
 */
2002
static int __perf_event_enable(void *info)
2003
{
2004 2005 2006
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
	struct perf_event *leader = event->group_leader;
P
Peter Zijlstra 已提交
2007
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2008
	int err;
2009

2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
	/*
	 * There's a time window between 'ctx->is_active' check
	 * in perf_event_enable function and this place having:
	 *   - IRQs on
	 *   - ctx->lock unlocked
	 *
	 * where the task could be killed and 'ctx' deactivated
	 * by perf_event_exit_task.
	 */
	if (!ctx->is_active)
2020
		return -EINVAL;
2021

2022
	raw_spin_lock(&ctx->lock);
2023
	update_context_time(ctx);
2024

2025
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2026
		goto unlock;
S
Stephane Eranian 已提交
2027 2028 2029 2030

	/*
	 * set current task's cgroup time reference point
	 */
2031
	perf_cgroup_set_timestamp(current, ctx);
S
Stephane Eranian 已提交
2032

2033
	__perf_event_mark_enabled(event);
2034

S
Stephane Eranian 已提交
2035 2036 2037
	if (!event_filter_match(event)) {
		if (is_cgroup_event(event))
			perf_cgroup_defer_enabled(event);
2038
		goto unlock;
S
Stephane Eranian 已提交
2039
	}
2040

2041
	/*
2042
	 * If the event is in a group and isn't the group leader,
2043
	 * then don't put it on unless the group is on.
2044
	 */
2045
	if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2046
		goto unlock;
2047

2048
	if (!group_can_go_on(event, cpuctx, 1)) {
2049
		err = -EEXIST;
2050
	} else {
2051
		if (event == leader)
2052
			err = group_sched_in(event, cpuctx, ctx);
2053
		else
2054
			err = event_sched_in(event, cpuctx, ctx);
2055
	}
2056 2057 2058

	if (err) {
		/*
2059
		 * If this event can't go on and it's part of a
2060 2061
		 * group, then the whole group has to come off.
		 */
2062
		if (leader != event) {
2063
			group_sched_out(leader, cpuctx, ctx);
2064 2065
			perf_cpu_hrtimer_restart(cpuctx);
		}
2066
		if (leader->attr.pinned) {
2067
			update_group_times(leader);
2068
			leader->state = PERF_EVENT_STATE_ERROR;
2069
		}
2070 2071
	}

P
Peter Zijlstra 已提交
2072
unlock:
2073
	raw_spin_unlock(&ctx->lock);
2074 2075

	return 0;
2076 2077 2078
}

/*
2079
 * Enable a event.
2080
 *
2081 2082
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
2083
 * remains valid.  This condition is satisfied when called through
2084 2085
 * perf_event_for_each_child or perf_event_for_each as described
 * for perf_event_disable.
2086
 */
2087
void perf_event_enable(struct perf_event *event)
2088
{
2089
	struct perf_event_context *ctx = event->ctx;
2090 2091 2092 2093
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
2094
		 * Enable the event on the cpu that it's on
2095
		 */
2096
		cpu_function_call(event->cpu, __perf_event_enable, event);
2097 2098 2099
		return;
	}

2100
	raw_spin_lock_irq(&ctx->lock);
2101
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2102 2103 2104
		goto out;

	/*
2105 2106
	 * If the event is in error state, clear that first.
	 * That way, if we see the event in error state below, we
2107 2108 2109 2110
	 * know that it has gone back into error state, as distinct
	 * from the task having been scheduled away before the
	 * cross-call arrived.
	 */
2111 2112
	if (event->state == PERF_EVENT_STATE_ERROR)
		event->state = PERF_EVENT_STATE_OFF;
2113

P
Peter Zijlstra 已提交
2114
retry:
2115
	if (!ctx->is_active) {
2116
		__perf_event_mark_enabled(event);
2117 2118 2119
		goto out;
	}

2120
	raw_spin_unlock_irq(&ctx->lock);
2121 2122 2123

	if (!task_function_call(task, __perf_event_enable, event))
		return;
2124

2125
	raw_spin_lock_irq(&ctx->lock);
2126 2127

	/*
2128
	 * If the context is active and the event is still off,
2129 2130
	 * we need to retry the cross-call.
	 */
2131 2132 2133 2134 2135 2136
	if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
		/*
		 * task could have been flipped by a concurrent
		 * perf_event_context_sched_out()
		 */
		task = ctx->task;
2137
		goto retry;
2138
	}
2139

P
Peter Zijlstra 已提交
2140
out:
2141
	raw_spin_unlock_irq(&ctx->lock);
2142
}
2143
EXPORT_SYMBOL_GPL(perf_event_enable);
2144

2145
int perf_event_refresh(struct perf_event *event, int refresh)
2146
{
2147
	/*
2148
	 * not supported on inherited events
2149
	 */
2150
	if (event->attr.inherit || !is_sampling_event(event))
2151 2152
		return -EINVAL;

2153 2154
	atomic_add(refresh, &event->event_limit);
	perf_event_enable(event);
2155 2156

	return 0;
2157
}
2158
EXPORT_SYMBOL_GPL(perf_event_refresh);
2159

2160 2161 2162
static void ctx_sched_out(struct perf_event_context *ctx,
			  struct perf_cpu_context *cpuctx,
			  enum event_type_t event_type)
2163
{
2164
	struct perf_event *event;
2165
	int is_active = ctx->is_active;
2166

2167
	ctx->is_active &= ~event_type;
2168
	if (likely(!ctx->nr_events))
2169 2170
		return;

2171
	update_context_time(ctx);
S
Stephane Eranian 已提交
2172
	update_cgrp_time_from_cpuctx(cpuctx);
2173
	if (!ctx->nr_active)
2174
		return;
2175

P
Peter Zijlstra 已提交
2176
	perf_pmu_disable(ctx->pmu);
2177
	if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2178 2179
		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2180
	}
2181

2182
	if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2183
		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2184
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2185
	}
P
Peter Zijlstra 已提交
2186
	perf_pmu_enable(ctx->pmu);
2187 2188
}

2189
/*
2190 2191 2192 2193 2194 2195
 * Test whether two contexts are equivalent, i.e. whether they have both been
 * cloned from the same version of the same context.
 *
 * Equivalence is measured using a generation number in the context that is
 * incremented on each modification to it; see unclone_ctx(), list_add_event()
 * and list_del_event().
2196
 */
2197 2198
static int context_equiv(struct perf_event_context *ctx1,
			 struct perf_event_context *ctx2)
2199
{
2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221
	/* Pinning disables the swap optimization */
	if (ctx1->pin_count || ctx2->pin_count)
		return 0;

	/* If ctx1 is the parent of ctx2 */
	if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
		return 1;

	/* If ctx2 is the parent of ctx1 */
	if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
		return 1;

	/*
	 * If ctx1 and ctx2 have the same parent; we flatten the parent
	 * hierarchy, see perf_event_init_context().
	 */
	if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
			ctx1->parent_gen == ctx2->parent_gen)
		return 1;

	/* Unmatched */
	return 0;
2222 2223
}

2224 2225
static void __perf_event_sync_stat(struct perf_event *event,
				     struct perf_event *next_event)
2226 2227 2228
{
	u64 value;

2229
	if (!event->attr.inherit_stat)
2230 2231 2232
		return;

	/*
2233
	 * Update the event value, we cannot use perf_event_read()
2234 2235
	 * because we're in the middle of a context switch and have IRQs
	 * disabled, which upsets smp_call_function_single(), however
2236
	 * we know the event must be on the current CPU, therefore we
2237 2238
	 * don't need to use it.
	 */
2239 2240
	switch (event->state) {
	case PERF_EVENT_STATE_ACTIVE:
2241 2242
		event->pmu->read(event);
		/* fall-through */
2243

2244 2245
	case PERF_EVENT_STATE_INACTIVE:
		update_event_times(event);
2246 2247 2248 2249 2250 2251 2252
		break;

	default:
		break;
	}

	/*
2253
	 * In order to keep per-task stats reliable we need to flip the event
2254 2255
	 * values when we flip the contexts.
	 */
2256 2257 2258
	value = local64_read(&next_event->count);
	value = local64_xchg(&event->count, value);
	local64_set(&next_event->count, value);
2259

2260 2261
	swap(event->total_time_enabled, next_event->total_time_enabled);
	swap(event->total_time_running, next_event->total_time_running);
2262

2263
	/*
2264
	 * Since we swizzled the values, update the user visible data too.
2265
	 */
2266 2267
	perf_event_update_userpage(event);
	perf_event_update_userpage(next_event);
2268 2269
}

2270 2271
static void perf_event_sync_stat(struct perf_event_context *ctx,
				   struct perf_event_context *next_ctx)
2272
{
2273
	struct perf_event *event, *next_event;
2274 2275 2276 2277

	if (!ctx->nr_stat)
		return;

2278 2279
	update_context_time(ctx);

2280 2281
	event = list_first_entry(&ctx->event_list,
				   struct perf_event, event_entry);
2282

2283 2284
	next_event = list_first_entry(&next_ctx->event_list,
					struct perf_event, event_entry);
2285

2286 2287
	while (&event->event_entry != &ctx->event_list &&
	       &next_event->event_entry != &next_ctx->event_list) {
2288

2289
		__perf_event_sync_stat(event, next_event);
2290

2291 2292
		event = list_next_entry(event, event_entry);
		next_event = list_next_entry(next_event, event_entry);
2293 2294 2295
	}
}

2296 2297
static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
					 struct task_struct *next)
T
Thomas Gleixner 已提交
2298
{
P
Peter Zijlstra 已提交
2299
	struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2300
	struct perf_event_context *next_ctx;
2301
	struct perf_event_context *parent, *next_parent;
P
Peter Zijlstra 已提交
2302
	struct perf_cpu_context *cpuctx;
2303
	int do_switch = 1;
T
Thomas Gleixner 已提交
2304

P
Peter Zijlstra 已提交
2305 2306
	if (likely(!ctx))
		return;
2307

P
Peter Zijlstra 已提交
2308 2309
	cpuctx = __get_cpu_context(ctx);
	if (!cpuctx->task_ctx)
T
Thomas Gleixner 已提交
2310 2311
		return;

2312
	rcu_read_lock();
P
Peter Zijlstra 已提交
2313
	next_ctx = next->perf_event_ctxp[ctxn];
2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324
	if (!next_ctx)
		goto unlock;

	parent = rcu_dereference(ctx->parent_ctx);
	next_parent = rcu_dereference(next_ctx->parent_ctx);

	/* If neither context have a parent context; they cannot be clones. */
	if (!parent && !next_parent)
		goto unlock;

	if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2325 2326 2327 2328 2329 2330 2331 2332 2333
		/*
		 * Looks like the two contexts are clones, so we might be
		 * able to optimize the context switch.  We lock both
		 * contexts and check that they are clones under the
		 * lock (including re-checking that neither has been
		 * uncloned in the meantime).  It doesn't matter which
		 * order we take the locks because no other cpu could
		 * be trying to lock both of these tasks.
		 */
2334 2335
		raw_spin_lock(&ctx->lock);
		raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2336
		if (context_equiv(ctx, next_ctx)) {
2337 2338
			/*
			 * XXX do we need a memory barrier of sorts
2339
			 * wrt to rcu_dereference() of perf_event_ctxp
2340
			 */
P
Peter Zijlstra 已提交
2341 2342
			task->perf_event_ctxp[ctxn] = next_ctx;
			next->perf_event_ctxp[ctxn] = ctx;
2343 2344 2345
			ctx->task = next;
			next_ctx->task = task;
			do_switch = 0;
2346

2347
			perf_event_sync_stat(ctx, next_ctx);
2348
		}
2349 2350
		raw_spin_unlock(&next_ctx->lock);
		raw_spin_unlock(&ctx->lock);
2351
	}
2352
unlock:
2353
	rcu_read_unlock();
2354

2355
	if (do_switch) {
2356
		raw_spin_lock(&ctx->lock);
2357
		ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2358
		cpuctx->task_ctx = NULL;
2359
		raw_spin_unlock(&ctx->lock);
2360
	}
T
Thomas Gleixner 已提交
2361 2362
}

P
Peter Zijlstra 已提交
2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376
#define for_each_task_context_nr(ctxn)					\
	for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)

/*
 * Called from scheduler to remove the events of the current task,
 * with interrupts disabled.
 *
 * We stop each event and update the event value in event->count.
 *
 * This does not protect us against NMI, but disable()
 * sets the disabled bit in the control field of event _before_
 * accessing the event control register. If a NMI hits, then it will
 * not restart the event.
 */
2377 2378
void __perf_event_task_sched_out(struct task_struct *task,
				 struct task_struct *next)
P
Peter Zijlstra 已提交
2379 2380 2381 2382 2383
{
	int ctxn;

	for_each_task_context_nr(ctxn)
		perf_event_context_sched_out(task, ctxn, next);
S
Stephane Eranian 已提交
2384 2385 2386 2387 2388 2389 2390

	/*
	 * if cgroup events exist on this CPU, then we need
	 * to check if we have to switch out PMU state.
	 * cgroup event are system-wide mode only
	 */
	if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2391
		perf_cgroup_sched_out(task, next);
P
Peter Zijlstra 已提交
2392 2393
}

2394
static void task_ctx_sched_out(struct perf_event_context *ctx)
2395
{
P
Peter Zijlstra 已提交
2396
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2397

2398 2399
	if (!cpuctx->task_ctx)
		return;
2400 2401 2402 2403

	if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
		return;

2404
	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2405 2406 2407
	cpuctx->task_ctx = NULL;
}

2408 2409 2410 2411 2412 2413 2414
/*
 * Called with IRQs disabled
 */
static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
			      enum event_type_t event_type)
{
	ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2415 2416
}

2417
static void
2418
ctx_pinned_sched_in(struct perf_event_context *ctx,
2419
		    struct perf_cpu_context *cpuctx)
T
Thomas Gleixner 已提交
2420
{
2421
	struct perf_event *event;
T
Thomas Gleixner 已提交
2422

2423 2424
	list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
		if (event->state <= PERF_EVENT_STATE_OFF)
2425
			continue;
2426
		if (!event_filter_match(event))
2427 2428
			continue;

S
Stephane Eranian 已提交
2429 2430 2431 2432
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

2433
		if (group_can_go_on(event, cpuctx, 1))
2434
			group_sched_in(event, cpuctx, ctx);
2435 2436 2437 2438 2439

		/*
		 * If this pinned group hasn't been scheduled,
		 * put it in error state.
		 */
2440 2441 2442
		if (event->state == PERF_EVENT_STATE_INACTIVE) {
			update_group_times(event);
			event->state = PERF_EVENT_STATE_ERROR;
2443
		}
2444
	}
2445 2446 2447 2448
}

static void
ctx_flexible_sched_in(struct perf_event_context *ctx,
2449
		      struct perf_cpu_context *cpuctx)
2450 2451 2452
{
	struct perf_event *event;
	int can_add_hw = 1;
2453

2454 2455 2456
	list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
		/* Ignore events in OFF or ERROR state */
		if (event->state <= PERF_EVENT_STATE_OFF)
2457
			continue;
2458 2459
		/*
		 * Listen to the 'cpu' scheduling filter constraint
2460
		 * of events:
2461
		 */
2462
		if (!event_filter_match(event))
T
Thomas Gleixner 已提交
2463 2464
			continue;

S
Stephane Eranian 已提交
2465 2466 2467 2468
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

P
Peter Zijlstra 已提交
2469
		if (group_can_go_on(event, cpuctx, can_add_hw)) {
2470
			if (group_sched_in(event, cpuctx, ctx))
2471
				can_add_hw = 0;
P
Peter Zijlstra 已提交
2472
		}
T
Thomas Gleixner 已提交
2473
	}
2474 2475 2476 2477 2478
}

static void
ctx_sched_in(struct perf_event_context *ctx,
	     struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2479 2480
	     enum event_type_t event_type,
	     struct task_struct *task)
2481
{
S
Stephane Eranian 已提交
2482
	u64 now;
2483
	int is_active = ctx->is_active;
S
Stephane Eranian 已提交
2484

2485
	ctx->is_active |= event_type;
2486
	if (likely(!ctx->nr_events))
2487
		return;
2488

S
Stephane Eranian 已提交
2489 2490
	now = perf_clock();
	ctx->timestamp = now;
2491
	perf_cgroup_set_timestamp(task, ctx);
2492 2493 2494 2495
	/*
	 * First go through the list and put on any pinned groups
	 * in order to give them the best chance of going on.
	 */
2496
	if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2497
		ctx_pinned_sched_in(ctx, cpuctx);
2498 2499

	/* Then walk through the lower prio flexible groups */
2500
	if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2501
		ctx_flexible_sched_in(ctx, cpuctx);
2502 2503
}

2504
static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2505 2506
			     enum event_type_t event_type,
			     struct task_struct *task)
2507 2508 2509
{
	struct perf_event_context *ctx = &cpuctx->ctx;

S
Stephane Eranian 已提交
2510
	ctx_sched_in(ctx, cpuctx, event_type, task);
2511 2512
}

S
Stephane Eranian 已提交
2513 2514
static void perf_event_context_sched_in(struct perf_event_context *ctx,
					struct task_struct *task)
2515
{
P
Peter Zijlstra 已提交
2516
	struct perf_cpu_context *cpuctx;
2517

P
Peter Zijlstra 已提交
2518
	cpuctx = __get_cpu_context(ctx);
2519 2520 2521
	if (cpuctx->task_ctx == ctx)
		return;

2522
	perf_ctx_lock(cpuctx, ctx);
P
Peter Zijlstra 已提交
2523
	perf_pmu_disable(ctx->pmu);
2524 2525 2526 2527 2528 2529 2530
	/*
	 * We want to keep the following priority order:
	 * cpu pinned (that don't need to move), task pinned,
	 * cpu flexible, task flexible.
	 */
	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);

2531 2532
	if (ctx->nr_events)
		cpuctx->task_ctx = ctx;
2533

2534 2535
	perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);

2536 2537 2538
	perf_pmu_enable(ctx->pmu);
	perf_ctx_unlock(cpuctx, ctx);

2539 2540 2541 2542
	/*
	 * Since these rotations are per-cpu, we need to ensure the
	 * cpu-context we got scheduled on is actually rotating.
	 */
P
Peter Zijlstra 已提交
2543
	perf_pmu_rotate_start(ctx->pmu);
2544 2545
}

2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603
/*
 * When sampling the branck stack in system-wide, it may be necessary
 * to flush the stack on context switch. This happens when the branch
 * stack does not tag its entries with the pid of the current task.
 * Otherwise it becomes impossible to associate a branch entry with a
 * task. This ambiguity is more likely to appear when the branch stack
 * supports priv level filtering and the user sets it to monitor only
 * at the user level (which could be a useful measurement in system-wide
 * mode). In that case, the risk is high of having a branch stack with
 * branch from multiple tasks. Flushing may mean dropping the existing
 * entries or stashing them somewhere in the PMU specific code layer.
 *
 * This function provides the context switch callback to the lower code
 * layer. It is invoked ONLY when there is at least one system-wide context
 * with at least one active event using taken branch sampling.
 */
static void perf_branch_stack_sched_in(struct task_struct *prev,
				       struct task_struct *task)
{
	struct perf_cpu_context *cpuctx;
	struct pmu *pmu;
	unsigned long flags;

	/* no need to flush branch stack if not changing task */
	if (prev == task)
		return;

	local_irq_save(flags);

	rcu_read_lock();

	list_for_each_entry_rcu(pmu, &pmus, entry) {
		cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);

		/*
		 * check if the context has at least one
		 * event using PERF_SAMPLE_BRANCH_STACK
		 */
		if (cpuctx->ctx.nr_branch_stack > 0
		    && pmu->flush_branch_stack) {

			perf_ctx_lock(cpuctx, cpuctx->task_ctx);

			perf_pmu_disable(pmu);

			pmu->flush_branch_stack();

			perf_pmu_enable(pmu);

			perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
		}
	}

	rcu_read_unlock();

	local_irq_restore(flags);
}

P
Peter Zijlstra 已提交
2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614
/*
 * Called from scheduler to add the events of the current task
 * with interrupts disabled.
 *
 * We restore the event value and then enable it.
 *
 * This does not protect us against NMI, but enable()
 * sets the enabled bit in the control field of event _before_
 * accessing the event control register. If a NMI hits, then it will
 * keep the event running.
 */
2615 2616
void __perf_event_task_sched_in(struct task_struct *prev,
				struct task_struct *task)
P
Peter Zijlstra 已提交
2617 2618 2619 2620 2621 2622 2623 2624 2625
{
	struct perf_event_context *ctx;
	int ctxn;

	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (likely(!ctx))
			continue;

S
Stephane Eranian 已提交
2626
		perf_event_context_sched_in(ctx, task);
P
Peter Zijlstra 已提交
2627
	}
S
Stephane Eranian 已提交
2628 2629 2630 2631 2632 2633
	/*
	 * if cgroup events exist on this CPU, then we need
	 * to check if we have to switch in PMU state.
	 * cgroup event are system-wide mode only
	 */
	if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2634
		perf_cgroup_sched_in(prev, task);
2635 2636 2637 2638

	/* check for system-wide branch_stack events */
	if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
		perf_branch_stack_sched_in(prev, task);
2639 2640
}

2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667
static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
{
	u64 frequency = event->attr.sample_freq;
	u64 sec = NSEC_PER_SEC;
	u64 divisor, dividend;

	int count_fls, nsec_fls, frequency_fls, sec_fls;

	count_fls = fls64(count);
	nsec_fls = fls64(nsec);
	frequency_fls = fls64(frequency);
	sec_fls = 30;

	/*
	 * We got @count in @nsec, with a target of sample_freq HZ
	 * the target period becomes:
	 *
	 *             @count * 10^9
	 * period = -------------------
	 *          @nsec * sample_freq
	 *
	 */

	/*
	 * Reduce accuracy by one bit such that @a and @b converge
	 * to a similar magnitude.
	 */
2668
#define REDUCE_FLS(a, b)		\
2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707
do {					\
	if (a##_fls > b##_fls) {	\
		a >>= 1;		\
		a##_fls--;		\
	} else {			\
		b >>= 1;		\
		b##_fls--;		\
	}				\
} while (0)

	/*
	 * Reduce accuracy until either term fits in a u64, then proceed with
	 * the other, so that finally we can do a u64/u64 division.
	 */
	while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
		REDUCE_FLS(nsec, frequency);
		REDUCE_FLS(sec, count);
	}

	if (count_fls + sec_fls > 64) {
		divisor = nsec * frequency;

		while (count_fls + sec_fls > 64) {
			REDUCE_FLS(count, sec);
			divisor >>= 1;
		}

		dividend = count * sec;
	} else {
		dividend = count * sec;

		while (nsec_fls + frequency_fls > 64) {
			REDUCE_FLS(nsec, frequency);
			dividend >>= 1;
		}

		divisor = nsec * frequency;
	}

2708 2709 2710
	if (!divisor)
		return dividend;

2711 2712 2713
	return div64_u64(dividend, divisor);
}

2714 2715 2716
static DEFINE_PER_CPU(int, perf_throttled_count);
static DEFINE_PER_CPU(u64, perf_throttled_seq);

2717
static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2718
{
2719
	struct hw_perf_event *hwc = &event->hw;
2720
	s64 period, sample_period;
2721 2722
	s64 delta;

2723
	period = perf_calculate_period(event, nsec, count);
2724 2725 2726 2727 2728 2729 2730 2731 2732 2733

	delta = (s64)(period - hwc->sample_period);
	delta = (delta + 7) / 8; /* low pass filter */

	sample_period = hwc->sample_period + delta;

	if (!sample_period)
		sample_period = 1;

	hwc->sample_period = sample_period;
2734

2735
	if (local64_read(&hwc->period_left) > 8*sample_period) {
2736 2737 2738
		if (disable)
			event->pmu->stop(event, PERF_EF_UPDATE);

2739
		local64_set(&hwc->period_left, 0);
2740 2741 2742

		if (disable)
			event->pmu->start(event, PERF_EF_RELOAD);
2743
	}
2744 2745
}

2746 2747 2748 2749 2750 2751 2752
/*
 * combine freq adjustment with unthrottling to avoid two passes over the
 * events. At the same time, make sure, having freq events does not change
 * the rate of unthrottling as that would introduce bias.
 */
static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
					   int needs_unthr)
2753
{
2754 2755
	struct perf_event *event;
	struct hw_perf_event *hwc;
2756
	u64 now, period = TICK_NSEC;
2757
	s64 delta;
2758

2759 2760 2761 2762 2763 2764
	/*
	 * only need to iterate over all events iff:
	 * - context have events in frequency mode (needs freq adjust)
	 * - there are events to unthrottle on this cpu
	 */
	if (!(ctx->nr_freq || needs_unthr))
2765 2766
		return;

2767
	raw_spin_lock(&ctx->lock);
2768
	perf_pmu_disable(ctx->pmu);
2769

2770
	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2771
		if (event->state != PERF_EVENT_STATE_ACTIVE)
2772 2773
			continue;

2774
		if (!event_filter_match(event))
2775 2776
			continue;

2777 2778
		perf_pmu_disable(event->pmu);

2779
		hwc = &event->hw;
2780

2781
		if (hwc->interrupts == MAX_INTERRUPTS) {
2782
			hwc->interrupts = 0;
2783
			perf_log_throttle(event, 1);
P
Peter Zijlstra 已提交
2784
			event->pmu->start(event, 0);
2785 2786
		}

2787
		if (!event->attr.freq || !event->attr.sample_freq)
2788
			goto next;
2789

2790 2791 2792 2793 2794
		/*
		 * stop the event and update event->count
		 */
		event->pmu->stop(event, PERF_EF_UPDATE);

2795
		now = local64_read(&event->count);
2796 2797
		delta = now - hwc->freq_count_stamp;
		hwc->freq_count_stamp = now;
2798

2799 2800 2801
		/*
		 * restart the event
		 * reload only if value has changed
2802 2803 2804
		 * we have stopped the event so tell that
		 * to perf_adjust_period() to avoid stopping it
		 * twice.
2805
		 */
2806
		if (delta > 0)
2807
			perf_adjust_period(event, period, delta, false);
2808 2809

		event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2810 2811
	next:
		perf_pmu_enable(event->pmu);
2812
	}
2813

2814
	perf_pmu_enable(ctx->pmu);
2815
	raw_spin_unlock(&ctx->lock);
2816 2817
}

2818
/*
2819
 * Round-robin a context's events:
2820
 */
2821
static void rotate_ctx(struct perf_event_context *ctx)
T
Thomas Gleixner 已提交
2822
{
2823 2824 2825 2826 2827 2828
	/*
	 * Rotate the first entry last of non-pinned groups. Rotation might be
	 * disabled by the inheritance code.
	 */
	if (!ctx->rotate_disable)
		list_rotate_left(&ctx->flexible_groups);
2829 2830
}

2831
/*
2832 2833 2834
 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
 * because they're strictly cpu affine and rotate_start is called with IRQs
 * disabled, while rotate_context is called from IRQ context.
2835
 */
2836
static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2837
{
P
Peter Zijlstra 已提交
2838
	struct perf_event_context *ctx = NULL;
2839
	int rotate = 0, remove = 1;
2840

2841
	if (cpuctx->ctx.nr_events) {
2842
		remove = 0;
2843 2844 2845
		if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
			rotate = 1;
	}
2846

P
Peter Zijlstra 已提交
2847
	ctx = cpuctx->task_ctx;
2848
	if (ctx && ctx->nr_events) {
2849
		remove = 0;
2850 2851 2852
		if (ctx->nr_events != ctx->nr_active)
			rotate = 1;
	}
2853

2854
	if (!rotate)
2855 2856
		goto done;

2857
	perf_ctx_lock(cpuctx, cpuctx->task_ctx);
P
Peter Zijlstra 已提交
2858
	perf_pmu_disable(cpuctx->ctx.pmu);
2859

2860 2861 2862
	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
	if (ctx)
		ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
T
Thomas Gleixner 已提交
2863

2864 2865 2866
	rotate_ctx(&cpuctx->ctx);
	if (ctx)
		rotate_ctx(ctx);
2867

2868
	perf_event_sched_in(cpuctx, ctx, current);
2869

2870 2871
	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2872
done:
2873 2874
	if (remove)
		list_del_init(&cpuctx->rotation_list);
2875 2876

	return rotate;
2877 2878
}

2879 2880 2881
#ifdef CONFIG_NO_HZ_FULL
bool perf_event_can_stop_tick(void)
{
2882
	if (atomic_read(&nr_freq_events) ||
2883
	    __this_cpu_read(perf_throttled_count))
2884
		return false;
2885 2886
	else
		return true;
2887 2888 2889
}
#endif

2890 2891 2892 2893
void perf_event_task_tick(void)
{
	struct list_head *head = &__get_cpu_var(rotation_list);
	struct perf_cpu_context *cpuctx, *tmp;
2894 2895
	struct perf_event_context *ctx;
	int throttled;
2896

2897 2898
	WARN_ON(!irqs_disabled());

2899 2900 2901
	__this_cpu_inc(perf_throttled_seq);
	throttled = __this_cpu_xchg(perf_throttled_count, 0);

2902
	list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2903 2904 2905 2906 2907 2908
		ctx = &cpuctx->ctx;
		perf_adjust_freq_unthr_context(ctx, throttled);

		ctx = cpuctx->task_ctx;
		if (ctx)
			perf_adjust_freq_unthr_context(ctx, throttled);
2909
	}
T
Thomas Gleixner 已提交
2910 2911
}

2912 2913 2914 2915 2916 2917 2918 2919 2920 2921
static int event_enable_on_exec(struct perf_event *event,
				struct perf_event_context *ctx)
{
	if (!event->attr.enable_on_exec)
		return 0;

	event->attr.enable_on_exec = 0;
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
		return 0;

2922
	__perf_event_mark_enabled(event);
2923 2924 2925 2926

	return 1;
}

2927
/*
2928
 * Enable all of a task's events that have been marked enable-on-exec.
2929 2930
 * This expects task == current.
 */
P
Peter Zijlstra 已提交
2931
static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2932
{
2933
	struct perf_event *event;
2934 2935
	unsigned long flags;
	int enabled = 0;
2936
	int ret;
2937 2938

	local_irq_save(flags);
2939
	if (!ctx || !ctx->nr_events)
2940 2941
		goto out;

2942 2943 2944 2945 2946 2947 2948
	/*
	 * We must ctxsw out cgroup events to avoid conflict
	 * when invoking perf_task_event_sched_in() later on
	 * in this function. Otherwise we end up trying to
	 * ctxswin cgroup events which are already scheduled
	 * in.
	 */
2949
	perf_cgroup_sched_out(current, NULL);
2950

2951
	raw_spin_lock(&ctx->lock);
2952
	task_ctx_sched_out(ctx);
2953

2954
	list_for_each_entry(event, &ctx->event_list, event_entry) {
2955 2956 2957
		ret = event_enable_on_exec(event, ctx);
		if (ret)
			enabled = 1;
2958 2959 2960
	}

	/*
2961
	 * Unclone this context if we enabled any event.
2962
	 */
2963 2964
	if (enabled)
		unclone_ctx(ctx);
2965

2966
	raw_spin_unlock(&ctx->lock);
2967

2968 2969 2970
	/*
	 * Also calls ctxswin for cgroup events, if any:
	 */
S
Stephane Eranian 已提交
2971
	perf_event_context_sched_in(ctx, ctx->task);
P
Peter Zijlstra 已提交
2972
out:
2973 2974 2975
	local_irq_restore(flags);
}

2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991
void perf_event_exec(void)
{
	struct perf_event_context *ctx;
	int ctxn;

	rcu_read_lock();
	for_each_task_context_nr(ctxn) {
		ctx = current->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;

		perf_event_enable_on_exec(ctx);
	}
	rcu_read_unlock();
}

T
Thomas Gleixner 已提交
2992
/*
2993
 * Cross CPU call to read the hardware event
T
Thomas Gleixner 已提交
2994
 */
2995
static void __perf_event_read(void *info)
T
Thomas Gleixner 已提交
2996
{
2997 2998
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
2999
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
I
Ingo Molnar 已提交
3000

3001 3002 3003 3004
	/*
	 * If this is a task context, we need to check whether it is
	 * the current task context of this cpu.  If not it has been
	 * scheduled out before the smp call arrived.  In that case
3005 3006
	 * event->count would have been updated to a recent sample
	 * when the event was scheduled out.
3007 3008 3009 3010
	 */
	if (ctx->task && cpuctx->task_ctx != ctx)
		return;

3011
	raw_spin_lock(&ctx->lock);
S
Stephane Eranian 已提交
3012
	if (ctx->is_active) {
3013
		update_context_time(ctx);
S
Stephane Eranian 已提交
3014 3015
		update_cgrp_time_from_event(event);
	}
3016
	update_event_times(event);
3017 3018
	if (event->state == PERF_EVENT_STATE_ACTIVE)
		event->pmu->read(event);
3019
	raw_spin_unlock(&ctx->lock);
T
Thomas Gleixner 已提交
3020 3021
}

P
Peter Zijlstra 已提交
3022 3023
static inline u64 perf_event_count(struct perf_event *event)
{
3024
	return local64_read(&event->count) + atomic64_read(&event->child_count);
P
Peter Zijlstra 已提交
3025 3026
}

3027
static u64 perf_event_read(struct perf_event *event)
T
Thomas Gleixner 已提交
3028 3029
{
	/*
3030 3031
	 * If event is enabled and currently active on a CPU, update the
	 * value in the event structure:
T
Thomas Gleixner 已提交
3032
	 */
3033 3034 3035 3036
	if (event->state == PERF_EVENT_STATE_ACTIVE) {
		smp_call_function_single(event->oncpu,
					 __perf_event_read, event, 1);
	} else if (event->state == PERF_EVENT_STATE_INACTIVE) {
P
Peter Zijlstra 已提交
3037 3038 3039
		struct perf_event_context *ctx = event->ctx;
		unsigned long flags;

3040
		raw_spin_lock_irqsave(&ctx->lock, flags);
3041 3042 3043 3044 3045
		/*
		 * may read while context is not active
		 * (e.g., thread is blocked), in that case
		 * we cannot update context time
		 */
S
Stephane Eranian 已提交
3046
		if (ctx->is_active) {
3047
			update_context_time(ctx);
S
Stephane Eranian 已提交
3048 3049
			update_cgrp_time_from_event(event);
		}
3050
		update_event_times(event);
3051
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
T
Thomas Gleixner 已提交
3052 3053
	}

P
Peter Zijlstra 已提交
3054
	return perf_event_count(event);
T
Thomas Gleixner 已提交
3055 3056
}

3057
/*
3058
 * Initialize the perf_event context in a task_struct:
3059
 */
3060
static void __perf_event_init_context(struct perf_event_context *ctx)
3061
{
3062
	raw_spin_lock_init(&ctx->lock);
3063
	mutex_init(&ctx->mutex);
3064 3065
	INIT_LIST_HEAD(&ctx->pinned_groups);
	INIT_LIST_HEAD(&ctx->flexible_groups);
3066 3067
	INIT_LIST_HEAD(&ctx->event_list);
	atomic_set(&ctx->refcount, 1);
3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082
}

static struct perf_event_context *
alloc_perf_context(struct pmu *pmu, struct task_struct *task)
{
	struct perf_event_context *ctx;

	ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
	if (!ctx)
		return NULL;

	__perf_event_init_context(ctx);
	if (task) {
		ctx->task = task;
		get_task_struct(task);
T
Thomas Gleixner 已提交
3083
	}
3084 3085 3086
	ctx->pmu = pmu;

	return ctx;
3087 3088
}

3089 3090 3091 3092 3093
static struct task_struct *
find_lively_task_by_vpid(pid_t vpid)
{
	struct task_struct *task;
	int err;
T
Thomas Gleixner 已提交
3094 3095

	rcu_read_lock();
3096
	if (!vpid)
T
Thomas Gleixner 已提交
3097 3098
		task = current;
	else
3099
		task = find_task_by_vpid(vpid);
T
Thomas Gleixner 已提交
3100 3101 3102 3103 3104 3105 3106 3107
	if (task)
		get_task_struct(task);
	rcu_read_unlock();

	if (!task)
		return ERR_PTR(-ESRCH);

	/* Reuse ptrace permission checks for now. */
3108 3109 3110 3111
	err = -EACCES;
	if (!ptrace_may_access(task, PTRACE_MODE_READ))
		goto errout;

3112 3113 3114 3115 3116 3117 3118
	return task;
errout:
	put_task_struct(task);
	return ERR_PTR(err);

}

3119 3120 3121
/*
 * Returns a matching context with refcount and pincount.
 */
P
Peter Zijlstra 已提交
3122
static struct perf_event_context *
M
Matt Helsley 已提交
3123
find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
T
Thomas Gleixner 已提交
3124
{
3125
	struct perf_event_context *ctx;
3126
	struct perf_cpu_context *cpuctx;
3127
	unsigned long flags;
P
Peter Zijlstra 已提交
3128
	int ctxn, err;
T
Thomas Gleixner 已提交
3129

3130
	if (!task) {
3131
		/* Must be root to operate on a CPU event: */
3132
		if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
T
Thomas Gleixner 已提交
3133 3134 3135
			return ERR_PTR(-EACCES);

		/*
3136
		 * We could be clever and allow to attach a event to an
T
Thomas Gleixner 已提交
3137 3138 3139
		 * offline CPU and activate it when the CPU comes up, but
		 * that's for later.
		 */
3140
		if (!cpu_online(cpu))
T
Thomas Gleixner 已提交
3141 3142
			return ERR_PTR(-ENODEV);

P
Peter Zijlstra 已提交
3143
		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
T
Thomas Gleixner 已提交
3144
		ctx = &cpuctx->ctx;
3145
		get_ctx(ctx);
3146
		++ctx->pin_count;
T
Thomas Gleixner 已提交
3147 3148 3149 3150

		return ctx;
	}

P
Peter Zijlstra 已提交
3151 3152 3153 3154 3155
	err = -EINVAL;
	ctxn = pmu->task_ctx_nr;
	if (ctxn < 0)
		goto errout;

P
Peter Zijlstra 已提交
3156
retry:
P
Peter Zijlstra 已提交
3157
	ctx = perf_lock_task_context(task, ctxn, &flags);
3158
	if (ctx) {
3159
		unclone_ctx(ctx);
3160
		++ctx->pin_count;
3161
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
3162
	} else {
3163
		ctx = alloc_perf_context(pmu, task);
3164 3165 3166
		err = -ENOMEM;
		if (!ctx)
			goto errout;
3167

3168 3169 3170 3171 3172 3173 3174 3175 3176 3177
		err = 0;
		mutex_lock(&task->perf_event_mutex);
		/*
		 * If it has already passed perf_event_exit_task().
		 * we must see PF_EXITING, it takes this mutex too.
		 */
		if (task->flags & PF_EXITING)
			err = -ESRCH;
		else if (task->perf_event_ctxp[ctxn])
			err = -EAGAIN;
3178
		else {
3179
			get_ctx(ctx);
3180
			++ctx->pin_count;
3181
			rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3182
		}
3183 3184 3185
		mutex_unlock(&task->perf_event_mutex);

		if (unlikely(err)) {
3186
			put_ctx(ctx);
3187 3188 3189 3190

			if (err == -EAGAIN)
				goto retry;
			goto errout;
3191 3192 3193
		}
	}

T
Thomas Gleixner 已提交
3194
	return ctx;
3195

P
Peter Zijlstra 已提交
3196
errout:
3197
	return ERR_PTR(err);
T
Thomas Gleixner 已提交
3198 3199
}

L
Li Zefan 已提交
3200 3201
static void perf_event_free_filter(struct perf_event *event);

3202
static void free_event_rcu(struct rcu_head *head)
P
Peter Zijlstra 已提交
3203
{
3204
	struct perf_event *event;
P
Peter Zijlstra 已提交
3205

3206 3207 3208
	event = container_of(head, struct perf_event, rcu_head);
	if (event->ns)
		put_pid_ns(event->ns);
L
Li Zefan 已提交
3209
	perf_event_free_filter(event);
3210
	kfree(event);
P
Peter Zijlstra 已提交
3211 3212
}

3213
static void ring_buffer_put(struct ring_buffer *rb);
3214 3215
static void ring_buffer_attach(struct perf_event *event,
			       struct ring_buffer *rb);
3216

3217
static void unaccount_event_cpu(struct perf_event *event, int cpu)
3218
{
3219 3220 3221 3222 3223 3224 3225 3226 3227 3228
	if (event->parent)
		return;

	if (has_branch_stack(event)) {
		if (!(event->attach_state & PERF_ATTACH_TASK))
			atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
	}
	if (is_cgroup_event(event))
		atomic_dec(&per_cpu(perf_cgroup_events, cpu));
}
3229

3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242
static void unaccount_event(struct perf_event *event)
{
	if (event->parent)
		return;

	if (event->attach_state & PERF_ATTACH_TASK)
		static_key_slow_dec_deferred(&perf_sched_events);
	if (event->attr.mmap || event->attr.mmap_data)
		atomic_dec(&nr_mmap_events);
	if (event->attr.comm)
		atomic_dec(&nr_comm_events);
	if (event->attr.task)
		atomic_dec(&nr_task_events);
3243 3244
	if (event->attr.freq)
		atomic_dec(&nr_freq_events);
3245 3246 3247 3248 3249 3250 3251
	if (is_cgroup_event(event))
		static_key_slow_dec_deferred(&perf_sched_events);
	if (has_branch_stack(event))
		static_key_slow_dec_deferred(&perf_sched_events);

	unaccount_event_cpu(event, event->cpu);
}
3252

3253 3254
static void __free_event(struct perf_event *event)
{
3255
	if (!event->parent) {
3256 3257
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
			put_callchain_buffers();
3258
	}
3259

3260 3261 3262 3263 3264 3265
	if (event->destroy)
		event->destroy(event);

	if (event->ctx)
		put_ctx(event->ctx);

3266 3267 3268
	if (event->pmu)
		module_put(event->pmu->module);

3269 3270
	call_rcu(&event->rcu_head, free_event_rcu);
}
P
Peter Zijlstra 已提交
3271 3272

static void _free_event(struct perf_event *event)
3273
{
3274
	irq_work_sync(&event->pending);
3275

3276
	unaccount_event(event);
3277

3278
	if (event->rb) {
3279 3280 3281 3282 3283 3284 3285
		/*
		 * Can happen when we close an event with re-directed output.
		 *
		 * Since we have a 0 refcount, perf_mmap_close() will skip
		 * over us; possibly making our ring_buffer_put() the last.
		 */
		mutex_lock(&event->mmap_mutex);
3286
		ring_buffer_attach(event, NULL);
3287
		mutex_unlock(&event->mmap_mutex);
3288 3289
	}

S
Stephane Eranian 已提交
3290 3291 3292
	if (is_cgroup_event(event))
		perf_detach_cgroup(event);

3293
	__free_event(event);
3294 3295
}

P
Peter Zijlstra 已提交
3296 3297 3298 3299 3300
/*
 * Used to free events which have a known refcount of 1, such as in error paths
 * where the event isn't exposed yet and inherited events.
 */
static void free_event(struct perf_event *event)
T
Thomas Gleixner 已提交
3301
{
P
Peter Zijlstra 已提交
3302 3303 3304 3305 3306 3307
	if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
				"unexpected event refcount: %ld; ptr=%p\n",
				atomic_long_read(&event->refcount), event)) {
		/* leak to avoid use-after-free */
		return;
	}
T
Thomas Gleixner 已提交
3308

P
Peter Zijlstra 已提交
3309
	_free_event(event);
T
Thomas Gleixner 已提交
3310 3311
}

3312 3313 3314
/*
 * Called when the last reference to the file is gone.
 */
3315
static void put_event(struct perf_event *event)
3316
{
P
Peter Zijlstra 已提交
3317
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
3318
	struct task_struct *owner;
3319

3320 3321
	if (!atomic_long_dec_and_test(&event->refcount))
		return;
3322

P
Peter Zijlstra 已提交
3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355
	rcu_read_lock();
	owner = ACCESS_ONCE(event->owner);
	/*
	 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
	 * !owner it means the list deletion is complete and we can indeed
	 * free this event, otherwise we need to serialize on
	 * owner->perf_event_mutex.
	 */
	smp_read_barrier_depends();
	if (owner) {
		/*
		 * Since delayed_put_task_struct() also drops the last
		 * task reference we can safely take a new reference
		 * while holding the rcu_read_lock().
		 */
		get_task_struct(owner);
	}
	rcu_read_unlock();

	if (owner) {
		mutex_lock(&owner->perf_event_mutex);
		/*
		 * We have to re-check the event->owner field, if it is cleared
		 * we raced with perf_event_exit_task(), acquiring the mutex
		 * ensured they're done, and we can proceed with freeing the
		 * event.
		 */
		if (event->owner)
			list_del_init(&event->owner_entry);
		mutex_unlock(&owner->perf_event_mutex);
		put_task_struct(owner);
	}

P
Peter Zijlstra 已提交
3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373
	WARN_ON_ONCE(ctx->parent_ctx);
	/*
	 * There are two ways this annotation is useful:
	 *
	 *  1) there is a lock recursion from perf_event_exit_task
	 *     see the comment there.
	 *
	 *  2) there is a lock-inversion with mmap_sem through
	 *     perf_event_read_group(), which takes faults while
	 *     holding ctx->mutex, however this is called after
	 *     the last filedesc died, so there is no possibility
	 *     to trigger the AB-BA case.
	 */
	mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
	perf_remove_from_context(event, true);
	mutex_unlock(&ctx->mutex);

	_free_event(event);
3374 3375
}

P
Peter Zijlstra 已提交
3376 3377 3378 3379 3380 3381 3382
int perf_event_release_kernel(struct perf_event *event)
{
	put_event(event);
	return 0;
}
EXPORT_SYMBOL_GPL(perf_event_release_kernel);

3383 3384 3385 3386
static int perf_release(struct inode *inode, struct file *file)
{
	put_event(file->private_data);
	return 0;
3387 3388
}

3389
u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3390
{
3391
	struct perf_event *child;
3392 3393
	u64 total = 0;

3394 3395 3396
	*enabled = 0;
	*running = 0;

3397
	mutex_lock(&event->child_mutex);
3398
	total += perf_event_read(event);
3399 3400 3401 3402 3403 3404
	*enabled += event->total_time_enabled +
			atomic64_read(&event->child_total_time_enabled);
	*running += event->total_time_running +
			atomic64_read(&event->child_total_time_running);

	list_for_each_entry(child, &event->child_list, child_list) {
3405
		total += perf_event_read(child);
3406 3407 3408
		*enabled += child->total_time_enabled;
		*running += child->total_time_running;
	}
3409
	mutex_unlock(&event->child_mutex);
3410 3411 3412

	return total;
}
3413
EXPORT_SYMBOL_GPL(perf_event_read_value);
3414

3415
static int perf_event_read_group(struct perf_event *event,
3416 3417
				   u64 read_format, char __user *buf)
{
3418
	struct perf_event *leader = event->group_leader, *sub;
3419 3420
	int n = 0, size = 0, ret = -EFAULT;
	struct perf_event_context *ctx = leader->ctx;
3421
	u64 values[5];
3422
	u64 count, enabled, running;
3423

3424
	mutex_lock(&ctx->mutex);
3425
	count = perf_event_read_value(leader, &enabled, &running);
3426 3427

	values[n++] = 1 + leader->nr_siblings;
3428 3429 3430 3431
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
		values[n++] = enabled;
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
		values[n++] = running;
3432 3433 3434
	values[n++] = count;
	if (read_format & PERF_FORMAT_ID)
		values[n++] = primary_event_id(leader);
3435 3436 3437 3438

	size = n * sizeof(u64);

	if (copy_to_user(buf, values, size))
3439
		goto unlock;
3440

3441
	ret = size;
3442

3443
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3444
		n = 0;
3445

3446
		values[n++] = perf_event_read_value(sub, &enabled, &running);
3447 3448 3449 3450 3451
		if (read_format & PERF_FORMAT_ID)
			values[n++] = primary_event_id(sub);

		size = n * sizeof(u64);

3452
		if (copy_to_user(buf + ret, values, size)) {
3453 3454 3455
			ret = -EFAULT;
			goto unlock;
		}
3456 3457

		ret += size;
3458
	}
3459 3460
unlock:
	mutex_unlock(&ctx->mutex);
3461

3462
	return ret;
3463 3464
}

3465
static int perf_event_read_one(struct perf_event *event,
3466 3467
				 u64 read_format, char __user *buf)
{
3468
	u64 enabled, running;
3469 3470 3471
	u64 values[4];
	int n = 0;

3472 3473 3474 3475 3476
	values[n++] = perf_event_read_value(event, &enabled, &running);
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
		values[n++] = enabled;
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
		values[n++] = running;
3477
	if (read_format & PERF_FORMAT_ID)
3478
		values[n++] = primary_event_id(event);
3479 3480 3481 3482 3483 3484 3485

	if (copy_to_user(buf, values, n * sizeof(u64)))
		return -EFAULT;

	return n * sizeof(u64);
}

T
Thomas Gleixner 已提交
3486
/*
3487
 * Read the performance event - simple non blocking version for now
T
Thomas Gleixner 已提交
3488 3489
 */
static ssize_t
3490
perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
T
Thomas Gleixner 已提交
3491
{
3492
	u64 read_format = event->attr.read_format;
3493
	int ret;
T
Thomas Gleixner 已提交
3494

3495
	/*
3496
	 * Return end-of-file for a read on a event that is in
3497 3498 3499
	 * error state (i.e. because it was pinned but it couldn't be
	 * scheduled on to the CPU at some point).
	 */
3500
	if (event->state == PERF_EVENT_STATE_ERROR)
3501 3502
		return 0;

3503
	if (count < event->read_size)
3504 3505
		return -ENOSPC;

3506
	WARN_ON_ONCE(event->ctx->parent_ctx);
3507
	if (read_format & PERF_FORMAT_GROUP)
3508
		ret = perf_event_read_group(event, read_format, buf);
3509
	else
3510
		ret = perf_event_read_one(event, read_format, buf);
T
Thomas Gleixner 已提交
3511

3512
	return ret;
T
Thomas Gleixner 已提交
3513 3514 3515 3516 3517
}

static ssize_t
perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
3518
	struct perf_event *event = file->private_data;
T
Thomas Gleixner 已提交
3519

3520
	return perf_read_hw(event, buf, count);
T
Thomas Gleixner 已提交
3521 3522 3523 3524
}

static unsigned int perf_poll(struct file *file, poll_table *wait)
{
3525
	struct perf_event *event = file->private_data;
3526
	struct ring_buffer *rb;
3527
	unsigned int events = POLL_HUP;
P
Peter Zijlstra 已提交
3528

3529
	/*
3530 3531
	 * Pin the event->rb by taking event->mmap_mutex; otherwise
	 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3532 3533
	 */
	mutex_lock(&event->mmap_mutex);
3534 3535
	rb = event->rb;
	if (rb)
3536
		events = atomic_xchg(&rb->poll, 0);
3537 3538
	mutex_unlock(&event->mmap_mutex);

3539
	poll_wait(file, &event->waitq, wait);
T
Thomas Gleixner 已提交
3540 3541 3542 3543

	return events;
}

3544
static void perf_event_reset(struct perf_event *event)
3545
{
3546
	(void)perf_event_read(event);
3547
	local64_set(&event->count, 0);
3548
	perf_event_update_userpage(event);
P
Peter Zijlstra 已提交
3549 3550
}

3551
/*
3552 3553 3554 3555
 * Holding the top-level event's child_mutex means that any
 * descendant process that has inherited this event will block
 * in sync_child_event if it goes to exit, thus satisfying the
 * task existence requirements of perf_event_enable/disable.
3556
 */
3557 3558
static void perf_event_for_each_child(struct perf_event *event,
					void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3559
{
3560
	struct perf_event *child;
P
Peter Zijlstra 已提交
3561

3562 3563 3564 3565
	WARN_ON_ONCE(event->ctx->parent_ctx);
	mutex_lock(&event->child_mutex);
	func(event);
	list_for_each_entry(child, &event->child_list, child_list)
P
Peter Zijlstra 已提交
3566
		func(child);
3567
	mutex_unlock(&event->child_mutex);
P
Peter Zijlstra 已提交
3568 3569
}

3570 3571
static void perf_event_for_each(struct perf_event *event,
				  void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3572
{
3573 3574
	struct perf_event_context *ctx = event->ctx;
	struct perf_event *sibling;
P
Peter Zijlstra 已提交
3575

3576 3577
	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
3578
	event = event->group_leader;
3579

3580 3581
	perf_event_for_each_child(event, func);
	list_for_each_entry(sibling, &event->sibling_list, group_entry)
3582
		perf_event_for_each_child(sibling, func);
3583
	mutex_unlock(&ctx->mutex);
3584 3585
}

3586
static int perf_event_period(struct perf_event *event, u64 __user *arg)
3587
{
3588
	struct perf_event_context *ctx = event->ctx;
3589
	int ret = 0, active;
3590 3591
	u64 value;

3592
	if (!is_sampling_event(event))
3593 3594
		return -EINVAL;

3595
	if (copy_from_user(&value, arg, sizeof(value)))
3596 3597 3598 3599 3600
		return -EFAULT;

	if (!value)
		return -EINVAL;

3601
	raw_spin_lock_irq(&ctx->lock);
3602 3603
	if (event->attr.freq) {
		if (value > sysctl_perf_event_sample_rate) {
3604 3605 3606 3607
			ret = -EINVAL;
			goto unlock;
		}

3608
		event->attr.sample_freq = value;
3609
	} else {
3610 3611
		event->attr.sample_period = value;
		event->hw.sample_period = value;
3612
	}
3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626

	active = (event->state == PERF_EVENT_STATE_ACTIVE);
	if (active) {
		perf_pmu_disable(ctx->pmu);
		event->pmu->stop(event, PERF_EF_UPDATE);
	}

	local64_set(&event->hw.period_left, 0);

	if (active) {
		event->pmu->start(event, PERF_EF_RELOAD);
		perf_pmu_enable(ctx->pmu);
	}

3627
unlock:
3628
	raw_spin_unlock_irq(&ctx->lock);
3629 3630 3631 3632

	return ret;
}

3633 3634
static const struct file_operations perf_fops;

3635
static inline int perf_fget_light(int fd, struct fd *p)
3636
{
3637 3638 3639
	struct fd f = fdget(fd);
	if (!f.file)
		return -EBADF;
3640

3641 3642 3643
	if (f.file->f_op != &perf_fops) {
		fdput(f);
		return -EBADF;
3644
	}
3645 3646
	*p = f;
	return 0;
3647 3648 3649 3650
}

static int perf_event_set_output(struct perf_event *event,
				 struct perf_event *output_event);
L
Li Zefan 已提交
3651
static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3652

3653 3654
static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
3655 3656
	struct perf_event *event = file->private_data;
	void (*func)(struct perf_event *);
P
Peter Zijlstra 已提交
3657
	u32 flags = arg;
3658 3659

	switch (cmd) {
3660 3661
	case PERF_EVENT_IOC_ENABLE:
		func = perf_event_enable;
3662
		break;
3663 3664
	case PERF_EVENT_IOC_DISABLE:
		func = perf_event_disable;
3665
		break;
3666 3667
	case PERF_EVENT_IOC_RESET:
		func = perf_event_reset;
3668
		break;
P
Peter Zijlstra 已提交
3669

3670 3671
	case PERF_EVENT_IOC_REFRESH:
		return perf_event_refresh(event, arg);
3672

3673 3674
	case PERF_EVENT_IOC_PERIOD:
		return perf_event_period(event, (u64 __user *)arg);
3675

3676 3677 3678 3679 3680 3681 3682 3683 3684
	case PERF_EVENT_IOC_ID:
	{
		u64 id = primary_event_id(event);

		if (copy_to_user((void __user *)arg, &id, sizeof(id)))
			return -EFAULT;
		return 0;
	}

3685
	case PERF_EVENT_IOC_SET_OUTPUT:
3686 3687 3688
	{
		int ret;
		if (arg != -1) {
3689 3690 3691 3692 3693 3694 3695 3696 3697 3698
			struct perf_event *output_event;
			struct fd output;
			ret = perf_fget_light(arg, &output);
			if (ret)
				return ret;
			output_event = output.file->private_data;
			ret = perf_event_set_output(event, output_event);
			fdput(output);
		} else {
			ret = perf_event_set_output(event, NULL);
3699 3700 3701
		}
		return ret;
	}
3702

L
Li Zefan 已提交
3703 3704 3705
	case PERF_EVENT_IOC_SET_FILTER:
		return perf_event_set_filter(event, (void __user *)arg);

3706
	default:
P
Peter Zijlstra 已提交
3707
		return -ENOTTY;
3708
	}
P
Peter Zijlstra 已提交
3709 3710

	if (flags & PERF_IOC_FLAG_GROUP)
3711
		perf_event_for_each(event, func);
P
Peter Zijlstra 已提交
3712
	else
3713
		perf_event_for_each_child(event, func);
P
Peter Zijlstra 已提交
3714 3715

	return 0;
3716 3717
}

3718
int perf_event_task_enable(void)
3719
{
3720
	struct perf_event *event;
3721

3722 3723 3724 3725
	mutex_lock(&current->perf_event_mutex);
	list_for_each_entry(event, &current->perf_event_list, owner_entry)
		perf_event_for_each_child(event, perf_event_enable);
	mutex_unlock(&current->perf_event_mutex);
3726 3727 3728 3729

	return 0;
}

3730
int perf_event_task_disable(void)
3731
{
3732
	struct perf_event *event;
3733

3734 3735 3736 3737
	mutex_lock(&current->perf_event_mutex);
	list_for_each_entry(event, &current->perf_event_list, owner_entry)
		perf_event_for_each_child(event, perf_event_disable);
	mutex_unlock(&current->perf_event_mutex);
3738 3739 3740 3741

	return 0;
}

3742
static int perf_event_index(struct perf_event *event)
3743
{
P
Peter Zijlstra 已提交
3744 3745 3746
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;

3747
	if (event->state != PERF_EVENT_STATE_ACTIVE)
3748 3749
		return 0;

3750
	return event->pmu->event_idx(event);
3751 3752
}

3753
static void calc_timer_values(struct perf_event *event,
3754
				u64 *now,
3755 3756
				u64 *enabled,
				u64 *running)
3757
{
3758
	u64 ctx_time;
3759

3760 3761
	*now = perf_clock();
	ctx_time = event->shadow_ctx_time + *now;
3762 3763 3764 3765
	*enabled = ctx_time - event->tstamp_enabled;
	*running = ctx_time - event->tstamp_running;
}

3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785
static void perf_event_init_userpage(struct perf_event *event)
{
	struct perf_event_mmap_page *userpg;
	struct ring_buffer *rb;

	rcu_read_lock();
	rb = rcu_dereference(event->rb);
	if (!rb)
		goto unlock;

	userpg = rb->user_page;

	/* Allow new userspace to detect that bit 0 is deprecated */
	userpg->cap_bit0_is_deprecated = 1;
	userpg->size = offsetof(struct perf_event_mmap_page, __reserved);

unlock:
	rcu_read_unlock();
}

3786
void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3787 3788 3789
{
}

3790 3791 3792 3793 3794
/*
 * Callers need to ensure there can be no nesting of this function, otherwise
 * the seqlock logic goes bad. We can not serialize this because the arch
 * code calls this from NMI context.
 */
3795
void perf_event_update_userpage(struct perf_event *event)
3796
{
3797
	struct perf_event_mmap_page *userpg;
3798
	struct ring_buffer *rb;
3799
	u64 enabled, running, now;
3800 3801

	rcu_read_lock();
3802 3803 3804 3805
	rb = rcu_dereference(event->rb);
	if (!rb)
		goto unlock;

3806 3807 3808 3809 3810 3811 3812 3813 3814
	/*
	 * compute total_time_enabled, total_time_running
	 * based on snapshot values taken when the event
	 * was last scheduled in.
	 *
	 * we cannot simply called update_context_time()
	 * because of locking issue as we can be called in
	 * NMI context
	 */
3815
	calc_timer_values(event, &now, &enabled, &running);
3816

3817
	userpg = rb->user_page;
3818 3819 3820 3821 3822
	/*
	 * Disable preemption so as to not let the corresponding user-space
	 * spin too long if we get preempted.
	 */
	preempt_disable();
3823
	++userpg->lock;
3824
	barrier();
3825
	userpg->index = perf_event_index(event);
P
Peter Zijlstra 已提交
3826
	userpg->offset = perf_event_count(event);
3827
	if (userpg->index)
3828
		userpg->offset -= local64_read(&event->hw.prev_count);
3829

3830
	userpg->time_enabled = enabled +
3831
			atomic64_read(&event->child_total_time_enabled);
3832

3833
	userpg->time_running = running +
3834
			atomic64_read(&event->child_total_time_running);
3835

3836
	arch_perf_update_userpage(userpg, now);
3837

3838
	barrier();
3839
	++userpg->lock;
3840
	preempt_enable();
3841
unlock:
3842
	rcu_read_unlock();
3843 3844
}

3845 3846 3847
static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
{
	struct perf_event *event = vma->vm_file->private_data;
3848
	struct ring_buffer *rb;
3849 3850 3851 3852 3853 3854 3855 3856 3857
	int ret = VM_FAULT_SIGBUS;

	if (vmf->flags & FAULT_FLAG_MKWRITE) {
		if (vmf->pgoff == 0)
			ret = 0;
		return ret;
	}

	rcu_read_lock();
3858 3859
	rb = rcu_dereference(event->rb);
	if (!rb)
3860 3861 3862 3863 3864
		goto unlock;

	if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
		goto unlock;

3865
	vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879
	if (!vmf->page)
		goto unlock;

	get_page(vmf->page);
	vmf->page->mapping = vma->vm_file->f_mapping;
	vmf->page->index   = vmf->pgoff;

	ret = 0;
unlock:
	rcu_read_unlock();

	return ret;
}

3880 3881 3882
static void ring_buffer_attach(struct perf_event *event,
			       struct ring_buffer *rb)
{
3883
	struct ring_buffer *old_rb = NULL;
3884 3885
	unsigned long flags;

3886 3887 3888 3889 3890 3891
	if (event->rb) {
		/*
		 * Should be impossible, we set this when removing
		 * event->rb_entry and wait/clear when adding event->rb_entry.
		 */
		WARN_ON_ONCE(event->rcu_pending);
3892

3893 3894 3895
		old_rb = event->rb;
		event->rcu_batches = get_state_synchronize_rcu();
		event->rcu_pending = 1;
3896

3897 3898 3899 3900
		spin_lock_irqsave(&old_rb->event_lock, flags);
		list_del_rcu(&event->rb_entry);
		spin_unlock_irqrestore(&old_rb->event_lock, flags);
	}
3901

3902 3903 3904 3905
	if (event->rcu_pending && rb) {
		cond_synchronize_rcu(event->rcu_batches);
		event->rcu_pending = 0;
	}
3906

3907 3908 3909 3910 3911 3912 3913
	if (rb) {
		spin_lock_irqsave(&rb->event_lock, flags);
		list_add_rcu(&event->rb_entry, &rb->event_list);
		spin_unlock_irqrestore(&rb->event_lock, flags);
	}

	rcu_assign_pointer(event->rb, rb);
3914

3915 3916 3917 3918 3919 3920 3921 3922 3923
	if (old_rb) {
		ring_buffer_put(old_rb);
		/*
		 * Since we detached before setting the new rb, so that we
		 * could attach the new rb, we could have missed a wakeup.
		 * Provide it now.
		 */
		wake_up_all(&event->waitq);
	}
3924 3925 3926 3927 3928 3929 3930 3931
}

static void ring_buffer_wakeup(struct perf_event *event)
{
	struct ring_buffer *rb;

	rcu_read_lock();
	rb = rcu_dereference(event->rb);
3932 3933 3934 3935
	if (rb) {
		list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
			wake_up_all(&event->waitq);
	}
3936 3937 3938
	rcu_read_unlock();
}

3939
static void rb_free_rcu(struct rcu_head *rcu_head)
3940
{
3941
	struct ring_buffer *rb;
3942

3943 3944
	rb = container_of(rcu_head, struct ring_buffer, rcu_head);
	rb_free(rb);
3945 3946
}

3947
static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3948
{
3949
	struct ring_buffer *rb;
3950

3951
	rcu_read_lock();
3952 3953 3954 3955
	rb = rcu_dereference(event->rb);
	if (rb) {
		if (!atomic_inc_not_zero(&rb->refcount))
			rb = NULL;
3956 3957 3958
	}
	rcu_read_unlock();

3959
	return rb;
3960 3961
}

3962
static void ring_buffer_put(struct ring_buffer *rb)
3963
{
3964
	if (!atomic_dec_and_test(&rb->refcount))
3965
		return;
3966

3967
	WARN_ON_ONCE(!list_empty(&rb->event_list));
3968

3969
	call_rcu(&rb->rcu_head, rb_free_rcu);
3970 3971 3972 3973
}

static void perf_mmap_open(struct vm_area_struct *vma)
{
3974
	struct perf_event *event = vma->vm_file->private_data;
3975

3976
	atomic_inc(&event->mmap_count);
3977
	atomic_inc(&event->rb->mmap_count);
3978 3979
}

3980 3981 3982 3983 3984 3985 3986 3987
/*
 * A buffer can be mmap()ed multiple times; either directly through the same
 * event, or through other events by use of perf_event_set_output().
 *
 * In order to undo the VM accounting done by perf_mmap() we need to destroy
 * the buffer here, where we still have a VM context. This means we need
 * to detach all events redirecting to us.
 */
3988 3989
static void perf_mmap_close(struct vm_area_struct *vma)
{
3990
	struct perf_event *event = vma->vm_file->private_data;
3991

3992
	struct ring_buffer *rb = ring_buffer_get(event);
3993 3994 3995
	struct user_struct *mmap_user = rb->mmap_user;
	int mmap_locked = rb->mmap_locked;
	unsigned long size = perf_data_size(rb);
3996

3997 3998 3999
	atomic_dec(&rb->mmap_count);

	if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4000
		goto out_put;
4001

4002
	ring_buffer_attach(event, NULL);
4003 4004 4005
	mutex_unlock(&event->mmap_mutex);

	/* If there's still other mmap()s of this buffer, we're done. */
4006 4007
	if (atomic_read(&rb->mmap_count))
		goto out_put;
4008

4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024
	/*
	 * No other mmap()s, detach from all other events that might redirect
	 * into the now unreachable buffer. Somewhat complicated by the
	 * fact that rb::event_lock otherwise nests inside mmap_mutex.
	 */
again:
	rcu_read_lock();
	list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
		if (!atomic_long_inc_not_zero(&event->refcount)) {
			/*
			 * This event is en-route to free_event() which will
			 * detach it and remove it from the list.
			 */
			continue;
		}
		rcu_read_unlock();
4025

4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036
		mutex_lock(&event->mmap_mutex);
		/*
		 * Check we didn't race with perf_event_set_output() which can
		 * swizzle the rb from under us while we were waiting to
		 * acquire mmap_mutex.
		 *
		 * If we find a different rb; ignore this event, a next
		 * iteration will no longer find it on the list. We have to
		 * still restart the iteration to make sure we're not now
		 * iterating the wrong list.
		 */
4037 4038 4039
		if (event->rb == rb)
			ring_buffer_attach(event, NULL);

4040
		mutex_unlock(&event->mmap_mutex);
4041
		put_event(event);
4042

4043 4044 4045 4046 4047
		/*
		 * Restart the iteration; either we're on the wrong list or
		 * destroyed its integrity by doing a deletion.
		 */
		goto again;
4048
	}
4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063
	rcu_read_unlock();

	/*
	 * It could be there's still a few 0-ref events on the list; they'll
	 * get cleaned up by free_event() -- they'll also still have their
	 * ref on the rb and will free it whenever they are done with it.
	 *
	 * Aside from that, this buffer is 'fully' detached and unmapped,
	 * undo the VM accounting.
	 */

	atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
	vma->vm_mm->pinned_vm -= mmap_locked;
	free_uid(mmap_user);

4064
out_put:
4065
	ring_buffer_put(rb); /* could be last */
4066 4067
}

4068
static const struct vm_operations_struct perf_mmap_vmops = {
4069 4070 4071 4072
	.open		= perf_mmap_open,
	.close		= perf_mmap_close,
	.fault		= perf_mmap_fault,
	.page_mkwrite	= perf_mmap_fault,
4073 4074 4075 4076
};

static int perf_mmap(struct file *file, struct vm_area_struct *vma)
{
4077
	struct perf_event *event = file->private_data;
4078
	unsigned long user_locked, user_lock_limit;
4079
	struct user_struct *user = current_user();
4080
	unsigned long locked, lock_limit;
4081
	struct ring_buffer *rb;
4082 4083
	unsigned long vma_size;
	unsigned long nr_pages;
4084
	long user_extra, extra;
4085
	int ret = 0, flags = 0;
4086

4087 4088 4089
	/*
	 * Don't allow mmap() of inherited per-task counters. This would
	 * create a performance issue due to all children writing to the
4090
	 * same rb.
4091 4092 4093 4094
	 */
	if (event->cpu == -1 && event->attr.inherit)
		return -EINVAL;

4095
	if (!(vma->vm_flags & VM_SHARED))
4096
		return -EINVAL;
4097 4098 4099 4100

	vma_size = vma->vm_end - vma->vm_start;
	nr_pages = (vma_size / PAGE_SIZE) - 1;

4101
	/*
4102
	 * If we have rb pages ensure they're a power-of-two number, so we
4103 4104 4105
	 * can do bitmasks instead of modulo.
	 */
	if (nr_pages != 0 && !is_power_of_2(nr_pages))
4106 4107
		return -EINVAL;

4108
	if (vma_size != PAGE_SIZE * (1 + nr_pages))
4109 4110
		return -EINVAL;

4111 4112
	if (vma->vm_pgoff != 0)
		return -EINVAL;
4113

4114
	WARN_ON_ONCE(event->ctx->parent_ctx);
4115
again:
4116
	mutex_lock(&event->mmap_mutex);
4117
	if (event->rb) {
4118
		if (event->rb->nr_pages != nr_pages) {
4119
			ret = -EINVAL;
4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132
			goto unlock;
		}

		if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
			/*
			 * Raced against perf_mmap_close() through
			 * perf_event_set_output(). Try again, hope for better
			 * luck.
			 */
			mutex_unlock(&event->mmap_mutex);
			goto again;
		}

4133 4134 4135
		goto unlock;
	}

4136
	user_extra = nr_pages + 1;
4137
	user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
I
Ingo Molnar 已提交
4138 4139 4140 4141 4142 4143

	/*
	 * Increase the limit linearly with more CPUs:
	 */
	user_lock_limit *= num_online_cpus();

4144
	user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4145

4146 4147 4148
	extra = 0;
	if (user_locked > user_lock_limit)
		extra = user_locked - user_lock_limit;
4149

4150
	lock_limit = rlimit(RLIMIT_MEMLOCK);
4151
	lock_limit >>= PAGE_SHIFT;
4152
	locked = vma->vm_mm->pinned_vm + extra;
4153

4154 4155
	if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
		!capable(CAP_IPC_LOCK)) {
4156 4157 4158
		ret = -EPERM;
		goto unlock;
	}
4159

4160
	WARN_ON(event->rb);
4161

4162
	if (vma->vm_flags & VM_WRITE)
4163
		flags |= RING_BUFFER_WRITABLE;
4164

4165 4166 4167 4168
	rb = rb_alloc(nr_pages, 
		event->attr.watermark ? event->attr.wakeup_watermark : 0,
		event->cpu, flags);

4169
	if (!rb) {
4170
		ret = -ENOMEM;
4171
		goto unlock;
4172
	}
P
Peter Zijlstra 已提交
4173

4174
	atomic_set(&rb->mmap_count, 1);
P
Peter Zijlstra 已提交
4175 4176
	rb->mmap_locked = extra;
	rb->mmap_user = get_current_user();
4177

4178
	atomic_long_add(user_extra, &user->locked_vm);
P
Peter Zijlstra 已提交
4179 4180
	vma->vm_mm->pinned_vm += extra;

4181
	ring_buffer_attach(event, rb);
4182

4183
	perf_event_init_userpage(event);
4184 4185
	perf_event_update_userpage(event);

4186
unlock:
4187 4188
	if (!ret)
		atomic_inc(&event->mmap_count);
4189
	mutex_unlock(&event->mmap_mutex);
4190

4191 4192 4193 4194
	/*
	 * Since pinned accounting is per vm we cannot allow fork() to copy our
	 * vma.
	 */
P
Peter Zijlstra 已提交
4195
	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4196
	vma->vm_ops = &perf_mmap_vmops;
4197 4198

	return ret;
4199 4200
}

P
Peter Zijlstra 已提交
4201 4202
static int perf_fasync(int fd, struct file *filp, int on)
{
A
Al Viro 已提交
4203
	struct inode *inode = file_inode(filp);
4204
	struct perf_event *event = filp->private_data;
P
Peter Zijlstra 已提交
4205 4206 4207
	int retval;

	mutex_lock(&inode->i_mutex);
4208
	retval = fasync_helper(fd, filp, on, &event->fasync);
P
Peter Zijlstra 已提交
4209 4210 4211 4212 4213 4214 4215 4216
	mutex_unlock(&inode->i_mutex);

	if (retval < 0)
		return retval;

	return 0;
}

T
Thomas Gleixner 已提交
4217
static const struct file_operations perf_fops = {
4218
	.llseek			= no_llseek,
T
Thomas Gleixner 已提交
4219 4220 4221
	.release		= perf_release,
	.read			= perf_read,
	.poll			= perf_poll,
4222 4223
	.unlocked_ioctl		= perf_ioctl,
	.compat_ioctl		= perf_ioctl,
4224
	.mmap			= perf_mmap,
P
Peter Zijlstra 已提交
4225
	.fasync			= perf_fasync,
T
Thomas Gleixner 已提交
4226 4227
};

4228
/*
4229
 * Perf event wakeup
4230 4231 4232 4233 4234
 *
 * If there's data, ensure we set the poll() state and publish everything
 * to user-space before waking everybody up.
 */

4235
void perf_event_wakeup(struct perf_event *event)
4236
{
4237
	ring_buffer_wakeup(event);
4238

4239 4240 4241
	if (event->pending_kill) {
		kill_fasync(&event->fasync, SIGIO, event->pending_kill);
		event->pending_kill = 0;
4242
	}
4243 4244
}

4245
static void perf_pending_event(struct irq_work *entry)
4246
{
4247 4248
	struct perf_event *event = container_of(entry,
			struct perf_event, pending);
4249

4250 4251 4252
	if (event->pending_disable) {
		event->pending_disable = 0;
		__perf_event_disable(event);
4253 4254
	}

4255 4256 4257
	if (event->pending_wakeup) {
		event->pending_wakeup = 0;
		perf_event_wakeup(event);
4258 4259 4260
	}
}

4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281
/*
 * We assume there is only KVM supporting the callbacks.
 * Later on, we might change it to a list if there is
 * another virtualization implementation supporting the callbacks.
 */
struct perf_guest_info_callbacks *perf_guest_cbs;

int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
{
	perf_guest_cbs = cbs;
	return 0;
}
EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);

int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
{
	perf_guest_cbs = NULL;
	return 0;
}
EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);

4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312
static void
perf_output_sample_regs(struct perf_output_handle *handle,
			struct pt_regs *regs, u64 mask)
{
	int bit;

	for_each_set_bit(bit, (const unsigned long *) &mask,
			 sizeof(mask) * BITS_PER_BYTE) {
		u64 val;

		val = perf_reg_value(regs, bit);
		perf_output_put(handle, val);
	}
}

static void perf_sample_regs_user(struct perf_regs_user *regs_user,
				  struct pt_regs *regs)
{
	if (!user_mode(regs)) {
		if (current->mm)
			regs = task_pt_regs(current);
		else
			regs = NULL;
	}

	if (regs) {
		regs_user->regs = regs;
		regs_user->abi  = perf_reg_abi(current);
	}
}

4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407
/*
 * Get remaining task size from user stack pointer.
 *
 * It'd be better to take stack vma map and limit this more
 * precisly, but there's no way to get it safely under interrupt,
 * so using TASK_SIZE as limit.
 */
static u64 perf_ustack_task_size(struct pt_regs *regs)
{
	unsigned long addr = perf_user_stack_pointer(regs);

	if (!addr || addr >= TASK_SIZE)
		return 0;

	return TASK_SIZE - addr;
}

static u16
perf_sample_ustack_size(u16 stack_size, u16 header_size,
			struct pt_regs *regs)
{
	u64 task_size;

	/* No regs, no stack pointer, no dump. */
	if (!regs)
		return 0;

	/*
	 * Check if we fit in with the requested stack size into the:
	 * - TASK_SIZE
	 *   If we don't, we limit the size to the TASK_SIZE.
	 *
	 * - remaining sample size
	 *   If we don't, we customize the stack size to
	 *   fit in to the remaining sample size.
	 */

	task_size  = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
	stack_size = min(stack_size, (u16) task_size);

	/* Current header size plus static size and dynamic size. */
	header_size += 2 * sizeof(u64);

	/* Do we fit in with the current stack dump size? */
	if ((u16) (header_size + stack_size) < header_size) {
		/*
		 * If we overflow the maximum size for the sample,
		 * we customize the stack dump size to fit in.
		 */
		stack_size = USHRT_MAX - header_size - sizeof(u64);
		stack_size = round_up(stack_size, sizeof(u64));
	}

	return stack_size;
}

static void
perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
			  struct pt_regs *regs)
{
	/* Case of a kernel thread, nothing to dump */
	if (!regs) {
		u64 size = 0;
		perf_output_put(handle, size);
	} else {
		unsigned long sp;
		unsigned int rem;
		u64 dyn_size;

		/*
		 * We dump:
		 * static size
		 *   - the size requested by user or the best one we can fit
		 *     in to the sample max size
		 * data
		 *   - user stack dump data
		 * dynamic size
		 *   - the actual dumped size
		 */

		/* Static size. */
		perf_output_put(handle, dump_size);

		/* Data. */
		sp = perf_user_stack_pointer(regs);
		rem = __output_copy_user(handle, (void *) sp, dump_size);
		dyn_size = dump_size - rem;

		perf_output_skip(handle, rem);

		/* Dynamic size. */
		perf_output_put(handle, dyn_size);
	}
}

4408 4409 4410
static void __perf_event_header__init_id(struct perf_event_header *header,
					 struct perf_sample_data *data,
					 struct perf_event *event)
4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425
{
	u64 sample_type = event->attr.sample_type;

	data->type = sample_type;
	header->size += event->id_header_size;

	if (sample_type & PERF_SAMPLE_TID) {
		/* namespace issues */
		data->tid_entry.pid = perf_event_pid(event, current);
		data->tid_entry.tid = perf_event_tid(event, current);
	}

	if (sample_type & PERF_SAMPLE_TIME)
		data->time = perf_clock();

4426
	if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437
		data->id = primary_event_id(event);

	if (sample_type & PERF_SAMPLE_STREAM_ID)
		data->stream_id = event->id;

	if (sample_type & PERF_SAMPLE_CPU) {
		data->cpu_entry.cpu	 = raw_smp_processor_id();
		data->cpu_entry.reserved = 0;
	}
}

4438 4439 4440
void perf_event_header__init_id(struct perf_event_header *header,
				struct perf_sample_data *data,
				struct perf_event *event)
4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464
{
	if (event->attr.sample_id_all)
		__perf_event_header__init_id(header, data, event);
}

static void __perf_event__output_id_sample(struct perf_output_handle *handle,
					   struct perf_sample_data *data)
{
	u64 sample_type = data->type;

	if (sample_type & PERF_SAMPLE_TID)
		perf_output_put(handle, data->tid_entry);

	if (sample_type & PERF_SAMPLE_TIME)
		perf_output_put(handle, data->time);

	if (sample_type & PERF_SAMPLE_ID)
		perf_output_put(handle, data->id);

	if (sample_type & PERF_SAMPLE_STREAM_ID)
		perf_output_put(handle, data->stream_id);

	if (sample_type & PERF_SAMPLE_CPU)
		perf_output_put(handle, data->cpu_entry);
4465 4466 4467

	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);
4468 4469
}

4470 4471 4472
void perf_event__output_id_sample(struct perf_event *event,
				  struct perf_output_handle *handle,
				  struct perf_sample_data *sample)
4473 4474 4475 4476 4477
{
	if (event->attr.sample_id_all)
		__perf_event__output_id_sample(handle, sample);
}

4478
static void perf_output_read_one(struct perf_output_handle *handle,
4479 4480
				 struct perf_event *event,
				 u64 enabled, u64 running)
4481
{
4482
	u64 read_format = event->attr.read_format;
4483 4484 4485
	u64 values[4];
	int n = 0;

P
Peter Zijlstra 已提交
4486
	values[n++] = perf_event_count(event);
4487
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4488
		values[n++] = enabled +
4489
			atomic64_read(&event->child_total_time_enabled);
4490 4491
	}
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4492
		values[n++] = running +
4493
			atomic64_read(&event->child_total_time_running);
4494 4495
	}
	if (read_format & PERF_FORMAT_ID)
4496
		values[n++] = primary_event_id(event);
4497

4498
	__output_copy(handle, values, n * sizeof(u64));
4499 4500 4501
}

/*
4502
 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4503 4504
 */
static void perf_output_read_group(struct perf_output_handle *handle,
4505 4506
			    struct perf_event *event,
			    u64 enabled, u64 running)
4507
{
4508 4509
	struct perf_event *leader = event->group_leader, *sub;
	u64 read_format = event->attr.read_format;
4510 4511 4512 4513 4514 4515
	u64 values[5];
	int n = 0;

	values[n++] = 1 + leader->nr_siblings;

	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4516
		values[n++] = enabled;
4517 4518

	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4519
		values[n++] = running;
4520

4521
	if (leader != event)
4522 4523
		leader->pmu->read(leader);

P
Peter Zijlstra 已提交
4524
	values[n++] = perf_event_count(leader);
4525
	if (read_format & PERF_FORMAT_ID)
4526
		values[n++] = primary_event_id(leader);
4527

4528
	__output_copy(handle, values, n * sizeof(u64));
4529

4530
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4531 4532
		n = 0;

4533 4534
		if ((sub != event) &&
		    (sub->state == PERF_EVENT_STATE_ACTIVE))
4535 4536
			sub->pmu->read(sub);

P
Peter Zijlstra 已提交
4537
		values[n++] = perf_event_count(sub);
4538
		if (read_format & PERF_FORMAT_ID)
4539
			values[n++] = primary_event_id(sub);
4540

4541
		__output_copy(handle, values, n * sizeof(u64));
4542 4543 4544
	}
}

4545 4546 4547
#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
				 PERF_FORMAT_TOTAL_TIME_RUNNING)

4548
static void perf_output_read(struct perf_output_handle *handle,
4549
			     struct perf_event *event)
4550
{
4551
	u64 enabled = 0, running = 0, now;
4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562
	u64 read_format = event->attr.read_format;

	/*
	 * compute total_time_enabled, total_time_running
	 * based on snapshot values taken when the event
	 * was last scheduled in.
	 *
	 * we cannot simply called update_context_time()
	 * because of locking issue as we are called in
	 * NMI context
	 */
4563
	if (read_format & PERF_FORMAT_TOTAL_TIMES)
4564
		calc_timer_values(event, &now, &enabled, &running);
4565

4566
	if (event->attr.read_format & PERF_FORMAT_GROUP)
4567
		perf_output_read_group(handle, event, enabled, running);
4568
	else
4569
		perf_output_read_one(handle, event, enabled, running);
4570 4571
}

4572 4573 4574
void perf_output_sample(struct perf_output_handle *handle,
			struct perf_event_header *header,
			struct perf_sample_data *data,
4575
			struct perf_event *event)
4576 4577 4578 4579 4580
{
	u64 sample_type = data->type;

	perf_output_put(handle, *header);

4581 4582 4583
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);

4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608
	if (sample_type & PERF_SAMPLE_IP)
		perf_output_put(handle, data->ip);

	if (sample_type & PERF_SAMPLE_TID)
		perf_output_put(handle, data->tid_entry);

	if (sample_type & PERF_SAMPLE_TIME)
		perf_output_put(handle, data->time);

	if (sample_type & PERF_SAMPLE_ADDR)
		perf_output_put(handle, data->addr);

	if (sample_type & PERF_SAMPLE_ID)
		perf_output_put(handle, data->id);

	if (sample_type & PERF_SAMPLE_STREAM_ID)
		perf_output_put(handle, data->stream_id);

	if (sample_type & PERF_SAMPLE_CPU)
		perf_output_put(handle, data->cpu_entry);

	if (sample_type & PERF_SAMPLE_PERIOD)
		perf_output_put(handle, data->period);

	if (sample_type & PERF_SAMPLE_READ)
4609
		perf_output_read(handle, event);
4610 4611 4612 4613 4614 4615 4616 4617 4618 4619

	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
		if (data->callchain) {
			int size = 1;

			if (data->callchain)
				size += data->callchain->nr;

			size *= sizeof(u64);

4620
			__output_copy(handle, data->callchain, size);
4621 4622 4623 4624 4625 4626 4627 4628 4629
		} else {
			u64 nr = 0;
			perf_output_put(handle, nr);
		}
	}

	if (sample_type & PERF_SAMPLE_RAW) {
		if (data->raw) {
			perf_output_put(handle, data->raw->size);
4630 4631
			__output_copy(handle, data->raw->data,
					   data->raw->size);
4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642
		} else {
			struct {
				u32	size;
				u32	data;
			} raw = {
				.size = sizeof(u32),
				.data = 0,
			};
			perf_output_put(handle, raw);
		}
	}
4643

4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660
	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
		if (data->br_stack) {
			size_t size;

			size = data->br_stack->nr
			     * sizeof(struct perf_branch_entry);

			perf_output_put(handle, data->br_stack->nr);
			perf_output_copy(handle, data->br_stack->entries, size);
		} else {
			/*
			 * we always store at least the value of nr
			 */
			u64 nr = 0;
			perf_output_put(handle, nr);
		}
	}
4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677

	if (sample_type & PERF_SAMPLE_REGS_USER) {
		u64 abi = data->regs_user.abi;

		/*
		 * If there are no regs to dump, notice it through
		 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
		 */
		perf_output_put(handle, abi);

		if (abi) {
			u64 mask = event->attr.sample_regs_user;
			perf_output_sample_regs(handle,
						data->regs_user.regs,
						mask);
		}
	}
4678

4679
	if (sample_type & PERF_SAMPLE_STACK_USER) {
4680 4681 4682
		perf_output_sample_ustack(handle,
					  data->stack_user_size,
					  data->regs_user.regs);
4683
	}
A
Andi Kleen 已提交
4684 4685 4686

	if (sample_type & PERF_SAMPLE_WEIGHT)
		perf_output_put(handle, data->weight);
4687 4688 4689

	if (sample_type & PERF_SAMPLE_DATA_SRC)
		perf_output_put(handle, data->data_src.val);
4690

A
Andi Kleen 已提交
4691 4692 4693
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		perf_output_put(handle, data->txn);

4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706
	if (!event->attr.watermark) {
		int wakeup_events = event->attr.wakeup_events;

		if (wakeup_events) {
			struct ring_buffer *rb = handle->rb;
			int events = local_inc_return(&rb->events);

			if (events >= wakeup_events) {
				local_sub(wakeup_events, &rb->events);
				local_inc(&rb->wakeup);
			}
		}
	}
4707 4708 4709 4710
}

void perf_prepare_sample(struct perf_event_header *header,
			 struct perf_sample_data *data,
4711
			 struct perf_event *event,
4712
			 struct pt_regs *regs)
4713
{
4714
	u64 sample_type = event->attr.sample_type;
4715

4716
	header->type = PERF_RECORD_SAMPLE;
4717
	header->size = sizeof(*header) + event->header_size;
4718 4719 4720

	header->misc = 0;
	header->misc |= perf_misc_flags(regs);
4721

4722
	__perf_event_header__init_id(header, data, event);
4723

4724
	if (sample_type & PERF_SAMPLE_IP)
4725 4726
		data->ip = perf_instruction_pointer(regs);

4727
	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4728
		int size = 1;
4729

4730
		data->callchain = perf_callchain(event, regs);
4731 4732 4733 4734 4735

		if (data->callchain)
			size += data->callchain->nr;

		header->size += size * sizeof(u64);
4736 4737
	}

4738
	if (sample_type & PERF_SAMPLE_RAW) {
4739 4740 4741 4742 4743 4744 4745 4746
		int size = sizeof(u32);

		if (data->raw)
			size += data->raw->size;
		else
			size += sizeof(u32);

		WARN_ON_ONCE(size & (sizeof(u64)-1));
4747
		header->size += size;
4748
	}
4749 4750 4751 4752 4753 4754 4755 4756 4757

	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
		int size = sizeof(u64); /* nr */
		if (data->br_stack) {
			size += data->br_stack->nr
			      * sizeof(struct perf_branch_entry);
		}
		header->size += size;
	}
4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771

	if (sample_type & PERF_SAMPLE_REGS_USER) {
		/* regs dump ABI info */
		int size = sizeof(u64);

		perf_sample_regs_user(&data->regs_user, regs);

		if (data->regs_user.regs) {
			u64 mask = event->attr.sample_regs_user;
			size += hweight64(mask) * sizeof(u64);
		}

		header->size += size;
	}
4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800

	if (sample_type & PERF_SAMPLE_STACK_USER) {
		/*
		 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
		 * processed as the last one or have additional check added
		 * in case new sample type is added, because we could eat
		 * up the rest of the sample size.
		 */
		struct perf_regs_user *uregs = &data->regs_user;
		u16 stack_size = event->attr.sample_stack_user;
		u16 size = sizeof(u64);

		if (!uregs->abi)
			perf_sample_regs_user(uregs, regs);

		stack_size = perf_sample_ustack_size(stack_size, header->size,
						     uregs->regs);

		/*
		 * If there is something to dump, add space for the dump
		 * itself and for the field that tells the dynamic size,
		 * which is how many have been actually dumped.
		 */
		if (stack_size)
			size += sizeof(u64) + stack_size;

		data->stack_user_size = stack_size;
		header->size += size;
	}
4801
}
4802

4803
static void perf_event_output(struct perf_event *event,
4804 4805 4806 4807 4808
				struct perf_sample_data *data,
				struct pt_regs *regs)
{
	struct perf_output_handle handle;
	struct perf_event_header header;
4809

4810 4811 4812
	/* protect the callchain buffers */
	rcu_read_lock();

4813
	perf_prepare_sample(&header, data, event, regs);
P
Peter Zijlstra 已提交
4814

4815
	if (perf_output_begin(&handle, event, header.size))
4816
		goto exit;
4817

4818
	perf_output_sample(&handle, &header, data, event);
4819

4820
	perf_output_end(&handle);
4821 4822 4823

exit:
	rcu_read_unlock();
4824 4825
}

4826
/*
4827
 * read event_id
4828 4829 4830 4831 4832 4833 4834 4835 4836 4837
 */

struct perf_read_event {
	struct perf_event_header	header;

	u32				pid;
	u32				tid;
};

static void
4838
perf_event_read_event(struct perf_event *event,
4839 4840 4841
			struct task_struct *task)
{
	struct perf_output_handle handle;
4842
	struct perf_sample_data sample;
4843
	struct perf_read_event read_event = {
4844
		.header = {
4845
			.type = PERF_RECORD_READ,
4846
			.misc = 0,
4847
			.size = sizeof(read_event) + event->read_size,
4848
		},
4849 4850
		.pid = perf_event_pid(event, task),
		.tid = perf_event_tid(event, task),
4851
	};
4852
	int ret;
4853

4854
	perf_event_header__init_id(&read_event.header, &sample, event);
4855
	ret = perf_output_begin(&handle, event, read_event.header.size);
4856 4857 4858
	if (ret)
		return;

4859
	perf_output_put(&handle, read_event);
4860
	perf_output_read(&handle, event);
4861
	perf_event__output_id_sample(event, &handle, &sample);
4862

4863 4864 4865
	perf_output_end(&handle);
}

4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879
typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);

static void
perf_event_aux_ctx(struct perf_event_context *ctx,
		   perf_event_aux_output_cb output,
		   void *data)
{
	struct perf_event *event;

	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
		if (event->state < PERF_EVENT_STATE_INACTIVE)
			continue;
		if (!event_filter_match(event))
			continue;
4880
		output(event, data);
4881 4882 4883 4884
	}
}

static void
4885
perf_event_aux(perf_event_aux_output_cb output, void *data,
4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897
	       struct perf_event_context *task_ctx)
{
	struct perf_cpu_context *cpuctx;
	struct perf_event_context *ctx;
	struct pmu *pmu;
	int ctxn;

	rcu_read_lock();
	list_for_each_entry_rcu(pmu, &pmus, entry) {
		cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
		if (cpuctx->unique_pmu != pmu)
			goto next;
4898
		perf_event_aux_ctx(&cpuctx->ctx, output, data);
4899 4900 4901 4902 4903 4904 4905
		if (task_ctx)
			goto next;
		ctxn = pmu->task_ctx_nr;
		if (ctxn < 0)
			goto next;
		ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
		if (ctx)
4906
			perf_event_aux_ctx(ctx, output, data);
4907 4908 4909 4910 4911 4912
next:
		put_cpu_ptr(pmu->pmu_cpu_context);
	}

	if (task_ctx) {
		preempt_disable();
4913
		perf_event_aux_ctx(task_ctx, output, data);
4914 4915 4916 4917 4918
		preempt_enable();
	}
	rcu_read_unlock();
}

P
Peter Zijlstra 已提交
4919
/*
P
Peter Zijlstra 已提交
4920 4921
 * task tracking -- fork/exit
 *
4922
 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
P
Peter Zijlstra 已提交
4923 4924
 */

P
Peter Zijlstra 已提交
4925
struct perf_task_event {
4926
	struct task_struct		*task;
4927
	struct perf_event_context	*task_ctx;
P
Peter Zijlstra 已提交
4928 4929 4930 4931 4932 4933

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				ppid;
P
Peter Zijlstra 已提交
4934 4935
		u32				tid;
		u32				ptid;
4936
		u64				time;
4937
	} event_id;
P
Peter Zijlstra 已提交
4938 4939
};

4940 4941
static int perf_event_task_match(struct perf_event *event)
{
4942 4943 4944
	return event->attr.comm  || event->attr.mmap ||
	       event->attr.mmap2 || event->attr.mmap_data ||
	       event->attr.task;
4945 4946
}

4947
static void perf_event_task_output(struct perf_event *event,
4948
				   void *data)
P
Peter Zijlstra 已提交
4949
{
4950
	struct perf_task_event *task_event = data;
P
Peter Zijlstra 已提交
4951
	struct perf_output_handle handle;
4952
	struct perf_sample_data	sample;
P
Peter Zijlstra 已提交
4953
	struct task_struct *task = task_event->task;
4954
	int ret, size = task_event->event_id.header.size;
4955

4956 4957 4958
	if (!perf_event_task_match(event))
		return;

4959
	perf_event_header__init_id(&task_event->event_id.header, &sample, event);
P
Peter Zijlstra 已提交
4960

4961
	ret = perf_output_begin(&handle, event,
4962
				task_event->event_id.header.size);
4963
	if (ret)
4964
		goto out;
P
Peter Zijlstra 已提交
4965

4966 4967
	task_event->event_id.pid = perf_event_pid(event, task);
	task_event->event_id.ppid = perf_event_pid(event, current);
P
Peter Zijlstra 已提交
4968

4969 4970
	task_event->event_id.tid = perf_event_tid(event, task);
	task_event->event_id.ptid = perf_event_tid(event, current);
P
Peter Zijlstra 已提交
4971

4972
	perf_output_put(&handle, task_event->event_id);
4973

4974 4975
	perf_event__output_id_sample(event, &handle, &sample);

P
Peter Zijlstra 已提交
4976
	perf_output_end(&handle);
4977 4978
out:
	task_event->event_id.header.size = size;
P
Peter Zijlstra 已提交
4979 4980
}

4981 4982
static void perf_event_task(struct task_struct *task,
			      struct perf_event_context *task_ctx,
4983
			      int new)
P
Peter Zijlstra 已提交
4984
{
P
Peter Zijlstra 已提交
4985
	struct perf_task_event task_event;
P
Peter Zijlstra 已提交
4986

4987 4988 4989
	if (!atomic_read(&nr_comm_events) &&
	    !atomic_read(&nr_mmap_events) &&
	    !atomic_read(&nr_task_events))
P
Peter Zijlstra 已提交
4990 4991
		return;

P
Peter Zijlstra 已提交
4992
	task_event = (struct perf_task_event){
4993 4994
		.task	  = task,
		.task_ctx = task_ctx,
4995
		.event_id    = {
P
Peter Zijlstra 已提交
4996
			.header = {
4997
				.type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4998
				.misc = 0,
4999
				.size = sizeof(task_event.event_id),
P
Peter Zijlstra 已提交
5000
			},
5001 5002
			/* .pid  */
			/* .ppid */
P
Peter Zijlstra 已提交
5003 5004
			/* .tid  */
			/* .ptid */
P
Peter Zijlstra 已提交
5005
			.time = perf_clock(),
P
Peter Zijlstra 已提交
5006 5007 5008
		},
	};

5009
	perf_event_aux(perf_event_task_output,
5010 5011
		       &task_event,
		       task_ctx);
P
Peter Zijlstra 已提交
5012 5013
}

5014
void perf_event_fork(struct task_struct *task)
P
Peter Zijlstra 已提交
5015
{
5016
	perf_event_task(task, NULL, 1);
P
Peter Zijlstra 已提交
5017 5018
}

5019 5020 5021 5022 5023
/*
 * comm tracking
 */

struct perf_comm_event {
5024 5025
	struct task_struct	*task;
	char			*comm;
5026 5027 5028 5029 5030 5031 5032
	int			comm_size;

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
5033
	} event_id;
5034 5035
};

5036 5037 5038 5039 5040
static int perf_event_comm_match(struct perf_event *event)
{
	return event->attr.comm;
}

5041
static void perf_event_comm_output(struct perf_event *event,
5042
				   void *data)
5043
{
5044
	struct perf_comm_event *comm_event = data;
5045
	struct perf_output_handle handle;
5046
	struct perf_sample_data sample;
5047
	int size = comm_event->event_id.header.size;
5048 5049
	int ret;

5050 5051 5052
	if (!perf_event_comm_match(event))
		return;

5053 5054
	perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5055
				comm_event->event_id.header.size);
5056 5057

	if (ret)
5058
		goto out;
5059

5060 5061
	comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
	comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5062

5063
	perf_output_put(&handle, comm_event->event_id);
5064
	__output_copy(&handle, comm_event->comm,
5065
				   comm_event->comm_size);
5066 5067 5068

	perf_event__output_id_sample(event, &handle, &sample);

5069
	perf_output_end(&handle);
5070 5071
out:
	comm_event->event_id.header.size = size;
5072 5073
}

5074
static void perf_event_comm_event(struct perf_comm_event *comm_event)
5075
{
5076
	char comm[TASK_COMM_LEN];
5077 5078
	unsigned int size;

5079
	memset(comm, 0, sizeof(comm));
5080
	strlcpy(comm, comm_event->task->comm, sizeof(comm));
5081
	size = ALIGN(strlen(comm)+1, sizeof(u64));
5082 5083 5084 5085

	comm_event->comm = comm;
	comm_event->comm_size = size;

5086
	comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
P
Peter Zijlstra 已提交
5087

5088
	perf_event_aux(perf_event_comm_output,
5089 5090
		       comm_event,
		       NULL);
5091 5092
}

5093
void perf_event_comm(struct task_struct *task, bool exec)
5094
{
5095 5096
	struct perf_comm_event comm_event;

5097
	if (!atomic_read(&nr_comm_events))
5098
		return;
5099

5100
	comm_event = (struct perf_comm_event){
5101
		.task	= task,
5102 5103
		/* .comm      */
		/* .comm_size */
5104
		.event_id  = {
5105
			.header = {
5106
				.type = PERF_RECORD_COMM,
5107
				.misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5108 5109 5110 5111
				/* .size */
			},
			/* .pid */
			/* .tid */
5112 5113 5114
		},
	};

5115
	perf_event_comm_event(&comm_event);
5116 5117
}

5118 5119 5120 5121 5122
/*
 * mmap tracking
 */

struct perf_mmap_event {
5123 5124 5125 5126
	struct vm_area_struct	*vma;

	const char		*file_name;
	int			file_size;
5127 5128 5129
	int			maj, min;
	u64			ino;
	u64			ino_generation;
5130 5131 5132 5133 5134 5135 5136 5137 5138

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
		u64				start;
		u64				len;
		u64				pgoff;
5139
	} event_id;
5140 5141
};

5142 5143 5144 5145 5146 5147 5148 5149
static int perf_event_mmap_match(struct perf_event *event,
				 void *data)
{
	struct perf_mmap_event *mmap_event = data;
	struct vm_area_struct *vma = mmap_event->vma;
	int executable = vma->vm_flags & VM_EXEC;

	return (!executable && event->attr.mmap_data) ||
5150
	       (executable && (event->attr.mmap || event->attr.mmap2));
5151 5152
}

5153
static void perf_event_mmap_output(struct perf_event *event,
5154
				   void *data)
5155
{
5156
	struct perf_mmap_event *mmap_event = data;
5157
	struct perf_output_handle handle;
5158
	struct perf_sample_data sample;
5159
	int size = mmap_event->event_id.header.size;
5160
	int ret;
5161

5162 5163 5164
	if (!perf_event_mmap_match(event, data))
		return;

5165 5166 5167 5168 5169
	if (event->attr.mmap2) {
		mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
		mmap_event->event_id.header.size += sizeof(mmap_event->maj);
		mmap_event->event_id.header.size += sizeof(mmap_event->min);
		mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5170
		mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5171 5172
	}

5173 5174
	perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5175
				mmap_event->event_id.header.size);
5176
	if (ret)
5177
		goto out;
5178

5179 5180
	mmap_event->event_id.pid = perf_event_pid(event, current);
	mmap_event->event_id.tid = perf_event_tid(event, current);
5181

5182
	perf_output_put(&handle, mmap_event->event_id);
5183 5184 5185 5186 5187 5188 5189 5190

	if (event->attr.mmap2) {
		perf_output_put(&handle, mmap_event->maj);
		perf_output_put(&handle, mmap_event->min);
		perf_output_put(&handle, mmap_event->ino);
		perf_output_put(&handle, mmap_event->ino_generation);
	}

5191
	__output_copy(&handle, mmap_event->file_name,
5192
				   mmap_event->file_size);
5193 5194 5195

	perf_event__output_id_sample(event, &handle, &sample);

5196
	perf_output_end(&handle);
5197 5198
out:
	mmap_event->event_id.header.size = size;
5199 5200
}

5201
static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5202
{
5203 5204
	struct vm_area_struct *vma = mmap_event->vma;
	struct file *file = vma->vm_file;
5205 5206
	int maj = 0, min = 0;
	u64 ino = 0, gen = 0;
5207 5208 5209
	unsigned int size;
	char tmp[16];
	char *buf = NULL;
5210
	char *name;
5211

5212
	if (file) {
5213 5214
		struct inode *inode;
		dev_t dev;
5215

5216
		buf = kmalloc(PATH_MAX, GFP_KERNEL);
5217
		if (!buf) {
5218 5219
			name = "//enomem";
			goto cpy_name;
5220
		}
5221
		/*
5222
		 * d_path() works from the end of the rb backwards, so we
5223 5224 5225
		 * need to add enough zero bytes after the string to handle
		 * the 64bit alignment we do later.
		 */
5226
		name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5227
		if (IS_ERR(name)) {
5228 5229
			name = "//toolong";
			goto cpy_name;
5230
		}
5231 5232 5233 5234 5235 5236
		inode = file_inode(vma->vm_file);
		dev = inode->i_sb->s_dev;
		ino = inode->i_ino;
		gen = inode->i_generation;
		maj = MAJOR(dev);
		min = MINOR(dev);
5237
		goto got_name;
5238
	} else {
5239
		name = (char *)arch_vma_name(vma);
5240 5241
		if (name)
			goto cpy_name;
5242

5243
		if (vma->vm_start <= vma->vm_mm->start_brk &&
5244
				vma->vm_end >= vma->vm_mm->brk) {
5245 5246
			name = "[heap]";
			goto cpy_name;
5247 5248
		}
		if (vma->vm_start <= vma->vm_mm->start_stack &&
5249
				vma->vm_end >= vma->vm_mm->start_stack) {
5250 5251
			name = "[stack]";
			goto cpy_name;
5252 5253
		}

5254 5255
		name = "//anon";
		goto cpy_name;
5256 5257
	}

5258 5259 5260
cpy_name:
	strlcpy(tmp, name, sizeof(tmp));
	name = tmp;
5261
got_name:
5262 5263 5264 5265 5266 5267 5268 5269
	/*
	 * Since our buffer works in 8 byte units we need to align our string
	 * size to a multiple of 8. However, we must guarantee the tail end is
	 * zero'd out to avoid leaking random bits to userspace.
	 */
	size = strlen(name)+1;
	while (!IS_ALIGNED(size, sizeof(u64)))
		name[size++] = '\0';
5270 5271 5272

	mmap_event->file_name = name;
	mmap_event->file_size = size;
5273 5274 5275 5276
	mmap_event->maj = maj;
	mmap_event->min = min;
	mmap_event->ino = ino;
	mmap_event->ino_generation = gen;
5277

5278 5279 5280
	if (!(vma->vm_flags & VM_EXEC))
		mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;

5281
	mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5282

5283
	perf_event_aux(perf_event_mmap_output,
5284 5285
		       mmap_event,
		       NULL);
5286

5287 5288 5289
	kfree(buf);
}

5290
void perf_event_mmap(struct vm_area_struct *vma)
5291
{
5292 5293
	struct perf_mmap_event mmap_event;

5294
	if (!atomic_read(&nr_mmap_events))
5295 5296 5297
		return;

	mmap_event = (struct perf_mmap_event){
5298
		.vma	= vma,
5299 5300
		/* .file_name */
		/* .file_size */
5301
		.event_id  = {
5302
			.header = {
5303
				.type = PERF_RECORD_MMAP,
5304
				.misc = PERF_RECORD_MISC_USER,
5305 5306 5307 5308
				/* .size */
			},
			/* .pid */
			/* .tid */
5309 5310
			.start  = vma->vm_start,
			.len    = vma->vm_end - vma->vm_start,
5311
			.pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
5312
		},
5313 5314 5315 5316
		/* .maj (attr_mmap2 only) */
		/* .min (attr_mmap2 only) */
		/* .ino (attr_mmap2 only) */
		/* .ino_generation (attr_mmap2 only) */
5317 5318
	};

5319
	perf_event_mmap_event(&mmap_event);
5320 5321
}

5322 5323 5324 5325
/*
 * IRQ throttle logging
 */

5326
static void perf_log_throttle(struct perf_event *event, int enable)
5327 5328
{
	struct perf_output_handle handle;
5329
	struct perf_sample_data sample;
5330 5331 5332 5333 5334
	int ret;

	struct {
		struct perf_event_header	header;
		u64				time;
5335
		u64				id;
5336
		u64				stream_id;
5337 5338
	} throttle_event = {
		.header = {
5339
			.type = PERF_RECORD_THROTTLE,
5340 5341 5342
			.misc = 0,
			.size = sizeof(throttle_event),
		},
P
Peter Zijlstra 已提交
5343
		.time		= perf_clock(),
5344 5345
		.id		= primary_event_id(event),
		.stream_id	= event->id,
5346 5347
	};

5348
	if (enable)
5349
		throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5350

5351 5352 5353
	perf_event_header__init_id(&throttle_event.header, &sample, event);

	ret = perf_output_begin(&handle, event,
5354
				throttle_event.header.size);
5355 5356 5357 5358
	if (ret)
		return;

	perf_output_put(&handle, throttle_event);
5359
	perf_event__output_id_sample(event, &handle, &sample);
5360 5361 5362
	perf_output_end(&handle);
}

5363
/*
5364
 * Generic event overflow handling, sampling.
5365 5366
 */

5367
static int __perf_event_overflow(struct perf_event *event,
5368 5369
				   int throttle, struct perf_sample_data *data,
				   struct pt_regs *regs)
5370
{
5371 5372
	int events = atomic_read(&event->event_limit);
	struct hw_perf_event *hwc = &event->hw;
5373
	u64 seq;
5374 5375
	int ret = 0;

5376 5377 5378 5379 5380 5381 5382
	/*
	 * Non-sampling counters might still use the PMI to fold short
	 * hardware counters, ignore those.
	 */
	if (unlikely(!is_sampling_event(event)))
		return 0;

5383 5384 5385 5386 5387 5388 5389 5390 5391
	seq = __this_cpu_read(perf_throttled_seq);
	if (seq != hwc->interrupts_seq) {
		hwc->interrupts_seq = seq;
		hwc->interrupts = 1;
	} else {
		hwc->interrupts++;
		if (unlikely(throttle
			     && hwc->interrupts >= max_samples_per_tick)) {
			__this_cpu_inc(perf_throttled_count);
P
Peter Zijlstra 已提交
5392 5393
			hwc->interrupts = MAX_INTERRUPTS;
			perf_log_throttle(event, 0);
5394
			tick_nohz_full_kick();
5395 5396
			ret = 1;
		}
5397
	}
5398

5399
	if (event->attr.freq) {
P
Peter Zijlstra 已提交
5400
		u64 now = perf_clock();
5401
		s64 delta = now - hwc->freq_time_stamp;
5402

5403
		hwc->freq_time_stamp = now;
5404

5405
		if (delta > 0 && delta < 2*TICK_NSEC)
5406
			perf_adjust_period(event, delta, hwc->last_period, true);
5407 5408
	}

5409 5410
	/*
	 * XXX event_limit might not quite work as expected on inherited
5411
	 * events
5412 5413
	 */

5414 5415
	event->pending_kill = POLL_IN;
	if (events && atomic_dec_and_test(&event->event_limit)) {
5416
		ret = 1;
5417
		event->pending_kill = POLL_HUP;
5418 5419
		event->pending_disable = 1;
		irq_work_queue(&event->pending);
5420 5421
	}

5422
	if (event->overflow_handler)
5423
		event->overflow_handler(event, data, regs);
5424
	else
5425
		perf_event_output(event, data, regs);
5426

P
Peter Zijlstra 已提交
5427
	if (event->fasync && event->pending_kill) {
5428 5429
		event->pending_wakeup = 1;
		irq_work_queue(&event->pending);
P
Peter Zijlstra 已提交
5430 5431
	}

5432
	return ret;
5433 5434
}

5435
int perf_event_overflow(struct perf_event *event,
5436 5437
			  struct perf_sample_data *data,
			  struct pt_regs *regs)
5438
{
5439
	return __perf_event_overflow(event, 1, data, regs);
5440 5441
}

5442
/*
5443
 * Generic software event infrastructure
5444 5445
 */

5446 5447 5448 5449 5450 5451 5452
struct swevent_htable {
	struct swevent_hlist		*swevent_hlist;
	struct mutex			hlist_mutex;
	int				hlist_refcount;

	/* Recursion avoidance in each contexts */
	int				recursion[PERF_NR_CONTEXTS];
5453 5454 5455

	/* Keeps track of cpu being initialized/exited */
	bool				online;
5456 5457 5458 5459
};

static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);

5460
/*
5461 5462
 * We directly increment event->count and keep a second value in
 * event->hw.period_left to count intervals. This period event
5463 5464 5465 5466
 * is kept in the range [-sample_period, 0] so that we can use the
 * sign as trigger.
 */

5467
u64 perf_swevent_set_period(struct perf_event *event)
5468
{
5469
	struct hw_perf_event *hwc = &event->hw;
5470 5471 5472 5473 5474
	u64 period = hwc->last_period;
	u64 nr, offset;
	s64 old, val;

	hwc->last_period = hwc->sample_period;
5475 5476

again:
5477
	old = val = local64_read(&hwc->period_left);
5478 5479
	if (val < 0)
		return 0;
5480

5481 5482 5483
	nr = div64_u64(period + val, period);
	offset = nr * period;
	val -= offset;
5484
	if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5485
		goto again;
5486

5487
	return nr;
5488 5489
}

5490
static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5491
				    struct perf_sample_data *data,
5492
				    struct pt_regs *regs)
5493
{
5494
	struct hw_perf_event *hwc = &event->hw;
5495
	int throttle = 0;
5496

5497 5498
	if (!overflow)
		overflow = perf_swevent_set_period(event);
5499

5500 5501
	if (hwc->interrupts == MAX_INTERRUPTS)
		return;
5502

5503
	for (; overflow; overflow--) {
5504
		if (__perf_event_overflow(event, throttle,
5505
					    data, regs)) {
5506 5507 5508 5509 5510 5511
			/*
			 * We inhibit the overflow from happening when
			 * hwc->interrupts == MAX_INTERRUPTS.
			 */
			break;
		}
5512
		throttle = 1;
5513
	}
5514 5515
}

P
Peter Zijlstra 已提交
5516
static void perf_swevent_event(struct perf_event *event, u64 nr,
5517
			       struct perf_sample_data *data,
5518
			       struct pt_regs *regs)
5519
{
5520
	struct hw_perf_event *hwc = &event->hw;
5521

5522
	local64_add(nr, &event->count);
5523

5524 5525 5526
	if (!regs)
		return;

5527
	if (!is_sampling_event(event))
5528
		return;
5529

5530 5531 5532 5533 5534 5535
	if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
		data->period = nr;
		return perf_swevent_overflow(event, 1, data, regs);
	} else
		data->period = event->hw.last_period;

5536
	if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5537
		return perf_swevent_overflow(event, 1, data, regs);
5538

5539
	if (local64_add_negative(nr, &hwc->period_left))
5540
		return;
5541

5542
	perf_swevent_overflow(event, 0, data, regs);
5543 5544
}

5545 5546 5547
static int perf_exclude_event(struct perf_event *event,
			      struct pt_regs *regs)
{
P
Peter Zijlstra 已提交
5548
	if (event->hw.state & PERF_HES_STOPPED)
5549
		return 1;
P
Peter Zijlstra 已提交
5550

5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561
	if (regs) {
		if (event->attr.exclude_user && user_mode(regs))
			return 1;

		if (event->attr.exclude_kernel && !user_mode(regs))
			return 1;
	}

	return 0;
}

5562
static int perf_swevent_match(struct perf_event *event,
P
Peter Zijlstra 已提交
5563
				enum perf_type_id type,
L
Li Zefan 已提交
5564 5565 5566
				u32 event_id,
				struct perf_sample_data *data,
				struct pt_regs *regs)
5567
{
5568
	if (event->attr.type != type)
5569
		return 0;
5570

5571
	if (event->attr.config != event_id)
5572 5573
		return 0;

5574 5575
	if (perf_exclude_event(event, regs))
		return 0;
5576 5577 5578 5579

	return 1;
}

5580 5581 5582 5583 5584 5585 5586
static inline u64 swevent_hash(u64 type, u32 event_id)
{
	u64 val = event_id | (type << 32);

	return hash_64(val, SWEVENT_HLIST_BITS);
}

5587 5588
static inline struct hlist_head *
__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5589
{
5590 5591 5592 5593
	u64 hash = swevent_hash(type, event_id);

	return &hlist->heads[hash];
}
5594

5595 5596
/* For the read side: events when they trigger */
static inline struct hlist_head *
5597
find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5598 5599
{
	struct swevent_hlist *hlist;
5600

5601
	hlist = rcu_dereference(swhash->swevent_hlist);
5602 5603 5604
	if (!hlist)
		return NULL;

5605 5606 5607 5608 5609
	return __find_swevent_head(hlist, type, event_id);
}

/* For the event head insertion and removal in the hlist */
static inline struct hlist_head *
5610
find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5611 5612 5613 5614 5615 5616 5617 5618 5619 5620
{
	struct swevent_hlist *hlist;
	u32 event_id = event->attr.config;
	u64 type = event->attr.type;

	/*
	 * Event scheduling is always serialized against hlist allocation
	 * and release. Which makes the protected version suitable here.
	 * The context lock guarantees that.
	 */
5621
	hlist = rcu_dereference_protected(swhash->swevent_hlist,
5622 5623 5624 5625 5626
					  lockdep_is_held(&event->ctx->lock));
	if (!hlist)
		return NULL;

	return __find_swevent_head(hlist, type, event_id);
5627 5628 5629
}

static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5630
				    u64 nr,
5631 5632
				    struct perf_sample_data *data,
				    struct pt_regs *regs)
5633
{
5634
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5635
	struct perf_event *event;
5636
	struct hlist_head *head;
5637

5638
	rcu_read_lock();
5639
	head = find_swevent_head_rcu(swhash, type, event_id);
5640 5641 5642
	if (!head)
		goto end;

5643
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
L
Li Zefan 已提交
5644
		if (perf_swevent_match(event, type, event_id, data, regs))
5645
			perf_swevent_event(event, nr, data, regs);
5646
	}
5647 5648
end:
	rcu_read_unlock();
5649 5650
}

5651
int perf_swevent_get_recursion_context(void)
P
Peter Zijlstra 已提交
5652
{
5653
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
P
Peter Zijlstra 已提交
5654

5655
	return get_recursion_context(swhash->recursion);
P
Peter Zijlstra 已提交
5656
}
I
Ingo Molnar 已提交
5657
EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
P
Peter Zijlstra 已提交
5658

5659
inline void perf_swevent_put_recursion_context(int rctx)
5660
{
5661
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5662

5663
	put_recursion_context(swhash->recursion, rctx);
5664
}
5665

5666
void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5667
{
5668
	struct perf_sample_data data;
5669 5670
	int rctx;

5671
	preempt_disable_notrace();
5672 5673 5674
	rctx = perf_swevent_get_recursion_context();
	if (rctx < 0)
		return;
5675

5676
	perf_sample_data_init(&data, addr, 0);
5677

5678
	do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5679 5680

	perf_swevent_put_recursion_context(rctx);
5681
	preempt_enable_notrace();
5682 5683
}

5684
static void perf_swevent_read(struct perf_event *event)
5685 5686 5687
{
}

P
Peter Zijlstra 已提交
5688
static int perf_swevent_add(struct perf_event *event, int flags)
5689
{
5690
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5691
	struct hw_perf_event *hwc = &event->hw;
5692 5693
	struct hlist_head *head;

5694
	if (is_sampling_event(event)) {
5695
		hwc->last_period = hwc->sample_period;
5696
		perf_swevent_set_period(event);
5697
	}
5698

P
Peter Zijlstra 已提交
5699 5700
	hwc->state = !(flags & PERF_EF_START);

5701
	head = find_swevent_head(swhash, event);
5702 5703 5704 5705 5706 5707
	if (!head) {
		/*
		 * We can race with cpu hotplug code. Do not
		 * WARN if the cpu just got unplugged.
		 */
		WARN_ON_ONCE(swhash->online);
5708
		return -EINVAL;
5709
	}
5710 5711 5712

	hlist_add_head_rcu(&event->hlist_entry, head);

5713 5714 5715
	return 0;
}

P
Peter Zijlstra 已提交
5716
static void perf_swevent_del(struct perf_event *event, int flags)
5717
{
5718
	hlist_del_rcu(&event->hlist_entry);
5719 5720
}

P
Peter Zijlstra 已提交
5721
static void perf_swevent_start(struct perf_event *event, int flags)
5722
{
P
Peter Zijlstra 已提交
5723
	event->hw.state = 0;
5724
}
I
Ingo Molnar 已提交
5725

P
Peter Zijlstra 已提交
5726
static void perf_swevent_stop(struct perf_event *event, int flags)
5727
{
P
Peter Zijlstra 已提交
5728
	event->hw.state = PERF_HES_STOPPED;
5729 5730
}

5731 5732
/* Deref the hlist from the update side */
static inline struct swevent_hlist *
5733
swevent_hlist_deref(struct swevent_htable *swhash)
5734
{
5735 5736
	return rcu_dereference_protected(swhash->swevent_hlist,
					 lockdep_is_held(&swhash->hlist_mutex));
5737 5738
}

5739
static void swevent_hlist_release(struct swevent_htable *swhash)
5740
{
5741
	struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5742

5743
	if (!hlist)
5744 5745
		return;

5746
	rcu_assign_pointer(swhash->swevent_hlist, NULL);
5747
	kfree_rcu(hlist, rcu_head);
5748 5749 5750 5751
}

static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
{
5752
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5753

5754
	mutex_lock(&swhash->hlist_mutex);
5755

5756 5757
	if (!--swhash->hlist_refcount)
		swevent_hlist_release(swhash);
5758

5759
	mutex_unlock(&swhash->hlist_mutex);
5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771
}

static void swevent_hlist_put(struct perf_event *event)
{
	int cpu;

	for_each_possible_cpu(cpu)
		swevent_hlist_put_cpu(event, cpu);
}

static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
{
5772
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5773 5774
	int err = 0;

5775
	mutex_lock(&swhash->hlist_mutex);
5776

5777
	if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5778 5779 5780 5781 5782 5783 5784
		struct swevent_hlist *hlist;

		hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
		if (!hlist) {
			err = -ENOMEM;
			goto exit;
		}
5785
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
5786
	}
5787
	swhash->hlist_refcount++;
P
Peter Zijlstra 已提交
5788
exit:
5789
	mutex_unlock(&swhash->hlist_mutex);
5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809

	return err;
}

static int swevent_hlist_get(struct perf_event *event)
{
	int err;
	int cpu, failed_cpu;

	get_online_cpus();
	for_each_possible_cpu(cpu) {
		err = swevent_hlist_get_cpu(event, cpu);
		if (err) {
			failed_cpu = cpu;
			goto fail;
		}
	}
	put_online_cpus();

	return 0;
P
Peter Zijlstra 已提交
5810
fail:
5811 5812 5813 5814 5815 5816 5817 5818 5819 5820
	for_each_possible_cpu(cpu) {
		if (cpu == failed_cpu)
			break;
		swevent_hlist_put_cpu(event, cpu);
	}

	put_online_cpus();
	return err;
}

5821
struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5822

5823 5824 5825
static void sw_perf_event_destroy(struct perf_event *event)
{
	u64 event_id = event->attr.config;
5826

5827 5828
	WARN_ON(event->parent);

5829
	static_key_slow_dec(&perf_swevent_enabled[event_id]);
5830 5831 5832 5833 5834
	swevent_hlist_put(event);
}

static int perf_swevent_init(struct perf_event *event)
{
5835
	u64 event_id = event->attr.config;
5836 5837 5838 5839

	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

5840 5841 5842 5843 5844 5845
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

5846 5847 5848 5849 5850 5851 5852 5853 5854
	switch (event_id) {
	case PERF_COUNT_SW_CPU_CLOCK:
	case PERF_COUNT_SW_TASK_CLOCK:
		return -ENOENT;

	default:
		break;
	}

5855
	if (event_id >= PERF_COUNT_SW_MAX)
5856 5857 5858 5859 5860 5861 5862 5863 5864
		return -ENOENT;

	if (!event->parent) {
		int err;

		err = swevent_hlist_get(event);
		if (err)
			return err;

5865
		static_key_slow_inc(&perf_swevent_enabled[event_id]);
5866 5867 5868 5869 5870 5871
		event->destroy = sw_perf_event_destroy;
	}

	return 0;
}

5872 5873 5874 5875 5876
static int perf_swevent_event_idx(struct perf_event *event)
{
	return 0;
}

5877
static struct pmu perf_swevent = {
5878
	.task_ctx_nr	= perf_sw_context,
5879

5880
	.event_init	= perf_swevent_init,
P
Peter Zijlstra 已提交
5881 5882 5883 5884
	.add		= perf_swevent_add,
	.del		= perf_swevent_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
5885
	.read		= perf_swevent_read,
5886 5887

	.event_idx	= perf_swevent_event_idx,
5888 5889
};

5890 5891
#ifdef CONFIG_EVENT_TRACING

5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905
static int perf_tp_filter_match(struct perf_event *event,
				struct perf_sample_data *data)
{
	void *record = data->raw->data;

	if (likely(!event->filter) || filter_match_preds(event->filter, record))
		return 1;
	return 0;
}

static int perf_tp_event_match(struct perf_event *event,
				struct perf_sample_data *data,
				struct pt_regs *regs)
{
5906 5907
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;
5908 5909 5910 5911
	/*
	 * All tracepoints are from kernel-space.
	 */
	if (event->attr.exclude_kernel)
5912 5913 5914 5915 5916 5917 5918 5919 5920
		return 0;

	if (!perf_tp_filter_match(event, data))
		return 0;

	return 1;
}

void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5921 5922
		   struct pt_regs *regs, struct hlist_head *head, int rctx,
		   struct task_struct *task)
5923 5924
{
	struct perf_sample_data data;
5925 5926
	struct perf_event *event;

5927 5928 5929 5930 5931
	struct perf_raw_record raw = {
		.size = entry_size,
		.data = record,
	};

5932
	perf_sample_data_init(&data, addr, 0);
5933 5934
	data.raw = &raw;

5935
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
5936
		if (perf_tp_event_match(event, &data, regs))
5937
			perf_swevent_event(event, count, &data, regs);
5938
	}
5939

5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 5962 5963 5964
	/*
	 * If we got specified a target task, also iterate its context and
	 * deliver this event there too.
	 */
	if (task && task != current) {
		struct perf_event_context *ctx;
		struct trace_entry *entry = record;

		rcu_read_lock();
		ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
		if (!ctx)
			goto unlock;

		list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
			if (event->attr.type != PERF_TYPE_TRACEPOINT)
				continue;
			if (event->attr.config != entry->type)
				continue;
			if (perf_tp_event_match(event, &data, regs))
				perf_swevent_event(event, count, &data, regs);
		}
unlock:
		rcu_read_unlock();
	}

5965
	perf_swevent_put_recursion_context(rctx);
5966 5967 5968
}
EXPORT_SYMBOL_GPL(perf_tp_event);

5969
static void tp_perf_event_destroy(struct perf_event *event)
5970
{
5971
	perf_trace_destroy(event);
5972 5973
}

5974
static int perf_tp_event_init(struct perf_event *event)
5975
{
5976 5977
	int err;

5978 5979 5980
	if (event->attr.type != PERF_TYPE_TRACEPOINT)
		return -ENOENT;

5981 5982 5983 5984 5985 5986
	/*
	 * no branch sampling for tracepoint events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

5987 5988
	err = perf_trace_init(event);
	if (err)
5989
		return err;
5990

5991
	event->destroy = tp_perf_event_destroy;
5992

5993 5994 5995 5996
	return 0;
}

static struct pmu perf_tracepoint = {
5997 5998
	.task_ctx_nr	= perf_sw_context,

5999
	.event_init	= perf_tp_event_init,
P
Peter Zijlstra 已提交
6000 6001 6002 6003
	.add		= perf_trace_add,
	.del		= perf_trace_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
6004
	.read		= perf_swevent_read,
6005 6006

	.event_idx	= perf_swevent_event_idx,
6007 6008 6009 6010
};

static inline void perf_tp_register(void)
{
P
Peter Zijlstra 已提交
6011
	perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6012
}
L
Li Zefan 已提交
6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035 6036

static int perf_event_set_filter(struct perf_event *event, void __user *arg)
{
	char *filter_str;
	int ret;

	if (event->attr.type != PERF_TYPE_TRACEPOINT)
		return -EINVAL;

	filter_str = strndup_user(arg, PAGE_SIZE);
	if (IS_ERR(filter_str))
		return PTR_ERR(filter_str);

	ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);

	kfree(filter_str);
	return ret;
}

static void perf_event_free_filter(struct perf_event *event)
{
	ftrace_profile_free_filter(event);
}

6037
#else
L
Li Zefan 已提交
6038

6039
static inline void perf_tp_register(void)
6040 6041
{
}
L
Li Zefan 已提交
6042 6043 6044 6045 6046 6047 6048 6049 6050 6051

static int perf_event_set_filter(struct perf_event *event, void __user *arg)
{
	return -ENOENT;
}

static void perf_event_free_filter(struct perf_event *event)
{
}

6052
#endif /* CONFIG_EVENT_TRACING */
6053

6054
#ifdef CONFIG_HAVE_HW_BREAKPOINT
6055
void perf_bp_event(struct perf_event *bp, void *data)
6056
{
6057 6058 6059
	struct perf_sample_data sample;
	struct pt_regs *regs = data;

6060
	perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6061

P
Peter Zijlstra 已提交
6062
	if (!bp->hw.state && !perf_exclude_event(bp, regs))
6063
		perf_swevent_event(bp, 1, &sample, regs);
6064 6065 6066
}
#endif

6067 6068 6069
/*
 * hrtimer based swevent callback
 */
6070

6071
static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6072
{
6073 6074 6075 6076 6077
	enum hrtimer_restart ret = HRTIMER_RESTART;
	struct perf_sample_data data;
	struct pt_regs *regs;
	struct perf_event *event;
	u64 period;
6078

6079
	event = container_of(hrtimer, struct perf_event, hw.hrtimer);
P
Peter Zijlstra 已提交
6080 6081 6082 6083

	if (event->state != PERF_EVENT_STATE_ACTIVE)
		return HRTIMER_NORESTART;

6084
	event->pmu->read(event);
6085

6086
	perf_sample_data_init(&data, 0, event->hw.last_period);
6087 6088 6089
	regs = get_irq_regs();

	if (regs && !perf_exclude_event(event, regs)) {
6090
		if (!(event->attr.exclude_idle && is_idle_task(current)))
6091
			if (__perf_event_overflow(event, 1, &data, regs))
6092 6093
				ret = HRTIMER_NORESTART;
	}
6094

6095 6096
	period = max_t(u64, 10000, event->hw.sample_period);
	hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6097

6098
	return ret;
6099 6100
}

6101
static void perf_swevent_start_hrtimer(struct perf_event *event)
6102
{
6103
	struct hw_perf_event *hwc = &event->hw;
6104 6105 6106 6107
	s64 period;

	if (!is_sampling_event(event))
		return;
6108

6109 6110 6111 6112
	period = local64_read(&hwc->period_left);
	if (period) {
		if (period < 0)
			period = 10000;
P
Peter Zijlstra 已提交
6113

6114 6115 6116 6117 6118
		local64_set(&hwc->period_left, 0);
	} else {
		period = max_t(u64, 10000, hwc->sample_period);
	}
	__hrtimer_start_range_ns(&hwc->hrtimer,
6119
				ns_to_ktime(period), 0,
6120
				HRTIMER_MODE_REL_PINNED, 0);
6121
}
6122 6123

static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6124
{
6125 6126
	struct hw_perf_event *hwc = &event->hw;

6127
	if (is_sampling_event(event)) {
6128
		ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
P
Peter Zijlstra 已提交
6129
		local64_set(&hwc->period_left, ktime_to_ns(remaining));
6130 6131 6132

		hrtimer_cancel(&hwc->hrtimer);
	}
6133 6134
}

P
Peter Zijlstra 已提交
6135 6136 6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150 6151 6152 6153 6154
static void perf_swevent_init_hrtimer(struct perf_event *event)
{
	struct hw_perf_event *hwc = &event->hw;

	if (!is_sampling_event(event))
		return;

	hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
	hwc->hrtimer.function = perf_swevent_hrtimer;

	/*
	 * Since hrtimers have a fixed rate, we can do a static freq->period
	 * mapping and avoid the whole period adjust feedback stuff.
	 */
	if (event->attr.freq) {
		long freq = event->attr.sample_freq;

		event->attr.sample_period = NSEC_PER_SEC / freq;
		hwc->sample_period = event->attr.sample_period;
		local64_set(&hwc->period_left, hwc->sample_period);
6155
		hwc->last_period = hwc->sample_period;
P
Peter Zijlstra 已提交
6156 6157 6158 6159
		event->attr.freq = 0;
	}
}

6160 6161 6162 6163 6164
/*
 * Software event: cpu wall time clock
 */

static void cpu_clock_event_update(struct perf_event *event)
6165
{
6166 6167 6168
	s64 prev;
	u64 now;

P
Peter Zijlstra 已提交
6169
	now = local_clock();
6170 6171
	prev = local64_xchg(&event->hw.prev_count, now);
	local64_add(now - prev, &event->count);
6172 6173
}

P
Peter Zijlstra 已提交
6174
static void cpu_clock_event_start(struct perf_event *event, int flags)
6175
{
P
Peter Zijlstra 已提交
6176
	local64_set(&event->hw.prev_count, local_clock());
6177 6178 6179
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6180
static void cpu_clock_event_stop(struct perf_event *event, int flags)
6181
{
6182 6183 6184
	perf_swevent_cancel_hrtimer(event);
	cpu_clock_event_update(event);
}
6185

P
Peter Zijlstra 已提交
6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198
static int cpu_clock_event_add(struct perf_event *event, int flags)
{
	if (flags & PERF_EF_START)
		cpu_clock_event_start(event, flags);

	return 0;
}

static void cpu_clock_event_del(struct perf_event *event, int flags)
{
	cpu_clock_event_stop(event, flags);
}

6199 6200 6201 6202
static void cpu_clock_event_read(struct perf_event *event)
{
	cpu_clock_event_update(event);
}
6203

6204 6205 6206 6207 6208 6209 6210 6211
static int cpu_clock_event_init(struct perf_event *event)
{
	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

	if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
		return -ENOENT;

6212 6213 6214 6215 6216 6217
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6218 6219
	perf_swevent_init_hrtimer(event);

6220
	return 0;
6221 6222
}

6223
static struct pmu perf_cpu_clock = {
6224 6225
	.task_ctx_nr	= perf_sw_context,

6226
	.event_init	= cpu_clock_event_init,
P
Peter Zijlstra 已提交
6227 6228 6229 6230
	.add		= cpu_clock_event_add,
	.del		= cpu_clock_event_del,
	.start		= cpu_clock_event_start,
	.stop		= cpu_clock_event_stop,
6231
	.read		= cpu_clock_event_read,
6232 6233

	.event_idx	= perf_swevent_event_idx,
6234 6235 6236 6237 6238 6239 6240
};

/*
 * Software event: task time clock
 */

static void task_clock_event_update(struct perf_event *event, u64 now)
6241
{
6242 6243
	u64 prev;
	s64 delta;
6244

6245 6246 6247 6248
	prev = local64_xchg(&event->hw.prev_count, now);
	delta = now - prev;
	local64_add(delta, &event->count);
}
6249

P
Peter Zijlstra 已提交
6250
static void task_clock_event_start(struct perf_event *event, int flags)
6251
{
P
Peter Zijlstra 已提交
6252
	local64_set(&event->hw.prev_count, event->ctx->time);
6253 6254 6255
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6256
static void task_clock_event_stop(struct perf_event *event, int flags)
6257 6258 6259
{
	perf_swevent_cancel_hrtimer(event);
	task_clock_event_update(event, event->ctx->time);
P
Peter Zijlstra 已提交
6260 6261 6262 6263 6264 6265
}

static int task_clock_event_add(struct perf_event *event, int flags)
{
	if (flags & PERF_EF_START)
		task_clock_event_start(event, flags);
6266

P
Peter Zijlstra 已提交
6267 6268 6269 6270 6271 6272
	return 0;
}

static void task_clock_event_del(struct perf_event *event, int flags)
{
	task_clock_event_stop(event, PERF_EF_UPDATE);
6273 6274 6275 6276
}

static void task_clock_event_read(struct perf_event *event)
{
6277 6278 6279
	u64 now = perf_clock();
	u64 delta = now - event->ctx->timestamp;
	u64 time = event->ctx->time + delta;
6280 6281 6282 6283 6284

	task_clock_event_update(event, time);
}

static int task_clock_event_init(struct perf_event *event)
L
Li Zefan 已提交
6285
{
6286 6287 6288 6289 6290 6291
	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

	if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
		return -ENOENT;

6292 6293 6294 6295 6296 6297
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6298 6299
	perf_swevent_init_hrtimer(event);

6300
	return 0;
L
Li Zefan 已提交
6301 6302
}

6303
static struct pmu perf_task_clock = {
6304 6305
	.task_ctx_nr	= perf_sw_context,

6306
	.event_init	= task_clock_event_init,
P
Peter Zijlstra 已提交
6307 6308 6309 6310
	.add		= task_clock_event_add,
	.del		= task_clock_event_del,
	.start		= task_clock_event_start,
	.stop		= task_clock_event_stop,
6311
	.read		= task_clock_event_read,
6312 6313

	.event_idx	= perf_swevent_event_idx,
6314
};
L
Li Zefan 已提交
6315

P
Peter Zijlstra 已提交
6316
static void perf_pmu_nop_void(struct pmu *pmu)
6317 6318
{
}
L
Li Zefan 已提交
6319

P
Peter Zijlstra 已提交
6320
static int perf_pmu_nop_int(struct pmu *pmu)
L
Li Zefan 已提交
6321
{
P
Peter Zijlstra 已提交
6322
	return 0;
L
Li Zefan 已提交
6323 6324
}

P
Peter Zijlstra 已提交
6325
static void perf_pmu_start_txn(struct pmu *pmu)
L
Li Zefan 已提交
6326
{
P
Peter Zijlstra 已提交
6327
	perf_pmu_disable(pmu);
L
Li Zefan 已提交
6328 6329
}

P
Peter Zijlstra 已提交
6330 6331 6332 6333 6334
static int perf_pmu_commit_txn(struct pmu *pmu)
{
	perf_pmu_enable(pmu);
	return 0;
}
6335

P
Peter Zijlstra 已提交
6336
static void perf_pmu_cancel_txn(struct pmu *pmu)
6337
{
P
Peter Zijlstra 已提交
6338
	perf_pmu_enable(pmu);
6339 6340
}

6341 6342 6343 6344 6345
static int perf_event_idx_default(struct perf_event *event)
{
	return event->hw.idx + 1;
}

P
Peter Zijlstra 已提交
6346 6347 6348 6349
/*
 * Ensures all contexts with the same task_ctx_nr have the same
 * pmu_cpu_context too.
 */
6350
static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6351
{
P
Peter Zijlstra 已提交
6352
	struct pmu *pmu;
6353

P
Peter Zijlstra 已提交
6354 6355
	if (ctxn < 0)
		return NULL;
6356

P
Peter Zijlstra 已提交
6357 6358 6359 6360
	list_for_each_entry(pmu, &pmus, entry) {
		if (pmu->task_ctx_nr == ctxn)
			return pmu->pmu_cpu_context;
	}
6361

P
Peter Zijlstra 已提交
6362
	return NULL;
6363 6364
}

6365
static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6366
{
6367 6368 6369 6370 6371 6372 6373
	int cpu;

	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);

6374 6375
		if (cpuctx->unique_pmu == old_pmu)
			cpuctx->unique_pmu = pmu;
6376 6377 6378 6379 6380 6381
	}
}

static void free_pmu_context(struct pmu *pmu)
{
	struct pmu *i;
6382

P
Peter Zijlstra 已提交
6383
	mutex_lock(&pmus_lock);
6384
	/*
P
Peter Zijlstra 已提交
6385
	 * Like a real lame refcount.
6386
	 */
6387 6388 6389
	list_for_each_entry(i, &pmus, entry) {
		if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
			update_pmu_context(i, pmu);
P
Peter Zijlstra 已提交
6390
			goto out;
6391
		}
P
Peter Zijlstra 已提交
6392
	}
6393

6394
	free_percpu(pmu->pmu_cpu_context);
P
Peter Zijlstra 已提交
6395 6396
out:
	mutex_unlock(&pmus_lock);
6397
}
P
Peter Zijlstra 已提交
6398
static struct idr pmu_idr;
6399

P
Peter Zijlstra 已提交
6400 6401 6402 6403 6404 6405 6406
static ssize_t
type_show(struct device *dev, struct device_attribute *attr, char *page)
{
	struct pmu *pmu = dev_get_drvdata(dev);

	return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
}
6407
static DEVICE_ATTR_RO(type);
P
Peter Zijlstra 已提交
6408

6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451
static ssize_t
perf_event_mux_interval_ms_show(struct device *dev,
				struct device_attribute *attr,
				char *page)
{
	struct pmu *pmu = dev_get_drvdata(dev);

	return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
}

static ssize_t
perf_event_mux_interval_ms_store(struct device *dev,
				 struct device_attribute *attr,
				 const char *buf, size_t count)
{
	struct pmu *pmu = dev_get_drvdata(dev);
	int timer, cpu, ret;

	ret = kstrtoint(buf, 0, &timer);
	if (ret)
		return ret;

	if (timer < 1)
		return -EINVAL;

	/* same value, noting to do */
	if (timer == pmu->hrtimer_interval_ms)
		return count;

	pmu->hrtimer_interval_ms = timer;

	/* update all cpuctx for this PMU */
	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;
		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
		cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);

		if (hrtimer_active(&cpuctx->hrtimer))
			hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
	}

	return count;
}
6452
static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6453

6454 6455 6456 6457
static struct attribute *pmu_dev_attrs[] = {
	&dev_attr_type.attr,
	&dev_attr_perf_event_mux_interval_ms.attr,
	NULL,
P
Peter Zijlstra 已提交
6458
};
6459
ATTRIBUTE_GROUPS(pmu_dev);
P
Peter Zijlstra 已提交
6460 6461 6462 6463

static int pmu_bus_running;
static struct bus_type pmu_bus = {
	.name		= "event_source",
6464
	.dev_groups	= pmu_dev_groups,
P
Peter Zijlstra 已提交
6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477 6478 6479
};

static void pmu_dev_release(struct device *dev)
{
	kfree(dev);
}

static int pmu_dev_alloc(struct pmu *pmu)
{
	int ret = -ENOMEM;

	pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
	if (!pmu->dev)
		goto out;

6480
	pmu->dev->groups = pmu->attr_groups;
P
Peter Zijlstra 已提交
6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500
	device_initialize(pmu->dev);
	ret = dev_set_name(pmu->dev, "%s", pmu->name);
	if (ret)
		goto free_dev;

	dev_set_drvdata(pmu->dev, pmu);
	pmu->dev->bus = &pmu_bus;
	pmu->dev->release = pmu_dev_release;
	ret = device_add(pmu->dev);
	if (ret)
		goto free_dev;

out:
	return ret;

free_dev:
	put_device(pmu->dev);
	goto out;
}

6501
static struct lock_class_key cpuctx_mutex;
6502
static struct lock_class_key cpuctx_lock;
6503

6504
int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6505
{
P
Peter Zijlstra 已提交
6506
	int cpu, ret;
6507

6508
	mutex_lock(&pmus_lock);
P
Peter Zijlstra 已提交
6509 6510 6511 6512
	ret = -ENOMEM;
	pmu->pmu_disable_count = alloc_percpu(int);
	if (!pmu->pmu_disable_count)
		goto unlock;
6513

P
Peter Zijlstra 已提交
6514 6515 6516 6517 6518 6519
	pmu->type = -1;
	if (!name)
		goto skip_type;
	pmu->name = name;

	if (type < 0) {
T
Tejun Heo 已提交
6520 6521 6522
		type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
		if (type < 0) {
			ret = type;
P
Peter Zijlstra 已提交
6523 6524 6525 6526 6527
			goto free_pdc;
		}
	}
	pmu->type = type;

P
Peter Zijlstra 已提交
6528 6529 6530 6531 6532 6533
	if (pmu_bus_running) {
		ret = pmu_dev_alloc(pmu);
		if (ret)
			goto free_idr;
	}

P
Peter Zijlstra 已提交
6534
skip_type:
P
Peter Zijlstra 已提交
6535 6536 6537
	pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
	if (pmu->pmu_cpu_context)
		goto got_cpu_context;
6538

W
Wei Yongjun 已提交
6539
	ret = -ENOMEM;
P
Peter Zijlstra 已提交
6540 6541
	pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
	if (!pmu->pmu_cpu_context)
P
Peter Zijlstra 已提交
6542
		goto free_dev;
6543

P
Peter Zijlstra 已提交
6544 6545 6546 6547
	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6548
		__perf_event_init_context(&cpuctx->ctx);
6549
		lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6550
		lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6551
		cpuctx->ctx.type = cpu_context;
P
Peter Zijlstra 已提交
6552
		cpuctx->ctx.pmu = pmu;
6553 6554 6555

		__perf_cpu_hrtimer_init(cpuctx, cpu);

6556
		INIT_LIST_HEAD(&cpuctx->rotation_list);
6557
		cpuctx->unique_pmu = pmu;
P
Peter Zijlstra 已提交
6558
	}
6559

P
Peter Zijlstra 已提交
6560
got_cpu_context:
P
Peter Zijlstra 已提交
6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572 6573 6574
	if (!pmu->start_txn) {
		if (pmu->pmu_enable) {
			/*
			 * If we have pmu_enable/pmu_disable calls, install
			 * transaction stubs that use that to try and batch
			 * hardware accesses.
			 */
			pmu->start_txn  = perf_pmu_start_txn;
			pmu->commit_txn = perf_pmu_commit_txn;
			pmu->cancel_txn = perf_pmu_cancel_txn;
		} else {
			pmu->start_txn  = perf_pmu_nop_void;
			pmu->commit_txn = perf_pmu_nop_int;
			pmu->cancel_txn = perf_pmu_nop_void;
6575
		}
6576
	}
6577

P
Peter Zijlstra 已提交
6578 6579 6580 6581 6582
	if (!pmu->pmu_enable) {
		pmu->pmu_enable  = perf_pmu_nop_void;
		pmu->pmu_disable = perf_pmu_nop_void;
	}

6583 6584 6585
	if (!pmu->event_idx)
		pmu->event_idx = perf_event_idx_default;

6586
	list_add_rcu(&pmu->entry, &pmus);
P
Peter Zijlstra 已提交
6587 6588
	ret = 0;
unlock:
6589 6590
	mutex_unlock(&pmus_lock);

P
Peter Zijlstra 已提交
6591
	return ret;
P
Peter Zijlstra 已提交
6592

P
Peter Zijlstra 已提交
6593 6594 6595 6596
free_dev:
	device_del(pmu->dev);
	put_device(pmu->dev);

P
Peter Zijlstra 已提交
6597 6598 6599 6600
free_idr:
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);

P
Peter Zijlstra 已提交
6601 6602 6603
free_pdc:
	free_percpu(pmu->pmu_disable_count);
	goto unlock;
6604
}
6605
EXPORT_SYMBOL_GPL(perf_pmu_register);
6606

6607
void perf_pmu_unregister(struct pmu *pmu)
6608
{
6609 6610 6611
	mutex_lock(&pmus_lock);
	list_del_rcu(&pmu->entry);
	mutex_unlock(&pmus_lock);
6612

6613
	/*
P
Peter Zijlstra 已提交
6614 6615
	 * We dereference the pmu list under both SRCU and regular RCU, so
	 * synchronize against both of those.
6616
	 */
6617
	synchronize_srcu(&pmus_srcu);
P
Peter Zijlstra 已提交
6618
	synchronize_rcu();
6619

P
Peter Zijlstra 已提交
6620
	free_percpu(pmu->pmu_disable_count);
P
Peter Zijlstra 已提交
6621 6622
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);
P
Peter Zijlstra 已提交
6623 6624
	device_del(pmu->dev);
	put_device(pmu->dev);
6625
	free_pmu_context(pmu);
6626
}
6627
EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6628

6629 6630 6631 6632
struct pmu *perf_init_event(struct perf_event *event)
{
	struct pmu *pmu = NULL;
	int idx;
6633
	int ret;
6634 6635

	idx = srcu_read_lock(&pmus_srcu);
P
Peter Zijlstra 已提交
6636 6637 6638 6639

	rcu_read_lock();
	pmu = idr_find(&pmu_idr, event->attr.type);
	rcu_read_unlock();
6640
	if (pmu) {
6641 6642 6643 6644
		if (!try_module_get(pmu->module)) {
			pmu = ERR_PTR(-ENODEV);
			goto unlock;
		}
6645
		event->pmu = pmu;
6646 6647 6648
		ret = pmu->event_init(event);
		if (ret)
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6649
		goto unlock;
6650
	}
P
Peter Zijlstra 已提交
6651

6652
	list_for_each_entry_rcu(pmu, &pmus, entry) {
6653 6654 6655 6656
		if (!try_module_get(pmu->module)) {
			pmu = ERR_PTR(-ENODEV);
			goto unlock;
		}
6657
		event->pmu = pmu;
6658
		ret = pmu->event_init(event);
6659
		if (!ret)
P
Peter Zijlstra 已提交
6660
			goto unlock;
6661

6662 6663
		if (ret != -ENOENT) {
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6664
			goto unlock;
6665
		}
6666
	}
P
Peter Zijlstra 已提交
6667 6668
	pmu = ERR_PTR(-ENOENT);
unlock:
6669
	srcu_read_unlock(&pmus_srcu, idx);
6670

6671
	return pmu;
6672 6673
}

6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686
static void account_event_cpu(struct perf_event *event, int cpu)
{
	if (event->parent)
		return;

	if (has_branch_stack(event)) {
		if (!(event->attach_state & PERF_ATTACH_TASK))
			atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
	}
	if (is_cgroup_event(event))
		atomic_inc(&per_cpu(perf_cgroup_events, cpu));
}

6687 6688
static void account_event(struct perf_event *event)
{
6689 6690 6691
	if (event->parent)
		return;

6692 6693 6694 6695 6696 6697 6698 6699
	if (event->attach_state & PERF_ATTACH_TASK)
		static_key_slow_inc(&perf_sched_events.key);
	if (event->attr.mmap || event->attr.mmap_data)
		atomic_inc(&nr_mmap_events);
	if (event->attr.comm)
		atomic_inc(&nr_comm_events);
	if (event->attr.task)
		atomic_inc(&nr_task_events);
6700 6701 6702 6703
	if (event->attr.freq) {
		if (atomic_inc_return(&nr_freq_events) == 1)
			tick_nohz_full_kick_all();
	}
6704
	if (has_branch_stack(event))
6705
		static_key_slow_inc(&perf_sched_events.key);
6706
	if (is_cgroup_event(event))
6707
		static_key_slow_inc(&perf_sched_events.key);
6708 6709

	account_event_cpu(event, event->cpu);
6710 6711
}

T
Thomas Gleixner 已提交
6712
/*
6713
 * Allocate and initialize a event structure
T
Thomas Gleixner 已提交
6714
 */
6715
static struct perf_event *
6716
perf_event_alloc(struct perf_event_attr *attr, int cpu,
6717 6718 6719
		 struct task_struct *task,
		 struct perf_event *group_leader,
		 struct perf_event *parent_event,
6720 6721
		 perf_overflow_handler_t overflow_handler,
		 void *context)
T
Thomas Gleixner 已提交
6722
{
P
Peter Zijlstra 已提交
6723
	struct pmu *pmu;
6724 6725
	struct perf_event *event;
	struct hw_perf_event *hwc;
6726
	long err = -EINVAL;
T
Thomas Gleixner 已提交
6727

6728 6729 6730 6731 6732
	if ((unsigned)cpu >= nr_cpu_ids) {
		if (!task || cpu != -1)
			return ERR_PTR(-EINVAL);
	}

6733
	event = kzalloc(sizeof(*event), GFP_KERNEL);
6734
	if (!event)
6735
		return ERR_PTR(-ENOMEM);
T
Thomas Gleixner 已提交
6736

6737
	/*
6738
	 * Single events are their own group leaders, with an
6739 6740 6741
	 * empty sibling list:
	 */
	if (!group_leader)
6742
		group_leader = event;
6743

6744 6745
	mutex_init(&event->child_mutex);
	INIT_LIST_HEAD(&event->child_list);
6746

6747 6748 6749
	INIT_LIST_HEAD(&event->group_entry);
	INIT_LIST_HEAD(&event->event_entry);
	INIT_LIST_HEAD(&event->sibling_list);
6750
	INIT_LIST_HEAD(&event->rb_entry);
6751
	INIT_LIST_HEAD(&event->active_entry);
6752 6753
	INIT_HLIST_NODE(&event->hlist_entry);

6754

6755
	init_waitqueue_head(&event->waitq);
6756
	init_irq_work(&event->pending, perf_pending_event);
T
Thomas Gleixner 已提交
6757

6758
	mutex_init(&event->mmap_mutex);
6759

6760
	atomic_long_set(&event->refcount, 1);
6761 6762 6763 6764 6765
	event->cpu		= cpu;
	event->attr		= *attr;
	event->group_leader	= group_leader;
	event->pmu		= NULL;
	event->oncpu		= -1;
6766

6767
	event->parent		= parent_event;
6768

6769
	event->ns		= get_pid_ns(task_active_pid_ns(current));
6770
	event->id		= atomic64_inc_return(&perf_event_id);
6771

6772
	event->state		= PERF_EVENT_STATE_INACTIVE;
6773

6774 6775
	if (task) {
		event->attach_state = PERF_ATTACH_TASK;
6776 6777 6778

		if (attr->type == PERF_TYPE_TRACEPOINT)
			event->hw.tp_target = task;
6779 6780 6781 6782
#ifdef CONFIG_HAVE_HW_BREAKPOINT
		/*
		 * hw_breakpoint is a bit difficult here..
		 */
6783
		else if (attr->type == PERF_TYPE_BREAKPOINT)
6784 6785 6786 6787
			event->hw.bp_target = task;
#endif
	}

6788
	if (!overflow_handler && parent_event) {
6789
		overflow_handler = parent_event->overflow_handler;
6790 6791
		context = parent_event->overflow_handler_context;
	}
6792

6793
	event->overflow_handler	= overflow_handler;
6794
	event->overflow_handler_context = context;
6795

J
Jiri Olsa 已提交
6796
	perf_event__state_init(event);
6797

6798
	pmu = NULL;
6799

6800
	hwc = &event->hw;
6801
	hwc->sample_period = attr->sample_period;
6802
	if (attr->freq && attr->sample_freq)
6803
		hwc->sample_period = 1;
6804
	hwc->last_period = hwc->sample_period;
6805

6806
	local64_set(&hwc->period_left, hwc->sample_period);
6807

6808
	/*
6809
	 * we currently do not support PERF_FORMAT_GROUP on inherited events
6810
	 */
6811
	if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6812
		goto err_ns;
6813

6814
	pmu = perf_init_event(event);
6815
	if (!pmu)
6816 6817
		goto err_ns;
	else if (IS_ERR(pmu)) {
6818
		err = PTR_ERR(pmu);
6819
		goto err_ns;
I
Ingo Molnar 已提交
6820
	}
6821

6822
	if (!event->parent) {
6823 6824
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
			err = get_callchain_buffers();
6825 6826
			if (err)
				goto err_pmu;
6827
		}
6828
	}
6829

6830
	return event;
6831 6832 6833 6834

err_pmu:
	if (event->destroy)
		event->destroy(event);
6835
	module_put(pmu->module);
6836 6837 6838 6839 6840 6841
err_ns:
	if (event->ns)
		put_pid_ns(event->ns);
	kfree(event);

	return ERR_PTR(err);
T
Thomas Gleixner 已提交
6842 6843
}

6844 6845
static int perf_copy_attr(struct perf_event_attr __user *uattr,
			  struct perf_event_attr *attr)
6846 6847
{
	u32 size;
6848
	int ret;
6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872

	if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
		return -EFAULT;

	/*
	 * zero the full structure, so that a short copy will be nice.
	 */
	memset(attr, 0, sizeof(*attr));

	ret = get_user(size, &uattr->size);
	if (ret)
		return ret;

	if (size > PAGE_SIZE)	/* silly large */
		goto err_size;

	if (!size)		/* abi compat */
		size = PERF_ATTR_SIZE_VER0;

	if (size < PERF_ATTR_SIZE_VER0)
		goto err_size;

	/*
	 * If we're handed a bigger struct than we know of,
6873 6874 6875
	 * ensure all the unknown bits are 0 - i.e. new
	 * user-space does not rely on any kernel feature
	 * extensions we dont know about yet.
6876 6877
	 */
	if (size > sizeof(*attr)) {
6878 6879 6880
		unsigned char __user *addr;
		unsigned char __user *end;
		unsigned char val;
6881

6882 6883
		addr = (void __user *)uattr + sizeof(*attr);
		end  = (void __user *)uattr + size;
6884

6885
		for (; addr < end; addr++) {
6886 6887 6888 6889 6890 6891
			ret = get_user(val, addr);
			if (ret)
				return ret;
			if (val)
				goto err_size;
		}
6892
		size = sizeof(*attr);
6893 6894 6895 6896 6897 6898
	}

	ret = copy_from_user(attr, uattr, size);
	if (ret)
		return -EFAULT;

6899 6900 6901 6902
	/* disabled for now */
	if (attr->mmap2)
		return -EINVAL;

6903
	if (attr->__reserved_1)
6904 6905 6906 6907 6908 6909 6910 6911
		return -EINVAL;

	if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
		return -EINVAL;

	if (attr->read_format & ~(PERF_FORMAT_MAX-1))
		return -EINVAL;

6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939
	if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
		u64 mask = attr->branch_sample_type;

		/* only using defined bits */
		if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
			return -EINVAL;

		/* at least one branch bit must be set */
		if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
			return -EINVAL;

		/* propagate priv level, when not set for branch */
		if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {

			/* exclude_kernel checked on syscall entry */
			if (!attr->exclude_kernel)
				mask |= PERF_SAMPLE_BRANCH_KERNEL;

			if (!attr->exclude_user)
				mask |= PERF_SAMPLE_BRANCH_USER;

			if (!attr->exclude_hv)
				mask |= PERF_SAMPLE_BRANCH_HV;
			/*
			 * adjust user setting (for HW filter setup)
			 */
			attr->branch_sample_type = mask;
		}
6940 6941
		/* privileged levels capture (kernel, hv): check permissions */
		if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6942 6943
		    && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
6944
	}
6945

6946
	if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6947
		ret = perf_reg_validate(attr->sample_regs_user);
6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961 6962 6963 6964 6965
		if (ret)
			return ret;
	}

	if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
		if (!arch_perf_have_user_stack_dump())
			return -ENOSYS;

		/*
		 * We have __u32 type for the size, but so far
		 * we can only use __u16 as maximum due to the
		 * __u16 sample size limit.
		 */
		if (attr->sample_stack_user >= USHRT_MAX)
			ret = -EINVAL;
		else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
			ret = -EINVAL;
	}
6966

6967 6968 6969 6970 6971 6972 6973 6974 6975
out:
	return ret;

err_size:
	put_user(sizeof(*attr), &uattr->size);
	ret = -E2BIG;
	goto out;
}

6976 6977
static int
perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6978
{
6979
	struct ring_buffer *rb = NULL;
6980 6981
	int ret = -EINVAL;

6982
	if (!output_event)
6983 6984
		goto set;

6985 6986
	/* don't allow circular references */
	if (event == output_event)
6987 6988
		goto out;

6989 6990 6991 6992 6993 6994 6995
	/*
	 * Don't allow cross-cpu buffers
	 */
	if (output_event->cpu != event->cpu)
		goto out;

	/*
6996
	 * If its not a per-cpu rb, it must be the same task.
6997 6998 6999 7000
	 */
	if (output_event->cpu == -1 && output_event->ctx != event->ctx)
		goto out;

7001
set:
7002
	mutex_lock(&event->mmap_mutex);
7003 7004 7005
	/* Can't redirect output if we've got an active mmap() */
	if (atomic_read(&event->mmap_count))
		goto unlock;
7006

7007
	if (output_event) {
7008 7009 7010
		/* get the rb we want to redirect to */
		rb = ring_buffer_get(output_event);
		if (!rb)
7011
			goto unlock;
7012 7013
	}

7014
	ring_buffer_attach(event, rb);
7015

7016
	ret = 0;
7017 7018 7019
unlock:
	mutex_unlock(&event->mmap_mutex);

7020 7021 7022 7023
out:
	return ret;
}

T
Thomas Gleixner 已提交
7024
/**
7025
 * sys_perf_event_open - open a performance event, associate it to a task/cpu
I
Ingo Molnar 已提交
7026
 *
7027
 * @attr_uptr:	event_id type attributes for monitoring/sampling
T
Thomas Gleixner 已提交
7028
 * @pid:		target pid
I
Ingo Molnar 已提交
7029
 * @cpu:		target cpu
7030
 * @group_fd:		group leader event fd
T
Thomas Gleixner 已提交
7031
 */
7032 7033
SYSCALL_DEFINE5(perf_event_open,
		struct perf_event_attr __user *, attr_uptr,
7034
		pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
T
Thomas Gleixner 已提交
7035
{
7036 7037
	struct perf_event *group_leader = NULL, *output_event = NULL;
	struct perf_event *event, *sibling;
7038 7039 7040
	struct perf_event_attr attr;
	struct perf_event_context *ctx;
	struct file *event_file = NULL;
7041
	struct fd group = {NULL, 0};
M
Matt Helsley 已提交
7042
	struct task_struct *task = NULL;
7043
	struct pmu *pmu;
7044
	int event_fd;
7045
	int move_group = 0;
7046
	int err;
7047
	int f_flags = O_RDWR;
T
Thomas Gleixner 已提交
7048

7049
	/* for future expandability... */
S
Stephane Eranian 已提交
7050
	if (flags & ~PERF_FLAG_ALL)
7051 7052
		return -EINVAL;

7053 7054 7055
	err = perf_copy_attr(attr_uptr, &attr);
	if (err)
		return err;
7056

7057 7058 7059 7060 7061
	if (!attr.exclude_kernel) {
		if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
	}

7062
	if (attr.freq) {
7063
		if (attr.sample_freq > sysctl_perf_event_sample_rate)
7064
			return -EINVAL;
7065 7066 7067
	} else {
		if (attr.sample_period & (1ULL << 63))
			return -EINVAL;
7068 7069
	}

S
Stephane Eranian 已提交
7070 7071 7072 7073 7074 7075 7076 7077 7078
	/*
	 * In cgroup mode, the pid argument is used to pass the fd
	 * opened to the cgroup directory in cgroupfs. The cpu argument
	 * designates the cpu on which to monitor threads from that
	 * cgroup.
	 */
	if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
		return -EINVAL;

7079 7080 7081 7082
	if (flags & PERF_FLAG_FD_CLOEXEC)
		f_flags |= O_CLOEXEC;

	event_fd = get_unused_fd_flags(f_flags);
7083 7084 7085
	if (event_fd < 0)
		return event_fd;

7086
	if (group_fd != -1) {
7087 7088
		err = perf_fget_light(group_fd, &group);
		if (err)
7089
			goto err_fd;
7090
		group_leader = group.file->private_data;
7091 7092 7093 7094 7095 7096
		if (flags & PERF_FLAG_FD_OUTPUT)
			output_event = group_leader;
		if (flags & PERF_FLAG_FD_NO_GROUP)
			group_leader = NULL;
	}

S
Stephane Eranian 已提交
7097
	if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7098 7099 7100 7101 7102 7103 7104
		task = find_lively_task_by_vpid(pid);
		if (IS_ERR(task)) {
			err = PTR_ERR(task);
			goto err_group_fd;
		}
	}

7105 7106 7107 7108 7109 7110
	if (task && group_leader &&
	    group_leader->attr.inherit != attr.inherit) {
		err = -EINVAL;
		goto err_task;
	}

7111 7112
	get_online_cpus();

7113 7114
	event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
				 NULL, NULL);
7115 7116
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
7117
		goto err_cpus;
7118 7119
	}

S
Stephane Eranian 已提交
7120 7121
	if (flags & PERF_FLAG_PID_CGROUP) {
		err = perf_cgroup_connect(pid, event, &attr, group_leader);
7122 7123
		if (err) {
			__free_event(event);
7124
			goto err_cpus;
7125
		}
S
Stephane Eranian 已提交
7126 7127
	}

7128 7129 7130 7131 7132 7133 7134
	if (is_sampling_event(event)) {
		if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
			err = -ENOTSUPP;
			goto err_alloc;
		}
	}

7135 7136
	account_event(event);

7137 7138 7139 7140 7141
	/*
	 * Special case software events and allow them to be part of
	 * any hardware group.
	 */
	pmu = event->pmu;
7142 7143 7144 7145 7146 7147 7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164

	if (group_leader &&
	    (is_software_event(event) != is_software_event(group_leader))) {
		if (is_software_event(event)) {
			/*
			 * If event and group_leader are not both a software
			 * event, and event is, then group leader is not.
			 *
			 * Allow the addition of software events to !software
			 * groups, this is safe because software events never
			 * fail to schedule.
			 */
			pmu = group_leader->pmu;
		} else if (is_software_event(group_leader) &&
			   (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
			/*
			 * In case the group is a pure software group, and we
			 * try to add a hardware event, move the whole group to
			 * the hardware context.
			 */
			move_group = 1;
		}
	}
7165 7166 7167 7168

	/*
	 * Get the target context (task or percpu):
	 */
7169
	ctx = find_get_context(pmu, task, event->cpu);
7170 7171
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7172
		goto err_alloc;
7173 7174
	}

7175 7176 7177 7178 7179
	if (task) {
		put_task_struct(task);
		task = NULL;
	}

I
Ingo Molnar 已提交
7180
	/*
7181
	 * Look up the group leader (we will attach this event to it):
7182
	 */
7183
	if (group_leader) {
7184
		err = -EINVAL;
7185 7186

		/*
I
Ingo Molnar 已提交
7187 7188 7189 7190
		 * Do not allow a recursive hierarchy (this new sibling
		 * becoming part of another group-sibling):
		 */
		if (group_leader->group_leader != group_leader)
7191
			goto err_context;
I
Ingo Molnar 已提交
7192 7193 7194
		/*
		 * Do not allow to attach to a group in a different
		 * task or CPU context:
7195
		 */
7196 7197 7198 7199 7200 7201 7202 7203
		if (move_group) {
			if (group_leader->ctx->type != ctx->type)
				goto err_context;
		} else {
			if (group_leader->ctx != ctx)
				goto err_context;
		}

7204 7205 7206
		/*
		 * Only a group leader can be exclusive or pinned
		 */
7207
		if (attr.exclusive || attr.pinned)
7208
			goto err_context;
7209 7210 7211 7212 7213
	}

	if (output_event) {
		err = perf_event_set_output(event, output_event);
		if (err)
7214
			goto err_context;
7215
	}
T
Thomas Gleixner 已提交
7216

7217 7218
	event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
					f_flags);
7219 7220
	if (IS_ERR(event_file)) {
		err = PTR_ERR(event_file);
7221
		goto err_context;
7222
	}
7223

7224 7225 7226 7227
	if (move_group) {
		struct perf_event_context *gctx = group_leader->ctx;

		mutex_lock(&gctx->mutex);
7228
		perf_remove_from_context(group_leader, false);
J
Jiri Olsa 已提交
7229 7230 7231 7232 7233 7234 7235

		/*
		 * Removing from the context ends up with disabled
		 * event. What we want here is event in the initial
		 * startup state, ready to be add into new context.
		 */
		perf_event__state_init(group_leader);
7236 7237
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7238
			perf_remove_from_context(sibling, false);
J
Jiri Olsa 已提交
7239
			perf_event__state_init(sibling);
7240 7241 7242 7243
			put_ctx(gctx);
		}
		mutex_unlock(&gctx->mutex);
		put_ctx(gctx);
7244
	}
7245

7246
	WARN_ON_ONCE(ctx->parent_ctx);
7247
	mutex_lock(&ctx->mutex);
7248 7249

	if (move_group) {
7250
		synchronize_rcu();
7251
		perf_install_in_context(ctx, group_leader, event->cpu);
7252 7253 7254
		get_ctx(ctx);
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7255
			perf_install_in_context(ctx, sibling, event->cpu);
7256 7257 7258 7259
			get_ctx(ctx);
		}
	}

7260
	perf_install_in_context(ctx, event, event->cpu);
7261
	perf_unpin_context(ctx);
7262
	mutex_unlock(&ctx->mutex);
7263

7264 7265
	put_online_cpus();

7266
	event->owner = current;
P
Peter Zijlstra 已提交
7267

7268 7269 7270
	mutex_lock(&current->perf_event_mutex);
	list_add_tail(&event->owner_entry, &current->perf_event_list);
	mutex_unlock(&current->perf_event_mutex);
7271

7272 7273 7274 7275
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(event);
7276
	perf_event__id_header_size(event);
7277

7278 7279 7280 7281 7282 7283
	/*
	 * Drop the reference on the group_event after placing the
	 * new event on the sibling_list. This ensures destruction
	 * of the group leader will find the pointer to itself in
	 * perf_group_detach().
	 */
7284
	fdput(group);
7285 7286
	fd_install(event_fd, event_file);
	return event_fd;
T
Thomas Gleixner 已提交
7287

7288
err_context:
7289
	perf_unpin_context(ctx);
7290
	put_ctx(ctx);
7291
err_alloc:
7292
	free_event(event);
7293
err_cpus:
7294
	put_online_cpus();
7295
err_task:
P
Peter Zijlstra 已提交
7296 7297
	if (task)
		put_task_struct(task);
7298
err_group_fd:
7299
	fdput(group);
7300 7301
err_fd:
	put_unused_fd(event_fd);
7302
	return err;
T
Thomas Gleixner 已提交
7303 7304
}

7305 7306 7307 7308 7309
/**
 * perf_event_create_kernel_counter
 *
 * @attr: attributes of the counter to create
 * @cpu: cpu in which the counter is bound
M
Matt Helsley 已提交
7310
 * @task: task to profile (NULL for percpu)
7311 7312 7313
 */
struct perf_event *
perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
M
Matt Helsley 已提交
7314
				 struct task_struct *task,
7315 7316
				 perf_overflow_handler_t overflow_handler,
				 void *context)
7317 7318
{
	struct perf_event_context *ctx;
7319
	struct perf_event *event;
7320
	int err;
7321

7322 7323 7324
	/*
	 * Get the target context (task or percpu):
	 */
7325

7326 7327
	event = perf_event_alloc(attr, cpu, task, NULL, NULL,
				 overflow_handler, context);
7328 7329 7330 7331
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
		goto err;
	}
7332

7333 7334
	account_event(event);

M
Matt Helsley 已提交
7335
	ctx = find_get_context(event->pmu, task, cpu);
7336 7337
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7338
		goto err_free;
7339
	}
7340 7341 7342 7343

	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
	perf_install_in_context(ctx, event, cpu);
7344
	perf_unpin_context(ctx);
7345 7346 7347 7348
	mutex_unlock(&ctx->mutex);

	return event;

7349 7350 7351
err_free:
	free_event(event);
err:
7352
	return ERR_PTR(err);
7353
}
7354
EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7355

7356 7357 7358 7359 7360 7361 7362 7363 7364 7365 7366 7367 7368
void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
{
	struct perf_event_context *src_ctx;
	struct perf_event_context *dst_ctx;
	struct perf_event *event, *tmp;
	LIST_HEAD(events);

	src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
	dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;

	mutex_lock(&src_ctx->mutex);
	list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
				 event_entry) {
7369
		perf_remove_from_context(event, false);
7370
		unaccount_event_cpu(event, src_cpu);
7371
		put_ctx(src_ctx);
7372
		list_add(&event->migrate_entry, &events);
7373 7374 7375 7376 7377 7378
	}
	mutex_unlock(&src_ctx->mutex);

	synchronize_rcu();

	mutex_lock(&dst_ctx->mutex);
7379 7380
	list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
		list_del(&event->migrate_entry);
7381 7382
		if (event->state >= PERF_EVENT_STATE_OFF)
			event->state = PERF_EVENT_STATE_INACTIVE;
7383
		account_event_cpu(event, dst_cpu);
7384 7385 7386 7387 7388 7389 7390
		perf_install_in_context(dst_ctx, event, dst_cpu);
		get_ctx(dst_ctx);
	}
	mutex_unlock(&dst_ctx->mutex);
}
EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);

7391
static void sync_child_event(struct perf_event *child_event,
7392
			       struct task_struct *child)
7393
{
7394
	struct perf_event *parent_event = child_event->parent;
7395
	u64 child_val;
7396

7397 7398
	if (child_event->attr.inherit_stat)
		perf_event_read_event(child_event, child);
7399

P
Peter Zijlstra 已提交
7400
	child_val = perf_event_count(child_event);
7401 7402 7403 7404

	/*
	 * Add back the child's count to the parent's count:
	 */
7405
	atomic64_add(child_val, &parent_event->child_count);
7406 7407 7408 7409
	atomic64_add(child_event->total_time_enabled,
		     &parent_event->child_total_time_enabled);
	atomic64_add(child_event->total_time_running,
		     &parent_event->child_total_time_running);
7410 7411

	/*
7412
	 * Remove this event from the parent's list
7413
	 */
7414 7415 7416 7417
	WARN_ON_ONCE(parent_event->ctx->parent_ctx);
	mutex_lock(&parent_event->child_mutex);
	list_del_init(&child_event->child_list);
	mutex_unlock(&parent_event->child_mutex);
7418 7419

	/*
7420
	 * Release the parent event, if this was the last
7421 7422
	 * reference to it.
	 */
7423
	put_event(parent_event);
7424 7425
}

7426
static void
7427 7428
__perf_event_exit_task(struct perf_event *child_event,
			 struct perf_event_context *child_ctx,
7429
			 struct task_struct *child)
7430
{
7431
	perf_remove_from_context(child_event, true);
7432

7433
	/*
7434
	 * It can happen that the parent exits first, and has events
7435
	 * that are still around due to the child reference. These
7436
	 * events need to be zapped.
7437
	 */
7438
	if (child_event->parent) {
7439 7440
		sync_child_event(child_event, child);
		free_event(child_event);
7441
	}
7442 7443
}

P
Peter Zijlstra 已提交
7444
static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7445
{
7446
	struct perf_event *child_event, *next;
7447
	struct perf_event_context *child_ctx;
7448
	unsigned long flags;
7449

P
Peter Zijlstra 已提交
7450
	if (likely(!child->perf_event_ctxp[ctxn])) {
7451
		perf_event_task(child, NULL, 0);
7452
		return;
P
Peter Zijlstra 已提交
7453
	}
7454

7455
	local_irq_save(flags);
7456 7457 7458 7459 7460 7461
	/*
	 * We can't reschedule here because interrupts are disabled,
	 * and either child is current or it is a task that can't be
	 * scheduled, so we are now safe from rescheduling changing
	 * our context.
	 */
7462
	child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7463 7464 7465

	/*
	 * Take the context lock here so that if find_get_context is
7466
	 * reading child->perf_event_ctxp, we wait until it has
7467 7468
	 * incremented the context's refcount before we do put_ctx below.
	 */
7469
	raw_spin_lock(&child_ctx->lock);
7470
	task_ctx_sched_out(child_ctx);
P
Peter Zijlstra 已提交
7471
	child->perf_event_ctxp[ctxn] = NULL;
7472 7473 7474
	/*
	 * If this context is a clone; unclone it so it can't get
	 * swapped to another process while we're removing all
7475
	 * the events from it.
7476 7477
	 */
	unclone_ctx(child_ctx);
7478
	update_context_time(child_ctx);
7479
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7480 7481

	/*
7482 7483 7484
	 * Report the task dead after unscheduling the events so that we
	 * won't get any samples after PERF_RECORD_EXIT. We can however still
	 * get a few PERF_RECORD_READ events.
P
Peter Zijlstra 已提交
7485
	 */
7486
	perf_event_task(child, child_ctx, 0);
7487

7488 7489 7490
	/*
	 * We can recurse on the same lock type through:
	 *
7491 7492
	 *   __perf_event_exit_task()
	 *     sync_child_event()
7493 7494
	 *       put_event()
	 *         mutex_lock(&ctx->mutex)
7495 7496 7497
	 *
	 * But since its the parent context it won't be the same instance.
	 */
7498
	mutex_lock(&child_ctx->mutex);
7499

7500
	list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7501
		__perf_event_exit_task(child_event, child_ctx, child);
7502

7503 7504 7505
	mutex_unlock(&child_ctx->mutex);

	put_ctx(child_ctx);
7506 7507
}

P
Peter Zijlstra 已提交
7508 7509 7510 7511 7512
/*
 * When a child task exits, feed back event values to parent events.
 */
void perf_event_exit_task(struct task_struct *child)
{
P
Peter Zijlstra 已提交
7513
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7514 7515
	int ctxn;

P
Peter Zijlstra 已提交
7516 7517 7518 7519 7520 7521 7522 7523 7524 7525 7526 7527 7528 7529 7530
	mutex_lock(&child->perf_event_mutex);
	list_for_each_entry_safe(event, tmp, &child->perf_event_list,
				 owner_entry) {
		list_del_init(&event->owner_entry);

		/*
		 * Ensure the list deletion is visible before we clear
		 * the owner, closes a race against perf_release() where
		 * we need to serialize on the owner->perf_event_mutex.
		 */
		smp_wmb();
		event->owner = NULL;
	}
	mutex_unlock(&child->perf_event_mutex);

P
Peter Zijlstra 已提交
7531 7532 7533 7534
	for_each_task_context_nr(ctxn)
		perf_event_exit_task_context(child, ctxn);
}

7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546
static void perf_free_event(struct perf_event *event,
			    struct perf_event_context *ctx)
{
	struct perf_event *parent = event->parent;

	if (WARN_ON_ONCE(!parent))
		return;

	mutex_lock(&parent->child_mutex);
	list_del_init(&event->child_list);
	mutex_unlock(&parent->child_mutex);

7547
	put_event(parent);
7548

7549
	perf_group_detach(event);
7550 7551 7552 7553
	list_del_event(event, ctx);
	free_event(event);
}

7554 7555
/*
 * free an unexposed, unused context as created by inheritance by
P
Peter Zijlstra 已提交
7556
 * perf_event_init_task below, used by fork() in case of fail.
7557
 */
7558
void perf_event_free_task(struct task_struct *task)
7559
{
P
Peter Zijlstra 已提交
7560
	struct perf_event_context *ctx;
7561
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7562
	int ctxn;
7563

P
Peter Zijlstra 已提交
7564 7565 7566 7567
	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;
7568

P
Peter Zijlstra 已提交
7569
		mutex_lock(&ctx->mutex);
7570
again:
P
Peter Zijlstra 已提交
7571 7572 7573
		list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
				group_entry)
			perf_free_event(event, ctx);
7574

P
Peter Zijlstra 已提交
7575 7576 7577
		list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
				group_entry)
			perf_free_event(event, ctx);
7578

P
Peter Zijlstra 已提交
7579 7580 7581
		if (!list_empty(&ctx->pinned_groups) ||
				!list_empty(&ctx->flexible_groups))
			goto again;
7582

P
Peter Zijlstra 已提交
7583
		mutex_unlock(&ctx->mutex);
7584

P
Peter Zijlstra 已提交
7585 7586
		put_ctx(ctx);
	}
7587 7588
}

7589 7590 7591 7592 7593 7594 7595 7596
void perf_event_delayed_put(struct task_struct *task)
{
	int ctxn;

	for_each_task_context_nr(ctxn)
		WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
}

P
Peter Zijlstra 已提交
7597 7598 7599 7600 7601 7602 7603 7604 7605 7606 7607 7608
/*
 * inherit a event from parent task to child task:
 */
static struct perf_event *
inherit_event(struct perf_event *parent_event,
	      struct task_struct *parent,
	      struct perf_event_context *parent_ctx,
	      struct task_struct *child,
	      struct perf_event *group_leader,
	      struct perf_event_context *child_ctx)
{
	struct perf_event *child_event;
7609
	unsigned long flags;
P
Peter Zijlstra 已提交
7610 7611 7612 7613 7614 7615 7616 7617 7618 7619 7620 7621

	/*
	 * Instead of creating recursive hierarchies of events,
	 * we link inherited events back to the original parent,
	 * which has a filp for sure, which we use as the reference
	 * count:
	 */
	if (parent_event->parent)
		parent_event = parent_event->parent;

	child_event = perf_event_alloc(&parent_event->attr,
					   parent_event->cpu,
7622
					   child,
P
Peter Zijlstra 已提交
7623
					   group_leader, parent_event,
7624
				           NULL, NULL);
P
Peter Zijlstra 已提交
7625 7626
	if (IS_ERR(child_event))
		return child_event;
7627 7628 7629 7630 7631 7632

	if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
		free_event(child_event);
		return NULL;
	}

P
Peter Zijlstra 已提交
7633 7634 7635 7636 7637 7638 7639 7640 7641 7642 7643 7644 7645 7646 7647 7648 7649 7650 7651 7652 7653 7654 7655 7656
	get_ctx(child_ctx);

	/*
	 * Make the child state follow the state of the parent event,
	 * not its attr.disabled bit.  We hold the parent's mutex,
	 * so we won't race with perf_event_{en, dis}able_family.
	 */
	if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
		child_event->state = PERF_EVENT_STATE_INACTIVE;
	else
		child_event->state = PERF_EVENT_STATE_OFF;

	if (parent_event->attr.freq) {
		u64 sample_period = parent_event->hw.sample_period;
		struct hw_perf_event *hwc = &child_event->hw;

		hwc->sample_period = sample_period;
		hwc->last_period   = sample_period;

		local64_set(&hwc->period_left, sample_period);
	}

	child_event->ctx = child_ctx;
	child_event->overflow_handler = parent_event->overflow_handler;
7657 7658
	child_event->overflow_handler_context
		= parent_event->overflow_handler_context;
P
Peter Zijlstra 已提交
7659

7660 7661 7662 7663
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(child_event);
7664
	perf_event__id_header_size(child_event);
7665

P
Peter Zijlstra 已提交
7666 7667 7668
	/*
	 * Link it up in the child's context:
	 */
7669
	raw_spin_lock_irqsave(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7670
	add_event_to_ctx(child_event, child_ctx);
7671
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7672 7673 7674 7675 7676 7677 7678 7679 7680 7681 7682 7683 7684 7685 7686 7687 7688 7689 7690 7691 7692 7693 7694 7695 7696 7697 7698 7699 7700 7701 7702 7703 7704

	/*
	 * Link this into the parent event's child list
	 */
	WARN_ON_ONCE(parent_event->ctx->parent_ctx);
	mutex_lock(&parent_event->child_mutex);
	list_add_tail(&child_event->child_list, &parent_event->child_list);
	mutex_unlock(&parent_event->child_mutex);

	return child_event;
}

static int inherit_group(struct perf_event *parent_event,
	      struct task_struct *parent,
	      struct perf_event_context *parent_ctx,
	      struct task_struct *child,
	      struct perf_event_context *child_ctx)
{
	struct perf_event *leader;
	struct perf_event *sub;
	struct perf_event *child_ctr;

	leader = inherit_event(parent_event, parent, parent_ctx,
				 child, NULL, child_ctx);
	if (IS_ERR(leader))
		return PTR_ERR(leader);
	list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
		child_ctr = inherit_event(sub, parent, parent_ctx,
					    child, leader, child_ctx);
		if (IS_ERR(child_ctr))
			return PTR_ERR(child_ctr);
	}
	return 0;
7705 7706 7707 7708 7709
}

static int
inherit_task_group(struct perf_event *event, struct task_struct *parent,
		   struct perf_event_context *parent_ctx,
P
Peter Zijlstra 已提交
7710
		   struct task_struct *child, int ctxn,
7711 7712 7713
		   int *inherited_all)
{
	int ret;
P
Peter Zijlstra 已提交
7714
	struct perf_event_context *child_ctx;
7715 7716 7717 7718

	if (!event->attr.inherit) {
		*inherited_all = 0;
		return 0;
7719 7720
	}

7721
	child_ctx = child->perf_event_ctxp[ctxn];
7722 7723 7724 7725 7726 7727 7728
	if (!child_ctx) {
		/*
		 * This is executed from the parent task context, so
		 * inherit events that have been marked for cloning.
		 * First allocate and initialize a context for the
		 * child.
		 */
7729

7730
		child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7731 7732
		if (!child_ctx)
			return -ENOMEM;
7733

P
Peter Zijlstra 已提交
7734
		child->perf_event_ctxp[ctxn] = child_ctx;
7735 7736 7737 7738 7739 7740 7741 7742 7743
	}

	ret = inherit_group(event, parent, parent_ctx,
			    child, child_ctx);

	if (ret)
		*inherited_all = 0;

	return ret;
7744 7745
}

7746
/*
7747
 * Initialize the perf_event context in task_struct
7748
 */
P
Peter Zijlstra 已提交
7749
int perf_event_init_context(struct task_struct *child, int ctxn)
7750
{
7751
	struct perf_event_context *child_ctx, *parent_ctx;
7752 7753
	struct perf_event_context *cloned_ctx;
	struct perf_event *event;
7754
	struct task_struct *parent = current;
7755
	int inherited_all = 1;
7756
	unsigned long flags;
7757
	int ret = 0;
7758

P
Peter Zijlstra 已提交
7759
	if (likely(!parent->perf_event_ctxp[ctxn]))
7760 7761
		return 0;

7762
	/*
7763 7764
	 * If the parent's context is a clone, pin it so it won't get
	 * swapped under us.
7765
	 */
P
Peter Zijlstra 已提交
7766
	parent_ctx = perf_pin_task_context(parent, ctxn);
7767 7768
	if (!parent_ctx)
		return 0;
7769

7770 7771 7772 7773 7774 7775 7776
	/*
	 * No need to check if parent_ctx != NULL here; since we saw
	 * it non-NULL earlier, the only reason for it to become NULL
	 * is if we exit, and since we're currently in the middle of
	 * a fork we can't be exiting at the same time.
	 */

7777 7778 7779 7780
	/*
	 * Lock the parent list. No need to lock the child - not PID
	 * hashed yet and not running, so nobody can access it.
	 */
7781
	mutex_lock(&parent_ctx->mutex);
7782 7783 7784 7785 7786

	/*
	 * We dont have to disable NMIs - we are only looking at
	 * the list, not manipulating it:
	 */
7787
	list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
P
Peter Zijlstra 已提交
7788 7789
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7790 7791 7792
		if (ret)
			break;
	}
7793

7794 7795 7796 7797 7798 7799 7800 7801 7802
	/*
	 * We can't hold ctx->lock when iterating the ->flexible_group list due
	 * to allocations, but we need to prevent rotation because
	 * rotate_ctx() will change the list from interrupt context.
	 */
	raw_spin_lock_irqsave(&parent_ctx->lock, flags);
	parent_ctx->rotate_disable = 1;
	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);

7803
	list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
P
Peter Zijlstra 已提交
7804 7805
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7806
		if (ret)
7807
			break;
7808 7809
	}

7810 7811 7812
	raw_spin_lock_irqsave(&parent_ctx->lock, flags);
	parent_ctx->rotate_disable = 0;

P
Peter Zijlstra 已提交
7813
	child_ctx = child->perf_event_ctxp[ctxn];
7814

7815
	if (child_ctx && inherited_all) {
7816 7817 7818
		/*
		 * Mark the child context as a clone of the parent
		 * context, or of whatever the parent is a clone of.
P
Peter Zijlstra 已提交
7819 7820 7821
		 *
		 * Note that if the parent is a clone, the holding of
		 * parent_ctx->lock avoids it from being uncloned.
7822
		 */
P
Peter Zijlstra 已提交
7823
		cloned_ctx = parent_ctx->parent_ctx;
7824 7825
		if (cloned_ctx) {
			child_ctx->parent_ctx = cloned_ctx;
7826
			child_ctx->parent_gen = parent_ctx->parent_gen;
7827 7828 7829 7830 7831
		} else {
			child_ctx->parent_ctx = parent_ctx;
			child_ctx->parent_gen = parent_ctx->generation;
		}
		get_ctx(child_ctx->parent_ctx);
7832 7833
	}

P
Peter Zijlstra 已提交
7834
	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7835
	mutex_unlock(&parent_ctx->mutex);
7836

7837
	perf_unpin_context(parent_ctx);
7838
	put_ctx(parent_ctx);
7839

7840
	return ret;
7841 7842
}

P
Peter Zijlstra 已提交
7843 7844 7845 7846 7847 7848 7849
/*
 * Initialize the perf_event context in task_struct
 */
int perf_event_init_task(struct task_struct *child)
{
	int ctxn, ret;

7850 7851 7852 7853
	memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
	mutex_init(&child->perf_event_mutex);
	INIT_LIST_HEAD(&child->perf_event_list);

P
Peter Zijlstra 已提交
7854 7855 7856 7857 7858 7859 7860 7861 7862
	for_each_task_context_nr(ctxn) {
		ret = perf_event_init_context(child, ctxn);
		if (ret)
			return ret;
	}

	return 0;
}

7863 7864
static void __init perf_event_init_all_cpus(void)
{
7865
	struct swevent_htable *swhash;
7866 7867 7868
	int cpu;

	for_each_possible_cpu(cpu) {
7869 7870
		swhash = &per_cpu(swevent_htable, cpu);
		mutex_init(&swhash->hlist_mutex);
7871
		INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7872 7873 7874
	}
}

7875
static void perf_event_init_cpu(int cpu)
T
Thomas Gleixner 已提交
7876
{
P
Peter Zijlstra 已提交
7877
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
T
Thomas Gleixner 已提交
7878

7879
	mutex_lock(&swhash->hlist_mutex);
7880
	swhash->online = true;
7881
	if (swhash->hlist_refcount > 0) {
7882 7883
		struct swevent_hlist *hlist;

7884 7885 7886
		hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
		WARN_ON(!hlist);
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
7887
	}
7888
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
7889 7890
}

P
Peter Zijlstra 已提交
7891
#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7892
static void perf_pmu_rotate_stop(struct pmu *pmu)
T
Thomas Gleixner 已提交
7893
{
7894 7895 7896 7897 7898 7899 7900
	struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);

	WARN_ON(!irqs_disabled());

	list_del_init(&cpuctx->rotation_list);
}

P
Peter Zijlstra 已提交
7901
static void __perf_event_exit_context(void *__info)
T
Thomas Gleixner 已提交
7902
{
7903
	struct remove_event re = { .detach_group = false };
P
Peter Zijlstra 已提交
7904
	struct perf_event_context *ctx = __info;
T
Thomas Gleixner 已提交
7905

P
Peter Zijlstra 已提交
7906
	perf_pmu_rotate_stop(ctx->pmu);
7907

P
Peter Zijlstra 已提交
7908
	rcu_read_lock();
7909 7910
	list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
		__perf_remove_from_context(&re);
P
Peter Zijlstra 已提交
7911
	rcu_read_unlock();
T
Thomas Gleixner 已提交
7912
}
P
Peter Zijlstra 已提交
7913 7914 7915 7916 7917 7918 7919 7920 7921

static void perf_event_exit_cpu_context(int cpu)
{
	struct perf_event_context *ctx;
	struct pmu *pmu;
	int idx;

	idx = srcu_read_lock(&pmus_srcu);
	list_for_each_entry_rcu(pmu, &pmus, entry) {
7922
		ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
P
Peter Zijlstra 已提交
7923 7924 7925 7926 7927 7928 7929 7930

		mutex_lock(&ctx->mutex);
		smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
		mutex_unlock(&ctx->mutex);
	}
	srcu_read_unlock(&pmus_srcu, idx);
}

7931
static void perf_event_exit_cpu(int cpu)
T
Thomas Gleixner 已提交
7932
{
7933
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7934

P
Peter Zijlstra 已提交
7935 7936
	perf_event_exit_cpu_context(cpu);

7937
	mutex_lock(&swhash->hlist_mutex);
7938
	swhash->online = false;
7939 7940
	swevent_hlist_release(swhash);
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
7941 7942
}
#else
7943
static inline void perf_event_exit_cpu(int cpu) { }
T
Thomas Gleixner 已提交
7944 7945
#endif

P
Peter Zijlstra 已提交
7946 7947 7948 7949 7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960 7961 7962 7963 7964 7965
static int
perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
{
	int cpu;

	for_each_online_cpu(cpu)
		perf_event_exit_cpu(cpu);

	return NOTIFY_OK;
}

/*
 * Run the perf reboot notifier at the very last possible moment so that
 * the generic watchdog code runs as long as possible.
 */
static struct notifier_block perf_reboot_notifier = {
	.notifier_call = perf_reboot,
	.priority = INT_MIN,
};

7966
static int
T
Thomas Gleixner 已提交
7967 7968 7969 7970
perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
{
	unsigned int cpu = (long)hcpu;

7971
	switch (action & ~CPU_TASKS_FROZEN) {
T
Thomas Gleixner 已提交
7972 7973

	case CPU_UP_PREPARE:
P
Peter Zijlstra 已提交
7974
	case CPU_DOWN_FAILED:
7975
		perf_event_init_cpu(cpu);
T
Thomas Gleixner 已提交
7976 7977
		break;

P
Peter Zijlstra 已提交
7978
	case CPU_UP_CANCELED:
T
Thomas Gleixner 已提交
7979
	case CPU_DOWN_PREPARE:
7980
		perf_event_exit_cpu(cpu);
T
Thomas Gleixner 已提交
7981 7982 7983 7984 7985 7986 7987 7988
		break;
	default:
		break;
	}

	return NOTIFY_OK;
}

7989
void __init perf_event_init(void)
T
Thomas Gleixner 已提交
7990
{
7991 7992
	int ret;

P
Peter Zijlstra 已提交
7993 7994
	idr_init(&pmu_idr);

7995
	perf_event_init_all_cpus();
7996
	init_srcu_struct(&pmus_srcu);
P
Peter Zijlstra 已提交
7997 7998 7999
	perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
	perf_pmu_register(&perf_cpu_clock, NULL, -1);
	perf_pmu_register(&perf_task_clock, NULL, -1);
8000 8001
	perf_tp_register();
	perf_cpu_notifier(perf_cpu_notify);
P
Peter Zijlstra 已提交
8002
	register_reboot_notifier(&perf_reboot_notifier);
8003 8004 8005

	ret = init_hw_breakpoint();
	WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8006 8007 8008

	/* do not patch jump label more than once per second */
	jump_label_rate_limit(&perf_sched_events, HZ);
8009 8010 8011 8012 8013 8014 8015

	/*
	 * Build time assertion that we keep the data_head at the intended
	 * location.  IOW, validation we got the __reserved[] size right.
	 */
	BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
		     != 1024);
T
Thomas Gleixner 已提交
8016
}
P
Peter Zijlstra 已提交
8017 8018 8019 8020 8021 8022 8023 8024 8025 8026 8027 8028 8029 8030 8031 8032 8033 8034 8035 8036 8037 8038 8039 8040 8041 8042 8043 8044

static int __init perf_event_sysfs_init(void)
{
	struct pmu *pmu;
	int ret;

	mutex_lock(&pmus_lock);

	ret = bus_register(&pmu_bus);
	if (ret)
		goto unlock;

	list_for_each_entry(pmu, &pmus, entry) {
		if (!pmu->name || pmu->type < 0)
			continue;

		ret = pmu_dev_alloc(pmu);
		WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
	}
	pmu_bus_running = 1;
	ret = 0;

unlock:
	mutex_unlock(&pmus_lock);

	return ret;
}
device_initcall(perf_event_sysfs_init);
S
Stephane Eranian 已提交
8045 8046

#ifdef CONFIG_CGROUP_PERF
8047 8048
static struct cgroup_subsys_state *
perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
S
Stephane Eranian 已提交
8049 8050 8051
{
	struct perf_cgroup *jc;

8052
	jc = kzalloc(sizeof(*jc), GFP_KERNEL);
S
Stephane Eranian 已提交
8053 8054 8055 8056 8057 8058 8059 8060 8061 8062 8063 8064
	if (!jc)
		return ERR_PTR(-ENOMEM);

	jc->info = alloc_percpu(struct perf_cgroup_info);
	if (!jc->info) {
		kfree(jc);
		return ERR_PTR(-ENOMEM);
	}

	return &jc->css;
}

8065
static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
S
Stephane Eranian 已提交
8066
{
8067 8068
	struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);

S
Stephane Eranian 已提交
8069 8070 8071 8072 8073 8074 8075 8076 8077 8078 8079
	free_percpu(jc->info);
	kfree(jc);
}

static int __perf_cgroup_move(void *info)
{
	struct task_struct *task = info;
	perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
	return 0;
}

8080 8081
static void perf_cgroup_attach(struct cgroup_subsys_state *css,
			       struct cgroup_taskset *tset)
S
Stephane Eranian 已提交
8082
{
8083 8084
	struct task_struct *task;

8085
	cgroup_taskset_for_each(task, tset)
8086
		task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8087 8088
}

8089 8090
static void perf_cgroup_exit(struct cgroup_subsys_state *css,
			     struct cgroup_subsys_state *old_css,
8091
			     struct task_struct *task)
S
Stephane Eranian 已提交
8092 8093 8094 8095 8096 8097 8098 8099 8100
{
	/*
	 * cgroup_exit() is called in the copy_process() failure path.
	 * Ignore this case since the task hasn't ran yet, this avoids
	 * trying to poke a half freed task state from generic code.
	 */
	if (!(task->flags & PF_EXITING))
		return;

8101
	task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8102 8103
}

8104
struct cgroup_subsys perf_event_cgrp_subsys = {
8105 8106
	.css_alloc	= perf_cgroup_css_alloc,
	.css_free	= perf_cgroup_css_free,
8107
	.exit		= perf_cgroup_exit,
8108
	.attach		= perf_cgroup_attach,
S
Stephane Eranian 已提交
8109 8110
};
#endif /* CONFIG_CGROUP_PERF */