core.c 185.9 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>
T
Thomas Gleixner 已提交
42

43 44
#include "internal.h"

45 46
#include <asm/irq_regs.h>

47
struct remote_function_call {
48 49 50 51
	struct task_struct	*p;
	int			(*func)(void *info);
	void			*info;
	int			ret;
52 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
};

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 = {
85 86 87 88
		.p	= p,
		.func	= func,
		.info	= info,
		.ret	= -ESRCH, /* No such (running) process */
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
	};

	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 = {
109 110 111 112
		.p	= NULL,
		.func	= func,
		.info	= info,
		.ret	= -ENXIO, /* No such CPU */
113 114 115 116 117 118 119
	};

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

	return data.ret;
}

S
Stephane Eranian 已提交
120 121 122 123
#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
		       PERF_FLAG_FD_OUTPUT  |\
		       PERF_FLAG_PID_CGROUP)

124 125 126 127 128 129 130
/*
 * branch priv levels that need permission checks
 */
#define PERF_SAMPLE_BRANCH_PERM_PLM \
	(PERF_SAMPLE_BRANCH_KERNEL |\
	 PERF_SAMPLE_BRANCH_HV)

131 132 133 134 135 136
enum event_type_t {
	EVENT_FLEXIBLE = 0x1,
	EVENT_PINNED = 0x2,
	EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
};

S
Stephane Eranian 已提交
137 138 139 140
/*
 * perf_sched_events : >0 events exist
 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
 */
141
struct static_key_deferred perf_sched_events __read_mostly;
S
Stephane Eranian 已提交
142
static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
143
static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
S
Stephane Eranian 已提交
144

145 146 147
static atomic_t nr_mmap_events __read_mostly;
static atomic_t nr_comm_events __read_mostly;
static atomic_t nr_task_events __read_mostly;
148
static atomic_t nr_freq_events __read_mostly;
149

P
Peter Zijlstra 已提交
150 151 152 153
static LIST_HEAD(pmus);
static DEFINE_MUTEX(pmus_lock);
static struct srcu_struct pmus_srcu;

154
/*
155
 * perf event paranoia level:
156 157
 *  -1 - not paranoid at all
 *   0 - disallow raw tracepoint access for unpriv
158
 *   1 - disallow cpu events for unpriv
159
 *   2 - disallow kernel profiling for unpriv
160
 */
161
int sysctl_perf_event_paranoid __read_mostly = 1;
162

163 164
/* Minimum for 512 kiB + 1 user control page */
int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
165 166

/*
167
 * max perf event sample rate
168
 */
169 170 171 172 173 174 175 176 177
#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
178 179
static int perf_sample_allowed_ns __read_mostly =
	DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
180 181 182 183 184 185

void update_perf_cpu_limits(void)
{
	u64 tmp = perf_sample_period_ns;

	tmp *= sysctl_perf_cpu_time_max_percent;
186
	do_div(tmp, 100);
P
Peter Zijlstra 已提交
187
	ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
188
}
P
Peter Zijlstra 已提交
189

190 191
static int perf_rotate_context(struct perf_cpu_context *cpuctx);

P
Peter Zijlstra 已提交
192 193 194 195
int perf_proc_update_handler(struct ctl_table *table, int write,
		void __user *buffer, size_t *lenp,
		loff_t *ppos)
{
196
	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
P
Peter Zijlstra 已提交
197 198 199 200 201

	if (ret || !write)
		return ret;

	max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219
	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
220 221 222

	return 0;
}
223

224 225 226 227 228 229 230
/*
 * 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
231
static DEFINE_PER_CPU(u64, running_sample_length);
232 233 234 235

void perf_sample_event_took(u64 sample_len_ns)
{
	u64 avg_local_sample_len;
236
	u64 local_samples_len;
P
Peter Zijlstra 已提交
237
	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
238

P
Peter Zijlstra 已提交
239
	if (allowed_ns == 0)
240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
		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
255
	if (avg_local_sample_len <= allowed_ns)
256 257 258 259 260 261 262 263 264 265
		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;

	printk_ratelimited(KERN_WARNING
P
Peter Zijlstra 已提交
266
			"perf samples too long (%lld > %lld), lowering "
267
			"kernel.perf_event_max_sample_rate to %d\n",
P
Peter Zijlstra 已提交
268
			avg_local_sample_len, allowed_ns,
269 270 271 272 273
			sysctl_perf_event_sample_rate);

	update_perf_cpu_limits();
}

274
static atomic64_t perf_event_id;
275

276 277 278 279
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
280 281 282 283 284
			     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);
285

286
void __weak perf_event_print_debug(void)	{ }
T
Thomas Gleixner 已提交
287

288
extern __weak const char *perf_pmu_name(void)
T
Thomas Gleixner 已提交
289
{
290
	return "pmu";
T
Thomas Gleixner 已提交
291 292
}

293 294 295 296 297
static inline u64 perf_clock(void)
{
	return local_clock();
}

S
Stephane Eranian 已提交
298 299 300 301 302 303
static inline struct perf_cpu_context *
__get_cpu_context(struct perf_event_context *ctx)
{
	return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
}

304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319
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
320 321
#ifdef CONFIG_CGROUP_PERF

322 323 324 325 326 327 328 329 330 331 332
/*
 * 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;
333
	struct perf_cgroup_info	__percpu *info;
334 335
};

336 337 338 339 340
/*
 * 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
341 342 343
static inline struct perf_cgroup *
perf_cgroup_from_task(struct task_struct *task)
{
344 345
	return container_of(task_css(task, perf_subsys_id),
			    struct perf_cgroup, css);
S
Stephane Eranian 已提交
346 347 348 349 350 351 352 353
}

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);

354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369
	/* @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
370 371
}

372
static inline bool perf_tryget_cgroup(struct perf_event *event)
S
Stephane Eranian 已提交
373
{
374
	return css_tryget(&event->cgrp->css);
S
Stephane Eranian 已提交
375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422
}

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)
{
423 424
	struct perf_cgroup *cgrp;

S
Stephane Eranian 已提交
425
	/*
426 427
	 * ensure we access cgroup data only when needed and
	 * when we know the cgroup is pinned (css_get)
S
Stephane Eranian 已提交
428
	 */
429
	if (!is_cgroup_event(event))
S
Stephane Eranian 已提交
430 431
		return;

432 433 434 435 436 437
	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
438 439 440
}

static inline void
441 442
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
443 444 445 446
{
	struct perf_cgroup *cgrp;
	struct perf_cgroup_info *info;

447 448 449 450 451 452
	/*
	 * 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
453 454 455 456
		return;

	cgrp = perf_cgroup_from_task(task);
	info = this_cpu_ptr(cgrp->info);
457
	info->timestamp = ctx->timestamp;
S
Stephane Eranian 已提交
458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489
}

#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);
490 491
		if (cpuctx->unique_pmu != pmu)
			continue; /* ensure we process each cpuctx once */
S
Stephane Eranian 已提交
492 493 494 495 496 497 498 499 500

		/*
		 * 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) {
501 502
			perf_ctx_lock(cpuctx, cpuctx->task_ctx);
			perf_pmu_disable(cpuctx->ctx.pmu);
S
Stephane Eranian 已提交
503 504 505 506 507 508 509 510 511 512 513

			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) {
514
				WARN_ON_ONCE(cpuctx->cgrp);
515 516 517 518
				/*
				 * set cgrp before ctxsw in to allow
				 * event_filter_match() to not have to pass
				 * task around
S
Stephane Eranian 已提交
519 520 521 522
				 */
				cpuctx->cgrp = perf_cgroup_from_task(task);
				cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
			}
523 524
			perf_pmu_enable(cpuctx->ctx.pmu);
			perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
S
Stephane Eranian 已提交
525 526 527 528 529 530 531 532
		}
	}

	rcu_read_unlock();

	local_irq_restore(flags);
}

533 534
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
535
{
536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557
	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
558 559
}

560 561
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
562
{
563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580
	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
581 582 583 584 585 586 587 588
}

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;
589 590
	struct fd f = fdget(fd);
	int ret = 0;
S
Stephane Eranian 已提交
591

592
	if (!f.file)
S
Stephane Eranian 已提交
593 594
		return -EBADF;

595 596
	rcu_read_lock();

597
	css = css_from_dir(f.file->f_dentry, &perf_subsys);
598 599 600 601
	if (IS_ERR(css)) {
		ret = PTR_ERR(css);
		goto out;
	}
S
Stephane Eranian 已提交
602 603 604 605

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

606
	/* must be done before we fput() the file */
607 608 609 610 611
	if (!perf_tryget_cgroup(event)) {
		event->cgrp = NULL;
		ret = -ENOENT;
		goto out;
	}
612

S
Stephane Eranian 已提交
613 614 615 616 617 618 619 620 621
	/*
	 * 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;
	}
622
out:
623
	rcu_read_unlock();
624
	fdput(f);
S
Stephane Eranian 已提交
625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697
	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)
{
}

698 699
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
700 701 702
{
}

703 704
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
705 706 707 708 709 710 711 712 713 714 715
{
}

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
716 717
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747
{
}

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

748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810
/*
 * 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;
811
	int timer;
812 813 814 815 816

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

817 818 819 820 821 822 823 824 825
	/*
	 * 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);
826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847

	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
848
void perf_pmu_disable(struct pmu *pmu)
849
{
P
Peter Zijlstra 已提交
850 851 852
	int *count = this_cpu_ptr(pmu->pmu_disable_count);
	if (!(*count)++)
		pmu->pmu_disable(pmu);
853 854
}

P
Peter Zijlstra 已提交
855
void perf_pmu_enable(struct pmu *pmu)
856
{
P
Peter Zijlstra 已提交
857 858 859
	int *count = this_cpu_ptr(pmu->pmu_disable_count);
	if (!--(*count))
		pmu->pmu_enable(pmu);
860 861
}

862 863 864 865 866 867 868
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
869
static void perf_pmu_rotate_start(struct pmu *pmu)
870
{
P
Peter Zijlstra 已提交
871
	struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
872
	struct list_head *head = &__get_cpu_var(rotation_list);
873

874
	WARN_ON(!irqs_disabled());
875

876
	if (list_empty(&cpuctx->rotation_list))
877
		list_add(&cpuctx->rotation_list, head);
878 879
}

880
static void get_ctx(struct perf_event_context *ctx)
881
{
882
	WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
883 884
}

885
static void put_ctx(struct perf_event_context *ctx)
886
{
887 888 889
	if (atomic_dec_and_test(&ctx->refcount)) {
		if (ctx->parent_ctx)
			put_ctx(ctx->parent_ctx);
890 891
		if (ctx->task)
			put_task_struct(ctx->task);
892
		kfree_rcu(ctx, rcu_head);
893
	}
894 895
}

896
static void unclone_ctx(struct perf_event_context *ctx)
897 898 899 900 901
{
	if (ctx->parent_ctx) {
		put_ctx(ctx->parent_ctx);
		ctx->parent_ctx = NULL;
	}
902
	ctx->generation++;
903 904
}

905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926
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);
}

927
/*
928
 * If we inherit events we want to return the parent event id
929 930
 * to userspace.
 */
931
static u64 primary_event_id(struct perf_event *event)
932
{
933
	u64 id = event->id;
934

935 936
	if (event->parent)
		id = event->parent->id;
937 938 939 940

	return id;
}

941
/*
942
 * Get the perf_event_context for a task and lock it.
943 944 945
 * This has to cope with with the fact that until it is locked,
 * the context could get moved to another task.
 */
946
static struct perf_event_context *
P
Peter Zijlstra 已提交
947
perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
948
{
949
	struct perf_event_context *ctx;
950

P
Peter Zijlstra 已提交
951
retry:
952 953 954 955 956 957 958 959 960 961 962
	/*
	 * 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
963
	ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
964 965 966 967
	if (ctx) {
		/*
		 * If this context is a clone of another, it might
		 * get swapped for another underneath us by
968
		 * perf_event_task_sched_out, though the
969 970 971 972 973 974
		 * 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.
		 */
975
		raw_spin_lock_irqsave(&ctx->lock, *flags);
P
Peter Zijlstra 已提交
976
		if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
977
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
978 979
			rcu_read_unlock();
			preempt_enable();
980 981
			goto retry;
		}
982 983

		if (!atomic_inc_not_zero(&ctx->refcount)) {
984
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
985 986
			ctx = NULL;
		}
987 988
	}
	rcu_read_unlock();
989
	preempt_enable();
990 991 992 993 994 995 996 997
	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
998 999
static struct perf_event_context *
perf_pin_task_context(struct task_struct *task, int ctxn)
1000
{
1001
	struct perf_event_context *ctx;
1002 1003
	unsigned long flags;

P
Peter Zijlstra 已提交
1004
	ctx = perf_lock_task_context(task, ctxn, &flags);
1005 1006
	if (ctx) {
		++ctx->pin_count;
1007
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
1008 1009 1010 1011
	}
	return ctx;
}

1012
static void perf_unpin_context(struct perf_event_context *ctx)
1013 1014 1015
{
	unsigned long flags;

1016
	raw_spin_lock_irqsave(&ctx->lock, flags);
1017
	--ctx->pin_count;
1018
	raw_spin_unlock_irqrestore(&ctx->lock, flags);
1019 1020
}

1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031
/*
 * 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;
}

1032 1033 1034
static u64 perf_event_time(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;
S
Stephane Eranian 已提交
1035 1036 1037 1038

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

1039 1040 1041
	return ctx ? ctx->time : 0;
}

1042 1043
/*
 * Update the total_time_enabled and total_time_running fields for a event.
1044
 * The caller of this function needs to hold the ctx->lock.
1045 1046 1047 1048 1049 1050 1051 1052 1053
 */
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
1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064
	/*
	 * 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))
1065
		run_end = perf_cgroup_event_time(event);
S
Stephane Eranian 已提交
1066 1067
	else if (ctx->is_active)
		run_end = ctx->time;
1068 1069 1070 1071
	else
		run_end = event->tstamp_stopped;

	event->total_time_enabled = run_end - event->tstamp_enabled;
1072 1073 1074 1075

	if (event->state == PERF_EVENT_STATE_INACTIVE)
		run_end = event->tstamp_stopped;
	else
1076
		run_end = perf_event_time(event);
1077 1078

	event->total_time_running = run_end - event->tstamp_running;
S
Stephane Eranian 已提交
1079

1080 1081
}

1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093
/*
 * 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);
}

1094 1095 1096 1097 1098 1099 1100 1101 1102
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;
}

1103
/*
1104
 * Add a event from the lists for its context.
1105 1106
 * Must be called with ctx->mutex and ctx->lock held.
 */
1107
static void
1108
list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1109
{
1110 1111
	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
	event->attach_state |= PERF_ATTACH_CONTEXT;
1112 1113

	/*
1114 1115 1116
	 * 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.
1117
	 */
1118
	if (event->group_leader == event) {
1119 1120
		struct list_head *list;

1121 1122 1123
		if (is_software_event(event))
			event->group_flags |= PERF_GROUP_SOFTWARE;

1124 1125
		list = ctx_group_list(event, ctx);
		list_add_tail(&event->group_entry, list);
P
Peter Zijlstra 已提交
1126
	}
P
Peter Zijlstra 已提交
1127

1128
	if (is_cgroup_event(event))
S
Stephane Eranian 已提交
1129 1130
		ctx->nr_cgroups++;

1131 1132 1133
	if (has_branch_stack(event))
		ctx->nr_branch_stack++;

1134
	list_add_rcu(&event->event_entry, &ctx->event_list);
1135
	if (!ctx->nr_events)
P
Peter Zijlstra 已提交
1136
		perf_pmu_rotate_start(ctx->pmu);
1137 1138
	ctx->nr_events++;
	if (event->attr.inherit_stat)
1139
		ctx->nr_stat++;
1140 1141

	ctx->generation++;
1142 1143
}

J
Jiri Olsa 已提交
1144 1145 1146 1147 1148 1149 1150 1151 1152
/*
 * 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;
}

1153 1154 1155 1156 1157 1158 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
/*
 * 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);

1192 1193 1194 1195 1196 1197
	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
1198 1199 1200
	if (sample_type & PERF_SAMPLE_WEIGHT)
		size += sizeof(data->weight);

1201 1202 1203
	if (sample_type & PERF_SAMPLE_READ)
		size += event->read_size;

1204 1205 1206
	if (sample_type & PERF_SAMPLE_DATA_SRC)
		size += sizeof(data->data_src.val);

A
Andi Kleen 已提交
1207 1208 1209
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		size += sizeof(data->txn);

1210 1211 1212 1213 1214 1215 1216 1217 1218
	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;

1219 1220 1221 1222 1223 1224
	if (sample_type & PERF_SAMPLE_TID)
		size += sizeof(data->tid_entry);

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

1225 1226 1227
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		size += sizeof(data->id);

1228 1229 1230 1231 1232 1233 1234 1235 1236
	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);

1237
	event->id_header_size = size;
1238 1239
}

1240 1241
static void perf_group_attach(struct perf_event *event)
{
1242
	struct perf_event *group_leader = event->group_leader, *pos;
1243

P
Peter Zijlstra 已提交
1244 1245 1246 1247 1248 1249
	/*
	 * We can have double attach due to group movement in perf_event_open.
	 */
	if (event->attach_state & PERF_ATTACH_GROUP)
		return;

1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260
	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++;
1261 1262 1263 1264 1265

	perf_event__header_size(group_leader);

	list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
		perf_event__header_size(pos);
1266 1267
}

1268
/*
1269
 * Remove a event from the lists for its context.
1270
 * Must be called with ctx->mutex and ctx->lock held.
1271
 */
1272
static void
1273
list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1274
{
1275
	struct perf_cpu_context *cpuctx;
1276 1277 1278 1279
	/*
	 * We can have double detach due to exit/hot-unplug + close.
	 */
	if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1280
		return;
1281 1282 1283

	event->attach_state &= ~PERF_ATTACH_CONTEXT;

1284
	if (is_cgroup_event(event)) {
S
Stephane Eranian 已提交
1285
		ctx->nr_cgroups--;
1286 1287 1288 1289 1290 1291 1292 1293 1294
		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
1295

1296 1297 1298
	if (has_branch_stack(event))
		ctx->nr_branch_stack--;

1299 1300
	ctx->nr_events--;
	if (event->attr.inherit_stat)
1301
		ctx->nr_stat--;
1302

1303
	list_del_rcu(&event->event_entry);
1304

1305 1306
	if (event->group_leader == event)
		list_del_init(&event->group_entry);
P
Peter Zijlstra 已提交
1307

1308
	update_group_times(event);
1309 1310 1311 1312 1313 1314 1315 1316 1317 1318

	/*
	 * 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;
1319 1320

	ctx->generation++;
1321 1322
}

1323
static void perf_group_detach(struct perf_event *event)
1324 1325
{
	struct perf_event *sibling, *tmp;
1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341
	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--;
1342
		goto out;
1343 1344 1345 1346
	}

	if (!list_empty(&event->group_entry))
		list = &event->group_entry;
1347

1348
	/*
1349 1350
	 * If this was a group event with sibling events then
	 * upgrade the siblings to singleton events by adding them
1351
	 * to whatever list we are on.
1352
	 */
1353
	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1354 1355
		if (list)
			list_move_tail(&sibling->group_entry, list);
1356
		sibling->group_leader = sibling;
1357 1358 1359

		/* Inherit group flags from the previous leader */
		sibling->group_flags = event->group_flags;
1360
	}
1361 1362 1363 1364 1365 1366

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);
1367 1368
}

1369 1370 1371
static inline int
event_filter_match(struct perf_event *event)
{
S
Stephane Eranian 已提交
1372 1373
	return (event->cpu == -1 || event->cpu == smp_processor_id())
	    && perf_cgroup_match(event);
1374 1375
}

1376 1377
static void
event_sched_out(struct perf_event *event,
1378
		  struct perf_cpu_context *cpuctx,
1379
		  struct perf_event_context *ctx)
1380
{
1381
	u64 tstamp = perf_event_time(event);
1382 1383 1384 1385 1386 1387 1388 1389 1390
	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
1391
		delta = tstamp - event->tstamp_stopped;
1392
		event->tstamp_running += delta;
1393
		event->tstamp_stopped = tstamp;
1394 1395
	}

1396
	if (event->state != PERF_EVENT_STATE_ACTIVE)
1397
		return;
1398

1399 1400 1401 1402
	event->state = PERF_EVENT_STATE_INACTIVE;
	if (event->pending_disable) {
		event->pending_disable = 0;
		event->state = PERF_EVENT_STATE_OFF;
1403
	}
1404
	event->tstamp_stopped = tstamp;
P
Peter Zijlstra 已提交
1405
	event->pmu->del(event, 0);
1406
	event->oncpu = -1;
1407

1408
	if (!is_software_event(event))
1409 1410
		cpuctx->active_oncpu--;
	ctx->nr_active--;
1411 1412
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq--;
1413
	if (event->attr.exclusive || !cpuctx->active_oncpu)
1414 1415 1416
		cpuctx->exclusive = 0;
}

1417
static void
1418
group_sched_out(struct perf_event *group_event,
1419
		struct perf_cpu_context *cpuctx,
1420
		struct perf_event_context *ctx)
1421
{
1422
	struct perf_event *event;
1423
	int state = group_event->state;
1424

1425
	event_sched_out(group_event, cpuctx, ctx);
1426 1427 1428 1429

	/*
	 * Schedule out siblings (if any):
	 */
1430 1431
	list_for_each_entry(event, &group_event->sibling_list, group_entry)
		event_sched_out(event, cpuctx, ctx);
1432

1433
	if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1434 1435 1436
		cpuctx->exclusive = 0;
}

T
Thomas Gleixner 已提交
1437
/*
1438
 * Cross CPU call to remove a performance event
T
Thomas Gleixner 已提交
1439
 *
1440
 * We disable the event on the hardware level first. After that we
T
Thomas Gleixner 已提交
1441 1442
 * remove it from the context list.
 */
1443
static int __perf_remove_from_context(void *info)
T
Thomas Gleixner 已提交
1444
{
1445 1446
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1447
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
T
Thomas Gleixner 已提交
1448

1449
	raw_spin_lock(&ctx->lock);
1450 1451
	event_sched_out(event, cpuctx, ctx);
	list_del_event(event, ctx);
1452 1453 1454 1455
	if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
		ctx->is_active = 0;
		cpuctx->task_ctx = NULL;
	}
1456
	raw_spin_unlock(&ctx->lock);
1457 1458

	return 0;
T
Thomas Gleixner 已提交
1459 1460 1461 1462
}


/*
1463
 * Remove the event from a task's (or a CPU's) list of events.
T
Thomas Gleixner 已提交
1464
 *
1465
 * CPU events are removed with a smp call. For task events we only
T
Thomas Gleixner 已提交
1466
 * call when the task is on a CPU.
1467
 *
1468 1469
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1470 1471
 * 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.
1472
 * When called from perf_event_exit_task, it's OK because the
1473
 * context has been detached from its task.
T
Thomas Gleixner 已提交
1474
 */
1475
static void perf_remove_from_context(struct perf_event *event)
T
Thomas Gleixner 已提交
1476
{
1477
	struct perf_event_context *ctx = event->ctx;
T
Thomas Gleixner 已提交
1478 1479
	struct task_struct *task = ctx->task;

1480 1481
	lockdep_assert_held(&ctx->mutex);

T
Thomas Gleixner 已提交
1482 1483
	if (!task) {
		/*
1484
		 * Per cpu events are removed via an smp call and
1485
		 * the removal is always successful.
T
Thomas Gleixner 已提交
1486
		 */
1487
		cpu_function_call(event->cpu, __perf_remove_from_context, event);
T
Thomas Gleixner 已提交
1488 1489 1490 1491
		return;
	}

retry:
1492 1493
	if (!task_function_call(task, __perf_remove_from_context, event))
		return;
T
Thomas Gleixner 已提交
1494

1495
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1496
	/*
1497 1498
	 * 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
1499
	 */
1500
	if (ctx->is_active) {
1501
		raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1502 1503 1504 1505
		goto retry;
	}

	/*
1506 1507
	 * 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
1508
	 */
1509
	list_del_event(event, ctx);
1510
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1511 1512
}

1513
/*
1514
 * Cross CPU call to disable a performance event
1515
 */
1516
int __perf_event_disable(void *info)
1517
{
1518 1519
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1520
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1521 1522

	/*
1523 1524
	 * If this is a per-task event, need to check whether this
	 * event's task is the current task on this cpu.
1525 1526 1527
	 *
	 * Can trigger due to concurrent perf_event_context_sched_out()
	 * flipping contexts around.
1528
	 */
1529
	if (ctx->task && cpuctx->task_ctx != ctx)
1530
		return -EINVAL;
1531

1532
	raw_spin_lock(&ctx->lock);
1533 1534

	/*
1535
	 * If the event is on, turn it off.
1536 1537
	 * If it is in error state, leave it in error state.
	 */
1538
	if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1539
		update_context_time(ctx);
S
Stephane Eranian 已提交
1540
		update_cgrp_time_from_event(event);
1541 1542 1543
		update_group_times(event);
		if (event == event->group_leader)
			group_sched_out(event, cpuctx, ctx);
1544
		else
1545 1546
			event_sched_out(event, cpuctx, ctx);
		event->state = PERF_EVENT_STATE_OFF;
1547 1548
	}

1549
	raw_spin_unlock(&ctx->lock);
1550 1551

	return 0;
1552 1553 1554
}

/*
1555
 * Disable a event.
1556
 *
1557 1558
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1559
 * remains valid.  This condition is satisifed when called through
1560 1561 1562 1563
 * 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
1564
 * is the current context on this CPU and preemption is disabled,
1565
 * hence we can't get into perf_event_task_sched_out for this context.
1566
 */
1567
void perf_event_disable(struct perf_event *event)
1568
{
1569
	struct perf_event_context *ctx = event->ctx;
1570 1571 1572 1573
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
1574
		 * Disable the event on the cpu that it's on
1575
		 */
1576
		cpu_function_call(event->cpu, __perf_event_disable, event);
1577 1578 1579
		return;
	}

P
Peter Zijlstra 已提交
1580
retry:
1581 1582
	if (!task_function_call(task, __perf_event_disable, event))
		return;
1583

1584
	raw_spin_lock_irq(&ctx->lock);
1585
	/*
1586
	 * If the event is still active, we need to retry the cross-call.
1587
	 */
1588
	if (event->state == PERF_EVENT_STATE_ACTIVE) {
1589
		raw_spin_unlock_irq(&ctx->lock);
1590 1591 1592 1593 1594
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
1595 1596 1597 1598 1599 1600 1601
		goto retry;
	}

	/*
	 * Since we have the lock this context can't be scheduled
	 * in, so we can change the state safely.
	 */
1602 1603 1604
	if (event->state == PERF_EVENT_STATE_INACTIVE) {
		update_group_times(event);
		event->state = PERF_EVENT_STATE_OFF;
1605
	}
1606
	raw_spin_unlock_irq(&ctx->lock);
1607
}
1608
EXPORT_SYMBOL_GPL(perf_event_disable);
1609

S
Stephane Eranian 已提交
1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644
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
1645 1646 1647 1648
#define MAX_INTERRUPTS (~0ULL)

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

1649
static int
1650
event_sched_in(struct perf_event *event,
1651
		 struct perf_cpu_context *cpuctx,
1652
		 struct perf_event_context *ctx)
1653
{
1654 1655
	u64 tstamp = perf_event_time(event);

1656
	if (event->state <= PERF_EVENT_STATE_OFF)
1657 1658
		return 0;

1659
	event->state = PERF_EVENT_STATE_ACTIVE;
1660
	event->oncpu = smp_processor_id();
P
Peter Zijlstra 已提交
1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671

	/*
	 * 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;
	}

1672 1673 1674 1675 1676
	/*
	 * The new state must be visible before we turn it on in the hardware:
	 */
	smp_wmb();

P
Peter Zijlstra 已提交
1677
	if (event->pmu->add(event, PERF_EF_START)) {
1678 1679
		event->state = PERF_EVENT_STATE_INACTIVE;
		event->oncpu = -1;
1680 1681 1682
		return -EAGAIN;
	}

1683
	event->tstamp_running += tstamp - event->tstamp_stopped;
1684

S
Stephane Eranian 已提交
1685
	perf_set_shadow_time(event, ctx, tstamp);
1686

1687
	if (!is_software_event(event))
1688
		cpuctx->active_oncpu++;
1689
	ctx->nr_active++;
1690 1691
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq++;
1692

1693
	if (event->attr.exclusive)
1694 1695
		cpuctx->exclusive = 1;

1696 1697 1698
	return 0;
}

1699
static int
1700
group_sched_in(struct perf_event *group_event,
1701
	       struct perf_cpu_context *cpuctx,
1702
	       struct perf_event_context *ctx)
1703
{
1704
	struct perf_event *event, *partial_group = NULL;
P
Peter Zijlstra 已提交
1705
	struct pmu *pmu = group_event->pmu;
1706 1707
	u64 now = ctx->time;
	bool simulate = false;
1708

1709
	if (group_event->state == PERF_EVENT_STATE_OFF)
1710 1711
		return 0;

P
Peter Zijlstra 已提交
1712
	pmu->start_txn(pmu);
1713

1714
	if (event_sched_in(group_event, cpuctx, ctx)) {
P
Peter Zijlstra 已提交
1715
		pmu->cancel_txn(pmu);
1716
		perf_cpu_hrtimer_restart(cpuctx);
1717
		return -EAGAIN;
1718
	}
1719 1720 1721 1722

	/*
	 * Schedule in siblings as one group (if any):
	 */
1723
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1724
		if (event_sched_in(event, cpuctx, ctx)) {
1725
			partial_group = event;
1726 1727 1728 1729
			goto group_error;
		}
	}

1730
	if (!pmu->commit_txn(pmu))
1731
		return 0;
1732

1733 1734 1735 1736
group_error:
	/*
	 * Groups can be scheduled in as one unit only, so undo any
	 * partial group before returning:
1737 1738 1739 1740 1741 1742 1743 1744 1745 1746
	 * 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.
1747
	 */
1748 1749
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
		if (event == partial_group)
1750 1751 1752 1753 1754 1755 1756 1757
			simulate = true;

		if (simulate) {
			event->tstamp_running += now - event->tstamp_stopped;
			event->tstamp_stopped = now;
		} else {
			event_sched_out(event, cpuctx, ctx);
		}
1758
	}
1759
	event_sched_out(group_event, cpuctx, ctx);
1760

P
Peter Zijlstra 已提交
1761
	pmu->cancel_txn(pmu);
1762

1763 1764
	perf_cpu_hrtimer_restart(cpuctx);

1765 1766 1767
	return -EAGAIN;
}

1768
/*
1769
 * Work out whether we can put this event group on the CPU now.
1770
 */
1771
static int group_can_go_on(struct perf_event *event,
1772 1773 1774 1775
			   struct perf_cpu_context *cpuctx,
			   int can_add_hw)
{
	/*
1776
	 * Groups consisting entirely of software events can always go on.
1777
	 */
1778
	if (event->group_flags & PERF_GROUP_SOFTWARE)
1779 1780 1781
		return 1;
	/*
	 * If an exclusive group is already on, no other hardware
1782
	 * events can go on.
1783 1784 1785 1786 1787
	 */
	if (cpuctx->exclusive)
		return 0;
	/*
	 * If this group is exclusive and there are already
1788
	 * events on the CPU, it can't go on.
1789
	 */
1790
	if (event->attr.exclusive && cpuctx->active_oncpu)
1791 1792 1793 1794 1795 1796 1797 1798
		return 0;
	/*
	 * Otherwise, try to add it if all previous groups were able
	 * to go on.
	 */
	return can_add_hw;
}

1799 1800
static void add_event_to_ctx(struct perf_event *event,
			       struct perf_event_context *ctx)
1801
{
1802 1803
	u64 tstamp = perf_event_time(event);

1804
	list_add_event(event, ctx);
1805
	perf_group_attach(event);
1806 1807 1808
	event->tstamp_enabled = tstamp;
	event->tstamp_running = tstamp;
	event->tstamp_stopped = tstamp;
1809 1810
}

1811 1812 1813 1814 1815 1816
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);
1817

1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829
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
1830
/*
1831
 * Cross CPU call to install and enable a performance event
1832 1833
 *
 * Must be called with ctx->mutex held
T
Thomas Gleixner 已提交
1834
 */
1835
static int  __perf_install_in_context(void *info)
T
Thomas Gleixner 已提交
1836
{
1837 1838
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1839
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1840 1841 1842
	struct perf_event_context *task_ctx = cpuctx->task_ctx;
	struct task_struct *task = current;

1843
	perf_ctx_lock(cpuctx, task_ctx);
1844
	perf_pmu_disable(cpuctx->ctx.pmu);
T
Thomas Gleixner 已提交
1845 1846

	/*
1847
	 * If there was an active task_ctx schedule it out.
T
Thomas Gleixner 已提交
1848
	 */
1849
	if (task_ctx)
1850
		task_ctx_sched_out(task_ctx);
1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864

	/*
	 * 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;
1865 1866
		task = task_ctx->task;
	}
1867

1868
	cpu_ctx_sched_out(cpuctx, EVENT_ALL);
T
Thomas Gleixner 已提交
1869

1870
	update_context_time(ctx);
S
Stephane Eranian 已提交
1871 1872 1873 1874 1875 1876
	/*
	 * 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
1877

1878
	add_event_to_ctx(event, ctx);
T
Thomas Gleixner 已提交
1879

1880
	/*
1881
	 * Schedule everything back in
1882
	 */
1883
	perf_event_sched_in(cpuctx, task_ctx, task);
1884 1885 1886

	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, task_ctx);
1887 1888

	return 0;
T
Thomas Gleixner 已提交
1889 1890 1891
}

/*
1892
 * Attach a performance event to a context
T
Thomas Gleixner 已提交
1893
 *
1894 1895
 * First we add the event to the list with the hardware enable bit
 * in event->hw_config cleared.
T
Thomas Gleixner 已提交
1896
 *
1897
 * If the event is attached to a task which is on a CPU we use a smp
T
Thomas Gleixner 已提交
1898 1899 1900 1901
 * 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
1902 1903
perf_install_in_context(struct perf_event_context *ctx,
			struct perf_event *event,
T
Thomas Gleixner 已提交
1904 1905 1906 1907
			int cpu)
{
	struct task_struct *task = ctx->task;

1908 1909
	lockdep_assert_held(&ctx->mutex);

1910
	event->ctx = ctx;
1911 1912
	if (event->cpu != -1)
		event->cpu = cpu;
1913

T
Thomas Gleixner 已提交
1914 1915
	if (!task) {
		/*
1916
		 * Per cpu events are installed via an smp call and
1917
		 * the install is always successful.
T
Thomas Gleixner 已提交
1918
		 */
1919
		cpu_function_call(cpu, __perf_install_in_context, event);
T
Thomas Gleixner 已提交
1920 1921 1922 1923
		return;
	}

retry:
1924 1925
	if (!task_function_call(task, __perf_install_in_context, event))
		return;
T
Thomas Gleixner 已提交
1926

1927
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1928
	/*
1929 1930
	 * 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
1931
	 */
1932
	if (ctx->is_active) {
1933
		raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1934 1935 1936 1937
		goto retry;
	}

	/*
1938 1939
	 * 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
1940
	 */
1941
	add_event_to_ctx(event, ctx);
1942
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1943 1944
}

1945
/*
1946
 * Put a event into inactive state and update time fields.
1947 1948 1949 1950 1951 1952
 * 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.
 */
1953
static void __perf_event_mark_enabled(struct perf_event *event)
1954
{
1955
	struct perf_event *sub;
1956
	u64 tstamp = perf_event_time(event);
1957

1958
	event->state = PERF_EVENT_STATE_INACTIVE;
1959
	event->tstamp_enabled = tstamp - event->total_time_enabled;
P
Peter Zijlstra 已提交
1960
	list_for_each_entry(sub, &event->sibling_list, group_entry) {
1961 1962
		if (sub->state >= PERF_EVENT_STATE_INACTIVE)
			sub->tstamp_enabled = tstamp - sub->total_time_enabled;
P
Peter Zijlstra 已提交
1963
	}
1964 1965
}

1966
/*
1967
 * Cross CPU call to enable a performance event
1968
 */
1969
static int __perf_event_enable(void *info)
1970
{
1971 1972 1973
	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
1974
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1975
	int err;
1976

1977 1978 1979 1980 1981 1982 1983 1984 1985 1986
	/*
	 * 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)
1987
		return -EINVAL;
1988

1989
	raw_spin_lock(&ctx->lock);
1990
	update_context_time(ctx);
1991

1992
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
1993
		goto unlock;
S
Stephane Eranian 已提交
1994 1995 1996 1997

	/*
	 * set current task's cgroup time reference point
	 */
1998
	perf_cgroup_set_timestamp(current, ctx);
S
Stephane Eranian 已提交
1999

2000
	__perf_event_mark_enabled(event);
2001

S
Stephane Eranian 已提交
2002 2003 2004
	if (!event_filter_match(event)) {
		if (is_cgroup_event(event))
			perf_cgroup_defer_enabled(event);
2005
		goto unlock;
S
Stephane Eranian 已提交
2006
	}
2007

2008
	/*
2009
	 * If the event is in a group and isn't the group leader,
2010
	 * then don't put it on unless the group is on.
2011
	 */
2012
	if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2013
		goto unlock;
2014

2015
	if (!group_can_go_on(event, cpuctx, 1)) {
2016
		err = -EEXIST;
2017
	} else {
2018
		if (event == leader)
2019
			err = group_sched_in(event, cpuctx, ctx);
2020
		else
2021
			err = event_sched_in(event, cpuctx, ctx);
2022
	}
2023 2024 2025

	if (err) {
		/*
2026
		 * If this event can't go on and it's part of a
2027 2028
		 * group, then the whole group has to come off.
		 */
2029
		if (leader != event) {
2030
			group_sched_out(leader, cpuctx, ctx);
2031 2032
			perf_cpu_hrtimer_restart(cpuctx);
		}
2033
		if (leader->attr.pinned) {
2034
			update_group_times(leader);
2035
			leader->state = PERF_EVENT_STATE_ERROR;
2036
		}
2037 2038
	}

P
Peter Zijlstra 已提交
2039
unlock:
2040
	raw_spin_unlock(&ctx->lock);
2041 2042

	return 0;
2043 2044 2045
}

/*
2046
 * Enable a event.
2047
 *
2048 2049
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
2050
 * remains valid.  This condition is satisfied when called through
2051 2052
 * perf_event_for_each_child or perf_event_for_each as described
 * for perf_event_disable.
2053
 */
2054
void perf_event_enable(struct perf_event *event)
2055
{
2056
	struct perf_event_context *ctx = event->ctx;
2057 2058 2059 2060
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
2061
		 * Enable the event on the cpu that it's on
2062
		 */
2063
		cpu_function_call(event->cpu, __perf_event_enable, event);
2064 2065 2066
		return;
	}

2067
	raw_spin_lock_irq(&ctx->lock);
2068
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2069 2070 2071
		goto out;

	/*
2072 2073
	 * If the event is in error state, clear that first.
	 * That way, if we see the event in error state below, we
2074 2075 2076 2077
	 * know that it has gone back into error state, as distinct
	 * from the task having been scheduled away before the
	 * cross-call arrived.
	 */
2078 2079
	if (event->state == PERF_EVENT_STATE_ERROR)
		event->state = PERF_EVENT_STATE_OFF;
2080

P
Peter Zijlstra 已提交
2081
retry:
2082
	if (!ctx->is_active) {
2083
		__perf_event_mark_enabled(event);
2084 2085 2086
		goto out;
	}

2087
	raw_spin_unlock_irq(&ctx->lock);
2088 2089 2090

	if (!task_function_call(task, __perf_event_enable, event))
		return;
2091

2092
	raw_spin_lock_irq(&ctx->lock);
2093 2094

	/*
2095
	 * If the context is active and the event is still off,
2096 2097
	 * we need to retry the cross-call.
	 */
2098 2099 2100 2101 2102 2103
	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;
2104
		goto retry;
2105
	}
2106

P
Peter Zijlstra 已提交
2107
out:
2108
	raw_spin_unlock_irq(&ctx->lock);
2109
}
2110
EXPORT_SYMBOL_GPL(perf_event_enable);
2111

2112
int perf_event_refresh(struct perf_event *event, int refresh)
2113
{
2114
	/*
2115
	 * not supported on inherited events
2116
	 */
2117
	if (event->attr.inherit || !is_sampling_event(event))
2118 2119
		return -EINVAL;

2120 2121
	atomic_add(refresh, &event->event_limit);
	perf_event_enable(event);
2122 2123

	return 0;
2124
}
2125
EXPORT_SYMBOL_GPL(perf_event_refresh);
2126

2127 2128 2129
static void ctx_sched_out(struct perf_event_context *ctx,
			  struct perf_cpu_context *cpuctx,
			  enum event_type_t event_type)
2130
{
2131
	struct perf_event *event;
2132
	int is_active = ctx->is_active;
2133

2134
	ctx->is_active &= ~event_type;
2135
	if (likely(!ctx->nr_events))
2136 2137
		return;

2138
	update_context_time(ctx);
S
Stephane Eranian 已提交
2139
	update_cgrp_time_from_cpuctx(cpuctx);
2140
	if (!ctx->nr_active)
2141
		return;
2142

P
Peter Zijlstra 已提交
2143
	perf_pmu_disable(ctx->pmu);
2144
	if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2145 2146
		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2147
	}
2148

2149
	if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2150
		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2151
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2152
	}
P
Peter Zijlstra 已提交
2153
	perf_pmu_enable(ctx->pmu);
2154 2155
}

2156
/*
2157 2158 2159 2160 2161 2162
 * 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().
2163
 */
2164 2165
static int context_equiv(struct perf_event_context *ctx1,
			 struct perf_event_context *ctx2)
2166
{
2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188
	/* 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;
2189 2190
}

2191 2192
static void __perf_event_sync_stat(struct perf_event *event,
				     struct perf_event *next_event)
2193 2194 2195
{
	u64 value;

2196
	if (!event->attr.inherit_stat)
2197 2198 2199
		return;

	/*
2200
	 * Update the event value, we cannot use perf_event_read()
2201 2202
	 * because we're in the middle of a context switch and have IRQs
	 * disabled, which upsets smp_call_function_single(), however
2203
	 * we know the event must be on the current CPU, therefore we
2204 2205
	 * don't need to use it.
	 */
2206 2207
	switch (event->state) {
	case PERF_EVENT_STATE_ACTIVE:
2208 2209
		event->pmu->read(event);
		/* fall-through */
2210

2211 2212
	case PERF_EVENT_STATE_INACTIVE:
		update_event_times(event);
2213 2214 2215 2216 2217 2218 2219
		break;

	default:
		break;
	}

	/*
2220
	 * In order to keep per-task stats reliable we need to flip the event
2221 2222
	 * values when we flip the contexts.
	 */
2223 2224 2225
	value = local64_read(&next_event->count);
	value = local64_xchg(&event->count, value);
	local64_set(&next_event->count, value);
2226

2227 2228
	swap(event->total_time_enabled, next_event->total_time_enabled);
	swap(event->total_time_running, next_event->total_time_running);
2229

2230
	/*
2231
	 * Since we swizzled the values, update the user visible data too.
2232
	 */
2233 2234
	perf_event_update_userpage(event);
	perf_event_update_userpage(next_event);
2235 2236 2237 2238 2239
}

#define list_next_entry(pos, member) \
	list_entry(pos->member.next, typeof(*pos), member)

2240 2241
static void perf_event_sync_stat(struct perf_event_context *ctx,
				   struct perf_event_context *next_ctx)
2242
{
2243
	struct perf_event *event, *next_event;
2244 2245 2246 2247

	if (!ctx->nr_stat)
		return;

2248 2249
	update_context_time(ctx);

2250 2251
	event = list_first_entry(&ctx->event_list,
				   struct perf_event, event_entry);
2252

2253 2254
	next_event = list_first_entry(&next_ctx->event_list,
					struct perf_event, event_entry);
2255

2256 2257
	while (&event->event_entry != &ctx->event_list &&
	       &next_event->event_entry != &next_ctx->event_list) {
2258

2259
		__perf_event_sync_stat(event, next_event);
2260

2261 2262
		event = list_next_entry(event, event_entry);
		next_event = list_next_entry(next_event, event_entry);
2263 2264 2265
	}
}

2266 2267
static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
					 struct task_struct *next)
T
Thomas Gleixner 已提交
2268
{
P
Peter Zijlstra 已提交
2269
	struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2270
	struct perf_event_context *next_ctx;
2271
	struct perf_event_context *parent, *next_parent;
P
Peter Zijlstra 已提交
2272
	struct perf_cpu_context *cpuctx;
2273
	int do_switch = 1;
T
Thomas Gleixner 已提交
2274

P
Peter Zijlstra 已提交
2275 2276
	if (likely(!ctx))
		return;
2277

P
Peter Zijlstra 已提交
2278 2279
	cpuctx = __get_cpu_context(ctx);
	if (!cpuctx->task_ctx)
T
Thomas Gleixner 已提交
2280 2281
		return;

2282
	rcu_read_lock();
P
Peter Zijlstra 已提交
2283
	next_ctx = next->perf_event_ctxp[ctxn];
2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294
	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) {
2295 2296 2297 2298 2299 2300 2301 2302 2303
		/*
		 * 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.
		 */
2304 2305
		raw_spin_lock(&ctx->lock);
		raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2306
		if (context_equiv(ctx, next_ctx)) {
2307 2308
			/*
			 * XXX do we need a memory barrier of sorts
2309
			 * wrt to rcu_dereference() of perf_event_ctxp
2310
			 */
P
Peter Zijlstra 已提交
2311 2312
			task->perf_event_ctxp[ctxn] = next_ctx;
			next->perf_event_ctxp[ctxn] = ctx;
2313 2314 2315
			ctx->task = next;
			next_ctx->task = task;
			do_switch = 0;
2316

2317
			perf_event_sync_stat(ctx, next_ctx);
2318
		}
2319 2320
		raw_spin_unlock(&next_ctx->lock);
		raw_spin_unlock(&ctx->lock);
2321
	}
2322
unlock:
2323
	rcu_read_unlock();
2324

2325
	if (do_switch) {
2326
		raw_spin_lock(&ctx->lock);
2327
		ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2328
		cpuctx->task_ctx = NULL;
2329
		raw_spin_unlock(&ctx->lock);
2330
	}
T
Thomas Gleixner 已提交
2331 2332
}

P
Peter Zijlstra 已提交
2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346
#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.
 */
2347 2348
void __perf_event_task_sched_out(struct task_struct *task,
				 struct task_struct *next)
P
Peter Zijlstra 已提交
2349 2350 2351 2352 2353
{
	int ctxn;

	for_each_task_context_nr(ctxn)
		perf_event_context_sched_out(task, ctxn, next);
S
Stephane Eranian 已提交
2354 2355 2356 2357 2358 2359 2360

	/*
	 * 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)))
2361
		perf_cgroup_sched_out(task, next);
P
Peter Zijlstra 已提交
2362 2363
}

2364
static void task_ctx_sched_out(struct perf_event_context *ctx)
2365
{
P
Peter Zijlstra 已提交
2366
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2367

2368 2369
	if (!cpuctx->task_ctx)
		return;
2370 2371 2372 2373

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

2374
	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2375 2376 2377
	cpuctx->task_ctx = NULL;
}

2378 2379 2380 2381 2382 2383 2384
/*
 * 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);
2385 2386
}

2387
static void
2388
ctx_pinned_sched_in(struct perf_event_context *ctx,
2389
		    struct perf_cpu_context *cpuctx)
T
Thomas Gleixner 已提交
2390
{
2391
	struct perf_event *event;
T
Thomas Gleixner 已提交
2392

2393 2394
	list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
		if (event->state <= PERF_EVENT_STATE_OFF)
2395
			continue;
2396
		if (!event_filter_match(event))
2397 2398
			continue;

S
Stephane Eranian 已提交
2399 2400 2401 2402
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

2403
		if (group_can_go_on(event, cpuctx, 1))
2404
			group_sched_in(event, cpuctx, ctx);
2405 2406 2407 2408 2409

		/*
		 * If this pinned group hasn't been scheduled,
		 * put it in error state.
		 */
2410 2411 2412
		if (event->state == PERF_EVENT_STATE_INACTIVE) {
			update_group_times(event);
			event->state = PERF_EVENT_STATE_ERROR;
2413
		}
2414
	}
2415 2416 2417 2418
}

static void
ctx_flexible_sched_in(struct perf_event_context *ctx,
2419
		      struct perf_cpu_context *cpuctx)
2420 2421 2422
{
	struct perf_event *event;
	int can_add_hw = 1;
2423

2424 2425 2426
	list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
		/* Ignore events in OFF or ERROR state */
		if (event->state <= PERF_EVENT_STATE_OFF)
2427
			continue;
2428 2429
		/*
		 * Listen to the 'cpu' scheduling filter constraint
2430
		 * of events:
2431
		 */
2432
		if (!event_filter_match(event))
T
Thomas Gleixner 已提交
2433 2434
			continue;

S
Stephane Eranian 已提交
2435 2436 2437 2438
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

P
Peter Zijlstra 已提交
2439
		if (group_can_go_on(event, cpuctx, can_add_hw)) {
2440
			if (group_sched_in(event, cpuctx, ctx))
2441
				can_add_hw = 0;
P
Peter Zijlstra 已提交
2442
		}
T
Thomas Gleixner 已提交
2443
	}
2444 2445 2446 2447 2448
}

static void
ctx_sched_in(struct perf_event_context *ctx,
	     struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2449 2450
	     enum event_type_t event_type,
	     struct task_struct *task)
2451
{
S
Stephane Eranian 已提交
2452
	u64 now;
2453
	int is_active = ctx->is_active;
S
Stephane Eranian 已提交
2454

2455
	ctx->is_active |= event_type;
2456
	if (likely(!ctx->nr_events))
2457
		return;
2458

S
Stephane Eranian 已提交
2459 2460
	now = perf_clock();
	ctx->timestamp = now;
2461
	perf_cgroup_set_timestamp(task, ctx);
2462 2463 2464 2465
	/*
	 * First go through the list and put on any pinned groups
	 * in order to give them the best chance of going on.
	 */
2466
	if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2467
		ctx_pinned_sched_in(ctx, cpuctx);
2468 2469

	/* Then walk through the lower prio flexible groups */
2470
	if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2471
		ctx_flexible_sched_in(ctx, cpuctx);
2472 2473
}

2474
static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2475 2476
			     enum event_type_t event_type,
			     struct task_struct *task)
2477 2478 2479
{
	struct perf_event_context *ctx = &cpuctx->ctx;

S
Stephane Eranian 已提交
2480
	ctx_sched_in(ctx, cpuctx, event_type, task);
2481 2482
}

S
Stephane Eranian 已提交
2483 2484
static void perf_event_context_sched_in(struct perf_event_context *ctx,
					struct task_struct *task)
2485
{
P
Peter Zijlstra 已提交
2486
	struct perf_cpu_context *cpuctx;
2487

P
Peter Zijlstra 已提交
2488
	cpuctx = __get_cpu_context(ctx);
2489 2490 2491
	if (cpuctx->task_ctx == ctx)
		return;

2492
	perf_ctx_lock(cpuctx, ctx);
P
Peter Zijlstra 已提交
2493
	perf_pmu_disable(ctx->pmu);
2494 2495 2496 2497 2498 2499 2500
	/*
	 * 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);

2501 2502
	if (ctx->nr_events)
		cpuctx->task_ctx = ctx;
2503

2504 2505
	perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);

2506 2507 2508
	perf_pmu_enable(ctx->pmu);
	perf_ctx_unlock(cpuctx, ctx);

2509 2510 2511 2512
	/*
	 * 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
2513
	perf_pmu_rotate_start(ctx->pmu);
2514 2515
}

2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 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
/*
 * 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) {

			pmu = cpuctx->ctx.pmu;

			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
2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586
/*
 * 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.
 */
2587 2588
void __perf_event_task_sched_in(struct task_struct *prev,
				struct task_struct *task)
P
Peter Zijlstra 已提交
2589 2590 2591 2592 2593 2594 2595 2596 2597
{
	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
2598
		perf_event_context_sched_in(ctx, task);
P
Peter Zijlstra 已提交
2599
	}
S
Stephane Eranian 已提交
2600 2601 2602 2603 2604 2605
	/*
	 * 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)))
2606
		perf_cgroup_sched_in(prev, task);
2607 2608 2609 2610

	/* check for system-wide branch_stack events */
	if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
		perf_branch_stack_sched_in(prev, task);
2611 2612
}

2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639
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.
	 */
2640
#define REDUCE_FLS(a, b)		\
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 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679
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;
	}

2680 2681 2682
	if (!divisor)
		return dividend;

2683 2684 2685
	return div64_u64(dividend, divisor);
}

2686 2687 2688
static DEFINE_PER_CPU(int, perf_throttled_count);
static DEFINE_PER_CPU(u64, perf_throttled_seq);

2689
static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2690
{
2691
	struct hw_perf_event *hwc = &event->hw;
2692
	s64 period, sample_period;
2693 2694
	s64 delta;

2695
	period = perf_calculate_period(event, nsec, count);
2696 2697 2698 2699 2700 2701 2702 2703 2704 2705

	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;
2706

2707
	if (local64_read(&hwc->period_left) > 8*sample_period) {
2708 2709 2710
		if (disable)
			event->pmu->stop(event, PERF_EF_UPDATE);

2711
		local64_set(&hwc->period_left, 0);
2712 2713 2714

		if (disable)
			event->pmu->start(event, PERF_EF_RELOAD);
2715
	}
2716 2717
}

2718 2719 2720 2721 2722 2723 2724
/*
 * 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)
2725
{
2726 2727
	struct perf_event *event;
	struct hw_perf_event *hwc;
2728
	u64 now, period = TICK_NSEC;
2729
	s64 delta;
2730

2731 2732 2733 2734 2735 2736
	/*
	 * 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))
2737 2738
		return;

2739
	raw_spin_lock(&ctx->lock);
2740
	perf_pmu_disable(ctx->pmu);
2741

2742
	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2743
		if (event->state != PERF_EVENT_STATE_ACTIVE)
2744 2745
			continue;

2746
		if (!event_filter_match(event))
2747 2748
			continue;

2749
		hwc = &event->hw;
2750

2751
		if (hwc->interrupts == MAX_INTERRUPTS) {
2752
			hwc->interrupts = 0;
2753
			perf_log_throttle(event, 1);
P
Peter Zijlstra 已提交
2754
			event->pmu->start(event, 0);
2755 2756
		}

2757
		if (!event->attr.freq || !event->attr.sample_freq)
2758 2759
			continue;

2760 2761 2762 2763 2764
		/*
		 * stop the event and update event->count
		 */
		event->pmu->stop(event, PERF_EF_UPDATE);

2765
		now = local64_read(&event->count);
2766 2767
		delta = now - hwc->freq_count_stamp;
		hwc->freq_count_stamp = now;
2768

2769 2770 2771
		/*
		 * restart the event
		 * reload only if value has changed
2772 2773 2774
		 * we have stopped the event so tell that
		 * to perf_adjust_period() to avoid stopping it
		 * twice.
2775
		 */
2776
		if (delta > 0)
2777
			perf_adjust_period(event, period, delta, false);
2778 2779

		event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2780
	}
2781

2782
	perf_pmu_enable(ctx->pmu);
2783
	raw_spin_unlock(&ctx->lock);
2784 2785
}

2786
/*
2787
 * Round-robin a context's events:
2788
 */
2789
static void rotate_ctx(struct perf_event_context *ctx)
T
Thomas Gleixner 已提交
2790
{
2791 2792 2793 2794 2795 2796
	/*
	 * 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);
2797 2798
}

2799
/*
2800 2801 2802
 * 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.
2803
 */
2804
static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2805
{
P
Peter Zijlstra 已提交
2806
	struct perf_event_context *ctx = NULL;
2807
	int rotate = 0, remove = 1;
2808

2809
	if (cpuctx->ctx.nr_events) {
2810
		remove = 0;
2811 2812 2813
		if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
			rotate = 1;
	}
2814

P
Peter Zijlstra 已提交
2815
	ctx = cpuctx->task_ctx;
2816
	if (ctx && ctx->nr_events) {
2817
		remove = 0;
2818 2819 2820
		if (ctx->nr_events != ctx->nr_active)
			rotate = 1;
	}
2821

2822
	if (!rotate)
2823 2824
		goto done;

2825
	perf_ctx_lock(cpuctx, cpuctx->task_ctx);
P
Peter Zijlstra 已提交
2826
	perf_pmu_disable(cpuctx->ctx.pmu);
2827

2828 2829 2830
	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
	if (ctx)
		ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
T
Thomas Gleixner 已提交
2831

2832 2833 2834
	rotate_ctx(&cpuctx->ctx);
	if (ctx)
		rotate_ctx(ctx);
2835

2836
	perf_event_sched_in(cpuctx, ctx, current);
2837

2838 2839
	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2840
done:
2841 2842
	if (remove)
		list_del_init(&cpuctx->rotation_list);
2843 2844

	return rotate;
2845 2846
}

2847 2848 2849
#ifdef CONFIG_NO_HZ_FULL
bool perf_event_can_stop_tick(void)
{
2850
	if (atomic_read(&nr_freq_events) ||
2851
	    __this_cpu_read(perf_throttled_count))
2852
		return false;
2853 2854
	else
		return true;
2855 2856 2857
}
#endif

2858 2859 2860 2861
void perf_event_task_tick(void)
{
	struct list_head *head = &__get_cpu_var(rotation_list);
	struct perf_cpu_context *cpuctx, *tmp;
2862 2863
	struct perf_event_context *ctx;
	int throttled;
2864

2865 2866
	WARN_ON(!irqs_disabled());

2867 2868 2869
	__this_cpu_inc(perf_throttled_seq);
	throttled = __this_cpu_xchg(perf_throttled_count, 0);

2870
	list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2871 2872 2873 2874 2875 2876
		ctx = &cpuctx->ctx;
		perf_adjust_freq_unthr_context(ctx, throttled);

		ctx = cpuctx->task_ctx;
		if (ctx)
			perf_adjust_freq_unthr_context(ctx, throttled);
2877
	}
T
Thomas Gleixner 已提交
2878 2879
}

2880 2881 2882 2883 2884 2885 2886 2887 2888 2889
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;

2890
	__perf_event_mark_enabled(event);
2891 2892 2893 2894

	return 1;
}

2895
/*
2896
 * Enable all of a task's events that have been marked enable-on-exec.
2897 2898
 * This expects task == current.
 */
P
Peter Zijlstra 已提交
2899
static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2900
{
2901
	struct perf_event *event;
2902 2903
	unsigned long flags;
	int enabled = 0;
2904
	int ret;
2905 2906

	local_irq_save(flags);
2907
	if (!ctx || !ctx->nr_events)
2908 2909
		goto out;

2910 2911 2912 2913 2914 2915 2916
	/*
	 * 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.
	 */
2917
	perf_cgroup_sched_out(current, NULL);
2918

2919
	raw_spin_lock(&ctx->lock);
2920
	task_ctx_sched_out(ctx);
2921

2922
	list_for_each_entry(event, &ctx->event_list, event_entry) {
2923 2924 2925
		ret = event_enable_on_exec(event, ctx);
		if (ret)
			enabled = 1;
2926 2927 2928
	}

	/*
2929
	 * Unclone this context if we enabled any event.
2930
	 */
2931 2932
	if (enabled)
		unclone_ctx(ctx);
2933

2934
	raw_spin_unlock(&ctx->lock);
2935

2936 2937 2938
	/*
	 * Also calls ctxswin for cgroup events, if any:
	 */
S
Stephane Eranian 已提交
2939
	perf_event_context_sched_in(ctx, ctx->task);
P
Peter Zijlstra 已提交
2940
out:
2941 2942 2943
	local_irq_restore(flags);
}

T
Thomas Gleixner 已提交
2944
/*
2945
 * Cross CPU call to read the hardware event
T
Thomas Gleixner 已提交
2946
 */
2947
static void __perf_event_read(void *info)
T
Thomas Gleixner 已提交
2948
{
2949 2950
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
2951
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
I
Ingo Molnar 已提交
2952

2953 2954 2955 2956
	/*
	 * 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
2957 2958
	 * event->count would have been updated to a recent sample
	 * when the event was scheduled out.
2959 2960 2961 2962
	 */
	if (ctx->task && cpuctx->task_ctx != ctx)
		return;

2963
	raw_spin_lock(&ctx->lock);
S
Stephane Eranian 已提交
2964
	if (ctx->is_active) {
2965
		update_context_time(ctx);
S
Stephane Eranian 已提交
2966 2967
		update_cgrp_time_from_event(event);
	}
2968
	update_event_times(event);
2969 2970
	if (event->state == PERF_EVENT_STATE_ACTIVE)
		event->pmu->read(event);
2971
	raw_spin_unlock(&ctx->lock);
T
Thomas Gleixner 已提交
2972 2973
}

P
Peter Zijlstra 已提交
2974 2975
static inline u64 perf_event_count(struct perf_event *event)
{
2976
	return local64_read(&event->count) + atomic64_read(&event->child_count);
P
Peter Zijlstra 已提交
2977 2978
}

2979
static u64 perf_event_read(struct perf_event *event)
T
Thomas Gleixner 已提交
2980 2981
{
	/*
2982 2983
	 * If event is enabled and currently active on a CPU, update the
	 * value in the event structure:
T
Thomas Gleixner 已提交
2984
	 */
2985 2986 2987 2988
	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
2989 2990 2991
		struct perf_event_context *ctx = event->ctx;
		unsigned long flags;

2992
		raw_spin_lock_irqsave(&ctx->lock, flags);
2993 2994 2995 2996 2997
		/*
		 * 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
2998
		if (ctx->is_active) {
2999
			update_context_time(ctx);
S
Stephane Eranian 已提交
3000 3001
			update_cgrp_time_from_event(event);
		}
3002
		update_event_times(event);
3003
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
T
Thomas Gleixner 已提交
3004 3005
	}

P
Peter Zijlstra 已提交
3006
	return perf_event_count(event);
T
Thomas Gleixner 已提交
3007 3008
}

3009
/*
3010
 * Initialize the perf_event context in a task_struct:
3011
 */
3012
static void __perf_event_init_context(struct perf_event_context *ctx)
3013
{
3014
	raw_spin_lock_init(&ctx->lock);
3015
	mutex_init(&ctx->mutex);
3016 3017
	INIT_LIST_HEAD(&ctx->pinned_groups);
	INIT_LIST_HEAD(&ctx->flexible_groups);
3018 3019
	INIT_LIST_HEAD(&ctx->event_list);
	atomic_set(&ctx->refcount, 1);
3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034
}

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
3035
	}
3036 3037 3038
	ctx->pmu = pmu;

	return ctx;
3039 3040
}

3041 3042 3043 3044 3045
static struct task_struct *
find_lively_task_by_vpid(pid_t vpid)
{
	struct task_struct *task;
	int err;
T
Thomas Gleixner 已提交
3046 3047

	rcu_read_lock();
3048
	if (!vpid)
T
Thomas Gleixner 已提交
3049 3050
		task = current;
	else
3051
		task = find_task_by_vpid(vpid);
T
Thomas Gleixner 已提交
3052 3053 3054 3055 3056 3057 3058 3059
	if (task)
		get_task_struct(task);
	rcu_read_unlock();

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

	/* Reuse ptrace permission checks for now. */
3060 3061 3062 3063
	err = -EACCES;
	if (!ptrace_may_access(task, PTRACE_MODE_READ))
		goto errout;

3064 3065 3066 3067 3068 3069 3070
	return task;
errout:
	put_task_struct(task);
	return ERR_PTR(err);

}

3071 3072 3073
/*
 * Returns a matching context with refcount and pincount.
 */
P
Peter Zijlstra 已提交
3074
static struct perf_event_context *
M
Matt Helsley 已提交
3075
find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
T
Thomas Gleixner 已提交
3076
{
3077
	struct perf_event_context *ctx;
3078
	struct perf_cpu_context *cpuctx;
3079
	unsigned long flags;
P
Peter Zijlstra 已提交
3080
	int ctxn, err;
T
Thomas Gleixner 已提交
3081

3082
	if (!task) {
3083
		/* Must be root to operate on a CPU event: */
3084
		if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
T
Thomas Gleixner 已提交
3085 3086 3087
			return ERR_PTR(-EACCES);

		/*
3088
		 * We could be clever and allow to attach a event to an
T
Thomas Gleixner 已提交
3089 3090 3091
		 * offline CPU and activate it when the CPU comes up, but
		 * that's for later.
		 */
3092
		if (!cpu_online(cpu))
T
Thomas Gleixner 已提交
3093 3094
			return ERR_PTR(-ENODEV);

P
Peter Zijlstra 已提交
3095
		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
T
Thomas Gleixner 已提交
3096
		ctx = &cpuctx->ctx;
3097
		get_ctx(ctx);
3098
		++ctx->pin_count;
T
Thomas Gleixner 已提交
3099 3100 3101 3102

		return ctx;
	}

P
Peter Zijlstra 已提交
3103 3104 3105 3106 3107
	err = -EINVAL;
	ctxn = pmu->task_ctx_nr;
	if (ctxn < 0)
		goto errout;

P
Peter Zijlstra 已提交
3108
retry:
P
Peter Zijlstra 已提交
3109
	ctx = perf_lock_task_context(task, ctxn, &flags);
3110
	if (ctx) {
3111
		unclone_ctx(ctx);
3112
		++ctx->pin_count;
3113
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
3114
	} else {
3115
		ctx = alloc_perf_context(pmu, task);
3116 3117 3118
		err = -ENOMEM;
		if (!ctx)
			goto errout;
3119

3120 3121 3122 3123 3124 3125 3126 3127 3128 3129
		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;
3130
		else {
3131
			get_ctx(ctx);
3132
			++ctx->pin_count;
3133
			rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3134
		}
3135 3136 3137
		mutex_unlock(&task->perf_event_mutex);

		if (unlikely(err)) {
3138
			put_ctx(ctx);
3139 3140 3141 3142

			if (err == -EAGAIN)
				goto retry;
			goto errout;
3143 3144 3145
		}
	}

T
Thomas Gleixner 已提交
3146
	return ctx;
3147

P
Peter Zijlstra 已提交
3148
errout:
3149
	return ERR_PTR(err);
T
Thomas Gleixner 已提交
3150 3151
}

L
Li Zefan 已提交
3152 3153
static void perf_event_free_filter(struct perf_event *event);

3154
static void free_event_rcu(struct rcu_head *head)
P
Peter Zijlstra 已提交
3155
{
3156
	struct perf_event *event;
P
Peter Zijlstra 已提交
3157

3158 3159 3160
	event = container_of(head, struct perf_event, rcu_head);
	if (event->ns)
		put_pid_ns(event->ns);
L
Li Zefan 已提交
3161
	perf_event_free_filter(event);
3162
	kfree(event);
P
Peter Zijlstra 已提交
3163 3164
}

3165
static void ring_buffer_put(struct ring_buffer *rb);
3166
static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3167

3168
static void unaccount_event_cpu(struct perf_event *event, int cpu)
3169
{
3170 3171 3172 3173 3174 3175 3176 3177 3178 3179
	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));
}
3180

3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193
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);
3194 3195
	if (event->attr.freq)
		atomic_dec(&nr_freq_events);
3196 3197 3198 3199 3200 3201 3202
	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);
}
3203

3204 3205
static void __free_event(struct perf_event *event)
{
3206
	if (!event->parent) {
3207 3208
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
			put_callchain_buffers();
3209
	}
3210

3211 3212 3213 3214 3215 3216 3217 3218
	if (event->destroy)
		event->destroy(event);

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

	call_rcu(&event->rcu_head, free_event_rcu);
}
3219
static void free_event(struct perf_event *event)
3220
{
3221
	irq_work_sync(&event->pending);
3222

3223
	unaccount_event(event);
3224

3225
	if (event->rb) {
3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241
		struct ring_buffer *rb;

		/*
		 * 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);
		rb = event->rb;
		if (rb) {
			rcu_assign_pointer(event->rb, NULL);
			ring_buffer_detach(event, rb);
			ring_buffer_put(rb); /* could be last */
		}
		mutex_unlock(&event->mmap_mutex);
3242 3243
	}

S
Stephane Eranian 已提交
3244 3245 3246
	if (is_cgroup_event(event))
		perf_detach_cgroup(event);

P
Peter Zijlstra 已提交
3247

3248
	__free_event(event);
3249 3250
}

3251
int perf_event_release_kernel(struct perf_event *event)
T
Thomas Gleixner 已提交
3252
{
3253
	struct perf_event_context *ctx = event->ctx;
T
Thomas Gleixner 已提交
3254

3255
	WARN_ON_ONCE(ctx->parent_ctx);
3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268
	/*
	 * 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);
3269
	raw_spin_lock_irq(&ctx->lock);
3270
	perf_group_detach(event);
3271
	raw_spin_unlock_irq(&ctx->lock);
3272
	perf_remove_from_context(event);
3273
	mutex_unlock(&ctx->mutex);
T
Thomas Gleixner 已提交
3274

3275
	free_event(event);
T
Thomas Gleixner 已提交
3276 3277 3278

	return 0;
}
3279
EXPORT_SYMBOL_GPL(perf_event_release_kernel);
T
Thomas Gleixner 已提交
3280

3281 3282 3283
/*
 * Called when the last reference to the file is gone.
 */
3284
static void put_event(struct perf_event *event)
3285
{
P
Peter Zijlstra 已提交
3286
	struct task_struct *owner;
3287

3288 3289
	if (!atomic_long_dec_and_test(&event->refcount))
		return;
3290

P
Peter Zijlstra 已提交
3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323
	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);
	}

3324 3325 3326 3327 3328 3329 3330
	perf_event_release_kernel(event);
}

static int perf_release(struct inode *inode, struct file *file)
{
	put_event(file->private_data);
	return 0;
3331 3332
}

3333
u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3334
{
3335
	struct perf_event *child;
3336 3337
	u64 total = 0;

3338 3339 3340
	*enabled = 0;
	*running = 0;

3341
	mutex_lock(&event->child_mutex);
3342
	total += perf_event_read(event);
3343 3344 3345 3346 3347 3348
	*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) {
3349
		total += perf_event_read(child);
3350 3351 3352
		*enabled += child->total_time_enabled;
		*running += child->total_time_running;
	}
3353
	mutex_unlock(&event->child_mutex);
3354 3355 3356

	return total;
}
3357
EXPORT_SYMBOL_GPL(perf_event_read_value);
3358

3359
static int perf_event_read_group(struct perf_event *event,
3360 3361
				   u64 read_format, char __user *buf)
{
3362
	struct perf_event *leader = event->group_leader, *sub;
3363 3364
	int n = 0, size = 0, ret = -EFAULT;
	struct perf_event_context *ctx = leader->ctx;
3365
	u64 values[5];
3366
	u64 count, enabled, running;
3367

3368
	mutex_lock(&ctx->mutex);
3369
	count = perf_event_read_value(leader, &enabled, &running);
3370 3371

	values[n++] = 1 + leader->nr_siblings;
3372 3373 3374 3375
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
		values[n++] = enabled;
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
		values[n++] = running;
3376 3377 3378
	values[n++] = count;
	if (read_format & PERF_FORMAT_ID)
		values[n++] = primary_event_id(leader);
3379 3380 3381 3382

	size = n * sizeof(u64);

	if (copy_to_user(buf, values, size))
3383
		goto unlock;
3384

3385
	ret = size;
3386

3387
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3388
		n = 0;
3389

3390
		values[n++] = perf_event_read_value(sub, &enabled, &running);
3391 3392 3393 3394 3395
		if (read_format & PERF_FORMAT_ID)
			values[n++] = primary_event_id(sub);

		size = n * sizeof(u64);

3396
		if (copy_to_user(buf + ret, values, size)) {
3397 3398 3399
			ret = -EFAULT;
			goto unlock;
		}
3400 3401

		ret += size;
3402
	}
3403 3404
unlock:
	mutex_unlock(&ctx->mutex);
3405

3406
	return ret;
3407 3408
}

3409
static int perf_event_read_one(struct perf_event *event,
3410 3411
				 u64 read_format, char __user *buf)
{
3412
	u64 enabled, running;
3413 3414 3415
	u64 values[4];
	int n = 0;

3416 3417 3418 3419 3420
	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;
3421
	if (read_format & PERF_FORMAT_ID)
3422
		values[n++] = primary_event_id(event);
3423 3424 3425 3426 3427 3428 3429

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

	return n * sizeof(u64);
}

T
Thomas Gleixner 已提交
3430
/*
3431
 * Read the performance event - simple non blocking version for now
T
Thomas Gleixner 已提交
3432 3433
 */
static ssize_t
3434
perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
T
Thomas Gleixner 已提交
3435
{
3436
	u64 read_format = event->attr.read_format;
3437
	int ret;
T
Thomas Gleixner 已提交
3438

3439
	/*
3440
	 * Return end-of-file for a read on a event that is in
3441 3442 3443
	 * error state (i.e. because it was pinned but it couldn't be
	 * scheduled on to the CPU at some point).
	 */
3444
	if (event->state == PERF_EVENT_STATE_ERROR)
3445 3446
		return 0;

3447
	if (count < event->read_size)
3448 3449
		return -ENOSPC;

3450
	WARN_ON_ONCE(event->ctx->parent_ctx);
3451
	if (read_format & PERF_FORMAT_GROUP)
3452
		ret = perf_event_read_group(event, read_format, buf);
3453
	else
3454
		ret = perf_event_read_one(event, read_format, buf);
T
Thomas Gleixner 已提交
3455

3456
	return ret;
T
Thomas Gleixner 已提交
3457 3458 3459 3460 3461
}

static ssize_t
perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
3462
	struct perf_event *event = file->private_data;
T
Thomas Gleixner 已提交
3463

3464
	return perf_read_hw(event, buf, count);
T
Thomas Gleixner 已提交
3465 3466 3467 3468
}

static unsigned int perf_poll(struct file *file, poll_table *wait)
{
3469
	struct perf_event *event = file->private_data;
3470
	struct ring_buffer *rb;
3471
	unsigned int events = POLL_HUP;
P
Peter Zijlstra 已提交
3472

3473
	/*
3474 3475
	 * Pin the event->rb by taking event->mmap_mutex; otherwise
	 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3476 3477
	 */
	mutex_lock(&event->mmap_mutex);
3478 3479
	rb = event->rb;
	if (rb)
3480
		events = atomic_xchg(&rb->poll, 0);
3481 3482
	mutex_unlock(&event->mmap_mutex);

3483
	poll_wait(file, &event->waitq, wait);
T
Thomas Gleixner 已提交
3484 3485 3486 3487

	return events;
}

3488
static void perf_event_reset(struct perf_event *event)
3489
{
3490
	(void)perf_event_read(event);
3491
	local64_set(&event->count, 0);
3492
	perf_event_update_userpage(event);
P
Peter Zijlstra 已提交
3493 3494
}

3495
/*
3496 3497 3498 3499
 * 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.
3500
 */
3501 3502
static void perf_event_for_each_child(struct perf_event *event,
					void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3503
{
3504
	struct perf_event *child;
P
Peter Zijlstra 已提交
3505

3506 3507 3508 3509
	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
3510
		func(child);
3511
	mutex_unlock(&event->child_mutex);
P
Peter Zijlstra 已提交
3512 3513
}

3514 3515
static void perf_event_for_each(struct perf_event *event,
				  void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3516
{
3517 3518
	struct perf_event_context *ctx = event->ctx;
	struct perf_event *sibling;
P
Peter Zijlstra 已提交
3519

3520 3521
	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
3522
	event = event->group_leader;
3523

3524 3525
	perf_event_for_each_child(event, func);
	list_for_each_entry(sibling, &event->sibling_list, group_entry)
3526
		perf_event_for_each_child(sibling, func);
3527
	mutex_unlock(&ctx->mutex);
3528 3529
}

3530
static int perf_event_period(struct perf_event *event, u64 __user *arg)
3531
{
3532
	struct perf_event_context *ctx = event->ctx;
3533 3534 3535
	int ret = 0;
	u64 value;

3536
	if (!is_sampling_event(event))
3537 3538
		return -EINVAL;

3539
	if (copy_from_user(&value, arg, sizeof(value)))
3540 3541 3542 3543 3544
		return -EFAULT;

	if (!value)
		return -EINVAL;

3545
	raw_spin_lock_irq(&ctx->lock);
3546 3547
	if (event->attr.freq) {
		if (value > sysctl_perf_event_sample_rate) {
3548 3549 3550 3551
			ret = -EINVAL;
			goto unlock;
		}

3552
		event->attr.sample_freq = value;
3553
	} else {
3554 3555
		event->attr.sample_period = value;
		event->hw.sample_period = value;
3556 3557
	}
unlock:
3558
	raw_spin_unlock_irq(&ctx->lock);
3559 3560 3561 3562

	return ret;
}

3563 3564
static const struct file_operations perf_fops;

3565
static inline int perf_fget_light(int fd, struct fd *p)
3566
{
3567 3568 3569
	struct fd f = fdget(fd);
	if (!f.file)
		return -EBADF;
3570

3571 3572 3573
	if (f.file->f_op != &perf_fops) {
		fdput(f);
		return -EBADF;
3574
	}
3575 3576
	*p = f;
	return 0;
3577 3578 3579 3580
}

static int perf_event_set_output(struct perf_event *event,
				 struct perf_event *output_event);
L
Li Zefan 已提交
3581
static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3582

3583 3584
static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
3585 3586
	struct perf_event *event = file->private_data;
	void (*func)(struct perf_event *);
P
Peter Zijlstra 已提交
3587
	u32 flags = arg;
3588 3589

	switch (cmd) {
3590 3591
	case PERF_EVENT_IOC_ENABLE:
		func = perf_event_enable;
3592
		break;
3593 3594
	case PERF_EVENT_IOC_DISABLE:
		func = perf_event_disable;
3595
		break;
3596 3597
	case PERF_EVENT_IOC_RESET:
		func = perf_event_reset;
3598
		break;
P
Peter Zijlstra 已提交
3599

3600 3601
	case PERF_EVENT_IOC_REFRESH:
		return perf_event_refresh(event, arg);
3602

3603 3604
	case PERF_EVENT_IOC_PERIOD:
		return perf_event_period(event, (u64 __user *)arg);
3605

3606 3607 3608 3609 3610 3611 3612 3613 3614
	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;
	}

3615
	case PERF_EVENT_IOC_SET_OUTPUT:
3616 3617 3618
	{
		int ret;
		if (arg != -1) {
3619 3620 3621 3622 3623 3624 3625 3626 3627 3628
			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);
3629 3630 3631
		}
		return ret;
	}
3632

L
Li Zefan 已提交
3633 3634 3635
	case PERF_EVENT_IOC_SET_FILTER:
		return perf_event_set_filter(event, (void __user *)arg);

3636
	default:
P
Peter Zijlstra 已提交
3637
		return -ENOTTY;
3638
	}
P
Peter Zijlstra 已提交
3639 3640

	if (flags & PERF_IOC_FLAG_GROUP)
3641
		perf_event_for_each(event, func);
P
Peter Zijlstra 已提交
3642
	else
3643
		perf_event_for_each_child(event, func);
P
Peter Zijlstra 已提交
3644 3645

	return 0;
3646 3647
}

3648
int perf_event_task_enable(void)
3649
{
3650
	struct perf_event *event;
3651

3652 3653 3654 3655
	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);
3656 3657 3658 3659

	return 0;
}

3660
int perf_event_task_disable(void)
3661
{
3662
	struct perf_event *event;
3663

3664 3665 3666 3667
	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);
3668 3669 3670 3671

	return 0;
}

3672
static int perf_event_index(struct perf_event *event)
3673
{
P
Peter Zijlstra 已提交
3674 3675 3676
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;

3677
	if (event->state != PERF_EVENT_STATE_ACTIVE)
3678 3679
		return 0;

3680
	return event->pmu->event_idx(event);
3681 3682
}

3683
static void calc_timer_values(struct perf_event *event,
3684
				u64 *now,
3685 3686
				u64 *enabled,
				u64 *running)
3687
{
3688
	u64 ctx_time;
3689

3690 3691
	*now = perf_clock();
	ctx_time = event->shadow_ctx_time + *now;
3692 3693 3694 3695
	*enabled = ctx_time - event->tstamp_enabled;
	*running = ctx_time - event->tstamp_running;
}

3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715
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();
}

3716
void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3717 3718 3719
{
}

3720 3721 3722 3723 3724
/*
 * 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.
 */
3725
void perf_event_update_userpage(struct perf_event *event)
3726
{
3727
	struct perf_event_mmap_page *userpg;
3728
	struct ring_buffer *rb;
3729
	u64 enabled, running, now;
3730 3731

	rcu_read_lock();
3732 3733 3734 3735
	rb = rcu_dereference(event->rb);
	if (!rb)
		goto unlock;

3736 3737 3738 3739 3740 3741 3742 3743 3744
	/*
	 * 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
	 */
3745
	calc_timer_values(event, &now, &enabled, &running);
3746

3747
	userpg = rb->user_page;
3748 3749 3750 3751 3752
	/*
	 * Disable preemption so as to not let the corresponding user-space
	 * spin too long if we get preempted.
	 */
	preempt_disable();
3753
	++userpg->lock;
3754
	barrier();
3755
	userpg->index = perf_event_index(event);
P
Peter Zijlstra 已提交
3756
	userpg->offset = perf_event_count(event);
3757
	if (userpg->index)
3758
		userpg->offset -= local64_read(&event->hw.prev_count);
3759

3760
	userpg->time_enabled = enabled +
3761
			atomic64_read(&event->child_total_time_enabled);
3762

3763
	userpg->time_running = running +
3764
			atomic64_read(&event->child_total_time_running);
3765

3766
	arch_perf_update_userpage(userpg, now);
3767

3768
	barrier();
3769
	++userpg->lock;
3770
	preempt_enable();
3771
unlock:
3772
	rcu_read_unlock();
3773 3774
}

3775 3776 3777
static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
{
	struct perf_event *event = vma->vm_file->private_data;
3778
	struct ring_buffer *rb;
3779 3780 3781 3782 3783 3784 3785 3786 3787
	int ret = VM_FAULT_SIGBUS;

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

	rcu_read_lock();
3788 3789
	rb = rcu_dereference(event->rb);
	if (!rb)
3790 3791 3792 3793 3794
		goto unlock;

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

3795
	vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809
	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;
}

3810 3811 3812 3813 3814 3815 3816 3817 3818
static void ring_buffer_attach(struct perf_event *event,
			       struct ring_buffer *rb)
{
	unsigned long flags;

	if (!list_empty(&event->rb_entry))
		return;

	spin_lock_irqsave(&rb->event_lock, flags);
3819 3820
	if (list_empty(&event->rb_entry))
		list_add(&event->rb_entry, &rb->event_list);
3821 3822 3823
	spin_unlock_irqrestore(&rb->event_lock, flags);
}

3824
static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842
{
	unsigned long flags;

	if (list_empty(&event->rb_entry))
		return;

	spin_lock_irqsave(&rb->event_lock, flags);
	list_del_init(&event->rb_entry);
	wake_up_all(&event->waitq);
	spin_unlock_irqrestore(&rb->event_lock, flags);
}

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

	rcu_read_lock();
	rb = rcu_dereference(event->rb);
3843 3844 3845 3846
	if (rb) {
		list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
			wake_up_all(&event->waitq);
	}
3847 3848 3849
	rcu_read_unlock();
}

3850
static void rb_free_rcu(struct rcu_head *rcu_head)
3851
{
3852
	struct ring_buffer *rb;
3853

3854 3855
	rb = container_of(rcu_head, struct ring_buffer, rcu_head);
	rb_free(rb);
3856 3857
}

3858
static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3859
{
3860
	struct ring_buffer *rb;
3861

3862
	rcu_read_lock();
3863 3864 3865 3866
	rb = rcu_dereference(event->rb);
	if (rb) {
		if (!atomic_inc_not_zero(&rb->refcount))
			rb = NULL;
3867 3868 3869
	}
	rcu_read_unlock();

3870
	return rb;
3871 3872
}

3873
static void ring_buffer_put(struct ring_buffer *rb)
3874
{
3875
	if (!atomic_dec_and_test(&rb->refcount))
3876
		return;
3877

3878
	WARN_ON_ONCE(!list_empty(&rb->event_list));
3879

3880
	call_rcu(&rb->rcu_head, rb_free_rcu);
3881 3882 3883 3884
}

static void perf_mmap_open(struct vm_area_struct *vma)
{
3885
	struct perf_event *event = vma->vm_file->private_data;
3886

3887
	atomic_inc(&event->mmap_count);
3888
	atomic_inc(&event->rb->mmap_count);
3889 3890
}

3891 3892 3893 3894 3895 3896 3897 3898
/*
 * 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.
 */
3899 3900
static void perf_mmap_close(struct vm_area_struct *vma)
{
3901
	struct perf_event *event = vma->vm_file->private_data;
3902

3903 3904 3905 3906
	struct ring_buffer *rb = event->rb;
	struct user_struct *mmap_user = rb->mmap_user;
	int mmap_locked = rb->mmap_locked;
	unsigned long size = perf_data_size(rb);
3907

3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922
	atomic_dec(&rb->mmap_count);

	if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
		return;

	/* Detach current event from the buffer. */
	rcu_assign_pointer(event->rb, NULL);
	ring_buffer_detach(event, rb);
	mutex_unlock(&event->mmap_mutex);

	/* If there's still other mmap()s of this buffer, we're done. */
	if (atomic_read(&rb->mmap_count)) {
		ring_buffer_put(rb); /* can't be last */
		return;
	}
3923

3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939
	/*
	 * 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();
3940

3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955
		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.
		 */
		if (event->rb == rb) {
			rcu_assign_pointer(event->rb, NULL);
			ring_buffer_detach(event, rb);
			ring_buffer_put(rb); /* can't be last, we still have one */
P
Peter Zijlstra 已提交
3956
		}
3957
		mutex_unlock(&event->mmap_mutex);
3958
		put_event(event);
3959

3960 3961 3962 3963 3964
		/*
		 * Restart the iteration; either we're on the wrong list or
		 * destroyed its integrity by doing a deletion.
		 */
		goto again;
3965
	}
3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981
	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);

	ring_buffer_put(rb); /* could be last */
3982 3983
}

3984
static const struct vm_operations_struct perf_mmap_vmops = {
3985 3986 3987 3988
	.open		= perf_mmap_open,
	.close		= perf_mmap_close,
	.fault		= perf_mmap_fault,
	.page_mkwrite	= perf_mmap_fault,
3989 3990 3991 3992
};

static int perf_mmap(struct file *file, struct vm_area_struct *vma)
{
3993
	struct perf_event *event = file->private_data;
3994
	unsigned long user_locked, user_lock_limit;
3995
	struct user_struct *user = current_user();
3996
	unsigned long locked, lock_limit;
3997
	struct ring_buffer *rb;
3998 3999
	unsigned long vma_size;
	unsigned long nr_pages;
4000
	long user_extra, extra;
4001
	int ret = 0, flags = 0;
4002

4003 4004 4005
	/*
	 * Don't allow mmap() of inherited per-task counters. This would
	 * create a performance issue due to all children writing to the
4006
	 * same rb.
4007 4008 4009 4010
	 */
	if (event->cpu == -1 && event->attr.inherit)
		return -EINVAL;

4011
	if (!(vma->vm_flags & VM_SHARED))
4012
		return -EINVAL;
4013 4014 4015 4016

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

4017
	/*
4018
	 * If we have rb pages ensure they're a power-of-two number, so we
4019 4020 4021
	 * can do bitmasks instead of modulo.
	 */
	if (nr_pages != 0 && !is_power_of_2(nr_pages))
4022 4023
		return -EINVAL;

4024
	if (vma_size != PAGE_SIZE * (1 + nr_pages))
4025 4026
		return -EINVAL;

4027 4028
	if (vma->vm_pgoff != 0)
		return -EINVAL;
4029

4030
	WARN_ON_ONCE(event->ctx->parent_ctx);
4031
again:
4032
	mutex_lock(&event->mmap_mutex);
4033
	if (event->rb) {
4034
		if (event->rb->nr_pages != nr_pages) {
4035
			ret = -EINVAL;
4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048
			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;
		}

4049 4050 4051
		goto unlock;
	}

4052
	user_extra = nr_pages + 1;
4053
	user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
I
Ingo Molnar 已提交
4054 4055 4056 4057 4058 4059

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

4060
	user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4061

4062 4063 4064
	extra = 0;
	if (user_locked > user_lock_limit)
		extra = user_locked - user_lock_limit;
4065

4066
	lock_limit = rlimit(RLIMIT_MEMLOCK);
4067
	lock_limit >>= PAGE_SHIFT;
4068
	locked = vma->vm_mm->pinned_vm + extra;
4069

4070 4071
	if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
		!capable(CAP_IPC_LOCK)) {
4072 4073 4074
		ret = -EPERM;
		goto unlock;
	}
4075

4076
	WARN_ON(event->rb);
4077

4078
	if (vma->vm_flags & VM_WRITE)
4079
		flags |= RING_BUFFER_WRITABLE;
4080

4081 4082 4083 4084
	rb = rb_alloc(nr_pages, 
		event->attr.watermark ? event->attr.wakeup_watermark : 0,
		event->cpu, flags);

4085
	if (!rb) {
4086
		ret = -ENOMEM;
4087
		goto unlock;
4088
	}
P
Peter Zijlstra 已提交
4089

4090
	atomic_set(&rb->mmap_count, 1);
P
Peter Zijlstra 已提交
4091 4092
	rb->mmap_locked = extra;
	rb->mmap_user = get_current_user();
4093

4094
	atomic_long_add(user_extra, &user->locked_vm);
P
Peter Zijlstra 已提交
4095 4096
	vma->vm_mm->pinned_vm += extra;

4097
	ring_buffer_attach(event, rb);
P
Peter Zijlstra 已提交
4098
	rcu_assign_pointer(event->rb, rb);
4099

4100
	perf_event_init_userpage(event);
4101 4102
	perf_event_update_userpage(event);

4103
unlock:
4104 4105
	if (!ret)
		atomic_inc(&event->mmap_count);
4106
	mutex_unlock(&event->mmap_mutex);
4107

4108 4109 4110 4111
	/*
	 * Since pinned accounting is per vm we cannot allow fork() to copy our
	 * vma.
	 */
P
Peter Zijlstra 已提交
4112
	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4113
	vma->vm_ops = &perf_mmap_vmops;
4114 4115

	return ret;
4116 4117
}

P
Peter Zijlstra 已提交
4118 4119
static int perf_fasync(int fd, struct file *filp, int on)
{
A
Al Viro 已提交
4120
	struct inode *inode = file_inode(filp);
4121
	struct perf_event *event = filp->private_data;
P
Peter Zijlstra 已提交
4122 4123 4124
	int retval;

	mutex_lock(&inode->i_mutex);
4125
	retval = fasync_helper(fd, filp, on, &event->fasync);
P
Peter Zijlstra 已提交
4126 4127 4128 4129 4130 4131 4132 4133
	mutex_unlock(&inode->i_mutex);

	if (retval < 0)
		return retval;

	return 0;
}

T
Thomas Gleixner 已提交
4134
static const struct file_operations perf_fops = {
4135
	.llseek			= no_llseek,
T
Thomas Gleixner 已提交
4136 4137 4138
	.release		= perf_release,
	.read			= perf_read,
	.poll			= perf_poll,
4139 4140
	.unlocked_ioctl		= perf_ioctl,
	.compat_ioctl		= perf_ioctl,
4141
	.mmap			= perf_mmap,
P
Peter Zijlstra 已提交
4142
	.fasync			= perf_fasync,
T
Thomas Gleixner 已提交
4143 4144
};

4145
/*
4146
 * Perf event wakeup
4147 4148 4149 4150 4151
 *
 * If there's data, ensure we set the poll() state and publish everything
 * to user-space before waking everybody up.
 */

4152
void perf_event_wakeup(struct perf_event *event)
4153
{
4154
	ring_buffer_wakeup(event);
4155

4156 4157 4158
	if (event->pending_kill) {
		kill_fasync(&event->fasync, SIGIO, event->pending_kill);
		event->pending_kill = 0;
4159
	}
4160 4161
}

4162
static void perf_pending_event(struct irq_work *entry)
4163
{
4164 4165
	struct perf_event *event = container_of(entry,
			struct perf_event, pending);
4166

4167 4168 4169
	if (event->pending_disable) {
		event->pending_disable = 0;
		__perf_event_disable(event);
4170 4171
	}

4172 4173 4174
	if (event->pending_wakeup) {
		event->pending_wakeup = 0;
		perf_event_wakeup(event);
4175 4176 4177
	}
}

4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198
/*
 * 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);

4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229
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);
	}
}

4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 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 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324
/*
 * 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);
	}
}

4325 4326 4327
static void __perf_event_header__init_id(struct perf_event_header *header,
					 struct perf_sample_data *data,
					 struct perf_event *event)
4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342
{
	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();

4343
	if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354
		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;
	}
}

4355 4356 4357
void perf_event_header__init_id(struct perf_event_header *header,
				struct perf_sample_data *data,
				struct perf_event *event)
4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381
{
	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);
4382 4383 4384

	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);
4385 4386
}

4387 4388 4389
void perf_event__output_id_sample(struct perf_event *event,
				  struct perf_output_handle *handle,
				  struct perf_sample_data *sample)
4390 4391 4392 4393 4394
{
	if (event->attr.sample_id_all)
		__perf_event__output_id_sample(handle, sample);
}

4395
static void perf_output_read_one(struct perf_output_handle *handle,
4396 4397
				 struct perf_event *event,
				 u64 enabled, u64 running)
4398
{
4399
	u64 read_format = event->attr.read_format;
4400 4401 4402
	u64 values[4];
	int n = 0;

P
Peter Zijlstra 已提交
4403
	values[n++] = perf_event_count(event);
4404
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4405
		values[n++] = enabled +
4406
			atomic64_read(&event->child_total_time_enabled);
4407 4408
	}
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4409
		values[n++] = running +
4410
			atomic64_read(&event->child_total_time_running);
4411 4412
	}
	if (read_format & PERF_FORMAT_ID)
4413
		values[n++] = primary_event_id(event);
4414

4415
	__output_copy(handle, values, n * sizeof(u64));
4416 4417 4418
}

/*
4419
 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4420 4421
 */
static void perf_output_read_group(struct perf_output_handle *handle,
4422 4423
			    struct perf_event *event,
			    u64 enabled, u64 running)
4424
{
4425 4426
	struct perf_event *leader = event->group_leader, *sub;
	u64 read_format = event->attr.read_format;
4427 4428 4429 4430 4431 4432
	u64 values[5];
	int n = 0;

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

	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4433
		values[n++] = enabled;
4434 4435

	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4436
		values[n++] = running;
4437

4438
	if (leader != event)
4439 4440
		leader->pmu->read(leader);

P
Peter Zijlstra 已提交
4441
	values[n++] = perf_event_count(leader);
4442
	if (read_format & PERF_FORMAT_ID)
4443
		values[n++] = primary_event_id(leader);
4444

4445
	__output_copy(handle, values, n * sizeof(u64));
4446

4447
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4448 4449
		n = 0;

4450 4451
		if ((sub != event) &&
		    (sub->state == PERF_EVENT_STATE_ACTIVE))
4452 4453
			sub->pmu->read(sub);

P
Peter Zijlstra 已提交
4454
		values[n++] = perf_event_count(sub);
4455
		if (read_format & PERF_FORMAT_ID)
4456
			values[n++] = primary_event_id(sub);
4457

4458
		__output_copy(handle, values, n * sizeof(u64));
4459 4460 4461
	}
}

4462 4463 4464
#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
				 PERF_FORMAT_TOTAL_TIME_RUNNING)

4465
static void perf_output_read(struct perf_output_handle *handle,
4466
			     struct perf_event *event)
4467
{
4468
	u64 enabled = 0, running = 0, now;
4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479
	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
	 */
4480
	if (read_format & PERF_FORMAT_TOTAL_TIMES)
4481
		calc_timer_values(event, &now, &enabled, &running);
4482

4483
	if (event->attr.read_format & PERF_FORMAT_GROUP)
4484
		perf_output_read_group(handle, event, enabled, running);
4485
	else
4486
		perf_output_read_one(handle, event, enabled, running);
4487 4488
}

4489 4490 4491
void perf_output_sample(struct perf_output_handle *handle,
			struct perf_event_header *header,
			struct perf_sample_data *data,
4492
			struct perf_event *event)
4493 4494 4495 4496 4497
{
	u64 sample_type = data->type;

	perf_output_put(handle, *header);

4498 4499 4500
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);

4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525
	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)
4526
		perf_output_read(handle, event);
4527 4528 4529 4530 4531 4532 4533 4534 4535 4536

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

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

			size *= sizeof(u64);

4537
			__output_copy(handle, data->callchain, size);
4538 4539 4540 4541 4542 4543 4544 4545 4546
		} 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);
4547 4548
			__output_copy(handle, data->raw->data,
					   data->raw->size);
4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559
		} else {
			struct {
				u32	size;
				u32	data;
			} raw = {
				.size = sizeof(u32),
				.data = 0,
			};
			perf_output_put(handle, raw);
		}
	}
4560

4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577
	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);
		}
	}
4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594

	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);
		}
	}
4595

4596
	if (sample_type & PERF_SAMPLE_STACK_USER) {
4597 4598 4599
		perf_output_sample_ustack(handle,
					  data->stack_user_size,
					  data->regs_user.regs);
4600
	}
A
Andi Kleen 已提交
4601 4602 4603

	if (sample_type & PERF_SAMPLE_WEIGHT)
		perf_output_put(handle, data->weight);
4604 4605 4606

	if (sample_type & PERF_SAMPLE_DATA_SRC)
		perf_output_put(handle, data->data_src.val);
4607

A
Andi Kleen 已提交
4608 4609 4610
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		perf_output_put(handle, data->txn);

4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623
	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);
			}
		}
	}
4624 4625 4626 4627
}

void perf_prepare_sample(struct perf_event_header *header,
			 struct perf_sample_data *data,
4628
			 struct perf_event *event,
4629
			 struct pt_regs *regs)
4630
{
4631
	u64 sample_type = event->attr.sample_type;
4632

4633
	header->type = PERF_RECORD_SAMPLE;
4634
	header->size = sizeof(*header) + event->header_size;
4635 4636 4637

	header->misc = 0;
	header->misc |= perf_misc_flags(regs);
4638

4639
	__perf_event_header__init_id(header, data, event);
4640

4641
	if (sample_type & PERF_SAMPLE_IP)
4642 4643
		data->ip = perf_instruction_pointer(regs);

4644
	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4645
		int size = 1;
4646

4647
		data->callchain = perf_callchain(event, regs);
4648 4649 4650 4651 4652

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

		header->size += size * sizeof(u64);
4653 4654
	}

4655
	if (sample_type & PERF_SAMPLE_RAW) {
4656 4657 4658 4659 4660 4661 4662 4663
		int size = sizeof(u32);

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

		WARN_ON_ONCE(size & (sizeof(u64)-1));
4664
		header->size += size;
4665
	}
4666 4667 4668 4669 4670 4671 4672 4673 4674

	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;
	}
4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688

	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;
	}
4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717

	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;
	}
4718
}
4719

4720
static void perf_event_output(struct perf_event *event,
4721 4722 4723 4724 4725
				struct perf_sample_data *data,
				struct pt_regs *regs)
{
	struct perf_output_handle handle;
	struct perf_event_header header;
4726

4727 4728 4729
	/* protect the callchain buffers */
	rcu_read_lock();

4730
	perf_prepare_sample(&header, data, event, regs);
P
Peter Zijlstra 已提交
4731

4732
	if (perf_output_begin(&handle, event, header.size))
4733
		goto exit;
4734

4735
	perf_output_sample(&handle, &header, data, event);
4736

4737
	perf_output_end(&handle);
4738 4739 4740

exit:
	rcu_read_unlock();
4741 4742
}

4743
/*
4744
 * read event_id
4745 4746 4747 4748 4749 4750 4751 4752 4753 4754
 */

struct perf_read_event {
	struct perf_event_header	header;

	u32				pid;
	u32				tid;
};

static void
4755
perf_event_read_event(struct perf_event *event,
4756 4757 4758
			struct task_struct *task)
{
	struct perf_output_handle handle;
4759
	struct perf_sample_data sample;
4760
	struct perf_read_event read_event = {
4761
		.header = {
4762
			.type = PERF_RECORD_READ,
4763
			.misc = 0,
4764
			.size = sizeof(read_event) + event->read_size,
4765
		},
4766 4767
		.pid = perf_event_pid(event, task),
		.tid = perf_event_tid(event, task),
4768
	};
4769
	int ret;
4770

4771
	perf_event_header__init_id(&read_event.header, &sample, event);
4772
	ret = perf_output_begin(&handle, event, read_event.header.size);
4773 4774 4775
	if (ret)
		return;

4776
	perf_output_put(&handle, read_event);
4777
	perf_output_read(&handle, event);
4778
	perf_event__output_id_sample(event, &handle, &sample);
4779

4780 4781 4782
	perf_output_end(&handle);
}

4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796
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;
4797
		output(event, data);
4798 4799 4800 4801
	}
}

static void
4802
perf_event_aux(perf_event_aux_output_cb output, void *data,
4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814
	       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;
4815
		perf_event_aux_ctx(&cpuctx->ctx, output, data);
4816 4817 4818 4819 4820 4821 4822
		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)
4823
			perf_event_aux_ctx(ctx, output, data);
4824 4825 4826 4827 4828 4829
next:
		put_cpu_ptr(pmu->pmu_cpu_context);
	}

	if (task_ctx) {
		preempt_disable();
4830
		perf_event_aux_ctx(task_ctx, output, data);
4831 4832 4833 4834 4835
		preempt_enable();
	}
	rcu_read_unlock();
}

P
Peter Zijlstra 已提交
4836
/*
P
Peter Zijlstra 已提交
4837 4838
 * task tracking -- fork/exit
 *
4839
 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
P
Peter Zijlstra 已提交
4840 4841
 */

P
Peter Zijlstra 已提交
4842
struct perf_task_event {
4843
	struct task_struct		*task;
4844
	struct perf_event_context	*task_ctx;
P
Peter Zijlstra 已提交
4845 4846 4847 4848 4849 4850

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				ppid;
P
Peter Zijlstra 已提交
4851 4852
		u32				tid;
		u32				ptid;
4853
		u64				time;
4854
	} event_id;
P
Peter Zijlstra 已提交
4855 4856
};

4857 4858
static int perf_event_task_match(struct perf_event *event)
{
4859 4860 4861
	return event->attr.comm  || event->attr.mmap ||
	       event->attr.mmap2 || event->attr.mmap_data ||
	       event->attr.task;
4862 4863
}

4864
static void perf_event_task_output(struct perf_event *event,
4865
				   void *data)
P
Peter Zijlstra 已提交
4866
{
4867
	struct perf_task_event *task_event = data;
P
Peter Zijlstra 已提交
4868
	struct perf_output_handle handle;
4869
	struct perf_sample_data	sample;
P
Peter Zijlstra 已提交
4870
	struct task_struct *task = task_event->task;
4871
	int ret, size = task_event->event_id.header.size;
4872

4873 4874 4875
	if (!perf_event_task_match(event))
		return;

4876
	perf_event_header__init_id(&task_event->event_id.header, &sample, event);
P
Peter Zijlstra 已提交
4877

4878
	ret = perf_output_begin(&handle, event,
4879
				task_event->event_id.header.size);
4880
	if (ret)
4881
		goto out;
P
Peter Zijlstra 已提交
4882

4883 4884
	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
4885

4886 4887
	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
4888

4889
	perf_output_put(&handle, task_event->event_id);
4890

4891 4892
	perf_event__output_id_sample(event, &handle, &sample);

P
Peter Zijlstra 已提交
4893
	perf_output_end(&handle);
4894 4895
out:
	task_event->event_id.header.size = size;
P
Peter Zijlstra 已提交
4896 4897
}

4898 4899
static void perf_event_task(struct task_struct *task,
			      struct perf_event_context *task_ctx,
4900
			      int new)
P
Peter Zijlstra 已提交
4901
{
P
Peter Zijlstra 已提交
4902
	struct perf_task_event task_event;
P
Peter Zijlstra 已提交
4903

4904 4905 4906
	if (!atomic_read(&nr_comm_events) &&
	    !atomic_read(&nr_mmap_events) &&
	    !atomic_read(&nr_task_events))
P
Peter Zijlstra 已提交
4907 4908
		return;

P
Peter Zijlstra 已提交
4909
	task_event = (struct perf_task_event){
4910 4911
		.task	  = task,
		.task_ctx = task_ctx,
4912
		.event_id    = {
P
Peter Zijlstra 已提交
4913
			.header = {
4914
				.type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4915
				.misc = 0,
4916
				.size = sizeof(task_event.event_id),
P
Peter Zijlstra 已提交
4917
			},
4918 4919
			/* .pid  */
			/* .ppid */
P
Peter Zijlstra 已提交
4920 4921
			/* .tid  */
			/* .ptid */
P
Peter Zijlstra 已提交
4922
			.time = perf_clock(),
P
Peter Zijlstra 已提交
4923 4924 4925
		},
	};

4926
	perf_event_aux(perf_event_task_output,
4927 4928
		       &task_event,
		       task_ctx);
P
Peter Zijlstra 已提交
4929 4930
}

4931
void perf_event_fork(struct task_struct *task)
P
Peter Zijlstra 已提交
4932
{
4933
	perf_event_task(task, NULL, 1);
P
Peter Zijlstra 已提交
4934 4935
}

4936 4937 4938 4939 4940
/*
 * comm tracking
 */

struct perf_comm_event {
4941 4942
	struct task_struct	*task;
	char			*comm;
4943 4944 4945 4946 4947 4948 4949
	int			comm_size;

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
4950
	} event_id;
4951 4952
};

4953 4954 4955 4956 4957
static int perf_event_comm_match(struct perf_event *event)
{
	return event->attr.comm;
}

4958
static void perf_event_comm_output(struct perf_event *event,
4959
				   void *data)
4960
{
4961
	struct perf_comm_event *comm_event = data;
4962
	struct perf_output_handle handle;
4963
	struct perf_sample_data sample;
4964
	int size = comm_event->event_id.header.size;
4965 4966
	int ret;

4967 4968 4969
	if (!perf_event_comm_match(event))
		return;

4970 4971
	perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
4972
				comm_event->event_id.header.size);
4973 4974

	if (ret)
4975
		goto out;
4976

4977 4978
	comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
	comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4979

4980
	perf_output_put(&handle, comm_event->event_id);
4981
	__output_copy(&handle, comm_event->comm,
4982
				   comm_event->comm_size);
4983 4984 4985

	perf_event__output_id_sample(event, &handle, &sample);

4986
	perf_output_end(&handle);
4987 4988
out:
	comm_event->event_id.header.size = size;
4989 4990
}

4991
static void perf_event_comm_event(struct perf_comm_event *comm_event)
4992
{
4993
	char comm[TASK_COMM_LEN];
4994 4995
	unsigned int size;

4996
	memset(comm, 0, sizeof(comm));
4997
	strlcpy(comm, comm_event->task->comm, sizeof(comm));
4998
	size = ALIGN(strlen(comm)+1, sizeof(u64));
4999 5000 5001 5002

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

5003
	comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
P
Peter Zijlstra 已提交
5004

5005
	perf_event_aux(perf_event_comm_output,
5006 5007
		       comm_event,
		       NULL);
5008 5009
}

5010
void perf_event_comm(struct task_struct *task)
5011
{
5012
	struct perf_comm_event comm_event;
P
Peter Zijlstra 已提交
5013 5014
	struct perf_event_context *ctx;
	int ctxn;
5015

5016
	rcu_read_lock();
P
Peter Zijlstra 已提交
5017 5018 5019 5020
	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;
5021

P
Peter Zijlstra 已提交
5022 5023
		perf_event_enable_on_exec(ctx);
	}
5024
	rcu_read_unlock();
5025

5026
	if (!atomic_read(&nr_comm_events))
5027
		return;
5028

5029
	comm_event = (struct perf_comm_event){
5030
		.task	= task,
5031 5032
		/* .comm      */
		/* .comm_size */
5033
		.event_id  = {
5034
			.header = {
5035
				.type = PERF_RECORD_COMM,
5036 5037 5038 5039 5040
				.misc = 0,
				/* .size */
			},
			/* .pid */
			/* .tid */
5041 5042 5043
		},
	};

5044
	perf_event_comm_event(&comm_event);
5045 5046
}

5047 5048 5049 5050 5051
/*
 * mmap tracking
 */

struct perf_mmap_event {
5052 5053 5054 5055
	struct vm_area_struct	*vma;

	const char		*file_name;
	int			file_size;
5056 5057 5058
	int			maj, min;
	u64			ino;
	u64			ino_generation;
5059 5060 5061 5062 5063 5064 5065 5066 5067

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
		u64				start;
		u64				len;
		u64				pgoff;
5068
	} event_id;
5069 5070
};

5071 5072 5073 5074 5075 5076 5077 5078
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) ||
5079
	       (executable && (event->attr.mmap || event->attr.mmap2));
5080 5081
}

5082
static void perf_event_mmap_output(struct perf_event *event,
5083
				   void *data)
5084
{
5085
	struct perf_mmap_event *mmap_event = data;
5086
	struct perf_output_handle handle;
5087
	struct perf_sample_data sample;
5088
	int size = mmap_event->event_id.header.size;
5089
	int ret;
5090

5091 5092 5093
	if (!perf_event_mmap_match(event, data))
		return;

5094 5095 5096 5097 5098
	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);
5099
		mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5100 5101
	}

5102 5103
	perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5104
				mmap_event->event_id.header.size);
5105
	if (ret)
5106
		goto out;
5107

5108 5109
	mmap_event->event_id.pid = perf_event_pid(event, current);
	mmap_event->event_id.tid = perf_event_tid(event, current);
5110

5111
	perf_output_put(&handle, mmap_event->event_id);
5112 5113 5114 5115 5116 5117 5118 5119

	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);
	}

5120
	__output_copy(&handle, mmap_event->file_name,
5121
				   mmap_event->file_size);
5122 5123 5124

	perf_event__output_id_sample(event, &handle, &sample);

5125
	perf_output_end(&handle);
5126 5127
out:
	mmap_event->event_id.header.size = size;
5128 5129
}

5130
static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5131
{
5132 5133
	struct vm_area_struct *vma = mmap_event->vma;
	struct file *file = vma->vm_file;
5134 5135
	int maj = 0, min = 0;
	u64 ino = 0, gen = 0;
5136 5137 5138
	unsigned int size;
	char tmp[16];
	char *buf = NULL;
5139
	char *name;
5140

5141
	if (file) {
5142 5143
		struct inode *inode;
		dev_t dev;
5144

5145
		buf = kmalloc(PATH_MAX, GFP_KERNEL);
5146 5147 5148 5149
		if (!buf) {
			name = strncpy(tmp, "//enomem", sizeof(tmp));
			goto got_name;
		}
5150 5151 5152 5153 5154 5155
		/*
		 * d_path() works from the end of the rb backwards, so we
		 * need to add enough zero bytes after the string to handle
		 * the 64bit alignment we do later.
		 */
		name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5156 5157 5158 5159
		if (IS_ERR(name)) {
			name = strncpy(tmp, "//toolong", sizeof(tmp));
			goto got_name;
		}
5160 5161 5162 5163 5164 5165 5166
		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);

5167
	} else {
5168
		name = (char *)arch_vma_name(vma);
5169 5170
		if (name) {
			name = strncpy(tmp, name, sizeof(tmp) - 1);
5171
			tmp[sizeof(tmp) - 1] = '\0';
5172
			goto got_name;
5173
		}
5174

5175
		if (vma->vm_start <= vma->vm_mm->start_brk &&
5176 5177 5178
				vma->vm_end >= vma->vm_mm->brk) {
			name = strncpy(tmp, "[heap]", sizeof(tmp));
			goto got_name;
5179 5180
		}
		if (vma->vm_start <= vma->vm_mm->start_stack &&
5181 5182 5183
				vma->vm_end >= vma->vm_mm->start_stack) {
			name = strncpy(tmp, "[stack]", sizeof(tmp));
			goto got_name;
5184 5185
		}

5186 5187 5188 5189 5190
		name = strncpy(tmp, "//anon", sizeof(tmp));
		goto got_name;
	}

got_name:
5191 5192 5193 5194 5195 5196 5197 5198
	/*
	 * 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';
5199 5200 5201

	mmap_event->file_name = name;
	mmap_event->file_size = size;
5202 5203 5204 5205
	mmap_event->maj = maj;
	mmap_event->min = min;
	mmap_event->ino = ino;
	mmap_event->ino_generation = gen;
5206

5207 5208 5209
	if (!(vma->vm_flags & VM_EXEC))
		mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;

5210
	mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5211

5212
	perf_event_aux(perf_event_mmap_output,
5213 5214
		       mmap_event,
		       NULL);
5215

5216 5217 5218
	kfree(buf);
}

5219
void perf_event_mmap(struct vm_area_struct *vma)
5220
{
5221 5222
	struct perf_mmap_event mmap_event;

5223
	if (!atomic_read(&nr_mmap_events))
5224 5225 5226
		return;

	mmap_event = (struct perf_mmap_event){
5227
		.vma	= vma,
5228 5229
		/* .file_name */
		/* .file_size */
5230
		.event_id  = {
5231
			.header = {
5232
				.type = PERF_RECORD_MMAP,
5233
				.misc = PERF_RECORD_MISC_USER,
5234 5235 5236 5237
				/* .size */
			},
			/* .pid */
			/* .tid */
5238 5239
			.start  = vma->vm_start,
			.len    = vma->vm_end - vma->vm_start,
5240
			.pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
5241
		},
5242 5243 5244 5245
		/* .maj (attr_mmap2 only) */
		/* .min (attr_mmap2 only) */
		/* .ino (attr_mmap2 only) */
		/* .ino_generation (attr_mmap2 only) */
5246 5247
	};

5248
	perf_event_mmap_event(&mmap_event);
5249 5250
}

5251 5252 5253 5254
/*
 * IRQ throttle logging
 */

5255
static void perf_log_throttle(struct perf_event *event, int enable)
5256 5257
{
	struct perf_output_handle handle;
5258
	struct perf_sample_data sample;
5259 5260 5261 5262 5263
	int ret;

	struct {
		struct perf_event_header	header;
		u64				time;
5264
		u64				id;
5265
		u64				stream_id;
5266 5267
	} throttle_event = {
		.header = {
5268
			.type = PERF_RECORD_THROTTLE,
5269 5270 5271
			.misc = 0,
			.size = sizeof(throttle_event),
		},
P
Peter Zijlstra 已提交
5272
		.time		= perf_clock(),
5273 5274
		.id		= primary_event_id(event),
		.stream_id	= event->id,
5275 5276
	};

5277
	if (enable)
5278
		throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5279

5280 5281 5282
	perf_event_header__init_id(&throttle_event.header, &sample, event);

	ret = perf_output_begin(&handle, event,
5283
				throttle_event.header.size);
5284 5285 5286 5287
	if (ret)
		return;

	perf_output_put(&handle, throttle_event);
5288
	perf_event__output_id_sample(event, &handle, &sample);
5289 5290 5291
	perf_output_end(&handle);
}

5292
/*
5293
 * Generic event overflow handling, sampling.
5294 5295
 */

5296
static int __perf_event_overflow(struct perf_event *event,
5297 5298
				   int throttle, struct perf_sample_data *data,
				   struct pt_regs *regs)
5299
{
5300 5301
	int events = atomic_read(&event->event_limit);
	struct hw_perf_event *hwc = &event->hw;
5302
	u64 seq;
5303 5304
	int ret = 0;

5305 5306 5307 5308 5309 5310 5311
	/*
	 * Non-sampling counters might still use the PMI to fold short
	 * hardware counters, ignore those.
	 */
	if (unlikely(!is_sampling_event(event)))
		return 0;

5312 5313 5314 5315 5316 5317 5318 5319 5320
	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
5321 5322
			hwc->interrupts = MAX_INTERRUPTS;
			perf_log_throttle(event, 0);
5323
			tick_nohz_full_kick();
5324 5325
			ret = 1;
		}
5326
	}
5327

5328
	if (event->attr.freq) {
P
Peter Zijlstra 已提交
5329
		u64 now = perf_clock();
5330
		s64 delta = now - hwc->freq_time_stamp;
5331

5332
		hwc->freq_time_stamp = now;
5333

5334
		if (delta > 0 && delta < 2*TICK_NSEC)
5335
			perf_adjust_period(event, delta, hwc->last_period, true);
5336 5337
	}

5338 5339
	/*
	 * XXX event_limit might not quite work as expected on inherited
5340
	 * events
5341 5342
	 */

5343 5344
	event->pending_kill = POLL_IN;
	if (events && atomic_dec_and_test(&event->event_limit)) {
5345
		ret = 1;
5346
		event->pending_kill = POLL_HUP;
5347 5348
		event->pending_disable = 1;
		irq_work_queue(&event->pending);
5349 5350
	}

5351
	if (event->overflow_handler)
5352
		event->overflow_handler(event, data, regs);
5353
	else
5354
		perf_event_output(event, data, regs);
5355

P
Peter Zijlstra 已提交
5356
	if (event->fasync && event->pending_kill) {
5357 5358
		event->pending_wakeup = 1;
		irq_work_queue(&event->pending);
P
Peter Zijlstra 已提交
5359 5360
	}

5361
	return ret;
5362 5363
}

5364
int perf_event_overflow(struct perf_event *event,
5365 5366
			  struct perf_sample_data *data,
			  struct pt_regs *regs)
5367
{
5368
	return __perf_event_overflow(event, 1, data, regs);
5369 5370
}

5371
/*
5372
 * Generic software event infrastructure
5373 5374
 */

5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385
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];
};

static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);

5386
/*
5387 5388
 * We directly increment event->count and keep a second value in
 * event->hw.period_left to count intervals. This period event
5389 5390 5391 5392
 * is kept in the range [-sample_period, 0] so that we can use the
 * sign as trigger.
 */

5393
u64 perf_swevent_set_period(struct perf_event *event)
5394
{
5395
	struct hw_perf_event *hwc = &event->hw;
5396 5397 5398 5399 5400
	u64 period = hwc->last_period;
	u64 nr, offset;
	s64 old, val;

	hwc->last_period = hwc->sample_period;
5401 5402

again:
5403
	old = val = local64_read(&hwc->period_left);
5404 5405
	if (val < 0)
		return 0;
5406

5407 5408 5409
	nr = div64_u64(period + val, period);
	offset = nr * period;
	val -= offset;
5410
	if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5411
		goto again;
5412

5413
	return nr;
5414 5415
}

5416
static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5417
				    struct perf_sample_data *data,
5418
				    struct pt_regs *regs)
5419
{
5420
	struct hw_perf_event *hwc = &event->hw;
5421
	int throttle = 0;
5422

5423 5424
	if (!overflow)
		overflow = perf_swevent_set_period(event);
5425

5426 5427
	if (hwc->interrupts == MAX_INTERRUPTS)
		return;
5428

5429
	for (; overflow; overflow--) {
5430
		if (__perf_event_overflow(event, throttle,
5431
					    data, regs)) {
5432 5433 5434 5435 5436 5437
			/*
			 * We inhibit the overflow from happening when
			 * hwc->interrupts == MAX_INTERRUPTS.
			 */
			break;
		}
5438
		throttle = 1;
5439
	}
5440 5441
}

P
Peter Zijlstra 已提交
5442
static void perf_swevent_event(struct perf_event *event, u64 nr,
5443
			       struct perf_sample_data *data,
5444
			       struct pt_regs *regs)
5445
{
5446
	struct hw_perf_event *hwc = &event->hw;
5447

5448
	local64_add(nr, &event->count);
5449

5450 5451 5452
	if (!regs)
		return;

5453
	if (!is_sampling_event(event))
5454
		return;
5455

5456 5457 5458 5459 5460 5461
	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;

5462
	if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5463
		return perf_swevent_overflow(event, 1, data, regs);
5464

5465
	if (local64_add_negative(nr, &hwc->period_left))
5466
		return;
5467

5468
	perf_swevent_overflow(event, 0, data, regs);
5469 5470
}

5471 5472 5473
static int perf_exclude_event(struct perf_event *event,
			      struct pt_regs *regs)
{
P
Peter Zijlstra 已提交
5474
	if (event->hw.state & PERF_HES_STOPPED)
5475
		return 1;
P
Peter Zijlstra 已提交
5476

5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487
	if (regs) {
		if (event->attr.exclude_user && user_mode(regs))
			return 1;

		if (event->attr.exclude_kernel && !user_mode(regs))
			return 1;
	}

	return 0;
}

5488
static int perf_swevent_match(struct perf_event *event,
P
Peter Zijlstra 已提交
5489
				enum perf_type_id type,
L
Li Zefan 已提交
5490 5491 5492
				u32 event_id,
				struct perf_sample_data *data,
				struct pt_regs *regs)
5493
{
5494
	if (event->attr.type != type)
5495
		return 0;
5496

5497
	if (event->attr.config != event_id)
5498 5499
		return 0;

5500 5501
	if (perf_exclude_event(event, regs))
		return 0;
5502 5503 5504 5505

	return 1;
}

5506 5507 5508 5509 5510 5511 5512
static inline u64 swevent_hash(u64 type, u32 event_id)
{
	u64 val = event_id | (type << 32);

	return hash_64(val, SWEVENT_HLIST_BITS);
}

5513 5514
static inline struct hlist_head *
__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5515
{
5516 5517 5518 5519
	u64 hash = swevent_hash(type, event_id);

	return &hlist->heads[hash];
}
5520

5521 5522
/* For the read side: events when they trigger */
static inline struct hlist_head *
5523
find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5524 5525
{
	struct swevent_hlist *hlist;
5526

5527
	hlist = rcu_dereference(swhash->swevent_hlist);
5528 5529 5530
	if (!hlist)
		return NULL;

5531 5532 5533 5534 5535
	return __find_swevent_head(hlist, type, event_id);
}

/* For the event head insertion and removal in the hlist */
static inline struct hlist_head *
5536
find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5537 5538 5539 5540 5541 5542 5543 5544 5545 5546
{
	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.
	 */
5547
	hlist = rcu_dereference_protected(swhash->swevent_hlist,
5548 5549 5550 5551 5552
					  lockdep_is_held(&event->ctx->lock));
	if (!hlist)
		return NULL;

	return __find_swevent_head(hlist, type, event_id);
5553 5554 5555
}

static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5556
				    u64 nr,
5557 5558
				    struct perf_sample_data *data,
				    struct pt_regs *regs)
5559
{
5560
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5561
	struct perf_event *event;
5562
	struct hlist_head *head;
5563

5564
	rcu_read_lock();
5565
	head = find_swevent_head_rcu(swhash, type, event_id);
5566 5567 5568
	if (!head)
		goto end;

5569
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
L
Li Zefan 已提交
5570
		if (perf_swevent_match(event, type, event_id, data, regs))
5571
			perf_swevent_event(event, nr, data, regs);
5572
	}
5573 5574
end:
	rcu_read_unlock();
5575 5576
}

5577
int perf_swevent_get_recursion_context(void)
P
Peter Zijlstra 已提交
5578
{
5579
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
P
Peter Zijlstra 已提交
5580

5581
	return get_recursion_context(swhash->recursion);
P
Peter Zijlstra 已提交
5582
}
I
Ingo Molnar 已提交
5583
EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
P
Peter Zijlstra 已提交
5584

5585
inline void perf_swevent_put_recursion_context(int rctx)
5586
{
5587
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5588

5589
	put_recursion_context(swhash->recursion, rctx);
5590
}
5591

5592
void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5593
{
5594
	struct perf_sample_data data;
5595 5596
	int rctx;

5597
	preempt_disable_notrace();
5598 5599 5600
	rctx = perf_swevent_get_recursion_context();
	if (rctx < 0)
		return;
5601

5602
	perf_sample_data_init(&data, addr, 0);
5603

5604
	do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5605 5606

	perf_swevent_put_recursion_context(rctx);
5607
	preempt_enable_notrace();
5608 5609
}

5610
static void perf_swevent_read(struct perf_event *event)
5611 5612 5613
{
}

P
Peter Zijlstra 已提交
5614
static int perf_swevent_add(struct perf_event *event, int flags)
5615
{
5616
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5617
	struct hw_perf_event *hwc = &event->hw;
5618 5619
	struct hlist_head *head;

5620
	if (is_sampling_event(event)) {
5621
		hwc->last_period = hwc->sample_period;
5622
		perf_swevent_set_period(event);
5623
	}
5624

P
Peter Zijlstra 已提交
5625 5626
	hwc->state = !(flags & PERF_EF_START);

5627
	head = find_swevent_head(swhash, event);
5628 5629 5630 5631 5632
	if (WARN_ON_ONCE(!head))
		return -EINVAL;

	hlist_add_head_rcu(&event->hlist_entry, head);

5633 5634 5635
	return 0;
}

P
Peter Zijlstra 已提交
5636
static void perf_swevent_del(struct perf_event *event, int flags)
5637
{
5638
	hlist_del_rcu(&event->hlist_entry);
5639 5640
}

P
Peter Zijlstra 已提交
5641
static void perf_swevent_start(struct perf_event *event, int flags)
5642
{
P
Peter Zijlstra 已提交
5643
	event->hw.state = 0;
5644
}
I
Ingo Molnar 已提交
5645

P
Peter Zijlstra 已提交
5646
static void perf_swevent_stop(struct perf_event *event, int flags)
5647
{
P
Peter Zijlstra 已提交
5648
	event->hw.state = PERF_HES_STOPPED;
5649 5650
}

5651 5652
/* Deref the hlist from the update side */
static inline struct swevent_hlist *
5653
swevent_hlist_deref(struct swevent_htable *swhash)
5654
{
5655 5656
	return rcu_dereference_protected(swhash->swevent_hlist,
					 lockdep_is_held(&swhash->hlist_mutex));
5657 5658
}

5659
static void swevent_hlist_release(struct swevent_htable *swhash)
5660
{
5661
	struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5662

5663
	if (!hlist)
5664 5665
		return;

5666
	rcu_assign_pointer(swhash->swevent_hlist, NULL);
5667
	kfree_rcu(hlist, rcu_head);
5668 5669 5670 5671
}

static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
{
5672
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5673

5674
	mutex_lock(&swhash->hlist_mutex);
5675

5676 5677
	if (!--swhash->hlist_refcount)
		swevent_hlist_release(swhash);
5678

5679
	mutex_unlock(&swhash->hlist_mutex);
5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696
}

static void swevent_hlist_put(struct perf_event *event)
{
	int cpu;

	if (event->cpu != -1) {
		swevent_hlist_put_cpu(event, event->cpu);
		return;
	}

	for_each_possible_cpu(cpu)
		swevent_hlist_put_cpu(event, cpu);
}

static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
{
5697
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5698 5699
	int err = 0;

5700
	mutex_lock(&swhash->hlist_mutex);
5701

5702
	if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5703 5704 5705 5706 5707 5708 5709
		struct swevent_hlist *hlist;

		hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
		if (!hlist) {
			err = -ENOMEM;
			goto exit;
		}
5710
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
5711
	}
5712
	swhash->hlist_refcount++;
P
Peter Zijlstra 已提交
5713
exit:
5714
	mutex_unlock(&swhash->hlist_mutex);
5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737

	return err;
}

static int swevent_hlist_get(struct perf_event *event)
{
	int err;
	int cpu, failed_cpu;

	if (event->cpu != -1)
		return swevent_hlist_get_cpu(event, event->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
5738
fail:
5739 5740 5741 5742 5743 5744 5745 5746 5747 5748
	for_each_possible_cpu(cpu) {
		if (cpu == failed_cpu)
			break;
		swevent_hlist_put_cpu(event, cpu);
	}

	put_online_cpus();
	return err;
}

5749
struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5750

5751 5752 5753
static void sw_perf_event_destroy(struct perf_event *event)
{
	u64 event_id = event->attr.config;
5754

5755 5756
	WARN_ON(event->parent);

5757
	static_key_slow_dec(&perf_swevent_enabled[event_id]);
5758 5759 5760 5761 5762
	swevent_hlist_put(event);
}

static int perf_swevent_init(struct perf_event *event)
{
5763
	u64 event_id = event->attr.config;
5764 5765 5766 5767

	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

5768 5769 5770 5771 5772 5773
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

5774 5775 5776 5777 5778 5779 5780 5781 5782
	switch (event_id) {
	case PERF_COUNT_SW_CPU_CLOCK:
	case PERF_COUNT_SW_TASK_CLOCK:
		return -ENOENT;

	default:
		break;
	}

5783
	if (event_id >= PERF_COUNT_SW_MAX)
5784 5785 5786 5787 5788 5789 5790 5791 5792
		return -ENOENT;

	if (!event->parent) {
		int err;

		err = swevent_hlist_get(event);
		if (err)
			return err;

5793
		static_key_slow_inc(&perf_swevent_enabled[event_id]);
5794 5795 5796 5797 5798 5799
		event->destroy = sw_perf_event_destroy;
	}

	return 0;
}

5800 5801 5802 5803 5804
static int perf_swevent_event_idx(struct perf_event *event)
{
	return 0;
}

5805
static struct pmu perf_swevent = {
5806
	.task_ctx_nr	= perf_sw_context,
5807

5808
	.event_init	= perf_swevent_init,
P
Peter Zijlstra 已提交
5809 5810 5811 5812
	.add		= perf_swevent_add,
	.del		= perf_swevent_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
5813
	.read		= perf_swevent_read,
5814 5815

	.event_idx	= perf_swevent_event_idx,
5816 5817
};

5818 5819
#ifdef CONFIG_EVENT_TRACING

5820 5821 5822 5823 5824 5825 5826 5827 5828 5829 5830 5831 5832 5833
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)
{
5834 5835
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;
5836 5837 5838 5839
	/*
	 * All tracepoints are from kernel-space.
	 */
	if (event->attr.exclude_kernel)
5840 5841 5842 5843 5844 5845 5846 5847 5848
		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,
5849 5850
		   struct pt_regs *regs, struct hlist_head *head, int rctx,
		   struct task_struct *task)
5851 5852
{
	struct perf_sample_data data;
5853 5854
	struct perf_event *event;

5855 5856 5857 5858 5859
	struct perf_raw_record raw = {
		.size = entry_size,
		.data = record,
	};

5860
	perf_sample_data_init(&data, addr, 0);
5861 5862
	data.raw = &raw;

5863
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
5864
		if (perf_tp_event_match(event, &data, regs))
5865
			perf_swevent_event(event, count, &data, regs);
5866
	}
5867

5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892
	/*
	 * 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();
	}

5893
	perf_swevent_put_recursion_context(rctx);
5894 5895 5896
}
EXPORT_SYMBOL_GPL(perf_tp_event);

5897
static void tp_perf_event_destroy(struct perf_event *event)
5898
{
5899
	perf_trace_destroy(event);
5900 5901
}

5902
static int perf_tp_event_init(struct perf_event *event)
5903
{
5904 5905
	int err;

5906 5907 5908
	if (event->attr.type != PERF_TYPE_TRACEPOINT)
		return -ENOENT;

5909 5910 5911 5912 5913 5914
	/*
	 * no branch sampling for tracepoint events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

5915 5916
	err = perf_trace_init(event);
	if (err)
5917
		return err;
5918

5919
	event->destroy = tp_perf_event_destroy;
5920

5921 5922 5923 5924
	return 0;
}

static struct pmu perf_tracepoint = {
5925 5926
	.task_ctx_nr	= perf_sw_context,

5927
	.event_init	= perf_tp_event_init,
P
Peter Zijlstra 已提交
5928 5929 5930 5931
	.add		= perf_trace_add,
	.del		= perf_trace_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
5932
	.read		= perf_swevent_read,
5933 5934

	.event_idx	= perf_swevent_event_idx,
5935 5936 5937 5938
};

static inline void perf_tp_register(void)
{
P
Peter Zijlstra 已提交
5939
	perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5940
}
L
Li Zefan 已提交
5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 5962 5963 5964

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);
}

5965
#else
L
Li Zefan 已提交
5966

5967
static inline void perf_tp_register(void)
5968 5969
{
}
L
Li Zefan 已提交
5970 5971 5972 5973 5974 5975 5976 5977 5978 5979

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)
{
}

5980
#endif /* CONFIG_EVENT_TRACING */
5981

5982
#ifdef CONFIG_HAVE_HW_BREAKPOINT
5983
void perf_bp_event(struct perf_event *bp, void *data)
5984
{
5985 5986 5987
	struct perf_sample_data sample;
	struct pt_regs *regs = data;

5988
	perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5989

P
Peter Zijlstra 已提交
5990
	if (!bp->hw.state && !perf_exclude_event(bp, regs))
5991
		perf_swevent_event(bp, 1, &sample, regs);
5992 5993 5994
}
#endif

5995 5996 5997
/*
 * hrtimer based swevent callback
 */
5998

5999
static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6000
{
6001 6002 6003 6004 6005
	enum hrtimer_restart ret = HRTIMER_RESTART;
	struct perf_sample_data data;
	struct pt_regs *regs;
	struct perf_event *event;
	u64 period;
6006

6007
	event = container_of(hrtimer, struct perf_event, hw.hrtimer);
P
Peter Zijlstra 已提交
6008 6009 6010 6011

	if (event->state != PERF_EVENT_STATE_ACTIVE)
		return HRTIMER_NORESTART;

6012
	event->pmu->read(event);
6013

6014
	perf_sample_data_init(&data, 0, event->hw.last_period);
6015 6016 6017
	regs = get_irq_regs();

	if (regs && !perf_exclude_event(event, regs)) {
6018
		if (!(event->attr.exclude_idle && is_idle_task(current)))
6019
			if (__perf_event_overflow(event, 1, &data, regs))
6020 6021
				ret = HRTIMER_NORESTART;
	}
6022

6023 6024
	period = max_t(u64, 10000, event->hw.sample_period);
	hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6025

6026
	return ret;
6027 6028
}

6029
static void perf_swevent_start_hrtimer(struct perf_event *event)
6030
{
6031
	struct hw_perf_event *hwc = &event->hw;
6032 6033 6034 6035
	s64 period;

	if (!is_sampling_event(event))
		return;
6036

6037 6038 6039 6040
	period = local64_read(&hwc->period_left);
	if (period) {
		if (period < 0)
			period = 10000;
P
Peter Zijlstra 已提交
6041

6042 6043 6044 6045 6046
		local64_set(&hwc->period_left, 0);
	} else {
		period = max_t(u64, 10000, hwc->sample_period);
	}
	__hrtimer_start_range_ns(&hwc->hrtimer,
6047
				ns_to_ktime(period), 0,
6048
				HRTIMER_MODE_REL_PINNED, 0);
6049
}
6050 6051

static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6052
{
6053 6054
	struct hw_perf_event *hwc = &event->hw;

6055
	if (is_sampling_event(event)) {
6056
		ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
P
Peter Zijlstra 已提交
6057
		local64_set(&hwc->period_left, ktime_to_ns(remaining));
6058 6059 6060

		hrtimer_cancel(&hwc->hrtimer);
	}
6061 6062
}

P
Peter Zijlstra 已提交
6063 6064 6065 6066 6067 6068 6069 6070 6071 6072 6073 6074 6075 6076 6077 6078 6079 6080 6081 6082
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);
6083
		hwc->last_period = hwc->sample_period;
P
Peter Zijlstra 已提交
6084 6085 6086 6087
		event->attr.freq = 0;
	}
}

6088 6089 6090 6091 6092
/*
 * Software event: cpu wall time clock
 */

static void cpu_clock_event_update(struct perf_event *event)
6093
{
6094 6095 6096
	s64 prev;
	u64 now;

P
Peter Zijlstra 已提交
6097
	now = local_clock();
6098 6099
	prev = local64_xchg(&event->hw.prev_count, now);
	local64_add(now - prev, &event->count);
6100 6101
}

P
Peter Zijlstra 已提交
6102
static void cpu_clock_event_start(struct perf_event *event, int flags)
6103
{
P
Peter Zijlstra 已提交
6104
	local64_set(&event->hw.prev_count, local_clock());
6105 6106 6107
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6108
static void cpu_clock_event_stop(struct perf_event *event, int flags)
6109
{
6110 6111 6112
	perf_swevent_cancel_hrtimer(event);
	cpu_clock_event_update(event);
}
6113

P
Peter Zijlstra 已提交
6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126
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);
}

6127 6128 6129 6130
static void cpu_clock_event_read(struct perf_event *event)
{
	cpu_clock_event_update(event);
}
6131

6132 6133 6134 6135 6136 6137 6138 6139
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;

6140 6141 6142 6143 6144 6145
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6146 6147
	perf_swevent_init_hrtimer(event);

6148
	return 0;
6149 6150
}

6151
static struct pmu perf_cpu_clock = {
6152 6153
	.task_ctx_nr	= perf_sw_context,

6154
	.event_init	= cpu_clock_event_init,
P
Peter Zijlstra 已提交
6155 6156 6157 6158
	.add		= cpu_clock_event_add,
	.del		= cpu_clock_event_del,
	.start		= cpu_clock_event_start,
	.stop		= cpu_clock_event_stop,
6159
	.read		= cpu_clock_event_read,
6160 6161

	.event_idx	= perf_swevent_event_idx,
6162 6163 6164 6165 6166 6167 6168
};

/*
 * Software event: task time clock
 */

static void task_clock_event_update(struct perf_event *event, u64 now)
6169
{
6170 6171
	u64 prev;
	s64 delta;
6172

6173 6174 6175 6176
	prev = local64_xchg(&event->hw.prev_count, now);
	delta = now - prev;
	local64_add(delta, &event->count);
}
6177

P
Peter Zijlstra 已提交
6178
static void task_clock_event_start(struct perf_event *event, int flags)
6179
{
P
Peter Zijlstra 已提交
6180
	local64_set(&event->hw.prev_count, event->ctx->time);
6181 6182 6183
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6184
static void task_clock_event_stop(struct perf_event *event, int flags)
6185 6186 6187
{
	perf_swevent_cancel_hrtimer(event);
	task_clock_event_update(event, event->ctx->time);
P
Peter Zijlstra 已提交
6188 6189 6190 6191 6192 6193
}

static int task_clock_event_add(struct perf_event *event, int flags)
{
	if (flags & PERF_EF_START)
		task_clock_event_start(event, flags);
6194

P
Peter Zijlstra 已提交
6195 6196 6197 6198 6199 6200
	return 0;
}

static void task_clock_event_del(struct perf_event *event, int flags)
{
	task_clock_event_stop(event, PERF_EF_UPDATE);
6201 6202 6203 6204
}

static void task_clock_event_read(struct perf_event *event)
{
6205 6206 6207
	u64 now = perf_clock();
	u64 delta = now - event->ctx->timestamp;
	u64 time = event->ctx->time + delta;
6208 6209 6210 6211 6212

	task_clock_event_update(event, time);
}

static int task_clock_event_init(struct perf_event *event)
L
Li Zefan 已提交
6213
{
6214 6215 6216 6217 6218 6219
	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

	if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
		return -ENOENT;

6220 6221 6222 6223 6224 6225
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6226 6227
	perf_swevent_init_hrtimer(event);

6228
	return 0;
L
Li Zefan 已提交
6229 6230
}

6231
static struct pmu perf_task_clock = {
6232 6233
	.task_ctx_nr	= perf_sw_context,

6234
	.event_init	= task_clock_event_init,
P
Peter Zijlstra 已提交
6235 6236 6237 6238
	.add		= task_clock_event_add,
	.del		= task_clock_event_del,
	.start		= task_clock_event_start,
	.stop		= task_clock_event_stop,
6239
	.read		= task_clock_event_read,
6240 6241

	.event_idx	= perf_swevent_event_idx,
6242
};
L
Li Zefan 已提交
6243

P
Peter Zijlstra 已提交
6244
static void perf_pmu_nop_void(struct pmu *pmu)
6245 6246
{
}
L
Li Zefan 已提交
6247

P
Peter Zijlstra 已提交
6248
static int perf_pmu_nop_int(struct pmu *pmu)
L
Li Zefan 已提交
6249
{
P
Peter Zijlstra 已提交
6250
	return 0;
L
Li Zefan 已提交
6251 6252
}

P
Peter Zijlstra 已提交
6253
static void perf_pmu_start_txn(struct pmu *pmu)
L
Li Zefan 已提交
6254
{
P
Peter Zijlstra 已提交
6255
	perf_pmu_disable(pmu);
L
Li Zefan 已提交
6256 6257
}

P
Peter Zijlstra 已提交
6258 6259 6260 6261 6262
static int perf_pmu_commit_txn(struct pmu *pmu)
{
	perf_pmu_enable(pmu);
	return 0;
}
6263

P
Peter Zijlstra 已提交
6264
static void perf_pmu_cancel_txn(struct pmu *pmu)
6265
{
P
Peter Zijlstra 已提交
6266
	perf_pmu_enable(pmu);
6267 6268
}

6269 6270 6271 6272 6273
static int perf_event_idx_default(struct perf_event *event)
{
	return event->hw.idx + 1;
}

P
Peter Zijlstra 已提交
6274 6275 6276 6277 6278
/*
 * Ensures all contexts with the same task_ctx_nr have the same
 * pmu_cpu_context too.
 */
static void *find_pmu_context(int ctxn)
6279
{
P
Peter Zijlstra 已提交
6280
	struct pmu *pmu;
6281

P
Peter Zijlstra 已提交
6282 6283
	if (ctxn < 0)
		return NULL;
6284

P
Peter Zijlstra 已提交
6285 6286 6287 6288
	list_for_each_entry(pmu, &pmus, entry) {
		if (pmu->task_ctx_nr == ctxn)
			return pmu->pmu_cpu_context;
	}
6289

P
Peter Zijlstra 已提交
6290
	return NULL;
6291 6292
}

6293
static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6294
{
6295 6296 6297 6298 6299 6300 6301
	int cpu;

	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);

6302 6303
		if (cpuctx->unique_pmu == old_pmu)
			cpuctx->unique_pmu = pmu;
6304 6305 6306 6307 6308 6309
	}
}

static void free_pmu_context(struct pmu *pmu)
{
	struct pmu *i;
6310

P
Peter Zijlstra 已提交
6311
	mutex_lock(&pmus_lock);
6312
	/*
P
Peter Zijlstra 已提交
6313
	 * Like a real lame refcount.
6314
	 */
6315 6316 6317
	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
6318
			goto out;
6319
		}
P
Peter Zijlstra 已提交
6320
	}
6321

6322
	free_percpu(pmu->pmu_cpu_context);
P
Peter Zijlstra 已提交
6323 6324
out:
	mutex_unlock(&pmus_lock);
6325
}
P
Peter Zijlstra 已提交
6326
static struct idr pmu_idr;
6327

P
Peter Zijlstra 已提交
6328 6329 6330 6331 6332 6333 6334 6335
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);
}

6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379
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;
}

P
Peter Zijlstra 已提交
6380
static struct device_attribute pmu_dev_attrs[] = {
6381 6382 6383
	__ATTR_RO(type),
	__ATTR_RW(perf_event_mux_interval_ms),
	__ATTR_NULL,
P
Peter Zijlstra 已提交
6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404
};

static int pmu_bus_running;
static struct bus_type pmu_bus = {
	.name		= "event_source",
	.dev_attrs	= pmu_dev_attrs,
};

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;

6405
	pmu->dev->groups = pmu->attr_groups;
P
Peter Zijlstra 已提交
6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425
	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;
}

6426
static struct lock_class_key cpuctx_mutex;
6427
static struct lock_class_key cpuctx_lock;
6428

6429
int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6430
{
P
Peter Zijlstra 已提交
6431
	int cpu, ret;
6432

6433
	mutex_lock(&pmus_lock);
P
Peter Zijlstra 已提交
6434 6435 6436 6437
	ret = -ENOMEM;
	pmu->pmu_disable_count = alloc_percpu(int);
	if (!pmu->pmu_disable_count)
		goto unlock;
6438

P
Peter Zijlstra 已提交
6439 6440 6441 6442 6443 6444
	pmu->type = -1;
	if (!name)
		goto skip_type;
	pmu->name = name;

	if (type < 0) {
T
Tejun Heo 已提交
6445 6446 6447
		type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
		if (type < 0) {
			ret = type;
P
Peter Zijlstra 已提交
6448 6449 6450 6451 6452
			goto free_pdc;
		}
	}
	pmu->type = type;

P
Peter Zijlstra 已提交
6453 6454 6455 6456 6457 6458
	if (pmu_bus_running) {
		ret = pmu_dev_alloc(pmu);
		if (ret)
			goto free_idr;
	}

P
Peter Zijlstra 已提交
6459
skip_type:
P
Peter Zijlstra 已提交
6460 6461 6462
	pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
	if (pmu->pmu_cpu_context)
		goto got_cpu_context;
6463

W
Wei Yongjun 已提交
6464
	ret = -ENOMEM;
P
Peter Zijlstra 已提交
6465 6466
	pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
	if (!pmu->pmu_cpu_context)
P
Peter Zijlstra 已提交
6467
		goto free_dev;
6468

P
Peter Zijlstra 已提交
6469 6470 6471 6472
	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6473
		__perf_event_init_context(&cpuctx->ctx);
6474
		lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6475
		lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6476
		cpuctx->ctx.type = cpu_context;
P
Peter Zijlstra 已提交
6477
		cpuctx->ctx.pmu = pmu;
6478 6479 6480

		__perf_cpu_hrtimer_init(cpuctx, cpu);

6481
		INIT_LIST_HEAD(&cpuctx->rotation_list);
6482
		cpuctx->unique_pmu = pmu;
P
Peter Zijlstra 已提交
6483
	}
6484

P
Peter Zijlstra 已提交
6485
got_cpu_context:
P
Peter Zijlstra 已提交
6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499
	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;
6500
		}
6501
	}
6502

P
Peter Zijlstra 已提交
6503 6504 6505 6506 6507
	if (!pmu->pmu_enable) {
		pmu->pmu_enable  = perf_pmu_nop_void;
		pmu->pmu_disable = perf_pmu_nop_void;
	}

6508 6509 6510
	if (!pmu->event_idx)
		pmu->event_idx = perf_event_idx_default;

6511
	list_add_rcu(&pmu->entry, &pmus);
P
Peter Zijlstra 已提交
6512 6513
	ret = 0;
unlock:
6514 6515
	mutex_unlock(&pmus_lock);

P
Peter Zijlstra 已提交
6516
	return ret;
P
Peter Zijlstra 已提交
6517

P
Peter Zijlstra 已提交
6518 6519 6520 6521
free_dev:
	device_del(pmu->dev);
	put_device(pmu->dev);

P
Peter Zijlstra 已提交
6522 6523 6524 6525
free_idr:
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);

P
Peter Zijlstra 已提交
6526 6527 6528
free_pdc:
	free_percpu(pmu->pmu_disable_count);
	goto unlock;
6529 6530
}

6531
void perf_pmu_unregister(struct pmu *pmu)
6532
{
6533 6534 6535
	mutex_lock(&pmus_lock);
	list_del_rcu(&pmu->entry);
	mutex_unlock(&pmus_lock);
6536

6537
	/*
P
Peter Zijlstra 已提交
6538 6539
	 * We dereference the pmu list under both SRCU and regular RCU, so
	 * synchronize against both of those.
6540
	 */
6541
	synchronize_srcu(&pmus_srcu);
P
Peter Zijlstra 已提交
6542
	synchronize_rcu();
6543

P
Peter Zijlstra 已提交
6544
	free_percpu(pmu->pmu_disable_count);
P
Peter Zijlstra 已提交
6545 6546
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);
P
Peter Zijlstra 已提交
6547 6548
	device_del(pmu->dev);
	put_device(pmu->dev);
6549
	free_pmu_context(pmu);
6550
}
6551

6552 6553 6554 6555
struct pmu *perf_init_event(struct perf_event *event)
{
	struct pmu *pmu = NULL;
	int idx;
6556
	int ret;
6557 6558

	idx = srcu_read_lock(&pmus_srcu);
P
Peter Zijlstra 已提交
6559 6560 6561 6562

	rcu_read_lock();
	pmu = idr_find(&pmu_idr, event->attr.type);
	rcu_read_unlock();
6563
	if (pmu) {
6564
		event->pmu = pmu;
6565 6566 6567
		ret = pmu->event_init(event);
		if (ret)
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6568
		goto unlock;
6569
	}
P
Peter Zijlstra 已提交
6570

6571
	list_for_each_entry_rcu(pmu, &pmus, entry) {
6572
		event->pmu = pmu;
6573
		ret = pmu->event_init(event);
6574
		if (!ret)
P
Peter Zijlstra 已提交
6575
			goto unlock;
6576

6577 6578
		if (ret != -ENOENT) {
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6579
			goto unlock;
6580
		}
6581
	}
P
Peter Zijlstra 已提交
6582 6583
	pmu = ERR_PTR(-ENOENT);
unlock:
6584
	srcu_read_unlock(&pmus_srcu, idx);
6585

6586
	return pmu;
6587 6588
}

6589 6590 6591 6592 6593 6594 6595 6596 6597 6598 6599 6600 6601
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));
}

6602 6603
static void account_event(struct perf_event *event)
{
6604 6605 6606
	if (event->parent)
		return;

6607 6608 6609 6610 6611 6612 6613 6614
	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);
6615 6616 6617 6618
	if (event->attr.freq) {
		if (atomic_inc_return(&nr_freq_events) == 1)
			tick_nohz_full_kick_all();
	}
6619
	if (has_branch_stack(event))
6620
		static_key_slow_inc(&perf_sched_events.key);
6621
	if (is_cgroup_event(event))
6622
		static_key_slow_inc(&perf_sched_events.key);
6623 6624

	account_event_cpu(event, event->cpu);
6625 6626
}

T
Thomas Gleixner 已提交
6627
/*
6628
 * Allocate and initialize a event structure
T
Thomas Gleixner 已提交
6629
 */
6630
static struct perf_event *
6631
perf_event_alloc(struct perf_event_attr *attr, int cpu,
6632 6633 6634
		 struct task_struct *task,
		 struct perf_event *group_leader,
		 struct perf_event *parent_event,
6635 6636
		 perf_overflow_handler_t overflow_handler,
		 void *context)
T
Thomas Gleixner 已提交
6637
{
P
Peter Zijlstra 已提交
6638
	struct pmu *pmu;
6639 6640
	struct perf_event *event;
	struct hw_perf_event *hwc;
6641
	long err = -EINVAL;
T
Thomas Gleixner 已提交
6642

6643 6644 6645 6646 6647
	if ((unsigned)cpu >= nr_cpu_ids) {
		if (!task || cpu != -1)
			return ERR_PTR(-EINVAL);
	}

6648
	event = kzalloc(sizeof(*event), GFP_KERNEL);
6649
	if (!event)
6650
		return ERR_PTR(-ENOMEM);
T
Thomas Gleixner 已提交
6651

6652
	/*
6653
	 * Single events are their own group leaders, with an
6654 6655 6656
	 * empty sibling list:
	 */
	if (!group_leader)
6657
		group_leader = event;
6658

6659 6660
	mutex_init(&event->child_mutex);
	INIT_LIST_HEAD(&event->child_list);
6661

6662 6663 6664
	INIT_LIST_HEAD(&event->group_entry);
	INIT_LIST_HEAD(&event->event_entry);
	INIT_LIST_HEAD(&event->sibling_list);
6665 6666
	INIT_LIST_HEAD(&event->rb_entry);

6667
	init_waitqueue_head(&event->waitq);
6668
	init_irq_work(&event->pending, perf_pending_event);
T
Thomas Gleixner 已提交
6669

6670
	mutex_init(&event->mmap_mutex);
6671

6672
	atomic_long_set(&event->refcount, 1);
6673 6674 6675 6676 6677
	event->cpu		= cpu;
	event->attr		= *attr;
	event->group_leader	= group_leader;
	event->pmu		= NULL;
	event->oncpu		= -1;
6678

6679
	event->parent		= parent_event;
6680

6681
	event->ns		= get_pid_ns(task_active_pid_ns(current));
6682
	event->id		= atomic64_inc_return(&perf_event_id);
6683

6684
	event->state		= PERF_EVENT_STATE_INACTIVE;
6685

6686 6687
	if (task) {
		event->attach_state = PERF_ATTACH_TASK;
6688 6689 6690

		if (attr->type == PERF_TYPE_TRACEPOINT)
			event->hw.tp_target = task;
6691 6692 6693 6694
#ifdef CONFIG_HAVE_HW_BREAKPOINT
		/*
		 * hw_breakpoint is a bit difficult here..
		 */
6695
		else if (attr->type == PERF_TYPE_BREAKPOINT)
6696 6697 6698 6699
			event->hw.bp_target = task;
#endif
	}

6700
	if (!overflow_handler && parent_event) {
6701
		overflow_handler = parent_event->overflow_handler;
6702 6703
		context = parent_event->overflow_handler_context;
	}
6704

6705
	event->overflow_handler	= overflow_handler;
6706
	event->overflow_handler_context = context;
6707

J
Jiri Olsa 已提交
6708
	perf_event__state_init(event);
6709

6710
	pmu = NULL;
6711

6712
	hwc = &event->hw;
6713
	hwc->sample_period = attr->sample_period;
6714
	if (attr->freq && attr->sample_freq)
6715
		hwc->sample_period = 1;
6716
	hwc->last_period = hwc->sample_period;
6717

6718
	local64_set(&hwc->period_left, hwc->sample_period);
6719

6720
	/*
6721
	 * we currently do not support PERF_FORMAT_GROUP on inherited events
6722
	 */
6723
	if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6724
		goto err_ns;
6725

6726
	pmu = perf_init_event(event);
6727
	if (!pmu)
6728 6729
		goto err_ns;
	else if (IS_ERR(pmu)) {
6730
		err = PTR_ERR(pmu);
6731
		goto err_ns;
I
Ingo Molnar 已提交
6732
	}
6733

6734
	if (!event->parent) {
6735 6736
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
			err = get_callchain_buffers();
6737 6738
			if (err)
				goto err_pmu;
6739
		}
6740
	}
6741

6742
	return event;
6743 6744 6745 6746 6747 6748 6749 6750 6751 6752

err_pmu:
	if (event->destroy)
		event->destroy(event);
err_ns:
	if (event->ns)
		put_pid_ns(event->ns);
	kfree(event);

	return ERR_PTR(err);
T
Thomas Gleixner 已提交
6753 6754
}

6755 6756
static int perf_copy_attr(struct perf_event_attr __user *uattr,
			  struct perf_event_attr *attr)
6757 6758
{
	u32 size;
6759
	int ret;
6760 6761 6762 6763 6764 6765 6766 6767 6768 6769 6770 6771 6772 6773 6774 6775 6776 6777 6778 6779 6780 6781 6782 6783

	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,
6784 6785 6786
	 * 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.
6787 6788
	 */
	if (size > sizeof(*attr)) {
6789 6790 6791
		unsigned char __user *addr;
		unsigned char __user *end;
		unsigned char val;
6792

6793 6794
		addr = (void __user *)uattr + sizeof(*attr);
		end  = (void __user *)uattr + size;
6795

6796
		for (; addr < end; addr++) {
6797 6798 6799 6800 6801 6802
			ret = get_user(val, addr);
			if (ret)
				return ret;
			if (val)
				goto err_size;
		}
6803
		size = sizeof(*attr);
6804 6805 6806 6807 6808 6809
	}

	ret = copy_from_user(attr, uattr, size);
	if (ret)
		return -EFAULT;

6810 6811 6812 6813
	/* disabled for now */
	if (attr->mmap2)
		return -EINVAL;

6814
	if (attr->__reserved_1)
6815 6816 6817 6818 6819 6820 6821 6822
		return -EINVAL;

	if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
		return -EINVAL;

	if (attr->read_format & ~(PERF_FORMAT_MAX-1))
		return -EINVAL;

6823 6824 6825 6826 6827 6828 6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846 6847 6848 6849 6850
	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;
		}
6851 6852
		/* privileged levels capture (kernel, hv): check permissions */
		if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6853 6854
		    && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
6855
	}
6856

6857
	if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6858
		ret = perf_reg_validate(attr->sample_regs_user);
6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872 6873 6874 6875 6876
		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;
	}
6877

6878 6879 6880 6881 6882 6883 6884 6885 6886
out:
	return ret;

err_size:
	put_user(sizeof(*attr), &uattr->size);
	ret = -E2BIG;
	goto out;
}

6887 6888
static int
perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6889
{
6890
	struct ring_buffer *rb = NULL, *old_rb = NULL;
6891 6892
	int ret = -EINVAL;

6893
	if (!output_event)
6894 6895
		goto set;

6896 6897
	/* don't allow circular references */
	if (event == output_event)
6898 6899
		goto out;

6900 6901 6902 6903 6904 6905 6906
	/*
	 * Don't allow cross-cpu buffers
	 */
	if (output_event->cpu != event->cpu)
		goto out;

	/*
6907
	 * If its not a per-cpu rb, it must be the same task.
6908 6909 6910 6911
	 */
	if (output_event->cpu == -1 && output_event->ctx != event->ctx)
		goto out;

6912
set:
6913
	mutex_lock(&event->mmap_mutex);
6914 6915 6916
	/* Can't redirect output if we've got an active mmap() */
	if (atomic_read(&event->mmap_count))
		goto unlock;
6917

6918 6919
	old_rb = event->rb;

6920
	if (output_event) {
6921 6922 6923
		/* get the rb we want to redirect to */
		rb = ring_buffer_get(output_event);
		if (!rb)
6924
			goto unlock;
6925 6926
	}

6927 6928
	if (old_rb)
		ring_buffer_detach(event, old_rb);
6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944

	if (rb)
		ring_buffer_attach(event, rb);

	rcu_assign_pointer(event->rb, rb);

	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);
	}

6945
	ret = 0;
6946 6947 6948
unlock:
	mutex_unlock(&event->mmap_mutex);

6949 6950 6951 6952
out:
	return ret;
}

T
Thomas Gleixner 已提交
6953
/**
6954
 * sys_perf_event_open - open a performance event, associate it to a task/cpu
I
Ingo Molnar 已提交
6955
 *
6956
 * @attr_uptr:	event_id type attributes for monitoring/sampling
T
Thomas Gleixner 已提交
6957
 * @pid:		target pid
I
Ingo Molnar 已提交
6958
 * @cpu:		target cpu
6959
 * @group_fd:		group leader event fd
T
Thomas Gleixner 已提交
6960
 */
6961 6962
SYSCALL_DEFINE5(perf_event_open,
		struct perf_event_attr __user *, attr_uptr,
6963
		pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
T
Thomas Gleixner 已提交
6964
{
6965 6966
	struct perf_event *group_leader = NULL, *output_event = NULL;
	struct perf_event *event, *sibling;
6967 6968 6969
	struct perf_event_attr attr;
	struct perf_event_context *ctx;
	struct file *event_file = NULL;
6970
	struct fd group = {NULL, 0};
M
Matt Helsley 已提交
6971
	struct task_struct *task = NULL;
6972
	struct pmu *pmu;
6973
	int event_fd;
6974
	int move_group = 0;
6975
	int err;
T
Thomas Gleixner 已提交
6976

6977
	/* for future expandability... */
S
Stephane Eranian 已提交
6978
	if (flags & ~PERF_FLAG_ALL)
6979 6980
		return -EINVAL;

6981 6982 6983
	err = perf_copy_attr(attr_uptr, &attr);
	if (err)
		return err;
6984

6985 6986 6987 6988 6989
	if (!attr.exclude_kernel) {
		if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
	}

6990
	if (attr.freq) {
6991
		if (attr.sample_freq > sysctl_perf_event_sample_rate)
6992 6993 6994
			return -EINVAL;
	}

S
Stephane Eranian 已提交
6995 6996 6997 6998 6999 7000 7001 7002 7003
	/*
	 * 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;

7004
	event_fd = get_unused_fd();
7005 7006 7007
	if (event_fd < 0)
		return event_fd;

7008
	if (group_fd != -1) {
7009 7010
		err = perf_fget_light(group_fd, &group);
		if (err)
7011
			goto err_fd;
7012
		group_leader = group.file->private_data;
7013 7014 7015 7016 7017 7018
		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
7019
	if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7020 7021 7022 7023 7024 7025 7026
		task = find_lively_task_by_vpid(pid);
		if (IS_ERR(task)) {
			err = PTR_ERR(task);
			goto err_group_fd;
		}
	}

7027 7028
	get_online_cpus();

7029 7030
	event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
				 NULL, NULL);
7031 7032
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
7033
		goto err_task;
7034 7035
	}

S
Stephane Eranian 已提交
7036 7037
	if (flags & PERF_FLAG_PID_CGROUP) {
		err = perf_cgroup_connect(pid, event, &attr, group_leader);
7038 7039 7040 7041
		if (err) {
			__free_event(event);
			goto err_task;
		}
S
Stephane Eranian 已提交
7042 7043
	}

7044 7045
	account_event(event);

7046 7047 7048 7049 7050
	/*
	 * Special case software events and allow them to be part of
	 * any hardware group.
	 */
	pmu = event->pmu;
7051 7052 7053 7054 7055 7056 7057 7058 7059 7060 7061 7062 7063 7064 7065 7066 7067 7068 7069 7070 7071 7072 7073

	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;
		}
	}
7074 7075 7076 7077

	/*
	 * Get the target context (task or percpu):
	 */
7078
	ctx = find_get_context(pmu, task, event->cpu);
7079 7080
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7081
		goto err_alloc;
7082 7083
	}

7084 7085 7086 7087 7088
	if (task) {
		put_task_struct(task);
		task = NULL;
	}

I
Ingo Molnar 已提交
7089
	/*
7090
	 * Look up the group leader (we will attach this event to it):
7091
	 */
7092
	if (group_leader) {
7093
		err = -EINVAL;
7094 7095

		/*
I
Ingo Molnar 已提交
7096 7097 7098 7099
		 * Do not allow a recursive hierarchy (this new sibling
		 * becoming part of another group-sibling):
		 */
		if (group_leader->group_leader != group_leader)
7100
			goto err_context;
I
Ingo Molnar 已提交
7101 7102 7103
		/*
		 * Do not allow to attach to a group in a different
		 * task or CPU context:
7104
		 */
7105 7106 7107 7108 7109 7110 7111 7112
		if (move_group) {
			if (group_leader->ctx->type != ctx->type)
				goto err_context;
		} else {
			if (group_leader->ctx != ctx)
				goto err_context;
		}

7113 7114 7115
		/*
		 * Only a group leader can be exclusive or pinned
		 */
7116
		if (attr.exclusive || attr.pinned)
7117
			goto err_context;
7118 7119 7120 7121 7122
	}

	if (output_event) {
		err = perf_event_set_output(event, output_event);
		if (err)
7123
			goto err_context;
7124
	}
T
Thomas Gleixner 已提交
7125

7126 7127 7128
	event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
	if (IS_ERR(event_file)) {
		err = PTR_ERR(event_file);
7129
		goto err_context;
7130
	}
7131

7132 7133 7134 7135
	if (move_group) {
		struct perf_event_context *gctx = group_leader->ctx;

		mutex_lock(&gctx->mutex);
7136
		perf_remove_from_context(group_leader);
J
Jiri Olsa 已提交
7137 7138 7139 7140 7141 7142 7143

		/*
		 * 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);
7144 7145
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7146
			perf_remove_from_context(sibling);
J
Jiri Olsa 已提交
7147
			perf_event__state_init(sibling);
7148 7149 7150 7151
			put_ctx(gctx);
		}
		mutex_unlock(&gctx->mutex);
		put_ctx(gctx);
7152
	}
7153

7154
	WARN_ON_ONCE(ctx->parent_ctx);
7155
	mutex_lock(&ctx->mutex);
7156 7157

	if (move_group) {
7158
		synchronize_rcu();
7159
		perf_install_in_context(ctx, group_leader, event->cpu);
7160 7161 7162
		get_ctx(ctx);
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7163
			perf_install_in_context(ctx, sibling, event->cpu);
7164 7165 7166 7167
			get_ctx(ctx);
		}
	}

7168
	perf_install_in_context(ctx, event, event->cpu);
7169
	perf_unpin_context(ctx);
7170
	mutex_unlock(&ctx->mutex);
7171

7172 7173
	put_online_cpus();

7174
	event->owner = current;
P
Peter Zijlstra 已提交
7175

7176 7177 7178
	mutex_lock(&current->perf_event_mutex);
	list_add_tail(&event->owner_entry, &current->perf_event_list);
	mutex_unlock(&current->perf_event_mutex);
7179

7180 7181 7182 7183
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(event);
7184
	perf_event__id_header_size(event);
7185

7186 7187 7188 7189 7190 7191
	/*
	 * 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().
	 */
7192
	fdput(group);
7193 7194
	fd_install(event_fd, event_file);
	return event_fd;
T
Thomas Gleixner 已提交
7195

7196
err_context:
7197
	perf_unpin_context(ctx);
7198
	put_ctx(ctx);
7199
err_alloc:
7200
	free_event(event);
P
Peter Zijlstra 已提交
7201
err_task:
7202
	put_online_cpus();
P
Peter Zijlstra 已提交
7203 7204
	if (task)
		put_task_struct(task);
7205
err_group_fd:
7206
	fdput(group);
7207 7208
err_fd:
	put_unused_fd(event_fd);
7209
	return err;
T
Thomas Gleixner 已提交
7210 7211
}

7212 7213 7214 7215 7216
/**
 * 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
7217
 * @task: task to profile (NULL for percpu)
7218 7219 7220
 */
struct perf_event *
perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
M
Matt Helsley 已提交
7221
				 struct task_struct *task,
7222 7223
				 perf_overflow_handler_t overflow_handler,
				 void *context)
7224 7225
{
	struct perf_event_context *ctx;
7226
	struct perf_event *event;
7227
	int err;
7228

7229 7230 7231
	/*
	 * Get the target context (task or percpu):
	 */
7232

7233 7234
	event = perf_event_alloc(attr, cpu, task, NULL, NULL,
				 overflow_handler, context);
7235 7236 7237 7238
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
		goto err;
	}
7239

7240 7241
	account_event(event);

M
Matt Helsley 已提交
7242
	ctx = find_get_context(event->pmu, task, cpu);
7243 7244
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7245
		goto err_free;
7246
	}
7247 7248 7249 7250

	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
	perf_install_in_context(ctx, event, cpu);
7251
	perf_unpin_context(ctx);
7252 7253 7254 7255
	mutex_unlock(&ctx->mutex);

	return event;

7256 7257 7258
err_free:
	free_event(event);
err:
7259
	return ERR_PTR(err);
7260
}
7261
EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7262

7263 7264 7265 7266 7267 7268 7269 7270 7271 7272 7273 7274 7275 7276
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) {
		perf_remove_from_context(event);
7277
		unaccount_event_cpu(event, src_cpu);
7278
		put_ctx(src_ctx);
7279
		list_add(&event->migrate_entry, &events);
7280 7281 7282 7283 7284 7285
	}
	mutex_unlock(&src_ctx->mutex);

	synchronize_rcu();

	mutex_lock(&dst_ctx->mutex);
7286 7287
	list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
		list_del(&event->migrate_entry);
7288 7289
		if (event->state >= PERF_EVENT_STATE_OFF)
			event->state = PERF_EVENT_STATE_INACTIVE;
7290
		account_event_cpu(event, dst_cpu);
7291 7292 7293 7294 7295 7296 7297
		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);

7298
static void sync_child_event(struct perf_event *child_event,
7299
			       struct task_struct *child)
7300
{
7301
	struct perf_event *parent_event = child_event->parent;
7302
	u64 child_val;
7303

7304 7305
	if (child_event->attr.inherit_stat)
		perf_event_read_event(child_event, child);
7306

P
Peter Zijlstra 已提交
7307
	child_val = perf_event_count(child_event);
7308 7309 7310 7311

	/*
	 * Add back the child's count to the parent's count:
	 */
7312
	atomic64_add(child_val, &parent_event->child_count);
7313 7314 7315 7316
	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);
7317 7318

	/*
7319
	 * Remove this event from the parent's list
7320
	 */
7321 7322 7323 7324
	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);
7325 7326

	/*
7327
	 * Release the parent event, if this was the last
7328 7329
	 * reference to it.
	 */
7330
	put_event(parent_event);
7331 7332
}

7333
static void
7334 7335
__perf_event_exit_task(struct perf_event *child_event,
			 struct perf_event_context *child_ctx,
7336
			 struct task_struct *child)
7337
{
7338 7339 7340 7341 7342
	if (child_event->parent) {
		raw_spin_lock_irq(&child_ctx->lock);
		perf_group_detach(child_event);
		raw_spin_unlock_irq(&child_ctx->lock);
	}
7343

7344
	perf_remove_from_context(child_event);
7345

7346
	/*
7347
	 * It can happen that the parent exits first, and has events
7348
	 * that are still around due to the child reference. These
7349
	 * events need to be zapped.
7350
	 */
7351
	if (child_event->parent) {
7352 7353
		sync_child_event(child_event, child);
		free_event(child_event);
7354
	}
7355 7356
}

P
Peter Zijlstra 已提交
7357
static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7358
{
7359 7360
	struct perf_event *child_event, *tmp;
	struct perf_event_context *child_ctx;
7361
	unsigned long flags;
7362

P
Peter Zijlstra 已提交
7363
	if (likely(!child->perf_event_ctxp[ctxn])) {
7364
		perf_event_task(child, NULL, 0);
7365
		return;
P
Peter Zijlstra 已提交
7366
	}
7367

7368
	local_irq_save(flags);
7369 7370 7371 7372 7373 7374
	/*
	 * 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.
	 */
7375
	child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7376 7377 7378

	/*
	 * Take the context lock here so that if find_get_context is
7379
	 * reading child->perf_event_ctxp, we wait until it has
7380 7381
	 * incremented the context's refcount before we do put_ctx below.
	 */
7382
	raw_spin_lock(&child_ctx->lock);
7383
	task_ctx_sched_out(child_ctx);
P
Peter Zijlstra 已提交
7384
	child->perf_event_ctxp[ctxn] = NULL;
7385 7386 7387
	/*
	 * If this context is a clone; unclone it so it can't get
	 * swapped to another process while we're removing all
7388
	 * the events from it.
7389 7390
	 */
	unclone_ctx(child_ctx);
7391
	update_context_time(child_ctx);
7392
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7393 7394

	/*
7395 7396 7397
	 * 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
7398
	 */
7399
	perf_event_task(child, child_ctx, 0);
7400

7401 7402 7403
	/*
	 * We can recurse on the same lock type through:
	 *
7404 7405
	 *   __perf_event_exit_task()
	 *     sync_child_event()
7406 7407
	 *       put_event()
	 *         mutex_lock(&ctx->mutex)
7408 7409 7410
	 *
	 * But since its the parent context it won't be the same instance.
	 */
7411
	mutex_lock(&child_ctx->mutex);
7412

7413
again:
7414 7415 7416 7417 7418
	list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
				 group_entry)
		__perf_event_exit_task(child_event, child_ctx, child);

	list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7419
				 group_entry)
7420
		__perf_event_exit_task(child_event, child_ctx, child);
7421 7422

	/*
7423
	 * If the last event was a group event, it will have appended all
7424 7425 7426
	 * its siblings to the list, but we obtained 'tmp' before that which
	 * will still point to the list head terminating the iteration.
	 */
7427 7428
	if (!list_empty(&child_ctx->pinned_groups) ||
	    !list_empty(&child_ctx->flexible_groups))
7429
		goto again;
7430 7431 7432 7433

	mutex_unlock(&child_ctx->mutex);

	put_ctx(child_ctx);
7434 7435
}

P
Peter Zijlstra 已提交
7436 7437 7438 7439 7440
/*
 * 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
7441
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7442 7443
	int ctxn;

P
Peter Zijlstra 已提交
7444 7445 7446 7447 7448 7449 7450 7451 7452 7453 7454 7455 7456 7457 7458
	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
7459 7460 7461 7462
	for_each_task_context_nr(ctxn)
		perf_event_exit_task_context(child, ctxn);
}

7463 7464 7465 7466 7467 7468 7469 7470 7471 7472 7473 7474
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);

7475
	put_event(parent);
7476

7477
	perf_group_detach(event);
7478 7479 7480 7481
	list_del_event(event, ctx);
	free_event(event);
}

7482 7483
/*
 * free an unexposed, unused context as created by inheritance by
P
Peter Zijlstra 已提交
7484
 * perf_event_init_task below, used by fork() in case of fail.
7485
 */
7486
void perf_event_free_task(struct task_struct *task)
7487
{
P
Peter Zijlstra 已提交
7488
	struct perf_event_context *ctx;
7489
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7490
	int ctxn;
7491

P
Peter Zijlstra 已提交
7492 7493 7494 7495
	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;
7496

P
Peter Zijlstra 已提交
7497
		mutex_lock(&ctx->mutex);
7498
again:
P
Peter Zijlstra 已提交
7499 7500 7501
		list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
				group_entry)
			perf_free_event(event, ctx);
7502

P
Peter Zijlstra 已提交
7503 7504 7505
		list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
				group_entry)
			perf_free_event(event, ctx);
7506

P
Peter Zijlstra 已提交
7507 7508 7509
		if (!list_empty(&ctx->pinned_groups) ||
				!list_empty(&ctx->flexible_groups))
			goto again;
7510

P
Peter Zijlstra 已提交
7511
		mutex_unlock(&ctx->mutex);
7512

P
Peter Zijlstra 已提交
7513 7514
		put_ctx(ctx);
	}
7515 7516
}

7517 7518 7519 7520 7521 7522 7523 7524
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
7525 7526 7527 7528 7529 7530 7531 7532 7533 7534 7535 7536
/*
 * 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;
7537
	unsigned long flags;
P
Peter Zijlstra 已提交
7538 7539 7540 7541 7542 7543 7544 7545 7546 7547 7548 7549

	/*
	 * 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,
7550
					   child,
P
Peter Zijlstra 已提交
7551
					   group_leader, parent_event,
7552
				           NULL, NULL);
P
Peter Zijlstra 已提交
7553 7554
	if (IS_ERR(child_event))
		return child_event;
7555 7556 7557 7558 7559 7560

	if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
		free_event(child_event);
		return NULL;
	}

P
Peter Zijlstra 已提交
7561 7562 7563 7564 7565 7566 7567 7568 7569 7570 7571 7572 7573 7574 7575 7576 7577 7578 7579 7580 7581 7582 7583 7584
	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;
7585 7586
	child_event->overflow_handler_context
		= parent_event->overflow_handler_context;
P
Peter Zijlstra 已提交
7587

7588 7589 7590 7591
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(child_event);
7592
	perf_event__id_header_size(child_event);
7593

P
Peter Zijlstra 已提交
7594 7595 7596
	/*
	 * Link it up in the child's context:
	 */
7597
	raw_spin_lock_irqsave(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7598
	add_event_to_ctx(child_event, child_ctx);
7599
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7600 7601 7602 7603 7604 7605 7606 7607 7608 7609 7610 7611 7612 7613 7614 7615 7616 7617 7618 7619 7620 7621 7622 7623 7624 7625 7626 7627 7628 7629 7630 7631 7632

	/*
	 * 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;
7633 7634 7635 7636 7637
}

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
7638
		   struct task_struct *child, int ctxn,
7639 7640 7641
		   int *inherited_all)
{
	int ret;
P
Peter Zijlstra 已提交
7642
	struct perf_event_context *child_ctx;
7643 7644 7645 7646

	if (!event->attr.inherit) {
		*inherited_all = 0;
		return 0;
7647 7648
	}

7649
	child_ctx = child->perf_event_ctxp[ctxn];
7650 7651 7652 7653 7654 7655 7656
	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.
		 */
7657

7658
		child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7659 7660
		if (!child_ctx)
			return -ENOMEM;
7661

P
Peter Zijlstra 已提交
7662
		child->perf_event_ctxp[ctxn] = child_ctx;
7663 7664 7665 7666 7667 7668 7669 7670 7671
	}

	ret = inherit_group(event, parent, parent_ctx,
			    child, child_ctx);

	if (ret)
		*inherited_all = 0;

	return ret;
7672 7673
}

7674
/*
7675
 * Initialize the perf_event context in task_struct
7676
 */
P
Peter Zijlstra 已提交
7677
int perf_event_init_context(struct task_struct *child, int ctxn)
7678
{
7679
	struct perf_event_context *child_ctx, *parent_ctx;
7680 7681
	struct perf_event_context *cloned_ctx;
	struct perf_event *event;
7682
	struct task_struct *parent = current;
7683
	int inherited_all = 1;
7684
	unsigned long flags;
7685
	int ret = 0;
7686

P
Peter Zijlstra 已提交
7687
	if (likely(!parent->perf_event_ctxp[ctxn]))
7688 7689
		return 0;

7690
	/*
7691 7692
	 * If the parent's context is a clone, pin it so it won't get
	 * swapped under us.
7693
	 */
P
Peter Zijlstra 已提交
7694
	parent_ctx = perf_pin_task_context(parent, ctxn);
7695

7696 7697 7698 7699 7700 7701 7702
	/*
	 * 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.
	 */

7703 7704 7705 7706
	/*
	 * Lock the parent list. No need to lock the child - not PID
	 * hashed yet and not running, so nobody can access it.
	 */
7707
	mutex_lock(&parent_ctx->mutex);
7708 7709 7710 7711 7712

	/*
	 * We dont have to disable NMIs - we are only looking at
	 * the list, not manipulating it:
	 */
7713
	list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
P
Peter Zijlstra 已提交
7714 7715
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7716 7717 7718
		if (ret)
			break;
	}
7719

7720 7721 7722 7723 7724 7725 7726 7727 7728
	/*
	 * 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);

7729
	list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
P
Peter Zijlstra 已提交
7730 7731
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7732
		if (ret)
7733
			break;
7734 7735
	}

7736 7737 7738
	raw_spin_lock_irqsave(&parent_ctx->lock, flags);
	parent_ctx->rotate_disable = 0;

P
Peter Zijlstra 已提交
7739
	child_ctx = child->perf_event_ctxp[ctxn];
7740

7741
	if (child_ctx && inherited_all) {
7742 7743 7744
		/*
		 * 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
7745 7746 7747
		 *
		 * Note that if the parent is a clone, the holding of
		 * parent_ctx->lock avoids it from being uncloned.
7748
		 */
P
Peter Zijlstra 已提交
7749
		cloned_ctx = parent_ctx->parent_ctx;
7750 7751
		if (cloned_ctx) {
			child_ctx->parent_ctx = cloned_ctx;
7752
			child_ctx->parent_gen = parent_ctx->parent_gen;
7753 7754 7755 7756 7757
		} else {
			child_ctx->parent_ctx = parent_ctx;
			child_ctx->parent_gen = parent_ctx->generation;
		}
		get_ctx(child_ctx->parent_ctx);
7758 7759
	}

P
Peter Zijlstra 已提交
7760
	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7761
	mutex_unlock(&parent_ctx->mutex);
7762

7763
	perf_unpin_context(parent_ctx);
7764
	put_ctx(parent_ctx);
7765

7766
	return ret;
7767 7768
}

P
Peter Zijlstra 已提交
7769 7770 7771 7772 7773 7774 7775
/*
 * Initialize the perf_event context in task_struct
 */
int perf_event_init_task(struct task_struct *child)
{
	int ctxn, ret;

7776 7777 7778 7779
	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
7780 7781 7782 7783 7784 7785 7786 7787 7788
	for_each_task_context_nr(ctxn) {
		ret = perf_event_init_context(child, ctxn);
		if (ret)
			return ret;
	}

	return 0;
}

7789 7790
static void __init perf_event_init_all_cpus(void)
{
7791
	struct swevent_htable *swhash;
7792 7793 7794
	int cpu;

	for_each_possible_cpu(cpu) {
7795 7796
		swhash = &per_cpu(swevent_htable, cpu);
		mutex_init(&swhash->hlist_mutex);
7797
		INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7798 7799 7800
	}
}

7801
static void perf_event_init_cpu(int cpu)
T
Thomas Gleixner 已提交
7802
{
P
Peter Zijlstra 已提交
7803
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
T
Thomas Gleixner 已提交
7804

7805
	mutex_lock(&swhash->hlist_mutex);
7806
	if (swhash->hlist_refcount > 0) {
7807 7808
		struct swevent_hlist *hlist;

7809 7810 7811
		hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
		WARN_ON(!hlist);
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
7812
	}
7813
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
7814 7815
}

P
Peter Zijlstra 已提交
7816
#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7817
static void perf_pmu_rotate_stop(struct pmu *pmu)
T
Thomas Gleixner 已提交
7818
{
7819 7820 7821 7822 7823 7824 7825
	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
7826
static void __perf_event_exit_context(void *__info)
T
Thomas Gleixner 已提交
7827
{
P
Peter Zijlstra 已提交
7828
	struct perf_event_context *ctx = __info;
7829
	struct perf_event *event, *tmp;
T
Thomas Gleixner 已提交
7830

P
Peter Zijlstra 已提交
7831
	perf_pmu_rotate_stop(ctx->pmu);
7832

7833
	list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7834
		__perf_remove_from_context(event);
7835
	list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7836
		__perf_remove_from_context(event);
T
Thomas Gleixner 已提交
7837
}
P
Peter Zijlstra 已提交
7838 7839 7840 7841 7842 7843 7844 7845 7846

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) {
7847
		ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
P
Peter Zijlstra 已提交
7848 7849 7850 7851 7852 7853 7854 7855

		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);
}

7856
static void perf_event_exit_cpu(int cpu)
T
Thomas Gleixner 已提交
7857
{
7858
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7859

7860 7861 7862
	mutex_lock(&swhash->hlist_mutex);
	swevent_hlist_release(swhash);
	mutex_unlock(&swhash->hlist_mutex);
7863

P
Peter Zijlstra 已提交
7864
	perf_event_exit_cpu_context(cpu);
T
Thomas Gleixner 已提交
7865 7866
}
#else
7867
static inline void perf_event_exit_cpu(int cpu) { }
T
Thomas Gleixner 已提交
7868 7869
#endif

P
Peter Zijlstra 已提交
7870 7871 7872 7873 7874 7875 7876 7877 7878 7879 7880 7881 7882 7883 7884 7885 7886 7887 7888 7889
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,
};

7890
static int
T
Thomas Gleixner 已提交
7891 7892 7893 7894
perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
{
	unsigned int cpu = (long)hcpu;

7895
	switch (action & ~CPU_TASKS_FROZEN) {
T
Thomas Gleixner 已提交
7896 7897

	case CPU_UP_PREPARE:
P
Peter Zijlstra 已提交
7898
	case CPU_DOWN_FAILED:
7899
		perf_event_init_cpu(cpu);
T
Thomas Gleixner 已提交
7900 7901
		break;

P
Peter Zijlstra 已提交
7902
	case CPU_UP_CANCELED:
T
Thomas Gleixner 已提交
7903
	case CPU_DOWN_PREPARE:
7904
		perf_event_exit_cpu(cpu);
T
Thomas Gleixner 已提交
7905 7906 7907 7908 7909 7910 7911 7912
		break;
	default:
		break;
	}

	return NOTIFY_OK;
}

7913
void __init perf_event_init(void)
T
Thomas Gleixner 已提交
7914
{
7915 7916
	int ret;

P
Peter Zijlstra 已提交
7917 7918
	idr_init(&pmu_idr);

7919
	perf_event_init_all_cpus();
7920
	init_srcu_struct(&pmus_srcu);
P
Peter Zijlstra 已提交
7921 7922 7923
	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);
7924 7925
	perf_tp_register();
	perf_cpu_notifier(perf_cpu_notify);
P
Peter Zijlstra 已提交
7926
	register_reboot_notifier(&perf_reboot_notifier);
7927 7928 7929

	ret = init_hw_breakpoint();
	WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7930 7931 7932

	/* do not patch jump label more than once per second */
	jump_label_rate_limit(&perf_sched_events, HZ);
7933 7934 7935 7936 7937 7938 7939

	/*
	 * 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
7940
}
P
Peter Zijlstra 已提交
7941 7942 7943 7944 7945 7946 7947 7948 7949 7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960 7961 7962 7963 7964 7965 7966 7967 7968

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
7969 7970

#ifdef CONFIG_CGROUP_PERF
7971 7972
static struct cgroup_subsys_state *
perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
S
Stephane Eranian 已提交
7973 7974 7975
{
	struct perf_cgroup *jc;

7976
	jc = kzalloc(sizeof(*jc), GFP_KERNEL);
S
Stephane Eranian 已提交
7977 7978 7979 7980 7981 7982 7983 7984 7985 7986 7987 7988
	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;
}

7989
static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
S
Stephane Eranian 已提交
7990
{
7991 7992
	struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);

S
Stephane Eranian 已提交
7993 7994 7995 7996 7997 7998 7999 8000 8001 8002 8003
	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;
}

8004 8005
static void perf_cgroup_attach(struct cgroup_subsys_state *css,
			       struct cgroup_taskset *tset)
S
Stephane Eranian 已提交
8006
{
8007 8008
	struct task_struct *task;

8009
	cgroup_taskset_for_each(task, css, tset)
8010
		task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8011 8012
}

8013 8014
static void perf_cgroup_exit(struct cgroup_subsys_state *css,
			     struct cgroup_subsys_state *old_css,
8015
			     struct task_struct *task)
S
Stephane Eranian 已提交
8016 8017 8018 8019 8020 8021 8022 8023 8024
{
	/*
	 * 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;

8025
	task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8026 8027 8028
}

struct cgroup_subsys perf_subsys = {
8029 8030
	.name		= "perf_event",
	.subsys_id	= perf_subsys_id,
8031 8032
	.css_alloc	= perf_cgroup_css_alloc,
	.css_free	= perf_cgroup_css_free,
8033
	.exit		= perf_cgroup_exit,
8034
	.attach		= perf_cgroup_attach,
S
Stephane Eranian 已提交
8035 8036
};
#endif /* CONFIG_CGROUP_PERF */