core.c 185.5 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)
897 898 899 900 901 902 903
{
	if (ctx->parent_ctx) {
		put_ctx(ctx->parent_ctx);
		ctx->parent_ctx = NULL;
	}
}

904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925
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);
}

926
/*
927
 * If we inherit events we want to return the parent event id
928 929
 * to userspace.
 */
930
static u64 primary_event_id(struct perf_event *event)
931
{
932
	u64 id = event->id;
933

934 935
	if (event->parent)
		id = event->parent->id;
936 937 938 939

	return id;
}

940
/*
941
 * Get the perf_event_context for a task and lock it.
942 943 944
 * This has to cope with with the fact that until it is locked,
 * the context could get moved to another task.
 */
945
static struct perf_event_context *
P
Peter Zijlstra 已提交
946
perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
947
{
948
	struct perf_event_context *ctx;
949

P
Peter Zijlstra 已提交
950
retry:
951 952 953 954 955 956 957 958 959 960 961
	/*
	 * 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
962
	ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
963 964 965 966
	if (ctx) {
		/*
		 * If this context is a clone of another, it might
		 * get swapped for another underneath us by
967
		 * perf_event_task_sched_out, though the
968 969 970 971 972 973
		 * 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.
		 */
974
		raw_spin_lock_irqsave(&ctx->lock, *flags);
P
Peter Zijlstra 已提交
975
		if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
976
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
977 978
			rcu_read_unlock();
			preempt_enable();
979 980
			goto retry;
		}
981 982

		if (!atomic_inc_not_zero(&ctx->refcount)) {
983
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
984 985
			ctx = NULL;
		}
986 987
	}
	rcu_read_unlock();
988
	preempt_enable();
989 990 991 992 993 994 995 996
	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
997 998
static struct perf_event_context *
perf_pin_task_context(struct task_struct *task, int ctxn)
999
{
1000
	struct perf_event_context *ctx;
1001 1002
	unsigned long flags;

P
Peter Zijlstra 已提交
1003
	ctx = perf_lock_task_context(task, ctxn, &flags);
1004 1005
	if (ctx) {
		++ctx->pin_count;
1006
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
1007 1008 1009 1010
	}
	return ctx;
}

1011
static void perf_unpin_context(struct perf_event_context *ctx)
1012 1013 1014
{
	unsigned long flags;

1015
	raw_spin_lock_irqsave(&ctx->lock, flags);
1016
	--ctx->pin_count;
1017
	raw_spin_unlock_irqrestore(&ctx->lock, flags);
1018 1019
}

1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030
/*
 * 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;
}

1031 1032 1033
static u64 perf_event_time(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;
S
Stephane Eranian 已提交
1034 1035 1036 1037

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

1038 1039 1040
	return ctx ? ctx->time : 0;
}

1041 1042
/*
 * Update the total_time_enabled and total_time_running fields for a event.
1043
 * The caller of this function needs to hold the ctx->lock.
1044 1045 1046 1047 1048 1049 1050 1051 1052
 */
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
1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063
	/*
	 * 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))
1064
		run_end = perf_cgroup_event_time(event);
S
Stephane Eranian 已提交
1065 1066
	else if (ctx->is_active)
		run_end = ctx->time;
1067 1068 1069 1070
	else
		run_end = event->tstamp_stopped;

	event->total_time_enabled = run_end - event->tstamp_enabled;
1071 1072 1073 1074

	if (event->state == PERF_EVENT_STATE_INACTIVE)
		run_end = event->tstamp_stopped;
	else
1075
		run_end = perf_event_time(event);
1076 1077

	event->total_time_running = run_end - event->tstamp_running;
S
Stephane Eranian 已提交
1078

1079 1080
}

1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092
/*
 * 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);
}

1093 1094 1095 1096 1097 1098 1099 1100 1101
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;
}

1102
/*
1103
 * Add a event from the lists for its context.
1104 1105
 * Must be called with ctx->mutex and ctx->lock held.
 */
1106
static void
1107
list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1108
{
1109 1110
	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
	event->attach_state |= PERF_ATTACH_CONTEXT;
1111 1112

	/*
1113 1114 1115
	 * 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.
1116
	 */
1117
	if (event->group_leader == event) {
1118 1119
		struct list_head *list;

1120 1121 1122
		if (is_software_event(event))
			event->group_flags |= PERF_GROUP_SOFTWARE;

1123 1124
		list = ctx_group_list(event, ctx);
		list_add_tail(&event->group_entry, list);
P
Peter Zijlstra 已提交
1125
	}
P
Peter Zijlstra 已提交
1126

1127
	if (is_cgroup_event(event))
S
Stephane Eranian 已提交
1128 1129
		ctx->nr_cgroups++;

1130 1131 1132
	if (has_branch_stack(event))
		ctx->nr_branch_stack++;

1133
	list_add_rcu(&event->event_entry, &ctx->event_list);
1134
	if (!ctx->nr_events)
P
Peter Zijlstra 已提交
1135
		perf_pmu_rotate_start(ctx->pmu);
1136 1137
	ctx->nr_events++;
	if (event->attr.inherit_stat)
1138
		ctx->nr_stat++;
1139 1140
}

J
Jiri Olsa 已提交
1141 1142 1143 1144 1145 1146 1147 1148 1149
/*
 * 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;
}

1150 1151 1152 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
/*
 * 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);

1189 1190 1191 1192 1193 1194
	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
1195 1196 1197
	if (sample_type & PERF_SAMPLE_WEIGHT)
		size += sizeof(data->weight);

1198 1199 1200
	if (sample_type & PERF_SAMPLE_READ)
		size += event->read_size;

1201 1202 1203
	if (sample_type & PERF_SAMPLE_DATA_SRC)
		size += sizeof(data->data_src.val);

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

1207 1208 1209 1210 1211 1212 1213 1214 1215
	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;

1216 1217 1218 1219 1220 1221
	if (sample_type & PERF_SAMPLE_TID)
		size += sizeof(data->tid_entry);

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

1222 1223 1224
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		size += sizeof(data->id);

1225 1226 1227 1228 1229 1230 1231 1232 1233
	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);

1234
	event->id_header_size = size;
1235 1236
}

1237 1238
static void perf_group_attach(struct perf_event *event)
{
1239
	struct perf_event *group_leader = event->group_leader, *pos;
1240

P
Peter Zijlstra 已提交
1241 1242 1243 1244 1245 1246
	/*
	 * We can have double attach due to group movement in perf_event_open.
	 */
	if (event->attach_state & PERF_ATTACH_GROUP)
		return;

1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257
	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++;
1258 1259 1260 1261 1262

	perf_event__header_size(group_leader);

	list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
		perf_event__header_size(pos);
1263 1264
}

1265
/*
1266
 * Remove a event from the lists for its context.
1267
 * Must be called with ctx->mutex and ctx->lock held.
1268
 */
1269
static void
1270
list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1271
{
1272
	struct perf_cpu_context *cpuctx;
1273 1274 1275 1276
	/*
	 * We can have double detach due to exit/hot-unplug + close.
	 */
	if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1277
		return;
1278 1279 1280

	event->attach_state &= ~PERF_ATTACH_CONTEXT;

1281
	if (is_cgroup_event(event)) {
S
Stephane Eranian 已提交
1282
		ctx->nr_cgroups--;
1283 1284 1285 1286 1287 1288 1289 1290 1291
		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
1292

1293 1294 1295
	if (has_branch_stack(event))
		ctx->nr_branch_stack--;

1296 1297
	ctx->nr_events--;
	if (event->attr.inherit_stat)
1298
		ctx->nr_stat--;
1299

1300
	list_del_rcu(&event->event_entry);
1301

1302 1303
	if (event->group_leader == event)
		list_del_init(&event->group_entry);
P
Peter Zijlstra 已提交
1304

1305
	update_group_times(event);
1306 1307 1308 1309 1310 1311 1312 1313 1314 1315

	/*
	 * 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;
1316 1317
}

1318
static void perf_group_detach(struct perf_event *event)
1319 1320
{
	struct perf_event *sibling, *tmp;
1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336
	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--;
1337
		goto out;
1338 1339 1340 1341
	}

	if (!list_empty(&event->group_entry))
		list = &event->group_entry;
1342

1343
	/*
1344 1345
	 * If this was a group event with sibling events then
	 * upgrade the siblings to singleton events by adding them
1346
	 * to whatever list we are on.
1347
	 */
1348
	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1349 1350
		if (list)
			list_move_tail(&sibling->group_entry, list);
1351
		sibling->group_leader = sibling;
1352 1353 1354

		/* Inherit group flags from the previous leader */
		sibling->group_flags = event->group_flags;
1355
	}
1356 1357 1358 1359 1360 1361

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);
1362 1363
}

1364 1365 1366
static inline int
event_filter_match(struct perf_event *event)
{
S
Stephane Eranian 已提交
1367 1368
	return (event->cpu == -1 || event->cpu == smp_processor_id())
	    && perf_cgroup_match(event);
1369 1370
}

1371 1372
static void
event_sched_out(struct perf_event *event,
1373
		  struct perf_cpu_context *cpuctx,
1374
		  struct perf_event_context *ctx)
1375
{
1376
	u64 tstamp = perf_event_time(event);
1377 1378 1379 1380 1381 1382 1383 1384 1385
	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
1386
		delta = tstamp - event->tstamp_stopped;
1387
		event->tstamp_running += delta;
1388
		event->tstamp_stopped = tstamp;
1389 1390
	}

1391
	if (event->state != PERF_EVENT_STATE_ACTIVE)
1392
		return;
1393

1394 1395 1396 1397
	event->state = PERF_EVENT_STATE_INACTIVE;
	if (event->pending_disable) {
		event->pending_disable = 0;
		event->state = PERF_EVENT_STATE_OFF;
1398
	}
1399
	event->tstamp_stopped = tstamp;
P
Peter Zijlstra 已提交
1400
	event->pmu->del(event, 0);
1401
	event->oncpu = -1;
1402

1403
	if (!is_software_event(event))
1404 1405
		cpuctx->active_oncpu--;
	ctx->nr_active--;
1406 1407
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq--;
1408
	if (event->attr.exclusive || !cpuctx->active_oncpu)
1409 1410 1411
		cpuctx->exclusive = 0;
}

1412
static void
1413
group_sched_out(struct perf_event *group_event,
1414
		struct perf_cpu_context *cpuctx,
1415
		struct perf_event_context *ctx)
1416
{
1417
	struct perf_event *event;
1418
	int state = group_event->state;
1419

1420
	event_sched_out(group_event, cpuctx, ctx);
1421 1422 1423 1424

	/*
	 * Schedule out siblings (if any):
	 */
1425 1426
	list_for_each_entry(event, &group_event->sibling_list, group_entry)
		event_sched_out(event, cpuctx, ctx);
1427

1428
	if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1429 1430 1431
		cpuctx->exclusive = 0;
}

T
Thomas Gleixner 已提交
1432
/*
1433
 * Cross CPU call to remove a performance event
T
Thomas Gleixner 已提交
1434
 *
1435
 * We disable the event on the hardware level first. After that we
T
Thomas Gleixner 已提交
1436 1437
 * remove it from the context list.
 */
1438
static int __perf_remove_from_context(void *info)
T
Thomas Gleixner 已提交
1439
{
1440 1441
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1442
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
T
Thomas Gleixner 已提交
1443

1444
	raw_spin_lock(&ctx->lock);
1445 1446
	event_sched_out(event, cpuctx, ctx);
	list_del_event(event, ctx);
1447 1448 1449 1450
	if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
		ctx->is_active = 0;
		cpuctx->task_ctx = NULL;
	}
1451
	raw_spin_unlock(&ctx->lock);
1452 1453

	return 0;
T
Thomas Gleixner 已提交
1454 1455 1456 1457
}


/*
1458
 * Remove the event from a task's (or a CPU's) list of events.
T
Thomas Gleixner 已提交
1459
 *
1460
 * CPU events are removed with a smp call. For task events we only
T
Thomas Gleixner 已提交
1461
 * call when the task is on a CPU.
1462
 *
1463 1464
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1465 1466
 * 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.
1467
 * When called from perf_event_exit_task, it's OK because the
1468
 * context has been detached from its task.
T
Thomas Gleixner 已提交
1469
 */
1470
static void perf_remove_from_context(struct perf_event *event)
T
Thomas Gleixner 已提交
1471
{
1472
	struct perf_event_context *ctx = event->ctx;
T
Thomas Gleixner 已提交
1473 1474
	struct task_struct *task = ctx->task;

1475 1476
	lockdep_assert_held(&ctx->mutex);

T
Thomas Gleixner 已提交
1477 1478
	if (!task) {
		/*
1479
		 * Per cpu events are removed via an smp call and
1480
		 * the removal is always successful.
T
Thomas Gleixner 已提交
1481
		 */
1482
		cpu_function_call(event->cpu, __perf_remove_from_context, event);
T
Thomas Gleixner 已提交
1483 1484 1485 1486
		return;
	}

retry:
1487 1488
	if (!task_function_call(task, __perf_remove_from_context, event))
		return;
T
Thomas Gleixner 已提交
1489

1490
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1491
	/*
1492 1493
	 * 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
1494
	 */
1495
	if (ctx->is_active) {
1496
		raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1497 1498 1499 1500
		goto retry;
	}

	/*
1501 1502
	 * 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
1503
	 */
1504
	list_del_event(event, ctx);
1505
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1506 1507
}

1508
/*
1509
 * Cross CPU call to disable a performance event
1510
 */
1511
int __perf_event_disable(void *info)
1512
{
1513 1514
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1515
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1516 1517

	/*
1518 1519
	 * If this is a per-task event, need to check whether this
	 * event's task is the current task on this cpu.
1520 1521 1522
	 *
	 * Can trigger due to concurrent perf_event_context_sched_out()
	 * flipping contexts around.
1523
	 */
1524
	if (ctx->task && cpuctx->task_ctx != ctx)
1525
		return -EINVAL;
1526

1527
	raw_spin_lock(&ctx->lock);
1528 1529

	/*
1530
	 * If the event is on, turn it off.
1531 1532
	 * If it is in error state, leave it in error state.
	 */
1533
	if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1534
		update_context_time(ctx);
S
Stephane Eranian 已提交
1535
		update_cgrp_time_from_event(event);
1536 1537 1538
		update_group_times(event);
		if (event == event->group_leader)
			group_sched_out(event, cpuctx, ctx);
1539
		else
1540 1541
			event_sched_out(event, cpuctx, ctx);
		event->state = PERF_EVENT_STATE_OFF;
1542 1543
	}

1544
	raw_spin_unlock(&ctx->lock);
1545 1546

	return 0;
1547 1548 1549
}

/*
1550
 * Disable a event.
1551
 *
1552 1553
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1554
 * remains valid.  This condition is satisifed when called through
1555 1556 1557 1558
 * 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
1559
 * is the current context on this CPU and preemption is disabled,
1560
 * hence we can't get into perf_event_task_sched_out for this context.
1561
 */
1562
void perf_event_disable(struct perf_event *event)
1563
{
1564
	struct perf_event_context *ctx = event->ctx;
1565 1566 1567 1568
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
1569
		 * Disable the event on the cpu that it's on
1570
		 */
1571
		cpu_function_call(event->cpu, __perf_event_disable, event);
1572 1573 1574
		return;
	}

P
Peter Zijlstra 已提交
1575
retry:
1576 1577
	if (!task_function_call(task, __perf_event_disable, event))
		return;
1578

1579
	raw_spin_lock_irq(&ctx->lock);
1580
	/*
1581
	 * If the event is still active, we need to retry the cross-call.
1582
	 */
1583
	if (event->state == PERF_EVENT_STATE_ACTIVE) {
1584
		raw_spin_unlock_irq(&ctx->lock);
1585 1586 1587 1588 1589
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
1590 1591 1592 1593 1594 1595 1596
		goto retry;
	}

	/*
	 * Since we have the lock this context can't be scheduled
	 * in, so we can change the state safely.
	 */
1597 1598 1599
	if (event->state == PERF_EVENT_STATE_INACTIVE) {
		update_group_times(event);
		event->state = PERF_EVENT_STATE_OFF;
1600
	}
1601
	raw_spin_unlock_irq(&ctx->lock);
1602
}
1603
EXPORT_SYMBOL_GPL(perf_event_disable);
1604

S
Stephane Eranian 已提交
1605 1606 1607 1608 1609 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
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
1640 1641 1642 1643
#define MAX_INTERRUPTS (~0ULL)

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

1644
static int
1645
event_sched_in(struct perf_event *event,
1646
		 struct perf_cpu_context *cpuctx,
1647
		 struct perf_event_context *ctx)
1648
{
1649 1650
	u64 tstamp = perf_event_time(event);

1651
	if (event->state <= PERF_EVENT_STATE_OFF)
1652 1653
		return 0;

1654
	event->state = PERF_EVENT_STATE_ACTIVE;
1655
	event->oncpu = smp_processor_id();
P
Peter Zijlstra 已提交
1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666

	/*
	 * 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;
	}

1667 1668 1669 1670 1671
	/*
	 * The new state must be visible before we turn it on in the hardware:
	 */
	smp_wmb();

P
Peter Zijlstra 已提交
1672
	if (event->pmu->add(event, PERF_EF_START)) {
1673 1674
		event->state = PERF_EVENT_STATE_INACTIVE;
		event->oncpu = -1;
1675 1676 1677
		return -EAGAIN;
	}

1678
	event->tstamp_running += tstamp - event->tstamp_stopped;
1679

S
Stephane Eranian 已提交
1680
	perf_set_shadow_time(event, ctx, tstamp);
1681

1682
	if (!is_software_event(event))
1683
		cpuctx->active_oncpu++;
1684
	ctx->nr_active++;
1685 1686
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq++;
1687

1688
	if (event->attr.exclusive)
1689 1690
		cpuctx->exclusive = 1;

1691 1692 1693
	return 0;
}

1694
static int
1695
group_sched_in(struct perf_event *group_event,
1696
	       struct perf_cpu_context *cpuctx,
1697
	       struct perf_event_context *ctx)
1698
{
1699
	struct perf_event *event, *partial_group = NULL;
P
Peter Zijlstra 已提交
1700
	struct pmu *pmu = group_event->pmu;
1701 1702
	u64 now = ctx->time;
	bool simulate = false;
1703

1704
	if (group_event->state == PERF_EVENT_STATE_OFF)
1705 1706
		return 0;

P
Peter Zijlstra 已提交
1707
	pmu->start_txn(pmu);
1708

1709
	if (event_sched_in(group_event, cpuctx, ctx)) {
P
Peter Zijlstra 已提交
1710
		pmu->cancel_txn(pmu);
1711
		perf_cpu_hrtimer_restart(cpuctx);
1712
		return -EAGAIN;
1713
	}
1714 1715 1716 1717

	/*
	 * Schedule in siblings as one group (if any):
	 */
1718
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1719
		if (event_sched_in(event, cpuctx, ctx)) {
1720
			partial_group = event;
1721 1722 1723 1724
			goto group_error;
		}
	}

1725
	if (!pmu->commit_txn(pmu))
1726
		return 0;
1727

1728 1729 1730 1731
group_error:
	/*
	 * Groups can be scheduled in as one unit only, so undo any
	 * partial group before returning:
1732 1733 1734 1735 1736 1737 1738 1739 1740 1741
	 * 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.
1742
	 */
1743 1744
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
		if (event == partial_group)
1745 1746 1747 1748 1749 1750 1751 1752
			simulate = true;

		if (simulate) {
			event->tstamp_running += now - event->tstamp_stopped;
			event->tstamp_stopped = now;
		} else {
			event_sched_out(event, cpuctx, ctx);
		}
1753
	}
1754
	event_sched_out(group_event, cpuctx, ctx);
1755

P
Peter Zijlstra 已提交
1756
	pmu->cancel_txn(pmu);
1757

1758 1759
	perf_cpu_hrtimer_restart(cpuctx);

1760 1761 1762
	return -EAGAIN;
}

1763
/*
1764
 * Work out whether we can put this event group on the CPU now.
1765
 */
1766
static int group_can_go_on(struct perf_event *event,
1767 1768 1769 1770
			   struct perf_cpu_context *cpuctx,
			   int can_add_hw)
{
	/*
1771
	 * Groups consisting entirely of software events can always go on.
1772
	 */
1773
	if (event->group_flags & PERF_GROUP_SOFTWARE)
1774 1775 1776
		return 1;
	/*
	 * If an exclusive group is already on, no other hardware
1777
	 * events can go on.
1778 1779 1780 1781 1782
	 */
	if (cpuctx->exclusive)
		return 0;
	/*
	 * If this group is exclusive and there are already
1783
	 * events on the CPU, it can't go on.
1784
	 */
1785
	if (event->attr.exclusive && cpuctx->active_oncpu)
1786 1787 1788 1789 1790 1791 1792 1793
		return 0;
	/*
	 * Otherwise, try to add it if all previous groups were able
	 * to go on.
	 */
	return can_add_hw;
}

1794 1795
static void add_event_to_ctx(struct perf_event *event,
			       struct perf_event_context *ctx)
1796
{
1797 1798
	u64 tstamp = perf_event_time(event);

1799
	list_add_event(event, ctx);
1800
	perf_group_attach(event);
1801 1802 1803
	event->tstamp_enabled = tstamp;
	event->tstamp_running = tstamp;
	event->tstamp_stopped = tstamp;
1804 1805
}

1806 1807 1808 1809 1810 1811
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);
1812

1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824
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
1825
/*
1826
 * Cross CPU call to install and enable a performance event
1827 1828
 *
 * Must be called with ctx->mutex held
T
Thomas Gleixner 已提交
1829
 */
1830
static int  __perf_install_in_context(void *info)
T
Thomas Gleixner 已提交
1831
{
1832 1833
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1834
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1835 1836 1837
	struct perf_event_context *task_ctx = cpuctx->task_ctx;
	struct task_struct *task = current;

1838
	perf_ctx_lock(cpuctx, task_ctx);
1839
	perf_pmu_disable(cpuctx->ctx.pmu);
T
Thomas Gleixner 已提交
1840 1841

	/*
1842
	 * If there was an active task_ctx schedule it out.
T
Thomas Gleixner 已提交
1843
	 */
1844
	if (task_ctx)
1845
		task_ctx_sched_out(task_ctx);
1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859

	/*
	 * 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;
1860 1861
		task = task_ctx->task;
	}
1862

1863
	cpu_ctx_sched_out(cpuctx, EVENT_ALL);
T
Thomas Gleixner 已提交
1864

1865
	update_context_time(ctx);
S
Stephane Eranian 已提交
1866 1867 1868 1869 1870 1871
	/*
	 * 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
1872

1873
	add_event_to_ctx(event, ctx);
T
Thomas Gleixner 已提交
1874

1875
	/*
1876
	 * Schedule everything back in
1877
	 */
1878
	perf_event_sched_in(cpuctx, task_ctx, task);
1879 1880 1881

	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, task_ctx);
1882 1883

	return 0;
T
Thomas Gleixner 已提交
1884 1885 1886
}

/*
1887
 * Attach a performance event to a context
T
Thomas Gleixner 已提交
1888
 *
1889 1890
 * First we add the event to the list with the hardware enable bit
 * in event->hw_config cleared.
T
Thomas Gleixner 已提交
1891
 *
1892
 * If the event is attached to a task which is on a CPU we use a smp
T
Thomas Gleixner 已提交
1893 1894 1895 1896
 * 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
1897 1898
perf_install_in_context(struct perf_event_context *ctx,
			struct perf_event *event,
T
Thomas Gleixner 已提交
1899 1900 1901 1902
			int cpu)
{
	struct task_struct *task = ctx->task;

1903 1904
	lockdep_assert_held(&ctx->mutex);

1905
	event->ctx = ctx;
1906 1907
	if (event->cpu != -1)
		event->cpu = cpu;
1908

T
Thomas Gleixner 已提交
1909 1910
	if (!task) {
		/*
1911
		 * Per cpu events are installed via an smp call and
1912
		 * the install is always successful.
T
Thomas Gleixner 已提交
1913
		 */
1914
		cpu_function_call(cpu, __perf_install_in_context, event);
T
Thomas Gleixner 已提交
1915 1916 1917 1918
		return;
	}

retry:
1919 1920
	if (!task_function_call(task, __perf_install_in_context, event))
		return;
T
Thomas Gleixner 已提交
1921

1922
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1923
	/*
1924 1925
	 * 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
1926
	 */
1927
	if (ctx->is_active) {
1928
		raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1929 1930 1931 1932
		goto retry;
	}

	/*
1933 1934
	 * 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
1935
	 */
1936
	add_event_to_ctx(event, ctx);
1937
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1938 1939
}

1940
/*
1941
 * Put a event into inactive state and update time fields.
1942 1943 1944 1945 1946 1947
 * 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.
 */
1948
static void __perf_event_mark_enabled(struct perf_event *event)
1949
{
1950
	struct perf_event *sub;
1951
	u64 tstamp = perf_event_time(event);
1952

1953
	event->state = PERF_EVENT_STATE_INACTIVE;
1954
	event->tstamp_enabled = tstamp - event->total_time_enabled;
P
Peter Zijlstra 已提交
1955
	list_for_each_entry(sub, &event->sibling_list, group_entry) {
1956 1957
		if (sub->state >= PERF_EVENT_STATE_INACTIVE)
			sub->tstamp_enabled = tstamp - sub->total_time_enabled;
P
Peter Zijlstra 已提交
1958
	}
1959 1960
}

1961
/*
1962
 * Cross CPU call to enable a performance event
1963
 */
1964
static int __perf_event_enable(void *info)
1965
{
1966 1967 1968
	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
1969
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1970
	int err;
1971

1972 1973 1974 1975 1976 1977 1978 1979 1980 1981
	/*
	 * 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)
1982
		return -EINVAL;
1983

1984
	raw_spin_lock(&ctx->lock);
1985
	update_context_time(ctx);
1986

1987
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
1988
		goto unlock;
S
Stephane Eranian 已提交
1989 1990 1991 1992

	/*
	 * set current task's cgroup time reference point
	 */
1993
	perf_cgroup_set_timestamp(current, ctx);
S
Stephane Eranian 已提交
1994

1995
	__perf_event_mark_enabled(event);
1996

S
Stephane Eranian 已提交
1997 1998 1999
	if (!event_filter_match(event)) {
		if (is_cgroup_event(event))
			perf_cgroup_defer_enabled(event);
2000
		goto unlock;
S
Stephane Eranian 已提交
2001
	}
2002

2003
	/*
2004
	 * If the event is in a group and isn't the group leader,
2005
	 * then don't put it on unless the group is on.
2006
	 */
2007
	if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2008
		goto unlock;
2009

2010
	if (!group_can_go_on(event, cpuctx, 1)) {
2011
		err = -EEXIST;
2012
	} else {
2013
		if (event == leader)
2014
			err = group_sched_in(event, cpuctx, ctx);
2015
		else
2016
			err = event_sched_in(event, cpuctx, ctx);
2017
	}
2018 2019 2020

	if (err) {
		/*
2021
		 * If this event can't go on and it's part of a
2022 2023
		 * group, then the whole group has to come off.
		 */
2024
		if (leader != event) {
2025
			group_sched_out(leader, cpuctx, ctx);
2026 2027
			perf_cpu_hrtimer_restart(cpuctx);
		}
2028
		if (leader->attr.pinned) {
2029
			update_group_times(leader);
2030
			leader->state = PERF_EVENT_STATE_ERROR;
2031
		}
2032 2033
	}

P
Peter Zijlstra 已提交
2034
unlock:
2035
	raw_spin_unlock(&ctx->lock);
2036 2037

	return 0;
2038 2039 2040
}

/*
2041
 * Enable a event.
2042
 *
2043 2044
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
2045
 * remains valid.  This condition is satisfied when called through
2046 2047
 * perf_event_for_each_child or perf_event_for_each as described
 * for perf_event_disable.
2048
 */
2049
void perf_event_enable(struct perf_event *event)
2050
{
2051
	struct perf_event_context *ctx = event->ctx;
2052 2053 2054 2055
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
2056
		 * Enable the event on the cpu that it's on
2057
		 */
2058
		cpu_function_call(event->cpu, __perf_event_enable, event);
2059 2060 2061
		return;
	}

2062
	raw_spin_lock_irq(&ctx->lock);
2063
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2064 2065 2066
		goto out;

	/*
2067 2068
	 * If the event is in error state, clear that first.
	 * That way, if we see the event in error state below, we
2069 2070 2071 2072
	 * know that it has gone back into error state, as distinct
	 * from the task having been scheduled away before the
	 * cross-call arrived.
	 */
2073 2074
	if (event->state == PERF_EVENT_STATE_ERROR)
		event->state = PERF_EVENT_STATE_OFF;
2075

P
Peter Zijlstra 已提交
2076
retry:
2077
	if (!ctx->is_active) {
2078
		__perf_event_mark_enabled(event);
2079 2080 2081
		goto out;
	}

2082
	raw_spin_unlock_irq(&ctx->lock);
2083 2084 2085

	if (!task_function_call(task, __perf_event_enable, event))
		return;
2086

2087
	raw_spin_lock_irq(&ctx->lock);
2088 2089

	/*
2090
	 * If the context is active and the event is still off,
2091 2092
	 * we need to retry the cross-call.
	 */
2093 2094 2095 2096 2097 2098
	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;
2099
		goto retry;
2100
	}
2101

P
Peter Zijlstra 已提交
2102
out:
2103
	raw_spin_unlock_irq(&ctx->lock);
2104
}
2105
EXPORT_SYMBOL_GPL(perf_event_enable);
2106

2107
int perf_event_refresh(struct perf_event *event, int refresh)
2108
{
2109
	/*
2110
	 * not supported on inherited events
2111
	 */
2112
	if (event->attr.inherit || !is_sampling_event(event))
2113 2114
		return -EINVAL;

2115 2116
	atomic_add(refresh, &event->event_limit);
	perf_event_enable(event);
2117 2118

	return 0;
2119
}
2120
EXPORT_SYMBOL_GPL(perf_event_refresh);
2121

2122 2123 2124
static void ctx_sched_out(struct perf_event_context *ctx,
			  struct perf_cpu_context *cpuctx,
			  enum event_type_t event_type)
2125
{
2126
	struct perf_event *event;
2127
	int is_active = ctx->is_active;
2128

2129
	ctx->is_active &= ~event_type;
2130
	if (likely(!ctx->nr_events))
2131 2132
		return;

2133
	update_context_time(ctx);
S
Stephane Eranian 已提交
2134
	update_cgrp_time_from_cpuctx(cpuctx);
2135
	if (!ctx->nr_active)
2136
		return;
2137

P
Peter Zijlstra 已提交
2138
	perf_pmu_disable(ctx->pmu);
2139
	if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2140 2141
		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2142
	}
2143

2144
	if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2145
		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2146
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2147
	}
P
Peter Zijlstra 已提交
2148
	perf_pmu_enable(ctx->pmu);
2149 2150
}

2151 2152 2153
/*
 * Test whether two contexts are equivalent, i.e. whether they
 * have both been cloned from the same version of the same context
2154 2155 2156 2157
 * and they both have the same number of enabled events.
 * If the number of enabled events is the same, then the set
 * of enabled events should be the same, because these are both
 * inherited contexts, therefore we can't access individual events
2158
 * in them directly with an fd; we can only enable/disable all
2159
 * events via prctl, or enable/disable all events in a family
2160 2161
 * via ioctl, which will have the same effect on both contexts.
 */
2162 2163
static int context_equiv(struct perf_event_context *ctx1,
			 struct perf_event_context *ctx2)
2164 2165
{
	return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
2166
		&& ctx1->parent_gen == ctx2->parent_gen
2167
		&& !ctx1->pin_count && !ctx2->pin_count;
2168 2169
}

2170 2171
static void __perf_event_sync_stat(struct perf_event *event,
				     struct perf_event *next_event)
2172 2173 2174
{
	u64 value;

2175
	if (!event->attr.inherit_stat)
2176 2177 2178
		return;

	/*
2179
	 * Update the event value, we cannot use perf_event_read()
2180 2181
	 * because we're in the middle of a context switch and have IRQs
	 * disabled, which upsets smp_call_function_single(), however
2182
	 * we know the event must be on the current CPU, therefore we
2183 2184
	 * don't need to use it.
	 */
2185 2186
	switch (event->state) {
	case PERF_EVENT_STATE_ACTIVE:
2187 2188
		event->pmu->read(event);
		/* fall-through */
2189

2190 2191
	case PERF_EVENT_STATE_INACTIVE:
		update_event_times(event);
2192 2193 2194 2195 2196 2197 2198
		break;

	default:
		break;
	}

	/*
2199
	 * In order to keep per-task stats reliable we need to flip the event
2200 2201
	 * values when we flip the contexts.
	 */
2202 2203 2204
	value = local64_read(&next_event->count);
	value = local64_xchg(&event->count, value);
	local64_set(&next_event->count, value);
2205

2206 2207
	swap(event->total_time_enabled, next_event->total_time_enabled);
	swap(event->total_time_running, next_event->total_time_running);
2208

2209
	/*
2210
	 * Since we swizzled the values, update the user visible data too.
2211
	 */
2212 2213
	perf_event_update_userpage(event);
	perf_event_update_userpage(next_event);
2214 2215 2216 2217 2218
}

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

2219 2220
static void perf_event_sync_stat(struct perf_event_context *ctx,
				   struct perf_event_context *next_ctx)
2221
{
2222
	struct perf_event *event, *next_event;
2223 2224 2225 2226

	if (!ctx->nr_stat)
		return;

2227 2228
	update_context_time(ctx);

2229 2230
	event = list_first_entry(&ctx->event_list,
				   struct perf_event, event_entry);
2231

2232 2233
	next_event = list_first_entry(&next_ctx->event_list,
					struct perf_event, event_entry);
2234

2235 2236
	while (&event->event_entry != &ctx->event_list &&
	       &next_event->event_entry != &next_ctx->event_list) {
2237

2238
		__perf_event_sync_stat(event, next_event);
2239

2240 2241
		event = list_next_entry(event, event_entry);
		next_event = list_next_entry(next_event, event_entry);
2242 2243 2244
	}
}

2245 2246
static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
					 struct task_struct *next)
T
Thomas Gleixner 已提交
2247
{
P
Peter Zijlstra 已提交
2248
	struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2249 2250
	struct perf_event_context *next_ctx;
	struct perf_event_context *parent;
P
Peter Zijlstra 已提交
2251
	struct perf_cpu_context *cpuctx;
2252
	int do_switch = 1;
T
Thomas Gleixner 已提交
2253

P
Peter Zijlstra 已提交
2254 2255
	if (likely(!ctx))
		return;
2256

P
Peter Zijlstra 已提交
2257 2258
	cpuctx = __get_cpu_context(ctx);
	if (!cpuctx->task_ctx)
T
Thomas Gleixner 已提交
2259 2260
		return;

2261 2262
	rcu_read_lock();
	parent = rcu_dereference(ctx->parent_ctx);
P
Peter Zijlstra 已提交
2263
	next_ctx = next->perf_event_ctxp[ctxn];
2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274
	if (parent && next_ctx &&
	    rcu_dereference(next_ctx->parent_ctx) == parent) {
		/*
		 * 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.
		 */
2275 2276
		raw_spin_lock(&ctx->lock);
		raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2277
		if (context_equiv(ctx, next_ctx)) {
2278 2279
			/*
			 * XXX do we need a memory barrier of sorts
2280
			 * wrt to rcu_dereference() of perf_event_ctxp
2281
			 */
P
Peter Zijlstra 已提交
2282 2283
			task->perf_event_ctxp[ctxn] = next_ctx;
			next->perf_event_ctxp[ctxn] = ctx;
2284 2285 2286
			ctx->task = next;
			next_ctx->task = task;
			do_switch = 0;
2287

2288
			perf_event_sync_stat(ctx, next_ctx);
2289
		}
2290 2291
		raw_spin_unlock(&next_ctx->lock);
		raw_spin_unlock(&ctx->lock);
2292
	}
2293
	rcu_read_unlock();
2294

2295
	if (do_switch) {
2296
		raw_spin_lock(&ctx->lock);
2297
		ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2298
		cpuctx->task_ctx = NULL;
2299
		raw_spin_unlock(&ctx->lock);
2300
	}
T
Thomas Gleixner 已提交
2301 2302
}

P
Peter Zijlstra 已提交
2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316
#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.
 */
2317 2318
void __perf_event_task_sched_out(struct task_struct *task,
				 struct task_struct *next)
P
Peter Zijlstra 已提交
2319 2320 2321 2322 2323
{
	int ctxn;

	for_each_task_context_nr(ctxn)
		perf_event_context_sched_out(task, ctxn, next);
S
Stephane Eranian 已提交
2324 2325 2326 2327 2328 2329 2330

	/*
	 * 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)))
2331
		perf_cgroup_sched_out(task, next);
P
Peter Zijlstra 已提交
2332 2333
}

2334
static void task_ctx_sched_out(struct perf_event_context *ctx)
2335
{
P
Peter Zijlstra 已提交
2336
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2337

2338 2339
	if (!cpuctx->task_ctx)
		return;
2340 2341 2342 2343

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

2344
	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2345 2346 2347
	cpuctx->task_ctx = NULL;
}

2348 2349 2350 2351 2352 2353 2354
/*
 * 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);
2355 2356
}

2357
static void
2358
ctx_pinned_sched_in(struct perf_event_context *ctx,
2359
		    struct perf_cpu_context *cpuctx)
T
Thomas Gleixner 已提交
2360
{
2361
	struct perf_event *event;
T
Thomas Gleixner 已提交
2362

2363 2364
	list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
		if (event->state <= PERF_EVENT_STATE_OFF)
2365
			continue;
2366
		if (!event_filter_match(event))
2367 2368
			continue;

S
Stephane Eranian 已提交
2369 2370 2371 2372
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

2373
		if (group_can_go_on(event, cpuctx, 1))
2374
			group_sched_in(event, cpuctx, ctx);
2375 2376 2377 2378 2379

		/*
		 * If this pinned group hasn't been scheduled,
		 * put it in error state.
		 */
2380 2381 2382
		if (event->state == PERF_EVENT_STATE_INACTIVE) {
			update_group_times(event);
			event->state = PERF_EVENT_STATE_ERROR;
2383
		}
2384
	}
2385 2386 2387 2388
}

static void
ctx_flexible_sched_in(struct perf_event_context *ctx,
2389
		      struct perf_cpu_context *cpuctx)
2390 2391 2392
{
	struct perf_event *event;
	int can_add_hw = 1;
2393

2394 2395 2396
	list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
		/* Ignore events in OFF or ERROR state */
		if (event->state <= PERF_EVENT_STATE_OFF)
2397
			continue;
2398 2399
		/*
		 * Listen to the 'cpu' scheduling filter constraint
2400
		 * of events:
2401
		 */
2402
		if (!event_filter_match(event))
T
Thomas Gleixner 已提交
2403 2404
			continue;

S
Stephane Eranian 已提交
2405 2406 2407 2408
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

P
Peter Zijlstra 已提交
2409
		if (group_can_go_on(event, cpuctx, can_add_hw)) {
2410
			if (group_sched_in(event, cpuctx, ctx))
2411
				can_add_hw = 0;
P
Peter Zijlstra 已提交
2412
		}
T
Thomas Gleixner 已提交
2413
	}
2414 2415 2416 2417 2418
}

static void
ctx_sched_in(struct perf_event_context *ctx,
	     struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2419 2420
	     enum event_type_t event_type,
	     struct task_struct *task)
2421
{
S
Stephane Eranian 已提交
2422
	u64 now;
2423
	int is_active = ctx->is_active;
S
Stephane Eranian 已提交
2424

2425
	ctx->is_active |= event_type;
2426
	if (likely(!ctx->nr_events))
2427
		return;
2428

S
Stephane Eranian 已提交
2429 2430
	now = perf_clock();
	ctx->timestamp = now;
2431
	perf_cgroup_set_timestamp(task, ctx);
2432 2433 2434 2435
	/*
	 * First go through the list and put on any pinned groups
	 * in order to give them the best chance of going on.
	 */
2436
	if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2437
		ctx_pinned_sched_in(ctx, cpuctx);
2438 2439

	/* Then walk through the lower prio flexible groups */
2440
	if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2441
		ctx_flexible_sched_in(ctx, cpuctx);
2442 2443
}

2444
static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2445 2446
			     enum event_type_t event_type,
			     struct task_struct *task)
2447 2448 2449
{
	struct perf_event_context *ctx = &cpuctx->ctx;

S
Stephane Eranian 已提交
2450
	ctx_sched_in(ctx, cpuctx, event_type, task);
2451 2452
}

S
Stephane Eranian 已提交
2453 2454
static void perf_event_context_sched_in(struct perf_event_context *ctx,
					struct task_struct *task)
2455
{
P
Peter Zijlstra 已提交
2456
	struct perf_cpu_context *cpuctx;
2457

P
Peter Zijlstra 已提交
2458
	cpuctx = __get_cpu_context(ctx);
2459 2460 2461
	if (cpuctx->task_ctx == ctx)
		return;

2462
	perf_ctx_lock(cpuctx, ctx);
P
Peter Zijlstra 已提交
2463
	perf_pmu_disable(ctx->pmu);
2464 2465 2466 2467 2468 2469 2470
	/*
	 * 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);

2471 2472
	if (ctx->nr_events)
		cpuctx->task_ctx = ctx;
2473

2474 2475
	perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);

2476 2477 2478
	perf_pmu_enable(ctx->pmu);
	perf_ctx_unlock(cpuctx, ctx);

2479 2480 2481 2482
	/*
	 * 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
2483
	perf_pmu_rotate_start(ctx->pmu);
2484 2485
}

2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 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
/*
 * 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
2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556
/*
 * 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.
 */
2557 2558
void __perf_event_task_sched_in(struct task_struct *prev,
				struct task_struct *task)
P
Peter Zijlstra 已提交
2559 2560 2561 2562 2563 2564 2565 2566 2567
{
	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
2568
		perf_event_context_sched_in(ctx, task);
P
Peter Zijlstra 已提交
2569
	}
S
Stephane Eranian 已提交
2570 2571 2572 2573 2574 2575
	/*
	 * 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)))
2576
		perf_cgroup_sched_in(prev, task);
2577 2578 2579 2580

	/* check for system-wide branch_stack events */
	if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
		perf_branch_stack_sched_in(prev, task);
2581 2582
}

2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609
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.
	 */
2610
#define REDUCE_FLS(a, b)		\
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 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649
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;
	}

2650 2651 2652
	if (!divisor)
		return dividend;

2653 2654 2655
	return div64_u64(dividend, divisor);
}

2656 2657 2658
static DEFINE_PER_CPU(int, perf_throttled_count);
static DEFINE_PER_CPU(u64, perf_throttled_seq);

2659
static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2660
{
2661
	struct hw_perf_event *hwc = &event->hw;
2662
	s64 period, sample_period;
2663 2664
	s64 delta;

2665
	period = perf_calculate_period(event, nsec, count);
2666 2667 2668 2669 2670 2671 2672 2673 2674 2675

	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;
2676

2677
	if (local64_read(&hwc->period_left) > 8*sample_period) {
2678 2679 2680
		if (disable)
			event->pmu->stop(event, PERF_EF_UPDATE);

2681
		local64_set(&hwc->period_left, 0);
2682 2683 2684

		if (disable)
			event->pmu->start(event, PERF_EF_RELOAD);
2685
	}
2686 2687
}

2688 2689 2690 2691 2692 2693 2694
/*
 * 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)
2695
{
2696 2697
	struct perf_event *event;
	struct hw_perf_event *hwc;
2698
	u64 now, period = TICK_NSEC;
2699
	s64 delta;
2700

2701 2702 2703 2704 2705 2706
	/*
	 * 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))
2707 2708
		return;

2709
	raw_spin_lock(&ctx->lock);
2710
	perf_pmu_disable(ctx->pmu);
2711

2712
	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2713
		if (event->state != PERF_EVENT_STATE_ACTIVE)
2714 2715
			continue;

2716
		if (!event_filter_match(event))
2717 2718
			continue;

2719
		hwc = &event->hw;
2720

2721
		if (hwc->interrupts == MAX_INTERRUPTS) {
2722
			hwc->interrupts = 0;
2723
			perf_log_throttle(event, 1);
P
Peter Zijlstra 已提交
2724
			event->pmu->start(event, 0);
2725 2726
		}

2727
		if (!event->attr.freq || !event->attr.sample_freq)
2728 2729
			continue;

2730 2731 2732 2733 2734
		/*
		 * stop the event and update event->count
		 */
		event->pmu->stop(event, PERF_EF_UPDATE);

2735
		now = local64_read(&event->count);
2736 2737
		delta = now - hwc->freq_count_stamp;
		hwc->freq_count_stamp = now;
2738

2739 2740 2741
		/*
		 * restart the event
		 * reload only if value has changed
2742 2743 2744
		 * we have stopped the event so tell that
		 * to perf_adjust_period() to avoid stopping it
		 * twice.
2745
		 */
2746
		if (delta > 0)
2747
			perf_adjust_period(event, period, delta, false);
2748 2749

		event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2750
	}
2751

2752
	perf_pmu_enable(ctx->pmu);
2753
	raw_spin_unlock(&ctx->lock);
2754 2755
}

2756
/*
2757
 * Round-robin a context's events:
2758
 */
2759
static void rotate_ctx(struct perf_event_context *ctx)
T
Thomas Gleixner 已提交
2760
{
2761 2762 2763 2764 2765 2766
	/*
	 * 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);
2767 2768
}

2769
/*
2770 2771 2772
 * 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.
2773
 */
2774
static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2775
{
P
Peter Zijlstra 已提交
2776
	struct perf_event_context *ctx = NULL;
2777
	int rotate = 0, remove = 1;
2778

2779
	if (cpuctx->ctx.nr_events) {
2780
		remove = 0;
2781 2782 2783
		if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
			rotate = 1;
	}
2784

P
Peter Zijlstra 已提交
2785
	ctx = cpuctx->task_ctx;
2786
	if (ctx && ctx->nr_events) {
2787
		remove = 0;
2788 2789 2790
		if (ctx->nr_events != ctx->nr_active)
			rotate = 1;
	}
2791

2792
	if (!rotate)
2793 2794
		goto done;

2795
	perf_ctx_lock(cpuctx, cpuctx->task_ctx);
P
Peter Zijlstra 已提交
2796
	perf_pmu_disable(cpuctx->ctx.pmu);
2797

2798 2799 2800
	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
	if (ctx)
		ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
T
Thomas Gleixner 已提交
2801

2802 2803 2804
	rotate_ctx(&cpuctx->ctx);
	if (ctx)
		rotate_ctx(ctx);
2805

2806
	perf_event_sched_in(cpuctx, ctx, current);
2807

2808 2809
	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2810
done:
2811 2812
	if (remove)
		list_del_init(&cpuctx->rotation_list);
2813 2814

	return rotate;
2815 2816
}

2817 2818 2819
#ifdef CONFIG_NO_HZ_FULL
bool perf_event_can_stop_tick(void)
{
2820
	if (atomic_read(&nr_freq_events) ||
2821
	    __this_cpu_read(perf_throttled_count))
2822
		return false;
2823 2824
	else
		return true;
2825 2826 2827
}
#endif

2828 2829 2830 2831
void perf_event_task_tick(void)
{
	struct list_head *head = &__get_cpu_var(rotation_list);
	struct perf_cpu_context *cpuctx, *tmp;
2832 2833
	struct perf_event_context *ctx;
	int throttled;
2834

2835 2836
	WARN_ON(!irqs_disabled());

2837 2838 2839
	__this_cpu_inc(perf_throttled_seq);
	throttled = __this_cpu_xchg(perf_throttled_count, 0);

2840
	list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2841 2842 2843 2844 2845 2846
		ctx = &cpuctx->ctx;
		perf_adjust_freq_unthr_context(ctx, throttled);

		ctx = cpuctx->task_ctx;
		if (ctx)
			perf_adjust_freq_unthr_context(ctx, throttled);
2847
	}
T
Thomas Gleixner 已提交
2848 2849
}

2850 2851 2852 2853 2854 2855 2856 2857 2858 2859
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;

2860
	__perf_event_mark_enabled(event);
2861 2862 2863 2864

	return 1;
}

2865
/*
2866
 * Enable all of a task's events that have been marked enable-on-exec.
2867 2868
 * This expects task == current.
 */
P
Peter Zijlstra 已提交
2869
static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2870
{
2871
	struct perf_event *event;
2872 2873
	unsigned long flags;
	int enabled = 0;
2874
	int ret;
2875 2876

	local_irq_save(flags);
2877
	if (!ctx || !ctx->nr_events)
2878 2879
		goto out;

2880 2881 2882 2883 2884 2885 2886
	/*
	 * 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.
	 */
2887
	perf_cgroup_sched_out(current, NULL);
2888

2889
	raw_spin_lock(&ctx->lock);
2890
	task_ctx_sched_out(ctx);
2891

2892
	list_for_each_entry(event, &ctx->event_list, event_entry) {
2893 2894 2895
		ret = event_enable_on_exec(event, ctx);
		if (ret)
			enabled = 1;
2896 2897 2898
	}

	/*
2899
	 * Unclone this context if we enabled any event.
2900
	 */
2901 2902
	if (enabled)
		unclone_ctx(ctx);
2903

2904
	raw_spin_unlock(&ctx->lock);
2905

2906 2907 2908
	/*
	 * Also calls ctxswin for cgroup events, if any:
	 */
S
Stephane Eranian 已提交
2909
	perf_event_context_sched_in(ctx, ctx->task);
P
Peter Zijlstra 已提交
2910
out:
2911 2912 2913
	local_irq_restore(flags);
}

T
Thomas Gleixner 已提交
2914
/*
2915
 * Cross CPU call to read the hardware event
T
Thomas Gleixner 已提交
2916
 */
2917
static void __perf_event_read(void *info)
T
Thomas Gleixner 已提交
2918
{
2919 2920
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
2921
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
I
Ingo Molnar 已提交
2922

2923 2924 2925 2926
	/*
	 * 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
2927 2928
	 * event->count would have been updated to a recent sample
	 * when the event was scheduled out.
2929 2930 2931 2932
	 */
	if (ctx->task && cpuctx->task_ctx != ctx)
		return;

2933
	raw_spin_lock(&ctx->lock);
S
Stephane Eranian 已提交
2934
	if (ctx->is_active) {
2935
		update_context_time(ctx);
S
Stephane Eranian 已提交
2936 2937
		update_cgrp_time_from_event(event);
	}
2938
	update_event_times(event);
2939 2940
	if (event->state == PERF_EVENT_STATE_ACTIVE)
		event->pmu->read(event);
2941
	raw_spin_unlock(&ctx->lock);
T
Thomas Gleixner 已提交
2942 2943
}

P
Peter Zijlstra 已提交
2944 2945
static inline u64 perf_event_count(struct perf_event *event)
{
2946
	return local64_read(&event->count) + atomic64_read(&event->child_count);
P
Peter Zijlstra 已提交
2947 2948
}

2949
static u64 perf_event_read(struct perf_event *event)
T
Thomas Gleixner 已提交
2950 2951
{
	/*
2952 2953
	 * If event is enabled and currently active on a CPU, update the
	 * value in the event structure:
T
Thomas Gleixner 已提交
2954
	 */
2955 2956 2957 2958
	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
2959 2960 2961
		struct perf_event_context *ctx = event->ctx;
		unsigned long flags;

2962
		raw_spin_lock_irqsave(&ctx->lock, flags);
2963 2964 2965 2966 2967
		/*
		 * 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
2968
		if (ctx->is_active) {
2969
			update_context_time(ctx);
S
Stephane Eranian 已提交
2970 2971
			update_cgrp_time_from_event(event);
		}
2972
		update_event_times(event);
2973
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
T
Thomas Gleixner 已提交
2974 2975
	}

P
Peter Zijlstra 已提交
2976
	return perf_event_count(event);
T
Thomas Gleixner 已提交
2977 2978
}

2979
/*
2980
 * Initialize the perf_event context in a task_struct:
2981
 */
2982
static void __perf_event_init_context(struct perf_event_context *ctx)
2983
{
2984
	raw_spin_lock_init(&ctx->lock);
2985
	mutex_init(&ctx->mutex);
2986 2987
	INIT_LIST_HEAD(&ctx->pinned_groups);
	INIT_LIST_HEAD(&ctx->flexible_groups);
2988 2989
	INIT_LIST_HEAD(&ctx->event_list);
	atomic_set(&ctx->refcount, 1);
2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004
}

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
3005
	}
3006 3007 3008
	ctx->pmu = pmu;

	return ctx;
3009 3010
}

3011 3012 3013 3014 3015
static struct task_struct *
find_lively_task_by_vpid(pid_t vpid)
{
	struct task_struct *task;
	int err;
T
Thomas Gleixner 已提交
3016 3017

	rcu_read_lock();
3018
	if (!vpid)
T
Thomas Gleixner 已提交
3019 3020
		task = current;
	else
3021
		task = find_task_by_vpid(vpid);
T
Thomas Gleixner 已提交
3022 3023 3024 3025 3026 3027 3028 3029
	if (task)
		get_task_struct(task);
	rcu_read_unlock();

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

	/* Reuse ptrace permission checks for now. */
3030 3031 3032 3033
	err = -EACCES;
	if (!ptrace_may_access(task, PTRACE_MODE_READ))
		goto errout;

3034 3035 3036 3037 3038 3039 3040
	return task;
errout:
	put_task_struct(task);
	return ERR_PTR(err);

}

3041 3042 3043
/*
 * Returns a matching context with refcount and pincount.
 */
P
Peter Zijlstra 已提交
3044
static struct perf_event_context *
M
Matt Helsley 已提交
3045
find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
T
Thomas Gleixner 已提交
3046
{
3047
	struct perf_event_context *ctx;
3048
	struct perf_cpu_context *cpuctx;
3049
	unsigned long flags;
P
Peter Zijlstra 已提交
3050
	int ctxn, err;
T
Thomas Gleixner 已提交
3051

3052
	if (!task) {
3053
		/* Must be root to operate on a CPU event: */
3054
		if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
T
Thomas Gleixner 已提交
3055 3056 3057
			return ERR_PTR(-EACCES);

		/*
3058
		 * We could be clever and allow to attach a event to an
T
Thomas Gleixner 已提交
3059 3060 3061
		 * offline CPU and activate it when the CPU comes up, but
		 * that's for later.
		 */
3062
		if (!cpu_online(cpu))
T
Thomas Gleixner 已提交
3063 3064
			return ERR_PTR(-ENODEV);

P
Peter Zijlstra 已提交
3065
		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
T
Thomas Gleixner 已提交
3066
		ctx = &cpuctx->ctx;
3067
		get_ctx(ctx);
3068
		++ctx->pin_count;
T
Thomas Gleixner 已提交
3069 3070 3071 3072

		return ctx;
	}

P
Peter Zijlstra 已提交
3073 3074 3075 3076 3077
	err = -EINVAL;
	ctxn = pmu->task_ctx_nr;
	if (ctxn < 0)
		goto errout;

P
Peter Zijlstra 已提交
3078
retry:
P
Peter Zijlstra 已提交
3079
	ctx = perf_lock_task_context(task, ctxn, &flags);
3080
	if (ctx) {
3081
		unclone_ctx(ctx);
3082
		++ctx->pin_count;
3083
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
3084
	} else {
3085
		ctx = alloc_perf_context(pmu, task);
3086 3087 3088
		err = -ENOMEM;
		if (!ctx)
			goto errout;
3089

3090 3091 3092 3093 3094 3095 3096 3097 3098 3099
		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;
3100
		else {
3101
			get_ctx(ctx);
3102
			++ctx->pin_count;
3103
			rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3104
		}
3105 3106 3107
		mutex_unlock(&task->perf_event_mutex);

		if (unlikely(err)) {
3108
			put_ctx(ctx);
3109 3110 3111 3112

			if (err == -EAGAIN)
				goto retry;
			goto errout;
3113 3114 3115
		}
	}

T
Thomas Gleixner 已提交
3116
	return ctx;
3117

P
Peter Zijlstra 已提交
3118
errout:
3119
	return ERR_PTR(err);
T
Thomas Gleixner 已提交
3120 3121
}

L
Li Zefan 已提交
3122 3123
static void perf_event_free_filter(struct perf_event *event);

3124
static void free_event_rcu(struct rcu_head *head)
P
Peter Zijlstra 已提交
3125
{
3126
	struct perf_event *event;
P
Peter Zijlstra 已提交
3127

3128 3129 3130
	event = container_of(head, struct perf_event, rcu_head);
	if (event->ns)
		put_pid_ns(event->ns);
L
Li Zefan 已提交
3131
	perf_event_free_filter(event);
3132
	kfree(event);
P
Peter Zijlstra 已提交
3133 3134
}

3135
static void ring_buffer_put(struct ring_buffer *rb);
3136
static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3137

3138
static void unaccount_event_cpu(struct perf_event *event, int cpu)
3139
{
3140 3141 3142 3143 3144 3145 3146 3147 3148 3149
	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));
}
3150

3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163
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);
3164 3165
	if (event->attr.freq)
		atomic_dec(&nr_freq_events);
3166 3167 3168 3169 3170 3171 3172
	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);
}
3173

3174 3175
static void __free_event(struct perf_event *event)
{
3176
	if (!event->parent) {
3177 3178
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
			put_callchain_buffers();
3179
	}
3180

3181 3182 3183 3184 3185 3186 3187 3188
	if (event->destroy)
		event->destroy(event);

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

	call_rcu(&event->rcu_head, free_event_rcu);
}
3189
static void free_event(struct perf_event *event)
3190
{
3191
	irq_work_sync(&event->pending);
3192

3193
	unaccount_event(event);
3194

3195
	if (event->rb) {
3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211
		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);
3212 3213
	}

S
Stephane Eranian 已提交
3214 3215 3216
	if (is_cgroup_event(event))
		perf_detach_cgroup(event);

P
Peter Zijlstra 已提交
3217

3218
	__free_event(event);
3219 3220
}

3221
int perf_event_release_kernel(struct perf_event *event)
T
Thomas Gleixner 已提交
3222
{
3223
	struct perf_event_context *ctx = event->ctx;
T
Thomas Gleixner 已提交
3224

3225
	WARN_ON_ONCE(ctx->parent_ctx);
3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238
	/*
	 * 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);
3239
	raw_spin_lock_irq(&ctx->lock);
3240
	perf_group_detach(event);
3241
	raw_spin_unlock_irq(&ctx->lock);
3242
	perf_remove_from_context(event);
3243
	mutex_unlock(&ctx->mutex);
T
Thomas Gleixner 已提交
3244

3245
	free_event(event);
T
Thomas Gleixner 已提交
3246 3247 3248

	return 0;
}
3249
EXPORT_SYMBOL_GPL(perf_event_release_kernel);
T
Thomas Gleixner 已提交
3250

3251 3252 3253
/*
 * Called when the last reference to the file is gone.
 */
3254
static void put_event(struct perf_event *event)
3255
{
P
Peter Zijlstra 已提交
3256
	struct task_struct *owner;
3257

3258 3259
	if (!atomic_long_dec_and_test(&event->refcount))
		return;
3260

P
Peter Zijlstra 已提交
3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293
	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);
	}

3294 3295 3296 3297 3298 3299 3300
	perf_event_release_kernel(event);
}

static int perf_release(struct inode *inode, struct file *file)
{
	put_event(file->private_data);
	return 0;
3301 3302
}

3303
u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3304
{
3305
	struct perf_event *child;
3306 3307
	u64 total = 0;

3308 3309 3310
	*enabled = 0;
	*running = 0;

3311
	mutex_lock(&event->child_mutex);
3312
	total += perf_event_read(event);
3313 3314 3315 3316 3317 3318
	*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) {
3319
		total += perf_event_read(child);
3320 3321 3322
		*enabled += child->total_time_enabled;
		*running += child->total_time_running;
	}
3323
	mutex_unlock(&event->child_mutex);
3324 3325 3326

	return total;
}
3327
EXPORT_SYMBOL_GPL(perf_event_read_value);
3328

3329
static int perf_event_read_group(struct perf_event *event,
3330 3331
				   u64 read_format, char __user *buf)
{
3332
	struct perf_event *leader = event->group_leader, *sub;
3333 3334
	int n = 0, size = 0, ret = -EFAULT;
	struct perf_event_context *ctx = leader->ctx;
3335
	u64 values[5];
3336
	u64 count, enabled, running;
3337

3338
	mutex_lock(&ctx->mutex);
3339
	count = perf_event_read_value(leader, &enabled, &running);
3340 3341

	values[n++] = 1 + leader->nr_siblings;
3342 3343 3344 3345
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
		values[n++] = enabled;
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
		values[n++] = running;
3346 3347 3348
	values[n++] = count;
	if (read_format & PERF_FORMAT_ID)
		values[n++] = primary_event_id(leader);
3349 3350 3351 3352

	size = n * sizeof(u64);

	if (copy_to_user(buf, values, size))
3353
		goto unlock;
3354

3355
	ret = size;
3356

3357
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3358
		n = 0;
3359

3360
		values[n++] = perf_event_read_value(sub, &enabled, &running);
3361 3362 3363 3364 3365
		if (read_format & PERF_FORMAT_ID)
			values[n++] = primary_event_id(sub);

		size = n * sizeof(u64);

3366
		if (copy_to_user(buf + ret, values, size)) {
3367 3368 3369
			ret = -EFAULT;
			goto unlock;
		}
3370 3371

		ret += size;
3372
	}
3373 3374
unlock:
	mutex_unlock(&ctx->mutex);
3375

3376
	return ret;
3377 3378
}

3379
static int perf_event_read_one(struct perf_event *event,
3380 3381
				 u64 read_format, char __user *buf)
{
3382
	u64 enabled, running;
3383 3384 3385
	u64 values[4];
	int n = 0;

3386 3387 3388 3389 3390
	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;
3391
	if (read_format & PERF_FORMAT_ID)
3392
		values[n++] = primary_event_id(event);
3393 3394 3395 3396 3397 3398 3399

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

	return n * sizeof(u64);
}

T
Thomas Gleixner 已提交
3400
/*
3401
 * Read the performance event - simple non blocking version for now
T
Thomas Gleixner 已提交
3402 3403
 */
static ssize_t
3404
perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
T
Thomas Gleixner 已提交
3405
{
3406
	u64 read_format = event->attr.read_format;
3407
	int ret;
T
Thomas Gleixner 已提交
3408

3409
	/*
3410
	 * Return end-of-file for a read on a event that is in
3411 3412 3413
	 * error state (i.e. because it was pinned but it couldn't be
	 * scheduled on to the CPU at some point).
	 */
3414
	if (event->state == PERF_EVENT_STATE_ERROR)
3415 3416
		return 0;

3417
	if (count < event->read_size)
3418 3419
		return -ENOSPC;

3420
	WARN_ON_ONCE(event->ctx->parent_ctx);
3421
	if (read_format & PERF_FORMAT_GROUP)
3422
		ret = perf_event_read_group(event, read_format, buf);
3423
	else
3424
		ret = perf_event_read_one(event, read_format, buf);
T
Thomas Gleixner 已提交
3425

3426
	return ret;
T
Thomas Gleixner 已提交
3427 3428 3429 3430 3431
}

static ssize_t
perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
3432
	struct perf_event *event = file->private_data;
T
Thomas Gleixner 已提交
3433

3434
	return perf_read_hw(event, buf, count);
T
Thomas Gleixner 已提交
3435 3436 3437 3438
}

static unsigned int perf_poll(struct file *file, poll_table *wait)
{
3439
	struct perf_event *event = file->private_data;
3440
	struct ring_buffer *rb;
3441
	unsigned int events = POLL_HUP;
P
Peter Zijlstra 已提交
3442

3443
	/*
3444 3445
	 * Pin the event->rb by taking event->mmap_mutex; otherwise
	 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3446 3447
	 */
	mutex_lock(&event->mmap_mutex);
3448 3449
	rb = event->rb;
	if (rb)
3450
		events = atomic_xchg(&rb->poll, 0);
3451 3452
	mutex_unlock(&event->mmap_mutex);

3453
	poll_wait(file, &event->waitq, wait);
T
Thomas Gleixner 已提交
3454 3455 3456 3457

	return events;
}

3458
static void perf_event_reset(struct perf_event *event)
3459
{
3460
	(void)perf_event_read(event);
3461
	local64_set(&event->count, 0);
3462
	perf_event_update_userpage(event);
P
Peter Zijlstra 已提交
3463 3464
}

3465
/*
3466 3467 3468 3469
 * 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.
3470
 */
3471 3472
static void perf_event_for_each_child(struct perf_event *event,
					void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3473
{
3474
	struct perf_event *child;
P
Peter Zijlstra 已提交
3475

3476 3477 3478 3479
	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
3480
		func(child);
3481
	mutex_unlock(&event->child_mutex);
P
Peter Zijlstra 已提交
3482 3483
}

3484 3485
static void perf_event_for_each(struct perf_event *event,
				  void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3486
{
3487 3488
	struct perf_event_context *ctx = event->ctx;
	struct perf_event *sibling;
P
Peter Zijlstra 已提交
3489

3490 3491
	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
3492
	event = event->group_leader;
3493

3494 3495
	perf_event_for_each_child(event, func);
	list_for_each_entry(sibling, &event->sibling_list, group_entry)
3496
		perf_event_for_each_child(sibling, func);
3497
	mutex_unlock(&ctx->mutex);
3498 3499
}

3500
static int perf_event_period(struct perf_event *event, u64 __user *arg)
3501
{
3502
	struct perf_event_context *ctx = event->ctx;
3503 3504 3505
	int ret = 0;
	u64 value;

3506
	if (!is_sampling_event(event))
3507 3508
		return -EINVAL;

3509
	if (copy_from_user(&value, arg, sizeof(value)))
3510 3511 3512 3513 3514
		return -EFAULT;

	if (!value)
		return -EINVAL;

3515
	raw_spin_lock_irq(&ctx->lock);
3516 3517
	if (event->attr.freq) {
		if (value > sysctl_perf_event_sample_rate) {
3518 3519 3520 3521
			ret = -EINVAL;
			goto unlock;
		}

3522
		event->attr.sample_freq = value;
3523
	} else {
3524 3525
		event->attr.sample_period = value;
		event->hw.sample_period = value;
3526 3527
	}
unlock:
3528
	raw_spin_unlock_irq(&ctx->lock);
3529 3530 3531 3532

	return ret;
}

3533 3534
static const struct file_operations perf_fops;

3535
static inline int perf_fget_light(int fd, struct fd *p)
3536
{
3537 3538 3539
	struct fd f = fdget(fd);
	if (!f.file)
		return -EBADF;
3540

3541 3542 3543
	if (f.file->f_op != &perf_fops) {
		fdput(f);
		return -EBADF;
3544
	}
3545 3546
	*p = f;
	return 0;
3547 3548 3549 3550
}

static int perf_event_set_output(struct perf_event *event,
				 struct perf_event *output_event);
L
Li Zefan 已提交
3551
static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3552

3553 3554
static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
3555 3556
	struct perf_event *event = file->private_data;
	void (*func)(struct perf_event *);
P
Peter Zijlstra 已提交
3557
	u32 flags = arg;
3558 3559

	switch (cmd) {
3560 3561
	case PERF_EVENT_IOC_ENABLE:
		func = perf_event_enable;
3562
		break;
3563 3564
	case PERF_EVENT_IOC_DISABLE:
		func = perf_event_disable;
3565
		break;
3566 3567
	case PERF_EVENT_IOC_RESET:
		func = perf_event_reset;
3568
		break;
P
Peter Zijlstra 已提交
3569

3570 3571
	case PERF_EVENT_IOC_REFRESH:
		return perf_event_refresh(event, arg);
3572

3573 3574
	case PERF_EVENT_IOC_PERIOD:
		return perf_event_period(event, (u64 __user *)arg);
3575

3576 3577 3578 3579 3580 3581 3582 3583 3584
	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;
	}

3585
	case PERF_EVENT_IOC_SET_OUTPUT:
3586 3587 3588
	{
		int ret;
		if (arg != -1) {
3589 3590 3591 3592 3593 3594 3595 3596 3597 3598
			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);
3599 3600 3601
		}
		return ret;
	}
3602

L
Li Zefan 已提交
3603 3604 3605
	case PERF_EVENT_IOC_SET_FILTER:
		return perf_event_set_filter(event, (void __user *)arg);

3606
	default:
P
Peter Zijlstra 已提交
3607
		return -ENOTTY;
3608
	}
P
Peter Zijlstra 已提交
3609 3610

	if (flags & PERF_IOC_FLAG_GROUP)
3611
		perf_event_for_each(event, func);
P
Peter Zijlstra 已提交
3612
	else
3613
		perf_event_for_each_child(event, func);
P
Peter Zijlstra 已提交
3614 3615

	return 0;
3616 3617
}

3618
int perf_event_task_enable(void)
3619
{
3620
	struct perf_event *event;
3621

3622 3623 3624 3625
	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);
3626 3627 3628 3629

	return 0;
}

3630
int perf_event_task_disable(void)
3631
{
3632
	struct perf_event *event;
3633

3634 3635 3636 3637
	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);
3638 3639 3640 3641

	return 0;
}

3642
static int perf_event_index(struct perf_event *event)
3643
{
P
Peter Zijlstra 已提交
3644 3645 3646
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;

3647
	if (event->state != PERF_EVENT_STATE_ACTIVE)
3648 3649
		return 0;

3650
	return event->pmu->event_idx(event);
3651 3652
}

3653
static void calc_timer_values(struct perf_event *event,
3654
				u64 *now,
3655 3656
				u64 *enabled,
				u64 *running)
3657
{
3658
	u64 ctx_time;
3659

3660 3661
	*now = perf_clock();
	ctx_time = event->shadow_ctx_time + *now;
3662 3663 3664 3665
	*enabled = ctx_time - event->tstamp_enabled;
	*running = ctx_time - event->tstamp_running;
}

3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685
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();
}

3686
void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3687 3688 3689
{
}

3690 3691 3692 3693 3694
/*
 * 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.
 */
3695
void perf_event_update_userpage(struct perf_event *event)
3696
{
3697
	struct perf_event_mmap_page *userpg;
3698
	struct ring_buffer *rb;
3699
	u64 enabled, running, now;
3700 3701

	rcu_read_lock();
3702 3703 3704 3705
	rb = rcu_dereference(event->rb);
	if (!rb)
		goto unlock;

3706 3707 3708 3709 3710 3711 3712 3713 3714
	/*
	 * 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
	 */
3715
	calc_timer_values(event, &now, &enabled, &running);
3716

3717
	userpg = rb->user_page;
3718 3719 3720 3721 3722
	/*
	 * Disable preemption so as to not let the corresponding user-space
	 * spin too long if we get preempted.
	 */
	preempt_disable();
3723
	++userpg->lock;
3724
	barrier();
3725
	userpg->index = perf_event_index(event);
P
Peter Zijlstra 已提交
3726
	userpg->offset = perf_event_count(event);
3727
	if (userpg->index)
3728
		userpg->offset -= local64_read(&event->hw.prev_count);
3729

3730
	userpg->time_enabled = enabled +
3731
			atomic64_read(&event->child_total_time_enabled);
3732

3733
	userpg->time_running = running +
3734
			atomic64_read(&event->child_total_time_running);
3735

3736
	arch_perf_update_userpage(userpg, now);
3737

3738
	barrier();
3739
	++userpg->lock;
3740
	preempt_enable();
3741
unlock:
3742
	rcu_read_unlock();
3743 3744
}

3745 3746 3747
static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
{
	struct perf_event *event = vma->vm_file->private_data;
3748
	struct ring_buffer *rb;
3749 3750 3751 3752 3753 3754 3755 3756 3757
	int ret = VM_FAULT_SIGBUS;

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

	rcu_read_lock();
3758 3759
	rb = rcu_dereference(event->rb);
	if (!rb)
3760 3761 3762 3763 3764
		goto unlock;

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

3765
	vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779
	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;
}

3780 3781 3782 3783 3784 3785 3786 3787 3788
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);
3789 3790
	if (list_empty(&event->rb_entry))
		list_add(&event->rb_entry, &rb->event_list);
3791 3792 3793
	spin_unlock_irqrestore(&rb->event_lock, flags);
}

3794
static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812
{
	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);
3813 3814 3815 3816
	if (rb) {
		list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
			wake_up_all(&event->waitq);
	}
3817 3818 3819
	rcu_read_unlock();
}

3820
static void rb_free_rcu(struct rcu_head *rcu_head)
3821
{
3822
	struct ring_buffer *rb;
3823

3824 3825
	rb = container_of(rcu_head, struct ring_buffer, rcu_head);
	rb_free(rb);
3826 3827
}

3828
static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3829
{
3830
	struct ring_buffer *rb;
3831

3832
	rcu_read_lock();
3833 3834 3835 3836
	rb = rcu_dereference(event->rb);
	if (rb) {
		if (!atomic_inc_not_zero(&rb->refcount))
			rb = NULL;
3837 3838 3839
	}
	rcu_read_unlock();

3840
	return rb;
3841 3842
}

3843
static void ring_buffer_put(struct ring_buffer *rb)
3844
{
3845
	if (!atomic_dec_and_test(&rb->refcount))
3846
		return;
3847

3848
	WARN_ON_ONCE(!list_empty(&rb->event_list));
3849

3850
	call_rcu(&rb->rcu_head, rb_free_rcu);
3851 3852 3853 3854
}

static void perf_mmap_open(struct vm_area_struct *vma)
{
3855
	struct perf_event *event = vma->vm_file->private_data;
3856

3857
	atomic_inc(&event->mmap_count);
3858
	atomic_inc(&event->rb->mmap_count);
3859 3860
}

3861 3862 3863 3864 3865 3866 3867 3868
/*
 * 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.
 */
3869 3870
static void perf_mmap_close(struct vm_area_struct *vma)
{
3871
	struct perf_event *event = vma->vm_file->private_data;
3872

3873 3874 3875 3876
	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);
3877

3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892
	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;
	}
3893

3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909
	/*
	 * 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();
3910

3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925
		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
3926
		}
3927
		mutex_unlock(&event->mmap_mutex);
3928
		put_event(event);
3929

3930 3931 3932 3933 3934
		/*
		 * Restart the iteration; either we're on the wrong list or
		 * destroyed its integrity by doing a deletion.
		 */
		goto again;
3935
	}
3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951
	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 */
3952 3953
}

3954
static const struct vm_operations_struct perf_mmap_vmops = {
3955 3956 3957 3958
	.open		= perf_mmap_open,
	.close		= perf_mmap_close,
	.fault		= perf_mmap_fault,
	.page_mkwrite	= perf_mmap_fault,
3959 3960 3961 3962
};

static int perf_mmap(struct file *file, struct vm_area_struct *vma)
{
3963
	struct perf_event *event = file->private_data;
3964
	unsigned long user_locked, user_lock_limit;
3965
	struct user_struct *user = current_user();
3966
	unsigned long locked, lock_limit;
3967
	struct ring_buffer *rb;
3968 3969
	unsigned long vma_size;
	unsigned long nr_pages;
3970
	long user_extra, extra;
3971
	int ret = 0, flags = 0;
3972

3973 3974 3975
	/*
	 * Don't allow mmap() of inherited per-task counters. This would
	 * create a performance issue due to all children writing to the
3976
	 * same rb.
3977 3978 3979 3980
	 */
	if (event->cpu == -1 && event->attr.inherit)
		return -EINVAL;

3981
	if (!(vma->vm_flags & VM_SHARED))
3982
		return -EINVAL;
3983 3984 3985 3986

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

3987
	/*
3988
	 * If we have rb pages ensure they're a power-of-two number, so we
3989 3990 3991
	 * can do bitmasks instead of modulo.
	 */
	if (nr_pages != 0 && !is_power_of_2(nr_pages))
3992 3993
		return -EINVAL;

3994
	if (vma_size != PAGE_SIZE * (1 + nr_pages))
3995 3996
		return -EINVAL;

3997 3998
	if (vma->vm_pgoff != 0)
		return -EINVAL;
3999

4000
	WARN_ON_ONCE(event->ctx->parent_ctx);
4001
again:
4002
	mutex_lock(&event->mmap_mutex);
4003
	if (event->rb) {
4004
		if (event->rb->nr_pages != nr_pages) {
4005
			ret = -EINVAL;
4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018
			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;
		}

4019 4020 4021
		goto unlock;
	}

4022
	user_extra = nr_pages + 1;
4023
	user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
I
Ingo Molnar 已提交
4024 4025 4026 4027 4028 4029

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

4030
	user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4031

4032 4033 4034
	extra = 0;
	if (user_locked > user_lock_limit)
		extra = user_locked - user_lock_limit;
4035

4036
	lock_limit = rlimit(RLIMIT_MEMLOCK);
4037
	lock_limit >>= PAGE_SHIFT;
4038
	locked = vma->vm_mm->pinned_vm + extra;
4039

4040 4041
	if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
		!capable(CAP_IPC_LOCK)) {
4042 4043 4044
		ret = -EPERM;
		goto unlock;
	}
4045

4046
	WARN_ON(event->rb);
4047

4048
	if (vma->vm_flags & VM_WRITE)
4049
		flags |= RING_BUFFER_WRITABLE;
4050

4051 4052 4053 4054
	rb = rb_alloc(nr_pages, 
		event->attr.watermark ? event->attr.wakeup_watermark : 0,
		event->cpu, flags);

4055
	if (!rb) {
4056
		ret = -ENOMEM;
4057
		goto unlock;
4058
	}
P
Peter Zijlstra 已提交
4059

4060
	atomic_set(&rb->mmap_count, 1);
P
Peter Zijlstra 已提交
4061 4062
	rb->mmap_locked = extra;
	rb->mmap_user = get_current_user();
4063

4064
	atomic_long_add(user_extra, &user->locked_vm);
P
Peter Zijlstra 已提交
4065 4066
	vma->vm_mm->pinned_vm += extra;

4067
	ring_buffer_attach(event, rb);
P
Peter Zijlstra 已提交
4068
	rcu_assign_pointer(event->rb, rb);
4069

4070
	perf_event_init_userpage(event);
4071 4072
	perf_event_update_userpage(event);

4073
unlock:
4074 4075
	if (!ret)
		atomic_inc(&event->mmap_count);
4076
	mutex_unlock(&event->mmap_mutex);
4077

4078 4079 4080 4081
	/*
	 * Since pinned accounting is per vm we cannot allow fork() to copy our
	 * vma.
	 */
P
Peter Zijlstra 已提交
4082
	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4083
	vma->vm_ops = &perf_mmap_vmops;
4084 4085

	return ret;
4086 4087
}

P
Peter Zijlstra 已提交
4088 4089
static int perf_fasync(int fd, struct file *filp, int on)
{
A
Al Viro 已提交
4090
	struct inode *inode = file_inode(filp);
4091
	struct perf_event *event = filp->private_data;
P
Peter Zijlstra 已提交
4092 4093 4094
	int retval;

	mutex_lock(&inode->i_mutex);
4095
	retval = fasync_helper(fd, filp, on, &event->fasync);
P
Peter Zijlstra 已提交
4096 4097 4098 4099 4100 4101 4102 4103
	mutex_unlock(&inode->i_mutex);

	if (retval < 0)
		return retval;

	return 0;
}

T
Thomas Gleixner 已提交
4104
static const struct file_operations perf_fops = {
4105
	.llseek			= no_llseek,
T
Thomas Gleixner 已提交
4106 4107 4108
	.release		= perf_release,
	.read			= perf_read,
	.poll			= perf_poll,
4109 4110
	.unlocked_ioctl		= perf_ioctl,
	.compat_ioctl		= perf_ioctl,
4111
	.mmap			= perf_mmap,
P
Peter Zijlstra 已提交
4112
	.fasync			= perf_fasync,
T
Thomas Gleixner 已提交
4113 4114
};

4115
/*
4116
 * Perf event wakeup
4117 4118 4119 4120 4121
 *
 * If there's data, ensure we set the poll() state and publish everything
 * to user-space before waking everybody up.
 */

4122
void perf_event_wakeup(struct perf_event *event)
4123
{
4124
	ring_buffer_wakeup(event);
4125

4126 4127 4128
	if (event->pending_kill) {
		kill_fasync(&event->fasync, SIGIO, event->pending_kill);
		event->pending_kill = 0;
4129
	}
4130 4131
}

4132
static void perf_pending_event(struct irq_work *entry)
4133
{
4134 4135
	struct perf_event *event = container_of(entry,
			struct perf_event, pending);
4136

4137 4138 4139
	if (event->pending_disable) {
		event->pending_disable = 0;
		__perf_event_disable(event);
4140 4141
	}

4142 4143 4144
	if (event->pending_wakeup) {
		event->pending_wakeup = 0;
		perf_event_wakeup(event);
4145 4146 4147
	}
}

4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168
/*
 * 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);

4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199
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);
	}
}

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 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
/*
 * 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);
	}
}

4295 4296 4297
static void __perf_event_header__init_id(struct perf_event_header *header,
					 struct perf_sample_data *data,
					 struct perf_event *event)
4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312
{
	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();

4313
	if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324
		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;
	}
}

4325 4326 4327
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 4343 4344 4345 4346 4347 4348 4349 4350 4351
{
	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);
4352 4353 4354

	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);
4355 4356
}

4357 4358 4359
void perf_event__output_id_sample(struct perf_event *event,
				  struct perf_output_handle *handle,
				  struct perf_sample_data *sample)
4360 4361 4362 4363 4364
{
	if (event->attr.sample_id_all)
		__perf_event__output_id_sample(handle, sample);
}

4365
static void perf_output_read_one(struct perf_output_handle *handle,
4366 4367
				 struct perf_event *event,
				 u64 enabled, u64 running)
4368
{
4369
	u64 read_format = event->attr.read_format;
4370 4371 4372
	u64 values[4];
	int n = 0;

P
Peter Zijlstra 已提交
4373
	values[n++] = perf_event_count(event);
4374
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4375
		values[n++] = enabled +
4376
			atomic64_read(&event->child_total_time_enabled);
4377 4378
	}
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4379
		values[n++] = running +
4380
			atomic64_read(&event->child_total_time_running);
4381 4382
	}
	if (read_format & PERF_FORMAT_ID)
4383
		values[n++] = primary_event_id(event);
4384

4385
	__output_copy(handle, values, n * sizeof(u64));
4386 4387 4388
}

/*
4389
 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4390 4391
 */
static void perf_output_read_group(struct perf_output_handle *handle,
4392 4393
			    struct perf_event *event,
			    u64 enabled, u64 running)
4394
{
4395 4396
	struct perf_event *leader = event->group_leader, *sub;
	u64 read_format = event->attr.read_format;
4397 4398 4399 4400 4401 4402
	u64 values[5];
	int n = 0;

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

	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4403
		values[n++] = enabled;
4404 4405

	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4406
		values[n++] = running;
4407

4408
	if (leader != event)
4409 4410
		leader->pmu->read(leader);

P
Peter Zijlstra 已提交
4411
	values[n++] = perf_event_count(leader);
4412
	if (read_format & PERF_FORMAT_ID)
4413
		values[n++] = primary_event_id(leader);
4414

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

4417
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4418 4419
		n = 0;

4420 4421
		if ((sub != event) &&
		    (sub->state == PERF_EVENT_STATE_ACTIVE))
4422 4423
			sub->pmu->read(sub);

P
Peter Zijlstra 已提交
4424
		values[n++] = perf_event_count(sub);
4425
		if (read_format & PERF_FORMAT_ID)
4426
			values[n++] = primary_event_id(sub);
4427

4428
		__output_copy(handle, values, n * sizeof(u64));
4429 4430 4431
	}
}

4432 4433 4434
#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
				 PERF_FORMAT_TOTAL_TIME_RUNNING)

4435
static void perf_output_read(struct perf_output_handle *handle,
4436
			     struct perf_event *event)
4437
{
4438
	u64 enabled = 0, running = 0, now;
4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449
	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
	 */
4450
	if (read_format & PERF_FORMAT_TOTAL_TIMES)
4451
		calc_timer_values(event, &now, &enabled, &running);
4452

4453
	if (event->attr.read_format & PERF_FORMAT_GROUP)
4454
		perf_output_read_group(handle, event, enabled, running);
4455
	else
4456
		perf_output_read_one(handle, event, enabled, running);
4457 4458
}

4459 4460 4461
void perf_output_sample(struct perf_output_handle *handle,
			struct perf_event_header *header,
			struct perf_sample_data *data,
4462
			struct perf_event *event)
4463 4464 4465 4466 4467
{
	u64 sample_type = data->type;

	perf_output_put(handle, *header);

4468 4469 4470
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);

4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495
	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)
4496
		perf_output_read(handle, event);
4497 4498 4499 4500 4501 4502 4503 4504 4505 4506

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

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

			size *= sizeof(u64);

4507
			__output_copy(handle, data->callchain, size);
4508 4509 4510 4511 4512 4513 4514 4515 4516
		} 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);
4517 4518
			__output_copy(handle, data->raw->data,
					   data->raw->size);
4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529
		} else {
			struct {
				u32	size;
				u32	data;
			} raw = {
				.size = sizeof(u32),
				.data = 0,
			};
			perf_output_put(handle, raw);
		}
	}
4530

4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547
	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);
		}
	}
4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564

	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);
		}
	}
4565

4566
	if (sample_type & PERF_SAMPLE_STACK_USER) {
4567 4568 4569
		perf_output_sample_ustack(handle,
					  data->stack_user_size,
					  data->regs_user.regs);
4570
	}
A
Andi Kleen 已提交
4571 4572 4573

	if (sample_type & PERF_SAMPLE_WEIGHT)
		perf_output_put(handle, data->weight);
4574 4575 4576

	if (sample_type & PERF_SAMPLE_DATA_SRC)
		perf_output_put(handle, data->data_src.val);
4577

A
Andi Kleen 已提交
4578 4579 4580
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		perf_output_put(handle, data->txn);

4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593
	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);
			}
		}
	}
4594 4595 4596 4597
}

void perf_prepare_sample(struct perf_event_header *header,
			 struct perf_sample_data *data,
4598
			 struct perf_event *event,
4599
			 struct pt_regs *regs)
4600
{
4601
	u64 sample_type = event->attr.sample_type;
4602

4603
	header->type = PERF_RECORD_SAMPLE;
4604
	header->size = sizeof(*header) + event->header_size;
4605 4606 4607

	header->misc = 0;
	header->misc |= perf_misc_flags(regs);
4608

4609
	__perf_event_header__init_id(header, data, event);
4610

4611
	if (sample_type & PERF_SAMPLE_IP)
4612 4613
		data->ip = perf_instruction_pointer(regs);

4614
	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4615
		int size = 1;
4616

4617
		data->callchain = perf_callchain(event, regs);
4618 4619 4620 4621 4622

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

		header->size += size * sizeof(u64);
4623 4624
	}

4625
	if (sample_type & PERF_SAMPLE_RAW) {
4626 4627 4628 4629 4630 4631 4632 4633
		int size = sizeof(u32);

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

		WARN_ON_ONCE(size & (sizeof(u64)-1));
4634
		header->size += size;
4635
	}
4636 4637 4638 4639 4640 4641 4642 4643 4644

	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;
	}
4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658

	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;
	}
4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687

	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;
	}
4688
}
4689

4690
static void perf_event_output(struct perf_event *event,
4691 4692 4693 4694 4695
				struct perf_sample_data *data,
				struct pt_regs *regs)
{
	struct perf_output_handle handle;
	struct perf_event_header header;
4696

4697 4698 4699
	/* protect the callchain buffers */
	rcu_read_lock();

4700
	perf_prepare_sample(&header, data, event, regs);
P
Peter Zijlstra 已提交
4701

4702
	if (perf_output_begin(&handle, event, header.size))
4703
		goto exit;
4704

4705
	perf_output_sample(&handle, &header, data, event);
4706

4707
	perf_output_end(&handle);
4708 4709 4710

exit:
	rcu_read_unlock();
4711 4712
}

4713
/*
4714
 * read event_id
4715 4716 4717 4718 4719 4720 4721 4722 4723 4724
 */

struct perf_read_event {
	struct perf_event_header	header;

	u32				pid;
	u32				tid;
};

static void
4725
perf_event_read_event(struct perf_event *event,
4726 4727 4728
			struct task_struct *task)
{
	struct perf_output_handle handle;
4729
	struct perf_sample_data sample;
4730
	struct perf_read_event read_event = {
4731
		.header = {
4732
			.type = PERF_RECORD_READ,
4733
			.misc = 0,
4734
			.size = sizeof(read_event) + event->read_size,
4735
		},
4736 4737
		.pid = perf_event_pid(event, task),
		.tid = perf_event_tid(event, task),
4738
	};
4739
	int ret;
4740

4741
	perf_event_header__init_id(&read_event.header, &sample, event);
4742
	ret = perf_output_begin(&handle, event, read_event.header.size);
4743 4744 4745
	if (ret)
		return;

4746
	perf_output_put(&handle, read_event);
4747
	perf_output_read(&handle, event);
4748
	perf_event__output_id_sample(event, &handle, &sample);
4749

4750 4751 4752
	perf_output_end(&handle);
}

4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766
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;
4767
		output(event, data);
4768 4769 4770 4771
	}
}

static void
4772
perf_event_aux(perf_event_aux_output_cb output, void *data,
4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784
	       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;
4785
		perf_event_aux_ctx(&cpuctx->ctx, output, data);
4786 4787 4788 4789 4790 4791 4792
		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)
4793
			perf_event_aux_ctx(ctx, output, data);
4794 4795 4796 4797 4798 4799
next:
		put_cpu_ptr(pmu->pmu_cpu_context);
	}

	if (task_ctx) {
		preempt_disable();
4800
		perf_event_aux_ctx(task_ctx, output, data);
4801 4802 4803 4804 4805
		preempt_enable();
	}
	rcu_read_unlock();
}

P
Peter Zijlstra 已提交
4806
/*
P
Peter Zijlstra 已提交
4807 4808
 * task tracking -- fork/exit
 *
4809
 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
P
Peter Zijlstra 已提交
4810 4811
 */

P
Peter Zijlstra 已提交
4812
struct perf_task_event {
4813
	struct task_struct		*task;
4814
	struct perf_event_context	*task_ctx;
P
Peter Zijlstra 已提交
4815 4816 4817 4818 4819 4820

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				ppid;
P
Peter Zijlstra 已提交
4821 4822
		u32				tid;
		u32				ptid;
4823
		u64				time;
4824
	} event_id;
P
Peter Zijlstra 已提交
4825 4826
};

4827 4828
static int perf_event_task_match(struct perf_event *event)
{
4829 4830 4831
	return event->attr.comm  || event->attr.mmap ||
	       event->attr.mmap2 || event->attr.mmap_data ||
	       event->attr.task;
4832 4833
}

4834
static void perf_event_task_output(struct perf_event *event,
4835
				   void *data)
P
Peter Zijlstra 已提交
4836
{
4837
	struct perf_task_event *task_event = data;
P
Peter Zijlstra 已提交
4838
	struct perf_output_handle handle;
4839
	struct perf_sample_data	sample;
P
Peter Zijlstra 已提交
4840
	struct task_struct *task = task_event->task;
4841
	int ret, size = task_event->event_id.header.size;
4842

4843 4844 4845
	if (!perf_event_task_match(event))
		return;

4846
	perf_event_header__init_id(&task_event->event_id.header, &sample, event);
P
Peter Zijlstra 已提交
4847

4848
	ret = perf_output_begin(&handle, event,
4849
				task_event->event_id.header.size);
4850
	if (ret)
4851
		goto out;
P
Peter Zijlstra 已提交
4852

4853 4854
	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
4855

4856 4857
	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
4858

4859
	perf_output_put(&handle, task_event->event_id);
4860

4861 4862
	perf_event__output_id_sample(event, &handle, &sample);

P
Peter Zijlstra 已提交
4863
	perf_output_end(&handle);
4864 4865
out:
	task_event->event_id.header.size = size;
P
Peter Zijlstra 已提交
4866 4867
}

4868 4869
static void perf_event_task(struct task_struct *task,
			      struct perf_event_context *task_ctx,
4870
			      int new)
P
Peter Zijlstra 已提交
4871
{
P
Peter Zijlstra 已提交
4872
	struct perf_task_event task_event;
P
Peter Zijlstra 已提交
4873

4874 4875 4876
	if (!atomic_read(&nr_comm_events) &&
	    !atomic_read(&nr_mmap_events) &&
	    !atomic_read(&nr_task_events))
P
Peter Zijlstra 已提交
4877 4878
		return;

P
Peter Zijlstra 已提交
4879
	task_event = (struct perf_task_event){
4880 4881
		.task	  = task,
		.task_ctx = task_ctx,
4882
		.event_id    = {
P
Peter Zijlstra 已提交
4883
			.header = {
4884
				.type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4885
				.misc = 0,
4886
				.size = sizeof(task_event.event_id),
P
Peter Zijlstra 已提交
4887
			},
4888 4889
			/* .pid  */
			/* .ppid */
P
Peter Zijlstra 已提交
4890 4891
			/* .tid  */
			/* .ptid */
P
Peter Zijlstra 已提交
4892
			.time = perf_clock(),
P
Peter Zijlstra 已提交
4893 4894 4895
		},
	};

4896
	perf_event_aux(perf_event_task_output,
4897 4898
		       &task_event,
		       task_ctx);
P
Peter Zijlstra 已提交
4899 4900
}

4901
void perf_event_fork(struct task_struct *task)
P
Peter Zijlstra 已提交
4902
{
4903
	perf_event_task(task, NULL, 1);
P
Peter Zijlstra 已提交
4904 4905
}

4906 4907 4908 4909 4910
/*
 * comm tracking
 */

struct perf_comm_event {
4911 4912
	struct task_struct	*task;
	char			*comm;
4913 4914 4915 4916 4917 4918 4919
	int			comm_size;

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
4920
	} event_id;
4921 4922
};

4923 4924 4925 4926 4927
static int perf_event_comm_match(struct perf_event *event)
{
	return event->attr.comm;
}

4928
static void perf_event_comm_output(struct perf_event *event,
4929
				   void *data)
4930
{
4931
	struct perf_comm_event *comm_event = data;
4932
	struct perf_output_handle handle;
4933
	struct perf_sample_data sample;
4934
	int size = comm_event->event_id.header.size;
4935 4936
	int ret;

4937 4938 4939
	if (!perf_event_comm_match(event))
		return;

4940 4941
	perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
4942
				comm_event->event_id.header.size);
4943 4944

	if (ret)
4945
		goto out;
4946

4947 4948
	comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
	comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4949

4950
	perf_output_put(&handle, comm_event->event_id);
4951
	__output_copy(&handle, comm_event->comm,
4952
				   comm_event->comm_size);
4953 4954 4955

	perf_event__output_id_sample(event, &handle, &sample);

4956
	perf_output_end(&handle);
4957 4958
out:
	comm_event->event_id.header.size = size;
4959 4960
}

4961
static void perf_event_comm_event(struct perf_comm_event *comm_event)
4962
{
4963
	char comm[TASK_COMM_LEN];
4964 4965
	unsigned int size;

4966
	memset(comm, 0, sizeof(comm));
4967
	strlcpy(comm, comm_event->task->comm, sizeof(comm));
4968
	size = ALIGN(strlen(comm)+1, sizeof(u64));
4969 4970 4971 4972

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

4973
	comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
P
Peter Zijlstra 已提交
4974

4975
	perf_event_aux(perf_event_comm_output,
4976 4977
		       comm_event,
		       NULL);
4978 4979
}

4980
void perf_event_comm(struct task_struct *task)
4981
{
4982
	struct perf_comm_event comm_event;
P
Peter Zijlstra 已提交
4983 4984
	struct perf_event_context *ctx;
	int ctxn;
4985

4986
	rcu_read_lock();
P
Peter Zijlstra 已提交
4987 4988 4989 4990
	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;
4991

P
Peter Zijlstra 已提交
4992 4993
		perf_event_enable_on_exec(ctx);
	}
4994
	rcu_read_unlock();
4995

4996
	if (!atomic_read(&nr_comm_events))
4997
		return;
4998

4999
	comm_event = (struct perf_comm_event){
5000
		.task	= task,
5001 5002
		/* .comm      */
		/* .comm_size */
5003
		.event_id  = {
5004
			.header = {
5005
				.type = PERF_RECORD_COMM,
5006 5007 5008 5009 5010
				.misc = 0,
				/* .size */
			},
			/* .pid */
			/* .tid */
5011 5012 5013
		},
	};

5014
	perf_event_comm_event(&comm_event);
5015 5016
}

5017 5018 5019 5020 5021
/*
 * mmap tracking
 */

struct perf_mmap_event {
5022 5023 5024 5025
	struct vm_area_struct	*vma;

	const char		*file_name;
	int			file_size;
5026 5027 5028
	int			maj, min;
	u64			ino;
	u64			ino_generation;
5029 5030 5031 5032 5033 5034 5035 5036 5037

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
		u64				start;
		u64				len;
		u64				pgoff;
5038
	} event_id;
5039 5040
};

5041 5042 5043 5044 5045 5046 5047 5048
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) ||
5049
	       (executable && (event->attr.mmap || event->attr.mmap2));
5050 5051
}

5052
static void perf_event_mmap_output(struct perf_event *event,
5053
				   void *data)
5054
{
5055
	struct perf_mmap_event *mmap_event = data;
5056
	struct perf_output_handle handle;
5057
	struct perf_sample_data sample;
5058
	int size = mmap_event->event_id.header.size;
5059
	int ret;
5060

5061 5062 5063
	if (!perf_event_mmap_match(event, data))
		return;

5064 5065 5066 5067 5068
	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);
5069
		mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5070 5071
	}

5072 5073
	perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5074
				mmap_event->event_id.header.size);
5075
	if (ret)
5076
		goto out;
5077

5078 5079
	mmap_event->event_id.pid = perf_event_pid(event, current);
	mmap_event->event_id.tid = perf_event_tid(event, current);
5080

5081
	perf_output_put(&handle, mmap_event->event_id);
5082 5083 5084 5085 5086 5087 5088 5089

	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);
	}

5090
	__output_copy(&handle, mmap_event->file_name,
5091
				   mmap_event->file_size);
5092 5093 5094

	perf_event__output_id_sample(event, &handle, &sample);

5095
	perf_output_end(&handle);
5096 5097
out:
	mmap_event->event_id.header.size = size;
5098 5099
}

5100
static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5101
{
5102 5103
	struct vm_area_struct *vma = mmap_event->vma;
	struct file *file = vma->vm_file;
5104 5105
	int maj = 0, min = 0;
	u64 ino = 0, gen = 0;
5106 5107 5108
	unsigned int size;
	char tmp[16];
	char *buf = NULL;
5109
	char *name;
5110

5111
	if (file) {
5112 5113
		struct inode *inode;
		dev_t dev;
5114

5115
		buf = kmalloc(PATH_MAX, GFP_KERNEL);
5116 5117 5118 5119
		if (!buf) {
			name = strncpy(tmp, "//enomem", sizeof(tmp));
			goto got_name;
		}
5120 5121 5122 5123 5124 5125
		/*
		 * 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));
5126 5127 5128 5129
		if (IS_ERR(name)) {
			name = strncpy(tmp, "//toolong", sizeof(tmp));
			goto got_name;
		}
5130 5131 5132 5133 5134 5135 5136
		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);

5137
	} else {
5138
		name = (char *)arch_vma_name(vma);
5139 5140
		if (name) {
			name = strncpy(tmp, name, sizeof(tmp) - 1);
5141
			tmp[sizeof(tmp) - 1] = '\0';
5142
			goto got_name;
5143
		}
5144

5145
		if (vma->vm_start <= vma->vm_mm->start_brk &&
5146 5147 5148
				vma->vm_end >= vma->vm_mm->brk) {
			name = strncpy(tmp, "[heap]", sizeof(tmp));
			goto got_name;
5149 5150
		}
		if (vma->vm_start <= vma->vm_mm->start_stack &&
5151 5152 5153
				vma->vm_end >= vma->vm_mm->start_stack) {
			name = strncpy(tmp, "[stack]", sizeof(tmp));
			goto got_name;
5154 5155
		}

5156 5157 5158 5159 5160
		name = strncpy(tmp, "//anon", sizeof(tmp));
		goto got_name;
	}

got_name:
5161 5162 5163 5164 5165 5166 5167 5168
	/*
	 * 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';
5169 5170 5171

	mmap_event->file_name = name;
	mmap_event->file_size = size;
5172 5173 5174 5175
	mmap_event->maj = maj;
	mmap_event->min = min;
	mmap_event->ino = ino;
	mmap_event->ino_generation = gen;
5176

5177 5178 5179
	if (!(vma->vm_flags & VM_EXEC))
		mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;

5180
	mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5181

5182
	perf_event_aux(perf_event_mmap_output,
5183 5184
		       mmap_event,
		       NULL);
5185

5186 5187 5188
	kfree(buf);
}

5189
void perf_event_mmap(struct vm_area_struct *vma)
5190
{
5191 5192
	struct perf_mmap_event mmap_event;

5193
	if (!atomic_read(&nr_mmap_events))
5194 5195 5196
		return;

	mmap_event = (struct perf_mmap_event){
5197
		.vma	= vma,
5198 5199
		/* .file_name */
		/* .file_size */
5200
		.event_id  = {
5201
			.header = {
5202
				.type = PERF_RECORD_MMAP,
5203
				.misc = PERF_RECORD_MISC_USER,
5204 5205 5206 5207
				/* .size */
			},
			/* .pid */
			/* .tid */
5208 5209
			.start  = vma->vm_start,
			.len    = vma->vm_end - vma->vm_start,
5210
			.pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
5211
		},
5212 5213 5214 5215
		/* .maj (attr_mmap2 only) */
		/* .min (attr_mmap2 only) */
		/* .ino (attr_mmap2 only) */
		/* .ino_generation (attr_mmap2 only) */
5216 5217
	};

5218
	perf_event_mmap_event(&mmap_event);
5219 5220
}

5221 5222 5223 5224
/*
 * IRQ throttle logging
 */

5225
static void perf_log_throttle(struct perf_event *event, int enable)
5226 5227
{
	struct perf_output_handle handle;
5228
	struct perf_sample_data sample;
5229 5230 5231 5232 5233
	int ret;

	struct {
		struct perf_event_header	header;
		u64				time;
5234
		u64				id;
5235
		u64				stream_id;
5236 5237
	} throttle_event = {
		.header = {
5238
			.type = PERF_RECORD_THROTTLE,
5239 5240 5241
			.misc = 0,
			.size = sizeof(throttle_event),
		},
P
Peter Zijlstra 已提交
5242
		.time		= perf_clock(),
5243 5244
		.id		= primary_event_id(event),
		.stream_id	= event->id,
5245 5246
	};

5247
	if (enable)
5248
		throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5249

5250 5251 5252
	perf_event_header__init_id(&throttle_event.header, &sample, event);

	ret = perf_output_begin(&handle, event,
5253
				throttle_event.header.size);
5254 5255 5256 5257
	if (ret)
		return;

	perf_output_put(&handle, throttle_event);
5258
	perf_event__output_id_sample(event, &handle, &sample);
5259 5260 5261
	perf_output_end(&handle);
}

5262
/*
5263
 * Generic event overflow handling, sampling.
5264 5265
 */

5266
static int __perf_event_overflow(struct perf_event *event,
5267 5268
				   int throttle, struct perf_sample_data *data,
				   struct pt_regs *regs)
5269
{
5270 5271
	int events = atomic_read(&event->event_limit);
	struct hw_perf_event *hwc = &event->hw;
5272
	u64 seq;
5273 5274
	int ret = 0;

5275 5276 5277 5278 5279 5280 5281
	/*
	 * Non-sampling counters might still use the PMI to fold short
	 * hardware counters, ignore those.
	 */
	if (unlikely(!is_sampling_event(event)))
		return 0;

5282 5283 5284 5285 5286 5287 5288 5289 5290
	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
5291 5292
			hwc->interrupts = MAX_INTERRUPTS;
			perf_log_throttle(event, 0);
5293
			tick_nohz_full_kick();
5294 5295
			ret = 1;
		}
5296
	}
5297

5298
	if (event->attr.freq) {
P
Peter Zijlstra 已提交
5299
		u64 now = perf_clock();
5300
		s64 delta = now - hwc->freq_time_stamp;
5301

5302
		hwc->freq_time_stamp = now;
5303

5304
		if (delta > 0 && delta < 2*TICK_NSEC)
5305
			perf_adjust_period(event, delta, hwc->last_period, true);
5306 5307
	}

5308 5309
	/*
	 * XXX event_limit might not quite work as expected on inherited
5310
	 * events
5311 5312
	 */

5313 5314
	event->pending_kill = POLL_IN;
	if (events && atomic_dec_and_test(&event->event_limit)) {
5315
		ret = 1;
5316
		event->pending_kill = POLL_HUP;
5317 5318
		event->pending_disable = 1;
		irq_work_queue(&event->pending);
5319 5320
	}

5321
	if (event->overflow_handler)
5322
		event->overflow_handler(event, data, regs);
5323
	else
5324
		perf_event_output(event, data, regs);
5325

P
Peter Zijlstra 已提交
5326
	if (event->fasync && event->pending_kill) {
5327 5328
		event->pending_wakeup = 1;
		irq_work_queue(&event->pending);
P
Peter Zijlstra 已提交
5329 5330
	}

5331
	return ret;
5332 5333
}

5334
int perf_event_overflow(struct perf_event *event,
5335 5336
			  struct perf_sample_data *data,
			  struct pt_regs *regs)
5337
{
5338
	return __perf_event_overflow(event, 1, data, regs);
5339 5340
}

5341
/*
5342
 * Generic software event infrastructure
5343 5344
 */

5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355
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);

5356
/*
5357 5358
 * We directly increment event->count and keep a second value in
 * event->hw.period_left to count intervals. This period event
5359 5360 5361 5362
 * is kept in the range [-sample_period, 0] so that we can use the
 * sign as trigger.
 */

5363
u64 perf_swevent_set_period(struct perf_event *event)
5364
{
5365
	struct hw_perf_event *hwc = &event->hw;
5366 5367 5368 5369 5370
	u64 period = hwc->last_period;
	u64 nr, offset;
	s64 old, val;

	hwc->last_period = hwc->sample_period;
5371 5372

again:
5373
	old = val = local64_read(&hwc->period_left);
5374 5375
	if (val < 0)
		return 0;
5376

5377 5378 5379
	nr = div64_u64(period + val, period);
	offset = nr * period;
	val -= offset;
5380
	if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5381
		goto again;
5382

5383
	return nr;
5384 5385
}

5386
static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5387
				    struct perf_sample_data *data,
5388
				    struct pt_regs *regs)
5389
{
5390
	struct hw_perf_event *hwc = &event->hw;
5391
	int throttle = 0;
5392

5393 5394
	if (!overflow)
		overflow = perf_swevent_set_period(event);
5395

5396 5397
	if (hwc->interrupts == MAX_INTERRUPTS)
		return;
5398

5399
	for (; overflow; overflow--) {
5400
		if (__perf_event_overflow(event, throttle,
5401
					    data, regs)) {
5402 5403 5404 5405 5406 5407
			/*
			 * We inhibit the overflow from happening when
			 * hwc->interrupts == MAX_INTERRUPTS.
			 */
			break;
		}
5408
		throttle = 1;
5409
	}
5410 5411
}

P
Peter Zijlstra 已提交
5412
static void perf_swevent_event(struct perf_event *event, u64 nr,
5413
			       struct perf_sample_data *data,
5414
			       struct pt_regs *regs)
5415
{
5416
	struct hw_perf_event *hwc = &event->hw;
5417

5418
	local64_add(nr, &event->count);
5419

5420 5421 5422
	if (!regs)
		return;

5423
	if (!is_sampling_event(event))
5424
		return;
5425

5426 5427 5428 5429 5430 5431
	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;

5432
	if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5433
		return perf_swevent_overflow(event, 1, data, regs);
5434

5435
	if (local64_add_negative(nr, &hwc->period_left))
5436
		return;
5437

5438
	perf_swevent_overflow(event, 0, data, regs);
5439 5440
}

5441 5442 5443
static int perf_exclude_event(struct perf_event *event,
			      struct pt_regs *regs)
{
P
Peter Zijlstra 已提交
5444
	if (event->hw.state & PERF_HES_STOPPED)
5445
		return 1;
P
Peter Zijlstra 已提交
5446

5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457
	if (regs) {
		if (event->attr.exclude_user && user_mode(regs))
			return 1;

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

	return 0;
}

5458
static int perf_swevent_match(struct perf_event *event,
P
Peter Zijlstra 已提交
5459
				enum perf_type_id type,
L
Li Zefan 已提交
5460 5461 5462
				u32 event_id,
				struct perf_sample_data *data,
				struct pt_regs *regs)
5463
{
5464
	if (event->attr.type != type)
5465
		return 0;
5466

5467
	if (event->attr.config != event_id)
5468 5469
		return 0;

5470 5471
	if (perf_exclude_event(event, regs))
		return 0;
5472 5473 5474 5475

	return 1;
}

5476 5477 5478 5479 5480 5481 5482
static inline u64 swevent_hash(u64 type, u32 event_id)
{
	u64 val = event_id | (type << 32);

	return hash_64(val, SWEVENT_HLIST_BITS);
}

5483 5484
static inline struct hlist_head *
__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5485
{
5486 5487 5488 5489
	u64 hash = swevent_hash(type, event_id);

	return &hlist->heads[hash];
}
5490

5491 5492
/* For the read side: events when they trigger */
static inline struct hlist_head *
5493
find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5494 5495
{
	struct swevent_hlist *hlist;
5496

5497
	hlist = rcu_dereference(swhash->swevent_hlist);
5498 5499 5500
	if (!hlist)
		return NULL;

5501 5502 5503 5504 5505
	return __find_swevent_head(hlist, type, event_id);
}

/* For the event head insertion and removal in the hlist */
static inline struct hlist_head *
5506
find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5507 5508 5509 5510 5511 5512 5513 5514 5515 5516
{
	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.
	 */
5517
	hlist = rcu_dereference_protected(swhash->swevent_hlist,
5518 5519 5520 5521 5522
					  lockdep_is_held(&event->ctx->lock));
	if (!hlist)
		return NULL;

	return __find_swevent_head(hlist, type, event_id);
5523 5524 5525
}

static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5526
				    u64 nr,
5527 5528
				    struct perf_sample_data *data,
				    struct pt_regs *regs)
5529
{
5530
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5531
	struct perf_event *event;
5532
	struct hlist_head *head;
5533

5534
	rcu_read_lock();
5535
	head = find_swevent_head_rcu(swhash, type, event_id);
5536 5537 5538
	if (!head)
		goto end;

5539
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
L
Li Zefan 已提交
5540
		if (perf_swevent_match(event, type, event_id, data, regs))
5541
			perf_swevent_event(event, nr, data, regs);
5542
	}
5543 5544
end:
	rcu_read_unlock();
5545 5546
}

5547
int perf_swevent_get_recursion_context(void)
P
Peter Zijlstra 已提交
5548
{
5549
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
P
Peter Zijlstra 已提交
5550

5551
	return get_recursion_context(swhash->recursion);
P
Peter Zijlstra 已提交
5552
}
I
Ingo Molnar 已提交
5553
EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
P
Peter Zijlstra 已提交
5554

5555
inline void perf_swevent_put_recursion_context(int rctx)
5556
{
5557
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5558

5559
	put_recursion_context(swhash->recursion, rctx);
5560
}
5561

5562
void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5563
{
5564
	struct perf_sample_data data;
5565 5566
	int rctx;

5567
	preempt_disable_notrace();
5568 5569 5570
	rctx = perf_swevent_get_recursion_context();
	if (rctx < 0)
		return;
5571

5572
	perf_sample_data_init(&data, addr, 0);
5573

5574
	do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5575 5576

	perf_swevent_put_recursion_context(rctx);
5577
	preempt_enable_notrace();
5578 5579
}

5580
static void perf_swevent_read(struct perf_event *event)
5581 5582 5583
{
}

P
Peter Zijlstra 已提交
5584
static int perf_swevent_add(struct perf_event *event, int flags)
5585
{
5586
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5587
	struct hw_perf_event *hwc = &event->hw;
5588 5589
	struct hlist_head *head;

5590
	if (is_sampling_event(event)) {
5591
		hwc->last_period = hwc->sample_period;
5592
		perf_swevent_set_period(event);
5593
	}
5594

P
Peter Zijlstra 已提交
5595 5596
	hwc->state = !(flags & PERF_EF_START);

5597
	head = find_swevent_head(swhash, event);
5598 5599 5600 5601 5602
	if (WARN_ON_ONCE(!head))
		return -EINVAL;

	hlist_add_head_rcu(&event->hlist_entry, head);

5603 5604 5605
	return 0;
}

P
Peter Zijlstra 已提交
5606
static void perf_swevent_del(struct perf_event *event, int flags)
5607
{
5608
	hlist_del_rcu(&event->hlist_entry);
5609 5610
}

P
Peter Zijlstra 已提交
5611
static void perf_swevent_start(struct perf_event *event, int flags)
5612
{
P
Peter Zijlstra 已提交
5613
	event->hw.state = 0;
5614
}
I
Ingo Molnar 已提交
5615

P
Peter Zijlstra 已提交
5616
static void perf_swevent_stop(struct perf_event *event, int flags)
5617
{
P
Peter Zijlstra 已提交
5618
	event->hw.state = PERF_HES_STOPPED;
5619 5620
}

5621 5622
/* Deref the hlist from the update side */
static inline struct swevent_hlist *
5623
swevent_hlist_deref(struct swevent_htable *swhash)
5624
{
5625 5626
	return rcu_dereference_protected(swhash->swevent_hlist,
					 lockdep_is_held(&swhash->hlist_mutex));
5627 5628
}

5629
static void swevent_hlist_release(struct swevent_htable *swhash)
5630
{
5631
	struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5632

5633
	if (!hlist)
5634 5635
		return;

5636
	rcu_assign_pointer(swhash->swevent_hlist, NULL);
5637
	kfree_rcu(hlist, rcu_head);
5638 5639 5640 5641
}

static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
{
5642
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5643

5644
	mutex_lock(&swhash->hlist_mutex);
5645

5646 5647
	if (!--swhash->hlist_refcount)
		swevent_hlist_release(swhash);
5648

5649
	mutex_unlock(&swhash->hlist_mutex);
5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666
}

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)
{
5667
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5668 5669
	int err = 0;

5670
	mutex_lock(&swhash->hlist_mutex);
5671

5672
	if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5673 5674 5675 5676 5677 5678 5679
		struct swevent_hlist *hlist;

		hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
		if (!hlist) {
			err = -ENOMEM;
			goto exit;
		}
5680
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
5681
	}
5682
	swhash->hlist_refcount++;
P
Peter Zijlstra 已提交
5683
exit:
5684
	mutex_unlock(&swhash->hlist_mutex);
5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707

	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
5708
fail:
5709 5710 5711 5712 5713 5714 5715 5716 5717 5718
	for_each_possible_cpu(cpu) {
		if (cpu == failed_cpu)
			break;
		swevent_hlist_put_cpu(event, cpu);
	}

	put_online_cpus();
	return err;
}

5719
struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5720

5721 5722 5723
static void sw_perf_event_destroy(struct perf_event *event)
{
	u64 event_id = event->attr.config;
5724

5725 5726
	WARN_ON(event->parent);

5727
	static_key_slow_dec(&perf_swevent_enabled[event_id]);
5728 5729 5730 5731 5732
	swevent_hlist_put(event);
}

static int perf_swevent_init(struct perf_event *event)
{
5733
	u64 event_id = event->attr.config;
5734 5735 5736 5737

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

5738 5739 5740 5741 5742 5743
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

5744 5745 5746 5747 5748 5749 5750 5751 5752
	switch (event_id) {
	case PERF_COUNT_SW_CPU_CLOCK:
	case PERF_COUNT_SW_TASK_CLOCK:
		return -ENOENT;

	default:
		break;
	}

5753
	if (event_id >= PERF_COUNT_SW_MAX)
5754 5755 5756 5757 5758 5759 5760 5761 5762
		return -ENOENT;

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

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

5763
		static_key_slow_inc(&perf_swevent_enabled[event_id]);
5764 5765 5766 5767 5768 5769
		event->destroy = sw_perf_event_destroy;
	}

	return 0;
}

5770 5771 5772 5773 5774
static int perf_swevent_event_idx(struct perf_event *event)
{
	return 0;
}

5775
static struct pmu perf_swevent = {
5776
	.task_ctx_nr	= perf_sw_context,
5777

5778
	.event_init	= perf_swevent_init,
P
Peter Zijlstra 已提交
5779 5780 5781 5782
	.add		= perf_swevent_add,
	.del		= perf_swevent_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
5783
	.read		= perf_swevent_read,
5784 5785

	.event_idx	= perf_swevent_event_idx,
5786 5787
};

5788 5789
#ifdef CONFIG_EVENT_TRACING

5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803
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)
{
5804 5805
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;
5806 5807 5808 5809
	/*
	 * All tracepoints are from kernel-space.
	 */
	if (event->attr.exclude_kernel)
5810 5811 5812 5813 5814 5815 5816 5817 5818
		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,
5819 5820
		   struct pt_regs *regs, struct hlist_head *head, int rctx,
		   struct task_struct *task)
5821 5822
{
	struct perf_sample_data data;
5823 5824
	struct perf_event *event;

5825 5826 5827 5828 5829
	struct perf_raw_record raw = {
		.size = entry_size,
		.data = record,
	};

5830
	perf_sample_data_init(&data, addr, 0);
5831 5832
	data.raw = &raw;

5833
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
5834
		if (perf_tp_event_match(event, &data, regs))
5835
			perf_swevent_event(event, count, &data, regs);
5836
	}
5837

5838 5839 5840 5841 5842 5843 5844 5845 5846 5847 5848 5849 5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862
	/*
	 * 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();
	}

5863
	perf_swevent_put_recursion_context(rctx);
5864 5865 5866
}
EXPORT_SYMBOL_GPL(perf_tp_event);

5867
static void tp_perf_event_destroy(struct perf_event *event)
5868
{
5869
	perf_trace_destroy(event);
5870 5871
}

5872
static int perf_tp_event_init(struct perf_event *event)
5873
{
5874 5875
	int err;

5876 5877 5878
	if (event->attr.type != PERF_TYPE_TRACEPOINT)
		return -ENOENT;

5879 5880 5881 5882 5883 5884
	/*
	 * no branch sampling for tracepoint events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

5885 5886
	err = perf_trace_init(event);
	if (err)
5887
		return err;
5888

5889
	event->destroy = tp_perf_event_destroy;
5890

5891 5892 5893 5894
	return 0;
}

static struct pmu perf_tracepoint = {
5895 5896
	.task_ctx_nr	= perf_sw_context,

5897
	.event_init	= perf_tp_event_init,
P
Peter Zijlstra 已提交
5898 5899 5900 5901
	.add		= perf_trace_add,
	.del		= perf_trace_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
5902
	.read		= perf_swevent_read,
5903 5904

	.event_idx	= perf_swevent_event_idx,
5905 5906 5907 5908
};

static inline void perf_tp_register(void)
{
P
Peter Zijlstra 已提交
5909
	perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5910
}
L
Li Zefan 已提交
5911 5912 5913 5914 5915 5916 5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934

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);
}

5935
#else
L
Li Zefan 已提交
5936

5937
static inline void perf_tp_register(void)
5938 5939
{
}
L
Li Zefan 已提交
5940 5941 5942 5943 5944 5945 5946 5947 5948 5949

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)
{
}

5950
#endif /* CONFIG_EVENT_TRACING */
5951

5952
#ifdef CONFIG_HAVE_HW_BREAKPOINT
5953
void perf_bp_event(struct perf_event *bp, void *data)
5954
{
5955 5956 5957
	struct perf_sample_data sample;
	struct pt_regs *regs = data;

5958
	perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5959

P
Peter Zijlstra 已提交
5960
	if (!bp->hw.state && !perf_exclude_event(bp, regs))
5961
		perf_swevent_event(bp, 1, &sample, regs);
5962 5963 5964
}
#endif

5965 5966 5967
/*
 * hrtimer based swevent callback
 */
5968

5969
static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5970
{
5971 5972 5973 5974 5975
	enum hrtimer_restart ret = HRTIMER_RESTART;
	struct perf_sample_data data;
	struct pt_regs *regs;
	struct perf_event *event;
	u64 period;
5976

5977
	event = container_of(hrtimer, struct perf_event, hw.hrtimer);
P
Peter Zijlstra 已提交
5978 5979 5980 5981

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

5982
	event->pmu->read(event);
5983

5984
	perf_sample_data_init(&data, 0, event->hw.last_period);
5985 5986 5987
	regs = get_irq_regs();

	if (regs && !perf_exclude_event(event, regs)) {
5988
		if (!(event->attr.exclude_idle && is_idle_task(current)))
5989
			if (__perf_event_overflow(event, 1, &data, regs))
5990 5991
				ret = HRTIMER_NORESTART;
	}
5992

5993 5994
	period = max_t(u64, 10000, event->hw.sample_period);
	hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5995

5996
	return ret;
5997 5998
}

5999
static void perf_swevent_start_hrtimer(struct perf_event *event)
6000
{
6001
	struct hw_perf_event *hwc = &event->hw;
6002 6003 6004 6005
	s64 period;

	if (!is_sampling_event(event))
		return;
6006

6007 6008 6009 6010
	period = local64_read(&hwc->period_left);
	if (period) {
		if (period < 0)
			period = 10000;
P
Peter Zijlstra 已提交
6011

6012 6013 6014 6015 6016
		local64_set(&hwc->period_left, 0);
	} else {
		period = max_t(u64, 10000, hwc->sample_period);
	}
	__hrtimer_start_range_ns(&hwc->hrtimer,
6017
				ns_to_ktime(period), 0,
6018
				HRTIMER_MODE_REL_PINNED, 0);
6019
}
6020 6021

static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6022
{
6023 6024
	struct hw_perf_event *hwc = &event->hw;

6025
	if (is_sampling_event(event)) {
6026
		ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
P
Peter Zijlstra 已提交
6027
		local64_set(&hwc->period_left, ktime_to_ns(remaining));
6028 6029 6030

		hrtimer_cancel(&hwc->hrtimer);
	}
6031 6032
}

P
Peter Zijlstra 已提交
6033 6034 6035 6036 6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052
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);
6053
		hwc->last_period = hwc->sample_period;
P
Peter Zijlstra 已提交
6054 6055 6056 6057
		event->attr.freq = 0;
	}
}

6058 6059 6060 6061 6062
/*
 * Software event: cpu wall time clock
 */

static void cpu_clock_event_update(struct perf_event *event)
6063
{
6064 6065 6066
	s64 prev;
	u64 now;

P
Peter Zijlstra 已提交
6067
	now = local_clock();
6068 6069
	prev = local64_xchg(&event->hw.prev_count, now);
	local64_add(now - prev, &event->count);
6070 6071
}

P
Peter Zijlstra 已提交
6072
static void cpu_clock_event_start(struct perf_event *event, int flags)
6073
{
P
Peter Zijlstra 已提交
6074
	local64_set(&event->hw.prev_count, local_clock());
6075 6076 6077
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6078
static void cpu_clock_event_stop(struct perf_event *event, int flags)
6079
{
6080 6081 6082
	perf_swevent_cancel_hrtimer(event);
	cpu_clock_event_update(event);
}
6083

P
Peter Zijlstra 已提交
6084 6085 6086 6087 6088 6089 6090 6091 6092 6093 6094 6095 6096
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);
}

6097 6098 6099 6100
static void cpu_clock_event_read(struct perf_event *event)
{
	cpu_clock_event_update(event);
}
6101

6102 6103 6104 6105 6106 6107 6108 6109
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;

6110 6111 6112 6113 6114 6115
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6116 6117
	perf_swevent_init_hrtimer(event);

6118
	return 0;
6119 6120
}

6121
static struct pmu perf_cpu_clock = {
6122 6123
	.task_ctx_nr	= perf_sw_context,

6124
	.event_init	= cpu_clock_event_init,
P
Peter Zijlstra 已提交
6125 6126 6127 6128
	.add		= cpu_clock_event_add,
	.del		= cpu_clock_event_del,
	.start		= cpu_clock_event_start,
	.stop		= cpu_clock_event_stop,
6129
	.read		= cpu_clock_event_read,
6130 6131

	.event_idx	= perf_swevent_event_idx,
6132 6133 6134 6135 6136 6137 6138
};

/*
 * Software event: task time clock
 */

static void task_clock_event_update(struct perf_event *event, u64 now)
6139
{
6140 6141
	u64 prev;
	s64 delta;
6142

6143 6144 6145 6146
	prev = local64_xchg(&event->hw.prev_count, now);
	delta = now - prev;
	local64_add(delta, &event->count);
}
6147

P
Peter Zijlstra 已提交
6148
static void task_clock_event_start(struct perf_event *event, int flags)
6149
{
P
Peter Zijlstra 已提交
6150
	local64_set(&event->hw.prev_count, event->ctx->time);
6151 6152 6153
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6154
static void task_clock_event_stop(struct perf_event *event, int flags)
6155 6156 6157
{
	perf_swevent_cancel_hrtimer(event);
	task_clock_event_update(event, event->ctx->time);
P
Peter Zijlstra 已提交
6158 6159 6160 6161 6162 6163
}

static int task_clock_event_add(struct perf_event *event, int flags)
{
	if (flags & PERF_EF_START)
		task_clock_event_start(event, flags);
6164

P
Peter Zijlstra 已提交
6165 6166 6167 6168 6169 6170
	return 0;
}

static void task_clock_event_del(struct perf_event *event, int flags)
{
	task_clock_event_stop(event, PERF_EF_UPDATE);
6171 6172 6173 6174
}

static void task_clock_event_read(struct perf_event *event)
{
6175 6176 6177
	u64 now = perf_clock();
	u64 delta = now - event->ctx->timestamp;
	u64 time = event->ctx->time + delta;
6178 6179 6180 6181 6182

	task_clock_event_update(event, time);
}

static int task_clock_event_init(struct perf_event *event)
L
Li Zefan 已提交
6183
{
6184 6185 6186 6187 6188 6189
	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

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

6190 6191 6192 6193 6194 6195
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6196 6197
	perf_swevent_init_hrtimer(event);

6198
	return 0;
L
Li Zefan 已提交
6199 6200
}

6201
static struct pmu perf_task_clock = {
6202 6203
	.task_ctx_nr	= perf_sw_context,

6204
	.event_init	= task_clock_event_init,
P
Peter Zijlstra 已提交
6205 6206 6207 6208
	.add		= task_clock_event_add,
	.del		= task_clock_event_del,
	.start		= task_clock_event_start,
	.stop		= task_clock_event_stop,
6209
	.read		= task_clock_event_read,
6210 6211

	.event_idx	= perf_swevent_event_idx,
6212
};
L
Li Zefan 已提交
6213

P
Peter Zijlstra 已提交
6214
static void perf_pmu_nop_void(struct pmu *pmu)
6215 6216
{
}
L
Li Zefan 已提交
6217

P
Peter Zijlstra 已提交
6218
static int perf_pmu_nop_int(struct pmu *pmu)
L
Li Zefan 已提交
6219
{
P
Peter Zijlstra 已提交
6220
	return 0;
L
Li Zefan 已提交
6221 6222
}

P
Peter Zijlstra 已提交
6223
static void perf_pmu_start_txn(struct pmu *pmu)
L
Li Zefan 已提交
6224
{
P
Peter Zijlstra 已提交
6225
	perf_pmu_disable(pmu);
L
Li Zefan 已提交
6226 6227
}

P
Peter Zijlstra 已提交
6228 6229 6230 6231 6232
static int perf_pmu_commit_txn(struct pmu *pmu)
{
	perf_pmu_enable(pmu);
	return 0;
}
6233

P
Peter Zijlstra 已提交
6234
static void perf_pmu_cancel_txn(struct pmu *pmu)
6235
{
P
Peter Zijlstra 已提交
6236
	perf_pmu_enable(pmu);
6237 6238
}

6239 6240 6241 6242 6243
static int perf_event_idx_default(struct perf_event *event)
{
	return event->hw.idx + 1;
}

P
Peter Zijlstra 已提交
6244 6245 6246 6247 6248
/*
 * Ensures all contexts with the same task_ctx_nr have the same
 * pmu_cpu_context too.
 */
static void *find_pmu_context(int ctxn)
6249
{
P
Peter Zijlstra 已提交
6250
	struct pmu *pmu;
6251

P
Peter Zijlstra 已提交
6252 6253
	if (ctxn < 0)
		return NULL;
6254

P
Peter Zijlstra 已提交
6255 6256 6257 6258
	list_for_each_entry(pmu, &pmus, entry) {
		if (pmu->task_ctx_nr == ctxn)
			return pmu->pmu_cpu_context;
	}
6259

P
Peter Zijlstra 已提交
6260
	return NULL;
6261 6262
}

6263
static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6264
{
6265 6266 6267 6268 6269 6270 6271
	int cpu;

	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

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

6272 6273
		if (cpuctx->unique_pmu == old_pmu)
			cpuctx->unique_pmu = pmu;
6274 6275 6276 6277 6278 6279
	}
}

static void free_pmu_context(struct pmu *pmu)
{
	struct pmu *i;
6280

P
Peter Zijlstra 已提交
6281
	mutex_lock(&pmus_lock);
6282
	/*
P
Peter Zijlstra 已提交
6283
	 * Like a real lame refcount.
6284
	 */
6285 6286 6287
	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
6288
			goto out;
6289
		}
P
Peter Zijlstra 已提交
6290
	}
6291

6292
	free_percpu(pmu->pmu_cpu_context);
P
Peter Zijlstra 已提交
6293 6294
out:
	mutex_unlock(&pmus_lock);
6295
}
P
Peter Zijlstra 已提交
6296
static struct idr pmu_idr;
6297

P
Peter Zijlstra 已提交
6298 6299 6300 6301 6302 6303 6304 6305
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);
}

6306 6307 6308 6309 6310 6311 6312 6313 6314 6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349
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
6350
static struct device_attribute pmu_dev_attrs[] = {
6351 6352 6353
	__ATTR_RO(type),
	__ATTR_RW(perf_event_mux_interval_ms),
	__ATTR_NULL,
P
Peter Zijlstra 已提交
6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374
};

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;

6375
	pmu->dev->groups = pmu->attr_groups;
P
Peter Zijlstra 已提交
6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395
	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;
}

6396
static struct lock_class_key cpuctx_mutex;
6397
static struct lock_class_key cpuctx_lock;
6398

6399
int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6400
{
P
Peter Zijlstra 已提交
6401
	int cpu, ret;
6402

6403
	mutex_lock(&pmus_lock);
P
Peter Zijlstra 已提交
6404 6405 6406 6407
	ret = -ENOMEM;
	pmu->pmu_disable_count = alloc_percpu(int);
	if (!pmu->pmu_disable_count)
		goto unlock;
6408

P
Peter Zijlstra 已提交
6409 6410 6411 6412 6413 6414
	pmu->type = -1;
	if (!name)
		goto skip_type;
	pmu->name = name;

	if (type < 0) {
T
Tejun Heo 已提交
6415 6416 6417
		type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
		if (type < 0) {
			ret = type;
P
Peter Zijlstra 已提交
6418 6419 6420 6421 6422
			goto free_pdc;
		}
	}
	pmu->type = type;

P
Peter Zijlstra 已提交
6423 6424 6425 6426 6427 6428
	if (pmu_bus_running) {
		ret = pmu_dev_alloc(pmu);
		if (ret)
			goto free_idr;
	}

P
Peter Zijlstra 已提交
6429
skip_type:
P
Peter Zijlstra 已提交
6430 6431 6432
	pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
	if (pmu->pmu_cpu_context)
		goto got_cpu_context;
6433

W
Wei Yongjun 已提交
6434
	ret = -ENOMEM;
P
Peter Zijlstra 已提交
6435 6436
	pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
	if (!pmu->pmu_cpu_context)
P
Peter Zijlstra 已提交
6437
		goto free_dev;
6438

P
Peter Zijlstra 已提交
6439 6440 6441 6442
	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6443
		__perf_event_init_context(&cpuctx->ctx);
6444
		lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6445
		lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6446
		cpuctx->ctx.type = cpu_context;
P
Peter Zijlstra 已提交
6447
		cpuctx->ctx.pmu = pmu;
6448 6449 6450

		__perf_cpu_hrtimer_init(cpuctx, cpu);

6451
		INIT_LIST_HEAD(&cpuctx->rotation_list);
6452
		cpuctx->unique_pmu = pmu;
P
Peter Zijlstra 已提交
6453
	}
6454

P
Peter Zijlstra 已提交
6455
got_cpu_context:
P
Peter Zijlstra 已提交
6456 6457 6458 6459 6460 6461 6462 6463 6464 6465 6466 6467 6468 6469
	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;
6470
		}
6471
	}
6472

P
Peter Zijlstra 已提交
6473 6474 6475 6476 6477
	if (!pmu->pmu_enable) {
		pmu->pmu_enable  = perf_pmu_nop_void;
		pmu->pmu_disable = perf_pmu_nop_void;
	}

6478 6479 6480
	if (!pmu->event_idx)
		pmu->event_idx = perf_event_idx_default;

6481
	list_add_rcu(&pmu->entry, &pmus);
P
Peter Zijlstra 已提交
6482 6483
	ret = 0;
unlock:
6484 6485
	mutex_unlock(&pmus_lock);

P
Peter Zijlstra 已提交
6486
	return ret;
P
Peter Zijlstra 已提交
6487

P
Peter Zijlstra 已提交
6488 6489 6490 6491
free_dev:
	device_del(pmu->dev);
	put_device(pmu->dev);

P
Peter Zijlstra 已提交
6492 6493 6494 6495
free_idr:
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);

P
Peter Zijlstra 已提交
6496 6497 6498
free_pdc:
	free_percpu(pmu->pmu_disable_count);
	goto unlock;
6499 6500
}

6501
void perf_pmu_unregister(struct pmu *pmu)
6502
{
6503 6504 6505
	mutex_lock(&pmus_lock);
	list_del_rcu(&pmu->entry);
	mutex_unlock(&pmus_lock);
6506

6507
	/*
P
Peter Zijlstra 已提交
6508 6509
	 * We dereference the pmu list under both SRCU and regular RCU, so
	 * synchronize against both of those.
6510
	 */
6511
	synchronize_srcu(&pmus_srcu);
P
Peter Zijlstra 已提交
6512
	synchronize_rcu();
6513

P
Peter Zijlstra 已提交
6514
	free_percpu(pmu->pmu_disable_count);
P
Peter Zijlstra 已提交
6515 6516
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);
P
Peter Zijlstra 已提交
6517 6518
	device_del(pmu->dev);
	put_device(pmu->dev);
6519
	free_pmu_context(pmu);
6520
}
6521

6522 6523 6524 6525
struct pmu *perf_init_event(struct perf_event *event)
{
	struct pmu *pmu = NULL;
	int idx;
6526
	int ret;
6527 6528

	idx = srcu_read_lock(&pmus_srcu);
P
Peter Zijlstra 已提交
6529 6530 6531 6532

	rcu_read_lock();
	pmu = idr_find(&pmu_idr, event->attr.type);
	rcu_read_unlock();
6533
	if (pmu) {
6534
		event->pmu = pmu;
6535 6536 6537
		ret = pmu->event_init(event);
		if (ret)
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6538
		goto unlock;
6539
	}
P
Peter Zijlstra 已提交
6540

6541
	list_for_each_entry_rcu(pmu, &pmus, entry) {
6542
		event->pmu = pmu;
6543
		ret = pmu->event_init(event);
6544
		if (!ret)
P
Peter Zijlstra 已提交
6545
			goto unlock;
6546

6547 6548
		if (ret != -ENOENT) {
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6549
			goto unlock;
6550
		}
6551
	}
P
Peter Zijlstra 已提交
6552 6553
	pmu = ERR_PTR(-ENOENT);
unlock:
6554
	srcu_read_unlock(&pmus_srcu, idx);
6555

6556
	return pmu;
6557 6558
}

6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571
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));
}

6572 6573
static void account_event(struct perf_event *event)
{
6574 6575 6576
	if (event->parent)
		return;

6577 6578 6579 6580 6581 6582 6583 6584
	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);
6585 6586 6587 6588
	if (event->attr.freq) {
		if (atomic_inc_return(&nr_freq_events) == 1)
			tick_nohz_full_kick_all();
	}
6589
	if (has_branch_stack(event))
6590
		static_key_slow_inc(&perf_sched_events.key);
6591
	if (is_cgroup_event(event))
6592
		static_key_slow_inc(&perf_sched_events.key);
6593 6594

	account_event_cpu(event, event->cpu);
6595 6596
}

T
Thomas Gleixner 已提交
6597
/*
6598
 * Allocate and initialize a event structure
T
Thomas Gleixner 已提交
6599
 */
6600
static struct perf_event *
6601
perf_event_alloc(struct perf_event_attr *attr, int cpu,
6602 6603 6604
		 struct task_struct *task,
		 struct perf_event *group_leader,
		 struct perf_event *parent_event,
6605 6606
		 perf_overflow_handler_t overflow_handler,
		 void *context)
T
Thomas Gleixner 已提交
6607
{
P
Peter Zijlstra 已提交
6608
	struct pmu *pmu;
6609 6610
	struct perf_event *event;
	struct hw_perf_event *hwc;
6611
	long err = -EINVAL;
T
Thomas Gleixner 已提交
6612

6613 6614 6615 6616 6617
	if ((unsigned)cpu >= nr_cpu_ids) {
		if (!task || cpu != -1)
			return ERR_PTR(-EINVAL);
	}

6618
	event = kzalloc(sizeof(*event), GFP_KERNEL);
6619
	if (!event)
6620
		return ERR_PTR(-ENOMEM);
T
Thomas Gleixner 已提交
6621

6622
	/*
6623
	 * Single events are their own group leaders, with an
6624 6625 6626
	 * empty sibling list:
	 */
	if (!group_leader)
6627
		group_leader = event;
6628

6629 6630
	mutex_init(&event->child_mutex);
	INIT_LIST_HEAD(&event->child_list);
6631

6632 6633 6634
	INIT_LIST_HEAD(&event->group_entry);
	INIT_LIST_HEAD(&event->event_entry);
	INIT_LIST_HEAD(&event->sibling_list);
6635 6636
	INIT_LIST_HEAD(&event->rb_entry);

6637
	init_waitqueue_head(&event->waitq);
6638
	init_irq_work(&event->pending, perf_pending_event);
T
Thomas Gleixner 已提交
6639

6640
	mutex_init(&event->mmap_mutex);
6641

6642
	atomic_long_set(&event->refcount, 1);
6643 6644 6645 6646 6647
	event->cpu		= cpu;
	event->attr		= *attr;
	event->group_leader	= group_leader;
	event->pmu		= NULL;
	event->oncpu		= -1;
6648

6649
	event->parent		= parent_event;
6650

6651
	event->ns		= get_pid_ns(task_active_pid_ns(current));
6652
	event->id		= atomic64_inc_return(&perf_event_id);
6653

6654
	event->state		= PERF_EVENT_STATE_INACTIVE;
6655

6656 6657
	if (task) {
		event->attach_state = PERF_ATTACH_TASK;
6658 6659 6660

		if (attr->type == PERF_TYPE_TRACEPOINT)
			event->hw.tp_target = task;
6661 6662 6663 6664
#ifdef CONFIG_HAVE_HW_BREAKPOINT
		/*
		 * hw_breakpoint is a bit difficult here..
		 */
6665
		else if (attr->type == PERF_TYPE_BREAKPOINT)
6666 6667 6668 6669
			event->hw.bp_target = task;
#endif
	}

6670
	if (!overflow_handler && parent_event) {
6671
		overflow_handler = parent_event->overflow_handler;
6672 6673
		context = parent_event->overflow_handler_context;
	}
6674

6675
	event->overflow_handler	= overflow_handler;
6676
	event->overflow_handler_context = context;
6677

J
Jiri Olsa 已提交
6678
	perf_event__state_init(event);
6679

6680
	pmu = NULL;
6681

6682
	hwc = &event->hw;
6683
	hwc->sample_period = attr->sample_period;
6684
	if (attr->freq && attr->sample_freq)
6685
		hwc->sample_period = 1;
6686
	hwc->last_period = hwc->sample_period;
6687

6688
	local64_set(&hwc->period_left, hwc->sample_period);
6689

6690
	/*
6691
	 * we currently do not support PERF_FORMAT_GROUP on inherited events
6692
	 */
6693
	if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6694
		goto err_ns;
6695

6696
	pmu = perf_init_event(event);
6697
	if (!pmu)
6698 6699
		goto err_ns;
	else if (IS_ERR(pmu)) {
6700
		err = PTR_ERR(pmu);
6701
		goto err_ns;
I
Ingo Molnar 已提交
6702
	}
6703

6704
	if (!event->parent) {
6705 6706
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
			err = get_callchain_buffers();
6707 6708
			if (err)
				goto err_pmu;
6709
		}
6710
	}
6711

6712
	return event;
6713 6714 6715 6716 6717 6718 6719 6720 6721 6722

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
6723 6724
}

6725 6726
static int perf_copy_attr(struct perf_event_attr __user *uattr,
			  struct perf_event_attr *attr)
6727 6728
{
	u32 size;
6729
	int ret;
6730 6731 6732 6733 6734 6735 6736 6737 6738 6739 6740 6741 6742 6743 6744 6745 6746 6747 6748 6749 6750 6751 6752 6753

	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,
6754 6755 6756
	 * 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.
6757 6758
	 */
	if (size > sizeof(*attr)) {
6759 6760 6761
		unsigned char __user *addr;
		unsigned char __user *end;
		unsigned char val;
6762

6763 6764
		addr = (void __user *)uattr + sizeof(*attr);
		end  = (void __user *)uattr + size;
6765

6766
		for (; addr < end; addr++) {
6767 6768 6769 6770 6771 6772
			ret = get_user(val, addr);
			if (ret)
				return ret;
			if (val)
				goto err_size;
		}
6773
		size = sizeof(*attr);
6774 6775 6776 6777 6778 6779
	}

	ret = copy_from_user(attr, uattr, size);
	if (ret)
		return -EFAULT;

6780 6781 6782 6783
	/* disabled for now */
	if (attr->mmap2)
		return -EINVAL;

6784
	if (attr->__reserved_1)
6785 6786 6787 6788 6789 6790 6791 6792
		return -EINVAL;

	if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
		return -EINVAL;

	if (attr->read_format & ~(PERF_FORMAT_MAX-1))
		return -EINVAL;

6793 6794 6795 6796 6797 6798 6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815 6816 6817 6818 6819 6820
	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;
		}
6821 6822
		/* privileged levels capture (kernel, hv): check permissions */
		if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6823 6824
		    && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
6825
	}
6826

6827
	if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6828
		ret = perf_reg_validate(attr->sample_regs_user);
6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846
		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;
	}
6847

6848 6849 6850 6851 6852 6853 6854 6855 6856
out:
	return ret;

err_size:
	put_user(sizeof(*attr), &uattr->size);
	ret = -E2BIG;
	goto out;
}

6857 6858
static int
perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6859
{
6860
	struct ring_buffer *rb = NULL, *old_rb = NULL;
6861 6862
	int ret = -EINVAL;

6863
	if (!output_event)
6864 6865
		goto set;

6866 6867
	/* don't allow circular references */
	if (event == output_event)
6868 6869
		goto out;

6870 6871 6872 6873 6874 6875 6876
	/*
	 * Don't allow cross-cpu buffers
	 */
	if (output_event->cpu != event->cpu)
		goto out;

	/*
6877
	 * If its not a per-cpu rb, it must be the same task.
6878 6879 6880 6881
	 */
	if (output_event->cpu == -1 && output_event->ctx != event->ctx)
		goto out;

6882
set:
6883
	mutex_lock(&event->mmap_mutex);
6884 6885 6886
	/* Can't redirect output if we've got an active mmap() */
	if (atomic_read(&event->mmap_count))
		goto unlock;
6887

6888 6889
	old_rb = event->rb;

6890
	if (output_event) {
6891 6892 6893
		/* get the rb we want to redirect to */
		rb = ring_buffer_get(output_event);
		if (!rb)
6894
			goto unlock;
6895 6896
	}

6897 6898
	if (old_rb)
		ring_buffer_detach(event, old_rb);
6899 6900 6901 6902 6903 6904 6905 6906 6907 6908 6909 6910 6911 6912 6913 6914

	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);
	}

6915
	ret = 0;
6916 6917 6918
unlock:
	mutex_unlock(&event->mmap_mutex);

6919 6920 6921 6922
out:
	return ret;
}

T
Thomas Gleixner 已提交
6923
/**
6924
 * sys_perf_event_open - open a performance event, associate it to a task/cpu
I
Ingo Molnar 已提交
6925
 *
6926
 * @attr_uptr:	event_id type attributes for monitoring/sampling
T
Thomas Gleixner 已提交
6927
 * @pid:		target pid
I
Ingo Molnar 已提交
6928
 * @cpu:		target cpu
6929
 * @group_fd:		group leader event fd
T
Thomas Gleixner 已提交
6930
 */
6931 6932
SYSCALL_DEFINE5(perf_event_open,
		struct perf_event_attr __user *, attr_uptr,
6933
		pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
T
Thomas Gleixner 已提交
6934
{
6935 6936
	struct perf_event *group_leader = NULL, *output_event = NULL;
	struct perf_event *event, *sibling;
6937 6938 6939
	struct perf_event_attr attr;
	struct perf_event_context *ctx;
	struct file *event_file = NULL;
6940
	struct fd group = {NULL, 0};
M
Matt Helsley 已提交
6941
	struct task_struct *task = NULL;
6942
	struct pmu *pmu;
6943
	int event_fd;
6944
	int move_group = 0;
6945
	int err;
T
Thomas Gleixner 已提交
6946

6947
	/* for future expandability... */
S
Stephane Eranian 已提交
6948
	if (flags & ~PERF_FLAG_ALL)
6949 6950
		return -EINVAL;

6951 6952 6953
	err = perf_copy_attr(attr_uptr, &attr);
	if (err)
		return err;
6954

6955 6956 6957 6958 6959
	if (!attr.exclude_kernel) {
		if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
	}

6960
	if (attr.freq) {
6961
		if (attr.sample_freq > sysctl_perf_event_sample_rate)
6962 6963 6964
			return -EINVAL;
	}

S
Stephane Eranian 已提交
6965 6966 6967 6968 6969 6970 6971 6972 6973
	/*
	 * 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;

6974
	event_fd = get_unused_fd();
6975 6976 6977
	if (event_fd < 0)
		return event_fd;

6978
	if (group_fd != -1) {
6979 6980
		err = perf_fget_light(group_fd, &group);
		if (err)
6981
			goto err_fd;
6982
		group_leader = group.file->private_data;
6983 6984 6985 6986 6987 6988
		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
6989
	if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6990 6991 6992 6993 6994 6995 6996
		task = find_lively_task_by_vpid(pid);
		if (IS_ERR(task)) {
			err = PTR_ERR(task);
			goto err_group_fd;
		}
	}

6997 6998
	get_online_cpus();

6999 7000
	event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
				 NULL, NULL);
7001 7002
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
7003
		goto err_task;
7004 7005
	}

S
Stephane Eranian 已提交
7006 7007
	if (flags & PERF_FLAG_PID_CGROUP) {
		err = perf_cgroup_connect(pid, event, &attr, group_leader);
7008 7009 7010 7011
		if (err) {
			__free_event(event);
			goto err_task;
		}
S
Stephane Eranian 已提交
7012 7013
	}

7014 7015
	account_event(event);

7016 7017 7018 7019 7020
	/*
	 * Special case software events and allow them to be part of
	 * any hardware group.
	 */
	pmu = event->pmu;
7021 7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032 7033 7034 7035 7036 7037 7038 7039 7040 7041 7042 7043

	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;
		}
	}
7044 7045 7046 7047

	/*
	 * Get the target context (task or percpu):
	 */
7048
	ctx = find_get_context(pmu, task, event->cpu);
7049 7050
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7051
		goto err_alloc;
7052 7053
	}

7054 7055 7056 7057 7058
	if (task) {
		put_task_struct(task);
		task = NULL;
	}

I
Ingo Molnar 已提交
7059
	/*
7060
	 * Look up the group leader (we will attach this event to it):
7061
	 */
7062
	if (group_leader) {
7063
		err = -EINVAL;
7064 7065

		/*
I
Ingo Molnar 已提交
7066 7067 7068 7069
		 * Do not allow a recursive hierarchy (this new sibling
		 * becoming part of another group-sibling):
		 */
		if (group_leader->group_leader != group_leader)
7070
			goto err_context;
I
Ingo Molnar 已提交
7071 7072 7073
		/*
		 * Do not allow to attach to a group in a different
		 * task or CPU context:
7074
		 */
7075 7076 7077 7078 7079 7080 7081 7082
		if (move_group) {
			if (group_leader->ctx->type != ctx->type)
				goto err_context;
		} else {
			if (group_leader->ctx != ctx)
				goto err_context;
		}

7083 7084 7085
		/*
		 * Only a group leader can be exclusive or pinned
		 */
7086
		if (attr.exclusive || attr.pinned)
7087
			goto err_context;
7088 7089 7090 7091 7092
	}

	if (output_event) {
		err = perf_event_set_output(event, output_event);
		if (err)
7093
			goto err_context;
7094
	}
T
Thomas Gleixner 已提交
7095

7096 7097 7098
	event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
	if (IS_ERR(event_file)) {
		err = PTR_ERR(event_file);
7099
		goto err_context;
7100
	}
7101

7102 7103 7104 7105
	if (move_group) {
		struct perf_event_context *gctx = group_leader->ctx;

		mutex_lock(&gctx->mutex);
7106
		perf_remove_from_context(group_leader);
J
Jiri Olsa 已提交
7107 7108 7109 7110 7111 7112 7113

		/*
		 * 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);
7114 7115
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7116
			perf_remove_from_context(sibling);
J
Jiri Olsa 已提交
7117
			perf_event__state_init(sibling);
7118 7119 7120 7121
			put_ctx(gctx);
		}
		mutex_unlock(&gctx->mutex);
		put_ctx(gctx);
7122
	}
7123

7124
	WARN_ON_ONCE(ctx->parent_ctx);
7125
	mutex_lock(&ctx->mutex);
7126 7127

	if (move_group) {
7128
		synchronize_rcu();
7129
		perf_install_in_context(ctx, group_leader, event->cpu);
7130 7131 7132
		get_ctx(ctx);
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7133
			perf_install_in_context(ctx, sibling, event->cpu);
7134 7135 7136 7137
			get_ctx(ctx);
		}
	}

7138
	perf_install_in_context(ctx, event, event->cpu);
7139
	++ctx->generation;
7140
	perf_unpin_context(ctx);
7141
	mutex_unlock(&ctx->mutex);
7142

7143 7144
	put_online_cpus();

7145
	event->owner = current;
P
Peter Zijlstra 已提交
7146

7147 7148 7149
	mutex_lock(&current->perf_event_mutex);
	list_add_tail(&event->owner_entry, &current->perf_event_list);
	mutex_unlock(&current->perf_event_mutex);
7150

7151 7152 7153 7154
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(event);
7155
	perf_event__id_header_size(event);
7156

7157 7158 7159 7160 7161 7162
	/*
	 * 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().
	 */
7163
	fdput(group);
7164 7165
	fd_install(event_fd, event_file);
	return event_fd;
T
Thomas Gleixner 已提交
7166

7167
err_context:
7168
	perf_unpin_context(ctx);
7169
	put_ctx(ctx);
7170
err_alloc:
7171
	free_event(event);
P
Peter Zijlstra 已提交
7172
err_task:
7173
	put_online_cpus();
P
Peter Zijlstra 已提交
7174 7175
	if (task)
		put_task_struct(task);
7176
err_group_fd:
7177
	fdput(group);
7178 7179
err_fd:
	put_unused_fd(event_fd);
7180
	return err;
T
Thomas Gleixner 已提交
7181 7182
}

7183 7184 7185 7186 7187
/**
 * 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
7188
 * @task: task to profile (NULL for percpu)
7189 7190 7191
 */
struct perf_event *
perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
M
Matt Helsley 已提交
7192
				 struct task_struct *task,
7193 7194
				 perf_overflow_handler_t overflow_handler,
				 void *context)
7195 7196
{
	struct perf_event_context *ctx;
7197
	struct perf_event *event;
7198
	int err;
7199

7200 7201 7202
	/*
	 * Get the target context (task or percpu):
	 */
7203

7204 7205
	event = perf_event_alloc(attr, cpu, task, NULL, NULL,
				 overflow_handler, context);
7206 7207 7208 7209
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
		goto err;
	}
7210

7211 7212
	account_event(event);

M
Matt Helsley 已提交
7213
	ctx = find_get_context(event->pmu, task, cpu);
7214 7215
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7216
		goto err_free;
7217
	}
7218 7219 7220 7221 7222

	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
	perf_install_in_context(ctx, event, cpu);
	++ctx->generation;
7223
	perf_unpin_context(ctx);
7224 7225 7226 7227
	mutex_unlock(&ctx->mutex);

	return event;

7228 7229 7230
err_free:
	free_event(event);
err:
7231
	return ERR_PTR(err);
7232
}
7233
EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7234

7235 7236 7237 7238 7239 7240 7241 7242 7243 7244 7245 7246 7247 7248
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);
7249
		unaccount_event_cpu(event, src_cpu);
7250
		put_ctx(src_ctx);
7251
		list_add(&event->migrate_entry, &events);
7252 7253 7254 7255 7256 7257
	}
	mutex_unlock(&src_ctx->mutex);

	synchronize_rcu();

	mutex_lock(&dst_ctx->mutex);
7258 7259
	list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
		list_del(&event->migrate_entry);
7260 7261
		if (event->state >= PERF_EVENT_STATE_OFF)
			event->state = PERF_EVENT_STATE_INACTIVE;
7262
		account_event_cpu(event, dst_cpu);
7263 7264 7265 7266 7267 7268 7269
		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);

7270
static void sync_child_event(struct perf_event *child_event,
7271
			       struct task_struct *child)
7272
{
7273
	struct perf_event *parent_event = child_event->parent;
7274
	u64 child_val;
7275

7276 7277
	if (child_event->attr.inherit_stat)
		perf_event_read_event(child_event, child);
7278

P
Peter Zijlstra 已提交
7279
	child_val = perf_event_count(child_event);
7280 7281 7282 7283

	/*
	 * Add back the child's count to the parent's count:
	 */
7284
	atomic64_add(child_val, &parent_event->child_count);
7285 7286 7287 7288
	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);
7289 7290

	/*
7291
	 * Remove this event from the parent's list
7292
	 */
7293 7294 7295 7296
	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);
7297 7298

	/*
7299
	 * Release the parent event, if this was the last
7300 7301
	 * reference to it.
	 */
7302
	put_event(parent_event);
7303 7304
}

7305
static void
7306 7307
__perf_event_exit_task(struct perf_event *child_event,
			 struct perf_event_context *child_ctx,
7308
			 struct task_struct *child)
7309
{
7310 7311 7312 7313 7314
	if (child_event->parent) {
		raw_spin_lock_irq(&child_ctx->lock);
		perf_group_detach(child_event);
		raw_spin_unlock_irq(&child_ctx->lock);
	}
7315

7316
	perf_remove_from_context(child_event);
7317

7318
	/*
7319
	 * It can happen that the parent exits first, and has events
7320
	 * that are still around due to the child reference. These
7321
	 * events need to be zapped.
7322
	 */
7323
	if (child_event->parent) {
7324 7325
		sync_child_event(child_event, child);
		free_event(child_event);
7326
	}
7327 7328
}

P
Peter Zijlstra 已提交
7329
static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7330
{
7331 7332
	struct perf_event *child_event, *tmp;
	struct perf_event_context *child_ctx;
7333
	unsigned long flags;
7334

P
Peter Zijlstra 已提交
7335
	if (likely(!child->perf_event_ctxp[ctxn])) {
7336
		perf_event_task(child, NULL, 0);
7337
		return;
P
Peter Zijlstra 已提交
7338
	}
7339

7340
	local_irq_save(flags);
7341 7342 7343 7344 7345 7346
	/*
	 * 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.
	 */
7347
	child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7348 7349 7350

	/*
	 * Take the context lock here so that if find_get_context is
7351
	 * reading child->perf_event_ctxp, we wait until it has
7352 7353
	 * incremented the context's refcount before we do put_ctx below.
	 */
7354
	raw_spin_lock(&child_ctx->lock);
7355
	task_ctx_sched_out(child_ctx);
P
Peter Zijlstra 已提交
7356
	child->perf_event_ctxp[ctxn] = NULL;
7357 7358 7359
	/*
	 * If this context is a clone; unclone it so it can't get
	 * swapped to another process while we're removing all
7360
	 * the events from it.
7361 7362
	 */
	unclone_ctx(child_ctx);
7363
	update_context_time(child_ctx);
7364
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7365 7366

	/*
7367 7368 7369
	 * 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
7370
	 */
7371
	perf_event_task(child, child_ctx, 0);
7372

7373 7374 7375
	/*
	 * We can recurse on the same lock type through:
	 *
7376 7377
	 *   __perf_event_exit_task()
	 *     sync_child_event()
7378 7379
	 *       put_event()
	 *         mutex_lock(&ctx->mutex)
7380 7381 7382
	 *
	 * But since its the parent context it won't be the same instance.
	 */
7383
	mutex_lock(&child_ctx->mutex);
7384

7385
again:
7386 7387 7388 7389 7390
	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,
7391
				 group_entry)
7392
		__perf_event_exit_task(child_event, child_ctx, child);
7393 7394

	/*
7395
	 * If the last event was a group event, it will have appended all
7396 7397 7398
	 * its siblings to the list, but we obtained 'tmp' before that which
	 * will still point to the list head terminating the iteration.
	 */
7399 7400
	if (!list_empty(&child_ctx->pinned_groups) ||
	    !list_empty(&child_ctx->flexible_groups))
7401
		goto again;
7402 7403 7404 7405

	mutex_unlock(&child_ctx->mutex);

	put_ctx(child_ctx);
7406 7407
}

P
Peter Zijlstra 已提交
7408 7409 7410 7411 7412
/*
 * 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
7413
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7414 7415
	int ctxn;

P
Peter Zijlstra 已提交
7416 7417 7418 7419 7420 7421 7422 7423 7424 7425 7426 7427 7428 7429 7430
	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
7431 7432 7433 7434
	for_each_task_context_nr(ctxn)
		perf_event_exit_task_context(child, ctxn);
}

7435 7436 7437 7438 7439 7440 7441 7442 7443 7444 7445 7446
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);

7447
	put_event(parent);
7448

7449
	perf_group_detach(event);
7450 7451 7452 7453
	list_del_event(event, ctx);
	free_event(event);
}

7454 7455
/*
 * free an unexposed, unused context as created by inheritance by
P
Peter Zijlstra 已提交
7456
 * perf_event_init_task below, used by fork() in case of fail.
7457
 */
7458
void perf_event_free_task(struct task_struct *task)
7459
{
P
Peter Zijlstra 已提交
7460
	struct perf_event_context *ctx;
7461
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7462
	int ctxn;
7463

P
Peter Zijlstra 已提交
7464 7465 7466 7467
	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;
7468

P
Peter Zijlstra 已提交
7469
		mutex_lock(&ctx->mutex);
7470
again:
P
Peter Zijlstra 已提交
7471 7472 7473
		list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
				group_entry)
			perf_free_event(event, ctx);
7474

P
Peter Zijlstra 已提交
7475 7476 7477
		list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
				group_entry)
			perf_free_event(event, ctx);
7478

P
Peter Zijlstra 已提交
7479 7480 7481
		if (!list_empty(&ctx->pinned_groups) ||
				!list_empty(&ctx->flexible_groups))
			goto again;
7482

P
Peter Zijlstra 已提交
7483
		mutex_unlock(&ctx->mutex);
7484

P
Peter Zijlstra 已提交
7485 7486
		put_ctx(ctx);
	}
7487 7488
}

7489 7490 7491 7492 7493 7494 7495 7496
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
7497 7498 7499 7500 7501 7502 7503 7504 7505 7506 7507 7508
/*
 * 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;
7509
	unsigned long flags;
P
Peter Zijlstra 已提交
7510 7511 7512 7513 7514 7515 7516 7517 7518 7519 7520 7521

	/*
	 * 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,
7522
					   child,
P
Peter Zijlstra 已提交
7523
					   group_leader, parent_event,
7524
				           NULL, NULL);
P
Peter Zijlstra 已提交
7525 7526
	if (IS_ERR(child_event))
		return child_event;
7527 7528 7529 7530 7531 7532

	if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
		free_event(child_event);
		return NULL;
	}

P
Peter Zijlstra 已提交
7533 7534 7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546 7547 7548 7549 7550 7551 7552 7553 7554 7555 7556
	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;
7557 7558
	child_event->overflow_handler_context
		= parent_event->overflow_handler_context;
P
Peter Zijlstra 已提交
7559

7560 7561 7562 7563
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(child_event);
7564
	perf_event__id_header_size(child_event);
7565

P
Peter Zijlstra 已提交
7566 7567 7568
	/*
	 * Link it up in the child's context:
	 */
7569
	raw_spin_lock_irqsave(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7570
	add_event_to_ctx(child_event, child_ctx);
7571
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7572 7573 7574 7575 7576 7577 7578 7579 7580 7581 7582 7583 7584 7585 7586 7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601 7602 7603 7604

	/*
	 * 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;
7605 7606 7607 7608 7609
}

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
7610
		   struct task_struct *child, int ctxn,
7611 7612 7613
		   int *inherited_all)
{
	int ret;
P
Peter Zijlstra 已提交
7614
	struct perf_event_context *child_ctx;
7615 7616 7617 7618

	if (!event->attr.inherit) {
		*inherited_all = 0;
		return 0;
7619 7620
	}

7621
	child_ctx = child->perf_event_ctxp[ctxn];
7622 7623 7624 7625 7626 7627 7628
	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.
		 */
7629

7630
		child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7631 7632
		if (!child_ctx)
			return -ENOMEM;
7633

P
Peter Zijlstra 已提交
7634
		child->perf_event_ctxp[ctxn] = child_ctx;
7635 7636 7637 7638 7639 7640 7641 7642 7643
	}

	ret = inherit_group(event, parent, parent_ctx,
			    child, child_ctx);

	if (ret)
		*inherited_all = 0;

	return ret;
7644 7645
}

7646
/*
7647
 * Initialize the perf_event context in task_struct
7648
 */
P
Peter Zijlstra 已提交
7649
int perf_event_init_context(struct task_struct *child, int ctxn)
7650
{
7651
	struct perf_event_context *child_ctx, *parent_ctx;
7652 7653
	struct perf_event_context *cloned_ctx;
	struct perf_event *event;
7654
	struct task_struct *parent = current;
7655
	int inherited_all = 1;
7656
	unsigned long flags;
7657
	int ret = 0;
7658

P
Peter Zijlstra 已提交
7659
	if (likely(!parent->perf_event_ctxp[ctxn]))
7660 7661
		return 0;

7662
	/*
7663 7664
	 * If the parent's context is a clone, pin it so it won't get
	 * swapped under us.
7665
	 */
P
Peter Zijlstra 已提交
7666
	parent_ctx = perf_pin_task_context(parent, ctxn);
7667

7668 7669 7670 7671 7672 7673 7674
	/*
	 * 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.
	 */

7675 7676 7677 7678
	/*
	 * Lock the parent list. No need to lock the child - not PID
	 * hashed yet and not running, so nobody can access it.
	 */
7679
	mutex_lock(&parent_ctx->mutex);
7680 7681 7682 7683 7684

	/*
	 * We dont have to disable NMIs - we are only looking at
	 * the list, not manipulating it:
	 */
7685
	list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
P
Peter Zijlstra 已提交
7686 7687
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7688 7689 7690
		if (ret)
			break;
	}
7691

7692 7693 7694 7695 7696 7697 7698 7699 7700
	/*
	 * 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);

7701
	list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
P
Peter Zijlstra 已提交
7702 7703
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7704
		if (ret)
7705
			break;
7706 7707
	}

7708 7709 7710
	raw_spin_lock_irqsave(&parent_ctx->lock, flags);
	parent_ctx->rotate_disable = 0;

P
Peter Zijlstra 已提交
7711
	child_ctx = child->perf_event_ctxp[ctxn];
7712

7713
	if (child_ctx && inherited_all) {
7714 7715 7716
		/*
		 * 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
7717 7718 7719
		 *
		 * Note that if the parent is a clone, the holding of
		 * parent_ctx->lock avoids it from being uncloned.
7720
		 */
P
Peter Zijlstra 已提交
7721
		cloned_ctx = parent_ctx->parent_ctx;
7722 7723
		if (cloned_ctx) {
			child_ctx->parent_ctx = cloned_ctx;
7724
			child_ctx->parent_gen = parent_ctx->parent_gen;
7725 7726 7727 7728 7729
		} else {
			child_ctx->parent_ctx = parent_ctx;
			child_ctx->parent_gen = parent_ctx->generation;
		}
		get_ctx(child_ctx->parent_ctx);
7730 7731
	}

P
Peter Zijlstra 已提交
7732
	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7733
	mutex_unlock(&parent_ctx->mutex);
7734

7735
	perf_unpin_context(parent_ctx);
7736
	put_ctx(parent_ctx);
7737

7738
	return ret;
7739 7740
}

P
Peter Zijlstra 已提交
7741 7742 7743 7744 7745 7746 7747
/*
 * Initialize the perf_event context in task_struct
 */
int perf_event_init_task(struct task_struct *child)
{
	int ctxn, ret;

7748 7749 7750 7751
	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
7752 7753 7754 7755 7756 7757 7758 7759 7760
	for_each_task_context_nr(ctxn) {
		ret = perf_event_init_context(child, ctxn);
		if (ret)
			return ret;
	}

	return 0;
}

7761 7762
static void __init perf_event_init_all_cpus(void)
{
7763
	struct swevent_htable *swhash;
7764 7765 7766
	int cpu;

	for_each_possible_cpu(cpu) {
7767 7768
		swhash = &per_cpu(swevent_htable, cpu);
		mutex_init(&swhash->hlist_mutex);
7769
		INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7770 7771 7772
	}
}

7773
static void perf_event_init_cpu(int cpu)
T
Thomas Gleixner 已提交
7774
{
P
Peter Zijlstra 已提交
7775
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
T
Thomas Gleixner 已提交
7776

7777
	mutex_lock(&swhash->hlist_mutex);
7778
	if (swhash->hlist_refcount > 0) {
7779 7780
		struct swevent_hlist *hlist;

7781 7782 7783
		hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
		WARN_ON(!hlist);
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
7784
	}
7785
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
7786 7787
}

P
Peter Zijlstra 已提交
7788
#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7789
static void perf_pmu_rotate_stop(struct pmu *pmu)
T
Thomas Gleixner 已提交
7790
{
7791 7792 7793 7794 7795 7796 7797
	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
7798
static void __perf_event_exit_context(void *__info)
T
Thomas Gleixner 已提交
7799
{
P
Peter Zijlstra 已提交
7800
	struct perf_event_context *ctx = __info;
7801
	struct perf_event *event, *tmp;
T
Thomas Gleixner 已提交
7802

P
Peter Zijlstra 已提交
7803
	perf_pmu_rotate_stop(ctx->pmu);
7804

7805
	list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7806
		__perf_remove_from_context(event);
7807
	list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7808
		__perf_remove_from_context(event);
T
Thomas Gleixner 已提交
7809
}
P
Peter Zijlstra 已提交
7810 7811 7812 7813 7814 7815 7816 7817 7818

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) {
7819
		ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
P
Peter Zijlstra 已提交
7820 7821 7822 7823 7824 7825 7826 7827

		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);
}

7828
static void perf_event_exit_cpu(int cpu)
T
Thomas Gleixner 已提交
7829
{
7830
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7831

7832 7833 7834
	mutex_lock(&swhash->hlist_mutex);
	swevent_hlist_release(swhash);
	mutex_unlock(&swhash->hlist_mutex);
7835

P
Peter Zijlstra 已提交
7836
	perf_event_exit_cpu_context(cpu);
T
Thomas Gleixner 已提交
7837 7838
}
#else
7839
static inline void perf_event_exit_cpu(int cpu) { }
T
Thomas Gleixner 已提交
7840 7841
#endif

P
Peter Zijlstra 已提交
7842 7843 7844 7845 7846 7847 7848 7849 7850 7851 7852 7853 7854 7855 7856 7857 7858 7859 7860 7861
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,
};

7862
static int
T
Thomas Gleixner 已提交
7863 7864 7865 7866
perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
{
	unsigned int cpu = (long)hcpu;

7867
	switch (action & ~CPU_TASKS_FROZEN) {
T
Thomas Gleixner 已提交
7868 7869

	case CPU_UP_PREPARE:
P
Peter Zijlstra 已提交
7870
	case CPU_DOWN_FAILED:
7871
		perf_event_init_cpu(cpu);
T
Thomas Gleixner 已提交
7872 7873
		break;

P
Peter Zijlstra 已提交
7874
	case CPU_UP_CANCELED:
T
Thomas Gleixner 已提交
7875
	case CPU_DOWN_PREPARE:
7876
		perf_event_exit_cpu(cpu);
T
Thomas Gleixner 已提交
7877 7878 7879 7880 7881 7882 7883 7884
		break;
	default:
		break;
	}

	return NOTIFY_OK;
}

7885
void __init perf_event_init(void)
T
Thomas Gleixner 已提交
7886
{
7887 7888
	int ret;

P
Peter Zijlstra 已提交
7889 7890
	idr_init(&pmu_idr);

7891
	perf_event_init_all_cpus();
7892
	init_srcu_struct(&pmus_srcu);
P
Peter Zijlstra 已提交
7893 7894 7895
	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);
7896 7897
	perf_tp_register();
	perf_cpu_notifier(perf_cpu_notify);
P
Peter Zijlstra 已提交
7898
	register_reboot_notifier(&perf_reboot_notifier);
7899 7900 7901

	ret = init_hw_breakpoint();
	WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7902 7903 7904

	/* do not patch jump label more than once per second */
	jump_label_rate_limit(&perf_sched_events, HZ);
7905 7906 7907 7908 7909 7910 7911

	/*
	 * 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
7912
}
P
Peter Zijlstra 已提交
7913 7914 7915 7916 7917 7918 7919 7920 7921 7922 7923 7924 7925 7926 7927 7928 7929 7930 7931 7932 7933 7934 7935 7936 7937 7938 7939 7940

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
7941 7942

#ifdef CONFIG_CGROUP_PERF
7943 7944
static struct cgroup_subsys_state *
perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
S
Stephane Eranian 已提交
7945 7946 7947
{
	struct perf_cgroup *jc;

7948
	jc = kzalloc(sizeof(*jc), GFP_KERNEL);
S
Stephane Eranian 已提交
7949 7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960
	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;
}

7961
static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
S
Stephane Eranian 已提交
7962
{
7963 7964
	struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);

S
Stephane Eranian 已提交
7965 7966 7967 7968 7969 7970 7971 7972 7973 7974 7975
	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;
}

7976 7977
static void perf_cgroup_attach(struct cgroup_subsys_state *css,
			       struct cgroup_taskset *tset)
S
Stephane Eranian 已提交
7978
{
7979 7980
	struct task_struct *task;

7981
	cgroup_taskset_for_each(task, css, tset)
7982
		task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
7983 7984
}

7985 7986
static void perf_cgroup_exit(struct cgroup_subsys_state *css,
			     struct cgroup_subsys_state *old_css,
7987
			     struct task_struct *task)
S
Stephane Eranian 已提交
7988 7989 7990 7991 7992 7993 7994 7995 7996
{
	/*
	 * 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;

7997
	task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
7998 7999 8000
}

struct cgroup_subsys perf_subsys = {
8001 8002
	.name		= "perf_event",
	.subsys_id	= perf_subsys_id,
8003 8004
	.css_alloc	= perf_cgroup_css_alloc,
	.css_free	= perf_cgroup_css_free,
8005
	.exit		= perf_cgroup_exit,
8006
	.attach		= perf_cgroup_attach,
S
Stephane Eranian 已提交
8007 8008
};
#endif /* CONFIG_CGROUP_PERF */