core.c 190.0 KB
Newer Older
T
Thomas Gleixner 已提交
1
/*
I
Ingo Molnar 已提交
2
 * Performance events core code:
T
Thomas Gleixner 已提交
3
 *
4
 *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 6
 *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
 *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
A
Al Viro 已提交
7
 *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8
 *
I
Ingo Molnar 已提交
9
 * For licensing details see kernel-base/COPYING
T
Thomas Gleixner 已提交
10 11 12
 */

#include <linux/fs.h>
13
#include <linux/mm.h>
T
Thomas Gleixner 已提交
14 15
#include <linux/cpu.h>
#include <linux/smp.h>
P
Peter Zijlstra 已提交
16
#include <linux/idr.h>
17
#include <linux/file.h>
T
Thomas Gleixner 已提交
18
#include <linux/poll.h>
19
#include <linux/slab.h>
20
#include <linux/hash.h>
21
#include <linux/tick.h>
T
Thomas Gleixner 已提交
22
#include <linux/sysfs.h>
23
#include <linux/dcache.h>
T
Thomas Gleixner 已提交
24
#include <linux/percpu.h>
25
#include <linux/ptrace.h>
P
Peter Zijlstra 已提交
26
#include <linux/reboot.h>
27
#include <linux/vmstat.h>
P
Peter Zijlstra 已提交
28
#include <linux/device.h>
29
#include <linux/export.h>
30
#include <linux/vmalloc.h>
31 32
#include <linux/hardirq.h>
#include <linux/rculist.h>
T
Thomas Gleixner 已提交
33 34 35
#include <linux/uaccess.h>
#include <linux/syscalls.h>
#include <linux/anon_inodes.h>
I
Ingo Molnar 已提交
36
#include <linux/kernel_stat.h>
37
#include <linux/perf_event.h>
L
Li Zefan 已提交
38
#include <linux/ftrace_event.h>
39
#include <linux/hw_breakpoint.h>
40
#include <linux/mm_types.h>
41
#include <linux/cgroup.h>
42
#include <linux/module.h>
43
#include <linux/mman.h>
P
Pawel Moll 已提交
44
#include <linux/compat.h>
T
Thomas Gleixner 已提交
45

46 47
#include "internal.h"

48 49
#include <asm/irq_regs.h>

50
struct remote_function_call {
51 52 53 54
	struct task_struct	*p;
	int			(*func)(void *info);
	void			*info;
	int			ret;
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
};

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 = {
88 89 90 91
		.p	= p,
		.func	= func,
		.info	= info,
		.ret	= -ESRCH, /* No such (running) process */
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
	};

	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 = {
112 113 114 115
		.p	= NULL,
		.func	= func,
		.info	= info,
		.ret	= -ENXIO, /* No such CPU */
116 117 118 119 120 121 122
	};

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

	return data.ret;
}

S
Stephane Eranian 已提交
123 124
#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
		       PERF_FLAG_FD_OUTPUT  |\
125 126
		       PERF_FLAG_PID_CGROUP |\
		       PERF_FLAG_FD_CLOEXEC)
S
Stephane Eranian 已提交
127

128 129 130 131 132 133 134
/*
 * branch priv levels that need permission checks
 */
#define PERF_SAMPLE_BRANCH_PERM_PLM \
	(PERF_SAMPLE_BRANCH_KERNEL |\
	 PERF_SAMPLE_BRANCH_HV)

135 136 137 138 139 140
enum event_type_t {
	EVENT_FLEXIBLE = 0x1,
	EVENT_PINNED = 0x2,
	EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
};

S
Stephane Eranian 已提交
141 142 143 144
/*
 * perf_sched_events : >0 events exist
 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
 */
145
struct static_key_deferred perf_sched_events __read_mostly;
S
Stephane Eranian 已提交
146
static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
147
static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
S
Stephane Eranian 已提交
148

149 150 151
static atomic_t nr_mmap_events __read_mostly;
static atomic_t nr_comm_events __read_mostly;
static atomic_t nr_task_events __read_mostly;
152
static atomic_t nr_freq_events __read_mostly;
153

P
Peter Zijlstra 已提交
154 155 156 157
static LIST_HEAD(pmus);
static DEFINE_MUTEX(pmus_lock);
static struct srcu_struct pmus_srcu;

158
/*
159
 * perf event paranoia level:
160 161
 *  -1 - not paranoid at all
 *   0 - disallow raw tracepoint access for unpriv
162
 *   1 - disallow cpu events for unpriv
163
 *   2 - disallow kernel profiling for unpriv
164
 */
165
int sysctl_perf_event_paranoid __read_mostly = 1;
166

167 168
/* Minimum for 512 kiB + 1 user control page */
int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
169 170

/*
171
 * max perf event sample rate
172
 */
173 174 175 176 177 178 179 180 181
#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
182 183
static int perf_sample_allowed_ns __read_mostly =
	DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
184 185 186 187 188 189

void update_perf_cpu_limits(void)
{
	u64 tmp = perf_sample_period_ns;

	tmp *= sysctl_perf_cpu_time_max_percent;
190
	do_div(tmp, 100);
P
Peter Zijlstra 已提交
191
	ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
192
}
P
Peter Zijlstra 已提交
193

194 195
static int perf_rotate_context(struct perf_cpu_context *cpuctx);

P
Peter Zijlstra 已提交
196 197 198 199
int perf_proc_update_handler(struct ctl_table *table, int write,
		void __user *buffer, size_t *lenp,
		loff_t *ppos)
{
200
	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
P
Peter Zijlstra 已提交
201 202 203 204 205

	if (ret || !write)
		return ret;

	max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
	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
224 225 226

	return 0;
}
227

228 229 230 231 232 233 234
/*
 * 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
235
static DEFINE_PER_CPU(u64, running_sample_length);
236

237
static void perf_duration_warn(struct irq_work *w)
238
{
239
	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
240
	u64 avg_local_sample_len;
241
	u64 local_samples_len;
242 243 244 245 246 247 248

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

	printk_ratelimited(KERN_WARNING
			"perf interrupt took too long (%lld > %lld), lowering "
			"kernel.perf_event_max_sample_rate to %d\n",
249
			avg_local_sample_len, allowed_ns >> 1,
250 251 252 253 254 255 256
			sysctl_perf_event_sample_rate);
}

static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);

void perf_sample_event_took(u64 sample_len_ns)
{
P
Peter Zijlstra 已提交
257
	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
258 259
	u64 avg_local_sample_len;
	u64 local_samples_len;
260

P
Peter Zijlstra 已提交
261
	if (allowed_ns == 0)
262 263 264 265 266 267 268 269 270 271 272 273 274 275 276
		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
277
	if (avg_local_sample_len <= allowed_ns)
278 279 280 281 282 283 284 285 286 287
		return;

	if (max_samples_per_tick <= 1)
		return;

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

	update_perf_cpu_limits();
288

289 290 291 292 293 294
	if (!irq_work_queue(&perf_duration_work)) {
		early_printk("perf interrupt took too long (%lld > %lld), lowering "
			     "kernel.perf_event_max_sample_rate to %d\n",
			     avg_local_sample_len, allowed_ns >> 1,
			     sysctl_perf_event_sample_rate);
	}
295 296
}

297
static atomic64_t perf_event_id;
298

299 300 301 302
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
303 304 305 306 307
			     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);
308

309
void __weak perf_event_print_debug(void)	{ }
T
Thomas Gleixner 已提交
310

311
extern __weak const char *perf_pmu_name(void)
T
Thomas Gleixner 已提交
312
{
313
	return "pmu";
T
Thomas Gleixner 已提交
314 315
}

316 317 318 319 320
static inline u64 perf_clock(void)
{
	return local_clock();
}

S
Stephane Eranian 已提交
321 322 323 324 325 326
static inline struct perf_cpu_context *
__get_cpu_context(struct perf_event_context *ctx)
{
	return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
}

327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
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
343 344
#ifdef CONFIG_CGROUP_PERF

345 346 347 348 349 350 351 352 353 354 355
/*
 * 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;
356
	struct perf_cgroup_info	__percpu *info;
357 358
};

359 360 361 362 363
/*
 * 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
364 365 366
static inline struct perf_cgroup *
perf_cgroup_from_task(struct task_struct *task)
{
367
	return container_of(task_css(task, perf_event_cgrp_id),
368
			    struct perf_cgroup, css);
S
Stephane Eranian 已提交
369 370 371 372 373 374 375 376
}

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);

377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392
	/* @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
393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
}

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)
{
441 442
	struct perf_cgroup *cgrp;

S
Stephane Eranian 已提交
443
	/*
444 445
	 * ensure we access cgroup data only when needed and
	 * when we know the cgroup is pinned (css_get)
S
Stephane Eranian 已提交
446
	 */
447
	if (!is_cgroup_event(event))
S
Stephane Eranian 已提交
448 449
		return;

450 451 452 453 454 455
	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
456 457 458
}

static inline void
459 460
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
461 462 463 464
{
	struct perf_cgroup *cgrp;
	struct perf_cgroup_info *info;

465 466 467 468 469 470
	/*
	 * 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
471 472 473 474
		return;

	cgrp = perf_cgroup_from_task(task);
	info = this_cpu_ptr(cgrp->info);
475
	info->timestamp = ctx->timestamp;
S
Stephane Eranian 已提交
476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507
}

#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);
508 509
		if (cpuctx->unique_pmu != pmu)
			continue; /* ensure we process each cpuctx once */
S
Stephane Eranian 已提交
510 511 512 513 514 515 516 517 518

		/*
		 * 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) {
519 520
			perf_ctx_lock(cpuctx, cpuctx->task_ctx);
			perf_pmu_disable(cpuctx->ctx.pmu);
S
Stephane Eranian 已提交
521 522 523 524 525 526 527 528 529 530 531

			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) {
532
				WARN_ON_ONCE(cpuctx->cgrp);
533 534 535 536
				/*
				 * set cgrp before ctxsw in to allow
				 * event_filter_match() to not have to pass
				 * task around
S
Stephane Eranian 已提交
537 538 539 540
				 */
				cpuctx->cgrp = perf_cgroup_from_task(task);
				cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
			}
541 542
			perf_pmu_enable(cpuctx->ctx.pmu);
			perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
S
Stephane Eranian 已提交
543 544 545 546 547 548 549 550
		}
	}

	rcu_read_unlock();

	local_irq_restore(flags);
}

551 552
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
553
{
554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575
	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
576 577
}

578 579
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
580
{
581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598
	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
599 600 601 602 603 604 605 606
}

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;
607 608
	struct fd f = fdget(fd);
	int ret = 0;
S
Stephane Eranian 已提交
609

610
	if (!f.file)
S
Stephane Eranian 已提交
611 612
		return -EBADF;

613 614
	css = css_tryget_online_from_dir(f.file->f_dentry,
					 &perf_event_cgrp_subsys);
615 616 617 618
	if (IS_ERR(css)) {
		ret = PTR_ERR(css);
		goto out;
	}
S
Stephane Eranian 已提交
619 620 621 622 623 624 625 626 627 628 629 630 631

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

	/*
	 * all events in a group must monitor
	 * the same cgroup because a task belongs
	 * to only one perf cgroup at a time
	 */
	if (group_leader && group_leader->cgrp != cgrp) {
		perf_detach_cgroup(event);
		ret = -EINVAL;
	}
632
out:
633
	fdput(f);
S
Stephane Eranian 已提交
634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706
	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)
{
}

707 708
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
709 710 711
{
}

712 713
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
714 715 716 717 718 719 720 721 722 723 724
{
}

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
725 726
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756
{
}

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

757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819
/*
 * 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;
820
	int timer;
821 822 823 824 825

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

826 827 828 829 830 831 832 833 834
	/*
	 * 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);
835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856

	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
857
void perf_pmu_disable(struct pmu *pmu)
858
{
P
Peter Zijlstra 已提交
859 860 861
	int *count = this_cpu_ptr(pmu->pmu_disable_count);
	if (!(*count)++)
		pmu->pmu_disable(pmu);
862 863
}

P
Peter Zijlstra 已提交
864
void perf_pmu_enable(struct pmu *pmu)
865
{
P
Peter Zijlstra 已提交
866 867 868
	int *count = this_cpu_ptr(pmu->pmu_disable_count);
	if (!--(*count))
		pmu->pmu_enable(pmu);
869 870
}

871 872 873 874 875 876 877
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
878
static void perf_pmu_rotate_start(struct pmu *pmu)
879
{
P
Peter Zijlstra 已提交
880
	struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
881
	struct list_head *head = &__get_cpu_var(rotation_list);
882

883
	WARN_ON(!irqs_disabled());
884

885
	if (list_empty(&cpuctx->rotation_list))
886
		list_add(&cpuctx->rotation_list, head);
887 888
}

889
static void get_ctx(struct perf_event_context *ctx)
890
{
891
	WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
892 893
}

894
static void put_ctx(struct perf_event_context *ctx)
895
{
896 897 898
	if (atomic_dec_and_test(&ctx->refcount)) {
		if (ctx->parent_ctx)
			put_ctx(ctx->parent_ctx);
899 900
		if (ctx->task)
			put_task_struct(ctx->task);
901
		kfree_rcu(ctx, rcu_head);
902
	}
903 904
}

905
static void unclone_ctx(struct perf_event_context *ctx)
906 907 908 909 910
{
	if (ctx->parent_ctx) {
		put_ctx(ctx->parent_ctx);
		ctx->parent_ctx = NULL;
	}
911
	ctx->generation++;
912 913
}

914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935
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);
}

936
/*
937
 * If we inherit events we want to return the parent event id
938 939
 * to userspace.
 */
940
static u64 primary_event_id(struct perf_event *event)
941
{
942
	u64 id = event->id;
943

944 945
	if (event->parent)
		id = event->parent->id;
946 947 948 949

	return id;
}

950
/*
951
 * Get the perf_event_context for a task and lock it.
952 953 954
 * This has to cope with with the fact that until it is locked,
 * the context could get moved to another task.
 */
955
static struct perf_event_context *
P
Peter Zijlstra 已提交
956
perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
957
{
958
	struct perf_event_context *ctx;
959

P
Peter Zijlstra 已提交
960
retry:
961 962 963 964 965 966 967 968 969 970 971
	/*
	 * 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
972
	ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
973 974 975 976
	if (ctx) {
		/*
		 * If this context is a clone of another, it might
		 * get swapped for another underneath us by
977
		 * perf_event_task_sched_out, though the
978 979 980 981 982 983
		 * 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.
		 */
984
		raw_spin_lock_irqsave(&ctx->lock, *flags);
P
Peter Zijlstra 已提交
985
		if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
986
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
987 988
			rcu_read_unlock();
			preempt_enable();
989 990
			goto retry;
		}
991 992

		if (!atomic_inc_not_zero(&ctx->refcount)) {
993
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
994 995
			ctx = NULL;
		}
996 997
	}
	rcu_read_unlock();
998
	preempt_enable();
999 1000 1001 1002 1003 1004 1005 1006
	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
1007 1008
static struct perf_event_context *
perf_pin_task_context(struct task_struct *task, int ctxn)
1009
{
1010
	struct perf_event_context *ctx;
1011 1012
	unsigned long flags;

P
Peter Zijlstra 已提交
1013
	ctx = perf_lock_task_context(task, ctxn, &flags);
1014 1015
	if (ctx) {
		++ctx->pin_count;
1016
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
1017 1018 1019 1020
	}
	return ctx;
}

1021
static void perf_unpin_context(struct perf_event_context *ctx)
1022 1023 1024
{
	unsigned long flags;

1025
	raw_spin_lock_irqsave(&ctx->lock, flags);
1026
	--ctx->pin_count;
1027
	raw_spin_unlock_irqrestore(&ctx->lock, flags);
1028 1029
}

1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
/*
 * 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;
}

1041 1042 1043
static u64 perf_event_time(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;
S
Stephane Eranian 已提交
1044 1045 1046 1047

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

1048 1049 1050
	return ctx ? ctx->time : 0;
}

1051 1052
/*
 * Update the total_time_enabled and total_time_running fields for a event.
1053
 * The caller of this function needs to hold the ctx->lock.
1054 1055 1056 1057 1058 1059 1060 1061 1062
 */
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
1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073
	/*
	 * 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))
1074
		run_end = perf_cgroup_event_time(event);
S
Stephane Eranian 已提交
1075 1076
	else if (ctx->is_active)
		run_end = ctx->time;
1077 1078 1079 1080
	else
		run_end = event->tstamp_stopped;

	event->total_time_enabled = run_end - event->tstamp_enabled;
1081 1082 1083 1084

	if (event->state == PERF_EVENT_STATE_INACTIVE)
		run_end = event->tstamp_stopped;
	else
1085
		run_end = perf_event_time(event);
1086 1087

	event->total_time_running = run_end - event->tstamp_running;
S
Stephane Eranian 已提交
1088

1089 1090
}

1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
/*
 * 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);
}

1103 1104 1105 1106 1107 1108 1109 1110 1111
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;
}

1112
/*
1113
 * Add a event from the lists for its context.
1114 1115
 * Must be called with ctx->mutex and ctx->lock held.
 */
1116
static void
1117
list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1118
{
1119 1120
	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
	event->attach_state |= PERF_ATTACH_CONTEXT;
1121 1122

	/*
1123 1124 1125
	 * 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.
1126
	 */
1127
	if (event->group_leader == event) {
1128 1129
		struct list_head *list;

1130 1131 1132
		if (is_software_event(event))
			event->group_flags |= PERF_GROUP_SOFTWARE;

1133 1134
		list = ctx_group_list(event, ctx);
		list_add_tail(&event->group_entry, list);
P
Peter Zijlstra 已提交
1135
	}
P
Peter Zijlstra 已提交
1136

1137
	if (is_cgroup_event(event))
S
Stephane Eranian 已提交
1138 1139
		ctx->nr_cgroups++;

1140 1141 1142
	if (has_branch_stack(event))
		ctx->nr_branch_stack++;

1143
	list_add_rcu(&event->event_entry, &ctx->event_list);
1144
	if (!ctx->nr_events)
P
Peter Zijlstra 已提交
1145
		perf_pmu_rotate_start(ctx->pmu);
1146 1147
	ctx->nr_events++;
	if (event->attr.inherit_stat)
1148
		ctx->nr_stat++;
1149 1150

	ctx->generation++;
1151 1152
}

J
Jiri Olsa 已提交
1153 1154 1155 1156 1157 1158 1159 1160 1161
/*
 * 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;
}

1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200
/*
 * 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);

1201 1202 1203 1204 1205 1206
	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
1207 1208 1209
	if (sample_type & PERF_SAMPLE_WEIGHT)
		size += sizeof(data->weight);

1210 1211 1212
	if (sample_type & PERF_SAMPLE_READ)
		size += event->read_size;

1213 1214 1215
	if (sample_type & PERF_SAMPLE_DATA_SRC)
		size += sizeof(data->data_src.val);

A
Andi Kleen 已提交
1216 1217 1218
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		size += sizeof(data->txn);

1219 1220 1221 1222 1223 1224 1225 1226 1227
	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;

1228 1229 1230 1231 1232 1233
	if (sample_type & PERF_SAMPLE_TID)
		size += sizeof(data->tid_entry);

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

1234 1235 1236
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		size += sizeof(data->id);

1237 1238 1239 1240 1241 1242 1243 1244 1245
	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);

1246
	event->id_header_size = size;
1247 1248
}

1249 1250
static void perf_group_attach(struct perf_event *event)
{
1251
	struct perf_event *group_leader = event->group_leader, *pos;
1252

P
Peter Zijlstra 已提交
1253 1254 1255 1256 1257 1258
	/*
	 * We can have double attach due to group movement in perf_event_open.
	 */
	if (event->attach_state & PERF_ATTACH_GROUP)
		return;

1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269
	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++;
1270 1271 1272 1273 1274

	perf_event__header_size(group_leader);

	list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
		perf_event__header_size(pos);
1275 1276
}

1277
/*
1278
 * Remove a event from the lists for its context.
1279
 * Must be called with ctx->mutex and ctx->lock held.
1280
 */
1281
static void
1282
list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1283
{
1284
	struct perf_cpu_context *cpuctx;
1285 1286 1287 1288
	/*
	 * We can have double detach due to exit/hot-unplug + close.
	 */
	if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1289
		return;
1290 1291 1292

	event->attach_state &= ~PERF_ATTACH_CONTEXT;

1293
	if (is_cgroup_event(event)) {
S
Stephane Eranian 已提交
1294
		ctx->nr_cgroups--;
1295 1296 1297 1298 1299 1300 1301 1302 1303
		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
1304

1305 1306 1307
	if (has_branch_stack(event))
		ctx->nr_branch_stack--;

1308 1309
	ctx->nr_events--;
	if (event->attr.inherit_stat)
1310
		ctx->nr_stat--;
1311

1312
	list_del_rcu(&event->event_entry);
1313

1314 1315
	if (event->group_leader == event)
		list_del_init(&event->group_entry);
P
Peter Zijlstra 已提交
1316

1317
	update_group_times(event);
1318 1319 1320 1321 1322 1323 1324 1325 1326 1327

	/*
	 * 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;
1328 1329

	ctx->generation++;
1330 1331
}

1332
static void perf_group_detach(struct perf_event *event)
1333 1334
{
	struct perf_event *sibling, *tmp;
1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350
	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--;
1351
		goto out;
1352 1353 1354 1355
	}

	if (!list_empty(&event->group_entry))
		list = &event->group_entry;
1356

1357
	/*
1358 1359
	 * If this was a group event with sibling events then
	 * upgrade the siblings to singleton events by adding them
1360
	 * to whatever list we are on.
1361
	 */
1362
	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1363 1364
		if (list)
			list_move_tail(&sibling->group_entry, list);
1365
		sibling->group_leader = sibling;
1366 1367 1368

		/* Inherit group flags from the previous leader */
		sibling->group_flags = event->group_flags;
1369
	}
1370 1371 1372 1373 1374 1375

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);
1376 1377
}

1378 1379 1380
static inline int
event_filter_match(struct perf_event *event)
{
S
Stephane Eranian 已提交
1381 1382
	return (event->cpu == -1 || event->cpu == smp_processor_id())
	    && perf_cgroup_match(event);
1383 1384
}

1385 1386
static void
event_sched_out(struct perf_event *event,
1387
		  struct perf_cpu_context *cpuctx,
1388
		  struct perf_event_context *ctx)
1389
{
1390
	u64 tstamp = perf_event_time(event);
1391 1392 1393 1394 1395 1396 1397 1398 1399
	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
1400
		delta = tstamp - event->tstamp_stopped;
1401
		event->tstamp_running += delta;
1402
		event->tstamp_stopped = tstamp;
1403 1404
	}

1405
	if (event->state != PERF_EVENT_STATE_ACTIVE)
1406
		return;
1407

1408 1409
	perf_pmu_disable(event->pmu);

1410 1411 1412 1413
	event->state = PERF_EVENT_STATE_INACTIVE;
	if (event->pending_disable) {
		event->pending_disable = 0;
		event->state = PERF_EVENT_STATE_OFF;
1414
	}
1415
	event->tstamp_stopped = tstamp;
P
Peter Zijlstra 已提交
1416
	event->pmu->del(event, 0);
1417
	event->oncpu = -1;
1418

1419
	if (!is_software_event(event))
1420 1421
		cpuctx->active_oncpu--;
	ctx->nr_active--;
1422 1423
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq--;
1424
	if (event->attr.exclusive || !cpuctx->active_oncpu)
1425
		cpuctx->exclusive = 0;
1426 1427

	perf_pmu_enable(event->pmu);
1428 1429
}

1430
static void
1431
group_sched_out(struct perf_event *group_event,
1432
		struct perf_cpu_context *cpuctx,
1433
		struct perf_event_context *ctx)
1434
{
1435
	struct perf_event *event;
1436
	int state = group_event->state;
1437

1438
	event_sched_out(group_event, cpuctx, ctx);
1439 1440 1441 1442

	/*
	 * Schedule out siblings (if any):
	 */
1443 1444
	list_for_each_entry(event, &group_event->sibling_list, group_entry)
		event_sched_out(event, cpuctx, ctx);
1445

1446
	if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1447 1448 1449
		cpuctx->exclusive = 0;
}

1450 1451 1452 1453 1454
struct remove_event {
	struct perf_event *event;
	bool detach_group;
};

T
Thomas Gleixner 已提交
1455
/*
1456
 * Cross CPU call to remove a performance event
T
Thomas Gleixner 已提交
1457
 *
1458
 * We disable the event on the hardware level first. After that we
T
Thomas Gleixner 已提交
1459 1460
 * remove it from the context list.
 */
1461
static int __perf_remove_from_context(void *info)
T
Thomas Gleixner 已提交
1462
{
1463 1464
	struct remove_event *re = info;
	struct perf_event *event = re->event;
1465
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1466
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
T
Thomas Gleixner 已提交
1467

1468
	raw_spin_lock(&ctx->lock);
1469
	event_sched_out(event, cpuctx, ctx);
1470 1471
	if (re->detach_group)
		perf_group_detach(event);
1472
	list_del_event(event, ctx);
1473 1474 1475 1476
	if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
		ctx->is_active = 0;
		cpuctx->task_ctx = NULL;
	}
1477
	raw_spin_unlock(&ctx->lock);
1478 1479

	return 0;
T
Thomas Gleixner 已提交
1480 1481 1482 1483
}


/*
1484
 * Remove the event from a task's (or a CPU's) list of events.
T
Thomas Gleixner 已提交
1485
 *
1486
 * CPU events are removed with a smp call. For task events we only
T
Thomas Gleixner 已提交
1487
 * call when the task is on a CPU.
1488
 *
1489 1490
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1491 1492
 * 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.
1493
 * When called from perf_event_exit_task, it's OK because the
1494
 * context has been detached from its task.
T
Thomas Gleixner 已提交
1495
 */
1496
static void perf_remove_from_context(struct perf_event *event, bool detach_group)
T
Thomas Gleixner 已提交
1497
{
1498
	struct perf_event_context *ctx = event->ctx;
T
Thomas Gleixner 已提交
1499
	struct task_struct *task = ctx->task;
1500 1501 1502 1503
	struct remove_event re = {
		.event = event,
		.detach_group = detach_group,
	};
T
Thomas Gleixner 已提交
1504

1505 1506
	lockdep_assert_held(&ctx->mutex);

T
Thomas Gleixner 已提交
1507 1508
	if (!task) {
		/*
1509
		 * Per cpu events are removed via an smp call and
1510
		 * the removal is always successful.
T
Thomas Gleixner 已提交
1511
		 */
1512
		cpu_function_call(event->cpu, __perf_remove_from_context, &re);
T
Thomas Gleixner 已提交
1513 1514 1515 1516
		return;
	}

retry:
1517
	if (!task_function_call(task, __perf_remove_from_context, &re))
1518
		return;
T
Thomas Gleixner 已提交
1519

1520
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1521
	/*
1522 1523
	 * 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
1524
	 */
1525
	if (ctx->is_active) {
1526
		raw_spin_unlock_irq(&ctx->lock);
1527 1528 1529 1530 1531
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
T
Thomas Gleixner 已提交
1532 1533 1534 1535
		goto retry;
	}

	/*
1536 1537
	 * 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
1538
	 */
1539 1540
	if (detach_group)
		perf_group_detach(event);
1541
	list_del_event(event, ctx);
1542
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1543 1544
}

1545
/*
1546
 * Cross CPU call to disable a performance event
1547
 */
1548
int __perf_event_disable(void *info)
1549
{
1550 1551
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1552
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1553 1554

	/*
1555 1556
	 * If this is a per-task event, need to check whether this
	 * event's task is the current task on this cpu.
1557 1558 1559
	 *
	 * Can trigger due to concurrent perf_event_context_sched_out()
	 * flipping contexts around.
1560
	 */
1561
	if (ctx->task && cpuctx->task_ctx != ctx)
1562
		return -EINVAL;
1563

1564
	raw_spin_lock(&ctx->lock);
1565 1566

	/*
1567
	 * If the event is on, turn it off.
1568 1569
	 * If it is in error state, leave it in error state.
	 */
1570
	if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1571
		update_context_time(ctx);
S
Stephane Eranian 已提交
1572
		update_cgrp_time_from_event(event);
1573 1574 1575
		update_group_times(event);
		if (event == event->group_leader)
			group_sched_out(event, cpuctx, ctx);
1576
		else
1577 1578
			event_sched_out(event, cpuctx, ctx);
		event->state = PERF_EVENT_STATE_OFF;
1579 1580
	}

1581
	raw_spin_unlock(&ctx->lock);
1582 1583

	return 0;
1584 1585 1586
}

/*
1587
 * Disable a event.
1588
 *
1589 1590
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1591
 * remains valid.  This condition is satisifed when called through
1592 1593 1594 1595
 * 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
1596
 * is the current context on this CPU and preemption is disabled,
1597
 * hence we can't get into perf_event_task_sched_out for this context.
1598
 */
1599
void perf_event_disable(struct perf_event *event)
1600
{
1601
	struct perf_event_context *ctx = event->ctx;
1602 1603 1604 1605
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
1606
		 * Disable the event on the cpu that it's on
1607
		 */
1608
		cpu_function_call(event->cpu, __perf_event_disable, event);
1609 1610 1611
		return;
	}

P
Peter Zijlstra 已提交
1612
retry:
1613 1614
	if (!task_function_call(task, __perf_event_disable, event))
		return;
1615

1616
	raw_spin_lock_irq(&ctx->lock);
1617
	/*
1618
	 * If the event is still active, we need to retry the cross-call.
1619
	 */
1620
	if (event->state == PERF_EVENT_STATE_ACTIVE) {
1621
		raw_spin_unlock_irq(&ctx->lock);
1622 1623 1624 1625 1626
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
1627 1628 1629 1630 1631 1632 1633
		goto retry;
	}

	/*
	 * Since we have the lock this context can't be scheduled
	 * in, so we can change the state safely.
	 */
1634 1635 1636
	if (event->state == PERF_EVENT_STATE_INACTIVE) {
		update_group_times(event);
		event->state = PERF_EVENT_STATE_OFF;
1637
	}
1638
	raw_spin_unlock_irq(&ctx->lock);
1639
}
1640
EXPORT_SYMBOL_GPL(perf_event_disable);
1641

S
Stephane Eranian 已提交
1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676
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
1677 1678 1679 1680
#define MAX_INTERRUPTS (~0ULL)

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

1681
static int
1682
event_sched_in(struct perf_event *event,
1683
		 struct perf_cpu_context *cpuctx,
1684
		 struct perf_event_context *ctx)
1685
{
1686
	u64 tstamp = perf_event_time(event);
1687
	int ret = 0;
1688

1689 1690
	lockdep_assert_held(&ctx->lock);

1691
	if (event->state <= PERF_EVENT_STATE_OFF)
1692 1693
		return 0;

1694
	event->state = PERF_EVENT_STATE_ACTIVE;
1695
	event->oncpu = smp_processor_id();
P
Peter Zijlstra 已提交
1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706

	/*
	 * 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;
	}

1707 1708 1709 1710 1711
	/*
	 * The new state must be visible before we turn it on in the hardware:
	 */
	smp_wmb();

1712 1713
	perf_pmu_disable(event->pmu);

P
Peter Zijlstra 已提交
1714
	if (event->pmu->add(event, PERF_EF_START)) {
1715 1716
		event->state = PERF_EVENT_STATE_INACTIVE;
		event->oncpu = -1;
1717 1718
		ret = -EAGAIN;
		goto out;
1719 1720
	}

1721
	event->tstamp_running += tstamp - event->tstamp_stopped;
1722

S
Stephane Eranian 已提交
1723
	perf_set_shadow_time(event, ctx, tstamp);
1724

1725
	if (!is_software_event(event))
1726
		cpuctx->active_oncpu++;
1727
	ctx->nr_active++;
1728 1729
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq++;
1730

1731
	if (event->attr.exclusive)
1732 1733
		cpuctx->exclusive = 1;

1734 1735 1736 1737
out:
	perf_pmu_enable(event->pmu);

	return ret;
1738 1739
}

1740
static int
1741
group_sched_in(struct perf_event *group_event,
1742
	       struct perf_cpu_context *cpuctx,
1743
	       struct perf_event_context *ctx)
1744
{
1745
	struct perf_event *event, *partial_group = NULL;
P
Peter Zijlstra 已提交
1746
	struct pmu *pmu = ctx->pmu;
1747 1748
	u64 now = ctx->time;
	bool simulate = false;
1749

1750
	if (group_event->state == PERF_EVENT_STATE_OFF)
1751 1752
		return 0;

P
Peter Zijlstra 已提交
1753
	pmu->start_txn(pmu);
1754

1755
	if (event_sched_in(group_event, cpuctx, ctx)) {
P
Peter Zijlstra 已提交
1756
		pmu->cancel_txn(pmu);
1757
		perf_cpu_hrtimer_restart(cpuctx);
1758
		return -EAGAIN;
1759
	}
1760 1761 1762 1763

	/*
	 * Schedule in siblings as one group (if any):
	 */
1764
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1765
		if (event_sched_in(event, cpuctx, ctx)) {
1766
			partial_group = event;
1767 1768 1769 1770
			goto group_error;
		}
	}

1771
	if (!pmu->commit_txn(pmu))
1772
		return 0;
1773

1774 1775 1776 1777
group_error:
	/*
	 * Groups can be scheduled in as one unit only, so undo any
	 * partial group before returning:
1778 1779 1780 1781 1782 1783 1784 1785 1786 1787
	 * 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.
1788
	 */
1789 1790
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
		if (event == partial_group)
1791 1792 1793 1794 1795 1796 1797 1798
			simulate = true;

		if (simulate) {
			event->tstamp_running += now - event->tstamp_stopped;
			event->tstamp_stopped = now;
		} else {
			event_sched_out(event, cpuctx, ctx);
		}
1799
	}
1800
	event_sched_out(group_event, cpuctx, ctx);
1801

P
Peter Zijlstra 已提交
1802
	pmu->cancel_txn(pmu);
1803

1804 1805
	perf_cpu_hrtimer_restart(cpuctx);

1806 1807 1808
	return -EAGAIN;
}

1809
/*
1810
 * Work out whether we can put this event group on the CPU now.
1811
 */
1812
static int group_can_go_on(struct perf_event *event,
1813 1814 1815 1816
			   struct perf_cpu_context *cpuctx,
			   int can_add_hw)
{
	/*
1817
	 * Groups consisting entirely of software events can always go on.
1818
	 */
1819
	if (event->group_flags & PERF_GROUP_SOFTWARE)
1820 1821 1822
		return 1;
	/*
	 * If an exclusive group is already on, no other hardware
1823
	 * events can go on.
1824 1825 1826 1827 1828
	 */
	if (cpuctx->exclusive)
		return 0;
	/*
	 * If this group is exclusive and there are already
1829
	 * events on the CPU, it can't go on.
1830
	 */
1831
	if (event->attr.exclusive && cpuctx->active_oncpu)
1832 1833 1834 1835 1836 1837 1838 1839
		return 0;
	/*
	 * Otherwise, try to add it if all previous groups were able
	 * to go on.
	 */
	return can_add_hw;
}

1840 1841
static void add_event_to_ctx(struct perf_event *event,
			       struct perf_event_context *ctx)
1842
{
1843 1844
	u64 tstamp = perf_event_time(event);

1845
	list_add_event(event, ctx);
1846
	perf_group_attach(event);
1847 1848 1849
	event->tstamp_enabled = tstamp;
	event->tstamp_running = tstamp;
	event->tstamp_stopped = tstamp;
1850 1851
}

1852 1853 1854 1855 1856 1857
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);
1858

1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870
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
1871
/*
1872
 * Cross CPU call to install and enable a performance event
1873 1874
 *
 * Must be called with ctx->mutex held
T
Thomas Gleixner 已提交
1875
 */
1876
static int  __perf_install_in_context(void *info)
T
Thomas Gleixner 已提交
1877
{
1878 1879
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1880
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1881 1882 1883
	struct perf_event_context *task_ctx = cpuctx->task_ctx;
	struct task_struct *task = current;

1884
	perf_ctx_lock(cpuctx, task_ctx);
1885
	perf_pmu_disable(cpuctx->ctx.pmu);
T
Thomas Gleixner 已提交
1886 1887

	/*
1888
	 * If there was an active task_ctx schedule it out.
T
Thomas Gleixner 已提交
1889
	 */
1890
	if (task_ctx)
1891
		task_ctx_sched_out(task_ctx);
1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905

	/*
	 * 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;
1906 1907
		task = task_ctx->task;
	}
1908

1909
	cpu_ctx_sched_out(cpuctx, EVENT_ALL);
T
Thomas Gleixner 已提交
1910

1911
	update_context_time(ctx);
S
Stephane Eranian 已提交
1912 1913 1914 1915 1916 1917
	/*
	 * 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
1918

1919
	add_event_to_ctx(event, ctx);
T
Thomas Gleixner 已提交
1920

1921
	/*
1922
	 * Schedule everything back in
1923
	 */
1924
	perf_event_sched_in(cpuctx, task_ctx, task);
1925 1926 1927

	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, task_ctx);
1928 1929

	return 0;
T
Thomas Gleixner 已提交
1930 1931 1932
}

/*
1933
 * Attach a performance event to a context
T
Thomas Gleixner 已提交
1934
 *
1935 1936
 * First we add the event to the list with the hardware enable bit
 * in event->hw_config cleared.
T
Thomas Gleixner 已提交
1937
 *
1938
 * If the event is attached to a task which is on a CPU we use a smp
T
Thomas Gleixner 已提交
1939 1940 1941 1942
 * 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
1943 1944
perf_install_in_context(struct perf_event_context *ctx,
			struct perf_event *event,
T
Thomas Gleixner 已提交
1945 1946 1947 1948
			int cpu)
{
	struct task_struct *task = ctx->task;

1949 1950
	lockdep_assert_held(&ctx->mutex);

1951
	event->ctx = ctx;
1952 1953
	if (event->cpu != -1)
		event->cpu = cpu;
1954

T
Thomas Gleixner 已提交
1955 1956
	if (!task) {
		/*
1957
		 * Per cpu events are installed via an smp call and
1958
		 * the install is always successful.
T
Thomas Gleixner 已提交
1959
		 */
1960
		cpu_function_call(cpu, __perf_install_in_context, event);
T
Thomas Gleixner 已提交
1961 1962 1963 1964
		return;
	}

retry:
1965 1966
	if (!task_function_call(task, __perf_install_in_context, event))
		return;
T
Thomas Gleixner 已提交
1967

1968
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1969
	/*
1970 1971
	 * 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
1972
	 */
1973
	if (ctx->is_active) {
1974
		raw_spin_unlock_irq(&ctx->lock);
1975 1976 1977 1978 1979
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
T
Thomas Gleixner 已提交
1980 1981 1982 1983
		goto retry;
	}

	/*
1984 1985
	 * 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
1986
	 */
1987
	add_event_to_ctx(event, ctx);
1988
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1989 1990
}

1991
/*
1992
 * Put a event into inactive state and update time fields.
1993 1994 1995 1996 1997 1998
 * 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.
 */
1999
static void __perf_event_mark_enabled(struct perf_event *event)
2000
{
2001
	struct perf_event *sub;
2002
	u64 tstamp = perf_event_time(event);
2003

2004
	event->state = PERF_EVENT_STATE_INACTIVE;
2005
	event->tstamp_enabled = tstamp - event->total_time_enabled;
P
Peter Zijlstra 已提交
2006
	list_for_each_entry(sub, &event->sibling_list, group_entry) {
2007 2008
		if (sub->state >= PERF_EVENT_STATE_INACTIVE)
			sub->tstamp_enabled = tstamp - sub->total_time_enabled;
P
Peter Zijlstra 已提交
2009
	}
2010 2011
}

2012
/*
2013
 * Cross CPU call to enable a performance event
2014
 */
2015
static int __perf_event_enable(void *info)
2016
{
2017 2018 2019
	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
2020
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2021
	int err;
2022

2023 2024 2025 2026 2027 2028 2029 2030 2031 2032
	/*
	 * 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)
2033
		return -EINVAL;
2034

2035
	raw_spin_lock(&ctx->lock);
2036
	update_context_time(ctx);
2037

2038
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2039
		goto unlock;
S
Stephane Eranian 已提交
2040 2041 2042 2043

	/*
	 * set current task's cgroup time reference point
	 */
2044
	perf_cgroup_set_timestamp(current, ctx);
S
Stephane Eranian 已提交
2045

2046
	__perf_event_mark_enabled(event);
2047

S
Stephane Eranian 已提交
2048 2049 2050
	if (!event_filter_match(event)) {
		if (is_cgroup_event(event))
			perf_cgroup_defer_enabled(event);
2051
		goto unlock;
S
Stephane Eranian 已提交
2052
	}
2053

2054
	/*
2055
	 * If the event is in a group and isn't the group leader,
2056
	 * then don't put it on unless the group is on.
2057
	 */
2058
	if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2059
		goto unlock;
2060

2061
	if (!group_can_go_on(event, cpuctx, 1)) {
2062
		err = -EEXIST;
2063
	} else {
2064
		if (event == leader)
2065
			err = group_sched_in(event, cpuctx, ctx);
2066
		else
2067
			err = event_sched_in(event, cpuctx, ctx);
2068
	}
2069 2070 2071

	if (err) {
		/*
2072
		 * If this event can't go on and it's part of a
2073 2074
		 * group, then the whole group has to come off.
		 */
2075
		if (leader != event) {
2076
			group_sched_out(leader, cpuctx, ctx);
2077 2078
			perf_cpu_hrtimer_restart(cpuctx);
		}
2079
		if (leader->attr.pinned) {
2080
			update_group_times(leader);
2081
			leader->state = PERF_EVENT_STATE_ERROR;
2082
		}
2083 2084
	}

P
Peter Zijlstra 已提交
2085
unlock:
2086
	raw_spin_unlock(&ctx->lock);
2087 2088

	return 0;
2089 2090 2091
}

/*
2092
 * Enable a event.
2093
 *
2094 2095
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
2096
 * remains valid.  This condition is satisfied when called through
2097 2098
 * perf_event_for_each_child or perf_event_for_each as described
 * for perf_event_disable.
2099
 */
2100
void perf_event_enable(struct perf_event *event)
2101
{
2102
	struct perf_event_context *ctx = event->ctx;
2103 2104 2105 2106
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
2107
		 * Enable the event on the cpu that it's on
2108
		 */
2109
		cpu_function_call(event->cpu, __perf_event_enable, event);
2110 2111 2112
		return;
	}

2113
	raw_spin_lock_irq(&ctx->lock);
2114
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2115 2116 2117
		goto out;

	/*
2118 2119
	 * If the event is in error state, clear that first.
	 * That way, if we see the event in error state below, we
2120 2121 2122 2123
	 * know that it has gone back into error state, as distinct
	 * from the task having been scheduled away before the
	 * cross-call arrived.
	 */
2124 2125
	if (event->state == PERF_EVENT_STATE_ERROR)
		event->state = PERF_EVENT_STATE_OFF;
2126

P
Peter Zijlstra 已提交
2127
retry:
2128
	if (!ctx->is_active) {
2129
		__perf_event_mark_enabled(event);
2130 2131 2132
		goto out;
	}

2133
	raw_spin_unlock_irq(&ctx->lock);
2134 2135 2136

	if (!task_function_call(task, __perf_event_enable, event))
		return;
2137

2138
	raw_spin_lock_irq(&ctx->lock);
2139 2140

	/*
2141
	 * If the context is active and the event is still off,
2142 2143
	 * we need to retry the cross-call.
	 */
2144 2145 2146 2147 2148 2149
	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;
2150
		goto retry;
2151
	}
2152

P
Peter Zijlstra 已提交
2153
out:
2154
	raw_spin_unlock_irq(&ctx->lock);
2155
}
2156
EXPORT_SYMBOL_GPL(perf_event_enable);
2157

2158
int perf_event_refresh(struct perf_event *event, int refresh)
2159
{
2160
	/*
2161
	 * not supported on inherited events
2162
	 */
2163
	if (event->attr.inherit || !is_sampling_event(event))
2164 2165
		return -EINVAL;

2166 2167
	atomic_add(refresh, &event->event_limit);
	perf_event_enable(event);
2168 2169

	return 0;
2170
}
2171
EXPORT_SYMBOL_GPL(perf_event_refresh);
2172

2173 2174 2175
static void ctx_sched_out(struct perf_event_context *ctx,
			  struct perf_cpu_context *cpuctx,
			  enum event_type_t event_type)
2176
{
2177
	struct perf_event *event;
2178
	int is_active = ctx->is_active;
2179

2180
	ctx->is_active &= ~event_type;
2181
	if (likely(!ctx->nr_events))
2182 2183
		return;

2184
	update_context_time(ctx);
S
Stephane Eranian 已提交
2185
	update_cgrp_time_from_cpuctx(cpuctx);
2186
	if (!ctx->nr_active)
2187
		return;
2188

P
Peter Zijlstra 已提交
2189
	perf_pmu_disable(ctx->pmu);
2190
	if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2191 2192
		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2193
	}
2194

2195
	if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2196
		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2197
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2198
	}
P
Peter Zijlstra 已提交
2199
	perf_pmu_enable(ctx->pmu);
2200 2201
}

2202
/*
2203 2204 2205 2206 2207 2208
 * Test whether two contexts are equivalent, i.e. whether they have both been
 * cloned from the same version of the same context.
 *
 * Equivalence is measured using a generation number in the context that is
 * incremented on each modification to it; see unclone_ctx(), list_add_event()
 * and list_del_event().
2209
 */
2210 2211
static int context_equiv(struct perf_event_context *ctx1,
			 struct perf_event_context *ctx2)
2212
{
2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234
	/* Pinning disables the swap optimization */
	if (ctx1->pin_count || ctx2->pin_count)
		return 0;

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

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

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

	/* Unmatched */
	return 0;
2235 2236
}

2237 2238
static void __perf_event_sync_stat(struct perf_event *event,
				     struct perf_event *next_event)
2239 2240 2241
{
	u64 value;

2242
	if (!event->attr.inherit_stat)
2243 2244 2245
		return;

	/*
2246
	 * Update the event value, we cannot use perf_event_read()
2247 2248
	 * because we're in the middle of a context switch and have IRQs
	 * disabled, which upsets smp_call_function_single(), however
2249
	 * we know the event must be on the current CPU, therefore we
2250 2251
	 * don't need to use it.
	 */
2252 2253
	switch (event->state) {
	case PERF_EVENT_STATE_ACTIVE:
2254 2255
		event->pmu->read(event);
		/* fall-through */
2256

2257 2258
	case PERF_EVENT_STATE_INACTIVE:
		update_event_times(event);
2259 2260 2261 2262 2263 2264 2265
		break;

	default:
		break;
	}

	/*
2266
	 * In order to keep per-task stats reliable we need to flip the event
2267 2268
	 * values when we flip the contexts.
	 */
2269 2270 2271
	value = local64_read(&next_event->count);
	value = local64_xchg(&event->count, value);
	local64_set(&next_event->count, value);
2272

2273 2274
	swap(event->total_time_enabled, next_event->total_time_enabled);
	swap(event->total_time_running, next_event->total_time_running);
2275

2276
	/*
2277
	 * Since we swizzled the values, update the user visible data too.
2278
	 */
2279 2280
	perf_event_update_userpage(event);
	perf_event_update_userpage(next_event);
2281 2282
}

2283 2284
static void perf_event_sync_stat(struct perf_event_context *ctx,
				   struct perf_event_context *next_ctx)
2285
{
2286
	struct perf_event *event, *next_event;
2287 2288 2289 2290

	if (!ctx->nr_stat)
		return;

2291 2292
	update_context_time(ctx);

2293 2294
	event = list_first_entry(&ctx->event_list,
				   struct perf_event, event_entry);
2295

2296 2297
	next_event = list_first_entry(&next_ctx->event_list,
					struct perf_event, event_entry);
2298

2299 2300
	while (&event->event_entry != &ctx->event_list &&
	       &next_event->event_entry != &next_ctx->event_list) {
2301

2302
		__perf_event_sync_stat(event, next_event);
2303

2304 2305
		event = list_next_entry(event, event_entry);
		next_event = list_next_entry(next_event, event_entry);
2306 2307 2308
	}
}

2309 2310
static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
					 struct task_struct *next)
T
Thomas Gleixner 已提交
2311
{
P
Peter Zijlstra 已提交
2312
	struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2313
	struct perf_event_context *next_ctx;
2314
	struct perf_event_context *parent, *next_parent;
P
Peter Zijlstra 已提交
2315
	struct perf_cpu_context *cpuctx;
2316
	int do_switch = 1;
T
Thomas Gleixner 已提交
2317

P
Peter Zijlstra 已提交
2318 2319
	if (likely(!ctx))
		return;
2320

P
Peter Zijlstra 已提交
2321 2322
	cpuctx = __get_cpu_context(ctx);
	if (!cpuctx->task_ctx)
T
Thomas Gleixner 已提交
2323 2324
		return;

2325
	rcu_read_lock();
P
Peter Zijlstra 已提交
2326
	next_ctx = next->perf_event_ctxp[ctxn];
2327 2328 2329 2330 2331 2332 2333
	if (!next_ctx)
		goto unlock;

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

	/* If neither context have a parent context; they cannot be clones. */
2334
	if (!parent || !next_parent)
2335 2336 2337
		goto unlock;

	if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2338 2339 2340 2341 2342 2343 2344 2345 2346
		/*
		 * 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.
		 */
2347 2348
		raw_spin_lock(&ctx->lock);
		raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2349
		if (context_equiv(ctx, next_ctx)) {
2350 2351
			/*
			 * XXX do we need a memory barrier of sorts
2352
			 * wrt to rcu_dereference() of perf_event_ctxp
2353
			 */
P
Peter Zijlstra 已提交
2354 2355
			task->perf_event_ctxp[ctxn] = next_ctx;
			next->perf_event_ctxp[ctxn] = ctx;
2356 2357 2358
			ctx->task = next;
			next_ctx->task = task;
			do_switch = 0;
2359

2360
			perf_event_sync_stat(ctx, next_ctx);
2361
		}
2362 2363
		raw_spin_unlock(&next_ctx->lock);
		raw_spin_unlock(&ctx->lock);
2364
	}
2365
unlock:
2366
	rcu_read_unlock();
2367

2368
	if (do_switch) {
2369
		raw_spin_lock(&ctx->lock);
2370
		ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2371
		cpuctx->task_ctx = NULL;
2372
		raw_spin_unlock(&ctx->lock);
2373
	}
T
Thomas Gleixner 已提交
2374 2375
}

P
Peter Zijlstra 已提交
2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389
#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.
 */
2390 2391
void __perf_event_task_sched_out(struct task_struct *task,
				 struct task_struct *next)
P
Peter Zijlstra 已提交
2392 2393 2394 2395 2396
{
	int ctxn;

	for_each_task_context_nr(ctxn)
		perf_event_context_sched_out(task, ctxn, next);
S
Stephane Eranian 已提交
2397 2398 2399 2400 2401 2402 2403

	/*
	 * 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)))
2404
		perf_cgroup_sched_out(task, next);
P
Peter Zijlstra 已提交
2405 2406
}

2407
static void task_ctx_sched_out(struct perf_event_context *ctx)
2408
{
P
Peter Zijlstra 已提交
2409
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2410

2411 2412
	if (!cpuctx->task_ctx)
		return;
2413 2414 2415 2416

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

2417
	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2418 2419 2420
	cpuctx->task_ctx = NULL;
}

2421 2422 2423 2424 2425 2426 2427
/*
 * 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);
2428 2429
}

2430
static void
2431
ctx_pinned_sched_in(struct perf_event_context *ctx,
2432
		    struct perf_cpu_context *cpuctx)
T
Thomas Gleixner 已提交
2433
{
2434
	struct perf_event *event;
T
Thomas Gleixner 已提交
2435

2436 2437
	list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
		if (event->state <= PERF_EVENT_STATE_OFF)
2438
			continue;
2439
		if (!event_filter_match(event))
2440 2441
			continue;

S
Stephane Eranian 已提交
2442 2443 2444 2445
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

2446
		if (group_can_go_on(event, cpuctx, 1))
2447
			group_sched_in(event, cpuctx, ctx);
2448 2449 2450 2451 2452

		/*
		 * If this pinned group hasn't been scheduled,
		 * put it in error state.
		 */
2453 2454 2455
		if (event->state == PERF_EVENT_STATE_INACTIVE) {
			update_group_times(event);
			event->state = PERF_EVENT_STATE_ERROR;
2456
		}
2457
	}
2458 2459 2460 2461
}

static void
ctx_flexible_sched_in(struct perf_event_context *ctx,
2462
		      struct perf_cpu_context *cpuctx)
2463 2464 2465
{
	struct perf_event *event;
	int can_add_hw = 1;
2466

2467 2468 2469
	list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
		/* Ignore events in OFF or ERROR state */
		if (event->state <= PERF_EVENT_STATE_OFF)
2470
			continue;
2471 2472
		/*
		 * Listen to the 'cpu' scheduling filter constraint
2473
		 * of events:
2474
		 */
2475
		if (!event_filter_match(event))
T
Thomas Gleixner 已提交
2476 2477
			continue;

S
Stephane Eranian 已提交
2478 2479 2480 2481
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

P
Peter Zijlstra 已提交
2482
		if (group_can_go_on(event, cpuctx, can_add_hw)) {
2483
			if (group_sched_in(event, cpuctx, ctx))
2484
				can_add_hw = 0;
P
Peter Zijlstra 已提交
2485
		}
T
Thomas Gleixner 已提交
2486
	}
2487 2488 2489 2490 2491
}

static void
ctx_sched_in(struct perf_event_context *ctx,
	     struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2492 2493
	     enum event_type_t event_type,
	     struct task_struct *task)
2494
{
S
Stephane Eranian 已提交
2495
	u64 now;
2496
	int is_active = ctx->is_active;
S
Stephane Eranian 已提交
2497

2498
	ctx->is_active |= event_type;
2499
	if (likely(!ctx->nr_events))
2500
		return;
2501

S
Stephane Eranian 已提交
2502 2503
	now = perf_clock();
	ctx->timestamp = now;
2504
	perf_cgroup_set_timestamp(task, ctx);
2505 2506 2507 2508
	/*
	 * First go through the list and put on any pinned groups
	 * in order to give them the best chance of going on.
	 */
2509
	if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2510
		ctx_pinned_sched_in(ctx, cpuctx);
2511 2512

	/* Then walk through the lower prio flexible groups */
2513
	if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2514
		ctx_flexible_sched_in(ctx, cpuctx);
2515 2516
}

2517
static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2518 2519
			     enum event_type_t event_type,
			     struct task_struct *task)
2520 2521 2522
{
	struct perf_event_context *ctx = &cpuctx->ctx;

S
Stephane Eranian 已提交
2523
	ctx_sched_in(ctx, cpuctx, event_type, task);
2524 2525
}

S
Stephane Eranian 已提交
2526 2527
static void perf_event_context_sched_in(struct perf_event_context *ctx,
					struct task_struct *task)
2528
{
P
Peter Zijlstra 已提交
2529
	struct perf_cpu_context *cpuctx;
2530

P
Peter Zijlstra 已提交
2531
	cpuctx = __get_cpu_context(ctx);
2532 2533 2534
	if (cpuctx->task_ctx == ctx)
		return;

2535
	perf_ctx_lock(cpuctx, ctx);
P
Peter Zijlstra 已提交
2536
	perf_pmu_disable(ctx->pmu);
2537 2538 2539 2540 2541 2542 2543
	/*
	 * 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);

2544 2545
	if (ctx->nr_events)
		cpuctx->task_ctx = ctx;
2546

2547 2548
	perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);

2549 2550 2551
	perf_pmu_enable(ctx->pmu);
	perf_ctx_unlock(cpuctx, ctx);

2552 2553 2554 2555
	/*
	 * 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
2556
	perf_pmu_rotate_start(ctx->pmu);
2557 2558
}

2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616
/*
 * When sampling the branck stack in system-wide, it may be necessary
 * to flush the stack on context switch. This happens when the branch
 * stack does not tag its entries with the pid of the current task.
 * Otherwise it becomes impossible to associate a branch entry with a
 * task. This ambiguity is more likely to appear when the branch stack
 * supports priv level filtering and the user sets it to monitor only
 * at the user level (which could be a useful measurement in system-wide
 * mode). In that case, the risk is high of having a branch stack with
 * branch from multiple tasks. Flushing may mean dropping the existing
 * entries or stashing them somewhere in the PMU specific code layer.
 *
 * This function provides the context switch callback to the lower code
 * layer. It is invoked ONLY when there is at least one system-wide context
 * with at least one active event using taken branch sampling.
 */
static void perf_branch_stack_sched_in(struct task_struct *prev,
				       struct task_struct *task)
{
	struct perf_cpu_context *cpuctx;
	struct pmu *pmu;
	unsigned long flags;

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

	local_irq_save(flags);

	rcu_read_lock();

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

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

			perf_ctx_lock(cpuctx, cpuctx->task_ctx);

			perf_pmu_disable(pmu);

			pmu->flush_branch_stack();

			perf_pmu_enable(pmu);

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

	rcu_read_unlock();

	local_irq_restore(flags);
}

P
Peter Zijlstra 已提交
2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627
/*
 * 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.
 */
2628 2629
void __perf_event_task_sched_in(struct task_struct *prev,
				struct task_struct *task)
P
Peter Zijlstra 已提交
2630 2631 2632 2633 2634 2635 2636 2637 2638
{
	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
2639
		perf_event_context_sched_in(ctx, task);
P
Peter Zijlstra 已提交
2640
	}
S
Stephane Eranian 已提交
2641 2642 2643 2644 2645 2646
	/*
	 * 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)))
2647
		perf_cgroup_sched_in(prev, task);
2648 2649 2650 2651

	/* check for system-wide branch_stack events */
	if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
		perf_branch_stack_sched_in(prev, task);
2652 2653
}

2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680
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.
	 */
2681
#define REDUCE_FLS(a, b)		\
2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720
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;
	}

2721 2722 2723
	if (!divisor)
		return dividend;

2724 2725 2726
	return div64_u64(dividend, divisor);
}

2727 2728 2729
static DEFINE_PER_CPU(int, perf_throttled_count);
static DEFINE_PER_CPU(u64, perf_throttled_seq);

2730
static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2731
{
2732
	struct hw_perf_event *hwc = &event->hw;
2733
	s64 period, sample_period;
2734 2735
	s64 delta;

2736
	period = perf_calculate_period(event, nsec, count);
2737 2738 2739 2740 2741 2742 2743 2744 2745 2746

	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;
2747

2748
	if (local64_read(&hwc->period_left) > 8*sample_period) {
2749 2750 2751
		if (disable)
			event->pmu->stop(event, PERF_EF_UPDATE);

2752
		local64_set(&hwc->period_left, 0);
2753 2754 2755

		if (disable)
			event->pmu->start(event, PERF_EF_RELOAD);
2756
	}
2757 2758
}

2759 2760 2761 2762 2763 2764 2765
/*
 * 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)
2766
{
2767 2768
	struct perf_event *event;
	struct hw_perf_event *hwc;
2769
	u64 now, period = TICK_NSEC;
2770
	s64 delta;
2771

2772 2773 2774 2775 2776 2777
	/*
	 * 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))
2778 2779
		return;

2780
	raw_spin_lock(&ctx->lock);
2781
	perf_pmu_disable(ctx->pmu);
2782

2783
	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2784
		if (event->state != PERF_EVENT_STATE_ACTIVE)
2785 2786
			continue;

2787
		if (!event_filter_match(event))
2788 2789
			continue;

2790 2791
		perf_pmu_disable(event->pmu);

2792
		hwc = &event->hw;
2793

2794
		if (hwc->interrupts == MAX_INTERRUPTS) {
2795
			hwc->interrupts = 0;
2796
			perf_log_throttle(event, 1);
P
Peter Zijlstra 已提交
2797
			event->pmu->start(event, 0);
2798 2799
		}

2800
		if (!event->attr.freq || !event->attr.sample_freq)
2801
			goto next;
2802

2803 2804 2805 2806 2807
		/*
		 * stop the event and update event->count
		 */
		event->pmu->stop(event, PERF_EF_UPDATE);

2808
		now = local64_read(&event->count);
2809 2810
		delta = now - hwc->freq_count_stamp;
		hwc->freq_count_stamp = now;
2811

2812 2813 2814
		/*
		 * restart the event
		 * reload only if value has changed
2815 2816 2817
		 * we have stopped the event so tell that
		 * to perf_adjust_period() to avoid stopping it
		 * twice.
2818
		 */
2819
		if (delta > 0)
2820
			perf_adjust_period(event, period, delta, false);
2821 2822

		event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2823 2824
	next:
		perf_pmu_enable(event->pmu);
2825
	}
2826

2827
	perf_pmu_enable(ctx->pmu);
2828
	raw_spin_unlock(&ctx->lock);
2829 2830
}

2831
/*
2832
 * Round-robin a context's events:
2833
 */
2834
static void rotate_ctx(struct perf_event_context *ctx)
T
Thomas Gleixner 已提交
2835
{
2836 2837 2838 2839 2840 2841
	/*
	 * 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);
2842 2843
}

2844
/*
2845 2846 2847
 * 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.
2848
 */
2849
static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2850
{
P
Peter Zijlstra 已提交
2851
	struct perf_event_context *ctx = NULL;
2852
	int rotate = 0, remove = 1;
2853

2854
	if (cpuctx->ctx.nr_events) {
2855
		remove = 0;
2856 2857 2858
		if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
			rotate = 1;
	}
2859

P
Peter Zijlstra 已提交
2860
	ctx = cpuctx->task_ctx;
2861
	if (ctx && ctx->nr_events) {
2862
		remove = 0;
2863 2864 2865
		if (ctx->nr_events != ctx->nr_active)
			rotate = 1;
	}
2866

2867
	if (!rotate)
2868 2869
		goto done;

2870
	perf_ctx_lock(cpuctx, cpuctx->task_ctx);
P
Peter Zijlstra 已提交
2871
	perf_pmu_disable(cpuctx->ctx.pmu);
2872

2873 2874 2875
	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
	if (ctx)
		ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
T
Thomas Gleixner 已提交
2876

2877 2878 2879
	rotate_ctx(&cpuctx->ctx);
	if (ctx)
		rotate_ctx(ctx);
2880

2881
	perf_event_sched_in(cpuctx, ctx, current);
2882

2883 2884
	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2885
done:
2886 2887
	if (remove)
		list_del_init(&cpuctx->rotation_list);
2888 2889

	return rotate;
2890 2891
}

2892 2893 2894
#ifdef CONFIG_NO_HZ_FULL
bool perf_event_can_stop_tick(void)
{
2895
	if (atomic_read(&nr_freq_events) ||
2896
	    __this_cpu_read(perf_throttled_count))
2897
		return false;
2898 2899
	else
		return true;
2900 2901 2902
}
#endif

2903 2904 2905 2906
void perf_event_task_tick(void)
{
	struct list_head *head = &__get_cpu_var(rotation_list);
	struct perf_cpu_context *cpuctx, *tmp;
2907 2908
	struct perf_event_context *ctx;
	int throttled;
2909

2910 2911
	WARN_ON(!irqs_disabled());

2912 2913 2914
	__this_cpu_inc(perf_throttled_seq);
	throttled = __this_cpu_xchg(perf_throttled_count, 0);

2915
	list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2916 2917 2918 2919 2920 2921
		ctx = &cpuctx->ctx;
		perf_adjust_freq_unthr_context(ctx, throttled);

		ctx = cpuctx->task_ctx;
		if (ctx)
			perf_adjust_freq_unthr_context(ctx, throttled);
2922
	}
T
Thomas Gleixner 已提交
2923 2924
}

2925 2926 2927 2928 2929 2930 2931 2932 2933 2934
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;

2935
	__perf_event_mark_enabled(event);
2936 2937 2938 2939

	return 1;
}

2940
/*
2941
 * Enable all of a task's events that have been marked enable-on-exec.
2942 2943
 * This expects task == current.
 */
P
Peter Zijlstra 已提交
2944
static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2945
{
2946
	struct perf_event *event;
2947 2948
	unsigned long flags;
	int enabled = 0;
2949
	int ret;
2950 2951

	local_irq_save(flags);
2952
	if (!ctx || !ctx->nr_events)
2953 2954
		goto out;

2955 2956 2957 2958 2959 2960 2961
	/*
	 * 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.
	 */
2962
	perf_cgroup_sched_out(current, NULL);
2963

2964
	raw_spin_lock(&ctx->lock);
2965
	task_ctx_sched_out(ctx);
2966

2967
	list_for_each_entry(event, &ctx->event_list, event_entry) {
2968 2969 2970
		ret = event_enable_on_exec(event, ctx);
		if (ret)
			enabled = 1;
2971 2972 2973
	}

	/*
2974
	 * Unclone this context if we enabled any event.
2975
	 */
2976 2977
	if (enabled)
		unclone_ctx(ctx);
2978

2979
	raw_spin_unlock(&ctx->lock);
2980

2981 2982 2983
	/*
	 * Also calls ctxswin for cgroup events, if any:
	 */
S
Stephane Eranian 已提交
2984
	perf_event_context_sched_in(ctx, ctx->task);
P
Peter Zijlstra 已提交
2985
out:
2986 2987 2988
	local_irq_restore(flags);
}

2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004
void perf_event_exec(void)
{
	struct perf_event_context *ctx;
	int ctxn;

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

		perf_event_enable_on_exec(ctx);
	}
	rcu_read_unlock();
}

T
Thomas Gleixner 已提交
3005
/*
3006
 * Cross CPU call to read the hardware event
T
Thomas Gleixner 已提交
3007
 */
3008
static void __perf_event_read(void *info)
T
Thomas Gleixner 已提交
3009
{
3010 3011
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
3012
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
I
Ingo Molnar 已提交
3013

3014 3015 3016 3017
	/*
	 * 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
3018 3019
	 * event->count would have been updated to a recent sample
	 * when the event was scheduled out.
3020 3021 3022 3023
	 */
	if (ctx->task && cpuctx->task_ctx != ctx)
		return;

3024
	raw_spin_lock(&ctx->lock);
S
Stephane Eranian 已提交
3025
	if (ctx->is_active) {
3026
		update_context_time(ctx);
S
Stephane Eranian 已提交
3027 3028
		update_cgrp_time_from_event(event);
	}
3029
	update_event_times(event);
3030 3031
	if (event->state == PERF_EVENT_STATE_ACTIVE)
		event->pmu->read(event);
3032
	raw_spin_unlock(&ctx->lock);
T
Thomas Gleixner 已提交
3033 3034
}

P
Peter Zijlstra 已提交
3035 3036
static inline u64 perf_event_count(struct perf_event *event)
{
3037
	return local64_read(&event->count) + atomic64_read(&event->child_count);
P
Peter Zijlstra 已提交
3038 3039
}

3040
static u64 perf_event_read(struct perf_event *event)
T
Thomas Gleixner 已提交
3041 3042
{
	/*
3043 3044
	 * If event is enabled and currently active on a CPU, update the
	 * value in the event structure:
T
Thomas Gleixner 已提交
3045
	 */
3046 3047 3048 3049
	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
3050 3051 3052
		struct perf_event_context *ctx = event->ctx;
		unsigned long flags;

3053
		raw_spin_lock_irqsave(&ctx->lock, flags);
3054 3055 3056 3057 3058
		/*
		 * 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
3059
		if (ctx->is_active) {
3060
			update_context_time(ctx);
S
Stephane Eranian 已提交
3061 3062
			update_cgrp_time_from_event(event);
		}
3063
		update_event_times(event);
3064
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
T
Thomas Gleixner 已提交
3065 3066
	}

P
Peter Zijlstra 已提交
3067
	return perf_event_count(event);
T
Thomas Gleixner 已提交
3068 3069
}

3070
/*
3071
 * Initialize the perf_event context in a task_struct:
3072
 */
3073
static void __perf_event_init_context(struct perf_event_context *ctx)
3074
{
3075
	raw_spin_lock_init(&ctx->lock);
3076
	mutex_init(&ctx->mutex);
3077 3078
	INIT_LIST_HEAD(&ctx->pinned_groups);
	INIT_LIST_HEAD(&ctx->flexible_groups);
3079 3080
	INIT_LIST_HEAD(&ctx->event_list);
	atomic_set(&ctx->refcount, 1);
3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095
}

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
3096
	}
3097 3098 3099
	ctx->pmu = pmu;

	return ctx;
3100 3101
}

3102 3103 3104 3105 3106
static struct task_struct *
find_lively_task_by_vpid(pid_t vpid)
{
	struct task_struct *task;
	int err;
T
Thomas Gleixner 已提交
3107 3108

	rcu_read_lock();
3109
	if (!vpid)
T
Thomas Gleixner 已提交
3110 3111
		task = current;
	else
3112
		task = find_task_by_vpid(vpid);
T
Thomas Gleixner 已提交
3113 3114 3115 3116 3117 3118 3119 3120
	if (task)
		get_task_struct(task);
	rcu_read_unlock();

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

	/* Reuse ptrace permission checks for now. */
3121 3122 3123 3124
	err = -EACCES;
	if (!ptrace_may_access(task, PTRACE_MODE_READ))
		goto errout;

3125 3126 3127 3128 3129 3130 3131
	return task;
errout:
	put_task_struct(task);
	return ERR_PTR(err);

}

3132 3133 3134
/*
 * Returns a matching context with refcount and pincount.
 */
P
Peter Zijlstra 已提交
3135
static struct perf_event_context *
M
Matt Helsley 已提交
3136
find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
T
Thomas Gleixner 已提交
3137
{
3138
	struct perf_event_context *ctx;
3139
	struct perf_cpu_context *cpuctx;
3140
	unsigned long flags;
P
Peter Zijlstra 已提交
3141
	int ctxn, err;
T
Thomas Gleixner 已提交
3142

3143
	if (!task) {
3144
		/* Must be root to operate on a CPU event: */
3145
		if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
T
Thomas Gleixner 已提交
3146 3147 3148
			return ERR_PTR(-EACCES);

		/*
3149
		 * We could be clever and allow to attach a event to an
T
Thomas Gleixner 已提交
3150 3151 3152
		 * offline CPU and activate it when the CPU comes up, but
		 * that's for later.
		 */
3153
		if (!cpu_online(cpu))
T
Thomas Gleixner 已提交
3154 3155
			return ERR_PTR(-ENODEV);

P
Peter Zijlstra 已提交
3156
		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
T
Thomas Gleixner 已提交
3157
		ctx = &cpuctx->ctx;
3158
		get_ctx(ctx);
3159
		++ctx->pin_count;
T
Thomas Gleixner 已提交
3160 3161 3162 3163

		return ctx;
	}

P
Peter Zijlstra 已提交
3164 3165 3166 3167 3168
	err = -EINVAL;
	ctxn = pmu->task_ctx_nr;
	if (ctxn < 0)
		goto errout;

P
Peter Zijlstra 已提交
3169
retry:
P
Peter Zijlstra 已提交
3170
	ctx = perf_lock_task_context(task, ctxn, &flags);
3171
	if (ctx) {
3172
		unclone_ctx(ctx);
3173
		++ctx->pin_count;
3174
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
3175
	} else {
3176
		ctx = alloc_perf_context(pmu, task);
3177 3178 3179
		err = -ENOMEM;
		if (!ctx)
			goto errout;
3180

3181 3182 3183 3184 3185 3186 3187 3188 3189 3190
		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;
3191
		else {
3192
			get_ctx(ctx);
3193
			++ctx->pin_count;
3194
			rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3195
		}
3196 3197 3198
		mutex_unlock(&task->perf_event_mutex);

		if (unlikely(err)) {
3199
			put_ctx(ctx);
3200 3201 3202 3203

			if (err == -EAGAIN)
				goto retry;
			goto errout;
3204 3205 3206
		}
	}

T
Thomas Gleixner 已提交
3207
	return ctx;
3208

P
Peter Zijlstra 已提交
3209
errout:
3210
	return ERR_PTR(err);
T
Thomas Gleixner 已提交
3211 3212
}

L
Li Zefan 已提交
3213 3214
static void perf_event_free_filter(struct perf_event *event);

3215
static void free_event_rcu(struct rcu_head *head)
P
Peter Zijlstra 已提交
3216
{
3217
	struct perf_event *event;
P
Peter Zijlstra 已提交
3218

3219 3220 3221
	event = container_of(head, struct perf_event, rcu_head);
	if (event->ns)
		put_pid_ns(event->ns);
L
Li Zefan 已提交
3222
	perf_event_free_filter(event);
3223
	kfree(event);
P
Peter Zijlstra 已提交
3224 3225
}

3226
static void ring_buffer_put(struct ring_buffer *rb);
3227 3228
static void ring_buffer_attach(struct perf_event *event,
			       struct ring_buffer *rb);
3229

3230
static void unaccount_event_cpu(struct perf_event *event, int cpu)
3231
{
3232 3233 3234 3235 3236 3237 3238 3239 3240 3241
	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));
}
3242

3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255
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);
3256 3257
	if (event->attr.freq)
		atomic_dec(&nr_freq_events);
3258 3259 3260 3261 3262 3263 3264
	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);
}
3265

3266 3267
static void __free_event(struct perf_event *event)
{
3268
	if (!event->parent) {
3269 3270
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
			put_callchain_buffers();
3271
	}
3272

3273 3274 3275 3276 3277 3278
	if (event->destroy)
		event->destroy(event);

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

3279 3280 3281
	if (event->pmu)
		module_put(event->pmu->module);

3282 3283
	call_rcu(&event->rcu_head, free_event_rcu);
}
P
Peter Zijlstra 已提交
3284 3285

static void _free_event(struct perf_event *event)
3286
{
3287
	irq_work_sync(&event->pending);
3288

3289
	unaccount_event(event);
3290

3291
	if (event->rb) {
3292 3293 3294 3295 3296 3297 3298
		/*
		 * 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);
3299
		ring_buffer_attach(event, NULL);
3300
		mutex_unlock(&event->mmap_mutex);
3301 3302
	}

S
Stephane Eranian 已提交
3303 3304 3305
	if (is_cgroup_event(event))
		perf_detach_cgroup(event);

3306
	__free_event(event);
3307 3308
}

P
Peter Zijlstra 已提交
3309 3310 3311 3312 3313
/*
 * Used to free events which have a known refcount of 1, such as in error paths
 * where the event isn't exposed yet and inherited events.
 */
static void free_event(struct perf_event *event)
T
Thomas Gleixner 已提交
3314
{
P
Peter Zijlstra 已提交
3315 3316 3317 3318 3319 3320
	if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
				"unexpected event refcount: %ld; ptr=%p\n",
				atomic_long_read(&event->refcount), event)) {
		/* leak to avoid use-after-free */
		return;
	}
T
Thomas Gleixner 已提交
3321

P
Peter Zijlstra 已提交
3322
	_free_event(event);
T
Thomas Gleixner 已提交
3323 3324
}

3325 3326 3327
/*
 * Called when the last reference to the file is gone.
 */
3328
static void put_event(struct perf_event *event)
3329
{
P
Peter Zijlstra 已提交
3330
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
3331
	struct task_struct *owner;
3332

3333 3334
	if (!atomic_long_dec_and_test(&event->refcount))
		return;
3335

P
Peter Zijlstra 已提交
3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368
	rcu_read_lock();
	owner = ACCESS_ONCE(event->owner);
	/*
	 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
	 * !owner it means the list deletion is complete and we can indeed
	 * free this event, otherwise we need to serialize on
	 * owner->perf_event_mutex.
	 */
	smp_read_barrier_depends();
	if (owner) {
		/*
		 * Since delayed_put_task_struct() also drops the last
		 * task reference we can safely take a new reference
		 * while holding the rcu_read_lock().
		 */
		get_task_struct(owner);
	}
	rcu_read_unlock();

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

P
Peter Zijlstra 已提交
3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386
	WARN_ON_ONCE(ctx->parent_ctx);
	/*
	 * There are two ways this annotation is useful:
	 *
	 *  1) there is a lock recursion from perf_event_exit_task
	 *     see the comment there.
	 *
	 *  2) there is a lock-inversion with mmap_sem through
	 *     perf_event_read_group(), which takes faults while
	 *     holding ctx->mutex, however this is called after
	 *     the last filedesc died, so there is no possibility
	 *     to trigger the AB-BA case.
	 */
	mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
	perf_remove_from_context(event, true);
	mutex_unlock(&ctx->mutex);

	_free_event(event);
3387 3388
}

P
Peter Zijlstra 已提交
3389 3390 3391 3392 3393 3394 3395
int perf_event_release_kernel(struct perf_event *event)
{
	put_event(event);
	return 0;
}
EXPORT_SYMBOL_GPL(perf_event_release_kernel);

3396 3397 3398 3399
static int perf_release(struct inode *inode, struct file *file)
{
	put_event(file->private_data);
	return 0;
3400 3401
}

3402
u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3403
{
3404
	struct perf_event *child;
3405 3406
	u64 total = 0;

3407 3408 3409
	*enabled = 0;
	*running = 0;

3410
	mutex_lock(&event->child_mutex);
3411
	total += perf_event_read(event);
3412 3413 3414 3415 3416 3417
	*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) {
3418
		total += perf_event_read(child);
3419 3420 3421
		*enabled += child->total_time_enabled;
		*running += child->total_time_running;
	}
3422
	mutex_unlock(&event->child_mutex);
3423 3424 3425

	return total;
}
3426
EXPORT_SYMBOL_GPL(perf_event_read_value);
3427

3428
static int perf_event_read_group(struct perf_event *event,
3429 3430
				   u64 read_format, char __user *buf)
{
3431
	struct perf_event *leader = event->group_leader, *sub;
3432 3433
	int n = 0, size = 0, ret = -EFAULT;
	struct perf_event_context *ctx = leader->ctx;
3434
	u64 values[5];
3435
	u64 count, enabled, running;
3436

3437
	mutex_lock(&ctx->mutex);
3438
	count = perf_event_read_value(leader, &enabled, &running);
3439 3440

	values[n++] = 1 + leader->nr_siblings;
3441 3442 3443 3444
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
		values[n++] = enabled;
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
		values[n++] = running;
3445 3446 3447
	values[n++] = count;
	if (read_format & PERF_FORMAT_ID)
		values[n++] = primary_event_id(leader);
3448 3449 3450 3451

	size = n * sizeof(u64);

	if (copy_to_user(buf, values, size))
3452
		goto unlock;
3453

3454
	ret = size;
3455

3456
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3457
		n = 0;
3458

3459
		values[n++] = perf_event_read_value(sub, &enabled, &running);
3460 3461 3462 3463 3464
		if (read_format & PERF_FORMAT_ID)
			values[n++] = primary_event_id(sub);

		size = n * sizeof(u64);

3465
		if (copy_to_user(buf + ret, values, size)) {
3466 3467 3468
			ret = -EFAULT;
			goto unlock;
		}
3469 3470

		ret += size;
3471
	}
3472 3473
unlock:
	mutex_unlock(&ctx->mutex);
3474

3475
	return ret;
3476 3477
}

3478
static int perf_event_read_one(struct perf_event *event,
3479 3480
				 u64 read_format, char __user *buf)
{
3481
	u64 enabled, running;
3482 3483 3484
	u64 values[4];
	int n = 0;

3485 3486 3487 3488 3489
	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;
3490
	if (read_format & PERF_FORMAT_ID)
3491
		values[n++] = primary_event_id(event);
3492 3493 3494 3495 3496 3497 3498

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

	return n * sizeof(u64);
}

T
Thomas Gleixner 已提交
3499
/*
3500
 * Read the performance event - simple non blocking version for now
T
Thomas Gleixner 已提交
3501 3502
 */
static ssize_t
3503
perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
T
Thomas Gleixner 已提交
3504
{
3505
	u64 read_format = event->attr.read_format;
3506
	int ret;
T
Thomas Gleixner 已提交
3507

3508
	/*
3509
	 * Return end-of-file for a read on a event that is in
3510 3511 3512
	 * error state (i.e. because it was pinned but it couldn't be
	 * scheduled on to the CPU at some point).
	 */
3513
	if (event->state == PERF_EVENT_STATE_ERROR)
3514 3515
		return 0;

3516
	if (count < event->read_size)
3517 3518
		return -ENOSPC;

3519
	WARN_ON_ONCE(event->ctx->parent_ctx);
3520
	if (read_format & PERF_FORMAT_GROUP)
3521
		ret = perf_event_read_group(event, read_format, buf);
3522
	else
3523
		ret = perf_event_read_one(event, read_format, buf);
T
Thomas Gleixner 已提交
3524

3525
	return ret;
T
Thomas Gleixner 已提交
3526 3527 3528 3529 3530
}

static ssize_t
perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
3531
	struct perf_event *event = file->private_data;
T
Thomas Gleixner 已提交
3532

3533
	return perf_read_hw(event, buf, count);
T
Thomas Gleixner 已提交
3534 3535 3536 3537
}

static unsigned int perf_poll(struct file *file, poll_table *wait)
{
3538
	struct perf_event *event = file->private_data;
3539
	struct ring_buffer *rb;
3540
	unsigned int events = POLL_HUP;
P
Peter Zijlstra 已提交
3541

3542
	/*
3543 3544
	 * Pin the event->rb by taking event->mmap_mutex; otherwise
	 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3545 3546
	 */
	mutex_lock(&event->mmap_mutex);
3547 3548
	rb = event->rb;
	if (rb)
3549
		events = atomic_xchg(&rb->poll, 0);
3550 3551
	mutex_unlock(&event->mmap_mutex);

3552
	poll_wait(file, &event->waitq, wait);
T
Thomas Gleixner 已提交
3553 3554 3555 3556

	return events;
}

3557
static void perf_event_reset(struct perf_event *event)
3558
{
3559
	(void)perf_event_read(event);
3560
	local64_set(&event->count, 0);
3561
	perf_event_update_userpage(event);
P
Peter Zijlstra 已提交
3562 3563
}

3564
/*
3565 3566 3567 3568
 * 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.
3569
 */
3570 3571
static void perf_event_for_each_child(struct perf_event *event,
					void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3572
{
3573
	struct perf_event *child;
P
Peter Zijlstra 已提交
3574

3575 3576 3577 3578
	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
3579
		func(child);
3580
	mutex_unlock(&event->child_mutex);
P
Peter Zijlstra 已提交
3581 3582
}

3583 3584
static void perf_event_for_each(struct perf_event *event,
				  void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3585
{
3586 3587
	struct perf_event_context *ctx = event->ctx;
	struct perf_event *sibling;
P
Peter Zijlstra 已提交
3588

3589 3590
	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
3591
	event = event->group_leader;
3592

3593 3594
	perf_event_for_each_child(event, func);
	list_for_each_entry(sibling, &event->sibling_list, group_entry)
3595
		perf_event_for_each_child(sibling, func);
3596
	mutex_unlock(&ctx->mutex);
3597 3598
}

3599
static int perf_event_period(struct perf_event *event, u64 __user *arg)
3600
{
3601
	struct perf_event_context *ctx = event->ctx;
3602
	int ret = 0, active;
3603 3604
	u64 value;

3605
	if (!is_sampling_event(event))
3606 3607
		return -EINVAL;

3608
	if (copy_from_user(&value, arg, sizeof(value)))
3609 3610 3611 3612 3613
		return -EFAULT;

	if (!value)
		return -EINVAL;

3614
	raw_spin_lock_irq(&ctx->lock);
3615 3616
	if (event->attr.freq) {
		if (value > sysctl_perf_event_sample_rate) {
3617 3618 3619 3620
			ret = -EINVAL;
			goto unlock;
		}

3621
		event->attr.sample_freq = value;
3622
	} else {
3623 3624
		event->attr.sample_period = value;
		event->hw.sample_period = value;
3625
	}
3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639

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

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

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

3640
unlock:
3641
	raw_spin_unlock_irq(&ctx->lock);
3642 3643 3644 3645

	return ret;
}

3646 3647
static const struct file_operations perf_fops;

3648
static inline int perf_fget_light(int fd, struct fd *p)
3649
{
3650 3651 3652
	struct fd f = fdget(fd);
	if (!f.file)
		return -EBADF;
3653

3654 3655 3656
	if (f.file->f_op != &perf_fops) {
		fdput(f);
		return -EBADF;
3657
	}
3658 3659
	*p = f;
	return 0;
3660 3661 3662 3663
}

static int perf_event_set_output(struct perf_event *event,
				 struct perf_event *output_event);
L
Li Zefan 已提交
3664
static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3665

3666 3667
static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
3668 3669
	struct perf_event *event = file->private_data;
	void (*func)(struct perf_event *);
P
Peter Zijlstra 已提交
3670
	u32 flags = arg;
3671 3672

	switch (cmd) {
3673 3674
	case PERF_EVENT_IOC_ENABLE:
		func = perf_event_enable;
3675
		break;
3676 3677
	case PERF_EVENT_IOC_DISABLE:
		func = perf_event_disable;
3678
		break;
3679 3680
	case PERF_EVENT_IOC_RESET:
		func = perf_event_reset;
3681
		break;
P
Peter Zijlstra 已提交
3682

3683 3684
	case PERF_EVENT_IOC_REFRESH:
		return perf_event_refresh(event, arg);
3685

3686 3687
	case PERF_EVENT_IOC_PERIOD:
		return perf_event_period(event, (u64 __user *)arg);
3688

3689 3690 3691 3692 3693 3694 3695 3696 3697
	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;
	}

3698
	case PERF_EVENT_IOC_SET_OUTPUT:
3699 3700 3701
	{
		int ret;
		if (arg != -1) {
3702 3703 3704 3705 3706 3707 3708 3709 3710 3711
			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);
3712 3713 3714
		}
		return ret;
	}
3715

L
Li Zefan 已提交
3716 3717 3718
	case PERF_EVENT_IOC_SET_FILTER:
		return perf_event_set_filter(event, (void __user *)arg);

3719
	default:
P
Peter Zijlstra 已提交
3720
		return -ENOTTY;
3721
	}
P
Peter Zijlstra 已提交
3722 3723

	if (flags & PERF_IOC_FLAG_GROUP)
3724
		perf_event_for_each(event, func);
P
Peter Zijlstra 已提交
3725
	else
3726
		perf_event_for_each_child(event, func);
P
Peter Zijlstra 已提交
3727 3728

	return 0;
3729 3730
}

P
Pawel Moll 已提交
3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750
#ifdef CONFIG_COMPAT
static long perf_compat_ioctl(struct file *file, unsigned int cmd,
				unsigned long arg)
{
	switch (_IOC_NR(cmd)) {
	case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
	case _IOC_NR(PERF_EVENT_IOC_ID):
		/* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
		if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
			cmd &= ~IOCSIZE_MASK;
			cmd |= sizeof(void *) << IOCSIZE_SHIFT;
		}
		break;
	}
	return perf_ioctl(file, cmd, arg);
}
#else
# define perf_compat_ioctl NULL
#endif

3751
int perf_event_task_enable(void)
3752
{
3753
	struct perf_event *event;
3754

3755 3756 3757 3758
	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);
3759 3760 3761 3762

	return 0;
}

3763
int perf_event_task_disable(void)
3764
{
3765
	struct perf_event *event;
3766

3767 3768 3769 3770
	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);
3771 3772 3773 3774

	return 0;
}

3775
static int perf_event_index(struct perf_event *event)
3776
{
P
Peter Zijlstra 已提交
3777 3778 3779
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;

3780
	if (event->state != PERF_EVENT_STATE_ACTIVE)
3781 3782
		return 0;

3783
	return event->pmu->event_idx(event);
3784 3785
}

3786
static void calc_timer_values(struct perf_event *event,
3787
				u64 *now,
3788 3789
				u64 *enabled,
				u64 *running)
3790
{
3791
	u64 ctx_time;
3792

3793 3794
	*now = perf_clock();
	ctx_time = event->shadow_ctx_time + *now;
3795 3796 3797 3798
	*enabled = ctx_time - event->tstamp_enabled;
	*running = ctx_time - event->tstamp_running;
}

3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818
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();
}

3819
void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3820 3821 3822
{
}

3823 3824 3825 3826 3827
/*
 * 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.
 */
3828
void perf_event_update_userpage(struct perf_event *event)
3829
{
3830
	struct perf_event_mmap_page *userpg;
3831
	struct ring_buffer *rb;
3832
	u64 enabled, running, now;
3833 3834

	rcu_read_lock();
3835 3836 3837 3838
	rb = rcu_dereference(event->rb);
	if (!rb)
		goto unlock;

3839 3840 3841 3842 3843 3844 3845 3846 3847
	/*
	 * 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
	 */
3848
	calc_timer_values(event, &now, &enabled, &running);
3849

3850
	userpg = rb->user_page;
3851 3852 3853 3854 3855
	/*
	 * Disable preemption so as to not let the corresponding user-space
	 * spin too long if we get preempted.
	 */
	preempt_disable();
3856
	++userpg->lock;
3857
	barrier();
3858
	userpg->index = perf_event_index(event);
P
Peter Zijlstra 已提交
3859
	userpg->offset = perf_event_count(event);
3860
	if (userpg->index)
3861
		userpg->offset -= local64_read(&event->hw.prev_count);
3862

3863
	userpg->time_enabled = enabled +
3864
			atomic64_read(&event->child_total_time_enabled);
3865

3866
	userpg->time_running = running +
3867
			atomic64_read(&event->child_total_time_running);
3868

3869
	arch_perf_update_userpage(userpg, now);
3870

3871
	barrier();
3872
	++userpg->lock;
3873
	preempt_enable();
3874
unlock:
3875
	rcu_read_unlock();
3876 3877
}

3878 3879 3880
static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
{
	struct perf_event *event = vma->vm_file->private_data;
3881
	struct ring_buffer *rb;
3882 3883 3884 3885 3886 3887 3888 3889 3890
	int ret = VM_FAULT_SIGBUS;

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

	rcu_read_lock();
3891 3892
	rb = rcu_dereference(event->rb);
	if (!rb)
3893 3894 3895 3896 3897
		goto unlock;

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

3898
	vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912
	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;
}

3913 3914 3915
static void ring_buffer_attach(struct perf_event *event,
			       struct ring_buffer *rb)
{
3916
	struct ring_buffer *old_rb = NULL;
3917 3918
	unsigned long flags;

3919 3920 3921 3922 3923 3924
	if (event->rb) {
		/*
		 * Should be impossible, we set this when removing
		 * event->rb_entry and wait/clear when adding event->rb_entry.
		 */
		WARN_ON_ONCE(event->rcu_pending);
3925

3926 3927 3928
		old_rb = event->rb;
		event->rcu_batches = get_state_synchronize_rcu();
		event->rcu_pending = 1;
3929

3930 3931 3932 3933
		spin_lock_irqsave(&old_rb->event_lock, flags);
		list_del_rcu(&event->rb_entry);
		spin_unlock_irqrestore(&old_rb->event_lock, flags);
	}
3934

3935 3936 3937 3938
	if (event->rcu_pending && rb) {
		cond_synchronize_rcu(event->rcu_batches);
		event->rcu_pending = 0;
	}
3939

3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956
	if (rb) {
		spin_lock_irqsave(&rb->event_lock, flags);
		list_add_rcu(&event->rb_entry, &rb->event_list);
		spin_unlock_irqrestore(&rb->event_lock, flags);
	}

	rcu_assign_pointer(event->rb, rb);

	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);
	}
3957 3958 3959 3960 3961 3962 3963 3964
}

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

	rcu_read_lock();
	rb = rcu_dereference(event->rb);
3965 3966 3967 3968
	if (rb) {
		list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
			wake_up_all(&event->waitq);
	}
3969 3970 3971
	rcu_read_unlock();
}

3972
static void rb_free_rcu(struct rcu_head *rcu_head)
3973
{
3974
	struct ring_buffer *rb;
3975

3976 3977
	rb = container_of(rcu_head, struct ring_buffer, rcu_head);
	rb_free(rb);
3978 3979
}

3980
static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3981
{
3982
	struct ring_buffer *rb;
3983

3984
	rcu_read_lock();
3985 3986 3987 3988
	rb = rcu_dereference(event->rb);
	if (rb) {
		if (!atomic_inc_not_zero(&rb->refcount))
			rb = NULL;
3989 3990 3991
	}
	rcu_read_unlock();

3992
	return rb;
3993 3994
}

3995
static void ring_buffer_put(struct ring_buffer *rb)
3996
{
3997
	if (!atomic_dec_and_test(&rb->refcount))
3998
		return;
3999

4000
	WARN_ON_ONCE(!list_empty(&rb->event_list));
4001

4002
	call_rcu(&rb->rcu_head, rb_free_rcu);
4003 4004 4005 4006
}

static void perf_mmap_open(struct vm_area_struct *vma)
{
4007
	struct perf_event *event = vma->vm_file->private_data;
4008

4009
	atomic_inc(&event->mmap_count);
4010
	atomic_inc(&event->rb->mmap_count);
4011 4012
}

4013 4014 4015 4016 4017 4018 4019 4020
/*
 * 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.
 */
4021 4022
static void perf_mmap_close(struct vm_area_struct *vma)
{
4023
	struct perf_event *event = vma->vm_file->private_data;
4024

4025
	struct ring_buffer *rb = ring_buffer_get(event);
4026 4027 4028
	struct user_struct *mmap_user = rb->mmap_user;
	int mmap_locked = rb->mmap_locked;
	unsigned long size = perf_data_size(rb);
4029

4030 4031 4032
	atomic_dec(&rb->mmap_count);

	if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4033
		goto out_put;
4034

4035
	ring_buffer_attach(event, NULL);
4036 4037 4038
	mutex_unlock(&event->mmap_mutex);

	/* If there's still other mmap()s of this buffer, we're done. */
4039 4040
	if (atomic_read(&rb->mmap_count))
		goto out_put;
4041

4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057
	/*
	 * 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();
4058

4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069
		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.
		 */
4070 4071 4072
		if (event->rb == rb)
			ring_buffer_attach(event, NULL);

4073
		mutex_unlock(&event->mmap_mutex);
4074
		put_event(event);
4075

4076 4077 4078 4079 4080
		/*
		 * Restart the iteration; either we're on the wrong list or
		 * destroyed its integrity by doing a deletion.
		 */
		goto again;
4081
	}
4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096
	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);

4097
out_put:
4098
	ring_buffer_put(rb); /* could be last */
4099 4100
}

4101
static const struct vm_operations_struct perf_mmap_vmops = {
4102 4103 4104 4105
	.open		= perf_mmap_open,
	.close		= perf_mmap_close,
	.fault		= perf_mmap_fault,
	.page_mkwrite	= perf_mmap_fault,
4106 4107 4108 4109
};

static int perf_mmap(struct file *file, struct vm_area_struct *vma)
{
4110
	struct perf_event *event = file->private_data;
4111
	unsigned long user_locked, user_lock_limit;
4112
	struct user_struct *user = current_user();
4113
	unsigned long locked, lock_limit;
4114
	struct ring_buffer *rb;
4115 4116
	unsigned long vma_size;
	unsigned long nr_pages;
4117
	long user_extra, extra;
4118
	int ret = 0, flags = 0;
4119

4120 4121 4122
	/*
	 * Don't allow mmap() of inherited per-task counters. This would
	 * create a performance issue due to all children writing to the
4123
	 * same rb.
4124 4125 4126 4127
	 */
	if (event->cpu == -1 && event->attr.inherit)
		return -EINVAL;

4128
	if (!(vma->vm_flags & VM_SHARED))
4129
		return -EINVAL;
4130 4131 4132 4133

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

4134
	/*
4135
	 * If we have rb pages ensure they're a power-of-two number, so we
4136 4137 4138
	 * can do bitmasks instead of modulo.
	 */
	if (nr_pages != 0 && !is_power_of_2(nr_pages))
4139 4140
		return -EINVAL;

4141
	if (vma_size != PAGE_SIZE * (1 + nr_pages))
4142 4143
		return -EINVAL;

4144 4145
	if (vma->vm_pgoff != 0)
		return -EINVAL;
4146

4147
	WARN_ON_ONCE(event->ctx->parent_ctx);
4148
again:
4149
	mutex_lock(&event->mmap_mutex);
4150
	if (event->rb) {
4151
		if (event->rb->nr_pages != nr_pages) {
4152
			ret = -EINVAL;
4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165
			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;
		}

4166 4167 4168
		goto unlock;
	}

4169
	user_extra = nr_pages + 1;
4170
	user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
I
Ingo Molnar 已提交
4171 4172 4173 4174 4175 4176

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

4177
	user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4178

4179 4180 4181
	extra = 0;
	if (user_locked > user_lock_limit)
		extra = user_locked - user_lock_limit;
4182

4183
	lock_limit = rlimit(RLIMIT_MEMLOCK);
4184
	lock_limit >>= PAGE_SHIFT;
4185
	locked = vma->vm_mm->pinned_vm + extra;
4186

4187 4188
	if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
		!capable(CAP_IPC_LOCK)) {
4189 4190 4191
		ret = -EPERM;
		goto unlock;
	}
4192

4193
	WARN_ON(event->rb);
4194

4195
	if (vma->vm_flags & VM_WRITE)
4196
		flags |= RING_BUFFER_WRITABLE;
4197

4198 4199 4200 4201
	rb = rb_alloc(nr_pages, 
		event->attr.watermark ? event->attr.wakeup_watermark : 0,
		event->cpu, flags);

4202
	if (!rb) {
4203
		ret = -ENOMEM;
4204
		goto unlock;
4205
	}
P
Peter Zijlstra 已提交
4206

4207
	atomic_set(&rb->mmap_count, 1);
P
Peter Zijlstra 已提交
4208 4209
	rb->mmap_locked = extra;
	rb->mmap_user = get_current_user();
4210

4211
	atomic_long_add(user_extra, &user->locked_vm);
P
Peter Zijlstra 已提交
4212 4213
	vma->vm_mm->pinned_vm += extra;

4214
	ring_buffer_attach(event, rb);
4215

4216
	perf_event_init_userpage(event);
4217 4218
	perf_event_update_userpage(event);

4219
unlock:
4220 4221
	if (!ret)
		atomic_inc(&event->mmap_count);
4222
	mutex_unlock(&event->mmap_mutex);
4223

4224 4225 4226 4227
	/*
	 * Since pinned accounting is per vm we cannot allow fork() to copy our
	 * vma.
	 */
P
Peter Zijlstra 已提交
4228
	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4229
	vma->vm_ops = &perf_mmap_vmops;
4230 4231

	return ret;
4232 4233
}

P
Peter Zijlstra 已提交
4234 4235
static int perf_fasync(int fd, struct file *filp, int on)
{
A
Al Viro 已提交
4236
	struct inode *inode = file_inode(filp);
4237
	struct perf_event *event = filp->private_data;
P
Peter Zijlstra 已提交
4238 4239 4240
	int retval;

	mutex_lock(&inode->i_mutex);
4241
	retval = fasync_helper(fd, filp, on, &event->fasync);
P
Peter Zijlstra 已提交
4242 4243 4244 4245 4246 4247 4248 4249
	mutex_unlock(&inode->i_mutex);

	if (retval < 0)
		return retval;

	return 0;
}

T
Thomas Gleixner 已提交
4250
static const struct file_operations perf_fops = {
4251
	.llseek			= no_llseek,
T
Thomas Gleixner 已提交
4252 4253 4254
	.release		= perf_release,
	.read			= perf_read,
	.poll			= perf_poll,
4255
	.unlocked_ioctl		= perf_ioctl,
P
Pawel Moll 已提交
4256
	.compat_ioctl		= perf_compat_ioctl,
4257
	.mmap			= perf_mmap,
P
Peter Zijlstra 已提交
4258
	.fasync			= perf_fasync,
T
Thomas Gleixner 已提交
4259 4260
};

4261
/*
4262
 * Perf event wakeup
4263 4264 4265 4266 4267
 *
 * If there's data, ensure we set the poll() state and publish everything
 * to user-space before waking everybody up.
 */

4268
void perf_event_wakeup(struct perf_event *event)
4269
{
4270
	ring_buffer_wakeup(event);
4271

4272 4273 4274
	if (event->pending_kill) {
		kill_fasync(&event->fasync, SIGIO, event->pending_kill);
		event->pending_kill = 0;
4275
	}
4276 4277
}

4278
static void perf_pending_event(struct irq_work *entry)
4279
{
4280 4281
	struct perf_event *event = container_of(entry,
			struct perf_event, pending);
4282

4283 4284 4285
	if (event->pending_disable) {
		event->pending_disable = 0;
		__perf_event_disable(event);
4286 4287
	}

4288 4289 4290
	if (event->pending_wakeup) {
		event->pending_wakeup = 0;
		perf_event_wakeup(event);
4291 4292 4293
	}
}

4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314
/*
 * 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);

4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345
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);
	}
}

4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440
/*
 * 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);
	}
}

4441 4442 4443
static void __perf_event_header__init_id(struct perf_event_header *header,
					 struct perf_sample_data *data,
					 struct perf_event *event)
4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458
{
	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();

4459
	if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470
		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;
	}
}

4471 4472 4473
void perf_event_header__init_id(struct perf_event_header *header,
				struct perf_sample_data *data,
				struct perf_event *event)
4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497
{
	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);
4498 4499 4500

	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);
4501 4502
}

4503 4504 4505
void perf_event__output_id_sample(struct perf_event *event,
				  struct perf_output_handle *handle,
				  struct perf_sample_data *sample)
4506 4507 4508 4509 4510
{
	if (event->attr.sample_id_all)
		__perf_event__output_id_sample(handle, sample);
}

4511
static void perf_output_read_one(struct perf_output_handle *handle,
4512 4513
				 struct perf_event *event,
				 u64 enabled, u64 running)
4514
{
4515
	u64 read_format = event->attr.read_format;
4516 4517 4518
	u64 values[4];
	int n = 0;

P
Peter Zijlstra 已提交
4519
	values[n++] = perf_event_count(event);
4520
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4521
		values[n++] = enabled +
4522
			atomic64_read(&event->child_total_time_enabled);
4523 4524
	}
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4525
		values[n++] = running +
4526
			atomic64_read(&event->child_total_time_running);
4527 4528
	}
	if (read_format & PERF_FORMAT_ID)
4529
		values[n++] = primary_event_id(event);
4530

4531
	__output_copy(handle, values, n * sizeof(u64));
4532 4533 4534
}

/*
4535
 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4536 4537
 */
static void perf_output_read_group(struct perf_output_handle *handle,
4538 4539
			    struct perf_event *event,
			    u64 enabled, u64 running)
4540
{
4541 4542
	struct perf_event *leader = event->group_leader, *sub;
	u64 read_format = event->attr.read_format;
4543 4544 4545 4546 4547 4548
	u64 values[5];
	int n = 0;

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

	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4549
		values[n++] = enabled;
4550 4551

	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4552
		values[n++] = running;
4553

4554
	if (leader != event)
4555 4556
		leader->pmu->read(leader);

P
Peter Zijlstra 已提交
4557
	values[n++] = perf_event_count(leader);
4558
	if (read_format & PERF_FORMAT_ID)
4559
		values[n++] = primary_event_id(leader);
4560

4561
	__output_copy(handle, values, n * sizeof(u64));
4562

4563
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4564 4565
		n = 0;

4566 4567
		if ((sub != event) &&
		    (sub->state == PERF_EVENT_STATE_ACTIVE))
4568 4569
			sub->pmu->read(sub);

P
Peter Zijlstra 已提交
4570
		values[n++] = perf_event_count(sub);
4571
		if (read_format & PERF_FORMAT_ID)
4572
			values[n++] = primary_event_id(sub);
4573

4574
		__output_copy(handle, values, n * sizeof(u64));
4575 4576 4577
	}
}

4578 4579 4580
#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
				 PERF_FORMAT_TOTAL_TIME_RUNNING)

4581
static void perf_output_read(struct perf_output_handle *handle,
4582
			     struct perf_event *event)
4583
{
4584
	u64 enabled = 0, running = 0, now;
4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595
	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
	 */
4596
	if (read_format & PERF_FORMAT_TOTAL_TIMES)
4597
		calc_timer_values(event, &now, &enabled, &running);
4598

4599
	if (event->attr.read_format & PERF_FORMAT_GROUP)
4600
		perf_output_read_group(handle, event, enabled, running);
4601
	else
4602
		perf_output_read_one(handle, event, enabled, running);
4603 4604
}

4605 4606 4607
void perf_output_sample(struct perf_output_handle *handle,
			struct perf_event_header *header,
			struct perf_sample_data *data,
4608
			struct perf_event *event)
4609 4610 4611 4612 4613
{
	u64 sample_type = data->type;

	perf_output_put(handle, *header);

4614 4615 4616
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);

4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641
	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)
4642
		perf_output_read(handle, event);
4643 4644 4645 4646 4647 4648 4649 4650 4651 4652

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

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

			size *= sizeof(u64);

4653
			__output_copy(handle, data->callchain, size);
4654 4655 4656 4657 4658 4659 4660 4661 4662
		} 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);
4663 4664
			__output_copy(handle, data->raw->data,
					   data->raw->size);
4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675
		} else {
			struct {
				u32	size;
				u32	data;
			} raw = {
				.size = sizeof(u32),
				.data = 0,
			};
			perf_output_put(handle, raw);
		}
	}
4676

4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693
	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);
		}
	}
4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710

	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);
		}
	}
4711

4712
	if (sample_type & PERF_SAMPLE_STACK_USER) {
4713 4714 4715
		perf_output_sample_ustack(handle,
					  data->stack_user_size,
					  data->regs_user.regs);
4716
	}
A
Andi Kleen 已提交
4717 4718 4719

	if (sample_type & PERF_SAMPLE_WEIGHT)
		perf_output_put(handle, data->weight);
4720 4721 4722

	if (sample_type & PERF_SAMPLE_DATA_SRC)
		perf_output_put(handle, data->data_src.val);
4723

A
Andi Kleen 已提交
4724 4725 4726
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		perf_output_put(handle, data->txn);

4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739
	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);
			}
		}
	}
4740 4741 4742 4743
}

void perf_prepare_sample(struct perf_event_header *header,
			 struct perf_sample_data *data,
4744
			 struct perf_event *event,
4745
			 struct pt_regs *regs)
4746
{
4747
	u64 sample_type = event->attr.sample_type;
4748

4749
	header->type = PERF_RECORD_SAMPLE;
4750
	header->size = sizeof(*header) + event->header_size;
4751 4752 4753

	header->misc = 0;
	header->misc |= perf_misc_flags(regs);
4754

4755
	__perf_event_header__init_id(header, data, event);
4756

4757
	if (sample_type & PERF_SAMPLE_IP)
4758 4759
		data->ip = perf_instruction_pointer(regs);

4760
	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4761
		int size = 1;
4762

4763
		data->callchain = perf_callchain(event, regs);
4764 4765 4766 4767 4768

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

		header->size += size * sizeof(u64);
4769 4770
	}

4771
	if (sample_type & PERF_SAMPLE_RAW) {
4772 4773 4774 4775 4776 4777 4778 4779
		int size = sizeof(u32);

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

		WARN_ON_ONCE(size & (sizeof(u64)-1));
4780
		header->size += size;
4781
	}
4782 4783 4784 4785 4786 4787 4788 4789 4790

	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;
	}
4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804

	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;
	}
4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833

	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;
	}
4834
}
4835

4836
static void perf_event_output(struct perf_event *event,
4837 4838 4839 4840 4841
				struct perf_sample_data *data,
				struct pt_regs *regs)
{
	struct perf_output_handle handle;
	struct perf_event_header header;
4842

4843 4844 4845
	/* protect the callchain buffers */
	rcu_read_lock();

4846
	perf_prepare_sample(&header, data, event, regs);
P
Peter Zijlstra 已提交
4847

4848
	if (perf_output_begin(&handle, event, header.size))
4849
		goto exit;
4850

4851
	perf_output_sample(&handle, &header, data, event);
4852

4853
	perf_output_end(&handle);
4854 4855 4856

exit:
	rcu_read_unlock();
4857 4858
}

4859
/*
4860
 * read event_id
4861 4862 4863 4864 4865 4866 4867 4868 4869 4870
 */

struct perf_read_event {
	struct perf_event_header	header;

	u32				pid;
	u32				tid;
};

static void
4871
perf_event_read_event(struct perf_event *event,
4872 4873 4874
			struct task_struct *task)
{
	struct perf_output_handle handle;
4875
	struct perf_sample_data sample;
4876
	struct perf_read_event read_event = {
4877
		.header = {
4878
			.type = PERF_RECORD_READ,
4879
			.misc = 0,
4880
			.size = sizeof(read_event) + event->read_size,
4881
		},
4882 4883
		.pid = perf_event_pid(event, task),
		.tid = perf_event_tid(event, task),
4884
	};
4885
	int ret;
4886

4887
	perf_event_header__init_id(&read_event.header, &sample, event);
4888
	ret = perf_output_begin(&handle, event, read_event.header.size);
4889 4890 4891
	if (ret)
		return;

4892
	perf_output_put(&handle, read_event);
4893
	perf_output_read(&handle, event);
4894
	perf_event__output_id_sample(event, &handle, &sample);
4895

4896 4897 4898
	perf_output_end(&handle);
}

4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912
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;
4913
		output(event, data);
4914 4915 4916 4917
	}
}

static void
4918
perf_event_aux(perf_event_aux_output_cb output, void *data,
4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930
	       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;
4931
		perf_event_aux_ctx(&cpuctx->ctx, output, data);
4932 4933 4934 4935 4936 4937 4938
		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)
4939
			perf_event_aux_ctx(ctx, output, data);
4940 4941 4942 4943 4944 4945
next:
		put_cpu_ptr(pmu->pmu_cpu_context);
	}

	if (task_ctx) {
		preempt_disable();
4946
		perf_event_aux_ctx(task_ctx, output, data);
4947 4948 4949 4950 4951
		preempt_enable();
	}
	rcu_read_unlock();
}

P
Peter Zijlstra 已提交
4952
/*
P
Peter Zijlstra 已提交
4953 4954
 * task tracking -- fork/exit
 *
4955
 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
P
Peter Zijlstra 已提交
4956 4957
 */

P
Peter Zijlstra 已提交
4958
struct perf_task_event {
4959
	struct task_struct		*task;
4960
	struct perf_event_context	*task_ctx;
P
Peter Zijlstra 已提交
4961 4962 4963 4964 4965 4966

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				ppid;
P
Peter Zijlstra 已提交
4967 4968
		u32				tid;
		u32				ptid;
4969
		u64				time;
4970
	} event_id;
P
Peter Zijlstra 已提交
4971 4972
};

4973 4974
static int perf_event_task_match(struct perf_event *event)
{
4975 4976 4977
	return event->attr.comm  || event->attr.mmap ||
	       event->attr.mmap2 || event->attr.mmap_data ||
	       event->attr.task;
4978 4979
}

4980
static void perf_event_task_output(struct perf_event *event,
4981
				   void *data)
P
Peter Zijlstra 已提交
4982
{
4983
	struct perf_task_event *task_event = data;
P
Peter Zijlstra 已提交
4984
	struct perf_output_handle handle;
4985
	struct perf_sample_data	sample;
P
Peter Zijlstra 已提交
4986
	struct task_struct *task = task_event->task;
4987
	int ret, size = task_event->event_id.header.size;
4988

4989 4990 4991
	if (!perf_event_task_match(event))
		return;

4992
	perf_event_header__init_id(&task_event->event_id.header, &sample, event);
P
Peter Zijlstra 已提交
4993

4994
	ret = perf_output_begin(&handle, event,
4995
				task_event->event_id.header.size);
4996
	if (ret)
4997
		goto out;
P
Peter Zijlstra 已提交
4998

4999 5000
	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
5001

5002 5003
	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
5004

5005
	perf_output_put(&handle, task_event->event_id);
5006

5007 5008
	perf_event__output_id_sample(event, &handle, &sample);

P
Peter Zijlstra 已提交
5009
	perf_output_end(&handle);
5010 5011
out:
	task_event->event_id.header.size = size;
P
Peter Zijlstra 已提交
5012 5013
}

5014 5015
static void perf_event_task(struct task_struct *task,
			      struct perf_event_context *task_ctx,
5016
			      int new)
P
Peter Zijlstra 已提交
5017
{
P
Peter Zijlstra 已提交
5018
	struct perf_task_event task_event;
P
Peter Zijlstra 已提交
5019

5020 5021 5022
	if (!atomic_read(&nr_comm_events) &&
	    !atomic_read(&nr_mmap_events) &&
	    !atomic_read(&nr_task_events))
P
Peter Zijlstra 已提交
5023 5024
		return;

P
Peter Zijlstra 已提交
5025
	task_event = (struct perf_task_event){
5026 5027
		.task	  = task,
		.task_ctx = task_ctx,
5028
		.event_id    = {
P
Peter Zijlstra 已提交
5029
			.header = {
5030
				.type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5031
				.misc = 0,
5032
				.size = sizeof(task_event.event_id),
P
Peter Zijlstra 已提交
5033
			},
5034 5035
			/* .pid  */
			/* .ppid */
P
Peter Zijlstra 已提交
5036 5037
			/* .tid  */
			/* .ptid */
P
Peter Zijlstra 已提交
5038
			.time = perf_clock(),
P
Peter Zijlstra 已提交
5039 5040 5041
		},
	};

5042
	perf_event_aux(perf_event_task_output,
5043 5044
		       &task_event,
		       task_ctx);
P
Peter Zijlstra 已提交
5045 5046
}

5047
void perf_event_fork(struct task_struct *task)
P
Peter Zijlstra 已提交
5048
{
5049
	perf_event_task(task, NULL, 1);
P
Peter Zijlstra 已提交
5050 5051
}

5052 5053 5054 5055 5056
/*
 * comm tracking
 */

struct perf_comm_event {
5057 5058
	struct task_struct	*task;
	char			*comm;
5059 5060 5061 5062 5063 5064 5065
	int			comm_size;

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
5066
	} event_id;
5067 5068
};

5069 5070 5071 5072 5073
static int perf_event_comm_match(struct perf_event *event)
{
	return event->attr.comm;
}

5074
static void perf_event_comm_output(struct perf_event *event,
5075
				   void *data)
5076
{
5077
	struct perf_comm_event *comm_event = data;
5078
	struct perf_output_handle handle;
5079
	struct perf_sample_data sample;
5080
	int size = comm_event->event_id.header.size;
5081 5082
	int ret;

5083 5084 5085
	if (!perf_event_comm_match(event))
		return;

5086 5087
	perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5088
				comm_event->event_id.header.size);
5089 5090

	if (ret)
5091
		goto out;
5092

5093 5094
	comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
	comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5095

5096
	perf_output_put(&handle, comm_event->event_id);
5097
	__output_copy(&handle, comm_event->comm,
5098
				   comm_event->comm_size);
5099 5100 5101

	perf_event__output_id_sample(event, &handle, &sample);

5102
	perf_output_end(&handle);
5103 5104
out:
	comm_event->event_id.header.size = size;
5105 5106
}

5107
static void perf_event_comm_event(struct perf_comm_event *comm_event)
5108
{
5109
	char comm[TASK_COMM_LEN];
5110 5111
	unsigned int size;

5112
	memset(comm, 0, sizeof(comm));
5113
	strlcpy(comm, comm_event->task->comm, sizeof(comm));
5114
	size = ALIGN(strlen(comm)+1, sizeof(u64));
5115 5116 5117 5118

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

5119
	comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
P
Peter Zijlstra 已提交
5120

5121
	perf_event_aux(perf_event_comm_output,
5122 5123
		       comm_event,
		       NULL);
5124 5125
}

5126
void perf_event_comm(struct task_struct *task, bool exec)
5127
{
5128 5129
	struct perf_comm_event comm_event;

5130
	if (!atomic_read(&nr_comm_events))
5131
		return;
5132

5133
	comm_event = (struct perf_comm_event){
5134
		.task	= task,
5135 5136
		/* .comm      */
		/* .comm_size */
5137
		.event_id  = {
5138
			.header = {
5139
				.type = PERF_RECORD_COMM,
5140
				.misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5141 5142 5143 5144
				/* .size */
			},
			/* .pid */
			/* .tid */
5145 5146 5147
		},
	};

5148
	perf_event_comm_event(&comm_event);
5149 5150
}

5151 5152 5153 5154 5155
/*
 * mmap tracking
 */

struct perf_mmap_event {
5156 5157 5158 5159
	struct vm_area_struct	*vma;

	const char		*file_name;
	int			file_size;
5160 5161 5162
	int			maj, min;
	u64			ino;
	u64			ino_generation;
5163
	u32			prot, flags;
5164 5165 5166 5167 5168 5169 5170 5171 5172

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
		u64				start;
		u64				len;
		u64				pgoff;
5173
	} event_id;
5174 5175
};

5176 5177 5178 5179 5180 5181 5182 5183
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) ||
5184
	       (executable && (event->attr.mmap || event->attr.mmap2));
5185 5186
}

5187
static void perf_event_mmap_output(struct perf_event *event,
5188
				   void *data)
5189
{
5190
	struct perf_mmap_event *mmap_event = data;
5191
	struct perf_output_handle handle;
5192
	struct perf_sample_data sample;
5193
	int size = mmap_event->event_id.header.size;
5194
	int ret;
5195

5196 5197 5198
	if (!perf_event_mmap_match(event, data))
		return;

5199 5200 5201 5202 5203
	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);
5204
		mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5205 5206
		mmap_event->event_id.header.size += sizeof(mmap_event->prot);
		mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5207 5208
	}

5209 5210
	perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5211
				mmap_event->event_id.header.size);
5212
	if (ret)
5213
		goto out;
5214

5215 5216
	mmap_event->event_id.pid = perf_event_pid(event, current);
	mmap_event->event_id.tid = perf_event_tid(event, current);
5217

5218
	perf_output_put(&handle, mmap_event->event_id);
5219 5220 5221 5222 5223 5224

	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);
5225 5226
		perf_output_put(&handle, mmap_event->prot);
		perf_output_put(&handle, mmap_event->flags);
5227 5228
	}

5229
	__output_copy(&handle, mmap_event->file_name,
5230
				   mmap_event->file_size);
5231 5232 5233

	perf_event__output_id_sample(event, &handle, &sample);

5234
	perf_output_end(&handle);
5235 5236
out:
	mmap_event->event_id.header.size = size;
5237 5238
}

5239
static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5240
{
5241 5242
	struct vm_area_struct *vma = mmap_event->vma;
	struct file *file = vma->vm_file;
5243 5244
	int maj = 0, min = 0;
	u64 ino = 0, gen = 0;
5245
	u32 prot = 0, flags = 0;
5246 5247 5248
	unsigned int size;
	char tmp[16];
	char *buf = NULL;
5249
	char *name;
5250

5251
	if (file) {
5252 5253
		struct inode *inode;
		dev_t dev;
5254

5255
		buf = kmalloc(PATH_MAX, GFP_KERNEL);
5256
		if (!buf) {
5257 5258
			name = "//enomem";
			goto cpy_name;
5259
		}
5260
		/*
5261
		 * d_path() works from the end of the rb backwards, so we
5262 5263 5264
		 * need to add enough zero bytes after the string to handle
		 * the 64bit alignment we do later.
		 */
5265
		name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5266
		if (IS_ERR(name)) {
5267 5268
			name = "//toolong";
			goto cpy_name;
5269
		}
5270 5271 5272 5273 5274 5275
		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);
5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297

		if (vma->vm_flags & VM_READ)
			prot |= PROT_READ;
		if (vma->vm_flags & VM_WRITE)
			prot |= PROT_WRITE;
		if (vma->vm_flags & VM_EXEC)
			prot |= PROT_EXEC;

		if (vma->vm_flags & VM_MAYSHARE)
			flags = MAP_SHARED;
		else
			flags = MAP_PRIVATE;

		if (vma->vm_flags & VM_DENYWRITE)
			flags |= MAP_DENYWRITE;
		if (vma->vm_flags & VM_MAYEXEC)
			flags |= MAP_EXECUTABLE;
		if (vma->vm_flags & VM_LOCKED)
			flags |= MAP_LOCKED;
		if (vma->vm_flags & VM_HUGETLB)
			flags |= MAP_HUGETLB;

5298
		goto got_name;
5299
	} else {
5300 5301 5302 5303 5304 5305
		if (vma->vm_ops && vma->vm_ops->name) {
			name = (char *) vma->vm_ops->name(vma);
			if (name)
				goto cpy_name;
		}

5306
		name = (char *)arch_vma_name(vma);
5307 5308
		if (name)
			goto cpy_name;
5309

5310
		if (vma->vm_start <= vma->vm_mm->start_brk &&
5311
				vma->vm_end >= vma->vm_mm->brk) {
5312 5313
			name = "[heap]";
			goto cpy_name;
5314 5315
		}
		if (vma->vm_start <= vma->vm_mm->start_stack &&
5316
				vma->vm_end >= vma->vm_mm->start_stack) {
5317 5318
			name = "[stack]";
			goto cpy_name;
5319 5320
		}

5321 5322
		name = "//anon";
		goto cpy_name;
5323 5324
	}

5325 5326 5327
cpy_name:
	strlcpy(tmp, name, sizeof(tmp));
	name = tmp;
5328
got_name:
5329 5330 5331 5332 5333 5334 5335 5336
	/*
	 * 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';
5337 5338 5339

	mmap_event->file_name = name;
	mmap_event->file_size = size;
5340 5341 5342 5343
	mmap_event->maj = maj;
	mmap_event->min = min;
	mmap_event->ino = ino;
	mmap_event->ino_generation = gen;
5344 5345
	mmap_event->prot = prot;
	mmap_event->flags = flags;
5346

5347 5348 5349
	if (!(vma->vm_flags & VM_EXEC))
		mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;

5350
	mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5351

5352
	perf_event_aux(perf_event_mmap_output,
5353 5354
		       mmap_event,
		       NULL);
5355

5356 5357 5358
	kfree(buf);
}

5359
void perf_event_mmap(struct vm_area_struct *vma)
5360
{
5361 5362
	struct perf_mmap_event mmap_event;

5363
	if (!atomic_read(&nr_mmap_events))
5364 5365 5366
		return;

	mmap_event = (struct perf_mmap_event){
5367
		.vma	= vma,
5368 5369
		/* .file_name */
		/* .file_size */
5370
		.event_id  = {
5371
			.header = {
5372
				.type = PERF_RECORD_MMAP,
5373
				.misc = PERF_RECORD_MISC_USER,
5374 5375 5376 5377
				/* .size */
			},
			/* .pid */
			/* .tid */
5378 5379
			.start  = vma->vm_start,
			.len    = vma->vm_end - vma->vm_start,
5380
			.pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
5381
		},
5382 5383 5384 5385
		/* .maj (attr_mmap2 only) */
		/* .min (attr_mmap2 only) */
		/* .ino (attr_mmap2 only) */
		/* .ino_generation (attr_mmap2 only) */
5386 5387
		/* .prot (attr_mmap2 only) */
		/* .flags (attr_mmap2 only) */
5388 5389
	};

5390
	perf_event_mmap_event(&mmap_event);
5391 5392
}

5393 5394 5395 5396
/*
 * IRQ throttle logging
 */

5397
static void perf_log_throttle(struct perf_event *event, int enable)
5398 5399
{
	struct perf_output_handle handle;
5400
	struct perf_sample_data sample;
5401 5402 5403 5404 5405
	int ret;

	struct {
		struct perf_event_header	header;
		u64				time;
5406
		u64				id;
5407
		u64				stream_id;
5408 5409
	} throttle_event = {
		.header = {
5410
			.type = PERF_RECORD_THROTTLE,
5411 5412 5413
			.misc = 0,
			.size = sizeof(throttle_event),
		},
P
Peter Zijlstra 已提交
5414
		.time		= perf_clock(),
5415 5416
		.id		= primary_event_id(event),
		.stream_id	= event->id,
5417 5418
	};

5419
	if (enable)
5420
		throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5421

5422 5423 5424
	perf_event_header__init_id(&throttle_event.header, &sample, event);

	ret = perf_output_begin(&handle, event,
5425
				throttle_event.header.size);
5426 5427 5428 5429
	if (ret)
		return;

	perf_output_put(&handle, throttle_event);
5430
	perf_event__output_id_sample(event, &handle, &sample);
5431 5432 5433
	perf_output_end(&handle);
}

5434
/*
5435
 * Generic event overflow handling, sampling.
5436 5437
 */

5438
static int __perf_event_overflow(struct perf_event *event,
5439 5440
				   int throttle, struct perf_sample_data *data,
				   struct pt_regs *regs)
5441
{
5442 5443
	int events = atomic_read(&event->event_limit);
	struct hw_perf_event *hwc = &event->hw;
5444
	u64 seq;
5445 5446
	int ret = 0;

5447 5448 5449 5450 5451 5452 5453
	/*
	 * Non-sampling counters might still use the PMI to fold short
	 * hardware counters, ignore those.
	 */
	if (unlikely(!is_sampling_event(event)))
		return 0;

5454 5455 5456 5457 5458 5459 5460 5461 5462
	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
5463 5464
			hwc->interrupts = MAX_INTERRUPTS;
			perf_log_throttle(event, 0);
5465
			tick_nohz_full_kick();
5466 5467
			ret = 1;
		}
5468
	}
5469

5470
	if (event->attr.freq) {
P
Peter Zijlstra 已提交
5471
		u64 now = perf_clock();
5472
		s64 delta = now - hwc->freq_time_stamp;
5473

5474
		hwc->freq_time_stamp = now;
5475

5476
		if (delta > 0 && delta < 2*TICK_NSEC)
5477
			perf_adjust_period(event, delta, hwc->last_period, true);
5478 5479
	}

5480 5481
	/*
	 * XXX event_limit might not quite work as expected on inherited
5482
	 * events
5483 5484
	 */

5485 5486
	event->pending_kill = POLL_IN;
	if (events && atomic_dec_and_test(&event->event_limit)) {
5487
		ret = 1;
5488
		event->pending_kill = POLL_HUP;
5489 5490
		event->pending_disable = 1;
		irq_work_queue(&event->pending);
5491 5492
	}

5493
	if (event->overflow_handler)
5494
		event->overflow_handler(event, data, regs);
5495
	else
5496
		perf_event_output(event, data, regs);
5497

P
Peter Zijlstra 已提交
5498
	if (event->fasync && event->pending_kill) {
5499 5500
		event->pending_wakeup = 1;
		irq_work_queue(&event->pending);
P
Peter Zijlstra 已提交
5501 5502
	}

5503
	return ret;
5504 5505
}

5506
int perf_event_overflow(struct perf_event *event,
5507 5508
			  struct perf_sample_data *data,
			  struct pt_regs *regs)
5509
{
5510
	return __perf_event_overflow(event, 1, data, regs);
5511 5512
}

5513
/*
5514
 * Generic software event infrastructure
5515 5516
 */

5517 5518 5519 5520 5521 5522 5523
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];
5524 5525 5526

	/* Keeps track of cpu being initialized/exited */
	bool				online;
5527 5528 5529 5530
};

static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);

5531
/*
5532 5533
 * We directly increment event->count and keep a second value in
 * event->hw.period_left to count intervals. This period event
5534 5535 5536 5537
 * is kept in the range [-sample_period, 0] so that we can use the
 * sign as trigger.
 */

5538
u64 perf_swevent_set_period(struct perf_event *event)
5539
{
5540
	struct hw_perf_event *hwc = &event->hw;
5541 5542 5543 5544 5545
	u64 period = hwc->last_period;
	u64 nr, offset;
	s64 old, val;

	hwc->last_period = hwc->sample_period;
5546 5547

again:
5548
	old = val = local64_read(&hwc->period_left);
5549 5550
	if (val < 0)
		return 0;
5551

5552 5553 5554
	nr = div64_u64(period + val, period);
	offset = nr * period;
	val -= offset;
5555
	if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5556
		goto again;
5557

5558
	return nr;
5559 5560
}

5561
static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5562
				    struct perf_sample_data *data,
5563
				    struct pt_regs *regs)
5564
{
5565
	struct hw_perf_event *hwc = &event->hw;
5566
	int throttle = 0;
5567

5568 5569
	if (!overflow)
		overflow = perf_swevent_set_period(event);
5570

5571 5572
	if (hwc->interrupts == MAX_INTERRUPTS)
		return;
5573

5574
	for (; overflow; overflow--) {
5575
		if (__perf_event_overflow(event, throttle,
5576
					    data, regs)) {
5577 5578 5579 5580 5581 5582
			/*
			 * We inhibit the overflow from happening when
			 * hwc->interrupts == MAX_INTERRUPTS.
			 */
			break;
		}
5583
		throttle = 1;
5584
	}
5585 5586
}

P
Peter Zijlstra 已提交
5587
static void perf_swevent_event(struct perf_event *event, u64 nr,
5588
			       struct perf_sample_data *data,
5589
			       struct pt_regs *regs)
5590
{
5591
	struct hw_perf_event *hwc = &event->hw;
5592

5593
	local64_add(nr, &event->count);
5594

5595 5596 5597
	if (!regs)
		return;

5598
	if (!is_sampling_event(event))
5599
		return;
5600

5601 5602 5603 5604 5605 5606
	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;

5607
	if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5608
		return perf_swevent_overflow(event, 1, data, regs);
5609

5610
	if (local64_add_negative(nr, &hwc->period_left))
5611
		return;
5612

5613
	perf_swevent_overflow(event, 0, data, regs);
5614 5615
}

5616 5617 5618
static int perf_exclude_event(struct perf_event *event,
			      struct pt_regs *regs)
{
P
Peter Zijlstra 已提交
5619
	if (event->hw.state & PERF_HES_STOPPED)
5620
		return 1;
P
Peter Zijlstra 已提交
5621

5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632
	if (regs) {
		if (event->attr.exclude_user && user_mode(regs))
			return 1;

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

	return 0;
}

5633
static int perf_swevent_match(struct perf_event *event,
P
Peter Zijlstra 已提交
5634
				enum perf_type_id type,
L
Li Zefan 已提交
5635 5636 5637
				u32 event_id,
				struct perf_sample_data *data,
				struct pt_regs *regs)
5638
{
5639
	if (event->attr.type != type)
5640
		return 0;
5641

5642
	if (event->attr.config != event_id)
5643 5644
		return 0;

5645 5646
	if (perf_exclude_event(event, regs))
		return 0;
5647 5648 5649 5650

	return 1;
}

5651 5652 5653 5654 5655 5656 5657
static inline u64 swevent_hash(u64 type, u32 event_id)
{
	u64 val = event_id | (type << 32);

	return hash_64(val, SWEVENT_HLIST_BITS);
}

5658 5659
static inline struct hlist_head *
__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5660
{
5661 5662 5663 5664
	u64 hash = swevent_hash(type, event_id);

	return &hlist->heads[hash];
}
5665

5666 5667
/* For the read side: events when they trigger */
static inline struct hlist_head *
5668
find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5669 5670
{
	struct swevent_hlist *hlist;
5671

5672
	hlist = rcu_dereference(swhash->swevent_hlist);
5673 5674 5675
	if (!hlist)
		return NULL;

5676 5677 5678 5679 5680
	return __find_swevent_head(hlist, type, event_id);
}

/* For the event head insertion and removal in the hlist */
static inline struct hlist_head *
5681
find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5682 5683 5684 5685 5686 5687 5688 5689 5690 5691
{
	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.
	 */
5692
	hlist = rcu_dereference_protected(swhash->swevent_hlist,
5693 5694 5695 5696 5697
					  lockdep_is_held(&event->ctx->lock));
	if (!hlist)
		return NULL;

	return __find_swevent_head(hlist, type, event_id);
5698 5699 5700
}

static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5701
				    u64 nr,
5702 5703
				    struct perf_sample_data *data,
				    struct pt_regs *regs)
5704
{
5705
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5706
	struct perf_event *event;
5707
	struct hlist_head *head;
5708

5709
	rcu_read_lock();
5710
	head = find_swevent_head_rcu(swhash, type, event_id);
5711 5712 5713
	if (!head)
		goto end;

5714
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
L
Li Zefan 已提交
5715
		if (perf_swevent_match(event, type, event_id, data, regs))
5716
			perf_swevent_event(event, nr, data, regs);
5717
	}
5718 5719
end:
	rcu_read_unlock();
5720 5721
}

5722
int perf_swevent_get_recursion_context(void)
P
Peter Zijlstra 已提交
5723
{
5724
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
P
Peter Zijlstra 已提交
5725

5726
	return get_recursion_context(swhash->recursion);
P
Peter Zijlstra 已提交
5727
}
I
Ingo Molnar 已提交
5728
EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
P
Peter Zijlstra 已提交
5729

5730
inline void perf_swevent_put_recursion_context(int rctx)
5731
{
5732
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5733

5734
	put_recursion_context(swhash->recursion, rctx);
5735
}
5736

5737
void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5738
{
5739
	struct perf_sample_data data;
5740 5741
	int rctx;

5742
	preempt_disable_notrace();
5743 5744 5745
	rctx = perf_swevent_get_recursion_context();
	if (rctx < 0)
		return;
5746

5747
	perf_sample_data_init(&data, addr, 0);
5748

5749
	do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5750 5751

	perf_swevent_put_recursion_context(rctx);
5752
	preempt_enable_notrace();
5753 5754
}

5755
static void perf_swevent_read(struct perf_event *event)
5756 5757 5758
{
}

P
Peter Zijlstra 已提交
5759
static int perf_swevent_add(struct perf_event *event, int flags)
5760
{
5761
	struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5762
	struct hw_perf_event *hwc = &event->hw;
5763 5764
	struct hlist_head *head;

5765
	if (is_sampling_event(event)) {
5766
		hwc->last_period = hwc->sample_period;
5767
		perf_swevent_set_period(event);
5768
	}
5769

P
Peter Zijlstra 已提交
5770 5771
	hwc->state = !(flags & PERF_EF_START);

5772
	head = find_swevent_head(swhash, event);
5773 5774 5775 5776 5777 5778
	if (!head) {
		/*
		 * We can race with cpu hotplug code. Do not
		 * WARN if the cpu just got unplugged.
		 */
		WARN_ON_ONCE(swhash->online);
5779
		return -EINVAL;
5780
	}
5781 5782 5783

	hlist_add_head_rcu(&event->hlist_entry, head);

5784 5785 5786
	return 0;
}

P
Peter Zijlstra 已提交
5787
static void perf_swevent_del(struct perf_event *event, int flags)
5788
{
5789
	hlist_del_rcu(&event->hlist_entry);
5790 5791
}

P
Peter Zijlstra 已提交
5792
static void perf_swevent_start(struct perf_event *event, int flags)
5793
{
P
Peter Zijlstra 已提交
5794
	event->hw.state = 0;
5795
}
I
Ingo Molnar 已提交
5796

P
Peter Zijlstra 已提交
5797
static void perf_swevent_stop(struct perf_event *event, int flags)
5798
{
P
Peter Zijlstra 已提交
5799
	event->hw.state = PERF_HES_STOPPED;
5800 5801
}

5802 5803
/* Deref the hlist from the update side */
static inline struct swevent_hlist *
5804
swevent_hlist_deref(struct swevent_htable *swhash)
5805
{
5806 5807
	return rcu_dereference_protected(swhash->swevent_hlist,
					 lockdep_is_held(&swhash->hlist_mutex));
5808 5809
}

5810
static void swevent_hlist_release(struct swevent_htable *swhash)
5811
{
5812
	struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5813

5814
	if (!hlist)
5815 5816
		return;

5817
	rcu_assign_pointer(swhash->swevent_hlist, NULL);
5818
	kfree_rcu(hlist, rcu_head);
5819 5820 5821 5822
}

static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
{
5823
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5824

5825
	mutex_lock(&swhash->hlist_mutex);
5826

5827 5828
	if (!--swhash->hlist_refcount)
		swevent_hlist_release(swhash);
5829

5830
	mutex_unlock(&swhash->hlist_mutex);
5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842
}

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

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

static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
{
5843
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5844 5845
	int err = 0;

5846
	mutex_lock(&swhash->hlist_mutex);
5847

5848
	if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5849 5850 5851 5852 5853 5854 5855
		struct swevent_hlist *hlist;

		hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
		if (!hlist) {
			err = -ENOMEM;
			goto exit;
		}
5856
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
5857
	}
5858
	swhash->hlist_refcount++;
P
Peter Zijlstra 已提交
5859
exit:
5860
	mutex_unlock(&swhash->hlist_mutex);
5861 5862 5863 5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880

	return err;
}

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

	get_online_cpus();
	for_each_possible_cpu(cpu) {
		err = swevent_hlist_get_cpu(event, cpu);
		if (err) {
			failed_cpu = cpu;
			goto fail;
		}
	}
	put_online_cpus();

	return 0;
P
Peter Zijlstra 已提交
5881
fail:
5882 5883 5884 5885 5886 5887 5888 5889 5890 5891
	for_each_possible_cpu(cpu) {
		if (cpu == failed_cpu)
			break;
		swevent_hlist_put_cpu(event, cpu);
	}

	put_online_cpus();
	return err;
}

5892
struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5893

5894 5895 5896
static void sw_perf_event_destroy(struct perf_event *event)
{
	u64 event_id = event->attr.config;
5897

5898 5899
	WARN_ON(event->parent);

5900
	static_key_slow_dec(&perf_swevent_enabled[event_id]);
5901 5902 5903 5904 5905
	swevent_hlist_put(event);
}

static int perf_swevent_init(struct perf_event *event)
{
5906
	u64 event_id = event->attr.config;
5907 5908 5909 5910

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

5911 5912 5913 5914 5915 5916
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

5917 5918 5919 5920 5921 5922 5923 5924 5925
	switch (event_id) {
	case PERF_COUNT_SW_CPU_CLOCK:
	case PERF_COUNT_SW_TASK_CLOCK:
		return -ENOENT;

	default:
		break;
	}

5926
	if (event_id >= PERF_COUNT_SW_MAX)
5927 5928 5929 5930 5931 5932 5933 5934 5935
		return -ENOENT;

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

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

5936
		static_key_slow_inc(&perf_swevent_enabled[event_id]);
5937 5938 5939 5940 5941 5942
		event->destroy = sw_perf_event_destroy;
	}

	return 0;
}

5943 5944 5945 5946 5947
static int perf_swevent_event_idx(struct perf_event *event)
{
	return 0;
}

5948
static struct pmu perf_swevent = {
5949
	.task_ctx_nr	= perf_sw_context,
5950

5951
	.event_init	= perf_swevent_init,
P
Peter Zijlstra 已提交
5952 5953 5954 5955
	.add		= perf_swevent_add,
	.del		= perf_swevent_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
5956
	.read		= perf_swevent_read,
5957 5958

	.event_idx	= perf_swevent_event_idx,
5959 5960
};

5961 5962
#ifdef CONFIG_EVENT_TRACING

5963 5964 5965 5966 5967 5968 5969 5970 5971 5972 5973 5974 5975 5976
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)
{
5977 5978
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;
5979 5980 5981 5982
	/*
	 * All tracepoints are from kernel-space.
	 */
	if (event->attr.exclude_kernel)
5983 5984 5985 5986 5987 5988 5989 5990 5991
		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,
5992 5993
		   struct pt_regs *regs, struct hlist_head *head, int rctx,
		   struct task_struct *task)
5994 5995
{
	struct perf_sample_data data;
5996 5997
	struct perf_event *event;

5998 5999 6000 6001 6002
	struct perf_raw_record raw = {
		.size = entry_size,
		.data = record,
	};

6003
	perf_sample_data_init(&data, addr, 0);
6004 6005
	data.raw = &raw;

6006
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
6007
		if (perf_tp_event_match(event, &data, regs))
6008
			perf_swevent_event(event, count, &data, regs);
6009
	}
6010

6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035
	/*
	 * 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();
	}

6036
	perf_swevent_put_recursion_context(rctx);
6037 6038 6039
}
EXPORT_SYMBOL_GPL(perf_tp_event);

6040
static void tp_perf_event_destroy(struct perf_event *event)
6041
{
6042
	perf_trace_destroy(event);
6043 6044
}

6045
static int perf_tp_event_init(struct perf_event *event)
6046
{
6047 6048
	int err;

6049 6050 6051
	if (event->attr.type != PERF_TYPE_TRACEPOINT)
		return -ENOENT;

6052 6053 6054 6055 6056 6057
	/*
	 * no branch sampling for tracepoint events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

6058 6059
	err = perf_trace_init(event);
	if (err)
6060
		return err;
6061

6062
	event->destroy = tp_perf_event_destroy;
6063

6064 6065 6066 6067
	return 0;
}

static struct pmu perf_tracepoint = {
6068 6069
	.task_ctx_nr	= perf_sw_context,

6070
	.event_init	= perf_tp_event_init,
P
Peter Zijlstra 已提交
6071 6072 6073 6074
	.add		= perf_trace_add,
	.del		= perf_trace_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
6075
	.read		= perf_swevent_read,
6076 6077

	.event_idx	= perf_swevent_event_idx,
6078 6079 6080 6081
};

static inline void perf_tp_register(void)
{
P
Peter Zijlstra 已提交
6082
	perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6083
}
L
Li Zefan 已提交
6084 6085 6086 6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107

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);
}

6108
#else
L
Li Zefan 已提交
6109

6110
static inline void perf_tp_register(void)
6111 6112
{
}
L
Li Zefan 已提交
6113 6114 6115 6116 6117 6118 6119 6120 6121 6122

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)
{
}

6123
#endif /* CONFIG_EVENT_TRACING */
6124

6125
#ifdef CONFIG_HAVE_HW_BREAKPOINT
6126
void perf_bp_event(struct perf_event *bp, void *data)
6127
{
6128 6129 6130
	struct perf_sample_data sample;
	struct pt_regs *regs = data;

6131
	perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6132

P
Peter Zijlstra 已提交
6133
	if (!bp->hw.state && !perf_exclude_event(bp, regs))
6134
		perf_swevent_event(bp, 1, &sample, regs);
6135 6136 6137
}
#endif

6138 6139 6140
/*
 * hrtimer based swevent callback
 */
6141

6142
static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6143
{
6144 6145 6146 6147 6148
	enum hrtimer_restart ret = HRTIMER_RESTART;
	struct perf_sample_data data;
	struct pt_regs *regs;
	struct perf_event *event;
	u64 period;
6149

6150
	event = container_of(hrtimer, struct perf_event, hw.hrtimer);
P
Peter Zijlstra 已提交
6151 6152 6153 6154

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

6155
	event->pmu->read(event);
6156

6157
	perf_sample_data_init(&data, 0, event->hw.last_period);
6158 6159 6160
	regs = get_irq_regs();

	if (regs && !perf_exclude_event(event, regs)) {
6161
		if (!(event->attr.exclude_idle && is_idle_task(current)))
6162
			if (__perf_event_overflow(event, 1, &data, regs))
6163 6164
				ret = HRTIMER_NORESTART;
	}
6165

6166 6167
	period = max_t(u64, 10000, event->hw.sample_period);
	hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6168

6169
	return ret;
6170 6171
}

6172
static void perf_swevent_start_hrtimer(struct perf_event *event)
6173
{
6174
	struct hw_perf_event *hwc = &event->hw;
6175 6176 6177 6178
	s64 period;

	if (!is_sampling_event(event))
		return;
6179

6180 6181 6182 6183
	period = local64_read(&hwc->period_left);
	if (period) {
		if (period < 0)
			period = 10000;
P
Peter Zijlstra 已提交
6184

6185 6186 6187 6188 6189
		local64_set(&hwc->period_left, 0);
	} else {
		period = max_t(u64, 10000, hwc->sample_period);
	}
	__hrtimer_start_range_ns(&hwc->hrtimer,
6190
				ns_to_ktime(period), 0,
6191
				HRTIMER_MODE_REL_PINNED, 0);
6192
}
6193 6194

static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6195
{
6196 6197
	struct hw_perf_event *hwc = &event->hw;

6198
	if (is_sampling_event(event)) {
6199
		ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
P
Peter Zijlstra 已提交
6200
		local64_set(&hwc->period_left, ktime_to_ns(remaining));
6201 6202 6203

		hrtimer_cancel(&hwc->hrtimer);
	}
6204 6205
}

P
Peter Zijlstra 已提交
6206 6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225
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);
6226
		hwc->last_period = hwc->sample_period;
P
Peter Zijlstra 已提交
6227 6228 6229 6230
		event->attr.freq = 0;
	}
}

6231 6232 6233 6234 6235
/*
 * Software event: cpu wall time clock
 */

static void cpu_clock_event_update(struct perf_event *event)
6236
{
6237 6238 6239
	s64 prev;
	u64 now;

P
Peter Zijlstra 已提交
6240
	now = local_clock();
6241 6242
	prev = local64_xchg(&event->hw.prev_count, now);
	local64_add(now - prev, &event->count);
6243 6244
}

P
Peter Zijlstra 已提交
6245
static void cpu_clock_event_start(struct perf_event *event, int flags)
6246
{
P
Peter Zijlstra 已提交
6247
	local64_set(&event->hw.prev_count, local_clock());
6248 6249 6250
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6251
static void cpu_clock_event_stop(struct perf_event *event, int flags)
6252
{
6253 6254 6255
	perf_swevent_cancel_hrtimer(event);
	cpu_clock_event_update(event);
}
6256

P
Peter Zijlstra 已提交
6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268 6269
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);
}

6270 6271 6272 6273
static void cpu_clock_event_read(struct perf_event *event)
{
	cpu_clock_event_update(event);
}
6274

6275 6276 6277 6278 6279 6280 6281 6282
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;

6283 6284 6285 6286 6287 6288
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6289 6290
	perf_swevent_init_hrtimer(event);

6291
	return 0;
6292 6293
}

6294
static struct pmu perf_cpu_clock = {
6295 6296
	.task_ctx_nr	= perf_sw_context,

6297
	.event_init	= cpu_clock_event_init,
P
Peter Zijlstra 已提交
6298 6299 6300 6301
	.add		= cpu_clock_event_add,
	.del		= cpu_clock_event_del,
	.start		= cpu_clock_event_start,
	.stop		= cpu_clock_event_stop,
6302
	.read		= cpu_clock_event_read,
6303 6304

	.event_idx	= perf_swevent_event_idx,
6305 6306 6307 6308 6309 6310 6311
};

/*
 * Software event: task time clock
 */

static void task_clock_event_update(struct perf_event *event, u64 now)
6312
{
6313 6314
	u64 prev;
	s64 delta;
6315

6316 6317 6318 6319
	prev = local64_xchg(&event->hw.prev_count, now);
	delta = now - prev;
	local64_add(delta, &event->count);
}
6320

P
Peter Zijlstra 已提交
6321
static void task_clock_event_start(struct perf_event *event, int flags)
6322
{
P
Peter Zijlstra 已提交
6323
	local64_set(&event->hw.prev_count, event->ctx->time);
6324 6325 6326
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6327
static void task_clock_event_stop(struct perf_event *event, int flags)
6328 6329 6330
{
	perf_swevent_cancel_hrtimer(event);
	task_clock_event_update(event, event->ctx->time);
P
Peter Zijlstra 已提交
6331 6332 6333 6334 6335 6336
}

static int task_clock_event_add(struct perf_event *event, int flags)
{
	if (flags & PERF_EF_START)
		task_clock_event_start(event, flags);
6337

P
Peter Zijlstra 已提交
6338 6339 6340 6341 6342 6343
	return 0;
}

static void task_clock_event_del(struct perf_event *event, int flags)
{
	task_clock_event_stop(event, PERF_EF_UPDATE);
6344 6345 6346 6347
}

static void task_clock_event_read(struct perf_event *event)
{
6348 6349 6350
	u64 now = perf_clock();
	u64 delta = now - event->ctx->timestamp;
	u64 time = event->ctx->time + delta;
6351 6352 6353 6354 6355

	task_clock_event_update(event, time);
}

static int task_clock_event_init(struct perf_event *event)
L
Li Zefan 已提交
6356
{
6357 6358 6359 6360 6361 6362
	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

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

6363 6364 6365 6366 6367 6368
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6369 6370
	perf_swevent_init_hrtimer(event);

6371
	return 0;
L
Li Zefan 已提交
6372 6373
}

6374
static struct pmu perf_task_clock = {
6375 6376
	.task_ctx_nr	= perf_sw_context,

6377
	.event_init	= task_clock_event_init,
P
Peter Zijlstra 已提交
6378 6379 6380 6381
	.add		= task_clock_event_add,
	.del		= task_clock_event_del,
	.start		= task_clock_event_start,
	.stop		= task_clock_event_stop,
6382
	.read		= task_clock_event_read,
6383 6384

	.event_idx	= perf_swevent_event_idx,
6385
};
L
Li Zefan 已提交
6386

P
Peter Zijlstra 已提交
6387
static void perf_pmu_nop_void(struct pmu *pmu)
6388 6389
{
}
L
Li Zefan 已提交
6390

P
Peter Zijlstra 已提交
6391
static int perf_pmu_nop_int(struct pmu *pmu)
L
Li Zefan 已提交
6392
{
P
Peter Zijlstra 已提交
6393
	return 0;
L
Li Zefan 已提交
6394 6395
}

P
Peter Zijlstra 已提交
6396
static void perf_pmu_start_txn(struct pmu *pmu)
L
Li Zefan 已提交
6397
{
P
Peter Zijlstra 已提交
6398
	perf_pmu_disable(pmu);
L
Li Zefan 已提交
6399 6400
}

P
Peter Zijlstra 已提交
6401 6402 6403 6404 6405
static int perf_pmu_commit_txn(struct pmu *pmu)
{
	perf_pmu_enable(pmu);
	return 0;
}
6406

P
Peter Zijlstra 已提交
6407
static void perf_pmu_cancel_txn(struct pmu *pmu)
6408
{
P
Peter Zijlstra 已提交
6409
	perf_pmu_enable(pmu);
6410 6411
}

6412 6413 6414 6415 6416
static int perf_event_idx_default(struct perf_event *event)
{
	return event->hw.idx + 1;
}

P
Peter Zijlstra 已提交
6417 6418 6419 6420
/*
 * Ensures all contexts with the same task_ctx_nr have the same
 * pmu_cpu_context too.
 */
6421
static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6422
{
P
Peter Zijlstra 已提交
6423
	struct pmu *pmu;
6424

P
Peter Zijlstra 已提交
6425 6426
	if (ctxn < 0)
		return NULL;
6427

P
Peter Zijlstra 已提交
6428 6429 6430 6431
	list_for_each_entry(pmu, &pmus, entry) {
		if (pmu->task_ctx_nr == ctxn)
			return pmu->pmu_cpu_context;
	}
6432

P
Peter Zijlstra 已提交
6433
	return NULL;
6434 6435
}

6436
static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6437
{
6438 6439 6440 6441 6442 6443 6444
	int cpu;

	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

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

6445 6446
		if (cpuctx->unique_pmu == old_pmu)
			cpuctx->unique_pmu = pmu;
6447 6448 6449 6450 6451 6452
	}
}

static void free_pmu_context(struct pmu *pmu)
{
	struct pmu *i;
6453

P
Peter Zijlstra 已提交
6454
	mutex_lock(&pmus_lock);
6455
	/*
P
Peter Zijlstra 已提交
6456
	 * Like a real lame refcount.
6457
	 */
6458 6459 6460
	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
6461
			goto out;
6462
		}
P
Peter Zijlstra 已提交
6463
	}
6464

6465
	free_percpu(pmu->pmu_cpu_context);
P
Peter Zijlstra 已提交
6466 6467
out:
	mutex_unlock(&pmus_lock);
6468
}
P
Peter Zijlstra 已提交
6469
static struct idr pmu_idr;
6470

P
Peter Zijlstra 已提交
6471 6472 6473 6474 6475 6476 6477
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);
}
6478
static DEVICE_ATTR_RO(type);
P
Peter Zijlstra 已提交
6479

6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522
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;
}
6523
static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6524

6525 6526 6527 6528
static struct attribute *pmu_dev_attrs[] = {
	&dev_attr_type.attr,
	&dev_attr_perf_event_mux_interval_ms.attr,
	NULL,
P
Peter Zijlstra 已提交
6529
};
6530
ATTRIBUTE_GROUPS(pmu_dev);
P
Peter Zijlstra 已提交
6531 6532 6533 6534

static int pmu_bus_running;
static struct bus_type pmu_bus = {
	.name		= "event_source",
6535
	.dev_groups	= pmu_dev_groups,
P
Peter Zijlstra 已提交
6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550
};

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;

6551
	pmu->dev->groups = pmu->attr_groups;
P
Peter Zijlstra 已提交
6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571
	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;
}

6572
static struct lock_class_key cpuctx_mutex;
6573
static struct lock_class_key cpuctx_lock;
6574

6575
int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6576
{
P
Peter Zijlstra 已提交
6577
	int cpu, ret;
6578

6579
	mutex_lock(&pmus_lock);
P
Peter Zijlstra 已提交
6580 6581 6582 6583
	ret = -ENOMEM;
	pmu->pmu_disable_count = alloc_percpu(int);
	if (!pmu->pmu_disable_count)
		goto unlock;
6584

P
Peter Zijlstra 已提交
6585 6586 6587 6588 6589 6590
	pmu->type = -1;
	if (!name)
		goto skip_type;
	pmu->name = name;

	if (type < 0) {
T
Tejun Heo 已提交
6591 6592 6593
		type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
		if (type < 0) {
			ret = type;
P
Peter Zijlstra 已提交
6594 6595 6596 6597 6598
			goto free_pdc;
		}
	}
	pmu->type = type;

P
Peter Zijlstra 已提交
6599 6600 6601 6602 6603 6604
	if (pmu_bus_running) {
		ret = pmu_dev_alloc(pmu);
		if (ret)
			goto free_idr;
	}

P
Peter Zijlstra 已提交
6605
skip_type:
P
Peter Zijlstra 已提交
6606 6607 6608
	pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
	if (pmu->pmu_cpu_context)
		goto got_cpu_context;
6609

W
Wei Yongjun 已提交
6610
	ret = -ENOMEM;
P
Peter Zijlstra 已提交
6611 6612
	pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
	if (!pmu->pmu_cpu_context)
P
Peter Zijlstra 已提交
6613
		goto free_dev;
6614

P
Peter Zijlstra 已提交
6615 6616 6617 6618
	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6619
		__perf_event_init_context(&cpuctx->ctx);
6620
		lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6621
		lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6622
		cpuctx->ctx.type = cpu_context;
P
Peter Zijlstra 已提交
6623
		cpuctx->ctx.pmu = pmu;
6624 6625 6626

		__perf_cpu_hrtimer_init(cpuctx, cpu);

6627
		INIT_LIST_HEAD(&cpuctx->rotation_list);
6628
		cpuctx->unique_pmu = pmu;
P
Peter Zijlstra 已提交
6629
	}
6630

P
Peter Zijlstra 已提交
6631
got_cpu_context:
P
Peter Zijlstra 已提交
6632 6633 6634 6635 6636 6637 6638 6639 6640 6641 6642 6643 6644 6645
	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;
6646
		}
6647
	}
6648

P
Peter Zijlstra 已提交
6649 6650 6651 6652 6653
	if (!pmu->pmu_enable) {
		pmu->pmu_enable  = perf_pmu_nop_void;
		pmu->pmu_disable = perf_pmu_nop_void;
	}

6654 6655 6656
	if (!pmu->event_idx)
		pmu->event_idx = perf_event_idx_default;

6657
	list_add_rcu(&pmu->entry, &pmus);
P
Peter Zijlstra 已提交
6658 6659
	ret = 0;
unlock:
6660 6661
	mutex_unlock(&pmus_lock);

P
Peter Zijlstra 已提交
6662
	return ret;
P
Peter Zijlstra 已提交
6663

P
Peter Zijlstra 已提交
6664 6665 6666 6667
free_dev:
	device_del(pmu->dev);
	put_device(pmu->dev);

P
Peter Zijlstra 已提交
6668 6669 6670 6671
free_idr:
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);

P
Peter Zijlstra 已提交
6672 6673 6674
free_pdc:
	free_percpu(pmu->pmu_disable_count);
	goto unlock;
6675
}
6676
EXPORT_SYMBOL_GPL(perf_pmu_register);
6677

6678
void perf_pmu_unregister(struct pmu *pmu)
6679
{
6680 6681 6682
	mutex_lock(&pmus_lock);
	list_del_rcu(&pmu->entry);
	mutex_unlock(&pmus_lock);
6683

6684
	/*
P
Peter Zijlstra 已提交
6685 6686
	 * We dereference the pmu list under both SRCU and regular RCU, so
	 * synchronize against both of those.
6687
	 */
6688
	synchronize_srcu(&pmus_srcu);
P
Peter Zijlstra 已提交
6689
	synchronize_rcu();
6690

P
Peter Zijlstra 已提交
6691
	free_percpu(pmu->pmu_disable_count);
P
Peter Zijlstra 已提交
6692 6693
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);
P
Peter Zijlstra 已提交
6694 6695
	device_del(pmu->dev);
	put_device(pmu->dev);
6696
	free_pmu_context(pmu);
6697
}
6698
EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6699

6700 6701 6702 6703
struct pmu *perf_init_event(struct perf_event *event)
{
	struct pmu *pmu = NULL;
	int idx;
6704
	int ret;
6705 6706

	idx = srcu_read_lock(&pmus_srcu);
P
Peter Zijlstra 已提交
6707 6708 6709 6710

	rcu_read_lock();
	pmu = idr_find(&pmu_idr, event->attr.type);
	rcu_read_unlock();
6711
	if (pmu) {
6712 6713 6714 6715
		if (!try_module_get(pmu->module)) {
			pmu = ERR_PTR(-ENODEV);
			goto unlock;
		}
6716
		event->pmu = pmu;
6717 6718 6719
		ret = pmu->event_init(event);
		if (ret)
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6720
		goto unlock;
6721
	}
P
Peter Zijlstra 已提交
6722

6723
	list_for_each_entry_rcu(pmu, &pmus, entry) {
6724 6725 6726 6727
		if (!try_module_get(pmu->module)) {
			pmu = ERR_PTR(-ENODEV);
			goto unlock;
		}
6728
		event->pmu = pmu;
6729
		ret = pmu->event_init(event);
6730
		if (!ret)
P
Peter Zijlstra 已提交
6731
			goto unlock;
6732

6733 6734
		if (ret != -ENOENT) {
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6735
			goto unlock;
6736
		}
6737
	}
P
Peter Zijlstra 已提交
6738 6739
	pmu = ERR_PTR(-ENOENT);
unlock:
6740
	srcu_read_unlock(&pmus_srcu, idx);
6741

6742
	return pmu;
6743 6744
}

6745 6746 6747 6748 6749 6750 6751 6752 6753 6754 6755 6756 6757
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));
}

6758 6759
static void account_event(struct perf_event *event)
{
6760 6761 6762
	if (event->parent)
		return;

6763 6764 6765 6766 6767 6768 6769 6770
	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);
6771 6772 6773 6774
	if (event->attr.freq) {
		if (atomic_inc_return(&nr_freq_events) == 1)
			tick_nohz_full_kick_all();
	}
6775
	if (has_branch_stack(event))
6776
		static_key_slow_inc(&perf_sched_events.key);
6777
	if (is_cgroup_event(event))
6778
		static_key_slow_inc(&perf_sched_events.key);
6779 6780

	account_event_cpu(event, event->cpu);
6781 6782
}

T
Thomas Gleixner 已提交
6783
/*
6784
 * Allocate and initialize a event structure
T
Thomas Gleixner 已提交
6785
 */
6786
static struct perf_event *
6787
perf_event_alloc(struct perf_event_attr *attr, int cpu,
6788 6789 6790
		 struct task_struct *task,
		 struct perf_event *group_leader,
		 struct perf_event *parent_event,
6791 6792
		 perf_overflow_handler_t overflow_handler,
		 void *context)
T
Thomas Gleixner 已提交
6793
{
P
Peter Zijlstra 已提交
6794
	struct pmu *pmu;
6795 6796
	struct perf_event *event;
	struct hw_perf_event *hwc;
6797
	long err = -EINVAL;
T
Thomas Gleixner 已提交
6798

6799 6800 6801 6802 6803
	if ((unsigned)cpu >= nr_cpu_ids) {
		if (!task || cpu != -1)
			return ERR_PTR(-EINVAL);
	}

6804
	event = kzalloc(sizeof(*event), GFP_KERNEL);
6805
	if (!event)
6806
		return ERR_PTR(-ENOMEM);
T
Thomas Gleixner 已提交
6807

6808
	/*
6809
	 * Single events are their own group leaders, with an
6810 6811 6812
	 * empty sibling list:
	 */
	if (!group_leader)
6813
		group_leader = event;
6814

6815 6816
	mutex_init(&event->child_mutex);
	INIT_LIST_HEAD(&event->child_list);
6817

6818 6819 6820
	INIT_LIST_HEAD(&event->group_entry);
	INIT_LIST_HEAD(&event->event_entry);
	INIT_LIST_HEAD(&event->sibling_list);
6821
	INIT_LIST_HEAD(&event->rb_entry);
6822
	INIT_LIST_HEAD(&event->active_entry);
6823 6824
	INIT_HLIST_NODE(&event->hlist_entry);

6825

6826
	init_waitqueue_head(&event->waitq);
6827
	init_irq_work(&event->pending, perf_pending_event);
T
Thomas Gleixner 已提交
6828

6829
	mutex_init(&event->mmap_mutex);
6830

6831
	atomic_long_set(&event->refcount, 1);
6832 6833 6834 6835 6836
	event->cpu		= cpu;
	event->attr		= *attr;
	event->group_leader	= group_leader;
	event->pmu		= NULL;
	event->oncpu		= -1;
6837

6838
	event->parent		= parent_event;
6839

6840
	event->ns		= get_pid_ns(task_active_pid_ns(current));
6841
	event->id		= atomic64_inc_return(&perf_event_id);
6842

6843
	event->state		= PERF_EVENT_STATE_INACTIVE;
6844

6845 6846
	if (task) {
		event->attach_state = PERF_ATTACH_TASK;
6847 6848 6849

		if (attr->type == PERF_TYPE_TRACEPOINT)
			event->hw.tp_target = task;
6850 6851 6852 6853
#ifdef CONFIG_HAVE_HW_BREAKPOINT
		/*
		 * hw_breakpoint is a bit difficult here..
		 */
6854
		else if (attr->type == PERF_TYPE_BREAKPOINT)
6855 6856 6857 6858
			event->hw.bp_target = task;
#endif
	}

6859
	if (!overflow_handler && parent_event) {
6860
		overflow_handler = parent_event->overflow_handler;
6861 6862
		context = parent_event->overflow_handler_context;
	}
6863

6864
	event->overflow_handler	= overflow_handler;
6865
	event->overflow_handler_context = context;
6866

J
Jiri Olsa 已提交
6867
	perf_event__state_init(event);
6868

6869
	pmu = NULL;
6870

6871
	hwc = &event->hw;
6872
	hwc->sample_period = attr->sample_period;
6873
	if (attr->freq && attr->sample_freq)
6874
		hwc->sample_period = 1;
6875
	hwc->last_period = hwc->sample_period;
6876

6877
	local64_set(&hwc->period_left, hwc->sample_period);
6878

6879
	/*
6880
	 * we currently do not support PERF_FORMAT_GROUP on inherited events
6881
	 */
6882
	if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6883
		goto err_ns;
6884

6885
	pmu = perf_init_event(event);
6886
	if (!pmu)
6887 6888
		goto err_ns;
	else if (IS_ERR(pmu)) {
6889
		err = PTR_ERR(pmu);
6890
		goto err_ns;
I
Ingo Molnar 已提交
6891
	}
6892

6893
	if (!event->parent) {
6894 6895
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
			err = get_callchain_buffers();
6896 6897
			if (err)
				goto err_pmu;
6898
		}
6899
	}
6900

6901
	return event;
6902 6903 6904 6905

err_pmu:
	if (event->destroy)
		event->destroy(event);
6906
	module_put(pmu->module);
6907 6908 6909 6910 6911 6912
err_ns:
	if (event->ns)
		put_pid_ns(event->ns);
	kfree(event);

	return ERR_PTR(err);
T
Thomas Gleixner 已提交
6913 6914
}

6915 6916
static int perf_copy_attr(struct perf_event_attr __user *uattr,
			  struct perf_event_attr *attr)
6917 6918
{
	u32 size;
6919
	int ret;
6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943

	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,
6944 6945 6946
	 * 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.
6947 6948
	 */
	if (size > sizeof(*attr)) {
6949 6950 6951
		unsigned char __user *addr;
		unsigned char __user *end;
		unsigned char val;
6952

6953 6954
		addr = (void __user *)uattr + sizeof(*attr);
		end  = (void __user *)uattr + size;
6955

6956
		for (; addr < end; addr++) {
6957 6958 6959 6960 6961 6962
			ret = get_user(val, addr);
			if (ret)
				return ret;
			if (val)
				goto err_size;
		}
6963
		size = sizeof(*attr);
6964 6965 6966 6967 6968 6969
	}

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

6970
	if (attr->__reserved_1)
6971 6972 6973 6974 6975 6976 6977 6978
		return -EINVAL;

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

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

6979 6980 6981 6982 6983 6984 6985 6986 6987 6988 6989 6990 6991 6992 6993 6994 6995 6996 6997 6998 6999 7000 7001 7002 7003 7004 7005 7006
	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;
		}
7007 7008
		/* privileged levels capture (kernel, hv): check permissions */
		if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7009 7010
		    && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
7011
	}
7012

7013
	if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7014
		ret = perf_reg_validate(attr->sample_regs_user);
7015 7016 7017 7018 7019 7020 7021 7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032
		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;
	}
7033

7034 7035 7036 7037 7038 7039 7040 7041 7042
out:
	return ret;

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

7043 7044
static int
perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7045
{
7046
	struct ring_buffer *rb = NULL;
7047 7048
	int ret = -EINVAL;

7049
	if (!output_event)
7050 7051
		goto set;

7052 7053
	/* don't allow circular references */
	if (event == output_event)
7054 7055
		goto out;

7056 7057 7058 7059 7060 7061 7062
	/*
	 * Don't allow cross-cpu buffers
	 */
	if (output_event->cpu != event->cpu)
		goto out;

	/*
7063
	 * If its not a per-cpu rb, it must be the same task.
7064 7065 7066 7067
	 */
	if (output_event->cpu == -1 && output_event->ctx != event->ctx)
		goto out;

7068
set:
7069
	mutex_lock(&event->mmap_mutex);
7070 7071 7072
	/* Can't redirect output if we've got an active mmap() */
	if (atomic_read(&event->mmap_count))
		goto unlock;
7073

7074
	if (output_event) {
7075 7076 7077
		/* get the rb we want to redirect to */
		rb = ring_buffer_get(output_event);
		if (!rb)
7078
			goto unlock;
7079 7080
	}

7081
	ring_buffer_attach(event, rb);
7082

7083
	ret = 0;
7084 7085 7086
unlock:
	mutex_unlock(&event->mmap_mutex);

7087 7088 7089 7090
out:
	return ret;
}

T
Thomas Gleixner 已提交
7091
/**
7092
 * sys_perf_event_open - open a performance event, associate it to a task/cpu
I
Ingo Molnar 已提交
7093
 *
7094
 * @attr_uptr:	event_id type attributes for monitoring/sampling
T
Thomas Gleixner 已提交
7095
 * @pid:		target pid
I
Ingo Molnar 已提交
7096
 * @cpu:		target cpu
7097
 * @group_fd:		group leader event fd
T
Thomas Gleixner 已提交
7098
 */
7099 7100
SYSCALL_DEFINE5(perf_event_open,
		struct perf_event_attr __user *, attr_uptr,
7101
		pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
T
Thomas Gleixner 已提交
7102
{
7103 7104
	struct perf_event *group_leader = NULL, *output_event = NULL;
	struct perf_event *event, *sibling;
7105 7106 7107
	struct perf_event_attr attr;
	struct perf_event_context *ctx;
	struct file *event_file = NULL;
7108
	struct fd group = {NULL, 0};
M
Matt Helsley 已提交
7109
	struct task_struct *task = NULL;
7110
	struct pmu *pmu;
7111
	int event_fd;
7112
	int move_group = 0;
7113
	int err;
7114
	int f_flags = O_RDWR;
T
Thomas Gleixner 已提交
7115

7116
	/* for future expandability... */
S
Stephane Eranian 已提交
7117
	if (flags & ~PERF_FLAG_ALL)
7118 7119
		return -EINVAL;

7120 7121 7122
	err = perf_copy_attr(attr_uptr, &attr);
	if (err)
		return err;
7123

7124 7125 7126 7127 7128
	if (!attr.exclude_kernel) {
		if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
	}

7129
	if (attr.freq) {
7130
		if (attr.sample_freq > sysctl_perf_event_sample_rate)
7131
			return -EINVAL;
7132 7133 7134
	} else {
		if (attr.sample_period & (1ULL << 63))
			return -EINVAL;
7135 7136
	}

S
Stephane Eranian 已提交
7137 7138 7139 7140 7141 7142 7143 7144 7145
	/*
	 * 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;

7146 7147 7148 7149
	if (flags & PERF_FLAG_FD_CLOEXEC)
		f_flags |= O_CLOEXEC;

	event_fd = get_unused_fd_flags(f_flags);
7150 7151 7152
	if (event_fd < 0)
		return event_fd;

7153
	if (group_fd != -1) {
7154 7155
		err = perf_fget_light(group_fd, &group);
		if (err)
7156
			goto err_fd;
7157
		group_leader = group.file->private_data;
7158 7159 7160 7161 7162 7163
		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
7164
	if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7165 7166 7167 7168 7169 7170 7171
		task = find_lively_task_by_vpid(pid);
		if (IS_ERR(task)) {
			err = PTR_ERR(task);
			goto err_group_fd;
		}
	}

7172 7173 7174 7175 7176 7177
	if (task && group_leader &&
	    group_leader->attr.inherit != attr.inherit) {
		err = -EINVAL;
		goto err_task;
	}

7178 7179
	get_online_cpus();

7180 7181
	event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
				 NULL, NULL);
7182 7183
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
7184
		goto err_cpus;
7185 7186
	}

S
Stephane Eranian 已提交
7187 7188
	if (flags & PERF_FLAG_PID_CGROUP) {
		err = perf_cgroup_connect(pid, event, &attr, group_leader);
7189 7190
		if (err) {
			__free_event(event);
7191
			goto err_cpus;
7192
		}
S
Stephane Eranian 已提交
7193 7194
	}

7195 7196 7197 7198 7199 7200 7201
	if (is_sampling_event(event)) {
		if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
			err = -ENOTSUPP;
			goto err_alloc;
		}
	}

7202 7203
	account_event(event);

7204 7205 7206 7207 7208
	/*
	 * Special case software events and allow them to be part of
	 * any hardware group.
	 */
	pmu = event->pmu;
7209 7210 7211 7212 7213 7214 7215 7216 7217 7218 7219 7220 7221 7222 7223 7224 7225 7226 7227 7228 7229 7230 7231

	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;
		}
	}
7232 7233 7234 7235

	/*
	 * Get the target context (task or percpu):
	 */
7236
	ctx = find_get_context(pmu, task, event->cpu);
7237 7238
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7239
		goto err_alloc;
7240 7241
	}

7242 7243 7244 7245 7246
	if (task) {
		put_task_struct(task);
		task = NULL;
	}

I
Ingo Molnar 已提交
7247
	/*
7248
	 * Look up the group leader (we will attach this event to it):
7249
	 */
7250
	if (group_leader) {
7251
		err = -EINVAL;
7252 7253

		/*
I
Ingo Molnar 已提交
7254 7255 7256 7257
		 * Do not allow a recursive hierarchy (this new sibling
		 * becoming part of another group-sibling):
		 */
		if (group_leader->group_leader != group_leader)
7258
			goto err_context;
I
Ingo Molnar 已提交
7259 7260 7261
		/*
		 * Do not allow to attach to a group in a different
		 * task or CPU context:
7262
		 */
7263 7264 7265 7266 7267 7268 7269 7270
		if (move_group) {
			if (group_leader->ctx->type != ctx->type)
				goto err_context;
		} else {
			if (group_leader->ctx != ctx)
				goto err_context;
		}

7271 7272 7273
		/*
		 * Only a group leader can be exclusive or pinned
		 */
7274
		if (attr.exclusive || attr.pinned)
7275
			goto err_context;
7276 7277 7278 7279 7280
	}

	if (output_event) {
		err = perf_event_set_output(event, output_event);
		if (err)
7281
			goto err_context;
7282
	}
T
Thomas Gleixner 已提交
7283

7284 7285
	event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
					f_flags);
7286 7287
	if (IS_ERR(event_file)) {
		err = PTR_ERR(event_file);
7288
		goto err_context;
7289
	}
7290

7291 7292 7293 7294
	if (move_group) {
		struct perf_event_context *gctx = group_leader->ctx;

		mutex_lock(&gctx->mutex);
7295
		perf_remove_from_context(group_leader, false);
J
Jiri Olsa 已提交
7296 7297 7298 7299 7300 7301 7302

		/*
		 * 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);
7303 7304
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7305
			perf_remove_from_context(sibling, false);
J
Jiri Olsa 已提交
7306
			perf_event__state_init(sibling);
7307 7308 7309 7310
			put_ctx(gctx);
		}
		mutex_unlock(&gctx->mutex);
		put_ctx(gctx);
7311
	}
7312

7313
	WARN_ON_ONCE(ctx->parent_ctx);
7314
	mutex_lock(&ctx->mutex);
7315 7316

	if (move_group) {
7317
		synchronize_rcu();
7318
		perf_install_in_context(ctx, group_leader, event->cpu);
7319 7320 7321
		get_ctx(ctx);
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7322
			perf_install_in_context(ctx, sibling, event->cpu);
7323 7324 7325 7326
			get_ctx(ctx);
		}
	}

7327
	perf_install_in_context(ctx, event, event->cpu);
7328
	perf_unpin_context(ctx);
7329
	mutex_unlock(&ctx->mutex);
7330

7331 7332
	put_online_cpus();

7333
	event->owner = current;
P
Peter Zijlstra 已提交
7334

7335 7336 7337
	mutex_lock(&current->perf_event_mutex);
	list_add_tail(&event->owner_entry, &current->perf_event_list);
	mutex_unlock(&current->perf_event_mutex);
7338

7339 7340 7341 7342
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(event);
7343
	perf_event__id_header_size(event);
7344

7345 7346 7347 7348 7349 7350
	/*
	 * 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().
	 */
7351
	fdput(group);
7352 7353
	fd_install(event_fd, event_file);
	return event_fd;
T
Thomas Gleixner 已提交
7354

7355
err_context:
7356
	perf_unpin_context(ctx);
7357
	put_ctx(ctx);
7358
err_alloc:
7359
	free_event(event);
7360
err_cpus:
7361
	put_online_cpus();
7362
err_task:
P
Peter Zijlstra 已提交
7363 7364
	if (task)
		put_task_struct(task);
7365
err_group_fd:
7366
	fdput(group);
7367 7368
err_fd:
	put_unused_fd(event_fd);
7369
	return err;
T
Thomas Gleixner 已提交
7370 7371
}

7372 7373 7374 7375 7376
/**
 * 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
7377
 * @task: task to profile (NULL for percpu)
7378 7379 7380
 */
struct perf_event *
perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
M
Matt Helsley 已提交
7381
				 struct task_struct *task,
7382 7383
				 perf_overflow_handler_t overflow_handler,
				 void *context)
7384 7385
{
	struct perf_event_context *ctx;
7386
	struct perf_event *event;
7387
	int err;
7388

7389 7390 7391
	/*
	 * Get the target context (task or percpu):
	 */
7392

7393 7394
	event = perf_event_alloc(attr, cpu, task, NULL, NULL,
				 overflow_handler, context);
7395 7396 7397 7398
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
		goto err;
	}
7399

7400 7401
	account_event(event);

M
Matt Helsley 已提交
7402
	ctx = find_get_context(event->pmu, task, cpu);
7403 7404
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7405
		goto err_free;
7406
	}
7407 7408 7409 7410

	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
	perf_install_in_context(ctx, event, cpu);
7411
	perf_unpin_context(ctx);
7412 7413 7414 7415
	mutex_unlock(&ctx->mutex);

	return event;

7416 7417 7418
err_free:
	free_event(event);
err:
7419
	return ERR_PTR(err);
7420
}
7421
EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7422

7423 7424 7425 7426 7427 7428 7429 7430 7431 7432 7433 7434 7435
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) {
7436
		perf_remove_from_context(event, false);
7437
		unaccount_event_cpu(event, src_cpu);
7438
		put_ctx(src_ctx);
7439
		list_add(&event->migrate_entry, &events);
7440 7441 7442 7443 7444 7445
	}
	mutex_unlock(&src_ctx->mutex);

	synchronize_rcu();

	mutex_lock(&dst_ctx->mutex);
7446 7447
	list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
		list_del(&event->migrate_entry);
7448 7449
		if (event->state >= PERF_EVENT_STATE_OFF)
			event->state = PERF_EVENT_STATE_INACTIVE;
7450
		account_event_cpu(event, dst_cpu);
7451 7452 7453 7454 7455 7456 7457
		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);

7458
static void sync_child_event(struct perf_event *child_event,
7459
			       struct task_struct *child)
7460
{
7461
	struct perf_event *parent_event = child_event->parent;
7462
	u64 child_val;
7463

7464 7465
	if (child_event->attr.inherit_stat)
		perf_event_read_event(child_event, child);
7466

P
Peter Zijlstra 已提交
7467
	child_val = perf_event_count(child_event);
7468 7469 7470 7471

	/*
	 * Add back the child's count to the parent's count:
	 */
7472
	atomic64_add(child_val, &parent_event->child_count);
7473 7474 7475 7476
	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);
7477 7478

	/*
7479
	 * Remove this event from the parent's list
7480
	 */
7481 7482 7483 7484
	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);
7485 7486

	/*
7487
	 * Release the parent event, if this was the last
7488 7489
	 * reference to it.
	 */
7490
	put_event(parent_event);
7491 7492
}

7493
static void
7494 7495
__perf_event_exit_task(struct perf_event *child_event,
			 struct perf_event_context *child_ctx,
7496
			 struct task_struct *child)
7497
{
7498 7499 7500 7501 7502 7503 7504 7505 7506 7507 7508 7509 7510
	/*
	 * Do not destroy the 'original' grouping; because of the context
	 * switch optimization the original events could've ended up in a
	 * random child task.
	 *
	 * If we were to destroy the original group, all group related
	 * operations would cease to function properly after this random
	 * child dies.
	 *
	 * Do destroy all inherited groups, we don't care about those
	 * and being thorough is better.
	 */
	perf_remove_from_context(child_event, !!child_event->parent);
7511

7512
	/*
7513
	 * It can happen that the parent exits first, and has events
7514
	 * that are still around due to the child reference. These
7515
	 * events need to be zapped.
7516
	 */
7517
	if (child_event->parent) {
7518 7519
		sync_child_event(child_event, child);
		free_event(child_event);
7520
	}
7521 7522
}

P
Peter Zijlstra 已提交
7523
static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7524
{
7525
	struct perf_event *child_event, *next;
7526
	struct perf_event_context *child_ctx, *parent_ctx;
7527
	unsigned long flags;
7528

P
Peter Zijlstra 已提交
7529
	if (likely(!child->perf_event_ctxp[ctxn])) {
7530
		perf_event_task(child, NULL, 0);
7531
		return;
P
Peter Zijlstra 已提交
7532
	}
7533

7534
	local_irq_save(flags);
7535 7536 7537 7538 7539 7540
	/*
	 * 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.
	 */
7541
	child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7542 7543 7544

	/*
	 * Take the context lock here so that if find_get_context is
7545
	 * reading child->perf_event_ctxp, we wait until it has
7546 7547
	 * incremented the context's refcount before we do put_ctx below.
	 */
7548
	raw_spin_lock(&child_ctx->lock);
7549
	task_ctx_sched_out(child_ctx);
P
Peter Zijlstra 已提交
7550
	child->perf_event_ctxp[ctxn] = NULL;
7551 7552 7553 7554 7555 7556 7557 7558 7559

	/*
	 * In order to avoid freeing: child_ctx->parent_ctx->task
	 * under perf_event_context::lock, grab another reference.
	 */
	parent_ctx = child_ctx->parent_ctx;
	if (parent_ctx)
		get_ctx(parent_ctx);

7560 7561 7562
	/*
	 * If this context is a clone; unclone it so it can't get
	 * swapped to another process while we're removing all
7563
	 * the events from it.
7564 7565
	 */
	unclone_ctx(child_ctx);
7566
	update_context_time(child_ctx);
7567
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7568

7569 7570 7571 7572 7573 7574 7575
	/*
	 * Now that we no longer hold perf_event_context::lock, drop
	 * our extra child_ctx->parent_ctx reference.
	 */
	if (parent_ctx)
		put_ctx(parent_ctx);

P
Peter Zijlstra 已提交
7576
	/*
7577 7578 7579
	 * 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
7580
	 */
7581
	perf_event_task(child, child_ctx, 0);
7582

7583 7584 7585
	/*
	 * We can recurse on the same lock type through:
	 *
7586 7587
	 *   __perf_event_exit_task()
	 *     sync_child_event()
7588 7589
	 *       put_event()
	 *         mutex_lock(&ctx->mutex)
7590 7591 7592
	 *
	 * But since its the parent context it won't be the same instance.
	 */
7593
	mutex_lock(&child_ctx->mutex);
7594

7595
	list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7596
		__perf_event_exit_task(child_event, child_ctx, child);
7597

7598 7599 7600
	mutex_unlock(&child_ctx->mutex);

	put_ctx(child_ctx);
7601 7602
}

P
Peter Zijlstra 已提交
7603 7604 7605 7606 7607
/*
 * 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
7608
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7609 7610
	int ctxn;

P
Peter Zijlstra 已提交
7611 7612 7613 7614 7615 7616 7617 7618 7619 7620 7621 7622 7623 7624 7625
	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
7626 7627 7628 7629
	for_each_task_context_nr(ctxn)
		perf_event_exit_task_context(child, ctxn);
}

7630 7631 7632 7633 7634 7635 7636 7637 7638 7639 7640 7641
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);

7642
	put_event(parent);
7643

7644
	perf_group_detach(event);
7645 7646 7647 7648
	list_del_event(event, ctx);
	free_event(event);
}

7649 7650
/*
 * free an unexposed, unused context as created by inheritance by
P
Peter Zijlstra 已提交
7651
 * perf_event_init_task below, used by fork() in case of fail.
7652
 */
7653
void perf_event_free_task(struct task_struct *task)
7654
{
P
Peter Zijlstra 已提交
7655
	struct perf_event_context *ctx;
7656
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7657
	int ctxn;
7658

P
Peter Zijlstra 已提交
7659 7660 7661 7662
	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;
7663

P
Peter Zijlstra 已提交
7664
		mutex_lock(&ctx->mutex);
7665
again:
P
Peter Zijlstra 已提交
7666 7667 7668
		list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
				group_entry)
			perf_free_event(event, ctx);
7669

P
Peter Zijlstra 已提交
7670 7671 7672
		list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
				group_entry)
			perf_free_event(event, ctx);
7673

P
Peter Zijlstra 已提交
7674 7675 7676
		if (!list_empty(&ctx->pinned_groups) ||
				!list_empty(&ctx->flexible_groups))
			goto again;
7677

P
Peter Zijlstra 已提交
7678
		mutex_unlock(&ctx->mutex);
7679

P
Peter Zijlstra 已提交
7680 7681
		put_ctx(ctx);
	}
7682 7683
}

7684 7685 7686 7687 7688 7689 7690 7691
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
7692 7693 7694 7695 7696 7697 7698 7699 7700 7701 7702 7703
/*
 * 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;
7704
	unsigned long flags;
P
Peter Zijlstra 已提交
7705 7706 7707 7708 7709 7710 7711 7712 7713 7714 7715 7716

	/*
	 * 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,
7717
					   child,
P
Peter Zijlstra 已提交
7718
					   group_leader, parent_event,
7719
				           NULL, NULL);
P
Peter Zijlstra 已提交
7720 7721
	if (IS_ERR(child_event))
		return child_event;
7722 7723 7724 7725 7726 7727

	if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
		free_event(child_event);
		return NULL;
	}

P
Peter Zijlstra 已提交
7728 7729 7730 7731 7732 7733 7734 7735 7736 7737 7738 7739 7740 7741 7742 7743 7744 7745 7746 7747 7748 7749 7750 7751
	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;
7752 7753
	child_event->overflow_handler_context
		= parent_event->overflow_handler_context;
P
Peter Zijlstra 已提交
7754

7755 7756 7757 7758
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(child_event);
7759
	perf_event__id_header_size(child_event);
7760

P
Peter Zijlstra 已提交
7761 7762 7763
	/*
	 * Link it up in the child's context:
	 */
7764
	raw_spin_lock_irqsave(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7765
	add_event_to_ctx(child_event, child_ctx);
7766
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7767 7768 7769 7770 7771 7772 7773 7774 7775 7776 7777 7778 7779 7780 7781 7782 7783 7784 7785 7786 7787 7788 7789 7790 7791 7792 7793 7794 7795 7796 7797 7798 7799

	/*
	 * 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;
7800 7801 7802 7803 7804
}

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
7805
		   struct task_struct *child, int ctxn,
7806 7807 7808
		   int *inherited_all)
{
	int ret;
P
Peter Zijlstra 已提交
7809
	struct perf_event_context *child_ctx;
7810 7811 7812 7813

	if (!event->attr.inherit) {
		*inherited_all = 0;
		return 0;
7814 7815
	}

7816
	child_ctx = child->perf_event_ctxp[ctxn];
7817 7818 7819 7820 7821 7822 7823
	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.
		 */
7824

7825
		child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7826 7827
		if (!child_ctx)
			return -ENOMEM;
7828

P
Peter Zijlstra 已提交
7829
		child->perf_event_ctxp[ctxn] = child_ctx;
7830 7831 7832 7833 7834 7835 7836 7837 7838
	}

	ret = inherit_group(event, parent, parent_ctx,
			    child, child_ctx);

	if (ret)
		*inherited_all = 0;

	return ret;
7839 7840
}

7841
/*
7842
 * Initialize the perf_event context in task_struct
7843
 */
7844
static int perf_event_init_context(struct task_struct *child, int ctxn)
7845
{
7846
	struct perf_event_context *child_ctx, *parent_ctx;
7847 7848
	struct perf_event_context *cloned_ctx;
	struct perf_event *event;
7849
	struct task_struct *parent = current;
7850
	int inherited_all = 1;
7851
	unsigned long flags;
7852
	int ret = 0;
7853

P
Peter Zijlstra 已提交
7854
	if (likely(!parent->perf_event_ctxp[ctxn]))
7855 7856
		return 0;

7857
	/*
7858 7859
	 * If the parent's context is a clone, pin it so it won't get
	 * swapped under us.
7860
	 */
P
Peter Zijlstra 已提交
7861
	parent_ctx = perf_pin_task_context(parent, ctxn);
7862 7863
	if (!parent_ctx)
		return 0;
7864

7865 7866 7867 7868 7869 7870 7871
	/*
	 * 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.
	 */

7872 7873 7874 7875
	/*
	 * Lock the parent list. No need to lock the child - not PID
	 * hashed yet and not running, so nobody can access it.
	 */
7876
	mutex_lock(&parent_ctx->mutex);
7877 7878 7879 7880 7881

	/*
	 * We dont have to disable NMIs - we are only looking at
	 * the list, not manipulating it:
	 */
7882
	list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
P
Peter Zijlstra 已提交
7883 7884
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7885 7886 7887
		if (ret)
			break;
	}
7888

7889 7890 7891 7892 7893 7894 7895 7896 7897
	/*
	 * 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);

7898
	list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
P
Peter Zijlstra 已提交
7899 7900
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
7901
		if (ret)
7902
			break;
7903 7904
	}

7905 7906 7907
	raw_spin_lock_irqsave(&parent_ctx->lock, flags);
	parent_ctx->rotate_disable = 0;

P
Peter Zijlstra 已提交
7908
	child_ctx = child->perf_event_ctxp[ctxn];
7909

7910
	if (child_ctx && inherited_all) {
7911 7912 7913
		/*
		 * 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
7914 7915 7916
		 *
		 * Note that if the parent is a clone, the holding of
		 * parent_ctx->lock avoids it from being uncloned.
7917
		 */
P
Peter Zijlstra 已提交
7918
		cloned_ctx = parent_ctx->parent_ctx;
7919 7920
		if (cloned_ctx) {
			child_ctx->parent_ctx = cloned_ctx;
7921
			child_ctx->parent_gen = parent_ctx->parent_gen;
7922 7923 7924 7925 7926
		} else {
			child_ctx->parent_ctx = parent_ctx;
			child_ctx->parent_gen = parent_ctx->generation;
		}
		get_ctx(child_ctx->parent_ctx);
7927 7928
	}

P
Peter Zijlstra 已提交
7929
	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7930
	mutex_unlock(&parent_ctx->mutex);
7931

7932
	perf_unpin_context(parent_ctx);
7933
	put_ctx(parent_ctx);
7934

7935
	return ret;
7936 7937
}

P
Peter Zijlstra 已提交
7938 7939 7940 7941 7942 7943 7944
/*
 * Initialize the perf_event context in task_struct
 */
int perf_event_init_task(struct task_struct *child)
{
	int ctxn, ret;

7945 7946 7947 7948
	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
7949 7950 7951 7952 7953 7954 7955 7956 7957
	for_each_task_context_nr(ctxn) {
		ret = perf_event_init_context(child, ctxn);
		if (ret)
			return ret;
	}

	return 0;
}

7958 7959
static void __init perf_event_init_all_cpus(void)
{
7960
	struct swevent_htable *swhash;
7961 7962 7963
	int cpu;

	for_each_possible_cpu(cpu) {
7964 7965
		swhash = &per_cpu(swevent_htable, cpu);
		mutex_init(&swhash->hlist_mutex);
7966
		INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7967 7968 7969
	}
}

7970
static void perf_event_init_cpu(int cpu)
T
Thomas Gleixner 已提交
7971
{
P
Peter Zijlstra 已提交
7972
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
T
Thomas Gleixner 已提交
7973

7974
	mutex_lock(&swhash->hlist_mutex);
7975
	swhash->online = true;
7976
	if (swhash->hlist_refcount > 0) {
7977 7978
		struct swevent_hlist *hlist;

7979 7980 7981
		hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
		WARN_ON(!hlist);
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
7982
	}
7983
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
7984 7985
}

P
Peter Zijlstra 已提交
7986
#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7987
static void perf_pmu_rotate_stop(struct pmu *pmu)
T
Thomas Gleixner 已提交
7988
{
7989 7990 7991 7992 7993 7994 7995
	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
7996
static void __perf_event_exit_context(void *__info)
T
Thomas Gleixner 已提交
7997
{
7998
	struct remove_event re = { .detach_group = false };
P
Peter Zijlstra 已提交
7999
	struct perf_event_context *ctx = __info;
T
Thomas Gleixner 已提交
8000

P
Peter Zijlstra 已提交
8001
	perf_pmu_rotate_stop(ctx->pmu);
8002

P
Peter Zijlstra 已提交
8003
	rcu_read_lock();
8004 8005
	list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
		__perf_remove_from_context(&re);
P
Peter Zijlstra 已提交
8006
	rcu_read_unlock();
T
Thomas Gleixner 已提交
8007
}
P
Peter Zijlstra 已提交
8008 8009 8010 8011 8012 8013 8014 8015 8016

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) {
8017
		ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
P
Peter Zijlstra 已提交
8018 8019 8020 8021 8022 8023 8024 8025

		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);
}

8026
static void perf_event_exit_cpu(int cpu)
T
Thomas Gleixner 已提交
8027
{
8028
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8029

P
Peter Zijlstra 已提交
8030 8031
	perf_event_exit_cpu_context(cpu);

8032
	mutex_lock(&swhash->hlist_mutex);
8033
	swhash->online = false;
8034 8035
	swevent_hlist_release(swhash);
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
8036 8037
}
#else
8038
static inline void perf_event_exit_cpu(int cpu) { }
T
Thomas Gleixner 已提交
8039 8040
#endif

P
Peter Zijlstra 已提交
8041 8042 8043 8044 8045 8046 8047 8048 8049 8050 8051 8052 8053 8054 8055 8056 8057 8058 8059 8060
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,
};

8061
static int
T
Thomas Gleixner 已提交
8062 8063 8064 8065
perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
{
	unsigned int cpu = (long)hcpu;

8066
	switch (action & ~CPU_TASKS_FROZEN) {
T
Thomas Gleixner 已提交
8067 8068

	case CPU_UP_PREPARE:
P
Peter Zijlstra 已提交
8069
	case CPU_DOWN_FAILED:
8070
		perf_event_init_cpu(cpu);
T
Thomas Gleixner 已提交
8071 8072
		break;

P
Peter Zijlstra 已提交
8073
	case CPU_UP_CANCELED:
T
Thomas Gleixner 已提交
8074
	case CPU_DOWN_PREPARE:
8075
		perf_event_exit_cpu(cpu);
T
Thomas Gleixner 已提交
8076 8077 8078 8079 8080 8081 8082 8083
		break;
	default:
		break;
	}

	return NOTIFY_OK;
}

8084
void __init perf_event_init(void)
T
Thomas Gleixner 已提交
8085
{
8086 8087
	int ret;

P
Peter Zijlstra 已提交
8088 8089
	idr_init(&pmu_idr);

8090
	perf_event_init_all_cpus();
8091
	init_srcu_struct(&pmus_srcu);
P
Peter Zijlstra 已提交
8092 8093 8094
	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);
8095 8096
	perf_tp_register();
	perf_cpu_notifier(perf_cpu_notify);
P
Peter Zijlstra 已提交
8097
	register_reboot_notifier(&perf_reboot_notifier);
8098 8099 8100

	ret = init_hw_breakpoint();
	WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8101 8102 8103

	/* do not patch jump label more than once per second */
	jump_label_rate_limit(&perf_sched_events, HZ);
8104 8105 8106 8107 8108 8109 8110

	/*
	 * 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
8111
}
P
Peter Zijlstra 已提交
8112 8113 8114 8115 8116 8117 8118 8119 8120 8121 8122 8123 8124 8125 8126 8127 8128 8129 8130 8131 8132 8133 8134 8135 8136 8137 8138 8139

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
8140 8141

#ifdef CONFIG_CGROUP_PERF
8142 8143
static struct cgroup_subsys_state *
perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
S
Stephane Eranian 已提交
8144 8145 8146
{
	struct perf_cgroup *jc;

8147
	jc = kzalloc(sizeof(*jc), GFP_KERNEL);
S
Stephane Eranian 已提交
8148 8149 8150 8151 8152 8153 8154 8155 8156 8157 8158 8159
	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;
}

8160
static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
S
Stephane Eranian 已提交
8161
{
8162 8163
	struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);

S
Stephane Eranian 已提交
8164 8165 8166 8167 8168 8169 8170 8171 8172 8173 8174
	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;
}

8175 8176
static void perf_cgroup_attach(struct cgroup_subsys_state *css,
			       struct cgroup_taskset *tset)
S
Stephane Eranian 已提交
8177
{
8178 8179
	struct task_struct *task;

8180
	cgroup_taskset_for_each(task, tset)
8181
		task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8182 8183
}

8184 8185
static void perf_cgroup_exit(struct cgroup_subsys_state *css,
			     struct cgroup_subsys_state *old_css,
8186
			     struct task_struct *task)
S
Stephane Eranian 已提交
8187 8188 8189 8190 8191 8192 8193 8194 8195
{
	/*
	 * 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;

8196
	task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8197 8198
}

8199
struct cgroup_subsys perf_event_cgrp_subsys = {
8200 8201
	.css_alloc	= perf_cgroup_css_alloc,
	.css_free	= perf_cgroup_css_free,
8202
	.exit		= perf_cgroup_exit,
8203
	.attach		= perf_cgroup_attach,
S
Stephane Eranian 已提交
8204 8205
};
#endif /* CONFIG_CGROUP_PERF */