core.c 194.5 KB
Newer Older
T
Thomas Gleixner 已提交
1
/*
I
Ingo Molnar 已提交
2
 * Performance events core code:
T
Thomas Gleixner 已提交
3
 *
4
 *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 6
 *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
 *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
A
Al Viro 已提交
7
 *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8
 *
I
Ingo Molnar 已提交
9
 * For licensing details see kernel-base/COPYING
T
Thomas Gleixner 已提交
10 11 12
 */

#include <linux/fs.h>
13
#include <linux/mm.h>
T
Thomas Gleixner 已提交
14 15
#include <linux/cpu.h>
#include <linux/smp.h>
P
Peter Zijlstra 已提交
16
#include <linux/idr.h>
17
#include <linux/file.h>
T
Thomas Gleixner 已提交
18
#include <linux/poll.h>
19
#include <linux/slab.h>
20
#include <linux/hash.h>
21
#include <linux/tick.h>
T
Thomas Gleixner 已提交
22
#include <linux/sysfs.h>
23
#include <linux/dcache.h>
T
Thomas Gleixner 已提交
24
#include <linux/percpu.h>
25
#include <linux/ptrace.h>
P
Peter Zijlstra 已提交
26
#include <linux/reboot.h>
27
#include <linux/vmstat.h>
P
Peter Zijlstra 已提交
28
#include <linux/device.h>
29
#include <linux/export.h>
30
#include <linux/vmalloc.h>
31 32
#include <linux/hardirq.h>
#include <linux/rculist.h>
T
Thomas Gleixner 已提交
33 34 35
#include <linux/uaccess.h>
#include <linux/syscalls.h>
#include <linux/anon_inodes.h>
I
Ingo Molnar 已提交
36
#include <linux/kernel_stat.h>
37
#include <linux/perf_event.h>
L
Li Zefan 已提交
38
#include <linux/ftrace_event.h>
39
#include <linux/hw_breakpoint.h>
40
#include <linux/mm_types.h>
41
#include <linux/cgroup.h>
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 51
static struct workqueue_struct *perf_wq;

52
struct remote_function_call {
53 54 55 56
	struct task_struct	*p;
	int			(*func)(void *info);
	void			*info;
	int			ret;
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 88 89
};

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 = {
90 91 92 93
		.p	= p,
		.func	= func,
		.info	= info,
		.ret	= -ESRCH, /* No such (running) process */
94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
	};

	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 = {
114 115 116 117
		.p	= NULL,
		.func	= func,
		.info	= info,
		.ret	= -ENXIO, /* No such CPU */
118 119 120 121 122 123 124
	};

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

	return data.ret;
}

125 126 127 128 129 130 131
#define EVENT_OWNER_KERNEL ((void *) -1)

static bool is_kernel_event(struct perf_event *event)
{
	return event->owner == EVENT_OWNER_KERNEL;
}

S
Stephane Eranian 已提交
132 133
#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
		       PERF_FLAG_FD_OUTPUT  |\
134 135
		       PERF_FLAG_PID_CGROUP |\
		       PERF_FLAG_FD_CLOEXEC)
S
Stephane Eranian 已提交
136

137 138 139 140 141 142 143
/*
 * branch priv levels that need permission checks
 */
#define PERF_SAMPLE_BRANCH_PERM_PLM \
	(PERF_SAMPLE_BRANCH_KERNEL |\
	 PERF_SAMPLE_BRANCH_HV)

144 145 146 147 148 149
enum event_type_t {
	EVENT_FLEXIBLE = 0x1,
	EVENT_PINNED = 0x2,
	EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
};

S
Stephane Eranian 已提交
150 151 152 153
/*
 * perf_sched_events : >0 events exist
 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
 */
154
struct static_key_deferred perf_sched_events __read_mostly;
S
Stephane Eranian 已提交
155
static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
156
static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
S
Stephane Eranian 已提交
157

158 159 160
static atomic_t nr_mmap_events __read_mostly;
static atomic_t nr_comm_events __read_mostly;
static atomic_t nr_task_events __read_mostly;
161
static atomic_t nr_freq_events __read_mostly;
162

P
Peter Zijlstra 已提交
163 164 165 166
static LIST_HEAD(pmus);
static DEFINE_MUTEX(pmus_lock);
static struct srcu_struct pmus_srcu;

167
/*
168
 * perf event paranoia level:
169 170
 *  -1 - not paranoid at all
 *   0 - disallow raw tracepoint access for unpriv
171
 *   1 - disallow cpu events for unpriv
172
 *   2 - disallow kernel profiling for unpriv
173
 */
174
int sysctl_perf_event_paranoid __read_mostly = 1;
175

176 177
/* Minimum for 512 kiB + 1 user control page */
int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
178 179

/*
180
 * max perf event sample rate
181
 */
182 183 184 185 186 187 188 189 190
#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
191 192
static int perf_sample_allowed_ns __read_mostly =
	DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
193 194 195 196 197 198

void update_perf_cpu_limits(void)
{
	u64 tmp = perf_sample_period_ns;

	tmp *= sysctl_perf_cpu_time_max_percent;
199
	do_div(tmp, 100);
P
Peter Zijlstra 已提交
200
	ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
201
}
P
Peter Zijlstra 已提交
202

203 204
static int perf_rotate_context(struct perf_cpu_context *cpuctx);

P
Peter Zijlstra 已提交
205 206 207 208
int perf_proc_update_handler(struct ctl_table *table, int write,
		void __user *buffer, size_t *lenp,
		loff_t *ppos)
{
209
	int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
P
Peter Zijlstra 已提交
210 211 212 213 214

	if (ret || !write)
		return ret;

	max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
	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
233 234 235

	return 0;
}
236

237 238 239 240 241 242 243
/*
 * 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
244
static DEFINE_PER_CPU(u64, running_sample_length);
245

246
static void perf_duration_warn(struct irq_work *w)
247
{
248
	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
249
	u64 avg_local_sample_len;
250
	u64 local_samples_len;
251

252
	local_samples_len = __this_cpu_read(running_sample_length);
253 254 255 256 257
	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",
258
			avg_local_sample_len, allowed_ns >> 1,
259 260 261 262 263 264 265
			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
266
	u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
267 268
	u64 avg_local_sample_len;
	u64 local_samples_len;
269

P
Peter Zijlstra 已提交
270
	if (allowed_ns == 0)
271 272 273
		return;

	/* decay the counter by 1 average sample */
274
	local_samples_len = __this_cpu_read(running_sample_length);
275 276
	local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
	local_samples_len += sample_len_ns;
277
	__this_cpu_write(running_sample_length, local_samples_len);
278 279 280 281 282 283 284 285

	/*
	 * 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
286
	if (avg_local_sample_len <= allowed_ns)
287 288 289 290 291 292 293 294 295 296
		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();
297

298 299 300 301 302 303
	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);
	}
304 305
}

306
static atomic64_t perf_event_id;
307

308 309 310 311
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
312 313 314 315 316
			     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);
317

318
void __weak perf_event_print_debug(void)	{ }
T
Thomas Gleixner 已提交
319

320
extern __weak const char *perf_pmu_name(void)
T
Thomas Gleixner 已提交
321
{
322
	return "pmu";
T
Thomas Gleixner 已提交
323 324
}

325 326 327 328 329
static inline u64 perf_clock(void)
{
	return local_clock();
}

S
Stephane Eranian 已提交
330 331 332 333 334 335
static inline struct perf_cpu_context *
__get_cpu_context(struct perf_event_context *ctx)
{
	return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
}

336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351
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
352 353
#ifdef CONFIG_CGROUP_PERF

354 355 356 357 358 359 360 361 362 363 364
/*
 * 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;
365
	struct perf_cgroup_info	__percpu *info;
366 367
};

368 369 370 371 372
/*
 * 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
373 374 375
static inline struct perf_cgroup *
perf_cgroup_from_task(struct task_struct *task)
{
376
	return container_of(task_css(task, perf_event_cgrp_id),
377
			    struct perf_cgroup, css);
S
Stephane Eranian 已提交
378 379 380 381 382 383 384 385
}

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);

386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
	/* @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
402 403 404 405
}

static inline void perf_detach_cgroup(struct perf_event *event)
{
Z
Zefan Li 已提交
406
	css_put(&event->cgrp->css);
S
Stephane Eranian 已提交
407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444
	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)
{
445 446
	struct perf_cgroup *cgrp;

S
Stephane Eranian 已提交
447
	/*
448 449
	 * ensure we access cgroup data only when needed and
	 * when we know the cgroup is pinned (css_get)
S
Stephane Eranian 已提交
450
	 */
451
	if (!is_cgroup_event(event))
S
Stephane Eranian 已提交
452 453
		return;

454 455 456 457 458 459
	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
460 461 462
}

static inline void
463 464
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
465 466 467 468
{
	struct perf_cgroup *cgrp;
	struct perf_cgroup_info *info;

469 470 471 472 473 474
	/*
	 * 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
475 476 477 478
		return;

	cgrp = perf_cgroup_from_task(task);
	info = this_cpu_ptr(cgrp->info);
479
	info->timestamp = ctx->timestamp;
S
Stephane Eranian 已提交
480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511
}

#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);
512 513
		if (cpuctx->unique_pmu != pmu)
			continue; /* ensure we process each cpuctx once */
S
Stephane Eranian 已提交
514 515 516 517 518 519 520 521 522

		/*
		 * 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) {
523 524
			perf_ctx_lock(cpuctx, cpuctx->task_ctx);
			perf_pmu_disable(cpuctx->ctx.pmu);
S
Stephane Eranian 已提交
525 526 527 528 529 530 531 532 533 534 535

			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) {
536
				WARN_ON_ONCE(cpuctx->cgrp);
537 538 539 540
				/*
				 * set cgrp before ctxsw in to allow
				 * event_filter_match() to not have to pass
				 * task around
S
Stephane Eranian 已提交
541 542 543 544
				 */
				cpuctx->cgrp = perf_cgroup_from_task(task);
				cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
			}
545 546
			perf_pmu_enable(cpuctx->ctx.pmu);
			perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
S
Stephane Eranian 已提交
547 548 549 550 551 552 553 554
		}
	}

	rcu_read_unlock();

	local_irq_restore(flags);
}

555 556
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
557
{
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579
	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
580 581
}

582 583
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
584
{
585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602
	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
603 604 605 606 607 608 609 610
}

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;
611 612
	struct fd f = fdget(fd);
	int ret = 0;
S
Stephane Eranian 已提交
613

614
	if (!f.file)
S
Stephane Eranian 已提交
615 616
		return -EBADF;

A
Al Viro 已提交
617
	css = css_tryget_online_from_dir(f.file->f_path.dentry,
618
					 &perf_event_cgrp_subsys);
619 620 621 622
	if (IS_ERR(css)) {
		ret = PTR_ERR(css);
		goto out;
	}
S
Stephane Eranian 已提交
623 624 625 626 627 628 629 630 631 632 633 634 635

	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;
	}
636
out:
637
	fdput(f);
S
Stephane Eranian 已提交
638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710
	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)
{
}

711 712
static inline void perf_cgroup_sched_out(struct task_struct *task,
					 struct task_struct *next)
S
Stephane Eranian 已提交
713 714 715
{
}

716 717
static inline void perf_cgroup_sched_in(struct task_struct *prev,
					struct task_struct *task)
S
Stephane Eranian 已提交
718 719 720 721 722 723 724 725 726 727 728
{
}

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
729 730
perf_cgroup_set_timestamp(struct task_struct *task,
			  struct perf_event_context *ctx)
S
Stephane Eranian 已提交
731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760
{
}

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

761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823
/*
 * 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;
824
	int timer;
825 826 827 828 829

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

830 831 832 833 834 835 836 837 838
	/*
	 * 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);
839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860

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

P
Peter Zijlstra 已提交
868
void perf_pmu_enable(struct pmu *pmu)
869
{
P
Peter Zijlstra 已提交
870 871 872
	int *count = this_cpu_ptr(pmu->pmu_disable_count);
	if (!--(*count))
		pmu->pmu_enable(pmu);
873 874
}

875 876 877 878 879 880 881
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
882
static void perf_pmu_rotate_start(struct pmu *pmu)
883
{
P
Peter Zijlstra 已提交
884
	struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
885
	struct list_head *head = this_cpu_ptr(&rotation_list);
886

887
	WARN_ON(!irqs_disabled());
888

889
	if (list_empty(&cpuctx->rotation_list))
890
		list_add(&cpuctx->rotation_list, head);
891 892
}

893
static void get_ctx(struct perf_event_context *ctx)
894
{
895
	WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
896 897
}

898
static void put_ctx(struct perf_event_context *ctx)
899
{
900 901 902
	if (atomic_dec_and_test(&ctx->refcount)) {
		if (ctx->parent_ctx)
			put_ctx(ctx->parent_ctx);
903 904
		if (ctx->task)
			put_task_struct(ctx->task);
905
		kfree_rcu(ctx, rcu_head);
906
	}
907 908
}

909 910 911 912 913 914 915
/*
 * This must be done under the ctx->lock, such as to serialize against
 * context_equiv(), therefore we cannot call put_ctx() since that might end up
 * calling scheduler related locks and ctx->lock nests inside those.
 */
static __must_check struct perf_event_context *
unclone_ctx(struct perf_event_context *ctx)
916
{
917 918 919 920 921
	struct perf_event_context *parent_ctx = ctx->parent_ctx;

	lockdep_assert_held(&ctx->lock);

	if (parent_ctx)
922
		ctx->parent_ctx = NULL;
923
	ctx->generation++;
924 925

	return parent_ctx;
926 927
}

928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949
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);
}

950
/*
951
 * If we inherit events we want to return the parent event id
952 953
 * to userspace.
 */
954
static u64 primary_event_id(struct perf_event *event)
955
{
956
	u64 id = event->id;
957

958 959
	if (event->parent)
		id = event->parent->id;
960 961 962 963

	return id;
}

964
/*
965
 * Get the perf_event_context for a task and lock it.
966 967 968
 * This has to cope with with the fact that until it is locked,
 * the context could get moved to another task.
 */
969
static struct perf_event_context *
P
Peter Zijlstra 已提交
970
perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
971
{
972
	struct perf_event_context *ctx;
973

P
Peter Zijlstra 已提交
974
retry:
975 976 977 978 979 980 981 982 983 984 985
	/*
	 * 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
986
	ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
987 988 989 990
	if (ctx) {
		/*
		 * If this context is a clone of another, it might
		 * get swapped for another underneath us by
991
		 * perf_event_task_sched_out, though the
992 993 994 995 996 997
		 * 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.
		 */
998
		raw_spin_lock_irqsave(&ctx->lock, *flags);
P
Peter Zijlstra 已提交
999
		if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1000
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1001 1002
			rcu_read_unlock();
			preempt_enable();
1003 1004
			goto retry;
		}
1005 1006

		if (!atomic_inc_not_zero(&ctx->refcount)) {
1007
			raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1008 1009
			ctx = NULL;
		}
1010 1011
	}
	rcu_read_unlock();
1012
	preempt_enable();
1013 1014 1015 1016 1017 1018 1019 1020
	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
1021 1022
static struct perf_event_context *
perf_pin_task_context(struct task_struct *task, int ctxn)
1023
{
1024
	struct perf_event_context *ctx;
1025 1026
	unsigned long flags;

P
Peter Zijlstra 已提交
1027
	ctx = perf_lock_task_context(task, ctxn, &flags);
1028 1029
	if (ctx) {
		++ctx->pin_count;
1030
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
1031 1032 1033 1034
	}
	return ctx;
}

1035
static void perf_unpin_context(struct perf_event_context *ctx)
1036 1037 1038
{
	unsigned long flags;

1039
	raw_spin_lock_irqsave(&ctx->lock, flags);
1040
	--ctx->pin_count;
1041
	raw_spin_unlock_irqrestore(&ctx->lock, flags);
1042 1043
}

1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054
/*
 * 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;
}

1055 1056 1057
static u64 perf_event_time(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;
S
Stephane Eranian 已提交
1058 1059 1060 1061

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

1062 1063 1064
	return ctx ? ctx->time : 0;
}

1065 1066
/*
 * Update the total_time_enabled and total_time_running fields for a event.
1067
 * The caller of this function needs to hold the ctx->lock.
1068 1069 1070 1071 1072 1073 1074 1075 1076
 */
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
1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087
	/*
	 * 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))
1088
		run_end = perf_cgroup_event_time(event);
S
Stephane Eranian 已提交
1089 1090
	else if (ctx->is_active)
		run_end = ctx->time;
1091 1092 1093 1094
	else
		run_end = event->tstamp_stopped;

	event->total_time_enabled = run_end - event->tstamp_enabled;
1095 1096 1097 1098

	if (event->state == PERF_EVENT_STATE_INACTIVE)
		run_end = event->tstamp_stopped;
	else
1099
		run_end = perf_event_time(event);
1100 1101

	event->total_time_running = run_end - event->tstamp_running;
S
Stephane Eranian 已提交
1102

1103 1104
}

1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116
/*
 * 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);
}

1117 1118 1119 1120 1121 1122 1123 1124 1125
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;
}

1126
/*
1127
 * Add a event from the lists for its context.
1128 1129
 * Must be called with ctx->mutex and ctx->lock held.
 */
1130
static void
1131
list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1132
{
1133 1134
	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
	event->attach_state |= PERF_ATTACH_CONTEXT;
1135 1136

	/*
1137 1138 1139
	 * 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.
1140
	 */
1141
	if (event->group_leader == event) {
1142 1143
		struct list_head *list;

1144 1145 1146
		if (is_software_event(event))
			event->group_flags |= PERF_GROUP_SOFTWARE;

1147 1148
		list = ctx_group_list(event, ctx);
		list_add_tail(&event->group_entry, list);
P
Peter Zijlstra 已提交
1149
	}
P
Peter Zijlstra 已提交
1150

1151
	if (is_cgroup_event(event))
S
Stephane Eranian 已提交
1152 1153
		ctx->nr_cgroups++;

1154 1155 1156
	if (has_branch_stack(event))
		ctx->nr_branch_stack++;

1157
	list_add_rcu(&event->event_entry, &ctx->event_list);
1158
	if (!ctx->nr_events)
P
Peter Zijlstra 已提交
1159
		perf_pmu_rotate_start(ctx->pmu);
1160 1161
	ctx->nr_events++;
	if (event->attr.inherit_stat)
1162
		ctx->nr_stat++;
1163 1164

	ctx->generation++;
1165 1166
}

J
Jiri Olsa 已提交
1167 1168 1169 1170 1171 1172 1173 1174 1175
/*
 * 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;
}

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 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214
/*
 * 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);

1215 1216 1217 1218 1219 1220
	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
1221 1222 1223
	if (sample_type & PERF_SAMPLE_WEIGHT)
		size += sizeof(data->weight);

1224 1225 1226
	if (sample_type & PERF_SAMPLE_READ)
		size += event->read_size;

1227 1228 1229
	if (sample_type & PERF_SAMPLE_DATA_SRC)
		size += sizeof(data->data_src.val);

A
Andi Kleen 已提交
1230 1231 1232
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		size += sizeof(data->txn);

1233 1234 1235 1236 1237 1238 1239 1240 1241
	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;

1242 1243 1244 1245 1246 1247
	if (sample_type & PERF_SAMPLE_TID)
		size += sizeof(data->tid_entry);

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

1248 1249 1250
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		size += sizeof(data->id);

1251 1252 1253 1254 1255 1256 1257 1258 1259
	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);

1260
	event->id_header_size = size;
1261 1262
}

1263 1264
static void perf_group_attach(struct perf_event *event)
{
1265
	struct perf_event *group_leader = event->group_leader, *pos;
1266

P
Peter Zijlstra 已提交
1267 1268 1269 1270 1271 1272
	/*
	 * We can have double attach due to group movement in perf_event_open.
	 */
	if (event->attach_state & PERF_ATTACH_GROUP)
		return;

1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283
	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++;
1284 1285 1286 1287 1288

	perf_event__header_size(group_leader);

	list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
		perf_event__header_size(pos);
1289 1290
}

1291
/*
1292
 * Remove a event from the lists for its context.
1293
 * Must be called with ctx->mutex and ctx->lock held.
1294
 */
1295
static void
1296
list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1297
{
1298
	struct perf_cpu_context *cpuctx;
1299 1300 1301 1302
	/*
	 * We can have double detach due to exit/hot-unplug + close.
	 */
	if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1303
		return;
1304 1305 1306

	event->attach_state &= ~PERF_ATTACH_CONTEXT;

1307
	if (is_cgroup_event(event)) {
S
Stephane Eranian 已提交
1308
		ctx->nr_cgroups--;
1309 1310 1311 1312 1313 1314 1315 1316 1317
		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
1318

1319 1320 1321
	if (has_branch_stack(event))
		ctx->nr_branch_stack--;

1322 1323
	ctx->nr_events--;
	if (event->attr.inherit_stat)
1324
		ctx->nr_stat--;
1325

1326
	list_del_rcu(&event->event_entry);
1327

1328 1329
	if (event->group_leader == event)
		list_del_init(&event->group_entry);
P
Peter Zijlstra 已提交
1330

1331
	update_group_times(event);
1332 1333 1334 1335 1336 1337 1338 1339 1340 1341

	/*
	 * 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;
1342 1343

	ctx->generation++;
1344 1345
}

1346
static void perf_group_detach(struct perf_event *event)
1347 1348
{
	struct perf_event *sibling, *tmp;
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364
	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--;
1365
		goto out;
1366 1367 1368 1369
	}

	if (!list_empty(&event->group_entry))
		list = &event->group_entry;
1370

1371
	/*
1372 1373
	 * If this was a group event with sibling events then
	 * upgrade the siblings to singleton events by adding them
1374
	 * to whatever list we are on.
1375
	 */
1376
	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1377 1378
		if (list)
			list_move_tail(&sibling->group_entry, list);
1379
		sibling->group_leader = sibling;
1380 1381 1382

		/* Inherit group flags from the previous leader */
		sibling->group_flags = event->group_flags;
1383
	}
1384 1385 1386 1387 1388 1389

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);
1390 1391
}

1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430
/*
 * User event without the task.
 */
static bool is_orphaned_event(struct perf_event *event)
{
	return event && !is_kernel_event(event) && !event->owner;
}

/*
 * Event has a parent but parent's task finished and it's
 * alive only because of children holding refference.
 */
static bool is_orphaned_child(struct perf_event *event)
{
	return is_orphaned_event(event->parent);
}

static void orphans_remove_work(struct work_struct *work);

static void schedule_orphans_remove(struct perf_event_context *ctx)
{
	if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
		return;

	if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
		get_ctx(ctx);
		ctx->orphans_remove_sched = true;
	}
}

static int __init perf_workqueue_init(void)
{
	perf_wq = create_singlethread_workqueue("perf");
	WARN(!perf_wq, "failed to create perf workqueue\n");
	return perf_wq ? 0 : -1;
}

core_initcall(perf_workqueue_init);

1431 1432 1433
static inline int
event_filter_match(struct perf_event *event)
{
S
Stephane Eranian 已提交
1434 1435
	return (event->cpu == -1 || event->cpu == smp_processor_id())
	    && perf_cgroup_match(event);
1436 1437
}

1438 1439
static void
event_sched_out(struct perf_event *event,
1440
		  struct perf_cpu_context *cpuctx,
1441
		  struct perf_event_context *ctx)
1442
{
1443
	u64 tstamp = perf_event_time(event);
1444 1445 1446 1447 1448 1449 1450 1451 1452
	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
1453
		delta = tstamp - event->tstamp_stopped;
1454
		event->tstamp_running += delta;
1455
		event->tstamp_stopped = tstamp;
1456 1457
	}

1458
	if (event->state != PERF_EVENT_STATE_ACTIVE)
1459
		return;
1460

1461 1462
	perf_pmu_disable(event->pmu);

1463 1464 1465 1466
	event->state = PERF_EVENT_STATE_INACTIVE;
	if (event->pending_disable) {
		event->pending_disable = 0;
		event->state = PERF_EVENT_STATE_OFF;
1467
	}
1468
	event->tstamp_stopped = tstamp;
P
Peter Zijlstra 已提交
1469
	event->pmu->del(event, 0);
1470
	event->oncpu = -1;
1471

1472
	if (!is_software_event(event))
1473 1474
		cpuctx->active_oncpu--;
	ctx->nr_active--;
1475 1476
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq--;
1477
	if (event->attr.exclusive || !cpuctx->active_oncpu)
1478
		cpuctx->exclusive = 0;
1479

1480 1481 1482
	if (is_orphaned_child(event))
		schedule_orphans_remove(ctx);

1483
	perf_pmu_enable(event->pmu);
1484 1485
}

1486
static void
1487
group_sched_out(struct perf_event *group_event,
1488
		struct perf_cpu_context *cpuctx,
1489
		struct perf_event_context *ctx)
1490
{
1491
	struct perf_event *event;
1492
	int state = group_event->state;
1493

1494
	event_sched_out(group_event, cpuctx, ctx);
1495 1496 1497 1498

	/*
	 * Schedule out siblings (if any):
	 */
1499 1500
	list_for_each_entry(event, &group_event->sibling_list, group_entry)
		event_sched_out(event, cpuctx, ctx);
1501

1502
	if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1503 1504 1505
		cpuctx->exclusive = 0;
}

1506 1507 1508 1509 1510
struct remove_event {
	struct perf_event *event;
	bool detach_group;
};

T
Thomas Gleixner 已提交
1511
/*
1512
 * Cross CPU call to remove a performance event
T
Thomas Gleixner 已提交
1513
 *
1514
 * We disable the event on the hardware level first. After that we
T
Thomas Gleixner 已提交
1515 1516
 * remove it from the context list.
 */
1517
static int __perf_remove_from_context(void *info)
T
Thomas Gleixner 已提交
1518
{
1519 1520
	struct remove_event *re = info;
	struct perf_event *event = re->event;
1521
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1522
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
T
Thomas Gleixner 已提交
1523

1524
	raw_spin_lock(&ctx->lock);
1525
	event_sched_out(event, cpuctx, ctx);
1526 1527
	if (re->detach_group)
		perf_group_detach(event);
1528
	list_del_event(event, ctx);
1529 1530 1531 1532
	if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
		ctx->is_active = 0;
		cpuctx->task_ctx = NULL;
	}
1533
	raw_spin_unlock(&ctx->lock);
1534 1535

	return 0;
T
Thomas Gleixner 已提交
1536 1537 1538 1539
}


/*
1540
 * Remove the event from a task's (or a CPU's) list of events.
T
Thomas Gleixner 已提交
1541
 *
1542
 * CPU events are removed with a smp call. For task events we only
T
Thomas Gleixner 已提交
1543
 * call when the task is on a CPU.
1544
 *
1545 1546
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1547 1548
 * 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.
1549
 * When called from perf_event_exit_task, it's OK because the
1550
 * context has been detached from its task.
T
Thomas Gleixner 已提交
1551
 */
1552
static void perf_remove_from_context(struct perf_event *event, bool detach_group)
T
Thomas Gleixner 已提交
1553
{
1554
	struct perf_event_context *ctx = event->ctx;
T
Thomas Gleixner 已提交
1555
	struct task_struct *task = ctx->task;
1556 1557 1558 1559
	struct remove_event re = {
		.event = event,
		.detach_group = detach_group,
	};
T
Thomas Gleixner 已提交
1560

1561 1562
	lockdep_assert_held(&ctx->mutex);

T
Thomas Gleixner 已提交
1563 1564
	if (!task) {
		/*
1565 1566 1567 1568
		 * Per cpu events are removed via an smp call. The removal can
		 * fail if the CPU is currently offline, but in that case we
		 * already called __perf_remove_from_context from
		 * perf_event_exit_cpu.
T
Thomas Gleixner 已提交
1569
		 */
1570
		cpu_function_call(event->cpu, __perf_remove_from_context, &re);
T
Thomas Gleixner 已提交
1571 1572 1573 1574
		return;
	}

retry:
1575
	if (!task_function_call(task, __perf_remove_from_context, &re))
1576
		return;
T
Thomas Gleixner 已提交
1577

1578
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1579
	/*
1580 1581
	 * 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
1582
	 */
1583
	if (ctx->is_active) {
1584
		raw_spin_unlock_irq(&ctx->lock);
1585 1586 1587 1588 1589
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
T
Thomas Gleixner 已提交
1590 1591 1592 1593
		goto retry;
	}

	/*
1594 1595
	 * 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
1596
	 */
1597 1598
	if (detach_group)
		perf_group_detach(event);
1599
	list_del_event(event, ctx);
1600
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
1601 1602
}

1603
/*
1604
 * Cross CPU call to disable a performance event
1605
 */
1606
int __perf_event_disable(void *info)
1607
{
1608 1609
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1610
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1611 1612

	/*
1613 1614
	 * If this is a per-task event, need to check whether this
	 * event's task is the current task on this cpu.
1615 1616 1617
	 *
	 * Can trigger due to concurrent perf_event_context_sched_out()
	 * flipping contexts around.
1618
	 */
1619
	if (ctx->task && cpuctx->task_ctx != ctx)
1620
		return -EINVAL;
1621

1622
	raw_spin_lock(&ctx->lock);
1623 1624

	/*
1625
	 * If the event is on, turn it off.
1626 1627
	 * If it is in error state, leave it in error state.
	 */
1628
	if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1629
		update_context_time(ctx);
S
Stephane Eranian 已提交
1630
		update_cgrp_time_from_event(event);
1631 1632 1633
		update_group_times(event);
		if (event == event->group_leader)
			group_sched_out(event, cpuctx, ctx);
1634
		else
1635 1636
			event_sched_out(event, cpuctx, ctx);
		event->state = PERF_EVENT_STATE_OFF;
1637 1638
	}

1639
	raw_spin_unlock(&ctx->lock);
1640 1641

	return 0;
1642 1643 1644
}

/*
1645
 * Disable a event.
1646
 *
1647 1648
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
1649
 * remains valid.  This condition is satisifed when called through
1650 1651 1652 1653
 * 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
1654
 * is the current context on this CPU and preemption is disabled,
1655
 * hence we can't get into perf_event_task_sched_out for this context.
1656
 */
1657
void perf_event_disable(struct perf_event *event)
1658
{
1659
	struct perf_event_context *ctx = event->ctx;
1660 1661 1662 1663
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
1664
		 * Disable the event on the cpu that it's on
1665
		 */
1666
		cpu_function_call(event->cpu, __perf_event_disable, event);
1667 1668 1669
		return;
	}

P
Peter Zijlstra 已提交
1670
retry:
1671 1672
	if (!task_function_call(task, __perf_event_disable, event))
		return;
1673

1674
	raw_spin_lock_irq(&ctx->lock);
1675
	/*
1676
	 * If the event is still active, we need to retry the cross-call.
1677
	 */
1678
	if (event->state == PERF_EVENT_STATE_ACTIVE) {
1679
		raw_spin_unlock_irq(&ctx->lock);
1680 1681 1682 1683 1684
		/*
		 * Reload the task pointer, it might have been changed by
		 * a concurrent perf_event_context_sched_out().
		 */
		task = ctx->task;
1685 1686 1687 1688 1689 1690 1691
		goto retry;
	}

	/*
	 * Since we have the lock this context can't be scheduled
	 * in, so we can change the state safely.
	 */
1692 1693 1694
	if (event->state == PERF_EVENT_STATE_INACTIVE) {
		update_group_times(event);
		event->state = PERF_EVENT_STATE_OFF;
1695
	}
1696
	raw_spin_unlock_irq(&ctx->lock);
1697
}
1698
EXPORT_SYMBOL_GPL(perf_event_disable);
1699

S
Stephane Eranian 已提交
1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734
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
1735 1736 1737 1738
#define MAX_INTERRUPTS (~0ULL)

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

1739
static int
1740
event_sched_in(struct perf_event *event,
1741
		 struct perf_cpu_context *cpuctx,
1742
		 struct perf_event_context *ctx)
1743
{
1744
	u64 tstamp = perf_event_time(event);
1745
	int ret = 0;
1746

1747 1748
	lockdep_assert_held(&ctx->lock);

1749
	if (event->state <= PERF_EVENT_STATE_OFF)
1750 1751
		return 0;

1752
	event->state = PERF_EVENT_STATE_ACTIVE;
1753
	event->oncpu = smp_processor_id();
P
Peter Zijlstra 已提交
1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764

	/*
	 * 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;
	}

1765 1766 1767 1768 1769
	/*
	 * The new state must be visible before we turn it on in the hardware:
	 */
	smp_wmb();

1770 1771
	perf_pmu_disable(event->pmu);

P
Peter Zijlstra 已提交
1772
	if (event->pmu->add(event, PERF_EF_START)) {
1773 1774
		event->state = PERF_EVENT_STATE_INACTIVE;
		event->oncpu = -1;
1775 1776
		ret = -EAGAIN;
		goto out;
1777 1778
	}

1779
	event->tstamp_running += tstamp - event->tstamp_stopped;
1780

S
Stephane Eranian 已提交
1781
	perf_set_shadow_time(event, ctx, tstamp);
1782

1783
	if (!is_software_event(event))
1784
		cpuctx->active_oncpu++;
1785
	ctx->nr_active++;
1786 1787
	if (event->attr.freq && event->attr.sample_freq)
		ctx->nr_freq++;
1788

1789
	if (event->attr.exclusive)
1790 1791
		cpuctx->exclusive = 1;

1792 1793 1794
	if (is_orphaned_child(event))
		schedule_orphans_remove(ctx);

1795 1796 1797 1798
out:
	perf_pmu_enable(event->pmu);

	return ret;
1799 1800
}

1801
static int
1802
group_sched_in(struct perf_event *group_event,
1803
	       struct perf_cpu_context *cpuctx,
1804
	       struct perf_event_context *ctx)
1805
{
1806
	struct perf_event *event, *partial_group = NULL;
P
Peter Zijlstra 已提交
1807
	struct pmu *pmu = ctx->pmu;
1808 1809
	u64 now = ctx->time;
	bool simulate = false;
1810

1811
	if (group_event->state == PERF_EVENT_STATE_OFF)
1812 1813
		return 0;

P
Peter Zijlstra 已提交
1814
	pmu->start_txn(pmu);
1815

1816
	if (event_sched_in(group_event, cpuctx, ctx)) {
P
Peter Zijlstra 已提交
1817
		pmu->cancel_txn(pmu);
1818
		perf_cpu_hrtimer_restart(cpuctx);
1819
		return -EAGAIN;
1820
	}
1821 1822 1823 1824

	/*
	 * Schedule in siblings as one group (if any):
	 */
1825
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1826
		if (event_sched_in(event, cpuctx, ctx)) {
1827
			partial_group = event;
1828 1829 1830 1831
			goto group_error;
		}
	}

1832
	if (!pmu->commit_txn(pmu))
1833
		return 0;
1834

1835 1836 1837 1838
group_error:
	/*
	 * Groups can be scheduled in as one unit only, so undo any
	 * partial group before returning:
1839 1840 1841 1842 1843 1844 1845 1846 1847 1848
	 * 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.
1849
	 */
1850 1851
	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
		if (event == partial_group)
1852 1853 1854 1855 1856 1857 1858 1859
			simulate = true;

		if (simulate) {
			event->tstamp_running += now - event->tstamp_stopped;
			event->tstamp_stopped = now;
		} else {
			event_sched_out(event, cpuctx, ctx);
		}
1860
	}
1861
	event_sched_out(group_event, cpuctx, ctx);
1862

P
Peter Zijlstra 已提交
1863
	pmu->cancel_txn(pmu);
1864

1865 1866
	perf_cpu_hrtimer_restart(cpuctx);

1867 1868 1869
	return -EAGAIN;
}

1870
/*
1871
 * Work out whether we can put this event group on the CPU now.
1872
 */
1873
static int group_can_go_on(struct perf_event *event,
1874 1875 1876 1877
			   struct perf_cpu_context *cpuctx,
			   int can_add_hw)
{
	/*
1878
	 * Groups consisting entirely of software events can always go on.
1879
	 */
1880
	if (event->group_flags & PERF_GROUP_SOFTWARE)
1881 1882 1883
		return 1;
	/*
	 * If an exclusive group is already on, no other hardware
1884
	 * events can go on.
1885 1886 1887 1888 1889
	 */
	if (cpuctx->exclusive)
		return 0;
	/*
	 * If this group is exclusive and there are already
1890
	 * events on the CPU, it can't go on.
1891
	 */
1892
	if (event->attr.exclusive && cpuctx->active_oncpu)
1893 1894 1895 1896 1897 1898 1899 1900
		return 0;
	/*
	 * Otherwise, try to add it if all previous groups were able
	 * to go on.
	 */
	return can_add_hw;
}

1901 1902
static void add_event_to_ctx(struct perf_event *event,
			       struct perf_event_context *ctx)
1903
{
1904 1905
	u64 tstamp = perf_event_time(event);

1906
	list_add_event(event, ctx);
1907
	perf_group_attach(event);
1908 1909 1910
	event->tstamp_enabled = tstamp;
	event->tstamp_running = tstamp;
	event->tstamp_stopped = tstamp;
1911 1912
}

1913 1914 1915 1916 1917 1918
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);
1919

1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931
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
1932
/*
1933
 * Cross CPU call to install and enable a performance event
1934 1935
 *
 * Must be called with ctx->mutex held
T
Thomas Gleixner 已提交
1936
 */
1937
static int  __perf_install_in_context(void *info)
T
Thomas Gleixner 已提交
1938
{
1939 1940
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
1941
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1942 1943 1944
	struct perf_event_context *task_ctx = cpuctx->task_ctx;
	struct task_struct *task = current;

1945
	perf_ctx_lock(cpuctx, task_ctx);
1946
	perf_pmu_disable(cpuctx->ctx.pmu);
T
Thomas Gleixner 已提交
1947 1948

	/*
1949
	 * If there was an active task_ctx schedule it out.
T
Thomas Gleixner 已提交
1950
	 */
1951
	if (task_ctx)
1952
		task_ctx_sched_out(task_ctx);
1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966

	/*
	 * 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;
1967 1968
		task = task_ctx->task;
	}
1969

1970
	cpu_ctx_sched_out(cpuctx, EVENT_ALL);
T
Thomas Gleixner 已提交
1971

1972
	update_context_time(ctx);
S
Stephane Eranian 已提交
1973 1974 1975 1976 1977 1978
	/*
	 * 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
1979

1980
	add_event_to_ctx(event, ctx);
T
Thomas Gleixner 已提交
1981

1982
	/*
1983
	 * Schedule everything back in
1984
	 */
1985
	perf_event_sched_in(cpuctx, task_ctx, task);
1986 1987 1988

	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, task_ctx);
1989 1990

	return 0;
T
Thomas Gleixner 已提交
1991 1992 1993
}

/*
1994
 * Attach a performance event to a context
T
Thomas Gleixner 已提交
1995
 *
1996 1997
 * First we add the event to the list with the hardware enable bit
 * in event->hw_config cleared.
T
Thomas Gleixner 已提交
1998
 *
1999
 * If the event is attached to a task which is on a CPU we use a smp
T
Thomas Gleixner 已提交
2000 2001 2002 2003
 * 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
2004 2005
perf_install_in_context(struct perf_event_context *ctx,
			struct perf_event *event,
T
Thomas Gleixner 已提交
2006 2007 2008 2009
			int cpu)
{
	struct task_struct *task = ctx->task;

2010 2011
	lockdep_assert_held(&ctx->mutex);

2012
	event->ctx = ctx;
2013 2014
	if (event->cpu != -1)
		event->cpu = cpu;
2015

T
Thomas Gleixner 已提交
2016 2017
	if (!task) {
		/*
2018
		 * Per cpu events are installed via an smp call and
2019
		 * the install is always successful.
T
Thomas Gleixner 已提交
2020
		 */
2021
		cpu_function_call(cpu, __perf_install_in_context, event);
T
Thomas Gleixner 已提交
2022 2023 2024 2025
		return;
	}

retry:
2026 2027
	if (!task_function_call(task, __perf_install_in_context, event))
		return;
T
Thomas Gleixner 已提交
2028

2029
	raw_spin_lock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
2030
	/*
2031 2032
	 * 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
2033
	 */
2034
	if (ctx->is_active) {
2035
		raw_spin_unlock_irq(&ctx->lock);
2036 2037 2038 2039 2040
		/*
		 * 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
2041 2042 2043 2044
		goto retry;
	}

	/*
2045 2046
	 * 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
2047
	 */
2048
	add_event_to_ctx(event, ctx);
2049
	raw_spin_unlock_irq(&ctx->lock);
T
Thomas Gleixner 已提交
2050 2051
}

2052
/*
2053
 * Put a event into inactive state and update time fields.
2054 2055 2056 2057 2058 2059
 * 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.
 */
2060
static void __perf_event_mark_enabled(struct perf_event *event)
2061
{
2062
	struct perf_event *sub;
2063
	u64 tstamp = perf_event_time(event);
2064

2065
	event->state = PERF_EVENT_STATE_INACTIVE;
2066
	event->tstamp_enabled = tstamp - event->total_time_enabled;
P
Peter Zijlstra 已提交
2067
	list_for_each_entry(sub, &event->sibling_list, group_entry) {
2068 2069
		if (sub->state >= PERF_EVENT_STATE_INACTIVE)
			sub->tstamp_enabled = tstamp - sub->total_time_enabled;
P
Peter Zijlstra 已提交
2070
	}
2071 2072
}

2073
/*
2074
 * Cross CPU call to enable a performance event
2075
 */
2076
static int __perf_event_enable(void *info)
2077
{
2078 2079 2080
	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
2081
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2082
	int err;
2083

2084 2085 2086 2087 2088 2089 2090 2091 2092 2093
	/*
	 * 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)
2094
		return -EINVAL;
2095

2096
	raw_spin_lock(&ctx->lock);
2097
	update_context_time(ctx);
2098

2099
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2100
		goto unlock;
S
Stephane Eranian 已提交
2101 2102 2103 2104

	/*
	 * set current task's cgroup time reference point
	 */
2105
	perf_cgroup_set_timestamp(current, ctx);
S
Stephane Eranian 已提交
2106

2107
	__perf_event_mark_enabled(event);
2108

S
Stephane Eranian 已提交
2109 2110 2111
	if (!event_filter_match(event)) {
		if (is_cgroup_event(event))
			perf_cgroup_defer_enabled(event);
2112
		goto unlock;
S
Stephane Eranian 已提交
2113
	}
2114

2115
	/*
2116
	 * If the event is in a group and isn't the group leader,
2117
	 * then don't put it on unless the group is on.
2118
	 */
2119
	if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2120
		goto unlock;
2121

2122
	if (!group_can_go_on(event, cpuctx, 1)) {
2123
		err = -EEXIST;
2124
	} else {
2125
		if (event == leader)
2126
			err = group_sched_in(event, cpuctx, ctx);
2127
		else
2128
			err = event_sched_in(event, cpuctx, ctx);
2129
	}
2130 2131 2132

	if (err) {
		/*
2133
		 * If this event can't go on and it's part of a
2134 2135
		 * group, then the whole group has to come off.
		 */
2136
		if (leader != event) {
2137
			group_sched_out(leader, cpuctx, ctx);
2138 2139
			perf_cpu_hrtimer_restart(cpuctx);
		}
2140
		if (leader->attr.pinned) {
2141
			update_group_times(leader);
2142
			leader->state = PERF_EVENT_STATE_ERROR;
2143
		}
2144 2145
	}

P
Peter Zijlstra 已提交
2146
unlock:
2147
	raw_spin_unlock(&ctx->lock);
2148 2149

	return 0;
2150 2151 2152
}

/*
2153
 * Enable a event.
2154
 *
2155 2156
 * If event->ctx is a cloned context, callers must make sure that
 * every task struct that event->ctx->task could possibly point to
2157
 * remains valid.  This condition is satisfied when called through
2158 2159
 * perf_event_for_each_child or perf_event_for_each as described
 * for perf_event_disable.
2160
 */
2161
void perf_event_enable(struct perf_event *event)
2162
{
2163
	struct perf_event_context *ctx = event->ctx;
2164 2165 2166 2167
	struct task_struct *task = ctx->task;

	if (!task) {
		/*
2168
		 * Enable the event on the cpu that it's on
2169
		 */
2170
		cpu_function_call(event->cpu, __perf_event_enable, event);
2171 2172 2173
		return;
	}

2174
	raw_spin_lock_irq(&ctx->lock);
2175
	if (event->state >= PERF_EVENT_STATE_INACTIVE)
2176 2177 2178
		goto out;

	/*
2179 2180
	 * If the event is in error state, clear that first.
	 * That way, if we see the event in error state below, we
2181 2182 2183 2184
	 * know that it has gone back into error state, as distinct
	 * from the task having been scheduled away before the
	 * cross-call arrived.
	 */
2185 2186
	if (event->state == PERF_EVENT_STATE_ERROR)
		event->state = PERF_EVENT_STATE_OFF;
2187

P
Peter Zijlstra 已提交
2188
retry:
2189
	if (!ctx->is_active) {
2190
		__perf_event_mark_enabled(event);
2191 2192 2193
		goto out;
	}

2194
	raw_spin_unlock_irq(&ctx->lock);
2195 2196 2197

	if (!task_function_call(task, __perf_event_enable, event))
		return;
2198

2199
	raw_spin_lock_irq(&ctx->lock);
2200 2201

	/*
2202
	 * If the context is active and the event is still off,
2203 2204
	 * we need to retry the cross-call.
	 */
2205 2206 2207 2208 2209 2210
	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;
2211
		goto retry;
2212
	}
2213

P
Peter Zijlstra 已提交
2214
out:
2215
	raw_spin_unlock_irq(&ctx->lock);
2216
}
2217
EXPORT_SYMBOL_GPL(perf_event_enable);
2218

2219
int perf_event_refresh(struct perf_event *event, int refresh)
2220
{
2221
	/*
2222
	 * not supported on inherited events
2223
	 */
2224
	if (event->attr.inherit || !is_sampling_event(event))
2225 2226
		return -EINVAL;

2227 2228
	atomic_add(refresh, &event->event_limit);
	perf_event_enable(event);
2229 2230

	return 0;
2231
}
2232
EXPORT_SYMBOL_GPL(perf_event_refresh);
2233

2234 2235 2236
static void ctx_sched_out(struct perf_event_context *ctx,
			  struct perf_cpu_context *cpuctx,
			  enum event_type_t event_type)
2237
{
2238
	struct perf_event *event;
2239
	int is_active = ctx->is_active;
2240

2241
	ctx->is_active &= ~event_type;
2242
	if (likely(!ctx->nr_events))
2243 2244
		return;

2245
	update_context_time(ctx);
S
Stephane Eranian 已提交
2246
	update_cgrp_time_from_cpuctx(cpuctx);
2247
	if (!ctx->nr_active)
2248
		return;
2249

P
Peter Zijlstra 已提交
2250
	perf_pmu_disable(ctx->pmu);
2251
	if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2252 2253
		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2254
	}
2255

2256
	if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2257
		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2258
			group_sched_out(event, cpuctx, ctx);
P
Peter Zijlstra 已提交
2259
	}
P
Peter Zijlstra 已提交
2260
	perf_pmu_enable(ctx->pmu);
2261 2262
}

2263
/*
2264 2265 2266 2267 2268 2269
 * 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().
2270
 */
2271 2272
static int context_equiv(struct perf_event_context *ctx1,
			 struct perf_event_context *ctx2)
2273
{
2274 2275 2276
	lockdep_assert_held(&ctx1->lock);
	lockdep_assert_held(&ctx2->lock);

2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298
	/* 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;
2299 2300
}

2301 2302
static void __perf_event_sync_stat(struct perf_event *event,
				     struct perf_event *next_event)
2303 2304 2305
{
	u64 value;

2306
	if (!event->attr.inherit_stat)
2307 2308 2309
		return;

	/*
2310
	 * Update the event value, we cannot use perf_event_read()
2311 2312
	 * because we're in the middle of a context switch and have IRQs
	 * disabled, which upsets smp_call_function_single(), however
2313
	 * we know the event must be on the current CPU, therefore we
2314 2315
	 * don't need to use it.
	 */
2316 2317
	switch (event->state) {
	case PERF_EVENT_STATE_ACTIVE:
2318 2319
		event->pmu->read(event);
		/* fall-through */
2320

2321 2322
	case PERF_EVENT_STATE_INACTIVE:
		update_event_times(event);
2323 2324 2325 2326 2327 2328 2329
		break;

	default:
		break;
	}

	/*
2330
	 * In order to keep per-task stats reliable we need to flip the event
2331 2332
	 * values when we flip the contexts.
	 */
2333 2334 2335
	value = local64_read(&next_event->count);
	value = local64_xchg(&event->count, value);
	local64_set(&next_event->count, value);
2336

2337 2338
	swap(event->total_time_enabled, next_event->total_time_enabled);
	swap(event->total_time_running, next_event->total_time_running);
2339

2340
	/*
2341
	 * Since we swizzled the values, update the user visible data too.
2342
	 */
2343 2344
	perf_event_update_userpage(event);
	perf_event_update_userpage(next_event);
2345 2346
}

2347 2348
static void perf_event_sync_stat(struct perf_event_context *ctx,
				   struct perf_event_context *next_ctx)
2349
{
2350
	struct perf_event *event, *next_event;
2351 2352 2353 2354

	if (!ctx->nr_stat)
		return;

2355 2356
	update_context_time(ctx);

2357 2358
	event = list_first_entry(&ctx->event_list,
				   struct perf_event, event_entry);
2359

2360 2361
	next_event = list_first_entry(&next_ctx->event_list,
					struct perf_event, event_entry);
2362

2363 2364
	while (&event->event_entry != &ctx->event_list &&
	       &next_event->event_entry != &next_ctx->event_list) {
2365

2366
		__perf_event_sync_stat(event, next_event);
2367

2368 2369
		event = list_next_entry(event, event_entry);
		next_event = list_next_entry(next_event, event_entry);
2370 2371 2372
	}
}

2373 2374
static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
					 struct task_struct *next)
T
Thomas Gleixner 已提交
2375
{
P
Peter Zijlstra 已提交
2376
	struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2377
	struct perf_event_context *next_ctx;
2378
	struct perf_event_context *parent, *next_parent;
P
Peter Zijlstra 已提交
2379
	struct perf_cpu_context *cpuctx;
2380
	int do_switch = 1;
T
Thomas Gleixner 已提交
2381

P
Peter Zijlstra 已提交
2382 2383
	if (likely(!ctx))
		return;
2384

P
Peter Zijlstra 已提交
2385 2386
	cpuctx = __get_cpu_context(ctx);
	if (!cpuctx->task_ctx)
T
Thomas Gleixner 已提交
2387 2388
		return;

2389
	rcu_read_lock();
P
Peter Zijlstra 已提交
2390
	next_ctx = next->perf_event_ctxp[ctxn];
2391 2392 2393 2394 2395 2396 2397
	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. */
2398
	if (!parent && !next_parent)
2399 2400 2401
		goto unlock;

	if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2402 2403 2404 2405 2406 2407 2408 2409 2410
		/*
		 * 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.
		 */
2411 2412
		raw_spin_lock(&ctx->lock);
		raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2413
		if (context_equiv(ctx, next_ctx)) {
2414 2415
			/*
			 * XXX do we need a memory barrier of sorts
2416
			 * wrt to rcu_dereference() of perf_event_ctxp
2417
			 */
P
Peter Zijlstra 已提交
2418 2419
			task->perf_event_ctxp[ctxn] = next_ctx;
			next->perf_event_ctxp[ctxn] = ctx;
2420 2421 2422
			ctx->task = next;
			next_ctx->task = task;
			do_switch = 0;
2423

2424
			perf_event_sync_stat(ctx, next_ctx);
2425
		}
2426 2427
		raw_spin_unlock(&next_ctx->lock);
		raw_spin_unlock(&ctx->lock);
2428
	}
2429
unlock:
2430
	rcu_read_unlock();
2431

2432
	if (do_switch) {
2433
		raw_spin_lock(&ctx->lock);
2434
		ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2435
		cpuctx->task_ctx = NULL;
2436
		raw_spin_unlock(&ctx->lock);
2437
	}
T
Thomas Gleixner 已提交
2438 2439
}

P
Peter Zijlstra 已提交
2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453
#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.
 */
2454 2455
void __perf_event_task_sched_out(struct task_struct *task,
				 struct task_struct *next)
P
Peter Zijlstra 已提交
2456 2457 2458 2459 2460
{
	int ctxn;

	for_each_task_context_nr(ctxn)
		perf_event_context_sched_out(task, ctxn, next);
S
Stephane Eranian 已提交
2461 2462 2463 2464 2465 2466

	/*
	 * 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
	 */
2467
	if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2468
		perf_cgroup_sched_out(task, next);
P
Peter Zijlstra 已提交
2469 2470
}

2471
static void task_ctx_sched_out(struct perf_event_context *ctx)
2472
{
P
Peter Zijlstra 已提交
2473
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2474

2475 2476
	if (!cpuctx->task_ctx)
		return;
2477 2478 2479 2480

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

2481
	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2482 2483 2484
	cpuctx->task_ctx = NULL;
}

2485 2486 2487 2488 2489 2490 2491
/*
 * 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);
2492 2493
}

2494
static void
2495
ctx_pinned_sched_in(struct perf_event_context *ctx,
2496
		    struct perf_cpu_context *cpuctx)
T
Thomas Gleixner 已提交
2497
{
2498
	struct perf_event *event;
T
Thomas Gleixner 已提交
2499

2500 2501
	list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
		if (event->state <= PERF_EVENT_STATE_OFF)
2502
			continue;
2503
		if (!event_filter_match(event))
2504 2505
			continue;

S
Stephane Eranian 已提交
2506 2507 2508 2509
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

2510
		if (group_can_go_on(event, cpuctx, 1))
2511
			group_sched_in(event, cpuctx, ctx);
2512 2513 2514 2515 2516

		/*
		 * If this pinned group hasn't been scheduled,
		 * put it in error state.
		 */
2517 2518 2519
		if (event->state == PERF_EVENT_STATE_INACTIVE) {
			update_group_times(event);
			event->state = PERF_EVENT_STATE_ERROR;
2520
		}
2521
	}
2522 2523 2524 2525
}

static void
ctx_flexible_sched_in(struct perf_event_context *ctx,
2526
		      struct perf_cpu_context *cpuctx)
2527 2528 2529
{
	struct perf_event *event;
	int can_add_hw = 1;
2530

2531 2532 2533
	list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
		/* Ignore events in OFF or ERROR state */
		if (event->state <= PERF_EVENT_STATE_OFF)
2534
			continue;
2535 2536
		/*
		 * Listen to the 'cpu' scheduling filter constraint
2537
		 * of events:
2538
		 */
2539
		if (!event_filter_match(event))
T
Thomas Gleixner 已提交
2540 2541
			continue;

S
Stephane Eranian 已提交
2542 2543 2544 2545
		/* may need to reset tstamp_enabled */
		if (is_cgroup_event(event))
			perf_cgroup_mark_enabled(event, ctx);

P
Peter Zijlstra 已提交
2546
		if (group_can_go_on(event, cpuctx, can_add_hw)) {
2547
			if (group_sched_in(event, cpuctx, ctx))
2548
				can_add_hw = 0;
P
Peter Zijlstra 已提交
2549
		}
T
Thomas Gleixner 已提交
2550
	}
2551 2552 2553 2554 2555
}

static void
ctx_sched_in(struct perf_event_context *ctx,
	     struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2556 2557
	     enum event_type_t event_type,
	     struct task_struct *task)
2558
{
S
Stephane Eranian 已提交
2559
	u64 now;
2560
	int is_active = ctx->is_active;
S
Stephane Eranian 已提交
2561

2562
	ctx->is_active |= event_type;
2563
	if (likely(!ctx->nr_events))
2564
		return;
2565

S
Stephane Eranian 已提交
2566 2567
	now = perf_clock();
	ctx->timestamp = now;
2568
	perf_cgroup_set_timestamp(task, ctx);
2569 2570 2571 2572
	/*
	 * First go through the list and put on any pinned groups
	 * in order to give them the best chance of going on.
	 */
2573
	if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2574
		ctx_pinned_sched_in(ctx, cpuctx);
2575 2576

	/* Then walk through the lower prio flexible groups */
2577
	if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2578
		ctx_flexible_sched_in(ctx, cpuctx);
2579 2580
}

2581
static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
S
Stephane Eranian 已提交
2582 2583
			     enum event_type_t event_type,
			     struct task_struct *task)
2584 2585 2586
{
	struct perf_event_context *ctx = &cpuctx->ctx;

S
Stephane Eranian 已提交
2587
	ctx_sched_in(ctx, cpuctx, event_type, task);
2588 2589
}

S
Stephane Eranian 已提交
2590 2591
static void perf_event_context_sched_in(struct perf_event_context *ctx,
					struct task_struct *task)
2592
{
P
Peter Zijlstra 已提交
2593
	struct perf_cpu_context *cpuctx;
2594

P
Peter Zijlstra 已提交
2595
	cpuctx = __get_cpu_context(ctx);
2596 2597 2598
	if (cpuctx->task_ctx == ctx)
		return;

2599
	perf_ctx_lock(cpuctx, ctx);
P
Peter Zijlstra 已提交
2600
	perf_pmu_disable(ctx->pmu);
2601 2602 2603 2604 2605 2606 2607
	/*
	 * 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);

2608 2609
	if (ctx->nr_events)
		cpuctx->task_ctx = ctx;
2610

2611 2612
	perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);

2613 2614 2615
	perf_pmu_enable(ctx->pmu);
	perf_ctx_unlock(cpuctx, ctx);

2616 2617 2618 2619
	/*
	 * 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
2620
	perf_pmu_rotate_start(ctx->pmu);
2621 2622
}

2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680
/*
 * 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
2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691
/*
 * 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.
 */
2692 2693
void __perf_event_task_sched_in(struct task_struct *prev,
				struct task_struct *task)
P
Peter Zijlstra 已提交
2694 2695 2696 2697 2698 2699 2700 2701 2702
{
	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
2703
		perf_event_context_sched_in(ctx, task);
P
Peter Zijlstra 已提交
2704
	}
S
Stephane Eranian 已提交
2705 2706 2707 2708 2709
	/*
	 * 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
	 */
2710
	if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2711
		perf_cgroup_sched_in(prev, task);
2712 2713

	/* check for system-wide branch_stack events */
2714
	if (atomic_read(this_cpu_ptr(&perf_branch_stack_events)))
2715
		perf_branch_stack_sched_in(prev, task);
2716 2717
}

2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744
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.
	 */
2745
#define REDUCE_FLS(a, b)		\
2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784
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;
	}

2785 2786 2787
	if (!divisor)
		return dividend;

2788 2789 2790
	return div64_u64(dividend, divisor);
}

2791 2792 2793
static DEFINE_PER_CPU(int, perf_throttled_count);
static DEFINE_PER_CPU(u64, perf_throttled_seq);

2794
static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2795
{
2796
	struct hw_perf_event *hwc = &event->hw;
2797
	s64 period, sample_period;
2798 2799
	s64 delta;

2800
	period = perf_calculate_period(event, nsec, count);
2801 2802 2803 2804 2805 2806 2807 2808 2809 2810

	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;
2811

2812
	if (local64_read(&hwc->period_left) > 8*sample_period) {
2813 2814 2815
		if (disable)
			event->pmu->stop(event, PERF_EF_UPDATE);

2816
		local64_set(&hwc->period_left, 0);
2817 2818 2819

		if (disable)
			event->pmu->start(event, PERF_EF_RELOAD);
2820
	}
2821 2822
}

2823 2824 2825 2826 2827 2828 2829
/*
 * 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)
2830
{
2831 2832
	struct perf_event *event;
	struct hw_perf_event *hwc;
2833
	u64 now, period = TICK_NSEC;
2834
	s64 delta;
2835

2836 2837 2838 2839 2840 2841
	/*
	 * 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))
2842 2843
		return;

2844
	raw_spin_lock(&ctx->lock);
2845
	perf_pmu_disable(ctx->pmu);
2846

2847
	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2848
		if (event->state != PERF_EVENT_STATE_ACTIVE)
2849 2850
			continue;

2851
		if (!event_filter_match(event))
2852 2853
			continue;

2854 2855
		perf_pmu_disable(event->pmu);

2856
		hwc = &event->hw;
2857

2858
		if (hwc->interrupts == MAX_INTERRUPTS) {
2859
			hwc->interrupts = 0;
2860
			perf_log_throttle(event, 1);
P
Peter Zijlstra 已提交
2861
			event->pmu->start(event, 0);
2862 2863
		}

2864
		if (!event->attr.freq || !event->attr.sample_freq)
2865
			goto next;
2866

2867 2868 2869 2870 2871
		/*
		 * stop the event and update event->count
		 */
		event->pmu->stop(event, PERF_EF_UPDATE);

2872
		now = local64_read(&event->count);
2873 2874
		delta = now - hwc->freq_count_stamp;
		hwc->freq_count_stamp = now;
2875

2876 2877 2878
		/*
		 * restart the event
		 * reload only if value has changed
2879 2880 2881
		 * we have stopped the event so tell that
		 * to perf_adjust_period() to avoid stopping it
		 * twice.
2882
		 */
2883
		if (delta > 0)
2884
			perf_adjust_period(event, period, delta, false);
2885 2886

		event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2887 2888
	next:
		perf_pmu_enable(event->pmu);
2889
	}
2890

2891
	perf_pmu_enable(ctx->pmu);
2892
	raw_spin_unlock(&ctx->lock);
2893 2894
}

2895
/*
2896
 * Round-robin a context's events:
2897
 */
2898
static void rotate_ctx(struct perf_event_context *ctx)
T
Thomas Gleixner 已提交
2899
{
2900 2901 2902 2903 2904 2905
	/*
	 * 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);
2906 2907
}

2908
/*
2909 2910 2911
 * 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.
2912
 */
2913
static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2914
{
P
Peter Zijlstra 已提交
2915
	struct perf_event_context *ctx = NULL;
2916
	int rotate = 0, remove = 1;
2917

2918
	if (cpuctx->ctx.nr_events) {
2919
		remove = 0;
2920 2921 2922
		if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
			rotate = 1;
	}
2923

P
Peter Zijlstra 已提交
2924
	ctx = cpuctx->task_ctx;
2925
	if (ctx && ctx->nr_events) {
2926
		remove = 0;
2927 2928 2929
		if (ctx->nr_events != ctx->nr_active)
			rotate = 1;
	}
2930

2931
	if (!rotate)
2932 2933
		goto done;

2934
	perf_ctx_lock(cpuctx, cpuctx->task_ctx);
P
Peter Zijlstra 已提交
2935
	perf_pmu_disable(cpuctx->ctx.pmu);
2936

2937 2938 2939
	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
	if (ctx)
		ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
T
Thomas Gleixner 已提交
2940

2941 2942 2943
	rotate_ctx(&cpuctx->ctx);
	if (ctx)
		rotate_ctx(ctx);
2944

2945
	perf_event_sched_in(cpuctx, ctx, current);
2946

2947 2948
	perf_pmu_enable(cpuctx->ctx.pmu);
	perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2949
done:
2950 2951
	if (remove)
		list_del_init(&cpuctx->rotation_list);
2952 2953

	return rotate;
2954 2955
}

2956 2957 2958
#ifdef CONFIG_NO_HZ_FULL
bool perf_event_can_stop_tick(void)
{
2959
	if (atomic_read(&nr_freq_events) ||
2960
	    __this_cpu_read(perf_throttled_count))
2961
		return false;
2962 2963
	else
		return true;
2964 2965 2966
}
#endif

2967 2968
void perf_event_task_tick(void)
{
2969
	struct list_head *head = this_cpu_ptr(&rotation_list);
2970
	struct perf_cpu_context *cpuctx, *tmp;
2971 2972
	struct perf_event_context *ctx;
	int throttled;
2973

2974 2975
	WARN_ON(!irqs_disabled());

2976 2977 2978
	__this_cpu_inc(perf_throttled_seq);
	throttled = __this_cpu_xchg(perf_throttled_count, 0);

2979
	list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2980 2981 2982 2983 2984 2985
		ctx = &cpuctx->ctx;
		perf_adjust_freq_unthr_context(ctx, throttled);

		ctx = cpuctx->task_ctx;
		if (ctx)
			perf_adjust_freq_unthr_context(ctx, throttled);
2986
	}
T
Thomas Gleixner 已提交
2987 2988
}

2989 2990 2991 2992 2993 2994 2995 2996 2997 2998
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;

2999
	__perf_event_mark_enabled(event);
3000 3001 3002 3003

	return 1;
}

3004
/*
3005
 * Enable all of a task's events that have been marked enable-on-exec.
3006 3007
 * This expects task == current.
 */
P
Peter Zijlstra 已提交
3008
static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3009
{
3010
	struct perf_event_context *clone_ctx = NULL;
3011
	struct perf_event *event;
3012 3013
	unsigned long flags;
	int enabled = 0;
3014
	int ret;
3015 3016

	local_irq_save(flags);
3017
	if (!ctx || !ctx->nr_events)
3018 3019
		goto out;

3020 3021 3022 3023 3024 3025 3026
	/*
	 * 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.
	 */
3027
	perf_cgroup_sched_out(current, NULL);
3028

3029
	raw_spin_lock(&ctx->lock);
3030
	task_ctx_sched_out(ctx);
3031

3032
	list_for_each_entry(event, &ctx->event_list, event_entry) {
3033 3034 3035
		ret = event_enable_on_exec(event, ctx);
		if (ret)
			enabled = 1;
3036 3037 3038
	}

	/*
3039
	 * Unclone this context if we enabled any event.
3040
	 */
3041
	if (enabled)
3042
		clone_ctx = unclone_ctx(ctx);
3043

3044
	raw_spin_unlock(&ctx->lock);
3045

3046 3047 3048
	/*
	 * Also calls ctxswin for cgroup events, if any:
	 */
S
Stephane Eranian 已提交
3049
	perf_event_context_sched_in(ctx, ctx->task);
P
Peter Zijlstra 已提交
3050
out:
3051
	local_irq_restore(flags);
3052 3053 3054

	if (clone_ctx)
		put_ctx(clone_ctx);
3055 3056
}

3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072
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
3073
/*
3074
 * Cross CPU call to read the hardware event
T
Thomas Gleixner 已提交
3075
 */
3076
static void __perf_event_read(void *info)
T
Thomas Gleixner 已提交
3077
{
3078 3079
	struct perf_event *event = info;
	struct perf_event_context *ctx = event->ctx;
P
Peter Zijlstra 已提交
3080
	struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
I
Ingo Molnar 已提交
3081

3082 3083 3084 3085
	/*
	 * 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
3086 3087
	 * event->count would have been updated to a recent sample
	 * when the event was scheduled out.
3088 3089 3090 3091
	 */
	if (ctx->task && cpuctx->task_ctx != ctx)
		return;

3092
	raw_spin_lock(&ctx->lock);
S
Stephane Eranian 已提交
3093
	if (ctx->is_active) {
3094
		update_context_time(ctx);
S
Stephane Eranian 已提交
3095 3096
		update_cgrp_time_from_event(event);
	}
3097
	update_event_times(event);
3098 3099
	if (event->state == PERF_EVENT_STATE_ACTIVE)
		event->pmu->read(event);
3100
	raw_spin_unlock(&ctx->lock);
T
Thomas Gleixner 已提交
3101 3102
}

P
Peter Zijlstra 已提交
3103 3104
static inline u64 perf_event_count(struct perf_event *event)
{
3105
	return local64_read(&event->count) + atomic64_read(&event->child_count);
P
Peter Zijlstra 已提交
3106 3107
}

3108
static u64 perf_event_read(struct perf_event *event)
T
Thomas Gleixner 已提交
3109 3110
{
	/*
3111 3112
	 * If event is enabled and currently active on a CPU, update the
	 * value in the event structure:
T
Thomas Gleixner 已提交
3113
	 */
3114 3115 3116 3117
	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
3118 3119 3120
		struct perf_event_context *ctx = event->ctx;
		unsigned long flags;

3121
		raw_spin_lock_irqsave(&ctx->lock, flags);
3122 3123 3124 3125 3126
		/*
		 * 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
3127
		if (ctx->is_active) {
3128
			update_context_time(ctx);
S
Stephane Eranian 已提交
3129 3130
			update_cgrp_time_from_event(event);
		}
3131
		update_event_times(event);
3132
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
T
Thomas Gleixner 已提交
3133 3134
	}

P
Peter Zijlstra 已提交
3135
	return perf_event_count(event);
T
Thomas Gleixner 已提交
3136 3137
}

3138
/*
3139
 * Initialize the perf_event context in a task_struct:
3140
 */
3141
static void __perf_event_init_context(struct perf_event_context *ctx)
3142
{
3143
	raw_spin_lock_init(&ctx->lock);
3144
	mutex_init(&ctx->mutex);
3145 3146
	INIT_LIST_HEAD(&ctx->pinned_groups);
	INIT_LIST_HEAD(&ctx->flexible_groups);
3147 3148
	INIT_LIST_HEAD(&ctx->event_list);
	atomic_set(&ctx->refcount, 1);
3149
	INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164
}

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
3165
	}
3166 3167 3168
	ctx->pmu = pmu;

	return ctx;
3169 3170
}

3171 3172 3173 3174 3175
static struct task_struct *
find_lively_task_by_vpid(pid_t vpid)
{
	struct task_struct *task;
	int err;
T
Thomas Gleixner 已提交
3176 3177

	rcu_read_lock();
3178
	if (!vpid)
T
Thomas Gleixner 已提交
3179 3180
		task = current;
	else
3181
		task = find_task_by_vpid(vpid);
T
Thomas Gleixner 已提交
3182 3183 3184 3185 3186 3187 3188 3189
	if (task)
		get_task_struct(task);
	rcu_read_unlock();

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

	/* Reuse ptrace permission checks for now. */
3190 3191 3192 3193
	err = -EACCES;
	if (!ptrace_may_access(task, PTRACE_MODE_READ))
		goto errout;

3194 3195 3196 3197 3198 3199 3200
	return task;
errout:
	put_task_struct(task);
	return ERR_PTR(err);

}

3201 3202 3203
/*
 * Returns a matching context with refcount and pincount.
 */
P
Peter Zijlstra 已提交
3204
static struct perf_event_context *
M
Matt Helsley 已提交
3205
find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
T
Thomas Gleixner 已提交
3206
{
3207
	struct perf_event_context *ctx, *clone_ctx = NULL;
3208
	struct perf_cpu_context *cpuctx;
3209
	unsigned long flags;
P
Peter Zijlstra 已提交
3210
	int ctxn, err;
T
Thomas Gleixner 已提交
3211

3212
	if (!task) {
3213
		/* Must be root to operate on a CPU event: */
3214
		if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
T
Thomas Gleixner 已提交
3215 3216 3217
			return ERR_PTR(-EACCES);

		/*
3218
		 * We could be clever and allow to attach a event to an
T
Thomas Gleixner 已提交
3219 3220 3221
		 * offline CPU and activate it when the CPU comes up, but
		 * that's for later.
		 */
3222
		if (!cpu_online(cpu))
T
Thomas Gleixner 已提交
3223 3224
			return ERR_PTR(-ENODEV);

P
Peter Zijlstra 已提交
3225
		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
T
Thomas Gleixner 已提交
3226
		ctx = &cpuctx->ctx;
3227
		get_ctx(ctx);
3228
		++ctx->pin_count;
T
Thomas Gleixner 已提交
3229 3230 3231 3232

		return ctx;
	}

P
Peter Zijlstra 已提交
3233 3234 3235 3236 3237
	err = -EINVAL;
	ctxn = pmu->task_ctx_nr;
	if (ctxn < 0)
		goto errout;

P
Peter Zijlstra 已提交
3238
retry:
P
Peter Zijlstra 已提交
3239
	ctx = perf_lock_task_context(task, ctxn, &flags);
3240
	if (ctx) {
3241
		clone_ctx = unclone_ctx(ctx);
3242
		++ctx->pin_count;
3243
		raw_spin_unlock_irqrestore(&ctx->lock, flags);
3244 3245 3246

		if (clone_ctx)
			put_ctx(clone_ctx);
3247
	} else {
3248
		ctx = alloc_perf_context(pmu, task);
3249 3250 3251
		err = -ENOMEM;
		if (!ctx)
			goto errout;
3252

3253 3254 3255 3256 3257 3258 3259 3260 3261 3262
		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;
3263
		else {
3264
			get_ctx(ctx);
3265
			++ctx->pin_count;
3266
			rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3267
		}
3268 3269 3270
		mutex_unlock(&task->perf_event_mutex);

		if (unlikely(err)) {
3271
			put_ctx(ctx);
3272 3273 3274 3275

			if (err == -EAGAIN)
				goto retry;
			goto errout;
3276 3277 3278
		}
	}

T
Thomas Gleixner 已提交
3279
	return ctx;
3280

P
Peter Zijlstra 已提交
3281
errout:
3282
	return ERR_PTR(err);
T
Thomas Gleixner 已提交
3283 3284
}

L
Li Zefan 已提交
3285 3286
static void perf_event_free_filter(struct perf_event *event);

3287
static void free_event_rcu(struct rcu_head *head)
P
Peter Zijlstra 已提交
3288
{
3289
	struct perf_event *event;
P
Peter Zijlstra 已提交
3290

3291 3292 3293
	event = container_of(head, struct perf_event, rcu_head);
	if (event->ns)
		put_pid_ns(event->ns);
L
Li Zefan 已提交
3294
	perf_event_free_filter(event);
3295
	kfree(event);
P
Peter Zijlstra 已提交
3296 3297
}

3298
static void ring_buffer_put(struct ring_buffer *rb);
3299 3300
static void ring_buffer_attach(struct perf_event *event,
			       struct ring_buffer *rb);
3301

3302
static void unaccount_event_cpu(struct perf_event *event, int cpu)
3303
{
3304 3305 3306 3307 3308 3309 3310 3311 3312 3313
	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));
}
3314

3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327
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);
3328 3329
	if (event->attr.freq)
		atomic_dec(&nr_freq_events);
3330 3331 3332 3333 3334 3335 3336
	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);
}
3337

3338 3339
static void __free_event(struct perf_event *event)
{
3340
	if (!event->parent) {
3341 3342
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
			put_callchain_buffers();
3343
	}
3344

3345 3346 3347 3348 3349 3350
	if (event->destroy)
		event->destroy(event);

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

3351 3352 3353
	if (event->pmu)
		module_put(event->pmu->module);

3354 3355
	call_rcu(&event->rcu_head, free_event_rcu);
}
P
Peter Zijlstra 已提交
3356 3357

static void _free_event(struct perf_event *event)
3358
{
3359
	irq_work_sync(&event->pending);
3360

3361
	unaccount_event(event);
3362

3363
	if (event->rb) {
3364 3365 3366 3367 3368 3369 3370
		/*
		 * 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);
3371
		ring_buffer_attach(event, NULL);
3372
		mutex_unlock(&event->mmap_mutex);
3373 3374
	}

S
Stephane Eranian 已提交
3375 3376 3377
	if (is_cgroup_event(event))
		perf_detach_cgroup(event);

3378
	__free_event(event);
3379 3380
}

P
Peter Zijlstra 已提交
3381 3382 3383 3384 3385
/*
 * 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
3386
{
P
Peter Zijlstra 已提交
3387 3388 3389 3390 3391 3392
	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
3393

P
Peter Zijlstra 已提交
3394
	_free_event(event);
T
Thomas Gleixner 已提交
3395 3396
}

3397
/*
3398
 * Remove user event from the owner task.
3399
 */
3400
static void perf_remove_from_owner(struct perf_event *event)
3401
{
P
Peter Zijlstra 已提交
3402
	struct task_struct *owner;
3403

P
Peter Zijlstra 已提交
3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435
	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);
	}
3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449
}

/*
 * Called when the last reference to the file is gone.
 */
static void put_event(struct perf_event *event)
{
	struct perf_event_context *ctx = event->ctx;

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

	if (!is_kernel_event(event))
		perf_remove_from_owner(event);
P
Peter Zijlstra 已提交
3450

P
Peter Zijlstra 已提交
3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468
	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);
3469 3470
}

P
Peter Zijlstra 已提交
3471 3472 3473 3474 3475 3476 3477
int perf_event_release_kernel(struct perf_event *event)
{
	put_event(event);
	return 0;
}
EXPORT_SYMBOL_GPL(perf_event_release_kernel);

3478 3479 3480 3481
static int perf_release(struct inode *inode, struct file *file)
{
	put_event(file->private_data);
	return 0;
3482 3483
}

3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519
/*
 * Remove all orphanes events from the context.
 */
static void orphans_remove_work(struct work_struct *work)
{
	struct perf_event_context *ctx;
	struct perf_event *event, *tmp;

	ctx = container_of(work, struct perf_event_context,
			   orphans_remove.work);

	mutex_lock(&ctx->mutex);
	list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
		struct perf_event *parent_event = event->parent;

		if (!is_orphaned_child(event))
			continue;

		perf_remove_from_context(event, true);

		mutex_lock(&parent_event->child_mutex);
		list_del_init(&event->child_list);
		mutex_unlock(&parent_event->child_mutex);

		free_event(event);
		put_event(parent_event);
	}

	raw_spin_lock_irq(&ctx->lock);
	ctx->orphans_remove_sched = false;
	raw_spin_unlock_irq(&ctx->lock);
	mutex_unlock(&ctx->mutex);

	put_ctx(ctx);
}

3520
u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3521
{
3522
	struct perf_event *child;
3523 3524
	u64 total = 0;

3525 3526 3527
	*enabled = 0;
	*running = 0;

3528
	mutex_lock(&event->child_mutex);
3529
	total += perf_event_read(event);
3530 3531 3532 3533 3534 3535
	*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) {
3536
		total += perf_event_read(child);
3537 3538 3539
		*enabled += child->total_time_enabled;
		*running += child->total_time_running;
	}
3540
	mutex_unlock(&event->child_mutex);
3541 3542 3543

	return total;
}
3544
EXPORT_SYMBOL_GPL(perf_event_read_value);
3545

3546
static int perf_event_read_group(struct perf_event *event,
3547 3548
				   u64 read_format, char __user *buf)
{
3549
	struct perf_event *leader = event->group_leader, *sub;
3550 3551
	int n = 0, size = 0, ret = -EFAULT;
	struct perf_event_context *ctx = leader->ctx;
3552
	u64 values[5];
3553
	u64 count, enabled, running;
3554

3555
	mutex_lock(&ctx->mutex);
3556
	count = perf_event_read_value(leader, &enabled, &running);
3557 3558

	values[n++] = 1 + leader->nr_siblings;
3559 3560 3561 3562
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
		values[n++] = enabled;
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
		values[n++] = running;
3563 3564 3565
	values[n++] = count;
	if (read_format & PERF_FORMAT_ID)
		values[n++] = primary_event_id(leader);
3566 3567 3568 3569

	size = n * sizeof(u64);

	if (copy_to_user(buf, values, size))
3570
		goto unlock;
3571

3572
	ret = size;
3573

3574
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3575
		n = 0;
3576

3577
		values[n++] = perf_event_read_value(sub, &enabled, &running);
3578 3579 3580 3581 3582
		if (read_format & PERF_FORMAT_ID)
			values[n++] = primary_event_id(sub);

		size = n * sizeof(u64);

3583
		if (copy_to_user(buf + ret, values, size)) {
3584 3585 3586
			ret = -EFAULT;
			goto unlock;
		}
3587 3588

		ret += size;
3589
	}
3590 3591
unlock:
	mutex_unlock(&ctx->mutex);
3592

3593
	return ret;
3594 3595
}

3596
static int perf_event_read_one(struct perf_event *event,
3597 3598
				 u64 read_format, char __user *buf)
{
3599
	u64 enabled, running;
3600 3601 3602
	u64 values[4];
	int n = 0;

3603 3604 3605 3606 3607
	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;
3608
	if (read_format & PERF_FORMAT_ID)
3609
		values[n++] = primary_event_id(event);
3610 3611 3612 3613 3614 3615 3616

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

	return n * sizeof(u64);
}

3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629
static bool is_event_hup(struct perf_event *event)
{
	bool no_children;

	if (event->state != PERF_EVENT_STATE_EXIT)
		return false;

	mutex_lock(&event->child_mutex);
	no_children = list_empty(&event->child_list);
	mutex_unlock(&event->child_mutex);
	return no_children;
}

T
Thomas Gleixner 已提交
3630
/*
3631
 * Read the performance event - simple non blocking version for now
T
Thomas Gleixner 已提交
3632 3633
 */
static ssize_t
3634
perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
T
Thomas Gleixner 已提交
3635
{
3636
	u64 read_format = event->attr.read_format;
3637
	int ret;
T
Thomas Gleixner 已提交
3638

3639
	/*
3640
	 * Return end-of-file for a read on a event that is in
3641 3642 3643
	 * error state (i.e. because it was pinned but it couldn't be
	 * scheduled on to the CPU at some point).
	 */
3644
	if (event->state == PERF_EVENT_STATE_ERROR)
3645 3646
		return 0;

3647
	if (count < event->read_size)
3648 3649
		return -ENOSPC;

3650
	WARN_ON_ONCE(event->ctx->parent_ctx);
3651
	if (read_format & PERF_FORMAT_GROUP)
3652
		ret = perf_event_read_group(event, read_format, buf);
3653
	else
3654
		ret = perf_event_read_one(event, read_format, buf);
T
Thomas Gleixner 已提交
3655

3656
	return ret;
T
Thomas Gleixner 已提交
3657 3658 3659 3660 3661
}

static ssize_t
perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
3662
	struct perf_event *event = file->private_data;
T
Thomas Gleixner 已提交
3663

3664
	return perf_read_hw(event, buf, count);
T
Thomas Gleixner 已提交
3665 3666 3667 3668
}

static unsigned int perf_poll(struct file *file, poll_table *wait)
{
3669
	struct perf_event *event = file->private_data;
3670
	struct ring_buffer *rb;
3671
	unsigned int events = POLLHUP;
P
Peter Zijlstra 已提交
3672

3673
	poll_wait(file, &event->waitq, wait);
3674

3675
	if (is_event_hup(event))
3676
		return events;
P
Peter Zijlstra 已提交
3677

3678
	/*
3679 3680
	 * Pin the event->rb by taking event->mmap_mutex; otherwise
	 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3681 3682
	 */
	mutex_lock(&event->mmap_mutex);
3683 3684
	rb = event->rb;
	if (rb)
3685
		events = atomic_xchg(&rb->poll, 0);
3686
	mutex_unlock(&event->mmap_mutex);
T
Thomas Gleixner 已提交
3687 3688 3689
	return events;
}

3690
static void perf_event_reset(struct perf_event *event)
3691
{
3692
	(void)perf_event_read(event);
3693
	local64_set(&event->count, 0);
3694
	perf_event_update_userpage(event);
P
Peter Zijlstra 已提交
3695 3696
}

3697
/*
3698 3699 3700 3701
 * 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.
3702
 */
3703 3704
static void perf_event_for_each_child(struct perf_event *event,
					void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3705
{
3706
	struct perf_event *child;
P
Peter Zijlstra 已提交
3707

3708 3709 3710 3711
	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
3712
		func(child);
3713
	mutex_unlock(&event->child_mutex);
P
Peter Zijlstra 已提交
3714 3715
}

3716 3717
static void perf_event_for_each(struct perf_event *event,
				  void (*func)(struct perf_event *))
P
Peter Zijlstra 已提交
3718
{
3719 3720
	struct perf_event_context *ctx = event->ctx;
	struct perf_event *sibling;
P
Peter Zijlstra 已提交
3721

3722 3723
	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
3724
	event = event->group_leader;
3725

3726 3727
	perf_event_for_each_child(event, func);
	list_for_each_entry(sibling, &event->sibling_list, group_entry)
3728
		perf_event_for_each_child(sibling, func);
3729
	mutex_unlock(&ctx->mutex);
3730 3731
}

3732
static int perf_event_period(struct perf_event *event, u64 __user *arg)
3733
{
3734
	struct perf_event_context *ctx = event->ctx;
3735
	int ret = 0, active;
3736 3737
	u64 value;

3738
	if (!is_sampling_event(event))
3739 3740
		return -EINVAL;

3741
	if (copy_from_user(&value, arg, sizeof(value)))
3742 3743 3744 3745 3746
		return -EFAULT;

	if (!value)
		return -EINVAL;

3747
	raw_spin_lock_irq(&ctx->lock);
3748 3749
	if (event->attr.freq) {
		if (value > sysctl_perf_event_sample_rate) {
3750 3751 3752 3753
			ret = -EINVAL;
			goto unlock;
		}

3754
		event->attr.sample_freq = value;
3755
	} else {
3756 3757
		event->attr.sample_period = value;
		event->hw.sample_period = value;
3758
	}
3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772

	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);
	}

3773
unlock:
3774
	raw_spin_unlock_irq(&ctx->lock);
3775 3776 3777 3778

	return ret;
}

3779 3780
static const struct file_operations perf_fops;

3781
static inline int perf_fget_light(int fd, struct fd *p)
3782
{
3783 3784 3785
	struct fd f = fdget(fd);
	if (!f.file)
		return -EBADF;
3786

3787 3788 3789
	if (f.file->f_op != &perf_fops) {
		fdput(f);
		return -EBADF;
3790
	}
3791 3792
	*p = f;
	return 0;
3793 3794 3795 3796
}

static int perf_event_set_output(struct perf_event *event,
				 struct perf_event *output_event);
L
Li Zefan 已提交
3797
static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3798

3799 3800
static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
3801 3802
	struct perf_event *event = file->private_data;
	void (*func)(struct perf_event *);
P
Peter Zijlstra 已提交
3803
	u32 flags = arg;
3804 3805

	switch (cmd) {
3806 3807
	case PERF_EVENT_IOC_ENABLE:
		func = perf_event_enable;
3808
		break;
3809 3810
	case PERF_EVENT_IOC_DISABLE:
		func = perf_event_disable;
3811
		break;
3812 3813
	case PERF_EVENT_IOC_RESET:
		func = perf_event_reset;
3814
		break;
P
Peter Zijlstra 已提交
3815

3816 3817
	case PERF_EVENT_IOC_REFRESH:
		return perf_event_refresh(event, arg);
3818

3819 3820
	case PERF_EVENT_IOC_PERIOD:
		return perf_event_period(event, (u64 __user *)arg);
3821

3822 3823 3824 3825 3826 3827 3828 3829 3830
	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;
	}

3831
	case PERF_EVENT_IOC_SET_OUTPUT:
3832 3833 3834
	{
		int ret;
		if (arg != -1) {
3835 3836 3837 3838 3839 3840 3841 3842 3843 3844
			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);
3845 3846 3847
		}
		return ret;
	}
3848

L
Li Zefan 已提交
3849 3850 3851
	case PERF_EVENT_IOC_SET_FILTER:
		return perf_event_set_filter(event, (void __user *)arg);

3852
	default:
P
Peter Zijlstra 已提交
3853
		return -ENOTTY;
3854
	}
P
Peter Zijlstra 已提交
3855 3856

	if (flags & PERF_IOC_FLAG_GROUP)
3857
		perf_event_for_each(event, func);
P
Peter Zijlstra 已提交
3858
	else
3859
		perf_event_for_each_child(event, func);
P
Peter Zijlstra 已提交
3860 3861

	return 0;
3862 3863
}

P
Pawel Moll 已提交
3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883
#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

3884
int perf_event_task_enable(void)
3885
{
3886
	struct perf_event *event;
3887

3888 3889 3890 3891
	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);
3892 3893 3894 3895

	return 0;
}

3896
int perf_event_task_disable(void)
3897
{
3898
	struct perf_event *event;
3899

3900 3901 3902 3903
	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);
3904 3905 3906 3907

	return 0;
}

3908
static int perf_event_index(struct perf_event *event)
3909
{
P
Peter Zijlstra 已提交
3910 3911 3912
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;

3913
	if (event->state != PERF_EVENT_STATE_ACTIVE)
3914 3915
		return 0;

3916
	return event->pmu->event_idx(event);
3917 3918
}

3919
static void calc_timer_values(struct perf_event *event,
3920
				u64 *now,
3921 3922
				u64 *enabled,
				u64 *running)
3923
{
3924
	u64 ctx_time;
3925

3926 3927
	*now = perf_clock();
	ctx_time = event->shadow_ctx_time + *now;
3928 3929 3930 3931
	*enabled = ctx_time - event->tstamp_enabled;
	*running = ctx_time - event->tstamp_running;
}

3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951
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();
}

3952
void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3953 3954 3955
{
}

3956 3957 3958 3959 3960
/*
 * 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.
 */
3961
void perf_event_update_userpage(struct perf_event *event)
3962
{
3963
	struct perf_event_mmap_page *userpg;
3964
	struct ring_buffer *rb;
3965
	u64 enabled, running, now;
3966 3967

	rcu_read_lock();
3968 3969 3970 3971
	rb = rcu_dereference(event->rb);
	if (!rb)
		goto unlock;

3972 3973 3974 3975 3976 3977 3978 3979 3980
	/*
	 * 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
	 */
3981
	calc_timer_values(event, &now, &enabled, &running);
3982

3983
	userpg = rb->user_page;
3984 3985 3986 3987 3988
	/*
	 * Disable preemption so as to not let the corresponding user-space
	 * spin too long if we get preempted.
	 */
	preempt_disable();
3989
	++userpg->lock;
3990
	barrier();
3991
	userpg->index = perf_event_index(event);
P
Peter Zijlstra 已提交
3992
	userpg->offset = perf_event_count(event);
3993
	if (userpg->index)
3994
		userpg->offset -= local64_read(&event->hw.prev_count);
3995

3996
	userpg->time_enabled = enabled +
3997
			atomic64_read(&event->child_total_time_enabled);
3998

3999
	userpg->time_running = running +
4000
			atomic64_read(&event->child_total_time_running);
4001

4002
	arch_perf_update_userpage(userpg, now);
4003

4004
	barrier();
4005
	++userpg->lock;
4006
	preempt_enable();
4007
unlock:
4008
	rcu_read_unlock();
4009 4010
}

4011 4012 4013
static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
{
	struct perf_event *event = vma->vm_file->private_data;
4014
	struct ring_buffer *rb;
4015 4016 4017 4018 4019 4020 4021 4022 4023
	int ret = VM_FAULT_SIGBUS;

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

	rcu_read_lock();
4024 4025
	rb = rcu_dereference(event->rb);
	if (!rb)
4026 4027 4028 4029 4030
		goto unlock;

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

4031
	vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045
	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;
}

4046 4047 4048
static void ring_buffer_attach(struct perf_event *event,
			       struct ring_buffer *rb)
{
4049
	struct ring_buffer *old_rb = NULL;
4050 4051
	unsigned long flags;

4052 4053 4054 4055 4056 4057
	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);
4058

4059 4060 4061
		old_rb = event->rb;
		event->rcu_batches = get_state_synchronize_rcu();
		event->rcu_pending = 1;
4062

4063 4064 4065 4066
		spin_lock_irqsave(&old_rb->event_lock, flags);
		list_del_rcu(&event->rb_entry);
		spin_unlock_irqrestore(&old_rb->event_lock, flags);
	}
4067

4068 4069 4070 4071
	if (event->rcu_pending && rb) {
		cond_synchronize_rcu(event->rcu_batches);
		event->rcu_pending = 0;
	}
4072

4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089
	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);
	}
4090 4091 4092 4093 4094 4095 4096 4097
}

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

	rcu_read_lock();
	rb = rcu_dereference(event->rb);
4098 4099 4100 4101
	if (rb) {
		list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
			wake_up_all(&event->waitq);
	}
4102 4103 4104
	rcu_read_unlock();
}

4105
static void rb_free_rcu(struct rcu_head *rcu_head)
4106
{
4107
	struct ring_buffer *rb;
4108

4109 4110
	rb = container_of(rcu_head, struct ring_buffer, rcu_head);
	rb_free(rb);
4111 4112
}

4113
static struct ring_buffer *ring_buffer_get(struct perf_event *event)
4114
{
4115
	struct ring_buffer *rb;
4116

4117
	rcu_read_lock();
4118 4119 4120 4121
	rb = rcu_dereference(event->rb);
	if (rb) {
		if (!atomic_inc_not_zero(&rb->refcount))
			rb = NULL;
4122 4123 4124
	}
	rcu_read_unlock();

4125
	return rb;
4126 4127
}

4128
static void ring_buffer_put(struct ring_buffer *rb)
4129
{
4130
	if (!atomic_dec_and_test(&rb->refcount))
4131
		return;
4132

4133
	WARN_ON_ONCE(!list_empty(&rb->event_list));
4134

4135
	call_rcu(&rb->rcu_head, rb_free_rcu);
4136 4137 4138 4139
}

static void perf_mmap_open(struct vm_area_struct *vma)
{
4140
	struct perf_event *event = vma->vm_file->private_data;
4141

4142
	atomic_inc(&event->mmap_count);
4143
	atomic_inc(&event->rb->mmap_count);
4144 4145
}

4146 4147 4148 4149 4150 4151 4152 4153
/*
 * 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.
 */
4154 4155
static void perf_mmap_close(struct vm_area_struct *vma)
{
4156
	struct perf_event *event = vma->vm_file->private_data;
4157

4158
	struct ring_buffer *rb = ring_buffer_get(event);
4159 4160 4161
	struct user_struct *mmap_user = rb->mmap_user;
	int mmap_locked = rb->mmap_locked;
	unsigned long size = perf_data_size(rb);
4162

4163 4164 4165
	atomic_dec(&rb->mmap_count);

	if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4166
		goto out_put;
4167

4168
	ring_buffer_attach(event, NULL);
4169 4170 4171
	mutex_unlock(&event->mmap_mutex);

	/* If there's still other mmap()s of this buffer, we're done. */
4172 4173
	if (atomic_read(&rb->mmap_count))
		goto out_put;
4174

4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190
	/*
	 * 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();
4191

4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202
		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.
		 */
4203 4204 4205
		if (event->rb == rb)
			ring_buffer_attach(event, NULL);

4206
		mutex_unlock(&event->mmap_mutex);
4207
		put_event(event);
4208

4209 4210 4211 4212 4213
		/*
		 * Restart the iteration; either we're on the wrong list or
		 * destroyed its integrity by doing a deletion.
		 */
		goto again;
4214
	}
4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229
	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);

4230
out_put:
4231
	ring_buffer_put(rb); /* could be last */
4232 4233
}

4234
static const struct vm_operations_struct perf_mmap_vmops = {
4235 4236 4237 4238
	.open		= perf_mmap_open,
	.close		= perf_mmap_close,
	.fault		= perf_mmap_fault,
	.page_mkwrite	= perf_mmap_fault,
4239 4240 4241 4242
};

static int perf_mmap(struct file *file, struct vm_area_struct *vma)
{
4243
	struct perf_event *event = file->private_data;
4244
	unsigned long user_locked, user_lock_limit;
4245
	struct user_struct *user = current_user();
4246
	unsigned long locked, lock_limit;
4247
	struct ring_buffer *rb;
4248 4249
	unsigned long vma_size;
	unsigned long nr_pages;
4250
	long user_extra, extra;
4251
	int ret = 0, flags = 0;
4252

4253 4254 4255
	/*
	 * Don't allow mmap() of inherited per-task counters. This would
	 * create a performance issue due to all children writing to the
4256
	 * same rb.
4257 4258 4259 4260
	 */
	if (event->cpu == -1 && event->attr.inherit)
		return -EINVAL;

4261
	if (!(vma->vm_flags & VM_SHARED))
4262
		return -EINVAL;
4263 4264 4265 4266

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

4267
	/*
4268
	 * If we have rb pages ensure they're a power-of-two number, so we
4269 4270 4271
	 * can do bitmasks instead of modulo.
	 */
	if (nr_pages != 0 && !is_power_of_2(nr_pages))
4272 4273
		return -EINVAL;

4274
	if (vma_size != PAGE_SIZE * (1 + nr_pages))
4275 4276
		return -EINVAL;

4277 4278
	if (vma->vm_pgoff != 0)
		return -EINVAL;
4279

4280
	WARN_ON_ONCE(event->ctx->parent_ctx);
4281
again:
4282
	mutex_lock(&event->mmap_mutex);
4283
	if (event->rb) {
4284
		if (event->rb->nr_pages != nr_pages) {
4285
			ret = -EINVAL;
4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298
			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;
		}

4299 4300 4301
		goto unlock;
	}

4302
	user_extra = nr_pages + 1;
4303
	user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
I
Ingo Molnar 已提交
4304 4305 4306 4307 4308 4309

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

4310
	user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4311

4312 4313 4314
	extra = 0;
	if (user_locked > user_lock_limit)
		extra = user_locked - user_lock_limit;
4315

4316
	lock_limit = rlimit(RLIMIT_MEMLOCK);
4317
	lock_limit >>= PAGE_SHIFT;
4318
	locked = vma->vm_mm->pinned_vm + extra;
4319

4320 4321
	if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
		!capable(CAP_IPC_LOCK)) {
4322 4323 4324
		ret = -EPERM;
		goto unlock;
	}
4325

4326
	WARN_ON(event->rb);
4327

4328
	if (vma->vm_flags & VM_WRITE)
4329
		flags |= RING_BUFFER_WRITABLE;
4330

4331 4332 4333 4334
	rb = rb_alloc(nr_pages, 
		event->attr.watermark ? event->attr.wakeup_watermark : 0,
		event->cpu, flags);

4335
	if (!rb) {
4336
		ret = -ENOMEM;
4337
		goto unlock;
4338
	}
P
Peter Zijlstra 已提交
4339

4340
	atomic_set(&rb->mmap_count, 1);
P
Peter Zijlstra 已提交
4341 4342
	rb->mmap_locked = extra;
	rb->mmap_user = get_current_user();
4343

4344
	atomic_long_add(user_extra, &user->locked_vm);
P
Peter Zijlstra 已提交
4345 4346
	vma->vm_mm->pinned_vm += extra;

4347
	ring_buffer_attach(event, rb);
4348

4349
	perf_event_init_userpage(event);
4350 4351
	perf_event_update_userpage(event);

4352
unlock:
4353 4354
	if (!ret)
		atomic_inc(&event->mmap_count);
4355
	mutex_unlock(&event->mmap_mutex);
4356

4357 4358 4359 4360
	/*
	 * Since pinned accounting is per vm we cannot allow fork() to copy our
	 * vma.
	 */
P
Peter Zijlstra 已提交
4361
	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4362
	vma->vm_ops = &perf_mmap_vmops;
4363 4364

	return ret;
4365 4366
}

P
Peter Zijlstra 已提交
4367 4368
static int perf_fasync(int fd, struct file *filp, int on)
{
A
Al Viro 已提交
4369
	struct inode *inode = file_inode(filp);
4370
	struct perf_event *event = filp->private_data;
P
Peter Zijlstra 已提交
4371 4372 4373
	int retval;

	mutex_lock(&inode->i_mutex);
4374
	retval = fasync_helper(fd, filp, on, &event->fasync);
P
Peter Zijlstra 已提交
4375 4376 4377 4378 4379 4380 4381 4382
	mutex_unlock(&inode->i_mutex);

	if (retval < 0)
		return retval;

	return 0;
}

T
Thomas Gleixner 已提交
4383
static const struct file_operations perf_fops = {
4384
	.llseek			= no_llseek,
T
Thomas Gleixner 已提交
4385 4386 4387
	.release		= perf_release,
	.read			= perf_read,
	.poll			= perf_poll,
4388
	.unlocked_ioctl		= perf_ioctl,
P
Pawel Moll 已提交
4389
	.compat_ioctl		= perf_compat_ioctl,
4390
	.mmap			= perf_mmap,
P
Peter Zijlstra 已提交
4391
	.fasync			= perf_fasync,
T
Thomas Gleixner 已提交
4392 4393
};

4394
/*
4395
 * Perf event wakeup
4396 4397 4398 4399 4400
 *
 * If there's data, ensure we set the poll() state and publish everything
 * to user-space before waking everybody up.
 */

4401
void perf_event_wakeup(struct perf_event *event)
4402
{
4403
	ring_buffer_wakeup(event);
4404

4405 4406 4407
	if (event->pending_kill) {
		kill_fasync(&event->fasync, SIGIO, event->pending_kill);
		event->pending_kill = 0;
4408
	}
4409 4410
}

4411
static void perf_pending_event(struct irq_work *entry)
4412
{
4413 4414
	struct perf_event *event = container_of(entry,
			struct perf_event, pending);
4415

4416 4417 4418
	if (event->pending_disable) {
		event->pending_disable = 0;
		__perf_event_disable(event);
4419 4420
	}

4421 4422 4423
	if (event->pending_wakeup) {
		event->pending_wakeup = 0;
		perf_event_wakeup(event);
4424 4425 4426
	}
}

4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447
/*
 * 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);

4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462
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);
	}
}

4463
static void perf_sample_regs_user(struct perf_regs *regs_user,
4464 4465
				  struct pt_regs *regs,
				  struct pt_regs *regs_user_copy)
4466
{
4467 4468
	if (user_mode(regs)) {
		regs_user->abi = perf_reg_abi(current);
4469
		regs_user->regs = regs;
4470 4471
	} else if (current->mm) {
		perf_get_regs_user(regs_user, regs, regs_user_copy);
4472 4473 4474
	} else {
		regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
		regs_user->regs = NULL;
4475 4476 4477
	}
}

4478 4479 4480 4481 4482 4483 4484 4485
static void perf_sample_regs_intr(struct perf_regs *regs_intr,
				  struct pt_regs *regs)
{
	regs_intr->regs = regs;
	regs_intr->abi  = perf_reg_abi(current);
}


4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580
/*
 * 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);
	}
}

4581 4582 4583
static void __perf_event_header__init_id(struct perf_event_header *header,
					 struct perf_sample_data *data,
					 struct perf_event *event)
4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598
{
	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();

4599
	if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610
		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;
	}
}

4611 4612 4613
void perf_event_header__init_id(struct perf_event_header *header,
				struct perf_sample_data *data,
				struct perf_event *event)
4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637
{
	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);
4638 4639 4640

	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);
4641 4642
}

4643 4644 4645
void perf_event__output_id_sample(struct perf_event *event,
				  struct perf_output_handle *handle,
				  struct perf_sample_data *sample)
4646 4647 4648 4649 4650
{
	if (event->attr.sample_id_all)
		__perf_event__output_id_sample(handle, sample);
}

4651
static void perf_output_read_one(struct perf_output_handle *handle,
4652 4653
				 struct perf_event *event,
				 u64 enabled, u64 running)
4654
{
4655
	u64 read_format = event->attr.read_format;
4656 4657 4658
	u64 values[4];
	int n = 0;

P
Peter Zijlstra 已提交
4659
	values[n++] = perf_event_count(event);
4660
	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4661
		values[n++] = enabled +
4662
			atomic64_read(&event->child_total_time_enabled);
4663 4664
	}
	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4665
		values[n++] = running +
4666
			atomic64_read(&event->child_total_time_running);
4667 4668
	}
	if (read_format & PERF_FORMAT_ID)
4669
		values[n++] = primary_event_id(event);
4670

4671
	__output_copy(handle, values, n * sizeof(u64));
4672 4673 4674
}

/*
4675
 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4676 4677
 */
static void perf_output_read_group(struct perf_output_handle *handle,
4678 4679
			    struct perf_event *event,
			    u64 enabled, u64 running)
4680
{
4681 4682
	struct perf_event *leader = event->group_leader, *sub;
	u64 read_format = event->attr.read_format;
4683 4684 4685 4686 4687 4688
	u64 values[5];
	int n = 0;

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

	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4689
		values[n++] = enabled;
4690 4691

	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4692
		values[n++] = running;
4693

4694
	if (leader != event)
4695 4696
		leader->pmu->read(leader);

P
Peter Zijlstra 已提交
4697
	values[n++] = perf_event_count(leader);
4698
	if (read_format & PERF_FORMAT_ID)
4699
		values[n++] = primary_event_id(leader);
4700

4701
	__output_copy(handle, values, n * sizeof(u64));
4702

4703
	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4704 4705
		n = 0;

4706 4707
		if ((sub != event) &&
		    (sub->state == PERF_EVENT_STATE_ACTIVE))
4708 4709
			sub->pmu->read(sub);

P
Peter Zijlstra 已提交
4710
		values[n++] = perf_event_count(sub);
4711
		if (read_format & PERF_FORMAT_ID)
4712
			values[n++] = primary_event_id(sub);
4713

4714
		__output_copy(handle, values, n * sizeof(u64));
4715 4716 4717
	}
}

4718 4719 4720
#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
				 PERF_FORMAT_TOTAL_TIME_RUNNING)

4721
static void perf_output_read(struct perf_output_handle *handle,
4722
			     struct perf_event *event)
4723
{
4724
	u64 enabled = 0, running = 0, now;
4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735
	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
	 */
4736
	if (read_format & PERF_FORMAT_TOTAL_TIMES)
4737
		calc_timer_values(event, &now, &enabled, &running);
4738

4739
	if (event->attr.read_format & PERF_FORMAT_GROUP)
4740
		perf_output_read_group(handle, event, enabled, running);
4741
	else
4742
		perf_output_read_one(handle, event, enabled, running);
4743 4744
}

4745 4746 4747
void perf_output_sample(struct perf_output_handle *handle,
			struct perf_event_header *header,
			struct perf_sample_data *data,
4748
			struct perf_event *event)
4749 4750 4751 4752 4753
{
	u64 sample_type = data->type;

	perf_output_put(handle, *header);

4754 4755 4756
	if (sample_type & PERF_SAMPLE_IDENTIFIER)
		perf_output_put(handle, data->id);

4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781
	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)
4782
		perf_output_read(handle, event);
4783 4784 4785 4786 4787 4788 4789 4790 4791 4792

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

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

			size *= sizeof(u64);

4793
			__output_copy(handle, data->callchain, size);
4794 4795 4796 4797 4798 4799 4800 4801 4802
		} 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);
4803 4804
			__output_copy(handle, data->raw->data,
					   data->raw->size);
4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815
		} else {
			struct {
				u32	size;
				u32	data;
			} raw = {
				.size = sizeof(u32),
				.data = 0,
			};
			perf_output_put(handle, raw);
		}
	}
4816

4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833
	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);
		}
	}
4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850

	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);
		}
	}
4851

4852
	if (sample_type & PERF_SAMPLE_STACK_USER) {
4853 4854 4855
		perf_output_sample_ustack(handle,
					  data->stack_user_size,
					  data->regs_user.regs);
4856
	}
A
Andi Kleen 已提交
4857 4858 4859

	if (sample_type & PERF_SAMPLE_WEIGHT)
		perf_output_put(handle, data->weight);
4860 4861 4862

	if (sample_type & PERF_SAMPLE_DATA_SRC)
		perf_output_put(handle, data->data_src.val);
4863

A
Andi Kleen 已提交
4864 4865 4866
	if (sample_type & PERF_SAMPLE_TRANSACTION)
		perf_output_put(handle, data->txn);

4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883
	if (sample_type & PERF_SAMPLE_REGS_INTR) {
		u64 abi = data->regs_intr.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_intr;

			perf_output_sample_regs(handle,
						data->regs_intr.regs,
						mask);
		}
	}

4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896
	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);
			}
		}
	}
4897 4898 4899 4900
}

void perf_prepare_sample(struct perf_event_header *header,
			 struct perf_sample_data *data,
4901
			 struct perf_event *event,
4902
			 struct pt_regs *regs)
4903
{
4904
	u64 sample_type = event->attr.sample_type;
4905

4906
	header->type = PERF_RECORD_SAMPLE;
4907
	header->size = sizeof(*header) + event->header_size;
4908 4909 4910

	header->misc = 0;
	header->misc |= perf_misc_flags(regs);
4911

4912
	__perf_event_header__init_id(header, data, event);
4913

4914
	if (sample_type & PERF_SAMPLE_IP)
4915 4916
		data->ip = perf_instruction_pointer(regs);

4917
	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4918
		int size = 1;
4919

4920
		data->callchain = perf_callchain(event, regs);
4921 4922 4923 4924 4925

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

		header->size += size * sizeof(u64);
4926 4927
	}

4928
	if (sample_type & PERF_SAMPLE_RAW) {
4929 4930 4931 4932 4933 4934 4935 4936
		int size = sizeof(u32);

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

		WARN_ON_ONCE(size & (sizeof(u64)-1));
4937
		header->size += size;
4938
	}
4939 4940 4941 4942 4943 4944 4945 4946 4947

	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;
	}
4948

4949
	if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
4950 4951
		perf_sample_regs_user(&data->regs_user, regs,
				      &data->regs_user_copy);
4952

4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963
	if (sample_type & PERF_SAMPLE_REGS_USER) {
		/* regs dump ABI info */
		int size = sizeof(u64);

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

		header->size += size;
	}
4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975

	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.
		 */
		u16 stack_size = event->attr.sample_stack_user;
		u16 size = sizeof(u64);

		stack_size = perf_sample_ustack_size(stack_size, header->size,
4976
						     data->regs_user.regs);
4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988

		/*
		 * 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;
	}
4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003

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

		perf_sample_regs_intr(&data->regs_intr, regs);

		if (data->regs_intr.regs) {
			u64 mask = event->attr.sample_regs_intr;

			size += hweight64(mask) * sizeof(u64);
		}

		header->size += size;
	}
5004
}
5005

5006
static void perf_event_output(struct perf_event *event,
5007 5008 5009 5010 5011
				struct perf_sample_data *data,
				struct pt_regs *regs)
{
	struct perf_output_handle handle;
	struct perf_event_header header;
5012

5013 5014 5015
	/* protect the callchain buffers */
	rcu_read_lock();

5016
	perf_prepare_sample(&header, data, event, regs);
P
Peter Zijlstra 已提交
5017

5018
	if (perf_output_begin(&handle, event, header.size))
5019
		goto exit;
5020

5021
	perf_output_sample(&handle, &header, data, event);
5022

5023
	perf_output_end(&handle);
5024 5025 5026

exit:
	rcu_read_unlock();
5027 5028
}

5029
/*
5030
 * read event_id
5031 5032 5033 5034 5035 5036 5037 5038 5039 5040
 */

struct perf_read_event {
	struct perf_event_header	header;

	u32				pid;
	u32				tid;
};

static void
5041
perf_event_read_event(struct perf_event *event,
5042 5043 5044
			struct task_struct *task)
{
	struct perf_output_handle handle;
5045
	struct perf_sample_data sample;
5046
	struct perf_read_event read_event = {
5047
		.header = {
5048
			.type = PERF_RECORD_READ,
5049
			.misc = 0,
5050
			.size = sizeof(read_event) + event->read_size,
5051
		},
5052 5053
		.pid = perf_event_pid(event, task),
		.tid = perf_event_tid(event, task),
5054
	};
5055
	int ret;
5056

5057
	perf_event_header__init_id(&read_event.header, &sample, event);
5058
	ret = perf_output_begin(&handle, event, read_event.header.size);
5059 5060 5061
	if (ret)
		return;

5062
	perf_output_put(&handle, read_event);
5063
	perf_output_read(&handle, event);
5064
	perf_event__output_id_sample(event, &handle, &sample);
5065

5066 5067 5068
	perf_output_end(&handle);
}

5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082
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;
5083
		output(event, data);
5084 5085 5086 5087
	}
}

static void
5088
perf_event_aux(perf_event_aux_output_cb output, void *data,
5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100
	       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;
5101
		perf_event_aux_ctx(&cpuctx->ctx, output, data);
5102 5103 5104 5105 5106 5107 5108
		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)
5109
			perf_event_aux_ctx(ctx, output, data);
5110 5111 5112 5113 5114 5115
next:
		put_cpu_ptr(pmu->pmu_cpu_context);
	}

	if (task_ctx) {
		preempt_disable();
5116
		perf_event_aux_ctx(task_ctx, output, data);
5117 5118 5119 5120 5121
		preempt_enable();
	}
	rcu_read_unlock();
}

P
Peter Zijlstra 已提交
5122
/*
P
Peter Zijlstra 已提交
5123 5124
 * task tracking -- fork/exit
 *
5125
 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
P
Peter Zijlstra 已提交
5126 5127
 */

P
Peter Zijlstra 已提交
5128
struct perf_task_event {
5129
	struct task_struct		*task;
5130
	struct perf_event_context	*task_ctx;
P
Peter Zijlstra 已提交
5131 5132 5133 5134 5135 5136

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				ppid;
P
Peter Zijlstra 已提交
5137 5138
		u32				tid;
		u32				ptid;
5139
		u64				time;
5140
	} event_id;
P
Peter Zijlstra 已提交
5141 5142
};

5143 5144
static int perf_event_task_match(struct perf_event *event)
{
5145 5146 5147
	return event->attr.comm  || event->attr.mmap ||
	       event->attr.mmap2 || event->attr.mmap_data ||
	       event->attr.task;
5148 5149
}

5150
static void perf_event_task_output(struct perf_event *event,
5151
				   void *data)
P
Peter Zijlstra 已提交
5152
{
5153
	struct perf_task_event *task_event = data;
P
Peter Zijlstra 已提交
5154
	struct perf_output_handle handle;
5155
	struct perf_sample_data	sample;
P
Peter Zijlstra 已提交
5156
	struct task_struct *task = task_event->task;
5157
	int ret, size = task_event->event_id.header.size;
5158

5159 5160 5161
	if (!perf_event_task_match(event))
		return;

5162
	perf_event_header__init_id(&task_event->event_id.header, &sample, event);
P
Peter Zijlstra 已提交
5163

5164
	ret = perf_output_begin(&handle, event,
5165
				task_event->event_id.header.size);
5166
	if (ret)
5167
		goto out;
P
Peter Zijlstra 已提交
5168

5169 5170
	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
5171

5172 5173
	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
5174

5175
	perf_output_put(&handle, task_event->event_id);
5176

5177 5178
	perf_event__output_id_sample(event, &handle, &sample);

P
Peter Zijlstra 已提交
5179
	perf_output_end(&handle);
5180 5181
out:
	task_event->event_id.header.size = size;
P
Peter Zijlstra 已提交
5182 5183
}

5184 5185
static void perf_event_task(struct task_struct *task,
			      struct perf_event_context *task_ctx,
5186
			      int new)
P
Peter Zijlstra 已提交
5187
{
P
Peter Zijlstra 已提交
5188
	struct perf_task_event task_event;
P
Peter Zijlstra 已提交
5189

5190 5191 5192
	if (!atomic_read(&nr_comm_events) &&
	    !atomic_read(&nr_mmap_events) &&
	    !atomic_read(&nr_task_events))
P
Peter Zijlstra 已提交
5193 5194
		return;

P
Peter Zijlstra 已提交
5195
	task_event = (struct perf_task_event){
5196 5197
		.task	  = task,
		.task_ctx = task_ctx,
5198
		.event_id    = {
P
Peter Zijlstra 已提交
5199
			.header = {
5200
				.type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5201
				.misc = 0,
5202
				.size = sizeof(task_event.event_id),
P
Peter Zijlstra 已提交
5203
			},
5204 5205
			/* .pid  */
			/* .ppid */
P
Peter Zijlstra 已提交
5206 5207
			/* .tid  */
			/* .ptid */
P
Peter Zijlstra 已提交
5208
			.time = perf_clock(),
P
Peter Zijlstra 已提交
5209 5210 5211
		},
	};

5212
	perf_event_aux(perf_event_task_output,
5213 5214
		       &task_event,
		       task_ctx);
P
Peter Zijlstra 已提交
5215 5216
}

5217
void perf_event_fork(struct task_struct *task)
P
Peter Zijlstra 已提交
5218
{
5219
	perf_event_task(task, NULL, 1);
P
Peter Zijlstra 已提交
5220 5221
}

5222 5223 5224 5225 5226
/*
 * comm tracking
 */

struct perf_comm_event {
5227 5228
	struct task_struct	*task;
	char			*comm;
5229 5230 5231 5232 5233 5234 5235
	int			comm_size;

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
5236
	} event_id;
5237 5238
};

5239 5240 5241 5242 5243
static int perf_event_comm_match(struct perf_event *event)
{
	return event->attr.comm;
}

5244
static void perf_event_comm_output(struct perf_event *event,
5245
				   void *data)
5246
{
5247
	struct perf_comm_event *comm_event = data;
5248
	struct perf_output_handle handle;
5249
	struct perf_sample_data sample;
5250
	int size = comm_event->event_id.header.size;
5251 5252
	int ret;

5253 5254 5255
	if (!perf_event_comm_match(event))
		return;

5256 5257
	perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5258
				comm_event->event_id.header.size);
5259 5260

	if (ret)
5261
		goto out;
5262

5263 5264
	comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
	comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5265

5266
	perf_output_put(&handle, comm_event->event_id);
5267
	__output_copy(&handle, comm_event->comm,
5268
				   comm_event->comm_size);
5269 5270 5271

	perf_event__output_id_sample(event, &handle, &sample);

5272
	perf_output_end(&handle);
5273 5274
out:
	comm_event->event_id.header.size = size;
5275 5276
}

5277
static void perf_event_comm_event(struct perf_comm_event *comm_event)
5278
{
5279
	char comm[TASK_COMM_LEN];
5280 5281
	unsigned int size;

5282
	memset(comm, 0, sizeof(comm));
5283
	strlcpy(comm, comm_event->task->comm, sizeof(comm));
5284
	size = ALIGN(strlen(comm)+1, sizeof(u64));
5285 5286 5287 5288

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

5289
	comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
P
Peter Zijlstra 已提交
5290

5291
	perf_event_aux(perf_event_comm_output,
5292 5293
		       comm_event,
		       NULL);
5294 5295
}

5296
void perf_event_comm(struct task_struct *task, bool exec)
5297
{
5298 5299
	struct perf_comm_event comm_event;

5300
	if (!atomic_read(&nr_comm_events))
5301
		return;
5302

5303
	comm_event = (struct perf_comm_event){
5304
		.task	= task,
5305 5306
		/* .comm      */
		/* .comm_size */
5307
		.event_id  = {
5308
			.header = {
5309
				.type = PERF_RECORD_COMM,
5310
				.misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5311 5312 5313 5314
				/* .size */
			},
			/* .pid */
			/* .tid */
5315 5316 5317
		},
	};

5318
	perf_event_comm_event(&comm_event);
5319 5320
}

5321 5322 5323 5324 5325
/*
 * mmap tracking
 */

struct perf_mmap_event {
5326 5327 5328 5329
	struct vm_area_struct	*vma;

	const char		*file_name;
	int			file_size;
5330 5331 5332
	int			maj, min;
	u64			ino;
	u64			ino_generation;
5333
	u32			prot, flags;
5334 5335 5336 5337 5338 5339 5340 5341 5342

	struct {
		struct perf_event_header	header;

		u32				pid;
		u32				tid;
		u64				start;
		u64				len;
		u64				pgoff;
5343
	} event_id;
5344 5345
};

5346 5347 5348 5349 5350 5351 5352 5353
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) ||
5354
	       (executable && (event->attr.mmap || event->attr.mmap2));
5355 5356
}

5357
static void perf_event_mmap_output(struct perf_event *event,
5358
				   void *data)
5359
{
5360
	struct perf_mmap_event *mmap_event = data;
5361
	struct perf_output_handle handle;
5362
	struct perf_sample_data sample;
5363
	int size = mmap_event->event_id.header.size;
5364
	int ret;
5365

5366 5367 5368
	if (!perf_event_mmap_match(event, data))
		return;

5369 5370 5371 5372 5373
	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);
5374
		mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5375 5376
		mmap_event->event_id.header.size += sizeof(mmap_event->prot);
		mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5377 5378
	}

5379 5380
	perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
	ret = perf_output_begin(&handle, event,
5381
				mmap_event->event_id.header.size);
5382
	if (ret)
5383
		goto out;
5384

5385 5386
	mmap_event->event_id.pid = perf_event_pid(event, current);
	mmap_event->event_id.tid = perf_event_tid(event, current);
5387

5388
	perf_output_put(&handle, mmap_event->event_id);
5389 5390 5391 5392 5393 5394

	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);
5395 5396
		perf_output_put(&handle, mmap_event->prot);
		perf_output_put(&handle, mmap_event->flags);
5397 5398
	}

5399
	__output_copy(&handle, mmap_event->file_name,
5400
				   mmap_event->file_size);
5401 5402 5403

	perf_event__output_id_sample(event, &handle, &sample);

5404
	perf_output_end(&handle);
5405 5406
out:
	mmap_event->event_id.header.size = size;
5407 5408
}

5409
static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5410
{
5411 5412
	struct vm_area_struct *vma = mmap_event->vma;
	struct file *file = vma->vm_file;
5413 5414
	int maj = 0, min = 0;
	u64 ino = 0, gen = 0;
5415
	u32 prot = 0, flags = 0;
5416 5417 5418
	unsigned int size;
	char tmp[16];
	char *buf = NULL;
5419
	char *name;
5420

5421
	if (file) {
5422 5423
		struct inode *inode;
		dev_t dev;
5424

5425
		buf = kmalloc(PATH_MAX, GFP_KERNEL);
5426
		if (!buf) {
5427 5428
			name = "//enomem";
			goto cpy_name;
5429
		}
5430
		/*
5431
		 * d_path() works from the end of the rb backwards, so we
5432 5433 5434
		 * need to add enough zero bytes after the string to handle
		 * the 64bit alignment we do later.
		 */
5435
		name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5436
		if (IS_ERR(name)) {
5437 5438
			name = "//toolong";
			goto cpy_name;
5439
		}
5440 5441 5442 5443 5444 5445
		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);
5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467

		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;

5468
		goto got_name;
5469
	} else {
5470 5471 5472 5473 5474 5475
		if (vma->vm_ops && vma->vm_ops->name) {
			name = (char *) vma->vm_ops->name(vma);
			if (name)
				goto cpy_name;
		}

5476
		name = (char *)arch_vma_name(vma);
5477 5478
		if (name)
			goto cpy_name;
5479

5480
		if (vma->vm_start <= vma->vm_mm->start_brk &&
5481
				vma->vm_end >= vma->vm_mm->brk) {
5482 5483
			name = "[heap]";
			goto cpy_name;
5484 5485
		}
		if (vma->vm_start <= vma->vm_mm->start_stack &&
5486
				vma->vm_end >= vma->vm_mm->start_stack) {
5487 5488
			name = "[stack]";
			goto cpy_name;
5489 5490
		}

5491 5492
		name = "//anon";
		goto cpy_name;
5493 5494
	}

5495 5496 5497
cpy_name:
	strlcpy(tmp, name, sizeof(tmp));
	name = tmp;
5498
got_name:
5499 5500 5501 5502 5503 5504 5505 5506
	/*
	 * 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';
5507 5508 5509

	mmap_event->file_name = name;
	mmap_event->file_size = size;
5510 5511 5512 5513
	mmap_event->maj = maj;
	mmap_event->min = min;
	mmap_event->ino = ino;
	mmap_event->ino_generation = gen;
5514 5515
	mmap_event->prot = prot;
	mmap_event->flags = flags;
5516

5517 5518 5519
	if (!(vma->vm_flags & VM_EXEC))
		mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;

5520
	mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5521

5522
	perf_event_aux(perf_event_mmap_output,
5523 5524
		       mmap_event,
		       NULL);
5525

5526 5527 5528
	kfree(buf);
}

5529
void perf_event_mmap(struct vm_area_struct *vma)
5530
{
5531 5532
	struct perf_mmap_event mmap_event;

5533
	if (!atomic_read(&nr_mmap_events))
5534 5535 5536
		return;

	mmap_event = (struct perf_mmap_event){
5537
		.vma	= vma,
5538 5539
		/* .file_name */
		/* .file_size */
5540
		.event_id  = {
5541
			.header = {
5542
				.type = PERF_RECORD_MMAP,
5543
				.misc = PERF_RECORD_MISC_USER,
5544 5545 5546 5547
				/* .size */
			},
			/* .pid */
			/* .tid */
5548 5549
			.start  = vma->vm_start,
			.len    = vma->vm_end - vma->vm_start,
5550
			.pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
5551
		},
5552 5553 5554 5555
		/* .maj (attr_mmap2 only) */
		/* .min (attr_mmap2 only) */
		/* .ino (attr_mmap2 only) */
		/* .ino_generation (attr_mmap2 only) */
5556 5557
		/* .prot (attr_mmap2 only) */
		/* .flags (attr_mmap2 only) */
5558 5559
	};

5560
	perf_event_mmap_event(&mmap_event);
5561 5562
}

5563 5564 5565 5566
/*
 * IRQ throttle logging
 */

5567
static void perf_log_throttle(struct perf_event *event, int enable)
5568 5569
{
	struct perf_output_handle handle;
5570
	struct perf_sample_data sample;
5571 5572 5573 5574 5575
	int ret;

	struct {
		struct perf_event_header	header;
		u64				time;
5576
		u64				id;
5577
		u64				stream_id;
5578 5579
	} throttle_event = {
		.header = {
5580
			.type = PERF_RECORD_THROTTLE,
5581 5582 5583
			.misc = 0,
			.size = sizeof(throttle_event),
		},
P
Peter Zijlstra 已提交
5584
		.time		= perf_clock(),
5585 5586
		.id		= primary_event_id(event),
		.stream_id	= event->id,
5587 5588
	};

5589
	if (enable)
5590
		throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5591

5592 5593 5594
	perf_event_header__init_id(&throttle_event.header, &sample, event);

	ret = perf_output_begin(&handle, event,
5595
				throttle_event.header.size);
5596 5597 5598 5599
	if (ret)
		return;

	perf_output_put(&handle, throttle_event);
5600
	perf_event__output_id_sample(event, &handle, &sample);
5601 5602 5603
	perf_output_end(&handle);
}

5604
/*
5605
 * Generic event overflow handling, sampling.
5606 5607
 */

5608
static int __perf_event_overflow(struct perf_event *event,
5609 5610
				   int throttle, struct perf_sample_data *data,
				   struct pt_regs *regs)
5611
{
5612 5613
	int events = atomic_read(&event->event_limit);
	struct hw_perf_event *hwc = &event->hw;
5614
	u64 seq;
5615 5616
	int ret = 0;

5617 5618 5619 5620 5621 5622 5623
	/*
	 * Non-sampling counters might still use the PMI to fold short
	 * hardware counters, ignore those.
	 */
	if (unlikely(!is_sampling_event(event)))
		return 0;

5624 5625 5626 5627 5628 5629 5630 5631 5632
	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
5633 5634
			hwc->interrupts = MAX_INTERRUPTS;
			perf_log_throttle(event, 0);
5635
			tick_nohz_full_kick();
5636 5637
			ret = 1;
		}
5638
	}
5639

5640
	if (event->attr.freq) {
P
Peter Zijlstra 已提交
5641
		u64 now = perf_clock();
5642
		s64 delta = now - hwc->freq_time_stamp;
5643

5644
		hwc->freq_time_stamp = now;
5645

5646
		if (delta > 0 && delta < 2*TICK_NSEC)
5647
			perf_adjust_period(event, delta, hwc->last_period, true);
5648 5649
	}

5650 5651
	/*
	 * XXX event_limit might not quite work as expected on inherited
5652
	 * events
5653 5654
	 */

5655 5656
	event->pending_kill = POLL_IN;
	if (events && atomic_dec_and_test(&event->event_limit)) {
5657
		ret = 1;
5658
		event->pending_kill = POLL_HUP;
5659 5660
		event->pending_disable = 1;
		irq_work_queue(&event->pending);
5661 5662
	}

5663
	if (event->overflow_handler)
5664
		event->overflow_handler(event, data, regs);
5665
	else
5666
		perf_event_output(event, data, regs);
5667

P
Peter Zijlstra 已提交
5668
	if (event->fasync && event->pending_kill) {
5669 5670
		event->pending_wakeup = 1;
		irq_work_queue(&event->pending);
P
Peter Zijlstra 已提交
5671 5672
	}

5673
	return ret;
5674 5675
}

5676
int perf_event_overflow(struct perf_event *event,
5677 5678
			  struct perf_sample_data *data,
			  struct pt_regs *regs)
5679
{
5680
	return __perf_event_overflow(event, 1, data, regs);
5681 5682
}

5683
/*
5684
 * Generic software event infrastructure
5685 5686
 */

5687 5688 5689 5690 5691 5692 5693
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];
5694 5695 5696

	/* Keeps track of cpu being initialized/exited */
	bool				online;
5697 5698 5699 5700
};

static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);

5701
/*
5702 5703
 * We directly increment event->count and keep a second value in
 * event->hw.period_left to count intervals. This period event
5704 5705 5706 5707
 * is kept in the range [-sample_period, 0] so that we can use the
 * sign as trigger.
 */

5708
u64 perf_swevent_set_period(struct perf_event *event)
5709
{
5710
	struct hw_perf_event *hwc = &event->hw;
5711 5712 5713 5714 5715
	u64 period = hwc->last_period;
	u64 nr, offset;
	s64 old, val;

	hwc->last_period = hwc->sample_period;
5716 5717

again:
5718
	old = val = local64_read(&hwc->period_left);
5719 5720
	if (val < 0)
		return 0;
5721

5722 5723 5724
	nr = div64_u64(period + val, period);
	offset = nr * period;
	val -= offset;
5725
	if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5726
		goto again;
5727

5728
	return nr;
5729 5730
}

5731
static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5732
				    struct perf_sample_data *data,
5733
				    struct pt_regs *regs)
5734
{
5735
	struct hw_perf_event *hwc = &event->hw;
5736
	int throttle = 0;
5737

5738 5739
	if (!overflow)
		overflow = perf_swevent_set_period(event);
5740

5741 5742
	if (hwc->interrupts == MAX_INTERRUPTS)
		return;
5743

5744
	for (; overflow; overflow--) {
5745
		if (__perf_event_overflow(event, throttle,
5746
					    data, regs)) {
5747 5748 5749 5750 5751 5752
			/*
			 * We inhibit the overflow from happening when
			 * hwc->interrupts == MAX_INTERRUPTS.
			 */
			break;
		}
5753
		throttle = 1;
5754
	}
5755 5756
}

P
Peter Zijlstra 已提交
5757
static void perf_swevent_event(struct perf_event *event, u64 nr,
5758
			       struct perf_sample_data *data,
5759
			       struct pt_regs *regs)
5760
{
5761
	struct hw_perf_event *hwc = &event->hw;
5762

5763
	local64_add(nr, &event->count);
5764

5765 5766 5767
	if (!regs)
		return;

5768
	if (!is_sampling_event(event))
5769
		return;
5770

5771 5772 5773 5774 5775 5776
	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;

5777
	if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5778
		return perf_swevent_overflow(event, 1, data, regs);
5779

5780
	if (local64_add_negative(nr, &hwc->period_left))
5781
		return;
5782

5783
	perf_swevent_overflow(event, 0, data, regs);
5784 5785
}

5786 5787 5788
static int perf_exclude_event(struct perf_event *event,
			      struct pt_regs *regs)
{
P
Peter Zijlstra 已提交
5789
	if (event->hw.state & PERF_HES_STOPPED)
5790
		return 1;
P
Peter Zijlstra 已提交
5791

5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802
	if (regs) {
		if (event->attr.exclude_user && user_mode(regs))
			return 1;

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

	return 0;
}

5803
static int perf_swevent_match(struct perf_event *event,
P
Peter Zijlstra 已提交
5804
				enum perf_type_id type,
L
Li Zefan 已提交
5805 5806 5807
				u32 event_id,
				struct perf_sample_data *data,
				struct pt_regs *regs)
5808
{
5809
	if (event->attr.type != type)
5810
		return 0;
5811

5812
	if (event->attr.config != event_id)
5813 5814
		return 0;

5815 5816
	if (perf_exclude_event(event, regs))
		return 0;
5817 5818 5819 5820

	return 1;
}

5821 5822 5823 5824 5825 5826 5827
static inline u64 swevent_hash(u64 type, u32 event_id)
{
	u64 val = event_id | (type << 32);

	return hash_64(val, SWEVENT_HLIST_BITS);
}

5828 5829
static inline struct hlist_head *
__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5830
{
5831 5832 5833 5834
	u64 hash = swevent_hash(type, event_id);

	return &hlist->heads[hash];
}
5835

5836 5837
/* For the read side: events when they trigger */
static inline struct hlist_head *
5838
find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5839 5840
{
	struct swevent_hlist *hlist;
5841

5842
	hlist = rcu_dereference(swhash->swevent_hlist);
5843 5844 5845
	if (!hlist)
		return NULL;

5846 5847 5848 5849 5850
	return __find_swevent_head(hlist, type, event_id);
}

/* For the event head insertion and removal in the hlist */
static inline struct hlist_head *
5851
find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5852 5853 5854 5855 5856 5857 5858 5859 5860 5861
{
	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.
	 */
5862
	hlist = rcu_dereference_protected(swhash->swevent_hlist,
5863 5864 5865 5866 5867
					  lockdep_is_held(&event->ctx->lock));
	if (!hlist)
		return NULL;

	return __find_swevent_head(hlist, type, event_id);
5868 5869 5870
}

static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5871
				    u64 nr,
5872 5873
				    struct perf_sample_data *data,
				    struct pt_regs *regs)
5874
{
5875
	struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
5876
	struct perf_event *event;
5877
	struct hlist_head *head;
5878

5879
	rcu_read_lock();
5880
	head = find_swevent_head_rcu(swhash, type, event_id);
5881 5882 5883
	if (!head)
		goto end;

5884
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
L
Li Zefan 已提交
5885
		if (perf_swevent_match(event, type, event_id, data, regs))
5886
			perf_swevent_event(event, nr, data, regs);
5887
	}
5888 5889
end:
	rcu_read_unlock();
5890 5891
}

5892 5893
DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);

5894
int perf_swevent_get_recursion_context(void)
P
Peter Zijlstra 已提交
5895
{
5896
	struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
P
Peter Zijlstra 已提交
5897

5898
	return get_recursion_context(swhash->recursion);
P
Peter Zijlstra 已提交
5899
}
I
Ingo Molnar 已提交
5900
EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
P
Peter Zijlstra 已提交
5901

5902
inline void perf_swevent_put_recursion_context(int rctx)
5903
{
5904
	struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
5905

5906
	put_recursion_context(swhash->recursion, rctx);
5907
}
5908

5909
void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5910
{
5911
	struct perf_sample_data data;
5912

5913
	if (WARN_ON_ONCE(!regs))
5914
		return;
5915

5916
	perf_sample_data_init(&data, addr, 0);
5917
	do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929
}

void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
{
	int rctx;

	preempt_disable_notrace();
	rctx = perf_swevent_get_recursion_context();
	if (unlikely(rctx < 0))
		goto fail;

	___perf_sw_event(event_id, nr, regs, addr);
5930 5931

	perf_swevent_put_recursion_context(rctx);
5932
fail:
5933
	preempt_enable_notrace();
5934 5935
}

5936
static void perf_swevent_read(struct perf_event *event)
5937 5938 5939
{
}

P
Peter Zijlstra 已提交
5940
static int perf_swevent_add(struct perf_event *event, int flags)
5941
{
5942
	struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
5943
	struct hw_perf_event *hwc = &event->hw;
5944 5945
	struct hlist_head *head;

5946
	if (is_sampling_event(event)) {
5947
		hwc->last_period = hwc->sample_period;
5948
		perf_swevent_set_period(event);
5949
	}
5950

P
Peter Zijlstra 已提交
5951 5952
	hwc->state = !(flags & PERF_EF_START);

5953
	head = find_swevent_head(swhash, event);
5954 5955 5956 5957 5958 5959
	if (!head) {
		/*
		 * We can race with cpu hotplug code. Do not
		 * WARN if the cpu just got unplugged.
		 */
		WARN_ON_ONCE(swhash->online);
5960
		return -EINVAL;
5961
	}
5962 5963 5964

	hlist_add_head_rcu(&event->hlist_entry, head);

5965 5966 5967
	return 0;
}

P
Peter Zijlstra 已提交
5968
static void perf_swevent_del(struct perf_event *event, int flags)
5969
{
5970
	hlist_del_rcu(&event->hlist_entry);
5971 5972
}

P
Peter Zijlstra 已提交
5973
static void perf_swevent_start(struct perf_event *event, int flags)
5974
{
P
Peter Zijlstra 已提交
5975
	event->hw.state = 0;
5976
}
I
Ingo Molnar 已提交
5977

P
Peter Zijlstra 已提交
5978
static void perf_swevent_stop(struct perf_event *event, int flags)
5979
{
P
Peter Zijlstra 已提交
5980
	event->hw.state = PERF_HES_STOPPED;
5981 5982
}

5983 5984
/* Deref the hlist from the update side */
static inline struct swevent_hlist *
5985
swevent_hlist_deref(struct swevent_htable *swhash)
5986
{
5987 5988
	return rcu_dereference_protected(swhash->swevent_hlist,
					 lockdep_is_held(&swhash->hlist_mutex));
5989 5990
}

5991
static void swevent_hlist_release(struct swevent_htable *swhash)
5992
{
5993
	struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5994

5995
	if (!hlist)
5996 5997
		return;

5998
	RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
5999
	kfree_rcu(hlist, rcu_head);
6000 6001 6002 6003
}

static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
{
6004
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6005

6006
	mutex_lock(&swhash->hlist_mutex);
6007

6008 6009
	if (!--swhash->hlist_refcount)
		swevent_hlist_release(swhash);
6010

6011
	mutex_unlock(&swhash->hlist_mutex);
6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023
}

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)
{
6024
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6025 6026
	int err = 0;

6027
	mutex_lock(&swhash->hlist_mutex);
6028

6029
	if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6030 6031 6032 6033 6034 6035 6036
		struct swevent_hlist *hlist;

		hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
		if (!hlist) {
			err = -ENOMEM;
			goto exit;
		}
6037
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
6038
	}
6039
	swhash->hlist_refcount++;
P
Peter Zijlstra 已提交
6040
exit:
6041
	mutex_unlock(&swhash->hlist_mutex);
6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059 6060 6061

	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
6062
fail:
6063 6064 6065 6066 6067 6068 6069 6070 6071 6072
	for_each_possible_cpu(cpu) {
		if (cpu == failed_cpu)
			break;
		swevent_hlist_put_cpu(event, cpu);
	}

	put_online_cpus();
	return err;
}

6073
struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6074

6075 6076 6077
static void sw_perf_event_destroy(struct perf_event *event)
{
	u64 event_id = event->attr.config;
6078

6079 6080
	WARN_ON(event->parent);

6081
	static_key_slow_dec(&perf_swevent_enabled[event_id]);
6082 6083 6084 6085 6086
	swevent_hlist_put(event);
}

static int perf_swevent_init(struct perf_event *event)
{
6087
	u64 event_id = event->attr.config;
6088 6089 6090 6091

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

6092 6093 6094 6095 6096 6097
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

6098 6099 6100 6101 6102 6103 6104 6105 6106
	switch (event_id) {
	case PERF_COUNT_SW_CPU_CLOCK:
	case PERF_COUNT_SW_TASK_CLOCK:
		return -ENOENT;

	default:
		break;
	}

6107
	if (event_id >= PERF_COUNT_SW_MAX)
6108 6109 6110 6111 6112 6113 6114 6115 6116
		return -ENOENT;

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

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

6117
		static_key_slow_inc(&perf_swevent_enabled[event_id]);
6118 6119 6120 6121 6122 6123 6124
		event->destroy = sw_perf_event_destroy;
	}

	return 0;
}

static struct pmu perf_swevent = {
6125
	.task_ctx_nr	= perf_sw_context,
6126

6127
	.event_init	= perf_swevent_init,
P
Peter Zijlstra 已提交
6128 6129 6130 6131
	.add		= perf_swevent_add,
	.del		= perf_swevent_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
6132 6133 6134
	.read		= perf_swevent_read,
};

6135 6136
#ifdef CONFIG_EVENT_TRACING

6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150
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)
{
6151 6152
	if (event->hw.state & PERF_HES_STOPPED)
		return 0;
6153 6154 6155 6156
	/*
	 * All tracepoints are from kernel-space.
	 */
	if (event->attr.exclude_kernel)
6157 6158 6159 6160 6161 6162 6163 6164 6165
		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,
6166 6167
		   struct pt_regs *regs, struct hlist_head *head, int rctx,
		   struct task_struct *task)
6168 6169
{
	struct perf_sample_data data;
6170 6171
	struct perf_event *event;

6172 6173 6174 6175 6176
	struct perf_raw_record raw = {
		.size = entry_size,
		.data = record,
	};

6177
	perf_sample_data_init(&data, addr, 0);
6178 6179
	data.raw = &raw;

6180
	hlist_for_each_entry_rcu(event, head, hlist_entry) {
6181
		if (perf_tp_event_match(event, &data, regs))
6182
			perf_swevent_event(event, count, &data, regs);
6183
	}
6184

6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200 6201 6202 6203 6204 6205 6206 6207 6208 6209
	/*
	 * 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();
	}

6210
	perf_swevent_put_recursion_context(rctx);
6211 6212 6213
}
EXPORT_SYMBOL_GPL(perf_tp_event);

6214
static void tp_perf_event_destroy(struct perf_event *event)
6215
{
6216
	perf_trace_destroy(event);
6217 6218
}

6219
static int perf_tp_event_init(struct perf_event *event)
6220
{
6221 6222
	int err;

6223 6224 6225
	if (event->attr.type != PERF_TYPE_TRACEPOINT)
		return -ENOENT;

6226 6227 6228 6229 6230 6231
	/*
	 * no branch sampling for tracepoint events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

6232 6233
	err = perf_trace_init(event);
	if (err)
6234
		return err;
6235

6236
	event->destroy = tp_perf_event_destroy;
6237

6238 6239 6240 6241
	return 0;
}

static struct pmu perf_tracepoint = {
6242 6243
	.task_ctx_nr	= perf_sw_context,

6244
	.event_init	= perf_tp_event_init,
P
Peter Zijlstra 已提交
6245 6246 6247 6248
	.add		= perf_trace_add,
	.del		= perf_trace_del,
	.start		= perf_swevent_start,
	.stop		= perf_swevent_stop,
6249 6250 6251 6252 6253
	.read		= perf_swevent_read,
};

static inline void perf_tp_register(void)
{
P
Peter Zijlstra 已提交
6254
	perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6255
}
L
Li Zefan 已提交
6256 6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268 6269 6270 6271 6272 6273 6274 6275 6276 6277 6278 6279

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);
}

6280
#else
L
Li Zefan 已提交
6281

6282
static inline void perf_tp_register(void)
6283 6284
{
}
L
Li Zefan 已提交
6285 6286 6287 6288 6289 6290 6291 6292 6293 6294

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)
{
}

6295
#endif /* CONFIG_EVENT_TRACING */
6296

6297
#ifdef CONFIG_HAVE_HW_BREAKPOINT
6298
void perf_bp_event(struct perf_event *bp, void *data)
6299
{
6300 6301 6302
	struct perf_sample_data sample;
	struct pt_regs *regs = data;

6303
	perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6304

P
Peter Zijlstra 已提交
6305
	if (!bp->hw.state && !perf_exclude_event(bp, regs))
6306
		perf_swevent_event(bp, 1, &sample, regs);
6307 6308 6309
}
#endif

6310 6311 6312
/*
 * hrtimer based swevent callback
 */
6313

6314
static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6315
{
6316 6317 6318 6319 6320
	enum hrtimer_restart ret = HRTIMER_RESTART;
	struct perf_sample_data data;
	struct pt_regs *regs;
	struct perf_event *event;
	u64 period;
6321

6322
	event = container_of(hrtimer, struct perf_event, hw.hrtimer);
P
Peter Zijlstra 已提交
6323 6324 6325 6326

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

6327
	event->pmu->read(event);
6328

6329
	perf_sample_data_init(&data, 0, event->hw.last_period);
6330 6331 6332
	regs = get_irq_regs();

	if (regs && !perf_exclude_event(event, regs)) {
6333
		if (!(event->attr.exclude_idle && is_idle_task(current)))
6334
			if (__perf_event_overflow(event, 1, &data, regs))
6335 6336
				ret = HRTIMER_NORESTART;
	}
6337

6338 6339
	period = max_t(u64, 10000, event->hw.sample_period);
	hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6340

6341
	return ret;
6342 6343
}

6344
static void perf_swevent_start_hrtimer(struct perf_event *event)
6345
{
6346
	struct hw_perf_event *hwc = &event->hw;
6347 6348 6349 6350
	s64 period;

	if (!is_sampling_event(event))
		return;
6351

6352 6353 6354 6355
	period = local64_read(&hwc->period_left);
	if (period) {
		if (period < 0)
			period = 10000;
P
Peter Zijlstra 已提交
6356

6357 6358 6359 6360 6361
		local64_set(&hwc->period_left, 0);
	} else {
		period = max_t(u64, 10000, hwc->sample_period);
	}
	__hrtimer_start_range_ns(&hwc->hrtimer,
6362
				ns_to_ktime(period), 0,
6363
				HRTIMER_MODE_REL_PINNED, 0);
6364
}
6365 6366

static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6367
{
6368 6369
	struct hw_perf_event *hwc = &event->hw;

6370
	if (is_sampling_event(event)) {
6371
		ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
P
Peter Zijlstra 已提交
6372
		local64_set(&hwc->period_left, ktime_to_ns(remaining));
6373 6374 6375

		hrtimer_cancel(&hwc->hrtimer);
	}
6376 6377
}

P
Peter Zijlstra 已提交
6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397
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);
6398
		hwc->last_period = hwc->sample_period;
P
Peter Zijlstra 已提交
6399 6400 6401 6402
		event->attr.freq = 0;
	}
}

6403 6404 6405 6406 6407
/*
 * Software event: cpu wall time clock
 */

static void cpu_clock_event_update(struct perf_event *event)
6408
{
6409 6410 6411
	s64 prev;
	u64 now;

P
Peter Zijlstra 已提交
6412
	now = local_clock();
6413 6414
	prev = local64_xchg(&event->hw.prev_count, now);
	local64_add(now - prev, &event->count);
6415 6416
}

P
Peter Zijlstra 已提交
6417
static void cpu_clock_event_start(struct perf_event *event, int flags)
6418
{
P
Peter Zijlstra 已提交
6419
	local64_set(&event->hw.prev_count, local_clock());
6420 6421 6422
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6423
static void cpu_clock_event_stop(struct perf_event *event, int flags)
6424
{
6425 6426 6427
	perf_swevent_cancel_hrtimer(event);
	cpu_clock_event_update(event);
}
6428

P
Peter Zijlstra 已提交
6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441
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);
}

6442 6443 6444 6445
static void cpu_clock_event_read(struct perf_event *event)
{
	cpu_clock_event_update(event);
}
6446

6447 6448 6449 6450 6451 6452 6453 6454
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;

6455 6456 6457 6458 6459 6460
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6461 6462
	perf_swevent_init_hrtimer(event);

6463
	return 0;
6464 6465
}

6466
static struct pmu perf_cpu_clock = {
6467 6468
	.task_ctx_nr	= perf_sw_context,

6469
	.event_init	= cpu_clock_event_init,
P
Peter Zijlstra 已提交
6470 6471 6472 6473
	.add		= cpu_clock_event_add,
	.del		= cpu_clock_event_del,
	.start		= cpu_clock_event_start,
	.stop		= cpu_clock_event_stop,
6474 6475 6476 6477 6478 6479 6480 6481
	.read		= cpu_clock_event_read,
};

/*
 * Software event: task time clock
 */

static void task_clock_event_update(struct perf_event *event, u64 now)
6482
{
6483 6484
	u64 prev;
	s64 delta;
6485

6486 6487 6488 6489
	prev = local64_xchg(&event->hw.prev_count, now);
	delta = now - prev;
	local64_add(delta, &event->count);
}
6490

P
Peter Zijlstra 已提交
6491
static void task_clock_event_start(struct perf_event *event, int flags)
6492
{
P
Peter Zijlstra 已提交
6493
	local64_set(&event->hw.prev_count, event->ctx->time);
6494 6495 6496
	perf_swevent_start_hrtimer(event);
}

P
Peter Zijlstra 已提交
6497
static void task_clock_event_stop(struct perf_event *event, int flags)
6498 6499 6500
{
	perf_swevent_cancel_hrtimer(event);
	task_clock_event_update(event, event->ctx->time);
P
Peter Zijlstra 已提交
6501 6502 6503 6504 6505 6506
}

static int task_clock_event_add(struct perf_event *event, int flags)
{
	if (flags & PERF_EF_START)
		task_clock_event_start(event, flags);
6507

P
Peter Zijlstra 已提交
6508 6509 6510 6511 6512 6513
	return 0;
}

static void task_clock_event_del(struct perf_event *event, int flags)
{
	task_clock_event_stop(event, PERF_EF_UPDATE);
6514 6515 6516 6517
}

static void task_clock_event_read(struct perf_event *event)
{
6518 6519 6520
	u64 now = perf_clock();
	u64 delta = now - event->ctx->timestamp;
	u64 time = event->ctx->time + delta;
6521 6522 6523 6524 6525

	task_clock_event_update(event, time);
}

static int task_clock_event_init(struct perf_event *event)
L
Li Zefan 已提交
6526
{
6527 6528 6529 6530 6531 6532
	if (event->attr.type != PERF_TYPE_SOFTWARE)
		return -ENOENT;

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

6533 6534 6535 6536 6537 6538
	/*
	 * no branch sampling for software events
	 */
	if (has_branch_stack(event))
		return -EOPNOTSUPP;

P
Peter Zijlstra 已提交
6539 6540
	perf_swevent_init_hrtimer(event);

6541
	return 0;
L
Li Zefan 已提交
6542 6543
}

6544
static struct pmu perf_task_clock = {
6545 6546
	.task_ctx_nr	= perf_sw_context,

6547
	.event_init	= task_clock_event_init,
P
Peter Zijlstra 已提交
6548 6549 6550 6551
	.add		= task_clock_event_add,
	.del		= task_clock_event_del,
	.start		= task_clock_event_start,
	.stop		= task_clock_event_stop,
6552 6553
	.read		= task_clock_event_read,
};
L
Li Zefan 已提交
6554

P
Peter Zijlstra 已提交
6555
static void perf_pmu_nop_void(struct pmu *pmu)
6556 6557
{
}
L
Li Zefan 已提交
6558

P
Peter Zijlstra 已提交
6559
static int perf_pmu_nop_int(struct pmu *pmu)
L
Li Zefan 已提交
6560
{
P
Peter Zijlstra 已提交
6561
	return 0;
L
Li Zefan 已提交
6562 6563
}

P
Peter Zijlstra 已提交
6564
static void perf_pmu_start_txn(struct pmu *pmu)
L
Li Zefan 已提交
6565
{
P
Peter Zijlstra 已提交
6566
	perf_pmu_disable(pmu);
L
Li Zefan 已提交
6567 6568
}

P
Peter Zijlstra 已提交
6569 6570 6571 6572 6573
static int perf_pmu_commit_txn(struct pmu *pmu)
{
	perf_pmu_enable(pmu);
	return 0;
}
6574

P
Peter Zijlstra 已提交
6575
static void perf_pmu_cancel_txn(struct pmu *pmu)
6576
{
P
Peter Zijlstra 已提交
6577
	perf_pmu_enable(pmu);
6578 6579
}

6580 6581
static int perf_event_idx_default(struct perf_event *event)
{
6582
	return 0;
6583 6584
}

P
Peter Zijlstra 已提交
6585 6586 6587 6588
/*
 * Ensures all contexts with the same task_ctx_nr have the same
 * pmu_cpu_context too.
 */
6589
static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6590
{
P
Peter Zijlstra 已提交
6591
	struct pmu *pmu;
6592

P
Peter Zijlstra 已提交
6593 6594
	if (ctxn < 0)
		return NULL;
6595

P
Peter Zijlstra 已提交
6596 6597 6598 6599
	list_for_each_entry(pmu, &pmus, entry) {
		if (pmu->task_ctx_nr == ctxn)
			return pmu->pmu_cpu_context;
	}
6600

P
Peter Zijlstra 已提交
6601
	return NULL;
6602 6603
}

6604
static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6605
{
6606 6607 6608 6609 6610 6611 6612
	int cpu;

	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

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

6613 6614
		if (cpuctx->unique_pmu == old_pmu)
			cpuctx->unique_pmu = pmu;
6615 6616 6617 6618 6619 6620
	}
}

static void free_pmu_context(struct pmu *pmu)
{
	struct pmu *i;
6621

P
Peter Zijlstra 已提交
6622
	mutex_lock(&pmus_lock);
6623
	/*
P
Peter Zijlstra 已提交
6624
	 * Like a real lame refcount.
6625
	 */
6626 6627 6628
	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
6629
			goto out;
6630
		}
P
Peter Zijlstra 已提交
6631
	}
6632

6633
	free_percpu(pmu->pmu_cpu_context);
P
Peter Zijlstra 已提交
6634 6635
out:
	mutex_unlock(&pmus_lock);
6636
}
P
Peter Zijlstra 已提交
6637
static struct idr pmu_idr;
6638

P
Peter Zijlstra 已提交
6639 6640 6641 6642 6643 6644 6645
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);
}
6646
static DEVICE_ATTR_RO(type);
P
Peter Zijlstra 已提交
6647

6648 6649 6650 6651 6652 6653 6654 6655 6656 6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686 6687 6688 6689 6690
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;
}
6691
static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6692

6693 6694 6695 6696
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
6697
};
6698
ATTRIBUTE_GROUPS(pmu_dev);
P
Peter Zijlstra 已提交
6699 6700 6701 6702

static int pmu_bus_running;
static struct bus_type pmu_bus = {
	.name		= "event_source",
6703
	.dev_groups	= pmu_dev_groups,
P
Peter Zijlstra 已提交
6704 6705 6706 6707 6708 6709 6710 6711 6712 6713 6714 6715 6716 6717 6718
};

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;

6719
	pmu->dev->groups = pmu->attr_groups;
P
Peter Zijlstra 已提交
6720 6721 6722 6723 6724 6725 6726 6727 6728 6729 6730 6731 6732 6733 6734 6735 6736 6737 6738 6739
	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;
}

6740
static struct lock_class_key cpuctx_mutex;
6741
static struct lock_class_key cpuctx_lock;
6742

6743
int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6744
{
P
Peter Zijlstra 已提交
6745
	int cpu, ret;
6746

6747
	mutex_lock(&pmus_lock);
P
Peter Zijlstra 已提交
6748 6749 6750 6751
	ret = -ENOMEM;
	pmu->pmu_disable_count = alloc_percpu(int);
	if (!pmu->pmu_disable_count)
		goto unlock;
6752

P
Peter Zijlstra 已提交
6753 6754 6755 6756 6757 6758
	pmu->type = -1;
	if (!name)
		goto skip_type;
	pmu->name = name;

	if (type < 0) {
T
Tejun Heo 已提交
6759 6760 6761
		type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
		if (type < 0) {
			ret = type;
P
Peter Zijlstra 已提交
6762 6763 6764 6765 6766
			goto free_pdc;
		}
	}
	pmu->type = type;

P
Peter Zijlstra 已提交
6767 6768 6769 6770 6771 6772
	if (pmu_bus_running) {
		ret = pmu_dev_alloc(pmu);
		if (ret)
			goto free_idr;
	}

P
Peter Zijlstra 已提交
6773
skip_type:
P
Peter Zijlstra 已提交
6774 6775 6776
	pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
	if (pmu->pmu_cpu_context)
		goto got_cpu_context;
6777

W
Wei Yongjun 已提交
6778
	ret = -ENOMEM;
P
Peter Zijlstra 已提交
6779 6780
	pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
	if (!pmu->pmu_cpu_context)
P
Peter Zijlstra 已提交
6781
		goto free_dev;
6782

P
Peter Zijlstra 已提交
6783 6784 6785 6786
	for_each_possible_cpu(cpu) {
		struct perf_cpu_context *cpuctx;

		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6787
		__perf_event_init_context(&cpuctx->ctx);
6788
		lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6789
		lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
P
Peter Zijlstra 已提交
6790
		cpuctx->ctx.pmu = pmu;
6791 6792 6793

		__perf_cpu_hrtimer_init(cpuctx, cpu);

6794
		INIT_LIST_HEAD(&cpuctx->rotation_list);
6795
		cpuctx->unique_pmu = pmu;
P
Peter Zijlstra 已提交
6796
	}
6797

P
Peter Zijlstra 已提交
6798
got_cpu_context:
P
Peter Zijlstra 已提交
6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812
	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;
6813
		}
6814
	}
6815

P
Peter Zijlstra 已提交
6816 6817 6818 6819 6820
	if (!pmu->pmu_enable) {
		pmu->pmu_enable  = perf_pmu_nop_void;
		pmu->pmu_disable = perf_pmu_nop_void;
	}

6821 6822 6823
	if (!pmu->event_idx)
		pmu->event_idx = perf_event_idx_default;

6824
	list_add_rcu(&pmu->entry, &pmus);
P
Peter Zijlstra 已提交
6825 6826
	ret = 0;
unlock:
6827 6828
	mutex_unlock(&pmus_lock);

P
Peter Zijlstra 已提交
6829
	return ret;
P
Peter Zijlstra 已提交
6830

P
Peter Zijlstra 已提交
6831 6832 6833 6834
free_dev:
	device_del(pmu->dev);
	put_device(pmu->dev);

P
Peter Zijlstra 已提交
6835 6836 6837 6838
free_idr:
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);

P
Peter Zijlstra 已提交
6839 6840 6841
free_pdc:
	free_percpu(pmu->pmu_disable_count);
	goto unlock;
6842
}
6843
EXPORT_SYMBOL_GPL(perf_pmu_register);
6844

6845
void perf_pmu_unregister(struct pmu *pmu)
6846
{
6847 6848 6849
	mutex_lock(&pmus_lock);
	list_del_rcu(&pmu->entry);
	mutex_unlock(&pmus_lock);
6850

6851
	/*
P
Peter Zijlstra 已提交
6852 6853
	 * We dereference the pmu list under both SRCU and regular RCU, so
	 * synchronize against both of those.
6854
	 */
6855
	synchronize_srcu(&pmus_srcu);
P
Peter Zijlstra 已提交
6856
	synchronize_rcu();
6857

P
Peter Zijlstra 已提交
6858
	free_percpu(pmu->pmu_disable_count);
P
Peter Zijlstra 已提交
6859 6860
	if (pmu->type >= PERF_TYPE_MAX)
		idr_remove(&pmu_idr, pmu->type);
P
Peter Zijlstra 已提交
6861 6862
	device_del(pmu->dev);
	put_device(pmu->dev);
6863
	free_pmu_context(pmu);
6864
}
6865
EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6866

6867 6868 6869 6870
struct pmu *perf_init_event(struct perf_event *event)
{
	struct pmu *pmu = NULL;
	int idx;
6871
	int ret;
6872 6873

	idx = srcu_read_lock(&pmus_srcu);
P
Peter Zijlstra 已提交
6874 6875 6876 6877

	rcu_read_lock();
	pmu = idr_find(&pmu_idr, event->attr.type);
	rcu_read_unlock();
6878
	if (pmu) {
6879 6880 6881 6882
		if (!try_module_get(pmu->module)) {
			pmu = ERR_PTR(-ENODEV);
			goto unlock;
		}
6883
		event->pmu = pmu;
6884 6885 6886
		ret = pmu->event_init(event);
		if (ret)
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6887
		goto unlock;
6888
	}
P
Peter Zijlstra 已提交
6889

6890
	list_for_each_entry_rcu(pmu, &pmus, entry) {
6891 6892 6893 6894
		if (!try_module_get(pmu->module)) {
			pmu = ERR_PTR(-ENODEV);
			goto unlock;
		}
6895
		event->pmu = pmu;
6896
		ret = pmu->event_init(event);
6897
		if (!ret)
P
Peter Zijlstra 已提交
6898
			goto unlock;
6899

6900 6901
		if (ret != -ENOENT) {
			pmu = ERR_PTR(ret);
P
Peter Zijlstra 已提交
6902
			goto unlock;
6903
		}
6904
	}
P
Peter Zijlstra 已提交
6905 6906
	pmu = ERR_PTR(-ENOENT);
unlock:
6907
	srcu_read_unlock(&pmus_srcu, idx);
6908

6909
	return pmu;
6910 6911
}

6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924
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));
}

6925 6926
static void account_event(struct perf_event *event)
{
6927 6928 6929
	if (event->parent)
		return;

6930 6931 6932 6933 6934 6935 6936 6937
	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);
6938 6939 6940 6941
	if (event->attr.freq) {
		if (atomic_inc_return(&nr_freq_events) == 1)
			tick_nohz_full_kick_all();
	}
6942
	if (has_branch_stack(event))
6943
		static_key_slow_inc(&perf_sched_events.key);
6944
	if (is_cgroup_event(event))
6945
		static_key_slow_inc(&perf_sched_events.key);
6946 6947

	account_event_cpu(event, event->cpu);
6948 6949
}

T
Thomas Gleixner 已提交
6950
/*
6951
 * Allocate and initialize a event structure
T
Thomas Gleixner 已提交
6952
 */
6953
static struct perf_event *
6954
perf_event_alloc(struct perf_event_attr *attr, int cpu,
6955 6956 6957
		 struct task_struct *task,
		 struct perf_event *group_leader,
		 struct perf_event *parent_event,
6958 6959
		 perf_overflow_handler_t overflow_handler,
		 void *context)
T
Thomas Gleixner 已提交
6960
{
P
Peter Zijlstra 已提交
6961
	struct pmu *pmu;
6962 6963
	struct perf_event *event;
	struct hw_perf_event *hwc;
6964
	long err = -EINVAL;
T
Thomas Gleixner 已提交
6965

6966 6967 6968 6969 6970
	if ((unsigned)cpu >= nr_cpu_ids) {
		if (!task || cpu != -1)
			return ERR_PTR(-EINVAL);
	}

6971
	event = kzalloc(sizeof(*event), GFP_KERNEL);
6972
	if (!event)
6973
		return ERR_PTR(-ENOMEM);
T
Thomas Gleixner 已提交
6974

6975
	/*
6976
	 * Single events are their own group leaders, with an
6977 6978 6979
	 * empty sibling list:
	 */
	if (!group_leader)
6980
		group_leader = event;
6981

6982 6983
	mutex_init(&event->child_mutex);
	INIT_LIST_HEAD(&event->child_list);
6984

6985 6986 6987
	INIT_LIST_HEAD(&event->group_entry);
	INIT_LIST_HEAD(&event->event_entry);
	INIT_LIST_HEAD(&event->sibling_list);
6988
	INIT_LIST_HEAD(&event->rb_entry);
6989
	INIT_LIST_HEAD(&event->active_entry);
6990 6991
	INIT_HLIST_NODE(&event->hlist_entry);

6992

6993
	init_waitqueue_head(&event->waitq);
6994
	init_irq_work(&event->pending, perf_pending_event);
T
Thomas Gleixner 已提交
6995

6996
	mutex_init(&event->mmap_mutex);
6997

6998
	atomic_long_set(&event->refcount, 1);
6999 7000 7001 7002 7003
	event->cpu		= cpu;
	event->attr		= *attr;
	event->group_leader	= group_leader;
	event->pmu		= NULL;
	event->oncpu		= -1;
7004

7005
	event->parent		= parent_event;
7006

7007
	event->ns		= get_pid_ns(task_active_pid_ns(current));
7008
	event->id		= atomic64_inc_return(&perf_event_id);
7009

7010
	event->state		= PERF_EVENT_STATE_INACTIVE;
7011

7012 7013
	if (task) {
		event->attach_state = PERF_ATTACH_TASK;
7014 7015 7016

		if (attr->type == PERF_TYPE_TRACEPOINT)
			event->hw.tp_target = task;
7017 7018 7019 7020
#ifdef CONFIG_HAVE_HW_BREAKPOINT
		/*
		 * hw_breakpoint is a bit difficult here..
		 */
7021
		else if (attr->type == PERF_TYPE_BREAKPOINT)
7022 7023 7024 7025
			event->hw.bp_target = task;
#endif
	}

7026
	if (!overflow_handler && parent_event) {
7027
		overflow_handler = parent_event->overflow_handler;
7028 7029
		context = parent_event->overflow_handler_context;
	}
7030

7031
	event->overflow_handler	= overflow_handler;
7032
	event->overflow_handler_context = context;
7033

J
Jiri Olsa 已提交
7034
	perf_event__state_init(event);
7035

7036
	pmu = NULL;
7037

7038
	hwc = &event->hw;
7039
	hwc->sample_period = attr->sample_period;
7040
	if (attr->freq && attr->sample_freq)
7041
		hwc->sample_period = 1;
7042
	hwc->last_period = hwc->sample_period;
7043

7044
	local64_set(&hwc->period_left, hwc->sample_period);
7045

7046
	/*
7047
	 * we currently do not support PERF_FORMAT_GROUP on inherited events
7048
	 */
7049
	if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7050
		goto err_ns;
7051

7052
	pmu = perf_init_event(event);
7053
	if (!pmu)
7054 7055
		goto err_ns;
	else if (IS_ERR(pmu)) {
7056
		err = PTR_ERR(pmu);
7057
		goto err_ns;
I
Ingo Molnar 已提交
7058
	}
7059

7060
	if (!event->parent) {
7061 7062
		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
			err = get_callchain_buffers();
7063 7064
			if (err)
				goto err_pmu;
7065
		}
7066
	}
7067

7068
	return event;
7069 7070 7071 7072

err_pmu:
	if (event->destroy)
		event->destroy(event);
7073
	module_put(pmu->module);
7074 7075 7076 7077 7078 7079
err_ns:
	if (event->ns)
		put_pid_ns(event->ns);
	kfree(event);

	return ERR_PTR(err);
T
Thomas Gleixner 已提交
7080 7081
}

7082 7083
static int perf_copy_attr(struct perf_event_attr __user *uattr,
			  struct perf_event_attr *attr)
7084 7085
{
	u32 size;
7086
	int ret;
7087 7088 7089 7090 7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110

	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,
7111 7112 7113
	 * 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.
7114 7115
	 */
	if (size > sizeof(*attr)) {
7116 7117 7118
		unsigned char __user *addr;
		unsigned char __user *end;
		unsigned char val;
7119

7120 7121
		addr = (void __user *)uattr + sizeof(*attr);
		end  = (void __user *)uattr + size;
7122

7123
		for (; addr < end; addr++) {
7124 7125 7126 7127 7128 7129
			ret = get_user(val, addr);
			if (ret)
				return ret;
			if (val)
				goto err_size;
		}
7130
		size = sizeof(*attr);
7131 7132 7133 7134 7135 7136
	}

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

7137
	if (attr->__reserved_1)
7138 7139 7140 7141 7142 7143 7144 7145
		return -EINVAL;

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

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

7146 7147 7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164 7165 7166 7167 7168 7169 7170 7171 7172 7173
	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;
		}
7174 7175
		/* privileged levels capture (kernel, hv): check permissions */
		if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7176 7177
		    && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
7178
	}
7179

7180
	if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7181
		ret = perf_reg_validate(attr->sample_regs_user);
7182 7183 7184 7185 7186 7187 7188 7189 7190 7191 7192 7193 7194 7195 7196 7197 7198 7199
		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;
	}
7200

7201 7202
	if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
		ret = perf_reg_validate(attr->sample_regs_intr);
7203 7204 7205 7206 7207 7208 7209 7210 7211
out:
	return ret;

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

7212 7213
static int
perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7214
{
7215
	struct ring_buffer *rb = NULL;
7216 7217
	int ret = -EINVAL;

7218
	if (!output_event)
7219 7220
		goto set;

7221 7222
	/* don't allow circular references */
	if (event == output_event)
7223 7224
		goto out;

7225 7226 7227 7228 7229 7230 7231
	/*
	 * Don't allow cross-cpu buffers
	 */
	if (output_event->cpu != event->cpu)
		goto out;

	/*
7232
	 * If its not a per-cpu rb, it must be the same task.
7233 7234 7235 7236
	 */
	if (output_event->cpu == -1 && output_event->ctx != event->ctx)
		goto out;

7237
set:
7238
	mutex_lock(&event->mmap_mutex);
7239 7240 7241
	/* Can't redirect output if we've got an active mmap() */
	if (atomic_read(&event->mmap_count))
		goto unlock;
7242

7243
	if (output_event) {
7244 7245 7246
		/* get the rb we want to redirect to */
		rb = ring_buffer_get(output_event);
		if (!rb)
7247
			goto unlock;
7248 7249
	}

7250
	ring_buffer_attach(event, rb);
7251

7252
	ret = 0;
7253 7254 7255
unlock:
	mutex_unlock(&event->mmap_mutex);

7256 7257 7258 7259
out:
	return ret;
}

T
Thomas Gleixner 已提交
7260
/**
7261
 * sys_perf_event_open - open a performance event, associate it to a task/cpu
I
Ingo Molnar 已提交
7262
 *
7263
 * @attr_uptr:	event_id type attributes for monitoring/sampling
T
Thomas Gleixner 已提交
7264
 * @pid:		target pid
I
Ingo Molnar 已提交
7265
 * @cpu:		target cpu
7266
 * @group_fd:		group leader event fd
T
Thomas Gleixner 已提交
7267
 */
7268 7269
SYSCALL_DEFINE5(perf_event_open,
		struct perf_event_attr __user *, attr_uptr,
7270
		pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
T
Thomas Gleixner 已提交
7271
{
7272 7273
	struct perf_event *group_leader = NULL, *output_event = NULL;
	struct perf_event *event, *sibling;
7274 7275 7276
	struct perf_event_attr attr;
	struct perf_event_context *ctx;
	struct file *event_file = NULL;
7277
	struct fd group = {NULL, 0};
M
Matt Helsley 已提交
7278
	struct task_struct *task = NULL;
7279
	struct pmu *pmu;
7280
	int event_fd;
7281
	int move_group = 0;
7282
	int err;
7283
	int f_flags = O_RDWR;
T
Thomas Gleixner 已提交
7284

7285
	/* for future expandability... */
S
Stephane Eranian 已提交
7286
	if (flags & ~PERF_FLAG_ALL)
7287 7288
		return -EINVAL;

7289 7290 7291
	err = perf_copy_attr(attr_uptr, &attr);
	if (err)
		return err;
7292

7293 7294 7295 7296 7297
	if (!attr.exclude_kernel) {
		if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
			return -EACCES;
	}

7298
	if (attr.freq) {
7299
		if (attr.sample_freq > sysctl_perf_event_sample_rate)
7300
			return -EINVAL;
7301 7302 7303
	} else {
		if (attr.sample_period & (1ULL << 63))
			return -EINVAL;
7304 7305
	}

S
Stephane Eranian 已提交
7306 7307 7308 7309 7310 7311 7312 7313 7314
	/*
	 * 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;

7315 7316 7317 7318
	if (flags & PERF_FLAG_FD_CLOEXEC)
		f_flags |= O_CLOEXEC;

	event_fd = get_unused_fd_flags(f_flags);
7319 7320 7321
	if (event_fd < 0)
		return event_fd;

7322
	if (group_fd != -1) {
7323 7324
		err = perf_fget_light(group_fd, &group);
		if (err)
7325
			goto err_fd;
7326
		group_leader = group.file->private_data;
7327 7328 7329 7330 7331 7332
		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
7333
	if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7334 7335 7336 7337 7338 7339 7340
		task = find_lively_task_by_vpid(pid);
		if (IS_ERR(task)) {
			err = PTR_ERR(task);
			goto err_group_fd;
		}
	}

7341 7342 7343 7344 7345 7346
	if (task && group_leader &&
	    group_leader->attr.inherit != attr.inherit) {
		err = -EINVAL;
		goto err_task;
	}

7347 7348
	get_online_cpus();

7349 7350
	event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
				 NULL, NULL);
7351 7352
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
7353
		goto err_cpus;
7354 7355
	}

S
Stephane Eranian 已提交
7356 7357
	if (flags & PERF_FLAG_PID_CGROUP) {
		err = perf_cgroup_connect(pid, event, &attr, group_leader);
7358 7359
		if (err) {
			__free_event(event);
7360
			goto err_cpus;
7361
		}
S
Stephane Eranian 已提交
7362 7363
	}

7364 7365 7366 7367 7368 7369 7370
	if (is_sampling_event(event)) {
		if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
			err = -ENOTSUPP;
			goto err_alloc;
		}
	}

7371 7372
	account_event(event);

7373 7374 7375 7376 7377
	/*
	 * Special case software events and allow them to be part of
	 * any hardware group.
	 */
	pmu = event->pmu;
7378 7379 7380 7381 7382 7383 7384 7385 7386 7387 7388 7389 7390 7391 7392 7393 7394 7395 7396 7397 7398 7399 7400

	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;
		}
	}
7401 7402 7403 7404

	/*
	 * Get the target context (task or percpu):
	 */
7405
	ctx = find_get_context(pmu, task, event->cpu);
7406 7407
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7408
		goto err_alloc;
7409 7410
	}

7411 7412 7413 7414 7415
	if (task) {
		put_task_struct(task);
		task = NULL;
	}

I
Ingo Molnar 已提交
7416
	/*
7417
	 * Look up the group leader (we will attach this event to it):
7418
	 */
7419
	if (group_leader) {
7420
		err = -EINVAL;
7421 7422

		/*
I
Ingo Molnar 已提交
7423 7424 7425 7426
		 * Do not allow a recursive hierarchy (this new sibling
		 * becoming part of another group-sibling):
		 */
		if (group_leader->group_leader != group_leader)
7427
			goto err_context;
I
Ingo Molnar 已提交
7428 7429 7430
		/*
		 * Do not allow to attach to a group in a different
		 * task or CPU context:
7431
		 */
7432
		if (move_group) {
7433 7434 7435 7436 7437 7438 7439 7440 7441 7442 7443 7444 7445
			/*
			 * Make sure we're both on the same task, or both
			 * per-cpu events.
			 */
			if (group_leader->ctx->task != ctx->task)
				goto err_context;

			/*
			 * Make sure we're both events for the same CPU;
			 * grouping events for different CPUs is broken; since
			 * you can never concurrently schedule them anyhow.
			 */
			if (group_leader->cpu != event->cpu)
7446 7447 7448 7449 7450 7451
				goto err_context;
		} else {
			if (group_leader->ctx != ctx)
				goto err_context;
		}

7452 7453 7454
		/*
		 * Only a group leader can be exclusive or pinned
		 */
7455
		if (attr.exclusive || attr.pinned)
7456
			goto err_context;
7457 7458 7459 7460 7461
	}

	if (output_event) {
		err = perf_event_set_output(event, output_event);
		if (err)
7462
			goto err_context;
7463
	}
T
Thomas Gleixner 已提交
7464

7465 7466
	event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
					f_flags);
7467 7468
	if (IS_ERR(event_file)) {
		err = PTR_ERR(event_file);
7469
		goto err_context;
7470
	}
7471

7472 7473 7474 7475
	if (move_group) {
		struct perf_event_context *gctx = group_leader->ctx;

		mutex_lock(&gctx->mutex);
7476
		perf_remove_from_context(group_leader, false);
J
Jiri Olsa 已提交
7477 7478 7479 7480 7481 7482 7483

		/*
		 * 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);
7484 7485
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7486
			perf_remove_from_context(sibling, false);
J
Jiri Olsa 已提交
7487
			perf_event__state_init(sibling);
7488 7489 7490 7491
			put_ctx(gctx);
		}
		mutex_unlock(&gctx->mutex);
		put_ctx(gctx);
7492
	}
7493

7494
	WARN_ON_ONCE(ctx->parent_ctx);
7495
	mutex_lock(&ctx->mutex);
7496 7497

	if (move_group) {
7498
		synchronize_rcu();
7499
		perf_install_in_context(ctx, group_leader, group_leader->cpu);
7500 7501 7502
		get_ctx(ctx);
		list_for_each_entry(sibling, &group_leader->sibling_list,
				    group_entry) {
7503
			perf_install_in_context(ctx, sibling, sibling->cpu);
7504 7505 7506 7507
			get_ctx(ctx);
		}
	}

7508
	perf_install_in_context(ctx, event, event->cpu);
7509
	perf_unpin_context(ctx);
7510
	mutex_unlock(&ctx->mutex);
7511

7512 7513
	put_online_cpus();

7514
	event->owner = current;
P
Peter Zijlstra 已提交
7515

7516 7517 7518
	mutex_lock(&current->perf_event_mutex);
	list_add_tail(&event->owner_entry, &current->perf_event_list);
	mutex_unlock(&current->perf_event_mutex);
7519

7520 7521 7522 7523
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(event);
7524
	perf_event__id_header_size(event);
7525

7526 7527 7528 7529 7530 7531
	/*
	 * 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().
	 */
7532
	fdput(group);
7533 7534
	fd_install(event_fd, event_file);
	return event_fd;
T
Thomas Gleixner 已提交
7535

7536
err_context:
7537
	perf_unpin_context(ctx);
7538
	put_ctx(ctx);
7539
err_alloc:
7540
	free_event(event);
7541
err_cpus:
7542
	put_online_cpus();
7543
err_task:
P
Peter Zijlstra 已提交
7544 7545
	if (task)
		put_task_struct(task);
7546
err_group_fd:
7547
	fdput(group);
7548 7549
err_fd:
	put_unused_fd(event_fd);
7550
	return err;
T
Thomas Gleixner 已提交
7551 7552
}

7553 7554 7555 7556 7557
/**
 * 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
7558
 * @task: task to profile (NULL for percpu)
7559 7560 7561
 */
struct perf_event *
perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
M
Matt Helsley 已提交
7562
				 struct task_struct *task,
7563 7564
				 perf_overflow_handler_t overflow_handler,
				 void *context)
7565 7566
{
	struct perf_event_context *ctx;
7567
	struct perf_event *event;
7568
	int err;
7569

7570 7571 7572
	/*
	 * Get the target context (task or percpu):
	 */
7573

7574 7575
	event = perf_event_alloc(attr, cpu, task, NULL, NULL,
				 overflow_handler, context);
7576 7577 7578 7579
	if (IS_ERR(event)) {
		err = PTR_ERR(event);
		goto err;
	}
7580

7581 7582 7583
	/* Mark owner so we could distinguish it from user events. */
	event->owner = EVENT_OWNER_KERNEL;

7584 7585
	account_event(event);

M
Matt Helsley 已提交
7586
	ctx = find_get_context(event->pmu, task, cpu);
7587 7588
	if (IS_ERR(ctx)) {
		err = PTR_ERR(ctx);
7589
		goto err_free;
7590
	}
7591 7592 7593 7594

	WARN_ON_ONCE(ctx->parent_ctx);
	mutex_lock(&ctx->mutex);
	perf_install_in_context(ctx, event, cpu);
7595
	perf_unpin_context(ctx);
7596 7597 7598 7599
	mutex_unlock(&ctx->mutex);

	return event;

7600 7601 7602
err_free:
	free_event(event);
err:
7603
	return ERR_PTR(err);
7604
}
7605
EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7606

7607 7608 7609 7610 7611 7612 7613 7614 7615 7616 7617 7618 7619
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) {
7620
		perf_remove_from_context(event, false);
7621
		unaccount_event_cpu(event, src_cpu);
7622
		put_ctx(src_ctx);
7623
		list_add(&event->migrate_entry, &events);
7624 7625 7626 7627 7628 7629
	}
	mutex_unlock(&src_ctx->mutex);

	synchronize_rcu();

	mutex_lock(&dst_ctx->mutex);
7630 7631
	list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
		list_del(&event->migrate_entry);
7632 7633
		if (event->state >= PERF_EVENT_STATE_OFF)
			event->state = PERF_EVENT_STATE_INACTIVE;
7634
		account_event_cpu(event, dst_cpu);
7635 7636 7637 7638 7639 7640 7641
		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);

7642
static void sync_child_event(struct perf_event *child_event,
7643
			       struct task_struct *child)
7644
{
7645
	struct perf_event *parent_event = child_event->parent;
7646
	u64 child_val;
7647

7648 7649
	if (child_event->attr.inherit_stat)
		perf_event_read_event(child_event, child);
7650

P
Peter Zijlstra 已提交
7651
	child_val = perf_event_count(child_event);
7652 7653 7654 7655

	/*
	 * Add back the child's count to the parent's count:
	 */
7656
	atomic64_add(child_val, &parent_event->child_count);
7657 7658 7659 7660
	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);
7661 7662

	/*
7663
	 * Remove this event from the parent's list
7664
	 */
7665 7666 7667 7668
	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);
7669

7670 7671 7672 7673 7674 7675
	/*
	 * Make sure user/parent get notified, that we just
	 * lost one event.
	 */
	perf_event_wakeup(parent_event);

7676
	/*
7677
	 * Release the parent event, if this was the last
7678 7679
	 * reference to it.
	 */
7680
	put_event(parent_event);
7681 7682
}

7683
static void
7684 7685
__perf_event_exit_task(struct perf_event *child_event,
			 struct perf_event_context *child_ctx,
7686
			 struct task_struct *child)
7687
{
7688 7689 7690 7691 7692 7693 7694 7695 7696 7697 7698 7699 7700
	/*
	 * 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);
7701

7702
	/*
7703
	 * It can happen that the parent exits first, and has events
7704
	 * that are still around due to the child reference. These
7705
	 * events need to be zapped.
7706
	 */
7707
	if (child_event->parent) {
7708 7709
		sync_child_event(child_event, child);
		free_event(child_event);
7710 7711 7712
	} else {
		child_event->state = PERF_EVENT_STATE_EXIT;
		perf_event_wakeup(child_event);
7713
	}
7714 7715
}

P
Peter Zijlstra 已提交
7716
static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7717
{
7718
	struct perf_event *child_event, *next;
7719
	struct perf_event_context *child_ctx, *clone_ctx = NULL;
7720
	unsigned long flags;
7721

P
Peter Zijlstra 已提交
7722
	if (likely(!child->perf_event_ctxp[ctxn])) {
7723
		perf_event_task(child, NULL, 0);
7724
		return;
P
Peter Zijlstra 已提交
7725
	}
7726

7727
	local_irq_save(flags);
7728 7729 7730 7731 7732 7733
	/*
	 * 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.
	 */
7734
	child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7735 7736 7737

	/*
	 * Take the context lock here so that if find_get_context is
7738
	 * reading child->perf_event_ctxp, we wait until it has
7739 7740
	 * incremented the context's refcount before we do put_ctx below.
	 */
7741
	raw_spin_lock(&child_ctx->lock);
7742
	task_ctx_sched_out(child_ctx);
P
Peter Zijlstra 已提交
7743
	child->perf_event_ctxp[ctxn] = NULL;
7744

7745 7746 7747
	/*
	 * If this context is a clone; unclone it so it can't get
	 * swapped to another process while we're removing all
7748
	 * the events from it.
7749
	 */
7750
	clone_ctx = unclone_ctx(child_ctx);
7751
	update_context_time(child_ctx);
7752
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7753

7754 7755
	if (clone_ctx)
		put_ctx(clone_ctx);
7756

P
Peter Zijlstra 已提交
7757
	/*
7758 7759 7760
	 * 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
7761
	 */
7762
	perf_event_task(child, child_ctx, 0);
7763

7764 7765 7766
	/*
	 * We can recurse on the same lock type through:
	 *
7767 7768
	 *   __perf_event_exit_task()
	 *     sync_child_event()
7769 7770
	 *       put_event()
	 *         mutex_lock(&ctx->mutex)
7771 7772 7773
	 *
	 * But since its the parent context it won't be the same instance.
	 */
7774
	mutex_lock(&child_ctx->mutex);
7775

7776
	list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7777
		__perf_event_exit_task(child_event, child_ctx, child);
7778

7779 7780 7781
	mutex_unlock(&child_ctx->mutex);

	put_ctx(child_ctx);
7782 7783
}

P
Peter Zijlstra 已提交
7784 7785 7786 7787 7788
/*
 * 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
7789
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7790 7791
	int ctxn;

P
Peter Zijlstra 已提交
7792 7793 7794 7795 7796 7797 7798 7799 7800 7801 7802 7803 7804 7805 7806
	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
7807 7808 7809 7810
	for_each_task_context_nr(ctxn)
		perf_event_exit_task_context(child, ctxn);
}

7811 7812 7813 7814 7815 7816 7817 7818 7819 7820 7821 7822
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);

7823
	put_event(parent);
7824

7825
	perf_group_detach(event);
7826 7827 7828 7829
	list_del_event(event, ctx);
	free_event(event);
}

7830 7831
/*
 * free an unexposed, unused context as created by inheritance by
P
Peter Zijlstra 已提交
7832
 * perf_event_init_task below, used by fork() in case of fail.
7833
 */
7834
void perf_event_free_task(struct task_struct *task)
7835
{
P
Peter Zijlstra 已提交
7836
	struct perf_event_context *ctx;
7837
	struct perf_event *event, *tmp;
P
Peter Zijlstra 已提交
7838
	int ctxn;
7839

P
Peter Zijlstra 已提交
7840 7841 7842 7843
	for_each_task_context_nr(ctxn) {
		ctx = task->perf_event_ctxp[ctxn];
		if (!ctx)
			continue;
7844

P
Peter Zijlstra 已提交
7845
		mutex_lock(&ctx->mutex);
7846
again:
P
Peter Zijlstra 已提交
7847 7848 7849
		list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
				group_entry)
			perf_free_event(event, ctx);
7850

P
Peter Zijlstra 已提交
7851 7852 7853
		list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
				group_entry)
			perf_free_event(event, ctx);
7854

P
Peter Zijlstra 已提交
7855 7856 7857
		if (!list_empty(&ctx->pinned_groups) ||
				!list_empty(&ctx->flexible_groups))
			goto again;
7858

P
Peter Zijlstra 已提交
7859
		mutex_unlock(&ctx->mutex);
7860

P
Peter Zijlstra 已提交
7861 7862
		put_ctx(ctx);
	}
7863 7864
}

7865 7866 7867 7868 7869 7870 7871 7872
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
7873 7874 7875 7876 7877 7878 7879 7880 7881 7882 7883
/*
 * 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)
{
7884
	enum perf_event_active_state parent_state = parent_event->state;
P
Peter Zijlstra 已提交
7885
	struct perf_event *child_event;
7886
	unsigned long flags;
P
Peter Zijlstra 已提交
7887 7888 7889 7890 7891 7892 7893 7894 7895 7896 7897 7898

	/*
	 * 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,
7899
					   child,
P
Peter Zijlstra 已提交
7900
					   group_leader, parent_event,
7901
				           NULL, NULL);
P
Peter Zijlstra 已提交
7902 7903
	if (IS_ERR(child_event))
		return child_event;
7904

7905 7906
	if (is_orphaned_event(parent_event) ||
	    !atomic_long_inc_not_zero(&parent_event->refcount)) {
7907 7908 7909 7910
		free_event(child_event);
		return NULL;
	}

P
Peter Zijlstra 已提交
7911 7912 7913 7914 7915 7916 7917
	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.
	 */
7918
	if (parent_state >= PERF_EVENT_STATE_INACTIVE)
P
Peter Zijlstra 已提交
7919 7920 7921 7922 7923 7924 7925 7926 7927 7928 7929 7930 7931 7932 7933 7934
		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;
7935 7936
	child_event->overflow_handler_context
		= parent_event->overflow_handler_context;
P
Peter Zijlstra 已提交
7937

7938 7939 7940 7941
	/*
	 * Precalculate sample_data sizes
	 */
	perf_event__header_size(child_event);
7942
	perf_event__id_header_size(child_event);
7943

P
Peter Zijlstra 已提交
7944 7945 7946
	/*
	 * Link it up in the child's context:
	 */
7947
	raw_spin_lock_irqsave(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7948
	add_event_to_ctx(child_event, child_ctx);
7949
	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
P
Peter Zijlstra 已提交
7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960 7961 7962 7963 7964 7965 7966 7967 7968 7969 7970 7971 7972 7973 7974 7975 7976 7977 7978 7979 7980 7981 7982

	/*
	 * 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;
7983 7984 7985 7986 7987
}

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
7988
		   struct task_struct *child, int ctxn,
7989 7990 7991
		   int *inherited_all)
{
	int ret;
P
Peter Zijlstra 已提交
7992
	struct perf_event_context *child_ctx;
7993 7994 7995 7996

	if (!event->attr.inherit) {
		*inherited_all = 0;
		return 0;
7997 7998
	}

7999
	child_ctx = child->perf_event_ctxp[ctxn];
8000 8001 8002 8003 8004 8005 8006
	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.
		 */
8007

8008
		child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8009 8010
		if (!child_ctx)
			return -ENOMEM;
8011

P
Peter Zijlstra 已提交
8012
		child->perf_event_ctxp[ctxn] = child_ctx;
8013 8014 8015 8016 8017 8018 8019 8020 8021
	}

	ret = inherit_group(event, parent, parent_ctx,
			    child, child_ctx);

	if (ret)
		*inherited_all = 0;

	return ret;
8022 8023
}

8024
/*
8025
 * Initialize the perf_event context in task_struct
8026
 */
8027
static int perf_event_init_context(struct task_struct *child, int ctxn)
8028
{
8029
	struct perf_event_context *child_ctx, *parent_ctx;
8030 8031
	struct perf_event_context *cloned_ctx;
	struct perf_event *event;
8032
	struct task_struct *parent = current;
8033
	int inherited_all = 1;
8034
	unsigned long flags;
8035
	int ret = 0;
8036

P
Peter Zijlstra 已提交
8037
	if (likely(!parent->perf_event_ctxp[ctxn]))
8038 8039
		return 0;

8040
	/*
8041 8042
	 * If the parent's context is a clone, pin it so it won't get
	 * swapped under us.
8043
	 */
P
Peter Zijlstra 已提交
8044
	parent_ctx = perf_pin_task_context(parent, ctxn);
8045 8046
	if (!parent_ctx)
		return 0;
8047

8048 8049 8050 8051 8052 8053 8054
	/*
	 * 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.
	 */

8055 8056 8057 8058
	/*
	 * Lock the parent list. No need to lock the child - not PID
	 * hashed yet and not running, so nobody can access it.
	 */
8059
	mutex_lock(&parent_ctx->mutex);
8060 8061 8062 8063 8064

	/*
	 * We dont have to disable NMIs - we are only looking at
	 * the list, not manipulating it:
	 */
8065
	list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
P
Peter Zijlstra 已提交
8066 8067
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
8068 8069 8070
		if (ret)
			break;
	}
8071

8072 8073 8074 8075 8076 8077 8078 8079 8080
	/*
	 * 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);

8081
	list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
P
Peter Zijlstra 已提交
8082 8083
		ret = inherit_task_group(event, parent, parent_ctx,
					 child, ctxn, &inherited_all);
8084
		if (ret)
8085
			break;
8086 8087
	}

8088 8089 8090
	raw_spin_lock_irqsave(&parent_ctx->lock, flags);
	parent_ctx->rotate_disable = 0;

P
Peter Zijlstra 已提交
8091
	child_ctx = child->perf_event_ctxp[ctxn];
8092

8093
	if (child_ctx && inherited_all) {
8094 8095 8096
		/*
		 * 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
8097 8098 8099
		 *
		 * Note that if the parent is a clone, the holding of
		 * parent_ctx->lock avoids it from being uncloned.
8100
		 */
P
Peter Zijlstra 已提交
8101
		cloned_ctx = parent_ctx->parent_ctx;
8102 8103
		if (cloned_ctx) {
			child_ctx->parent_ctx = cloned_ctx;
8104
			child_ctx->parent_gen = parent_ctx->parent_gen;
8105 8106 8107 8108 8109
		} else {
			child_ctx->parent_ctx = parent_ctx;
			child_ctx->parent_gen = parent_ctx->generation;
		}
		get_ctx(child_ctx->parent_ctx);
8110 8111
	}

P
Peter Zijlstra 已提交
8112
	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8113
	mutex_unlock(&parent_ctx->mutex);
8114

8115
	perf_unpin_context(parent_ctx);
8116
	put_ctx(parent_ctx);
8117

8118
	return ret;
8119 8120
}

P
Peter Zijlstra 已提交
8121 8122 8123 8124 8125 8126 8127
/*
 * Initialize the perf_event context in task_struct
 */
int perf_event_init_task(struct task_struct *child)
{
	int ctxn, ret;

8128 8129 8130 8131
	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
8132 8133
	for_each_task_context_nr(ctxn) {
		ret = perf_event_init_context(child, ctxn);
P
Peter Zijlstra 已提交
8134 8135
		if (ret) {
			perf_event_free_task(child);
P
Peter Zijlstra 已提交
8136
			return ret;
P
Peter Zijlstra 已提交
8137
		}
P
Peter Zijlstra 已提交
8138 8139 8140 8141 8142
	}

	return 0;
}

8143 8144
static void __init perf_event_init_all_cpus(void)
{
8145
	struct swevent_htable *swhash;
8146 8147 8148
	int cpu;

	for_each_possible_cpu(cpu) {
8149 8150
		swhash = &per_cpu(swevent_htable, cpu);
		mutex_init(&swhash->hlist_mutex);
8151
		INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
8152 8153 8154
	}
}

8155
static void perf_event_init_cpu(int cpu)
T
Thomas Gleixner 已提交
8156
{
P
Peter Zijlstra 已提交
8157
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
T
Thomas Gleixner 已提交
8158

8159
	mutex_lock(&swhash->hlist_mutex);
8160
	swhash->online = true;
8161
	if (swhash->hlist_refcount > 0) {
8162 8163
		struct swevent_hlist *hlist;

8164 8165 8166
		hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
		WARN_ON(!hlist);
		rcu_assign_pointer(swhash->swevent_hlist, hlist);
8167
	}
8168
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
8169 8170
}

P
Peter Zijlstra 已提交
8171
#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8172
static void perf_pmu_rotate_stop(struct pmu *pmu)
T
Thomas Gleixner 已提交
8173
{
8174 8175 8176 8177 8178 8179 8180
	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
8181
static void __perf_event_exit_context(void *__info)
T
Thomas Gleixner 已提交
8182
{
8183
	struct remove_event re = { .detach_group = true };
P
Peter Zijlstra 已提交
8184
	struct perf_event_context *ctx = __info;
T
Thomas Gleixner 已提交
8185

P
Peter Zijlstra 已提交
8186
	perf_pmu_rotate_stop(ctx->pmu);
8187

P
Peter Zijlstra 已提交
8188
	rcu_read_lock();
8189 8190
	list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
		__perf_remove_from_context(&re);
P
Peter Zijlstra 已提交
8191
	rcu_read_unlock();
T
Thomas Gleixner 已提交
8192
}
P
Peter Zijlstra 已提交
8193 8194 8195 8196 8197 8198 8199 8200 8201

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) {
8202
		ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
P
Peter Zijlstra 已提交
8203 8204 8205 8206 8207 8208 8209 8210

		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);
}

8211
static void perf_event_exit_cpu(int cpu)
T
Thomas Gleixner 已提交
8212
{
8213
	struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8214

P
Peter Zijlstra 已提交
8215 8216
	perf_event_exit_cpu_context(cpu);

8217
	mutex_lock(&swhash->hlist_mutex);
8218
	swhash->online = false;
8219 8220
	swevent_hlist_release(swhash);
	mutex_unlock(&swhash->hlist_mutex);
T
Thomas Gleixner 已提交
8221 8222
}
#else
8223
static inline void perf_event_exit_cpu(int cpu) { }
T
Thomas Gleixner 已提交
8224 8225
#endif

P
Peter Zijlstra 已提交
8226 8227 8228 8229 8230 8231 8232 8233 8234 8235 8236 8237 8238 8239 8240 8241 8242 8243 8244 8245
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,
};

8246
static int
T
Thomas Gleixner 已提交
8247 8248 8249 8250
perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
{
	unsigned int cpu = (long)hcpu;

8251
	switch (action & ~CPU_TASKS_FROZEN) {
T
Thomas Gleixner 已提交
8252 8253

	case CPU_UP_PREPARE:
P
Peter Zijlstra 已提交
8254
	case CPU_DOWN_FAILED:
8255
		perf_event_init_cpu(cpu);
T
Thomas Gleixner 已提交
8256 8257
		break;

P
Peter Zijlstra 已提交
8258
	case CPU_UP_CANCELED:
T
Thomas Gleixner 已提交
8259
	case CPU_DOWN_PREPARE:
8260
		perf_event_exit_cpu(cpu);
T
Thomas Gleixner 已提交
8261 8262 8263 8264 8265 8266 8267 8268
		break;
	default:
		break;
	}

	return NOTIFY_OK;
}

8269
void __init perf_event_init(void)
T
Thomas Gleixner 已提交
8270
{
8271 8272
	int ret;

P
Peter Zijlstra 已提交
8273 8274
	idr_init(&pmu_idr);

8275
	perf_event_init_all_cpus();
8276
	init_srcu_struct(&pmus_srcu);
P
Peter Zijlstra 已提交
8277 8278 8279
	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);
8280 8281
	perf_tp_register();
	perf_cpu_notifier(perf_cpu_notify);
P
Peter Zijlstra 已提交
8282
	register_reboot_notifier(&perf_reboot_notifier);
8283 8284 8285

	ret = init_hw_breakpoint();
	WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8286 8287 8288

	/* do not patch jump label more than once per second */
	jump_label_rate_limit(&perf_sched_events, HZ);
8289 8290 8291 8292 8293 8294 8295

	/*
	 * 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
8296
}
P
Peter Zijlstra 已提交
8297 8298 8299 8300 8301 8302 8303 8304 8305 8306 8307 8308 8309 8310 8311 8312 8313 8314 8315 8316 8317 8318 8319 8320 8321 8322 8323 8324

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
8325 8326

#ifdef CONFIG_CGROUP_PERF
8327 8328
static struct cgroup_subsys_state *
perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
S
Stephane Eranian 已提交
8329 8330 8331
{
	struct perf_cgroup *jc;

8332
	jc = kzalloc(sizeof(*jc), GFP_KERNEL);
S
Stephane Eranian 已提交
8333 8334 8335 8336 8337 8338 8339 8340 8341 8342 8343 8344
	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;
}

8345
static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
S
Stephane Eranian 已提交
8346
{
8347 8348
	struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);

S
Stephane Eranian 已提交
8349 8350 8351 8352 8353 8354 8355 8356 8357 8358 8359
	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;
}

8360 8361
static void perf_cgroup_attach(struct cgroup_subsys_state *css,
			       struct cgroup_taskset *tset)
S
Stephane Eranian 已提交
8362
{
8363 8364
	struct task_struct *task;

8365
	cgroup_taskset_for_each(task, tset)
8366
		task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8367 8368
}

8369 8370
static void perf_cgroup_exit(struct cgroup_subsys_state *css,
			     struct cgroup_subsys_state *old_css,
8371
			     struct task_struct *task)
S
Stephane Eranian 已提交
8372 8373 8374 8375 8376 8377 8378 8379 8380
{
	/*
	 * 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;

8381
	task_function_call(task, __perf_cgroup_move, task);
S
Stephane Eranian 已提交
8382 8383
}

8384
struct cgroup_subsys perf_event_cgrp_subsys = {
8385 8386
	.css_alloc	= perf_cgroup_css_alloc,
	.css_free	= perf_cgroup_css_free,
8387
	.exit		= perf_cgroup_exit,
8388
	.attach		= perf_cgroup_attach,
S
Stephane Eranian 已提交
8389 8390
};
#endif /* CONFIG_CGROUP_PERF */