linux/kernel/events/core.c
<<
>>
Prefs
   1/*
   2 * Performance events core code:
   3 *
   4 *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   5 *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
   6 *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
   7 *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   8 *
   9 * For licensing details see kernel-base/COPYING
  10 */
  11
  12#include <linux/fs.h>
  13#include <linux/mm.h>
  14#include <linux/cpu.h>
  15#include <linux/smp.h>
  16#include <linux/idr.h>
  17#include <linux/file.h>
  18#include <linux/poll.h>
  19#include <linux/slab.h>
  20#include <linux/hash.h>
  21#include <linux/sysfs.h>
  22#include <linux/dcache.h>
  23#include <linux/percpu.h>
  24#include <linux/ptrace.h>
  25#include <linux/reboot.h>
  26#include <linux/vmstat.h>
  27#include <linux/device.h>
  28#include <linux/export.h>
  29#include <linux/vmalloc.h>
  30#include <linux/hardirq.h>
  31#include <linux/rculist.h>
  32#include <linux/uaccess.h>
  33#include <linux/syscalls.h>
  34#include <linux/anon_inodes.h>
  35#include <linux/kernel_stat.h>
  36#include <linux/perf_event.h>
  37#include <linux/ftrace_event.h>
  38#include <linux/hw_breakpoint.h>
  39#include <linux/mm_types.h>
  40
  41#include "internal.h"
  42
  43#include <asm/irq_regs.h>
  44
  45struct remote_function_call {
  46        struct task_struct      *p;
  47        int                     (*func)(void *info);
  48        void                    *info;
  49        int                     ret;
  50};
  51
  52static void remote_function(void *data)
  53{
  54        struct remote_function_call *tfc = data;
  55        struct task_struct *p = tfc->p;
  56
  57        if (p) {
  58                tfc->ret = -EAGAIN;
  59                if (task_cpu(p) != smp_processor_id() || !task_curr(p))
  60                        return;
  61        }
  62
  63        tfc->ret = tfc->func(tfc->info);
  64}
  65
  66/**
  67 * task_function_call - call a function on the cpu on which a task runs
  68 * @p:          the task to evaluate
  69 * @func:       the function to be called
  70 * @info:       the function call argument
  71 *
  72 * Calls the function @func when the task is currently running. This might
  73 * be on the current CPU, which just calls the function directly
  74 *
  75 * returns: @func return value, or
  76 *          -ESRCH  - when the process isn't running
  77 *          -EAGAIN - when the process moved away
  78 */
  79static int
  80task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
  81{
  82        struct remote_function_call data = {
  83                .p      = p,
  84                .func   = func,
  85                .info   = info,
  86                .ret    = -ESRCH, /* No such (running) process */
  87        };
  88
  89        if (task_curr(p))
  90                smp_call_function_single(task_cpu(p), remote_function, &data, 1);
  91
  92        return data.ret;
  93}
  94
  95/**
  96 * cpu_function_call - call a function on the cpu
  97 * @func:       the function to be called
  98 * @info:       the function call argument
  99 *
 100 * Calls the function @func on the remote cpu.
 101 *
 102 * returns: @func return value or -ENXIO when the cpu is offline
 103 */
 104static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
 105{
 106        struct remote_function_call data = {
 107                .p      = NULL,
 108                .func   = func,
 109                .info   = info,
 110                .ret    = -ENXIO, /* No such CPU */
 111        };
 112
 113        smp_call_function_single(cpu, remote_function, &data, 1);
 114
 115        return data.ret;
 116}
 117
 118#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
 119                       PERF_FLAG_FD_OUTPUT  |\
 120                       PERF_FLAG_PID_CGROUP)
 121
 122/*
 123 * branch priv levels that need permission checks
 124 */
 125#define PERF_SAMPLE_BRANCH_PERM_PLM \
 126        (PERF_SAMPLE_BRANCH_KERNEL |\
 127         PERF_SAMPLE_BRANCH_HV)
 128
 129enum event_type_t {
 130        EVENT_FLEXIBLE = 0x1,
 131        EVENT_PINNED = 0x2,
 132        EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
 133};
 134
 135/*
 136 * perf_sched_events : >0 events exist
 137 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
 138 */
 139struct static_key_deferred perf_sched_events __read_mostly;
 140static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
 141static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
 142
 143static atomic_t nr_mmap_events __read_mostly;
 144static atomic_t nr_comm_events __read_mostly;
 145static atomic_t nr_task_events __read_mostly;
 146
 147static LIST_HEAD(pmus);
 148static DEFINE_MUTEX(pmus_lock);
 149static struct srcu_struct pmus_srcu;
 150
 151/*
 152 * perf event paranoia level:
 153 *  -1 - not paranoid at all
 154 *   0 - disallow raw tracepoint access for unpriv
 155 *   1 - disallow cpu events for unpriv
 156 *   2 - disallow kernel profiling for unpriv
 157 */
 158int sysctl_perf_event_paranoid __read_mostly = 1;
 159
 160/* Minimum for 512 kiB + 1 user control page */
 161int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
 162
 163/*
 164 * max perf event sample rate
 165 */
 166#define DEFAULT_MAX_SAMPLE_RATE 100000
 167int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
 168static int max_samples_per_tick __read_mostly =
 169        DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
 170
 171int perf_proc_update_handler(struct ctl_table *table, int write,
 172                void __user *buffer, size_t *lenp,
 173                loff_t *ppos)
 174{
 175        int ret = proc_dointvec(table, write, buffer, lenp, ppos);
 176
 177        if (ret || !write)
 178                return ret;
 179
 180        max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
 181
 182        return 0;
 183}
 184
 185static atomic64_t perf_event_id;
 186
 187static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
 188                              enum event_type_t event_type);
 189
 190static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
 191                             enum event_type_t event_type,
 192                             struct task_struct *task);
 193
 194static void update_context_time(struct perf_event_context *ctx);
 195static u64 perf_event_time(struct perf_event *event);
 196
 197static void ring_buffer_attach(struct perf_event *event,
 198                               struct ring_buffer *rb);
 199
 200void __weak perf_event_print_debug(void)        { }
 201
 202extern __weak const char *perf_pmu_name(void)
 203{
 204        return "pmu";
 205}
 206
 207static inline u64 perf_clock(void)
 208{
 209        return local_clock();
 210}
 211
 212static inline struct perf_cpu_context *
 213__get_cpu_context(struct perf_event_context *ctx)
 214{
 215        return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
 216}
 217
 218static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
 219                          struct perf_event_context *ctx)
 220{
 221        raw_spin_lock(&cpuctx->ctx.lock);
 222        if (ctx)
 223                raw_spin_lock(&ctx->lock);
 224}
 225
 226static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
 227                            struct perf_event_context *ctx)
 228{
 229        if (ctx)
 230                raw_spin_unlock(&ctx->lock);
 231        raw_spin_unlock(&cpuctx->ctx.lock);
 232}
 233
 234#ifdef CONFIG_CGROUP_PERF
 235
 236/*
 237 * Must ensure cgroup is pinned (css_get) before calling
 238 * this function. In other words, we cannot call this function
 239 * if there is no cgroup event for the current CPU context.
 240 */
 241static inline struct perf_cgroup *
 242perf_cgroup_from_task(struct task_struct *task)
 243{
 244        return container_of(task_subsys_state(task, perf_subsys_id),
 245                        struct perf_cgroup, css);
 246}
 247
 248static inline bool
 249perf_cgroup_match(struct perf_event *event)
 250{
 251        struct perf_event_context *ctx = event->ctx;
 252        struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 253
 254        return !event->cgrp || event->cgrp == cpuctx->cgrp;
 255}
 256
 257static inline bool perf_tryget_cgroup(struct perf_event *event)
 258{
 259        return css_tryget(&event->cgrp->css);
 260}
 261
 262static inline void perf_put_cgroup(struct perf_event *event)
 263{
 264        css_put(&event->cgrp->css);
 265}
 266
 267static inline void perf_detach_cgroup(struct perf_event *event)
 268{
 269        perf_put_cgroup(event);
 270        event->cgrp = NULL;
 271}
 272
 273static inline int is_cgroup_event(struct perf_event *event)
 274{
 275        return event->cgrp != NULL;
 276}
 277
 278static inline u64 perf_cgroup_event_time(struct perf_event *event)
 279{
 280        struct perf_cgroup_info *t;
 281
 282        t = per_cpu_ptr(event->cgrp->info, event->cpu);
 283        return t->time;
 284}
 285
 286static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
 287{
 288        struct perf_cgroup_info *info;
 289        u64 now;
 290
 291        now = perf_clock();
 292
 293        info = this_cpu_ptr(cgrp->info);
 294
 295        info->time += now - info->timestamp;
 296        info->timestamp = now;
 297}
 298
 299static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
 300{
 301        struct perf_cgroup *cgrp_out = cpuctx->cgrp;
 302        if (cgrp_out)
 303                __update_cgrp_time(cgrp_out);
 304}
 305
 306static inline void update_cgrp_time_from_event(struct perf_event *event)
 307{
 308        struct perf_cgroup *cgrp;
 309
 310        /*
 311         * ensure we access cgroup data only when needed and
 312         * when we know the cgroup is pinned (css_get)
 313         */
 314        if (!is_cgroup_event(event))
 315                return;
 316
 317        cgrp = perf_cgroup_from_task(current);
 318        /*
 319         * Do not update time when cgroup is not active
 320         */
 321        if (cgrp == event->cgrp)
 322                __update_cgrp_time(event->cgrp);
 323}
 324
 325static inline void
 326perf_cgroup_set_timestamp(struct task_struct *task,
 327                          struct perf_event_context *ctx)
 328{
 329        struct perf_cgroup *cgrp;
 330        struct perf_cgroup_info *info;
 331
 332        /*
 333         * ctx->lock held by caller
 334         * ensure we do not access cgroup data
 335         * unless we have the cgroup pinned (css_get)
 336         */
 337        if (!task || !ctx->nr_cgroups)
 338                return;
 339
 340        cgrp = perf_cgroup_from_task(task);
 341        info = this_cpu_ptr(cgrp->info);
 342        info->timestamp = ctx->timestamp;
 343}
 344
 345#define PERF_CGROUP_SWOUT       0x1 /* cgroup switch out every event */
 346#define PERF_CGROUP_SWIN        0x2 /* cgroup switch in events based on task */
 347
 348/*
 349 * reschedule events based on the cgroup constraint of task.
 350 *
 351 * mode SWOUT : schedule out everything
 352 * mode SWIN : schedule in based on cgroup for next
 353 */
 354void perf_cgroup_switch(struct task_struct *task, int mode)
 355{
 356        struct perf_cpu_context *cpuctx;
 357        struct pmu *pmu;
 358        unsigned long flags;
 359
 360        /*
 361         * disable interrupts to avoid geting nr_cgroup
 362         * changes via __perf_event_disable(). Also
 363         * avoids preemption.
 364         */
 365        local_irq_save(flags);
 366
 367        /*
 368         * we reschedule only in the presence of cgroup
 369         * constrained events.
 370         */
 371        rcu_read_lock();
 372
 373        list_for_each_entry_rcu(pmu, &pmus, entry) {
 374                cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
 375                if (cpuctx->unique_pmu != pmu)
 376                        continue; /* ensure we process each cpuctx once */
 377
 378                /*
 379                 * perf_cgroup_events says at least one
 380                 * context on this CPU has cgroup events.
 381                 *
 382                 * ctx->nr_cgroups reports the number of cgroup
 383                 * events for a context.
 384                 */
 385                if (cpuctx->ctx.nr_cgroups > 0) {
 386                        perf_ctx_lock(cpuctx, cpuctx->task_ctx);
 387                        perf_pmu_disable(cpuctx->ctx.pmu);
 388
 389                        if (mode & PERF_CGROUP_SWOUT) {
 390                                cpu_ctx_sched_out(cpuctx, EVENT_ALL);
 391                                /*
 392                                 * must not be done before ctxswout due
 393                                 * to event_filter_match() in event_sched_out()
 394                                 */
 395                                cpuctx->cgrp = NULL;
 396                        }
 397
 398                        if (mode & PERF_CGROUP_SWIN) {
 399                                WARN_ON_ONCE(cpuctx->cgrp);
 400                                /*
 401                                 * set cgrp before ctxsw in to allow
 402                                 * event_filter_match() to not have to pass
 403                                 * task around
 404                                 */
 405                                cpuctx->cgrp = perf_cgroup_from_task(task);
 406                                cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
 407                        }
 408                        perf_pmu_enable(cpuctx->ctx.pmu);
 409                        perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
 410                }
 411        }
 412
 413        rcu_read_unlock();
 414
 415        local_irq_restore(flags);
 416}
 417
 418static inline void perf_cgroup_sched_out(struct task_struct *task,
 419                                         struct task_struct *next)
 420{
 421        struct perf_cgroup *cgrp1;
 422        struct perf_cgroup *cgrp2 = NULL;
 423
 424        /*
 425         * we come here when we know perf_cgroup_events > 0
 426         */
 427        cgrp1 = perf_cgroup_from_task(task);
 428
 429        /*
 430         * next is NULL when called from perf_event_enable_on_exec()
 431         * that will systematically cause a cgroup_switch()
 432         */
 433        if (next)
 434                cgrp2 = perf_cgroup_from_task(next);
 435
 436        /*
 437         * only schedule out current cgroup events if we know
 438         * that we are switching to a different cgroup. Otherwise,
 439         * do no touch the cgroup events.
 440         */
 441        if (cgrp1 != cgrp2)
 442                perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
 443}
 444
 445static inline void perf_cgroup_sched_in(struct task_struct *prev,
 446                                        struct task_struct *task)
 447{
 448        struct perf_cgroup *cgrp1;
 449        struct perf_cgroup *cgrp2 = NULL;
 450
 451        /*
 452         * we come here when we know perf_cgroup_events > 0
 453         */
 454        cgrp1 = perf_cgroup_from_task(task);
 455
 456        /* prev can never be NULL */
 457        cgrp2 = perf_cgroup_from_task(prev);
 458
 459        /*
 460         * only need to schedule in cgroup events if we are changing
 461         * cgroup during ctxsw. Cgroup events were not scheduled
 462         * out of ctxsw out if that was not the case.
 463         */
 464        if (cgrp1 != cgrp2)
 465                perf_cgroup_switch(task, PERF_CGROUP_SWIN);
 466}
 467
 468static inline int perf_cgroup_connect(int fd, struct perf_event *event,
 469                                      struct perf_event_attr *attr,
 470                                      struct perf_event *group_leader)
 471{
 472        struct perf_cgroup *cgrp;
 473        struct cgroup_subsys_state *css;
 474        struct fd f = fdget(fd);
 475        int ret = 0;
 476
 477        if (!f.file)
 478                return -EBADF;
 479
 480        css = cgroup_css_from_dir(f.file, perf_subsys_id);
 481        if (IS_ERR(css)) {
 482                ret = PTR_ERR(css);
 483                goto out;
 484        }
 485
 486        cgrp = container_of(css, struct perf_cgroup, css);
 487        event->cgrp = cgrp;
 488
 489        /* must be done before we fput() the file */
 490        if (!perf_tryget_cgroup(event)) {
 491                event->cgrp = NULL;
 492                ret = -ENOENT;
 493                goto out;
 494        }
 495
 496        /*
 497         * all events in a group must monitor
 498         * the same cgroup because a task belongs
 499         * to only one perf cgroup at a time
 500         */
 501        if (group_leader && group_leader->cgrp != cgrp) {
 502                perf_detach_cgroup(event);
 503                ret = -EINVAL;
 504        }
 505out:
 506        fdput(f);
 507        return ret;
 508}
 509
 510static inline void
 511perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
 512{
 513        struct perf_cgroup_info *t;
 514        t = per_cpu_ptr(event->cgrp->info, event->cpu);
 515        event->shadow_ctx_time = now - t->timestamp;
 516}
 517
 518static inline void
 519perf_cgroup_defer_enabled(struct perf_event *event)
 520{
 521        /*
 522         * when the current task's perf cgroup does not match
 523         * the event's, we need to remember to call the
 524         * perf_mark_enable() function the first time a task with
 525         * a matching perf cgroup is scheduled in.
 526         */
 527        if (is_cgroup_event(event) && !perf_cgroup_match(event))
 528                event->cgrp_defer_enabled = 1;
 529}
 530
 531static inline void
 532perf_cgroup_mark_enabled(struct perf_event *event,
 533                         struct perf_event_context *ctx)
 534{
 535        struct perf_event *sub;
 536        u64 tstamp = perf_event_time(event);
 537
 538        if (!event->cgrp_defer_enabled)
 539                return;
 540
 541        event->cgrp_defer_enabled = 0;
 542
 543        event->tstamp_enabled = tstamp - event->total_time_enabled;
 544        list_for_each_entry(sub, &event->sibling_list, group_entry) {
 545                if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
 546                        sub->tstamp_enabled = tstamp - sub->total_time_enabled;
 547                        sub->cgrp_defer_enabled = 0;
 548                }
 549        }
 550}
 551#else /* !CONFIG_CGROUP_PERF */
 552
 553static inline bool
 554perf_cgroup_match(struct perf_event *event)
 555{
 556        return true;
 557}
 558
 559static inline void perf_detach_cgroup(struct perf_event *event)
 560{}
 561
 562static inline int is_cgroup_event(struct perf_event *event)
 563{
 564        return 0;
 565}
 566
 567static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
 568{
 569        return 0;
 570}
 571
 572static inline void update_cgrp_time_from_event(struct perf_event *event)
 573{
 574}
 575
 576static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
 577{
 578}
 579
 580static inline void perf_cgroup_sched_out(struct task_struct *task,
 581                                         struct task_struct *next)
 582{
 583}
 584
 585static inline void perf_cgroup_sched_in(struct task_struct *prev,
 586                                        struct task_struct *task)
 587{
 588}
 589
 590static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
 591                                      struct perf_event_attr *attr,
 592                                      struct perf_event *group_leader)
 593{
 594        return -EINVAL;
 595}
 596
 597static inline void
 598perf_cgroup_set_timestamp(struct task_struct *task,
 599                          struct perf_event_context *ctx)
 600{
 601}
 602
 603void
 604perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
 605{
 606}
 607
 608static inline void
 609perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
 610{
 611}
 612
 613static inline u64 perf_cgroup_event_time(struct perf_event *event)
 614{
 615        return 0;
 616}
 617
 618static inline void
 619perf_cgroup_defer_enabled(struct perf_event *event)
 620{
 621}
 622
 623static inline void
 624perf_cgroup_mark_enabled(struct perf_event *event,
 625                         struct perf_event_context *ctx)
 626{
 627}
 628#endif
 629
 630void perf_pmu_disable(struct pmu *pmu)
 631{
 632        int *count = this_cpu_ptr(pmu->pmu_disable_count);
 633        if (!(*count)++)
 634                pmu->pmu_disable(pmu);
 635}
 636
 637void perf_pmu_enable(struct pmu *pmu)
 638{
 639        int *count = this_cpu_ptr(pmu->pmu_disable_count);
 640        if (!--(*count))
 641                pmu->pmu_enable(pmu);
 642}
 643
 644static DEFINE_PER_CPU(struct list_head, rotation_list);
 645
 646/*
 647 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
 648 * because they're strictly cpu affine and rotate_start is called with IRQs
 649 * disabled, while rotate_context is called from IRQ context.
 650 */
 651static void perf_pmu_rotate_start(struct pmu *pmu)
 652{
 653        struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
 654        struct list_head *head = &__get_cpu_var(rotation_list);
 655
 656        WARN_ON(!irqs_disabled());
 657
 658        if (list_empty(&cpuctx->rotation_list))
 659                list_add(&cpuctx->rotation_list, head);
 660}
 661
 662static void get_ctx(struct perf_event_context *ctx)
 663{
 664        WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
 665}
 666
 667static void put_ctx(struct perf_event_context *ctx)
 668{
 669        if (atomic_dec_and_test(&ctx->refcount)) {
 670                if (ctx->parent_ctx)
 671                        put_ctx(ctx->parent_ctx);
 672                if (ctx->task)
 673                        put_task_struct(ctx->task);
 674                kfree_rcu(ctx, rcu_head);
 675        }
 676}
 677
 678static void unclone_ctx(struct perf_event_context *ctx)
 679{
 680        if (ctx->parent_ctx) {
 681                put_ctx(ctx->parent_ctx);
 682                ctx->parent_ctx = NULL;
 683        }
 684}
 685
 686static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
 687{
 688        /*
 689         * only top level events have the pid namespace they were created in
 690         */
 691        if (event->parent)
 692                event = event->parent;
 693
 694        return task_tgid_nr_ns(p, event->ns);
 695}
 696
 697static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
 698{
 699        /*
 700         * only top level events have the pid namespace they were created in
 701         */
 702        if (event->parent)
 703                event = event->parent;
 704
 705        return task_pid_nr_ns(p, event->ns);
 706}
 707
 708/*
 709 * If we inherit events we want to return the parent event id
 710 * to userspace.
 711 */
 712static u64 primary_event_id(struct perf_event *event)
 713{
 714        u64 id = event->id;
 715
 716        if (event->parent)
 717                id = event->parent->id;
 718
 719        return id;
 720}
 721
 722/*
 723 * Get the perf_event_context for a task and lock it.
 724 * This has to cope with with the fact that until it is locked,
 725 * the context could get moved to another task.
 726 */
 727static struct perf_event_context *
 728perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
 729{
 730        struct perf_event_context *ctx;
 731
 732        rcu_read_lock();
 733retry:
 734        ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
 735        if (ctx) {
 736                /*
 737                 * If this context is a clone of another, it might
 738                 * get swapped for another underneath us by
 739                 * perf_event_task_sched_out, though the
 740                 * rcu_read_lock() protects us from any context
 741                 * getting freed.  Lock the context and check if it
 742                 * got swapped before we could get the lock, and retry
 743                 * if so.  If we locked the right context, then it
 744                 * can't get swapped on us any more.
 745                 */
 746                raw_spin_lock_irqsave(&ctx->lock, *flags);
 747                if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
 748                        raw_spin_unlock_irqrestore(&ctx->lock, *flags);
 749                        goto retry;
 750                }
 751
 752                if (!atomic_inc_not_zero(&ctx->refcount)) {
 753                        raw_spin_unlock_irqrestore(&ctx->lock, *flags);
 754                        ctx = NULL;
 755                }
 756        }
 757        rcu_read_unlock();
 758        return ctx;
 759}
 760
 761/*
 762 * Get the context for a task and increment its pin_count so it
 763 * can't get swapped to another task.  This also increments its
 764 * reference count so that the context can't get freed.
 765 */
 766static struct perf_event_context *
 767perf_pin_task_context(struct task_struct *task, int ctxn)
 768{
 769        struct perf_event_context *ctx;
 770        unsigned long flags;
 771
 772        ctx = perf_lock_task_context(task, ctxn, &flags);
 773        if (ctx) {
 774                ++ctx->pin_count;
 775                raw_spin_unlock_irqrestore(&ctx->lock, flags);
 776        }
 777        return ctx;
 778}
 779
 780static void perf_unpin_context(struct perf_event_context *ctx)
 781{
 782        unsigned long flags;
 783
 784        raw_spin_lock_irqsave(&ctx->lock, flags);
 785        --ctx->pin_count;
 786        raw_spin_unlock_irqrestore(&ctx->lock, flags);
 787}
 788
 789/*
 790 * Update the record of the current time in a context.
 791 */
 792static void update_context_time(struct perf_event_context *ctx)
 793{
 794        u64 now = perf_clock();
 795
 796        ctx->time += now - ctx->timestamp;
 797        ctx->timestamp = now;
 798}
 799
 800static u64 perf_event_time(struct perf_event *event)
 801{
 802        struct perf_event_context *ctx = event->ctx;
 803
 804        if (is_cgroup_event(event))
 805                return perf_cgroup_event_time(event);
 806
 807        return ctx ? ctx->time : 0;
 808}
 809
 810/*
 811 * Update the total_time_enabled and total_time_running fields for a event.
 812 * The caller of this function needs to hold the ctx->lock.
 813 */
 814static void update_event_times(struct perf_event *event)
 815{
 816        struct perf_event_context *ctx = event->ctx;
 817        u64 run_end;
 818
 819        if (event->state < PERF_EVENT_STATE_INACTIVE ||
 820            event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
 821                return;
 822        /*
 823         * in cgroup mode, time_enabled represents
 824         * the time the event was enabled AND active
 825         * tasks were in the monitored cgroup. This is
 826         * independent of the activity of the context as
 827         * there may be a mix of cgroup and non-cgroup events.
 828         *
 829         * That is why we treat cgroup events differently
 830         * here.
 831         */
 832        if (is_cgroup_event(event))
 833                run_end = perf_cgroup_event_time(event);
 834        else if (ctx->is_active)
 835                run_end = ctx->time;
 836        else
 837                run_end = event->tstamp_stopped;
 838
 839        event->total_time_enabled = run_end - event->tstamp_enabled;
 840
 841        if (event->state == PERF_EVENT_STATE_INACTIVE)
 842                run_end = event->tstamp_stopped;
 843        else
 844                run_end = perf_event_time(event);
 845
 846        event->total_time_running = run_end - event->tstamp_running;
 847
 848}
 849
 850/*
 851 * Update total_time_enabled and total_time_running for all events in a group.
 852 */
 853static void update_group_times(struct perf_event *leader)
 854{
 855        struct perf_event *event;
 856
 857        update_event_times(leader);
 858        list_for_each_entry(event, &leader->sibling_list, group_entry)
 859                update_event_times(event);
 860}
 861
 862static struct list_head *
 863ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
 864{
 865        if (event->attr.pinned)
 866                return &ctx->pinned_groups;
 867        else
 868                return &ctx->flexible_groups;
 869}
 870
 871/*
 872 * Add a event from the lists for its context.
 873 * Must be called with ctx->mutex and ctx->lock held.
 874 */
 875static void
 876list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 877{
 878        WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
 879        event->attach_state |= PERF_ATTACH_CONTEXT;
 880
 881        /*
 882         * If we're a stand alone event or group leader, we go to the context
 883         * list, group events are kept attached to the group so that
 884         * perf_group_detach can, at all times, locate all siblings.
 885         */
 886        if (event->group_leader == event) {
 887                struct list_head *list;
 888
 889                if (is_software_event(event))
 890                        event->group_flags |= PERF_GROUP_SOFTWARE;
 891
 892                list = ctx_group_list(event, ctx);
 893                list_add_tail(&event->group_entry, list);
 894        }
 895
 896        if (is_cgroup_event(event))
 897                ctx->nr_cgroups++;
 898
 899        if (has_branch_stack(event))
 900                ctx->nr_branch_stack++;
 901
 902        list_add_rcu(&event->event_entry, &ctx->event_list);
 903        if (!ctx->nr_events)
 904                perf_pmu_rotate_start(ctx->pmu);
 905        ctx->nr_events++;
 906        if (event->attr.inherit_stat)
 907                ctx->nr_stat++;
 908}
 909
 910/*
 911 * Called at perf_event creation and when events are attached/detached from a
 912 * group.
 913 */
 914static void perf_event__read_size(struct perf_event *event)
 915{
 916        int entry = sizeof(u64); /* value */
 917        int size = 0;
 918        int nr = 1;
 919
 920        if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 921                size += sizeof(u64);
 922
 923        if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 924                size += sizeof(u64);
 925
 926        if (event->attr.read_format & PERF_FORMAT_ID)
 927                entry += sizeof(u64);
 928
 929        if (event->attr.read_format & PERF_FORMAT_GROUP) {
 930                nr += event->group_leader->nr_siblings;
 931                size += sizeof(u64);
 932        }
 933
 934        size += entry * nr;
 935        event->read_size = size;
 936}
 937
 938static void perf_event__header_size(struct perf_event *event)
 939{
 940        struct perf_sample_data *data;
 941        u64 sample_type = event->attr.sample_type;
 942        u16 size = 0;
 943
 944        perf_event__read_size(event);
 945
 946        if (sample_type & PERF_SAMPLE_IP)
 947                size += sizeof(data->ip);
 948
 949        if (sample_type & PERF_SAMPLE_ADDR)
 950                size += sizeof(data->addr);
 951
 952        if (sample_type & PERF_SAMPLE_PERIOD)
 953                size += sizeof(data->period);
 954
 955        if (sample_type & PERF_SAMPLE_READ)
 956                size += event->read_size;
 957
 958        event->header_size = size;
 959}
 960
 961static void perf_event__id_header_size(struct perf_event *event)
 962{
 963        struct perf_sample_data *data;
 964        u64 sample_type = event->attr.sample_type;
 965        u16 size = 0;
 966
 967        if (sample_type & PERF_SAMPLE_TID)
 968                size += sizeof(data->tid_entry);
 969
 970        if (sample_type & PERF_SAMPLE_TIME)
 971                size += sizeof(data->time);
 972
 973        if (sample_type & PERF_SAMPLE_ID)
 974                size += sizeof(data->id);
 975
 976        if (sample_type & PERF_SAMPLE_STREAM_ID)
 977                size += sizeof(data->stream_id);
 978
 979        if (sample_type & PERF_SAMPLE_CPU)
 980                size += sizeof(data->cpu_entry);
 981
 982        event->id_header_size = size;
 983}
 984
 985static void perf_group_attach(struct perf_event *event)
 986{
 987        struct perf_event *group_leader = event->group_leader, *pos;
 988
 989        /*
 990         * We can have double attach due to group movement in perf_event_open.
 991         */
 992        if (event->attach_state & PERF_ATTACH_GROUP)
 993                return;
 994
 995        event->attach_state |= PERF_ATTACH_GROUP;
 996
 997        if (group_leader == event)
 998                return;
 999
1000        if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1001                        !is_software_event(event))
1002                group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1003
1004        list_add_tail(&event->group_entry, &group_leader->sibling_list);
1005        group_leader->nr_siblings++;
1006
1007        perf_event__header_size(group_leader);
1008
1009        list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1010                perf_event__header_size(pos);
1011}
1012
1013/*
1014 * Remove a event from the lists for its context.
1015 * Must be called with ctx->mutex and ctx->lock held.
1016 */
1017static void
1018list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1019{
1020        struct perf_cpu_context *cpuctx;
1021        /*
1022         * We can have double detach due to exit/hot-unplug + close.
1023         */
1024        if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1025                return;
1026
1027        event->attach_state &= ~PERF_ATTACH_CONTEXT;
1028
1029        if (is_cgroup_event(event)) {
1030                ctx->nr_cgroups--;
1031                cpuctx = __get_cpu_context(ctx);
1032                /*
1033                 * if there are no more cgroup events
1034                 * then cler cgrp to avoid stale pointer
1035                 * in update_cgrp_time_from_cpuctx()
1036                 */
1037                if (!ctx->nr_cgroups)
1038                        cpuctx->cgrp = NULL;
1039        }
1040
1041        if (has_branch_stack(event))
1042                ctx->nr_branch_stack--;
1043
1044        ctx->nr_events--;
1045        if (event->attr.inherit_stat)
1046                ctx->nr_stat--;
1047
1048        list_del_rcu(&event->event_entry);
1049
1050        if (event->group_leader == event)
1051                list_del_init(&event->group_entry);
1052
1053        update_group_times(event);
1054
1055        /*
1056         * If event was in error state, then keep it
1057         * that way, otherwise bogus counts will be
1058         * returned on read(). The only way to get out
1059         * of error state is by explicit re-enabling
1060         * of the event
1061         */
1062        if (event->state > PERF_EVENT_STATE_OFF)
1063                event->state = PERF_EVENT_STATE_OFF;
1064}
1065
1066static void perf_group_detach(struct perf_event *event)
1067{
1068        struct perf_event *sibling, *tmp;
1069        struct list_head *list = NULL;
1070
1071        /*
1072         * We can have double detach due to exit/hot-unplug + close.
1073         */
1074        if (!(event->attach_state & PERF_ATTACH_GROUP))
1075                return;
1076
1077        event->attach_state &= ~PERF_ATTACH_GROUP;
1078
1079        /*
1080         * If this is a sibling, remove it from its group.
1081         */
1082        if (event->group_leader != event) {
1083                list_del_init(&event->group_entry);
1084                event->group_leader->nr_siblings--;
1085                goto out;
1086        }
1087
1088        if (!list_empty(&event->group_entry))
1089                list = &event->group_entry;
1090
1091        /*
1092         * If this was a group event with sibling events then
1093         * upgrade the siblings to singleton events by adding them
1094         * to whatever list we are on.
1095         */
1096        list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1097                if (list)
1098                        list_move_tail(&sibling->group_entry, list);
1099                sibling->group_leader = sibling;
1100
1101                /* Inherit group flags from the previous leader */
1102                sibling->group_flags = event->group_flags;
1103        }
1104
1105out:
1106        perf_event__header_size(event->group_leader);
1107
1108        list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1109                perf_event__header_size(tmp);
1110}
1111
1112static inline int
1113event_filter_match(struct perf_event *event)
1114{
1115        return (event->cpu == -1 || event->cpu == smp_processor_id())
1116            && perf_cgroup_match(event);
1117}
1118
1119static void
1120event_sched_out(struct perf_event *event,
1121                  struct perf_cpu_context *cpuctx,
1122                  struct perf_event_context *ctx)
1123{
1124        u64 tstamp = perf_event_time(event);
1125        u64 delta;
1126        /*
1127         * An event which could not be activated because of
1128         * filter mismatch still needs to have its timings
1129         * maintained, otherwise bogus information is return
1130         * via read() for time_enabled, time_running:
1131         */
1132        if (event->state == PERF_EVENT_STATE_INACTIVE
1133            && !event_filter_match(event)) {
1134                delta = tstamp - event->tstamp_stopped;
1135                event->tstamp_running += delta;
1136                event->tstamp_stopped = tstamp;
1137        }
1138
1139        if (event->state != PERF_EVENT_STATE_ACTIVE)
1140                return;
1141
1142        event->state = PERF_EVENT_STATE_INACTIVE;
1143        if (event->pending_disable) {
1144                event->pending_disable = 0;
1145                event->state = PERF_EVENT_STATE_OFF;
1146        }
1147        event->tstamp_stopped = tstamp;
1148        event->pmu->del(event, 0);
1149        event->oncpu = -1;
1150
1151        if (!is_software_event(event))
1152                cpuctx->active_oncpu--;
1153        ctx->nr_active--;
1154        if (event->attr.freq && event->attr.sample_freq)
1155                ctx->nr_freq--;
1156        if (event->attr.exclusive || !cpuctx->active_oncpu)
1157                cpuctx->exclusive = 0;
1158}
1159
1160static void
1161group_sched_out(struct perf_event *group_event,
1162                struct perf_cpu_context *cpuctx,
1163                struct perf_event_context *ctx)
1164{
1165        struct perf_event *event;
1166        int state = group_event->state;
1167
1168        event_sched_out(group_event, cpuctx, ctx);
1169
1170        /*
1171         * Schedule out siblings (if any):
1172         */
1173        list_for_each_entry(event, &group_event->sibling_list, group_entry)
1174                event_sched_out(event, cpuctx, ctx);
1175
1176        if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1177                cpuctx->exclusive = 0;
1178}
1179
1180/*
1181 * Cross CPU call to remove a performance event
1182 *
1183 * We disable the event on the hardware level first. After that we
1184 * remove it from the context list.
1185 */
1186static int __perf_remove_from_context(void *info)
1187{
1188        struct perf_event *event = info;
1189        struct perf_event_context *ctx = event->ctx;
1190        struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1191
1192        raw_spin_lock(&ctx->lock);
1193        event_sched_out(event, cpuctx, ctx);
1194        list_del_event(event, ctx);
1195        if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1196                ctx->is_active = 0;
1197                cpuctx->task_ctx = NULL;
1198        }
1199        raw_spin_unlock(&ctx->lock);
1200
1201        return 0;
1202}
1203
1204
1205/*
1206 * Remove the event from a task's (or a CPU's) list of events.
1207 *
1208 * CPU events are removed with a smp call. For task events we only
1209 * call when the task is on a CPU.
1210 *
1211 * If event->ctx is a cloned context, callers must make sure that
1212 * every task struct that event->ctx->task could possibly point to
1213 * remains valid.  This is OK when called from perf_release since
1214 * that only calls us on the top-level context, which can't be a clone.
1215 * When called from perf_event_exit_task, it's OK because the
1216 * context has been detached from its task.
1217 */
1218static void perf_remove_from_context(struct perf_event *event)
1219{
1220        struct perf_event_context *ctx = event->ctx;
1221        struct task_struct *task = ctx->task;
1222
1223        lockdep_assert_held(&ctx->mutex);
1224
1225        if (!task) {
1226                /*
1227                 * Per cpu events are removed via an smp call and
1228                 * the removal is always successful.
1229                 */
1230                cpu_function_call(event->cpu, __perf_remove_from_context, event);
1231                return;
1232        }
1233
1234retry:
1235        if (!task_function_call(task, __perf_remove_from_context, event))
1236                return;
1237
1238        raw_spin_lock_irq(&ctx->lock);
1239        /*
1240         * If we failed to find a running task, but find the context active now
1241         * that we've acquired the ctx->lock, retry.
1242         */
1243        if (ctx->is_active) {
1244                raw_spin_unlock_irq(&ctx->lock);
1245                goto retry;
1246        }
1247
1248        /*
1249         * Since the task isn't running, its safe to remove the event, us
1250         * holding the ctx->lock ensures the task won't get scheduled in.
1251         */
1252        list_del_event(event, ctx);
1253        raw_spin_unlock_irq(&ctx->lock);
1254}
1255
1256/*
1257 * Cross CPU call to disable a performance event
1258 */
1259int __perf_event_disable(void *info)
1260{
1261        struct perf_event *event = info;
1262        struct perf_event_context *ctx = event->ctx;
1263        struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1264
1265        /*
1266         * If this is a per-task event, need to check whether this
1267         * event's task is the current task on this cpu.
1268         *
1269         * Can trigger due to concurrent perf_event_context_sched_out()
1270         * flipping contexts around.
1271         */
1272        if (ctx->task && cpuctx->task_ctx != ctx)
1273                return -EINVAL;
1274
1275        raw_spin_lock(&ctx->lock);
1276
1277        /*
1278         * If the event is on, turn it off.
1279         * If it is in error state, leave it in error state.
1280         */
1281        if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1282                update_context_time(ctx);
1283                update_cgrp_time_from_event(event);
1284                update_group_times(event);
1285                if (event == event->group_leader)
1286                        group_sched_out(event, cpuctx, ctx);
1287                else
1288                        event_sched_out(event, cpuctx, ctx);
1289                event->state = PERF_EVENT_STATE_OFF;
1290        }
1291
1292        raw_spin_unlock(&ctx->lock);
1293
1294        return 0;
1295}
1296
1297/*
1298 * Disable a event.
1299 *
1300 * If event->ctx is a cloned context, callers must make sure that
1301 * every task struct that event->ctx->task could possibly point to
1302 * remains valid.  This condition is satisifed when called through
1303 * perf_event_for_each_child or perf_event_for_each because they
1304 * hold the top-level event's child_mutex, so any descendant that
1305 * goes to exit will block in sync_child_event.
1306 * When called from perf_pending_event it's OK because event->ctx
1307 * is the current context on this CPU and preemption is disabled,
1308 * hence we can't get into perf_event_task_sched_out for this context.
1309 */
1310void perf_event_disable(struct perf_event *event)
1311{
1312        struct perf_event_context *ctx = event->ctx;
1313        struct task_struct *task = ctx->task;
1314
1315        if (!task) {
1316                /*
1317                 * Disable the event on the cpu that it's on
1318                 */
1319                cpu_function_call(event->cpu, __perf_event_disable, event);
1320                return;
1321        }
1322
1323retry:
1324        if (!task_function_call(task, __perf_event_disable, event))
1325                return;
1326
1327        raw_spin_lock_irq(&ctx->lock);
1328        /*
1329         * If the event is still active, we need to retry the cross-call.
1330         */
1331        if (event->state == PERF_EVENT_STATE_ACTIVE) {
1332                raw_spin_unlock_irq(&ctx->lock);
1333                /*
1334                 * Reload the task pointer, it might have been changed by
1335                 * a concurrent perf_event_context_sched_out().
1336                 */
1337                task = ctx->task;
1338                goto retry;
1339        }
1340
1341        /*
1342         * Since we have the lock this context can't be scheduled
1343         * in, so we can change the state safely.
1344         */
1345        if (event->state == PERF_EVENT_STATE_INACTIVE) {
1346                update_group_times(event);
1347                event->state = PERF_EVENT_STATE_OFF;
1348        }
1349        raw_spin_unlock_irq(&ctx->lock);
1350}
1351EXPORT_SYMBOL_GPL(perf_event_disable);
1352
1353static void perf_set_shadow_time(struct perf_event *event,
1354                                 struct perf_event_context *ctx,
1355                                 u64 tstamp)
1356{
1357        /*
1358         * use the correct time source for the time snapshot
1359         *
1360         * We could get by without this by leveraging the
1361         * fact that to get to this function, the caller
1362         * has most likely already called update_context_time()
1363         * and update_cgrp_time_xx() and thus both timestamp
1364         * are identical (or very close). Given that tstamp is,
1365         * already adjusted for cgroup, we could say that:
1366         *    tstamp - ctx->timestamp
1367         * is equivalent to
1368         *    tstamp - cgrp->timestamp.
1369         *
1370         * Then, in perf_output_read(), the calculation would
1371         * work with no changes because:
1372         * - event is guaranteed scheduled in
1373         * - no scheduled out in between
1374         * - thus the timestamp would be the same
1375         *
1376         * But this is a bit hairy.
1377         *
1378         * So instead, we have an explicit cgroup call to remain
1379         * within the time time source all along. We believe it
1380         * is cleaner and simpler to understand.
1381         */
1382        if (is_cgroup_event(event))
1383                perf_cgroup_set_shadow_time(event, tstamp);
1384        else
1385                event->shadow_ctx_time = tstamp - ctx->timestamp;
1386}
1387
1388#define MAX_INTERRUPTS (~0ULL)
1389
1390static void perf_log_throttle(struct perf_event *event, int enable);
1391
1392static int
1393event_sched_in(struct perf_event *event,
1394                 struct perf_cpu_context *cpuctx,
1395                 struct perf_event_context *ctx)
1396{
1397        u64 tstamp = perf_event_time(event);
1398
1399        if (event->state <= PERF_EVENT_STATE_OFF)
1400                return 0;
1401
1402        event->state = PERF_EVENT_STATE_ACTIVE;
1403        event->oncpu = smp_processor_id();
1404
1405        /*
1406         * Unthrottle events, since we scheduled we might have missed several
1407         * ticks already, also for a heavily scheduling task there is little
1408         * guarantee it'll get a tick in a timely manner.
1409         */
1410        if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1411                perf_log_throttle(event, 1);
1412                event->hw.interrupts = 0;
1413        }
1414
1415        /*
1416         * The new state must be visible before we turn it on in the hardware:
1417         */
1418        smp_wmb();
1419
1420        if (event->pmu->add(event, PERF_EF_START)) {
1421                event->state = PERF_EVENT_STATE_INACTIVE;
1422                event->oncpu = -1;
1423                return -EAGAIN;
1424        }
1425
1426        event->tstamp_running += tstamp - event->tstamp_stopped;
1427
1428        perf_set_shadow_time(event, ctx, tstamp);
1429
1430        if (!is_software_event(event))
1431                cpuctx->active_oncpu++;
1432        ctx->nr_active++;
1433        if (event->attr.freq && event->attr.sample_freq)
1434                ctx->nr_freq++;
1435
1436        if (event->attr.exclusive)
1437                cpuctx->exclusive = 1;
1438
1439        return 0;
1440}
1441
1442static int
1443group_sched_in(struct perf_event *group_event,
1444               struct perf_cpu_context *cpuctx,
1445               struct perf_event_context *ctx)
1446{
1447        struct perf_event *event, *partial_group = NULL;
1448        struct pmu *pmu = group_event->pmu;
1449        u64 now = ctx->time;
1450        bool simulate = false;
1451
1452        if (group_event->state == PERF_EVENT_STATE_OFF)
1453                return 0;
1454
1455        pmu->start_txn(pmu);
1456
1457        if (event_sched_in(group_event, cpuctx, ctx)) {
1458                pmu->cancel_txn(pmu);
1459                return -EAGAIN;
1460        }
1461
1462        /*
1463         * Schedule in siblings as one group (if any):
1464         */
1465        list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1466                if (event_sched_in(event, cpuctx, ctx)) {
1467                        partial_group = event;
1468                        goto group_error;
1469                }
1470        }
1471
1472        if (!pmu->commit_txn(pmu))
1473                return 0;
1474
1475group_error:
1476        /*
1477         * Groups can be scheduled in as one unit only, so undo any
1478         * partial group before returning:
1479         * The events up to the failed event are scheduled out normally,
1480         * tstamp_stopped will be updated.
1481         *
1482         * The failed events and the remaining siblings need to have
1483         * their timings updated as if they had gone thru event_sched_in()
1484         * and event_sched_out(). This is required to get consistent timings
1485         * across the group. This also takes care of the case where the group
1486         * could never be scheduled by ensuring tstamp_stopped is set to mark
1487         * the time the event was actually stopped, such that time delta
1488         * calculation in update_event_times() is correct.
1489         */
1490        list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1491                if (event == partial_group)
1492                        simulate = true;
1493
1494                if (simulate) {
1495                        event->tstamp_running += now - event->tstamp_stopped;
1496                        event->tstamp_stopped = now;
1497                } else {
1498                        event_sched_out(event, cpuctx, ctx);
1499                }
1500        }
1501        event_sched_out(group_event, cpuctx, ctx);
1502
1503        pmu->cancel_txn(pmu);
1504
1505        return -EAGAIN;
1506}
1507
1508/*
1509 * Work out whether we can put this event group on the CPU now.
1510 */
1511static int group_can_go_on(struct perf_event *event,
1512                           struct perf_cpu_context *cpuctx,
1513                           int can_add_hw)
1514{
1515        /*
1516         * Groups consisting entirely of software events can always go on.
1517         */
1518        if (event->group_flags & PERF_GROUP_SOFTWARE)
1519                return 1;
1520        /*
1521         * If an exclusive group is already on, no other hardware
1522         * events can go on.
1523         */
1524        if (cpuctx->exclusive)
1525                return 0;
1526        /*
1527         * If this group is exclusive and there are already
1528         * events on the CPU, it can't go on.
1529         */
1530        if (event->attr.exclusive && cpuctx->active_oncpu)
1531                return 0;
1532        /*
1533         * Otherwise, try to add it if all previous groups were able
1534         * to go on.
1535         */
1536        return can_add_hw;
1537}
1538
1539static void add_event_to_ctx(struct perf_event *event,
1540                               struct perf_event_context *ctx)
1541{
1542        u64 tstamp = perf_event_time(event);
1543
1544        list_add_event(event, ctx);
1545        perf_group_attach(event);
1546        event->tstamp_enabled = tstamp;
1547        event->tstamp_running = tstamp;
1548        event->tstamp_stopped = tstamp;
1549}
1550
1551static void task_ctx_sched_out(struct perf_event_context *ctx);
1552static void
1553ctx_sched_in(struct perf_event_context *ctx,
1554             struct perf_cpu_context *cpuctx,
1555             enum event_type_t event_type,
1556             struct task_struct *task);
1557
1558static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1559                                struct perf_event_context *ctx,
1560                                struct task_struct *task)
1561{
1562        cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1563        if (ctx)
1564                ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1565        cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1566        if (ctx)
1567                ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1568}
1569
1570/*
1571 * Cross CPU call to install and enable a performance event
1572 *
1573 * Must be called with ctx->mutex held
1574 */
1575static int  __perf_install_in_context(void *info)
1576{
1577        struct perf_event *event = info;
1578        struct perf_event_context *ctx = event->ctx;
1579        struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1580        struct perf_event_context *task_ctx = cpuctx->task_ctx;
1581        struct task_struct *task = current;
1582
1583        perf_ctx_lock(cpuctx, task_ctx);
1584        perf_pmu_disable(cpuctx->ctx.pmu);
1585
1586        /*
1587         * If there was an active task_ctx schedule it out.
1588         */
1589        if (task_ctx)
1590                task_ctx_sched_out(task_ctx);
1591
1592        /*
1593         * If the context we're installing events in is not the
1594         * active task_ctx, flip them.
1595         */
1596        if (ctx->task && task_ctx != ctx) {
1597                if (task_ctx)
1598                        raw_spin_unlock(&task_ctx->lock);
1599                raw_spin_lock(&ctx->lock);
1600                task_ctx = ctx;
1601        }
1602
1603        if (task_ctx) {
1604                cpuctx->task_ctx = task_ctx;
1605                task = task_ctx->task;
1606        }
1607
1608        cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1609
1610        update_context_time(ctx);
1611        /*
1612         * update cgrp time only if current cgrp
1613         * matches event->cgrp. Must be done before
1614         * calling add_event_to_ctx()
1615         */
1616        update_cgrp_time_from_event(event);
1617
1618        add_event_to_ctx(event, ctx);
1619
1620        /*
1621         * Schedule everything back in
1622         */
1623        perf_event_sched_in(cpuctx, task_ctx, task);
1624
1625        perf_pmu_enable(cpuctx->ctx.pmu);
1626        perf_ctx_unlock(cpuctx, task_ctx);
1627
1628        return 0;
1629}
1630
1631/*
1632 * Attach a performance event to a context
1633 *
1634 * First we add the event to the list with the hardware enable bit
1635 * in event->hw_config cleared.
1636 *
1637 * If the event is attached to a task which is on a CPU we use a smp
1638 * call to enable it in the task context. The task might have been
1639 * scheduled away, but we check this in the smp call again.
1640 */
1641static void
1642perf_install_in_context(struct perf_event_context *ctx,
1643                        struct perf_event *event,
1644                        int cpu)
1645{
1646        struct task_struct *task = ctx->task;
1647
1648        lockdep_assert_held(&ctx->mutex);
1649
1650        event->ctx = ctx;
1651        if (event->cpu != -1)
1652                event->cpu = cpu;
1653
1654        if (!task) {
1655                /*
1656                 * Per cpu events are installed via an smp call and
1657                 * the install is always successful.
1658                 */
1659                cpu_function_call(cpu, __perf_install_in_context, event);
1660                return;
1661        }
1662
1663retry:
1664        if (!task_function_call(task, __perf_install_in_context, event))
1665                return;
1666
1667        raw_spin_lock_irq(&ctx->lock);
1668        /*
1669         * If we failed to find a running task, but find the context active now
1670         * that we've acquired the ctx->lock, retry.
1671         */
1672        if (ctx->is_active) {
1673                raw_spin_unlock_irq(&ctx->lock);
1674                goto retry;
1675        }
1676
1677        /*
1678         * Since the task isn't running, its safe to add the event, us holding
1679         * the ctx->lock ensures the task won't get scheduled in.
1680         */
1681        add_event_to_ctx(event, ctx);
1682        raw_spin_unlock_irq(&ctx->lock);
1683}
1684
1685/*
1686 * Put a event into inactive state and update time fields.
1687 * Enabling the leader of a group effectively enables all
1688 * the group members that aren't explicitly disabled, so we
1689 * have to update their ->tstamp_enabled also.
1690 * Note: this works for group members as well as group leaders
1691 * since the non-leader members' sibling_lists will be empty.
1692 */
1693static void __perf_event_mark_enabled(struct perf_event *event)
1694{
1695        struct perf_event *sub;
1696        u64 tstamp = perf_event_time(event);
1697
1698        event->state = PERF_EVENT_STATE_INACTIVE;
1699        event->tstamp_enabled = tstamp - event->total_time_enabled;
1700        list_for_each_entry(sub, &event->sibling_list, group_entry) {
1701                if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1702                        sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1703        }
1704}
1705
1706/*
1707 * Cross CPU call to enable a performance event
1708 */
1709static int __perf_event_enable(void *info)
1710{
1711        struct perf_event *event = info;
1712        struct perf_event_context *ctx = event->ctx;
1713        struct perf_event *leader = event->group_leader;
1714        struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1715        int err;
1716
1717        if (WARN_ON_ONCE(!ctx->is_active))
1718                return -EINVAL;
1719
1720        raw_spin_lock(&ctx->lock);
1721        update_context_time(ctx);
1722
1723        if (event->state >= PERF_EVENT_STATE_INACTIVE)
1724                goto unlock;
1725
1726        /*
1727         * set current task's cgroup time reference point
1728         */
1729        perf_cgroup_set_timestamp(current, ctx);
1730
1731        __perf_event_mark_enabled(event);
1732
1733        if (!event_filter_match(event)) {
1734                if (is_cgroup_event(event))
1735                        perf_cgroup_defer_enabled(event);
1736                goto unlock;
1737        }
1738
1739        /*
1740         * If the event is in a group and isn't the group leader,
1741         * then don't put it on unless the group is on.
1742         */
1743        if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1744                goto unlock;
1745
1746        if (!group_can_go_on(event, cpuctx, 1)) {
1747                err = -EEXIST;
1748        } else {
1749                if (event == leader)
1750                        err = group_sched_in(event, cpuctx, ctx);
1751                else
1752                        err = event_sched_in(event, cpuctx, ctx);
1753        }
1754
1755        if (err) {
1756                /*
1757                 * If this event can't go on and it's part of a
1758                 * group, then the whole group has to come off.
1759                 */
1760                if (leader != event)
1761                        group_sched_out(leader, cpuctx, ctx);
1762                if (leader->attr.pinned) {
1763                        update_group_times(leader);
1764                        leader->state = PERF_EVENT_STATE_ERROR;
1765                }
1766        }
1767
1768unlock:
1769        raw_spin_unlock(&ctx->lock);
1770
1771        return 0;
1772}
1773
1774/*
1775 * Enable a event.
1776 *
1777 * If event->ctx is a cloned context, callers must make sure that
1778 * every task struct that event->ctx->task could possibly point to
1779 * remains valid.  This condition is satisfied when called through
1780 * perf_event_for_each_child or perf_event_for_each as described
1781 * for perf_event_disable.
1782 */
1783void perf_event_enable(struct perf_event *event)
1784{
1785        struct perf_event_context *ctx = event->ctx;
1786        struct task_struct *task = ctx->task;
1787
1788        if (!task) {
1789                /*
1790                 * Enable the event on the cpu that it's on
1791                 */
1792                cpu_function_call(event->cpu, __perf_event_enable, event);
1793                return;
1794        }
1795
1796        raw_spin_lock_irq(&ctx->lock);
1797        if (event->state >= PERF_EVENT_STATE_INACTIVE)
1798                goto out;
1799
1800        /*
1801         * If the event is in error state, clear that first.
1802         * That way, if we see the event in error state below, we
1803         * know that it has gone back into error state, as distinct
1804         * from the task having been scheduled away before the
1805         * cross-call arrived.
1806         */
1807        if (event->state == PERF_EVENT_STATE_ERROR)
1808                event->state = PERF_EVENT_STATE_OFF;
1809
1810retry:
1811        if (!ctx->is_active) {
1812                __perf_event_mark_enabled(event);
1813                goto out;
1814        }
1815
1816        raw_spin_unlock_irq(&ctx->lock);
1817
1818        if (!task_function_call(task, __perf_event_enable, event))
1819                return;
1820
1821        raw_spin_lock_irq(&ctx->lock);
1822
1823        /*
1824         * If the context is active and the event is still off,
1825         * we need to retry the cross-call.
1826         */
1827        if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1828                /*
1829                 * task could have been flipped by a concurrent
1830                 * perf_event_context_sched_out()
1831                 */
1832                task = ctx->task;
1833                goto retry;
1834        }
1835
1836out:
1837        raw_spin_unlock_irq(&ctx->lock);
1838}
1839EXPORT_SYMBOL_GPL(perf_event_enable);
1840
1841int perf_event_refresh(struct perf_event *event, int refresh)
1842{
1843        /*
1844         * not supported on inherited events
1845         */
1846        if (event->attr.inherit || !is_sampling_event(event))
1847                return -EINVAL;
1848
1849        atomic_add(refresh, &event->event_limit);
1850        perf_event_enable(event);
1851
1852        return 0;
1853}
1854EXPORT_SYMBOL_GPL(perf_event_refresh);
1855
1856static void ctx_sched_out(struct perf_event_context *ctx,
1857                          struct perf_cpu_context *cpuctx,
1858                          enum event_type_t event_type)
1859{
1860        struct perf_event *event;
1861        int is_active = ctx->is_active;
1862
1863        ctx->is_active &= ~event_type;
1864        if (likely(!ctx->nr_events))
1865                return;
1866
1867        update_context_time(ctx);
1868        update_cgrp_time_from_cpuctx(cpuctx);
1869        if (!ctx->nr_active)
1870                return;
1871
1872        perf_pmu_disable(ctx->pmu);
1873        if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1874                list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1875                        group_sched_out(event, cpuctx, ctx);
1876        }
1877
1878        if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1879                list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1880                        group_sched_out(event, cpuctx, ctx);
1881        }
1882        perf_pmu_enable(ctx->pmu);
1883}
1884
1885/*
1886 * Test whether two contexts are equivalent, i.e. whether they
1887 * have both been cloned from the same version of the same context
1888 * and they both have the same number of enabled events.
1889 * If the number of enabled events is the same, then the set
1890 * of enabled events should be the same, because these are both
1891 * inherited contexts, therefore we can't access individual events
1892 * in them directly with an fd; we can only enable/disable all
1893 * events via prctl, or enable/disable all events in a family
1894 * via ioctl, which will have the same effect on both contexts.
1895 */
1896static int context_equiv(struct perf_event_context *ctx1,
1897                         struct perf_event_context *ctx2)
1898{
1899        return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1900                && ctx1->parent_gen == ctx2->parent_gen
1901                && !ctx1->pin_count && !ctx2->pin_count;
1902}
1903
1904static void __perf_event_sync_stat(struct perf_event *event,
1905                                     struct perf_event *next_event)
1906{
1907        u64 value;
1908
1909        if (!event->attr.inherit_stat)
1910                return;
1911
1912        /*
1913         * Update the event value, we cannot use perf_event_read()
1914         * because we're in the middle of a context switch and have IRQs
1915         * disabled, which upsets smp_call_function_single(), however
1916         * we know the event must be on the current CPU, therefore we
1917         * don't need to use it.
1918         */
1919        switch (event->state) {
1920        case PERF_EVENT_STATE_ACTIVE:
1921                event->pmu->read(event);
1922                /* fall-through */
1923
1924        case PERF_EVENT_STATE_INACTIVE:
1925                update_event_times(event);
1926                break;
1927
1928        default:
1929                break;
1930        }
1931
1932        /*
1933         * In order to keep per-task stats reliable we need to flip the event
1934         * values when we flip the contexts.
1935         */
1936        value = local64_read(&next_event->count);
1937        value = local64_xchg(&event->count, value);
1938        local64_set(&next_event->count, value);
1939
1940        swap(event->total_time_enabled, next_event->total_time_enabled);
1941        swap(event->total_time_running, next_event->total_time_running);
1942
1943        /*
1944         * Since we swizzled the values, update the user visible data too.
1945         */
1946        perf_event_update_userpage(event);
1947        perf_event_update_userpage(next_event);
1948}
1949
1950#define list_next_entry(pos, member) \
1951        list_entry(pos->member.next, typeof(*pos), member)
1952
1953static void perf_event_sync_stat(struct perf_event_context *ctx,
1954                                   struct perf_event_context *next_ctx)
1955{
1956        struct perf_event *event, *next_event;
1957
1958        if (!ctx->nr_stat)
1959                return;
1960
1961        update_context_time(ctx);
1962
1963        event = list_first_entry(&ctx->event_list,
1964                                   struct perf_event, event_entry);
1965
1966        next_event = list_first_entry(&next_ctx->event_list,
1967                                        struct perf_event, event_entry);
1968
1969        while (&event->event_entry != &ctx->event_list &&
1970               &next_event->event_entry != &next_ctx->event_list) {
1971
1972                __perf_event_sync_stat(event, next_event);
1973
1974                event = list_next_entry(event, event_entry);
1975                next_event = list_next_entry(next_event, event_entry);
1976        }
1977}
1978
1979static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1980                                         struct task_struct *next)
1981{
1982        struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1983        struct perf_event_context *next_ctx;
1984        struct perf_event_context *parent;
1985        struct perf_cpu_context *cpuctx;
1986        int do_switch = 1;
1987
1988        if (likely(!ctx))
1989                return;
1990
1991        cpuctx = __get_cpu_context(ctx);
1992        if (!cpuctx->task_ctx)
1993                return;
1994
1995        rcu_read_lock();
1996        parent = rcu_dereference(ctx->parent_ctx);
1997        next_ctx = next->perf_event_ctxp[ctxn];
1998        if (parent && next_ctx &&
1999            rcu_dereference(next_ctx->parent_ctx) == parent) {
2000                /*
2001                 * Looks like the two contexts are clones, so we might be
2002                 * able to optimize the context switch.  We lock both
2003                 * contexts and check that they are clones under the
2004                 * lock (including re-checking that neither has been
2005                 * uncloned in the meantime).  It doesn't matter which
2006                 * order we take the locks because no other cpu could
2007                 * be trying to lock both of these tasks.
2008                 */
2009                raw_spin_lock(&ctx->lock);
2010                raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2011                if (context_equiv(ctx, next_ctx)) {
2012                        /*
2013                         * XXX do we need a memory barrier of sorts
2014                         * wrt to rcu_dereference() of perf_event_ctxp
2015                         */
2016                        task->perf_event_ctxp[ctxn] = next_ctx;
2017                        next->perf_event_ctxp[ctxn] = ctx;
2018                        ctx->task = next;
2019                        next_ctx->task = task;
2020                        do_switch = 0;
2021
2022                        perf_event_sync_stat(ctx, next_ctx);
2023                }
2024                raw_spin_unlock(&next_ctx->lock);
2025                raw_spin_unlock(&ctx->lock);
2026        }
2027        rcu_read_unlock();
2028
2029        if (do_switch) {
2030                raw_spin_lock(&ctx->lock);
2031                ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2032                cpuctx->task_ctx = NULL;
2033                raw_spin_unlock(&ctx->lock);
2034        }
2035}
2036
2037#define for_each_task_context_nr(ctxn)                                  \
2038        for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2039
2040/*
2041 * Called from scheduler to remove the events of the current task,
2042 * with interrupts disabled.
2043 *
2044 * We stop each event and update the event value in event->count.
2045 *
2046 * This does not protect us against NMI, but disable()
2047 * sets the disabled bit in the control field of event _before_
2048 * accessing the event control register. If a NMI hits, then it will
2049 * not restart the event.
2050 */
2051void __perf_event_task_sched_out(struct task_struct *task,
2052                                 struct task_struct *next)
2053{
2054        int ctxn;
2055
2056        for_each_task_context_nr(ctxn)
2057                perf_event_context_sched_out(task, ctxn, next);
2058
2059        /*
2060         * if cgroup events exist on this CPU, then we need
2061         * to check if we have to switch out PMU state.
2062         * cgroup event are system-wide mode only
2063         */
2064        if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2065                perf_cgroup_sched_out(task, next);
2066}
2067
2068static void task_ctx_sched_out(struct perf_event_context *ctx)
2069{
2070        struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2071
2072        if (!cpuctx->task_ctx)
2073                return;
2074
2075        if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2076                return;
2077
2078        ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2079        cpuctx->task_ctx = NULL;
2080}
2081
2082/*
2083 * Called with IRQs disabled
2084 */
2085static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2086                              enum event_type_t event_type)
2087{
2088        ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2089}
2090
2091static void
2092ctx_pinned_sched_in(struct perf_event_context *ctx,
2093                    struct perf_cpu_context *cpuctx)
2094{
2095        struct perf_event *event;
2096
2097        list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2098                if (event->state <= PERF_EVENT_STATE_OFF)
2099                        continue;
2100                if (!event_filter_match(event))
2101                        continue;
2102
2103                /* may need to reset tstamp_enabled */
2104                if (is_cgroup_event(event))
2105                        perf_cgroup_mark_enabled(event, ctx);
2106
2107                if (group_can_go_on(event, cpuctx, 1))
2108                        group_sched_in(event, cpuctx, ctx);
2109
2110                /*
2111                 * If this pinned group hasn't been scheduled,
2112                 * put it in error state.
2113                 */
2114                if (event->state == PERF_EVENT_STATE_INACTIVE) {
2115                        update_group_times(event);
2116                        event->state = PERF_EVENT_STATE_ERROR;
2117                }
2118        }
2119}
2120
2121static void
2122ctx_flexible_sched_in(struct perf_event_context *ctx,
2123                      struct perf_cpu_context *cpuctx)
2124{
2125        struct perf_event *event;
2126        int can_add_hw = 1;
2127
2128        list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2129                /* Ignore events in OFF or ERROR state */
2130                if (event->state <= PERF_EVENT_STATE_OFF)
2131                        continue;
2132                /*
2133                 * Listen to the 'cpu' scheduling filter constraint
2134                 * of events:
2135                 */
2136                if (!event_filter_match(event))
2137                        continue;
2138
2139                /* may need to reset tstamp_enabled */
2140                if (is_cgroup_event(event))
2141                        perf_cgroup_mark_enabled(event, ctx);
2142
2143                if (group_can_go_on(event, cpuctx, can_add_hw)) {
2144                        if (group_sched_in(event, cpuctx, ctx))
2145                                can_add_hw = 0;
2146                }
2147        }
2148}
2149
2150static void
2151ctx_sched_in(struct perf_event_context *ctx,
2152             struct perf_cpu_context *cpuctx,
2153             enum event_type_t event_type,
2154             struct task_struct *task)
2155{
2156        u64 now;
2157        int is_active = ctx->is_active;
2158
2159        ctx->is_active |= event_type;
2160        if (likely(!ctx->nr_events))
2161                return;
2162
2163        now = perf_clock();
2164        ctx->timestamp = now;
2165        perf_cgroup_set_timestamp(task, ctx);
2166        /*
2167         * First go through the list and put on any pinned groups
2168         * in order to give them the best chance of going on.
2169         */
2170        if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2171                ctx_pinned_sched_in(ctx, cpuctx);
2172
2173        /* Then walk through the lower prio flexible groups */
2174        if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2175                ctx_flexible_sched_in(ctx, cpuctx);
2176}
2177
2178static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2179                             enum event_type_t event_type,
2180                             struct task_struct *task)
2181{
2182        struct perf_event_context *ctx = &cpuctx->ctx;
2183
2184        ctx_sched_in(ctx, cpuctx, event_type, task);
2185}
2186
2187static void perf_event_context_sched_in(struct perf_event_context *ctx,
2188                                        struct task_struct *task)
2189{
2190        struct perf_cpu_context *cpuctx;
2191
2192        cpuctx = __get_cpu_context(ctx);
2193        if (cpuctx->task_ctx == ctx)
2194                return;
2195
2196        perf_ctx_lock(cpuctx, ctx);
2197        perf_pmu_disable(ctx->pmu);
2198        /*
2199         * We want to keep the following priority order:
2200         * cpu pinned (that don't need to move), task pinned,
2201         * cpu flexible, task flexible.
2202         */
2203        cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2204
2205        if (ctx->nr_events)
2206                cpuctx->task_ctx = ctx;
2207
2208        perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2209
2210        perf_pmu_enable(ctx->pmu);
2211        perf_ctx_unlock(cpuctx, ctx);
2212
2213        /*
2214         * Since these rotations are per-cpu, we need to ensure the
2215         * cpu-context we got scheduled on is actually rotating.
2216         */
2217        perf_pmu_rotate_start(ctx->pmu);
2218}
2219
2220/*
2221 * When sampling the branck stack in system-wide, it may be necessary
2222 * to flush the stack on context switch. This happens when the branch
2223 * stack does not tag its entries with the pid of the current task.
2224 * Otherwise it becomes impossible to associate a branch entry with a
2225 * task. This ambiguity is more likely to appear when the branch stack
2226 * supports priv level filtering and the user sets it to monitor only
2227 * at the user level (which could be a useful measurement in system-wide
2228 * mode). In that case, the risk is high of having a branch stack with
2229 * branch from multiple tasks. Flushing may mean dropping the existing
2230 * entries or stashing them somewhere in the PMU specific code layer.
2231 *
2232 * This function provides the context switch callback to the lower code
2233 * layer. It is invoked ONLY when there is at least one system-wide context
2234 * with at least one active event using taken branch sampling.
2235 */
2236static void perf_branch_stack_sched_in(struct task_struct *prev,
2237                                       struct task_struct *task)
2238{
2239        struct perf_cpu_context *cpuctx;
2240        struct pmu *pmu;
2241        unsigned long flags;
2242
2243        /* no need to flush branch stack if not changing task */
2244        if (prev == task)
2245                return;
2246
2247        local_irq_save(flags);
2248
2249        rcu_read_lock();
2250
2251        list_for_each_entry_rcu(pmu, &pmus, entry) {
2252                cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2253
2254                /*
2255                 * check if the context has at least one
2256                 * event using PERF_SAMPLE_BRANCH_STACK
2257                 */
2258                if (cpuctx->ctx.nr_branch_stack > 0
2259                    && pmu->flush_branch_stack) {
2260
2261                        pmu = cpuctx->ctx.pmu;
2262
2263                        perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2264
2265                        perf_pmu_disable(pmu);
2266
2267                        pmu->flush_branch_stack();
2268
2269                        perf_pmu_enable(pmu);
2270
2271                        perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2272                }
2273        }
2274
2275        rcu_read_unlock();
2276
2277        local_irq_restore(flags);
2278}
2279
2280/*
2281 * Called from scheduler to add the events of the current task
2282 * with interrupts disabled.
2283 *
2284 * We restore the event value and then enable it.
2285 *
2286 * This does not protect us against NMI, but enable()
2287 * sets the enabled bit in the control field of event _before_
2288 * accessing the event control register. If a NMI hits, then it will
2289 * keep the event running.
2290 */
2291void __perf_event_task_sched_in(struct task_struct *prev,
2292                                struct task_struct *task)
2293{
2294        struct perf_event_context *ctx;
2295        int ctxn;
2296
2297        for_each_task_context_nr(ctxn) {
2298                ctx = task->perf_event_ctxp[ctxn];
2299                if (likely(!ctx))
2300                        continue;
2301
2302                perf_event_context_sched_in(ctx, task);
2303        }
2304        /*
2305         * if cgroup events exist on this CPU, then we need
2306         * to check if we have to switch in PMU state.
2307         * cgroup event are system-wide mode only
2308         */
2309        if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2310                perf_cgroup_sched_in(prev, task);
2311
2312        /* check for system-wide branch_stack events */
2313        if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2314                perf_branch_stack_sched_in(prev, task);
2315}
2316
2317static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2318{
2319        u64 frequency = event->attr.sample_freq;
2320        u64 sec = NSEC_PER_SEC;
2321        u64 divisor, dividend;
2322
2323        int count_fls, nsec_fls, frequency_fls, sec_fls;
2324
2325        count_fls = fls64(count);
2326        nsec_fls = fls64(nsec);
2327        frequency_fls = fls64(frequency);
2328        sec_fls = 30;
2329
2330        /*
2331         * We got @count in @nsec, with a target of sample_freq HZ
2332         * the target period becomes:
2333         *
2334         *             @count * 10^9
2335         * period = -------------------
2336         *          @nsec * sample_freq
2337         *
2338         */
2339
2340        /*
2341         * Reduce accuracy by one bit such that @a and @b converge
2342         * to a similar magnitude.
2343         */
2344#define REDUCE_FLS(a, b)                \
2345do {                                    \
2346        if (a##_fls > b##_fls) {        \
2347                a >>= 1;                \
2348                a##_fls--;              \
2349        } else {                        \
2350                b >>= 1;                \
2351                b##_fls--;              \
2352        }                               \
2353} while (0)
2354
2355        /*
2356         * Reduce accuracy until either term fits in a u64, then proceed with
2357         * the other, so that finally we can do a u64/u64 division.
2358         */
2359        while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2360                REDUCE_FLS(nsec, frequency);
2361                REDUCE_FLS(sec, count);
2362        }
2363
2364        if (count_fls + sec_fls > 64) {
2365                divisor = nsec * frequency;
2366
2367                while (count_fls + sec_fls > 64) {
2368                        REDUCE_FLS(count, sec);
2369                        divisor >>= 1;
2370                }
2371
2372                dividend = count * sec;
2373        } else {
2374                dividend = count * sec;
2375
2376                while (nsec_fls + frequency_fls > 64) {
2377                        REDUCE_FLS(nsec, frequency);
2378                        dividend >>= 1;
2379                }
2380
2381                divisor = nsec * frequency;
2382        }
2383
2384        if (!divisor)
2385                return dividend;
2386
2387        return div64_u64(dividend, divisor);
2388}
2389
2390static DEFINE_PER_CPU(int, perf_throttled_count);
2391static DEFINE_PER_CPU(u64, perf_throttled_seq);
2392
2393static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2394{
2395        struct hw_perf_event *hwc = &event->hw;
2396        s64 period, sample_period;
2397        s64 delta;
2398
2399        period = perf_calculate_period(event, nsec, count);
2400
2401        delta = (s64)(period - hwc->sample_period);
2402        delta = (delta + 7) / 8; /* low pass filter */
2403
2404        sample_period = hwc->sample_period + delta;
2405
2406        if (!sample_period)
2407                sample_period = 1;
2408
2409        hwc->sample_period = sample_period;
2410
2411        if (local64_read(&hwc->period_left) > 8*sample_period) {
2412                if (disable)
2413                        event->pmu->stop(event, PERF_EF_UPDATE);
2414
2415                local64_set(&hwc->period_left, 0);
2416
2417                if (disable)
2418                        event->pmu->start(event, PERF_EF_RELOAD);
2419        }
2420}
2421
2422/*
2423 * combine freq adjustment with unthrottling to avoid two passes over the
2424 * events. At the same time, make sure, having freq events does not change
2425 * the rate of unthrottling as that would introduce bias.
2426 */
2427static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2428                                           int needs_unthr)
2429{
2430        struct perf_event *event;
2431        struct hw_perf_event *hwc;
2432        u64 now, period = TICK_NSEC;
2433        s64 delta;
2434
2435        /*
2436         * only need to iterate over all events iff:
2437         * - context have events in frequency mode (needs freq adjust)
2438         * - there are events to unthrottle on this cpu
2439         */
2440        if (!(ctx->nr_freq || needs_unthr))
2441                return;
2442
2443        raw_spin_lock(&ctx->lock);
2444        perf_pmu_disable(ctx->pmu);
2445
2446        list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2447                if (event->state != PERF_EVENT_STATE_ACTIVE)
2448                        continue;
2449
2450                if (!event_filter_match(event))
2451                        continue;
2452
2453                hwc = &event->hw;
2454
2455                if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2456                        hwc->interrupts = 0;
2457                        perf_log_throttle(event, 1);
2458                        event->pmu->start(event, 0);
2459                }
2460
2461                if (!event->attr.freq || !event->attr.sample_freq)
2462                        continue;
2463
2464                /*
2465                 * stop the event and update event->count
2466                 */
2467                event->pmu->stop(event, PERF_EF_UPDATE);
2468
2469                now = local64_read(&event->count);
2470                delta = now - hwc->freq_count_stamp;
2471                hwc->freq_count_stamp = now;
2472
2473                /*
2474                 * restart the event
2475                 * reload only if value has changed
2476                 * we have stopped the event so tell that
2477                 * to perf_adjust_period() to avoid stopping it
2478                 * twice.
2479                 */
2480                if (delta > 0)
2481                        perf_adjust_period(event, period, delta, false);
2482
2483                event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2484        }
2485
2486        perf_pmu_enable(ctx->pmu);
2487        raw_spin_unlock(&ctx->lock);
2488}
2489
2490/*
2491 * Round-robin a context's events:
2492 */
2493static void rotate_ctx(struct perf_event_context *ctx)
2494{
2495        /*
2496         * Rotate the first entry last of non-pinned groups. Rotation might be
2497         * disabled by the inheritance code.
2498         */
2499        if (!ctx->rotate_disable)
2500                list_rotate_left(&ctx->flexible_groups);
2501}
2502
2503/*
2504 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2505 * because they're strictly cpu affine and rotate_start is called with IRQs
2506 * disabled, while rotate_context is called from IRQ context.
2507 */
2508static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2509{
2510        struct perf_event_context *ctx = NULL;
2511        int rotate = 0, remove = 1;
2512
2513        if (cpuctx->ctx.nr_events) {
2514                remove = 0;
2515                if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2516                        rotate = 1;
2517        }
2518
2519        ctx = cpuctx->task_ctx;
2520        if (ctx && ctx->nr_events) {
2521                remove = 0;
2522                if (ctx->nr_events != ctx->nr_active)
2523                        rotate = 1;
2524        }
2525
2526        if (!rotate)
2527                goto done;
2528
2529        perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2530        perf_pmu_disable(cpuctx->ctx.pmu);
2531
2532        cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2533        if (ctx)
2534                ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2535
2536        rotate_ctx(&cpuctx->ctx);
2537        if (ctx)
2538                rotate_ctx(ctx);
2539
2540        perf_event_sched_in(cpuctx, ctx, current);
2541
2542        perf_pmu_enable(cpuctx->ctx.pmu);
2543        perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2544done:
2545        if (remove)
2546                list_del_init(&cpuctx->rotation_list);
2547}
2548
2549void perf_event_task_tick(void)
2550{
2551        struct list_head *head = &__get_cpu_var(rotation_list);
2552        struct perf_cpu_context *cpuctx, *tmp;
2553        struct perf_event_context *ctx;
2554        int throttled;
2555
2556        WARN_ON(!irqs_disabled());
2557
2558        __this_cpu_inc(perf_throttled_seq);
2559        throttled = __this_cpu_xchg(perf_throttled_count, 0);
2560
2561        list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2562                ctx = &cpuctx->ctx;
2563                perf_adjust_freq_unthr_context(ctx, throttled);
2564
2565                ctx = cpuctx->task_ctx;
2566                if (ctx)
2567                        perf_adjust_freq_unthr_context(ctx, throttled);
2568
2569                if (cpuctx->jiffies_interval == 1 ||
2570                                !(jiffies % cpuctx->jiffies_interval))
2571                        perf_rotate_context(cpuctx);
2572        }
2573}
2574
2575static int event_enable_on_exec(struct perf_event *event,
2576                                struct perf_event_context *ctx)
2577{
2578        if (!event->attr.enable_on_exec)
2579                return 0;
2580
2581        event->attr.enable_on_exec = 0;
2582        if (event->state >= PERF_EVENT_STATE_INACTIVE)
2583                return 0;
2584
2585        __perf_event_mark_enabled(event);
2586
2587        return 1;
2588}
2589
2590/*
2591 * Enable all of a task's events that have been marked enable-on-exec.
2592 * This expects task == current.
2593 */
2594static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2595{
2596        struct perf_event *event;
2597        unsigned long flags;
2598        int enabled = 0;
2599        int ret;
2600
2601        local_irq_save(flags);
2602        if (!ctx || !ctx->nr_events)
2603                goto out;
2604
2605        /*
2606         * We must ctxsw out cgroup events to avoid conflict
2607         * when invoking perf_task_event_sched_in() later on
2608         * in this function. Otherwise we end up trying to
2609         * ctxswin cgroup events which are already scheduled
2610         * in.
2611         */
2612        perf_cgroup_sched_out(current, NULL);
2613
2614        raw_spin_lock(&ctx->lock);
2615        task_ctx_sched_out(ctx);
2616
2617        list_for_each_entry(event, &ctx->event_list, event_entry) {
2618                ret = event_enable_on_exec(event, ctx);
2619                if (ret)
2620                        enabled = 1;
2621        }
2622
2623        /*
2624         * Unclone this context if we enabled any event.
2625         */
2626        if (enabled)
2627                unclone_ctx(ctx);
2628
2629        raw_spin_unlock(&ctx->lock);
2630
2631        /*
2632         * Also calls ctxswin for cgroup events, if any:
2633         */
2634        perf_event_context_sched_in(ctx, ctx->task);
2635out:
2636        local_irq_restore(flags);
2637}
2638
2639/*
2640 * Cross CPU call to read the hardware event
2641 */
2642static void __perf_event_read(void *info)
2643{
2644        struct perf_event *event = info;
2645        struct perf_event_context *ctx = event->ctx;
2646        struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2647
2648        /*
2649         * If this is a task context, we need to check whether it is
2650         * the current task context of this cpu.  If not it has been
2651         * scheduled out before the smp call arrived.  In that case
2652         * event->count would have been updated to a recent sample
2653         * when the event was scheduled out.
2654         */
2655        if (ctx->task && cpuctx->task_ctx != ctx)
2656                return;
2657
2658        raw_spin_lock(&ctx->lock);
2659        if (ctx->is_active) {
2660                update_context_time(ctx);
2661                update_cgrp_time_from_event(event);
2662        }
2663        update_event_times(event);
2664        if (event->state == PERF_EVENT_STATE_ACTIVE)
2665                event->pmu->read(event);
2666        raw_spin_unlock(&ctx->lock);
2667}
2668
2669static inline u64 perf_event_count(struct perf_event *event)
2670{
2671        return local64_read(&event->count) + atomic64_read(&event->child_count);
2672}
2673
2674static u64 perf_event_read(struct perf_event *event)
2675{
2676        /*
2677         * If event is enabled and currently active on a CPU, update the
2678         * value in the event structure:
2679         */
2680        if (event->state == PERF_EVENT_STATE_ACTIVE) {
2681                smp_call_function_single(event->oncpu,
2682                                         __perf_event_read, event, 1);
2683        } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2684                struct perf_event_context *ctx = event->ctx;
2685                unsigned long flags;
2686
2687                raw_spin_lock_irqsave(&ctx->lock, flags);
2688                /*
2689                 * may read while context is not active
2690                 * (e.g., thread is blocked), in that case
2691                 * we cannot update context time
2692                 */
2693                if (ctx->is_active) {
2694                        update_context_time(ctx);
2695                        update_cgrp_time_from_event(event);
2696                }
2697                update_event_times(event);
2698                raw_spin_unlock_irqrestore(&ctx->lock, flags);
2699        }
2700
2701        return perf_event_count(event);
2702}
2703
2704/*
2705 * Initialize the perf_event context in a task_struct:
2706 */
2707static void __perf_event_init_context(struct perf_event_context *ctx)
2708{
2709        raw_spin_lock_init(&ctx->lock);
2710        mutex_init(&ctx->mutex);
2711        INIT_LIST_HEAD(&ctx->pinned_groups);
2712        INIT_LIST_HEAD(&ctx->flexible_groups);
2713        INIT_LIST_HEAD(&ctx->event_list);
2714        atomic_set(&ctx->refcount, 1);
2715}
2716
2717static struct perf_event_context *
2718alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2719{
2720        struct perf_event_context *ctx;
2721
2722        ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2723        if (!ctx)
2724                return NULL;
2725
2726        __perf_event_init_context(ctx);
2727        if (task) {
2728                ctx->task = task;
2729                get_task_struct(task);
2730        }
2731        ctx->pmu = pmu;
2732
2733        return ctx;
2734}
2735
2736static struct task_struct *
2737find_lively_task_by_vpid(pid_t vpid)
2738{
2739        struct task_struct *task;
2740        int err;
2741
2742        rcu_read_lock();
2743        if (!vpid)
2744                task = current;
2745        else
2746                task = find_task_by_vpid(vpid);
2747        if (task)
2748                get_task_struct(task);
2749        rcu_read_unlock();
2750
2751        if (!task)
2752                return ERR_PTR(-ESRCH);
2753
2754        /* Reuse ptrace permission checks for now. */
2755        err = -EACCES;
2756        if (!ptrace_may_access(task, PTRACE_MODE_READ))
2757                goto errout;
2758
2759        return task;
2760errout:
2761        put_task_struct(task);
2762        return ERR_PTR(err);
2763
2764}
2765
2766/*
2767 * Returns a matching context with refcount and pincount.
2768 */
2769static struct perf_event_context *
2770find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2771{
2772        struct perf_event_context *ctx;
2773        struct perf_cpu_context *cpuctx;
2774        unsigned long flags;
2775        int ctxn, err;
2776
2777        if (!task) {
2778                /* Must be root to operate on a CPU event: */
2779                if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2780                        return ERR_PTR(-EACCES);
2781
2782                /*
2783                 * We could be clever and allow to attach a event to an
2784                 * offline CPU and activate it when the CPU comes up, but
2785                 * that's for later.
2786                 */
2787                if (!cpu_online(cpu))
2788                        return ERR_PTR(-ENODEV);
2789
2790                cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2791                ctx = &cpuctx->ctx;
2792                get_ctx(ctx);
2793                ++ctx->pin_count;
2794
2795                return ctx;
2796        }
2797
2798        err = -EINVAL;
2799        ctxn = pmu->task_ctx_nr;
2800        if (ctxn < 0)
2801                goto errout;
2802
2803retry:
2804        ctx = perf_lock_task_context(task, ctxn, &flags);
2805        if (ctx) {
2806                unclone_ctx(ctx);
2807                ++ctx->pin_count;
2808                raw_spin_unlock_irqrestore(&ctx->lock, flags);
2809        } else {
2810                ctx = alloc_perf_context(pmu, task);
2811                err = -ENOMEM;
2812                if (!ctx)
2813                        goto errout;
2814
2815                err = 0;
2816                mutex_lock(&task->perf_event_mutex);
2817                /*
2818                 * If it has already passed perf_event_exit_task().
2819                 * we must see PF_EXITING, it takes this mutex too.
2820                 */
2821                if (task->flags & PF_EXITING)
2822                        err = -ESRCH;
2823                else if (task->perf_event_ctxp[ctxn])
2824                        err = -EAGAIN;
2825                else {
2826                        get_ctx(ctx);
2827                        ++ctx->pin_count;
2828                        rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2829                }
2830                mutex_unlock(&task->perf_event_mutex);
2831
2832                if (unlikely(err)) {
2833                        put_ctx(ctx);
2834
2835                        if (err == -EAGAIN)
2836                                goto retry;
2837                        goto errout;
2838                }
2839        }
2840
2841        return ctx;
2842
2843errout:
2844        return ERR_PTR(err);
2845}
2846
2847static void perf_event_free_filter(struct perf_event *event);
2848
2849static void free_event_rcu(struct rcu_head *head)
2850{
2851        struct perf_event *event;
2852
2853        event = container_of(head, struct perf_event, rcu_head);
2854        if (event->ns)
2855                put_pid_ns(event->ns);
2856        perf_event_free_filter(event);
2857        kfree(event);
2858}
2859
2860static void ring_buffer_put(struct ring_buffer *rb);
2861
2862static void free_event(struct perf_event *event)
2863{
2864        irq_work_sync(&event->pending);
2865
2866        if (!event->parent) {
2867                if (event->attach_state & PERF_ATTACH_TASK)
2868                        static_key_slow_dec_deferred(&perf_sched_events);
2869                if (event->attr.mmap || event->attr.mmap_data)
2870                        atomic_dec(&nr_mmap_events);
2871                if (event->attr.comm)
2872                        atomic_dec(&nr_comm_events);
2873                if (event->attr.task)
2874                        atomic_dec(&nr_task_events);
2875                if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2876                        put_callchain_buffers();
2877                if (is_cgroup_event(event)) {
2878                        atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2879                        static_key_slow_dec_deferred(&perf_sched_events);
2880                }
2881
2882                if (has_branch_stack(event)) {
2883                        static_key_slow_dec_deferred(&perf_sched_events);
2884                        /* is system-wide event */
2885                        if (!(event->attach_state & PERF_ATTACH_TASK))
2886                                atomic_dec(&per_cpu(perf_branch_stack_events,
2887                                                    event->cpu));
2888                }
2889        }
2890
2891        if (event->rb) {
2892                ring_buffer_put(event->rb);
2893                event->rb = NULL;
2894        }
2895
2896        if (is_cgroup_event(event))
2897                perf_detach_cgroup(event);
2898
2899        if (event->destroy)
2900                event->destroy(event);
2901
2902        if (event->ctx)
2903                put_ctx(event->ctx);
2904
2905        call_rcu(&event->rcu_head, free_event_rcu);
2906}
2907
2908int perf_event_release_kernel(struct perf_event *event)
2909{
2910        struct perf_event_context *ctx = event->ctx;
2911
2912        WARN_ON_ONCE(ctx->parent_ctx);
2913        /*
2914         * There are two ways this annotation is useful:
2915         *
2916         *  1) there is a lock recursion from perf_event_exit_task
2917         *     see the comment there.
2918         *
2919         *  2) there is a lock-inversion with mmap_sem through
2920         *     perf_event_read_group(), which takes faults while
2921         *     holding ctx->mutex, however this is called after
2922         *     the last filedesc died, so there is no possibility
2923         *     to trigger the AB-BA case.
2924         */
2925        mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2926        raw_spin_lock_irq(&ctx->lock);
2927        perf_group_detach(event);
2928        raw_spin_unlock_irq(&ctx->lock);
2929        perf_remove_from_context(event);
2930        mutex_unlock(&ctx->mutex);
2931
2932        free_event(event);
2933
2934        return 0;
2935}
2936EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2937
2938/*
2939 * Called when the last reference to the file is gone.
2940 */
2941static void put_event(struct perf_event *event)
2942{
2943        struct task_struct *owner;
2944
2945        if (!atomic_long_dec_and_test(&event->refcount))
2946                return;
2947
2948        rcu_read_lock();
2949        owner = ACCESS_ONCE(event->owner);
2950        /*
2951         * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2952         * !owner it means the list deletion is complete and we can indeed
2953         * free this event, otherwise we need to serialize on
2954         * owner->perf_event_mutex.
2955         */
2956        smp_read_barrier_depends();
2957        if (owner) {
2958                /*
2959                 * Since delayed_put_task_struct() also drops the last
2960                 * task reference we can safely take a new reference
2961                 * while holding the rcu_read_lock().
2962                 */
2963                get_task_struct(owner);
2964        }
2965        rcu_read_unlock();
2966
2967        if (owner) {
2968                mutex_lock(&owner->perf_event_mutex);
2969                /*
2970                 * We have to re-check the event->owner field, if it is cleared
2971                 * we raced with perf_event_exit_task(), acquiring the mutex
2972                 * ensured they're done, and we can proceed with freeing the
2973                 * event.
2974                 */
2975                if (event->owner)
2976                        list_del_init(&event->owner_entry);
2977                mutex_unlock(&owner->perf_event_mutex);
2978                put_task_struct(owner);
2979        }
2980
2981        perf_event_release_kernel(event);
2982}
2983
2984static int perf_release(struct inode *inode, struct file *file)
2985{
2986        put_event(file->private_data);
2987        return 0;
2988}
2989
2990u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
2991{
2992        struct perf_event *child;
2993        u64 total = 0;
2994
2995        *enabled = 0;
2996        *running = 0;
2997
2998        mutex_lock(&event->child_mutex);
2999        total += perf_event_read(event);
3000        *enabled += event->total_time_enabled +
3001                        atomic64_read(&event->child_total_time_enabled);
3002        *running += event->total_time_running +
3003                        atomic64_read(&event->child_total_time_running);
3004
3005        list_for_each_entry(child, &event->child_list, child_list) {
3006                total += perf_event_read(child);
3007                *enabled += child->total_time_enabled;
3008                *running += child->total_time_running;
3009        }
3010        mutex_unlock(&event->child_mutex);
3011
3012        return total;
3013}
3014EXPORT_SYMBOL_GPL(perf_event_read_value);
3015
3016static int perf_event_read_group(struct perf_event *event,
3017                                   u64 read_format, char __user *buf)
3018{
3019        struct perf_event *leader = event->group_leader, *sub;
3020        int n = 0, size = 0, ret = -EFAULT;
3021        struct perf_event_context *ctx = leader->ctx;
3022        u64 values[5];
3023        u64 count, enabled, running;
3024
3025        mutex_lock(&ctx->mutex);
3026        count = perf_event_read_value(leader, &enabled, &running);
3027
3028        values[n++] = 1 + leader->nr_siblings;
3029        if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3030                values[n++] = enabled;
3031        if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3032                values[n++] = running;
3033        values[n++] = count;
3034        if (read_format & PERF_FORMAT_ID)
3035                values[n++] = primary_event_id(leader);
3036
3037        size = n * sizeof(u64);
3038
3039        if (copy_to_user(buf, values, size))
3040                goto unlock;
3041
3042        ret = size;
3043
3044        list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3045                n = 0;
3046
3047                values[n++] = perf_event_read_value(sub, &enabled, &running);
3048                if (read_format & PERF_FORMAT_ID)
3049                        values[n++] = primary_event_id(sub);
3050
3051                size = n * sizeof(u64);
3052
3053                if (copy_to_user(buf + ret, values, size)) {
3054                        ret = -EFAULT;
3055                        goto unlock;
3056                }
3057
3058                ret += size;
3059        }
3060unlock:
3061        mutex_unlock(&ctx->mutex);
3062
3063        return ret;
3064}
3065
3066static int perf_event_read_one(struct perf_event *event,
3067                                 u64 read_format, char __user *buf)
3068{
3069        u64 enabled, running;
3070        u64 values[4];
3071        int n = 0;
3072
3073        values[n++] = perf_event_read_value(event, &enabled, &running);
3074        if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3075                values[n++] = enabled;
3076        if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3077                values[n++] = running;
3078        if (read_format & PERF_FORMAT_ID)
3079                values[n++] = primary_event_id(event);
3080
3081        if (copy_to_user(buf, values, n * sizeof(u64)))
3082                return -EFAULT;
3083
3084        return n * sizeof(u64);
3085}
3086
3087/*
3088 * Read the performance event - simple non blocking version for now
3089 */
3090static ssize_t
3091perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3092{
3093        u64 read_format = event->attr.read_format;
3094        int ret;
3095
3096        /*
3097         * Return end-of-file for a read on a event that is in
3098         * error state (i.e. because it was pinned but it couldn't be
3099         * scheduled on to the CPU at some point).
3100         */
3101        if (event->state == PERF_EVENT_STATE_ERROR)
3102                return 0;
3103
3104        if (count < event->read_size)
3105                return -ENOSPC;
3106
3107        WARN_ON_ONCE(event->ctx->parent_ctx);
3108        if (read_format & PERF_FORMAT_GROUP)
3109                ret = perf_event_read_group(event, read_format, buf);
3110        else
3111                ret = perf_event_read_one(event, read_format, buf);
3112
3113        return ret;
3114}
3115
3116static ssize_t
3117perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3118{
3119        struct perf_event *event = file->private_data;
3120
3121        return perf_read_hw(event, buf, count);
3122}
3123
3124static unsigned int perf_poll(struct file *file, poll_table *wait)
3125{
3126        struct perf_event *event = file->private_data;
3127        struct ring_buffer *rb;
3128        unsigned int events = POLL_HUP;
3129
3130        /*
3131         * Race between perf_event_set_output() and perf_poll(): perf_poll()
3132         * grabs the rb reference but perf_event_set_output() overrides it.
3133         * Here is the timeline for two threads T1, T2:
3134         * t0: T1, rb = rcu_dereference(event->rb)
3135         * t1: T2, old_rb = event->rb
3136         * t2: T2, event->rb = new rb
3137         * t3: T2, ring_buffer_detach(old_rb)
3138         * t4: T1, ring_buffer_attach(rb1)
3139         * t5: T1, poll_wait(event->waitq)
3140         *
3141         * To avoid this problem, we grab mmap_mutex in perf_poll()
3142         * thereby ensuring that the assignment of the new ring buffer
3143         * and the detachment of the old buffer appear atomic to perf_poll()
3144         */
3145        mutex_lock(&event->mmap_mutex);
3146
3147        rcu_read_lock();
3148        rb = rcu_dereference(event->rb);
3149        if (rb) {
3150                ring_buffer_attach(event, rb);
3151                events = atomic_xchg(&rb->poll, 0);
3152        }
3153        rcu_read_unlock();
3154
3155        mutex_unlock(&event->mmap_mutex);
3156
3157        poll_wait(file, &event->waitq, wait);
3158
3159        return events;
3160}
3161
3162static void perf_event_reset(struct perf_event *event)
3163{
3164        (void)perf_event_read(event);
3165        local64_set(&event->count, 0);
3166        perf_event_update_userpage(event);
3167}
3168
3169/*
3170 * Holding the top-level event's child_mutex means that any
3171 * descendant process that has inherited this event will block
3172 * in sync_child_event if it goes to exit, thus satisfying the
3173 * task existence requirements of perf_event_enable/disable.
3174 */
3175static void perf_event_for_each_child(struct perf_event *event,
3176                                        void (*func)(struct perf_event *))
3177{
3178        struct perf_event *child;
3179
3180        WARN_ON_ONCE(event->ctx->parent_ctx);
3181        mutex_lock(&event->child_mutex);
3182        func(event);
3183        list_for_each_entry(child, &event->child_list, child_list)
3184                func(child);
3185        mutex_unlock(&event->child_mutex);
3186}
3187
3188static void perf_event_for_each(struct perf_event *event,
3189                                  void (*func)(struct perf_event *))
3190{
3191        struct perf_event_context *ctx = event->ctx;
3192        struct perf_event *sibling;
3193
3194        WARN_ON_ONCE(ctx->parent_ctx);
3195        mutex_lock(&ctx->mutex);
3196        event = event->group_leader;
3197
3198        perf_event_for_each_child(event, func);
3199        list_for_each_entry(sibling, &event->sibling_list, group_entry)
3200                perf_event_for_each_child(sibling, func);
3201        mutex_unlock(&ctx->mutex);
3202}
3203
3204static int perf_event_period(struct perf_event *event, u64 __user *arg)
3205{
3206        struct perf_event_context *ctx = event->ctx;
3207        int ret = 0;
3208        u64 value;
3209
3210        if (!is_sampling_event(event))
3211                return -EINVAL;
3212
3213        if (copy_from_user(&value, arg, sizeof(value)))
3214                return -EFAULT;
3215
3216        if (!value)
3217                return -EINVAL;
3218
3219        raw_spin_lock_irq(&ctx->lock);
3220        if (event->attr.freq) {
3221                if (value > sysctl_perf_event_sample_rate) {
3222                        ret = -EINVAL;
3223                        goto unlock;
3224                }
3225
3226                event->attr.sample_freq = value;
3227        } else {
3228                event->attr.sample_period = value;
3229                event->hw.sample_period = value;
3230        }
3231unlock:
3232        raw_spin_unlock_irq(&ctx->lock);
3233
3234        return ret;
3235}
3236
3237static const struct file_operations perf_fops;
3238
3239static inline int perf_fget_light(int fd, struct fd *p)
3240{
3241        struct fd f = fdget(fd);
3242        if (!f.file)
3243                return -EBADF;
3244
3245        if (f.file->f_op != &perf_fops) {
3246                fdput(f);
3247                return -EBADF;
3248        }
3249        *p = f;
3250        return 0;
3251}
3252
3253static int perf_event_set_output(struct perf_event *event,
3254                                 struct perf_event *output_event);
3255static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3256
3257static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3258{
3259        struct perf_event *event = file->private_data;
3260        void (*func)(struct perf_event *);
3261        u32 flags = arg;
3262
3263        switch (cmd) {
3264        case PERF_EVENT_IOC_ENABLE:
3265                func = perf_event_enable;
3266                break;
3267        case PERF_EVENT_IOC_DISABLE:
3268                func = perf_event_disable;
3269                break;
3270        case PERF_EVENT_IOC_RESET:
3271                func = perf_event_reset;
3272                break;
3273
3274        case PERF_EVENT_IOC_REFRESH:
3275                return perf_event_refresh(event, arg);
3276
3277        case PERF_EVENT_IOC_PERIOD:
3278                return perf_event_period(event, (u64 __user *)arg);
3279
3280        case PERF_EVENT_IOC_SET_OUTPUT:
3281        {
3282                int ret;
3283                if (arg != -1) {
3284                        struct perf_event *output_event;
3285                        struct fd output;
3286                        ret = perf_fget_light(arg, &output);
3287                        if (ret)
3288                                return ret;
3289                        output_event = output.file->private_data;
3290                        ret = perf_event_set_output(event, output_event);
3291                        fdput(output);
3292                } else {
3293                        ret = perf_event_set_output(event, NULL);
3294                }
3295                return ret;
3296        }
3297
3298        case PERF_EVENT_IOC_SET_FILTER:
3299                return perf_event_set_filter(event, (void __user *)arg);
3300
3301        default:
3302                return -ENOTTY;
3303        }
3304
3305        if (flags & PERF_IOC_FLAG_GROUP)
3306                perf_event_for_each(event, func);
3307        else
3308                perf_event_for_each_child(event, func);
3309
3310        return 0;
3311}
3312
3313int perf_event_task_enable(void)
3314{
3315        struct perf_event *event;
3316
3317        mutex_lock(&current->perf_event_mutex);
3318        list_for_each_entry(event, &current->perf_event_list, owner_entry)
3319                perf_event_for_each_child(event, perf_event_enable);
3320        mutex_unlock(&current->perf_event_mutex);
3321
3322        return 0;
3323}
3324
3325int perf_event_task_disable(void)
3326{
3327        struct perf_event *event;
3328
3329        mutex_lock(&current->perf_event_mutex);
3330        list_for_each_entry(event, &current->perf_event_list, owner_entry)
3331                perf_event_for_each_child(event, perf_event_disable);
3332        mutex_unlock(&current->perf_event_mutex);
3333
3334        return 0;
3335}
3336
3337static int perf_event_index(struct perf_event *event)
3338{
3339        if (event->hw.state & PERF_HES_STOPPED)
3340                return 0;
3341
3342        if (event->state != PERF_EVENT_STATE_ACTIVE)
3343                return 0;
3344
3345        return event->pmu->event_idx(event);
3346}
3347
3348static void calc_timer_values(struct perf_event *event,
3349                                u64 *now,
3350                                u64 *enabled,
3351                                u64 *running)
3352{
3353        u64 ctx_time;
3354
3355        *now = perf_clock();
3356        ctx_time = event->shadow_ctx_time + *now;
3357        *enabled = ctx_time - event->tstamp_enabled;
3358        *running = ctx_time - event->tstamp_running;
3359}
3360
3361void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3362{
3363}
3364
3365/*
3366 * Callers need to ensure there can be no nesting of this function, otherwise
3367 * the seqlock logic goes bad. We can not serialize this because the arch
3368 * code calls this from NMI context.
3369 */
3370void perf_event_update_userpage(struct perf_event *event)
3371{
3372        struct perf_event_mmap_page *userpg;
3373        struct ring_buffer *rb;
3374        u64 enabled, running, now;
3375
3376        rcu_read_lock();
3377        /*
3378         * compute total_time_enabled, total_time_running
3379         * based on snapshot values taken when the event
3380         * was last scheduled in.
3381         *
3382         * we cannot simply called update_context_time()
3383         * because of locking issue as we can be called in
3384         * NMI context
3385         */
3386        calc_timer_values(event, &now, &enabled, &running);
3387        rb = rcu_dereference(event->rb);
3388        if (!rb)
3389                goto unlock;
3390
3391        userpg = rb->user_page;
3392
3393        /*
3394         * Disable preemption so as to not let the corresponding user-space
3395         * spin too long if we get preempted.
3396         */
3397        preempt_disable();
3398        ++userpg->lock;
3399        barrier();
3400        userpg->index = perf_event_index(event);
3401        userpg->offset = perf_event_count(event);
3402        if (userpg->index)
3403                userpg->offset -= local64_read(&event->hw.prev_count);
3404
3405        userpg->time_enabled = enabled +
3406                        atomic64_read(&event->child_total_time_enabled);
3407
3408        userpg->time_running = running +
3409                        atomic64_read(&event->child_total_time_running);
3410
3411        arch_perf_update_userpage(userpg, now);
3412
3413        barrier();
3414        ++userpg->lock;
3415        preempt_enable();
3416unlock:
3417        rcu_read_unlock();
3418}
3419
3420static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3421{
3422        struct perf_event *event = vma->vm_file->private_data;
3423        struct ring_buffer *rb;
3424        int ret = VM_FAULT_SIGBUS;
3425
3426        if (vmf->flags & FAULT_FLAG_MKWRITE) {
3427                if (vmf->pgoff == 0)
3428                        ret = 0;
3429                return ret;
3430        }
3431
3432        rcu_read_lock();
3433        rb = rcu_dereference(event->rb);
3434        if (!rb)
3435                goto unlock;
3436
3437        if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3438                goto unlock;
3439
3440        vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3441        if (!vmf->page)
3442                goto unlock;
3443
3444        get_page(vmf->page);
3445        vmf->page->mapping = vma->vm_file->f_mapping;
3446        vmf->page->index   = vmf->pgoff;
3447
3448        ret = 0;
3449unlock:
3450        rcu_read_unlock();
3451
3452        return ret;
3453}
3454
3455static void ring_buffer_attach(struct perf_event *event,
3456                               struct ring_buffer *rb)
3457{
3458        unsigned long flags;
3459
3460        if (!list_empty(&event->rb_entry))
3461                return;
3462
3463        spin_lock_irqsave(&rb->event_lock, flags);
3464        if (!list_empty(&event->rb_entry))
3465                goto unlock;
3466
3467        list_add(&event->rb_entry, &rb->event_list);
3468unlock:
3469        spin_unlock_irqrestore(&rb->event_lock, flags);
3470}
3471
3472static void ring_buffer_detach(struct perf_event *event,
3473                               struct ring_buffer *rb)
3474{
3475        unsigned long flags;
3476
3477        if (list_empty(&event->rb_entry))
3478                return;
3479
3480        spin_lock_irqsave(&rb->event_lock, flags);
3481        list_del_init(&event->rb_entry);
3482        wake_up_all(&event->waitq);
3483        spin_unlock_irqrestore(&rb->event_lock, flags);
3484}
3485
3486static void ring_buffer_wakeup(struct perf_event *event)
3487{
3488        struct ring_buffer *rb;
3489
3490        rcu_read_lock();
3491        rb = rcu_dereference(event->rb);
3492        if (!rb)
3493                goto unlock;
3494
3495        list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3496                wake_up_all(&event->waitq);
3497
3498unlock:
3499        rcu_read_unlock();
3500}
3501
3502static void rb_free_rcu(struct rcu_head *rcu_head)
3503{
3504        struct ring_buffer *rb;
3505
3506        rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3507        rb_free(rb);
3508}
3509
3510static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3511{
3512        struct ring_buffer *rb;
3513
3514        rcu_read_lock();
3515        rb = rcu_dereference(event->rb);
3516        if (rb) {
3517                if (!atomic_inc_not_zero(&rb->refcount))
3518                        rb = NULL;
3519        }
3520        rcu_read_unlock();
3521
3522        return rb;
3523}
3524
3525static void ring_buffer_put(struct ring_buffer *rb)
3526{
3527        struct perf_event *event, *n;
3528        unsigned long flags;
3529
3530        if (!atomic_dec_and_test(&rb->refcount))
3531                return;
3532
3533        spin_lock_irqsave(&rb->event_lock, flags);
3534        list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3535                list_del_init(&event->rb_entry);
3536                wake_up_all(&event->waitq);
3537        }
3538        spin_unlock_irqrestore(&rb->event_lock, flags);
3539
3540        call_rcu(&rb->rcu_head, rb_free_rcu);
3541}
3542
3543static void perf_mmap_open(struct vm_area_struct *vma)
3544{
3545        struct perf_event *event = vma->vm_file->private_data;
3546
3547        atomic_inc(&event->mmap_count);
3548}
3549
3550static void perf_mmap_close(struct vm_area_struct *vma)
3551{
3552        struct perf_event *event = vma->vm_file->private_data;
3553
3554        if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3555                unsigned long size = perf_data_size(event->rb);
3556                struct user_struct *user = event->mmap_user;
3557                struct ring_buffer *rb = event->rb;
3558
3559                atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3560                vma->vm_mm->pinned_vm -= event->mmap_locked;
3561                rcu_assign_pointer(event->rb, NULL);
3562                ring_buffer_detach(event, rb);
3563                mutex_unlock(&event->mmap_mutex);
3564
3565                ring_buffer_put(rb);
3566                free_uid(user);
3567        }
3568}
3569
3570static const struct vm_operations_struct perf_mmap_vmops = {
3571        .open           = perf_mmap_open,
3572        .close          = perf_mmap_close,
3573        .fault          = perf_mmap_fault,
3574        .page_mkwrite   = perf_mmap_fault,
3575};
3576
3577static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3578{
3579        struct perf_event *event = file->private_data;
3580        unsigned long user_locked, user_lock_limit;
3581        struct user_struct *user = current_user();
3582        unsigned long locked, lock_limit;
3583        struct ring_buffer *rb;
3584        unsigned long vma_size;
3585        unsigned long nr_pages;
3586        long user_extra, extra;
3587        int ret = 0, flags = 0;
3588
3589        /*
3590         * Don't allow mmap() of inherited per-task counters. This would
3591         * create a performance issue due to all children writing to the
3592         * same rb.
3593         */
3594        if (event->cpu == -1 && event->attr.inherit)
3595                return -EINVAL;
3596
3597        if (!(vma->vm_flags & VM_SHARED))
3598                return -EINVAL;
3599
3600        vma_size = vma->vm_end - vma->vm_start;
3601        nr_pages = (vma_size / PAGE_SIZE) - 1;
3602
3603        /*
3604         * If we have rb pages ensure they're a power-of-two number, so we
3605         * can do bitmasks instead of modulo.
3606         */
3607        if (nr_pages != 0 && !is_power_of_2(nr_pages))
3608                return -EINVAL;
3609
3610        if (vma_size != PAGE_SIZE * (1 + nr_pages))
3611                return -EINVAL;
3612
3613        if (vma->vm_pgoff != 0)
3614                return -EINVAL;
3615
3616        WARN_ON_ONCE(event->ctx->parent_ctx);
3617        mutex_lock(&event->mmap_mutex);
3618        if (event->rb) {
3619                if (event->rb->nr_pages == nr_pages)
3620                        atomic_inc(&event->rb->refcount);
3621                else
3622                        ret = -EINVAL;
3623                goto unlock;
3624        }
3625
3626        user_extra = nr_pages + 1;
3627        user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3628
3629        /*
3630         * Increase the limit linearly with more CPUs:
3631         */
3632        user_lock_limit *= num_online_cpus();
3633
3634        user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3635
3636        extra = 0;
3637        if (user_locked > user_lock_limit)
3638                extra = user_locked - user_lock_limit;
3639
3640        lock_limit = rlimit(RLIMIT_MEMLOCK);
3641        lock_limit >>= PAGE_SHIFT;
3642        locked = vma->vm_mm->pinned_vm + extra;
3643
3644        if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3645                !capable(CAP_IPC_LOCK)) {
3646                ret = -EPERM;
3647                goto unlock;
3648        }
3649
3650        WARN_ON(event->rb);
3651
3652        if (vma->vm_flags & VM_WRITE)
3653                flags |= RING_BUFFER_WRITABLE;
3654
3655        rb = rb_alloc(nr_pages, 
3656                event->attr.watermark ? event->attr.wakeup_watermark : 0,
3657                event->cpu, flags);
3658
3659        if (!rb) {
3660                ret = -ENOMEM;
3661                goto unlock;
3662        }
3663        rcu_assign_pointer(event->rb, rb);
3664
3665        atomic_long_add(user_extra, &user->locked_vm);
3666        event->mmap_locked = extra;
3667        event->mmap_user = get_current_user();
3668        vma->vm_mm->pinned_vm += event->mmap_locked;
3669
3670        perf_event_update_userpage(event);
3671
3672unlock:
3673        if (!ret)
3674                atomic_inc(&event->mmap_count);
3675        mutex_unlock(&event->mmap_mutex);
3676
3677        vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3678        vma->vm_ops = &perf_mmap_vmops;
3679
3680        return ret;
3681}
3682
3683static int perf_fasync(int fd, struct file *filp, int on)
3684{
3685        struct inode *inode = filp->f_path.dentry->d_inode;
3686        struct perf_event *event = filp->private_data;
3687        int retval;
3688
3689        mutex_lock(&inode->i_mutex);
3690        retval = fasync_helper(fd, filp, on, &event->fasync);
3691        mutex_unlock(&inode->i_mutex);
3692
3693        if (retval < 0)
3694                return retval;
3695
3696        return 0;
3697}
3698
3699static const struct file_operations perf_fops = {
3700        .llseek                 = no_llseek,
3701        .release                = perf_release,
3702        .read                   = perf_read,
3703        .poll                   = perf_poll,
3704        .unlocked_ioctl         = perf_ioctl,
3705        .compat_ioctl           = perf_ioctl,
3706        .mmap                   = perf_mmap,
3707        .fasync                 = perf_fasync,
3708};
3709
3710/*
3711 * Perf event wakeup
3712 *
3713 * If there's data, ensure we set the poll() state and publish everything
3714 * to user-space before waking everybody up.
3715 */
3716
3717void perf_event_wakeup(struct perf_event *event)
3718{
3719        ring_buffer_wakeup(event);
3720
3721        if (event->pending_kill) {
3722                kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3723                event->pending_kill = 0;
3724        }
3725}
3726
3727static void perf_pending_event(struct irq_work *entry)
3728{
3729        struct perf_event *event = container_of(entry,
3730                        struct perf_event, pending);
3731
3732        if (event->pending_disable) {
3733                event->pending_disable = 0;
3734                __perf_event_disable(event);
3735        }
3736
3737        if (event->pending_wakeup) {
3738                event->pending_wakeup = 0;
3739                perf_event_wakeup(event);
3740        }
3741}
3742
3743/*
3744 * We assume there is only KVM supporting the callbacks.
3745 * Later on, we might change it to a list if there is
3746 * another virtualization implementation supporting the callbacks.
3747 */
3748struct perf_guest_info_callbacks *perf_guest_cbs;
3749
3750int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3751{
3752        perf_guest_cbs = cbs;
3753        return 0;
3754}
3755EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3756
3757int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3758{
3759        perf_guest_cbs = NULL;
3760        return 0;
3761}
3762EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3763
3764static void
3765perf_output_sample_regs(struct perf_output_handle *handle,
3766                        struct pt_regs *regs, u64 mask)
3767{
3768        int bit;
3769
3770        for_each_set_bit(bit, (const unsigned long *) &mask,
3771                         sizeof(mask) * BITS_PER_BYTE) {
3772                u64 val;
3773
3774                val = perf_reg_value(regs, bit);
3775                perf_output_put(handle, val);
3776        }
3777}
3778
3779static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3780                                  struct pt_regs *regs)
3781{
3782        if (!user_mode(regs)) {
3783                if (current->mm)
3784                        regs = task_pt_regs(current);
3785                else
3786                        regs = NULL;
3787        }
3788
3789        if (regs) {
3790                regs_user->regs = regs;
3791                regs_user->abi  = perf_reg_abi(current);
3792        }
3793}
3794
3795/*
3796 * Get remaining task size from user stack pointer.
3797 *
3798 * It'd be better to take stack vma map and limit this more
3799 * precisly, but there's no way to get it safely under interrupt,
3800 * so using TASK_SIZE as limit.
3801 */
3802static u64 perf_ustack_task_size(struct pt_regs *regs)
3803{
3804        unsigned long addr = perf_user_stack_pointer(regs);
3805
3806        if (!addr || addr >= TASK_SIZE)
3807                return 0;
3808
3809        return TASK_SIZE - addr;
3810}
3811
3812static u16
3813perf_sample_ustack_size(u16 stack_size, u16 header_size,
3814                        struct pt_regs *regs)
3815{
3816        u64 task_size;
3817
3818        /* No regs, no stack pointer, no dump. */
3819        if (!regs)
3820                return 0;
3821
3822        /*
3823         * Check if we fit in with the requested stack size into the:
3824         * - TASK_SIZE
3825         *   If we don't, we limit the size to the TASK_SIZE.
3826         *
3827         * - remaining sample size
3828         *   If we don't, we customize the stack size to
3829         *   fit in to the remaining sample size.
3830         */
3831
3832        task_size  = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3833        stack_size = min(stack_size, (u16) task_size);
3834
3835        /* Current header size plus static size and dynamic size. */
3836        header_size += 2 * sizeof(u64);
3837
3838        /* Do we fit in with the current stack dump size? */
3839        if ((u16) (header_size + stack_size) < header_size) {
3840                /*
3841                 * If we overflow the maximum size for the sample,
3842                 * we customize the stack dump size to fit in.
3843                 */
3844                stack_size = USHRT_MAX - header_size - sizeof(u64);
3845                stack_size = round_up(stack_size, sizeof(u64));
3846        }
3847
3848        return stack_size;
3849}
3850
3851static void
3852perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
3853                          struct pt_regs *regs)
3854{
3855        /* Case of a kernel thread, nothing to dump */
3856        if (!regs) {
3857                u64 size = 0;
3858                perf_output_put(handle, size);
3859        } else {
3860                unsigned long sp;
3861                unsigned int rem;
3862                u64 dyn_size;
3863
3864                /*
3865                 * We dump:
3866                 * static size
3867                 *   - the size requested by user or the best one we can fit
3868                 *     in to the sample max size
3869                 * data
3870                 *   - user stack dump data
3871                 * dynamic size
3872                 *   - the actual dumped size
3873                 */
3874
3875                /* Static size. */
3876                perf_output_put(handle, dump_size);
3877
3878                /* Data. */
3879                sp = perf_user_stack_pointer(regs);
3880                rem = __output_copy_user(handle, (void *) sp, dump_size);
3881                dyn_size = dump_size - rem;
3882
3883                perf_output_skip(handle, rem);
3884
3885                /* Dynamic size. */
3886                perf_output_put(handle, dyn_size);
3887        }
3888}
3889
3890static void __perf_event_header__init_id(struct perf_event_header *header,
3891                                         struct perf_sample_data *data,
3892                                         struct perf_event *event)
3893{
3894        u64 sample_type = event->attr.sample_type;
3895
3896        data->type = sample_type;
3897        header->size += event->id_header_size;
3898
3899        if (sample_type & PERF_SAMPLE_TID) {
3900                /* namespace issues */
3901                data->tid_entry.pid = perf_event_pid(event, current);
3902                data->tid_entry.tid = perf_event_tid(event, current);
3903        }
3904
3905        if (sample_type & PERF_SAMPLE_TIME)
3906                data->time = perf_clock();
3907
3908        if (sample_type & PERF_SAMPLE_ID)
3909                data->id = primary_event_id(event);
3910
3911        if (sample_type & PERF_SAMPLE_STREAM_ID)
3912                data->stream_id = event->id;
3913
3914        if (sample_type & PERF_SAMPLE_CPU) {
3915                data->cpu_entry.cpu      = raw_smp_processor_id();
3916                data->cpu_entry.reserved = 0;
3917        }
3918}
3919
3920void perf_event_header__init_id(struct perf_event_header *header,
3921                                struct perf_sample_data *data,
3922                                struct perf_event *event)
3923{
3924        if (event->attr.sample_id_all)
3925                __perf_event_header__init_id(header, data, event);
3926}
3927
3928static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3929                                           struct perf_sample_data *data)
3930{
3931        u64 sample_type = data->type;
3932
3933        if (sample_type & PERF_SAMPLE_TID)
3934                perf_output_put(handle, data->tid_entry);
3935
3936        if (sample_type & PERF_SAMPLE_TIME)
3937                perf_output_put(handle, data->time);
3938
3939        if (sample_type & PERF_SAMPLE_ID)
3940                perf_output_put(handle, data->id);
3941
3942        if (sample_type & PERF_SAMPLE_STREAM_ID)
3943                perf_output_put(handle, data->stream_id);
3944
3945        if (sample_type & PERF_SAMPLE_CPU)
3946                perf_output_put(handle, data->cpu_entry);
3947}
3948
3949void perf_event__output_id_sample(struct perf_event *event,
3950                                  struct perf_output_handle *handle,
3951                                  struct perf_sample_data *sample)
3952{
3953        if (event->attr.sample_id_all)
3954                __perf_event__output_id_sample(handle, sample);
3955}
3956
3957static void perf_output_read_one(struct perf_output_handle *handle,
3958                                 struct perf_event *event,
3959                                 u64 enabled, u64 running)
3960{
3961        u64 read_format = event->attr.read_format;
3962        u64 values[4];
3963        int n = 0;
3964
3965        values[n++] = perf_event_count(event);
3966        if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3967                values[n++] = enabled +
3968                        atomic64_read(&event->child_total_time_enabled);
3969        }
3970        if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3971                values[n++] = running +
3972                        atomic64_read(&event->child_total_time_running);
3973        }
3974        if (read_format & PERF_FORMAT_ID)
3975                values[n++] = primary_event_id(event);
3976
3977        __output_copy(handle, values, n * sizeof(u64));
3978}
3979
3980/*
3981 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3982 */
3983static void perf_output_read_group(struct perf_output_handle *handle,
3984                            struct perf_event *event,
3985                            u64 enabled, u64 running)
3986{
3987        struct perf_event *leader = event->group_leader, *sub;
3988        u64 read_format = event->attr.read_format;
3989        u64 values[5];
3990        int n = 0;
3991
3992        values[n++] = 1 + leader->nr_siblings;
3993
3994        if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3995                values[n++] = enabled;
3996
3997        if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3998                values[n++] = running;
3999
4000        if (leader != event)
4001                leader->pmu->read(leader);
4002
4003        values[n++] = perf_event_count(leader);
4004        if (read_format & PERF_FORMAT_ID)
4005                values[n++] = primary_event_id(leader);
4006
4007        __output_copy(handle, values, n * sizeof(u64));
4008
4009        list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4010                n = 0;
4011
4012                if (sub != event)
4013                        sub->pmu->read(sub);
4014
4015                values[n++] = perf_event_count(sub);
4016                if (read_format & PERF_FORMAT_ID)
4017                        values[n++] = primary_event_id(sub);
4018
4019                __output_copy(handle, values, n * sizeof(u64));
4020        }
4021}
4022
4023#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4024                                 PERF_FORMAT_TOTAL_TIME_RUNNING)
4025
4026static void perf_output_read(struct perf_output_handle *handle,
4027                             struct perf_event *event)
4028{
4029        u64 enabled = 0, running = 0, now;
4030        u64 read_format = event->attr.read_format;
4031
4032        /*
4033         * compute total_time_enabled, total_time_running
4034         * based on snapshot values taken when the event
4035         * was last scheduled in.
4036         *
4037         * we cannot simply called update_context_time()
4038         * because of locking issue as we are called in
4039         * NMI context
4040         */
4041        if (read_format & PERF_FORMAT_TOTAL_TIMES)
4042                calc_timer_values(event, &now, &enabled, &running);
4043
4044        if (event->attr.read_format & PERF_FORMAT_GROUP)
4045                perf_output_read_group(handle, event, enabled, running);
4046        else
4047                perf_output_read_one(handle, event, enabled, running);
4048}
4049
4050void perf_output_sample(struct perf_output_handle *handle,
4051                        struct perf_event_header *header,
4052                        struct perf_sample_data *data,
4053                        struct perf_event *event)
4054{
4055        u64 sample_type = data->type;
4056
4057        perf_output_put(handle, *header);
4058
4059        if (sample_type & PERF_SAMPLE_IP)
4060                perf_output_put(handle, data->ip);
4061
4062        if (sample_type & PERF_SAMPLE_TID)
4063                perf_output_put(handle, data->tid_entry);
4064
4065        if (sample_type & PERF_SAMPLE_TIME)
4066                perf_output_put(handle, data->time);
4067
4068        if (sample_type & PERF_SAMPLE_ADDR)
4069                perf_output_put(handle, data->addr);
4070
4071        if (sample_type & PERF_SAMPLE_ID)
4072                perf_output_put(handle, data->id);
4073
4074        if (sample_type & PERF_SAMPLE_STREAM_ID)
4075                perf_output_put(handle, data->stream_id);
4076
4077        if (sample_type & PERF_SAMPLE_CPU)
4078                perf_output_put(handle, data->cpu_entry);
4079
4080        if (sample_type & PERF_SAMPLE_PERIOD)
4081                perf_output_put(handle, data->period);
4082
4083        if (sample_type & PERF_SAMPLE_READ)
4084                perf_output_read(handle, event);
4085
4086        if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4087                if (data->callchain) {
4088                        int size = 1;
4089
4090                        if (data->callchain)
4091                                size += data->callchain->nr;
4092
4093                        size *= sizeof(u64);
4094
4095                        __output_copy(handle, data->callchain, size);
4096                } else {
4097                        u64 nr = 0;
4098                        perf_output_put(handle, nr);
4099                }
4100        }
4101
4102        if (sample_type & PERF_SAMPLE_RAW) {
4103                if (data->raw) {
4104                        perf_output_put(handle, data->raw->size);
4105                        __output_copy(handle, data->raw->data,
4106                                           data->raw->size);
4107                } else {
4108                        struct {
4109                                u32     size;
4110                                u32     data;
4111                        } raw = {
4112                                .size = sizeof(u32),
4113                                .data = 0,
4114                        };
4115                        perf_output_put(handle, raw);
4116                }
4117        }
4118
4119        if (!event->attr.watermark) {
4120                int wakeup_events = event->attr.wakeup_events;
4121
4122                if (wakeup_events) {
4123                        struct ring_buffer *rb = handle->rb;
4124                        int events = local_inc_return(&rb->events);
4125
4126                        if (events >= wakeup_events) {
4127                                local_sub(wakeup_events, &rb->events);
4128                                local_inc(&rb->wakeup);
4129                        }
4130                }
4131        }
4132
4133        if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4134                if (data->br_stack) {
4135                        size_t size;
4136
4137                        size = data->br_stack->nr
4138                             * sizeof(struct perf_branch_entry);
4139
4140                        perf_output_put(handle, data->br_stack->nr);
4141                        perf_output_copy(handle, data->br_stack->entries, size);
4142                } else {
4143                        /*
4144                         * we always store at least the value of nr
4145                         */
4146                        u64 nr = 0;
4147                        perf_output_put(handle, nr);
4148                }
4149        }
4150
4151        if (sample_type & PERF_SAMPLE_REGS_USER) {
4152                u64 abi = data->regs_user.abi;
4153
4154                /*
4155                 * If there are no regs to dump, notice it through
4156                 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4157                 */
4158                perf_output_put(handle, abi);
4159
4160                if (abi) {
4161                        u64 mask = event->attr.sample_regs_user;
4162                        perf_output_sample_regs(handle,
4163                                                data->regs_user.regs,
4164                                                mask);
4165                }
4166        }
4167
4168        if (sample_type & PERF_SAMPLE_STACK_USER)
4169                perf_output_sample_ustack(handle,
4170                                          data->stack_user_size,
4171                                          data->regs_user.regs);
4172}
4173
4174void perf_prepare_sample(struct perf_event_header *header,
4175                         struct perf_sample_data *data,
4176                         struct perf_event *event,
4177                         struct pt_regs *regs)
4178{
4179        u64 sample_type = event->attr.sample_type;
4180
4181        header->type = PERF_RECORD_SAMPLE;
4182        header->size = sizeof(*header) + event->header_size;
4183
4184        header->misc = 0;
4185        header->misc |= perf_misc_flags(regs);
4186
4187        __perf_event_header__init_id(header, data, event);
4188
4189        if (sample_type & PERF_SAMPLE_IP)
4190                data->ip = perf_instruction_pointer(regs);
4191
4192        if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4193                int size = 1;
4194
4195                data->callchain = perf_callchain(event, regs);
4196
4197                if (data->callchain)
4198                        size += data->callchain->nr;
4199
4200                header->size += size * sizeof(u64);
4201        }
4202
4203        if (sample_type & PERF_SAMPLE_RAW) {
4204                int size = sizeof(u32);
4205
4206                if (data->raw)
4207                        size += data->raw->size;
4208                else
4209                        size += sizeof(u32);
4210
4211                WARN_ON_ONCE(size & (sizeof(u64)-1));
4212                header->size += size;
4213        }
4214
4215        if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4216                int size = sizeof(u64); /* nr */
4217                if (data->br_stack) {
4218                        size += data->br_stack->nr
4219                              * sizeof(struct perf_branch_entry);
4220                }
4221                header->size += size;
4222        }
4223
4224        if (sample_type & PERF_SAMPLE_REGS_USER) {
4225                /* regs dump ABI info */
4226                int size = sizeof(u64);
4227
4228                perf_sample_regs_user(&data->regs_user, regs);
4229
4230                if (data->regs_user.regs) {
4231                        u64 mask = event->attr.sample_regs_user;
4232                        size += hweight64(mask) * sizeof(u64);
4233                }
4234
4235                header->size += size;
4236        }
4237
4238        if (sample_type & PERF_SAMPLE_STACK_USER) {
4239                /*
4240                 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4241                 * processed as the last one or have additional check added
4242                 * in case new sample type is added, because we could eat
4243                 * up the rest of the sample size.
4244                 */
4245                struct perf_regs_user *uregs = &data->regs_user;
4246                u16 stack_size = event->attr.sample_stack_user;
4247                u16 size = sizeof(u64);
4248
4249                if (!uregs->abi)
4250                        perf_sample_regs_user(uregs, regs);
4251
4252                stack_size = perf_sample_ustack_size(stack_size, header->size,
4253                                                     uregs->regs);
4254
4255                /*
4256                 * If there is something to dump, add space for the dump
4257                 * itself and for the field that tells the dynamic size,
4258                 * which is how many have been actually dumped.
4259                 */
4260                if (stack_size)
4261                        size += sizeof(u64) + stack_size;
4262
4263                data->stack_user_size = stack_size;
4264                header->size += size;
4265        }
4266}
4267
4268static void perf_event_output(struct perf_event *event,
4269                                struct perf_sample_data *data,
4270                                struct pt_regs *regs)
4271{
4272        struct perf_output_handle handle;
4273        struct perf_event_header header;
4274
4275        /* protect the callchain buffers */
4276        rcu_read_lock();
4277
4278        perf_prepare_sample(&header, data, event, regs);
4279
4280        if (perf_output_begin(&handle, event, header.size))
4281                goto exit;
4282
4283        perf_output_sample(&handle, &header, data, event);
4284
4285        perf_output_end(&handle);
4286
4287exit:
4288        rcu_read_unlock();
4289}
4290
4291/*
4292 * read event_id
4293 */
4294
4295struct perf_read_event {
4296        struct perf_event_header        header;
4297
4298        u32                             pid;
4299        u32                             tid;
4300};
4301
4302static void
4303perf_event_read_event(struct perf_event *event,
4304                        struct task_struct *task)
4305{
4306        struct perf_output_handle handle;
4307        struct perf_sample_data sample;
4308        struct perf_read_event read_event = {
4309                .header = {
4310                        .type = PERF_RECORD_READ,
4311                        .misc = 0,
4312                        .size = sizeof(read_event) + event->read_size,
4313                },
4314                .pid = perf_event_pid(event, task),
4315                .tid = perf_event_tid(event, task),
4316        };
4317        int ret;
4318
4319        perf_event_header__init_id(&read_event.header, &sample, event);
4320        ret = perf_output_begin(&handle, event, read_event.header.size);
4321        if (ret)
4322                return;
4323
4324        perf_output_put(&handle, read_event);
4325        perf_output_read(&handle, event);
4326        perf_event__output_id_sample(event, &handle, &sample);
4327
4328        perf_output_end(&handle);
4329}
4330
4331/*
4332 * task tracking -- fork/exit
4333 *
4334 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4335 */
4336
4337struct perf_task_event {
4338        struct task_struct              *task;
4339        struct perf_event_context       *task_ctx;
4340
4341        struct {
4342                struct perf_event_header        header;
4343
4344                u32                             pid;
4345                u32                             ppid;
4346                u32                             tid;
4347                u32                             ptid;
4348                u64                             time;
4349        } event_id;
4350};
4351
4352static void perf_event_task_output(struct perf_event *event,
4353                                     struct perf_task_event *task_event)
4354{
4355        struct perf_output_handle handle;
4356        struct perf_sample_data sample;
4357        struct task_struct *task = task_event->task;
4358        int ret, size = task_event->event_id.header.size;
4359
4360        perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4361
4362        ret = perf_output_begin(&handle, event,
4363                                task_event->event_id.header.size);
4364        if (ret)
4365                goto out;
4366
4367        task_event->event_id.pid = perf_event_pid(event, task);
4368        task_event->event_id.ppid = perf_event_pid(event, current);
4369
4370        task_event->event_id.tid = perf_event_tid(event, task);
4371        task_event->event_id.ptid = perf_event_tid(event, current);
4372
4373        perf_output_put(&handle, task_event->event_id);
4374
4375        perf_event__output_id_sample(event, &handle, &sample);
4376
4377        perf_output_end(&handle);
4378out:
4379        task_event->event_id.header.size = size;
4380}
4381
4382static int perf_event_task_match(struct perf_event *event)
4383{
4384        if (event->state < PERF_EVENT_STATE_INACTIVE)
4385                return 0;
4386
4387        if (!event_filter_match(event))
4388                return 0;
4389
4390        if (event->attr.comm || event->attr.mmap ||
4391            event->attr.mmap_data || event->attr.task)
4392                return 1;
4393
4394        return 0;
4395}
4396
4397static void perf_event_task_ctx(struct perf_event_context *ctx,
4398                                  struct perf_task_event *task_event)
4399{
4400        struct perf_event *event;
4401
4402        list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4403                if (perf_event_task_match(event))
4404                        perf_event_task_output(event, task_event);
4405        }
4406}
4407
4408static void perf_event_task_event(struct perf_task_event *task_event)
4409{
4410        struct perf_cpu_context *cpuctx;
4411        struct perf_event_context *ctx;
4412        struct pmu *pmu;
4413        int ctxn;
4414
4415        rcu_read_lock();
4416        list_for_each_entry_rcu(pmu, &pmus, entry) {
4417                cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4418                if (cpuctx->unique_pmu != pmu)
4419                        goto next;
4420                perf_event_task_ctx(&cpuctx->ctx, task_event);
4421
4422                ctx = task_event->task_ctx;
4423                if (!ctx) {
4424                        ctxn = pmu->task_ctx_nr;
4425                        if (ctxn < 0)
4426                                goto next;
4427                        ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4428                }
4429                if (ctx)
4430                        perf_event_task_ctx(ctx, task_event);
4431next:
4432                put_cpu_ptr(pmu->pmu_cpu_context);
4433        }
4434        rcu_read_unlock();
4435}
4436
4437static void perf_event_task(struct task_struct *task,
4438                              struct perf_event_context *task_ctx,
4439                              int new)
4440{
4441        struct perf_task_event task_event;
4442
4443        if (!atomic_read(&nr_comm_events) &&
4444            !atomic_read(&nr_mmap_events) &&
4445            !atomic_read(&nr_task_events))
4446                return;
4447
4448        task_event = (struct perf_task_event){
4449                .task     = task,
4450                .task_ctx = task_ctx,
4451                .event_id    = {
4452                        .header = {
4453                                .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4454                                .misc = 0,
4455                                .size = sizeof(task_event.event_id),
4456                        },
4457                        /* .pid  */
4458                        /* .ppid */
4459                        /* .tid  */
4460                        /* .ptid */
4461                        .time = perf_clock(),
4462                },
4463        };
4464
4465        perf_event_task_event(&task_event);
4466}
4467
4468void perf_event_fork(struct task_struct *task)
4469{
4470        perf_event_task(task, NULL, 1);
4471}
4472
4473/*
4474 * comm tracking
4475 */
4476
4477struct perf_comm_event {
4478        struct task_struct      *task;
4479        char                    *comm;
4480        int                     comm_size;
4481
4482        struct {
4483                struct perf_event_header        header;
4484
4485                u32                             pid;
4486                u32                             tid;
4487        } event_id;
4488};
4489
4490static void perf_event_comm_output(struct perf_event *event,
4491                                     struct perf_comm_event *comm_event)
4492{
4493        struct perf_output_handle handle;
4494        struct perf_sample_data sample;
4495        int size = comm_event->event_id.header.size;
4496        int ret;
4497
4498        perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4499        ret = perf_output_begin(&handle, event,
4500                                comm_event->event_id.header.size);
4501
4502        if (ret)
4503                goto out;
4504
4505        comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4506        comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4507
4508        perf_output_put(&handle, comm_event->event_id);
4509        __output_copy(&handle, comm_event->comm,
4510                                   comm_event->comm_size);
4511
4512        perf_event__output_id_sample(event, &handle, &sample);
4513
4514        perf_output_end(&handle);
4515out:
4516        comm_event->event_id.header.size = size;
4517}
4518
4519static int perf_event_comm_match(struct perf_event *event)
4520{
4521        if (event->state < PERF_EVENT_STATE_INACTIVE)
4522                return 0;
4523
4524        if (!event_filter_match(event))
4525                return 0;
4526
4527        if (event->attr.comm)
4528                return 1;
4529
4530        return 0;
4531}
4532
4533static void perf_event_comm_ctx(struct perf_event_context *ctx,
4534                                  struct perf_comm_event *comm_event)
4535{
4536        struct perf_event *event;
4537
4538        list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4539                if (perf_event_comm_match(event))
4540                        perf_event_comm_output(event, comm_event);
4541        }
4542}
4543
4544static void perf_event_comm_event(struct perf_comm_event *comm_event)
4545{
4546        struct perf_cpu_context *cpuctx;
4547        struct perf_event_context *ctx;
4548        char comm[TASK_COMM_LEN];
4549        unsigned int size;
4550        struct pmu *pmu;
4551        int ctxn;
4552
4553        memset(comm, 0, sizeof(comm));
4554        strlcpy(comm, comm_event->task->comm, sizeof(comm));
4555        size = ALIGN(strlen(comm)+1, sizeof(u64));
4556
4557        comm_event->comm = comm;
4558        comm_event->comm_size = size;
4559
4560        comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4561        rcu_read_lock();
4562        list_for_each_entry_rcu(pmu, &pmus, entry) {
4563                cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4564                if (cpuctx->unique_pmu != pmu)
4565                        goto next;
4566                perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4567
4568                ctxn = pmu->task_ctx_nr;
4569                if (ctxn < 0)
4570                        goto next;
4571
4572                ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4573                if (ctx)
4574                        perf_event_comm_ctx(ctx, comm_event);
4575next:
4576                put_cpu_ptr(pmu->pmu_cpu_context);
4577        }
4578        rcu_read_unlock();
4579}
4580
4581void perf_event_comm(struct task_struct *task)
4582{
4583        struct perf_comm_event comm_event;
4584        struct perf_event_context *ctx;
4585        int ctxn;
4586
4587        for_each_task_context_nr(ctxn) {
4588                ctx = task->perf_event_ctxp[ctxn];
4589                if (!ctx)
4590                        continue;
4591
4592                perf_event_enable_on_exec(ctx);
4593        }
4594
4595        if (!atomic_read(&nr_comm_events))
4596                return;
4597
4598        comm_event = (struct perf_comm_event){
4599                .task   = task,
4600                /* .comm      */
4601                /* .comm_size */
4602                .event_id  = {
4603                        .header = {
4604                                .type = PERF_RECORD_COMM,
4605                                .misc = 0,
4606                                /* .size */
4607                        },
4608                        /* .pid */
4609                        /* .tid */
4610                },
4611        };
4612
4613        perf_event_comm_event(&comm_event);
4614}
4615
4616/*
4617 * mmap tracking
4618 */
4619
4620struct perf_mmap_event {
4621        struct vm_area_struct   *vma;
4622
4623        const char              *file_name;
4624        int                     file_size;
4625
4626        struct {
4627                struct perf_event_header        header;
4628
4629                u32                             pid;
4630                u32                             tid;
4631                u64                             start;
4632                u64                             len;
4633                u64                             pgoff;
4634        } event_id;
4635};
4636
4637static void perf_event_mmap_output(struct perf_event *event,
4638                                     struct perf_mmap_event *mmap_event)
4639{
4640        struct perf_output_handle handle;
4641        struct perf_sample_data sample;
4642        int size = mmap_event->event_id.header.size;
4643        int ret;
4644
4645        perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4646        ret = perf_output_begin(&handle, event,
4647                                mmap_event->event_id.header.size);
4648        if (ret)
4649                goto out;
4650
4651        mmap_event->event_id.pid = perf_event_pid(event, current);
4652        mmap_event->event_id.tid = perf_event_tid(event, current);
4653
4654        perf_output_put(&handle, mmap_event->event_id);
4655        __output_copy(&handle, mmap_event->file_name,
4656                                   mmap_event->file_size);
4657
4658        perf_event__output_id_sample(event, &handle, &sample);
4659
4660        perf_output_end(&handle);
4661out:
4662        mmap_event->event_id.header.size = size;
4663}
4664
4665static int perf_event_mmap_match(struct perf_event *event,
4666                                   struct perf_mmap_event *mmap_event,
4667                                   int executable)
4668{
4669        if (event->state < PERF_EVENT_STATE_INACTIVE)
4670                return 0;
4671
4672        if (!event_filter_match(event))
4673                return 0;
4674
4675        if ((!executable && event->attr.mmap_data) ||
4676            (executable && event->attr.mmap))
4677                return 1;
4678
4679        return 0;
4680}
4681
4682static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4683                                  struct perf_mmap_event *mmap_event,
4684                                  int executable)
4685{
4686        struct perf_event *event;
4687
4688        list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4689                if (perf_event_mmap_match(event, mmap_event, executable))
4690                        perf_event_mmap_output(event, mmap_event);
4691        }
4692}
4693
4694static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4695{
4696        struct perf_cpu_context *cpuctx;
4697        struct perf_event_context *ctx;
4698        struct vm_area_struct *vma = mmap_event->vma;
4699        struct file *file = vma->vm_file;
4700        unsigned int size;
4701        char tmp[16];
4702        char *buf = NULL;
4703        const char *name;
4704        struct pmu *pmu;
4705        int ctxn;
4706
4707        memset(tmp, 0, sizeof(tmp));
4708
4709        if (file) {
4710                /*
4711                 * d_path works from the end of the rb backwards, so we
4712                 * need to add enough zero bytes after the string to handle
4713                 * the 64bit alignment we do later.
4714                 */
4715                buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4716                if (!buf) {
4717                        name = strncpy(tmp, "//enomem", sizeof(tmp));
4718                        goto got_name;
4719                }
4720                name = d_path(&file->f_path, buf, PATH_MAX);
4721                if (IS_ERR(name)) {
4722                        name = strncpy(tmp, "//toolong", sizeof(tmp));
4723                        goto got_name;
4724                }
4725        } else {
4726                if (arch_vma_name(mmap_event->vma)) {
4727                        name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4728                                       sizeof(tmp));
4729                        goto got_name;
4730                }
4731
4732                if (!vma->vm_mm) {
4733                        name = strncpy(tmp, "[vdso]", sizeof(tmp));
4734                        goto got_name;
4735                } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4736                                vma->vm_end >= vma->vm_mm->brk) {
4737                        name = strncpy(tmp, "[heap]", sizeof(tmp));
4738                        goto got_name;
4739                } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4740                                vma->vm_end >= vma->vm_mm->start_stack) {
4741                        name = strncpy(tmp, "[stack]", sizeof(tmp));
4742                        goto got_name;
4743                }
4744
4745                name = strncpy(tmp, "//anon", sizeof(tmp));
4746                goto got_name;
4747        }
4748
4749got_name:
4750        size = ALIGN(strlen(name)+1, sizeof(u64));
4751
4752        mmap_event->file_name = name;
4753        mmap_event->file_size = size;
4754
4755        mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4756
4757        rcu_read_lock();
4758        list_for_each_entry_rcu(pmu, &pmus, entry) {
4759                cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4760                if (cpuctx->unique_pmu != pmu)
4761                        goto next;
4762                perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4763                                        vma->vm_flags & VM_EXEC);
4764
4765                ctxn = pmu->task_ctx_nr;
4766                if (ctxn < 0)
4767                        goto next;
4768
4769                ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4770                if (ctx) {
4771                        perf_event_mmap_ctx(ctx, mmap_event,
4772                                        vma->vm_flags & VM_EXEC);
4773                }
4774next:
4775                put_cpu_ptr(pmu->pmu_cpu_context);
4776        }
4777        rcu_read_unlock();
4778
4779        kfree(buf);
4780}
4781
4782void perf_event_mmap(struct vm_area_struct *vma)
4783{
4784        struct perf_mmap_event mmap_event;
4785
4786        if (!atomic_read(&nr_mmap_events))
4787                return;
4788
4789        mmap_event = (struct perf_mmap_event){
4790                .vma    = vma,
4791                /* .file_name */
4792                /* .file_size */
4793                .event_id  = {
4794                        .header = {
4795                                .type = PERF_RECORD_MMAP,
4796                                .misc = PERF_RECORD_MISC_USER,
4797                                /* .size */
4798                        },
4799                        /* .pid */
4800                        /* .tid */
4801                        .start  = vma->vm_start,
4802                        .len    = vma->vm_end - vma->vm_start,
4803                        .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
4804                },
4805        };
4806
4807        perf_event_mmap_event(&mmap_event);
4808}
4809
4810/*
4811 * IRQ throttle logging
4812 */
4813
4814static void perf_log_throttle(struct perf_event *event, int enable)
4815{
4816        struct perf_output_handle handle;
4817        struct perf_sample_data sample;
4818        int ret;
4819
4820        struct {
4821                struct perf_event_header        header;
4822                u64                             time;
4823                u64                             id;
4824                u64                             stream_id;
4825        } throttle_event = {
4826                .header = {
4827                        .type = PERF_RECORD_THROTTLE,
4828                        .misc = 0,
4829                        .size = sizeof(throttle_event),
4830                },
4831                .time           = perf_clock(),
4832                .id             = primary_event_id(event),
4833                .stream_id      = event->id,
4834        };
4835
4836        if (enable)
4837                throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4838
4839        perf_event_header__init_id(&throttle_event.header, &sample, event);
4840
4841        ret = perf_output_begin(&handle, event,
4842                                throttle_event.header.size);
4843        if (ret)
4844                return;
4845
4846        perf_output_put(&handle, throttle_event);
4847        perf_event__output_id_sample(event, &handle, &sample);
4848        perf_output_end(&handle);
4849}
4850
4851/*
4852 * Generic event overflow handling, sampling.
4853 */
4854
4855static int __perf_event_overflow(struct perf_event *event,
4856                                   int throttle, struct perf_sample_data *data,
4857                                   struct pt_regs *regs)
4858{
4859        int events = atomic_read(&event->event_limit);
4860        struct hw_perf_event *hwc = &event->hw;
4861        u64 seq;
4862        int ret = 0;
4863
4864        /*
4865         * Non-sampling counters might still use the PMI to fold short
4866         * hardware counters, ignore those.
4867         */
4868        if (unlikely(!is_sampling_event(event)))
4869                return 0;
4870
4871        seq = __this_cpu_read(perf_throttled_seq);
4872        if (seq != hwc->interrupts_seq) {
4873                hwc->interrupts_seq = seq;
4874                hwc->interrupts = 1;
4875        } else {
4876                hwc->interrupts++;
4877                if (unlikely(throttle
4878                             && hwc->interrupts >= max_samples_per_tick)) {
4879                        __this_cpu_inc(perf_throttled_count);
4880                        hwc->interrupts = MAX_INTERRUPTS;
4881                        perf_log_throttle(event, 0);
4882                        ret = 1;
4883                }
4884        }
4885
4886        if (event->attr.freq) {
4887                u64 now = perf_clock();
4888                s64 delta = now - hwc->freq_time_stamp;
4889
4890                hwc->freq_time_stamp = now;
4891
4892                if (delta > 0 && delta < 2*TICK_NSEC)
4893                        perf_adjust_period(event, delta, hwc->last_period, true);
4894        }
4895
4896        /*
4897         * XXX event_limit might not quite work as expected on inherited
4898         * events
4899         */
4900
4901        event->pending_kill = POLL_IN;
4902        if (events && atomic_dec_and_test(&event->event_limit)) {
4903                ret = 1;
4904                event->pending_kill = POLL_HUP;
4905                event->pending_disable = 1;
4906                irq_work_queue(&event->pending);
4907        }
4908
4909        if (event->overflow_handler)
4910                event->overflow_handler(event, data, regs);
4911        else
4912                perf_event_output(event, data, regs);
4913
4914        if (event->fasync && event->pending_kill) {
4915                event->pending_wakeup = 1;
4916                irq_work_queue(&event->pending);
4917        }
4918
4919        return ret;
4920}
4921
4922int perf_event_overflow(struct perf_event *event,
4923                          struct perf_sample_data *data,
4924                          struct pt_regs *regs)
4925{
4926        return __perf_event_overflow(event, 1, data, regs);
4927}
4928
4929/*
4930 * Generic software event infrastructure
4931 */
4932
4933struct swevent_htable {
4934        struct swevent_hlist            *swevent_hlist;
4935        struct mutex                    hlist_mutex;
4936        int                             hlist_refcount;
4937
4938        /* Recursion avoidance in each contexts */
4939        int                             recursion[PERF_NR_CONTEXTS];
4940};
4941
4942static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4943
4944/*
4945 * We directly increment event->count and keep a second value in
4946 * event->hw.period_left to count intervals. This period event
4947 * is kept in the range [-sample_period, 0] so that we can use the
4948 * sign as trigger.
4949 */
4950
4951static u64 perf_swevent_set_period(struct perf_event *event)
4952{
4953        struct hw_perf_event *hwc = &event->hw;
4954        u64 period = hwc->last_period;
4955        u64 nr, offset;
4956        s64 old, val;
4957
4958        hwc->last_period = hwc->sample_period;
4959
4960again:
4961        old = val = local64_read(&hwc->period_left);
4962        if (val < 0)
4963                return 0;
4964
4965        nr = div64_u64(period + val, period);
4966        offset = nr * period;
4967        val -= offset;
4968        if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4969                goto again;
4970
4971        return nr;
4972}
4973
4974static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4975                                    struct perf_sample_data *data,
4976                                    struct pt_regs *regs)
4977{
4978        struct hw_perf_event *hwc = &event->hw;
4979        int throttle = 0;
4980
4981        if (!overflow)
4982                overflow = perf_swevent_set_period(event);
4983
4984        if (hwc->interrupts == MAX_INTERRUPTS)
4985                return;
4986
4987        for (; overflow; overflow--) {
4988                if (__perf_event_overflow(event, throttle,
4989                                            data, regs)) {
4990                        /*
4991                         * We inhibit the overflow from happening when
4992                         * hwc->interrupts == MAX_INTERRUPTS.
4993                         */
4994                        break;
4995                }
4996                throttle = 1;
4997        }
4998}
4999
5000static void perf_swevent_event(struct perf_event *event, u64 nr,
5001                               struct perf_sample_data *data,
5002                               struct pt_regs *regs)
5003{
5004        struct hw_perf_event *hwc = &event->hw;
5005
5006        local64_add(nr, &event->count);
5007
5008        if (!regs)
5009                return;
5010
5011        if (!is_sampling_event(event))
5012                return;
5013
5014        if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5015                data->period = nr;
5016                return perf_swevent_overflow(event, 1, data, regs);
5017        } else
5018                data->period = event->hw.last_period;
5019
5020        if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5021                return perf_swevent_overflow(event, 1, data, regs);
5022
5023        if (local64_add_negative(nr, &hwc->period_left))
5024                return;
5025
5026        perf_swevent_overflow(event, 0, data, regs);
5027}
5028
5029static int perf_exclude_event(struct perf_event *event,
5030                              struct pt_regs *regs)
5031{
5032        if (event->hw.state & PERF_HES_STOPPED)
5033                return 1;
5034
5035        if (regs) {
5036                if (event->attr.exclude_user && user_mode(regs))
5037                        return 1;
5038
5039                if (event->attr.exclude_kernel && !user_mode(regs))
5040                        return 1;
5041        }
5042
5043        return 0;
5044}
5045
5046static int perf_swevent_match(struct perf_event *event,
5047                                enum perf_type_id type,
5048                                u32 event_id,
5049                                struct perf_sample_data *data,
5050                                struct pt_regs *regs)
5051{
5052        if (event->attr.type != type)
5053                return 0;
5054
5055        if (event->attr.config != event_id)
5056                return 0;
5057
5058        if (perf_exclude_event(event, regs))
5059                return 0;
5060
5061        return 1;
5062}
5063
5064static inline u64 swevent_hash(u64 type, u32 event_id)
5065{
5066        u64 val = event_id | (type << 32);
5067
5068        return hash_64(val, SWEVENT_HLIST_BITS);
5069}
5070
5071static inline struct hlist_head *
5072__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5073{
5074        u64 hash = swevent_hash(type, event_id);
5075
5076        return &hlist->heads[hash];
5077}
5078
5079/* For the read side: events when they trigger */
5080static inline struct hlist_head *
5081find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5082{
5083        struct swevent_hlist *hlist;
5084
5085        hlist = rcu_dereference(swhash->swevent_hlist);
5086        if (!hlist)
5087                return NULL;
5088
5089        return __find_swevent_head(hlist, type, event_id);
5090}
5091
5092/* For the event head insertion and removal in the hlist */
5093static inline struct hlist_head *
5094find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5095{
5096        struct swevent_hlist *hlist;
5097        u32 event_id = event->attr.config;
5098        u64 type = event->attr.type;
5099
5100        /*
5101         * Event scheduling is always serialized against hlist allocation
5102         * and release. Which makes the protected version suitable here.
5103         * The context lock guarantees that.
5104         */
5105        hlist = rcu_dereference_protected(swhash->swevent_hlist,
5106                                          lockdep_is_held(&event->ctx->lock));
5107        if (!hlist)
5108                return NULL;
5109
5110        return __find_swevent_head(hlist, type, event_id);
5111}
5112
5113static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5114                                    u64 nr,
5115                                    struct perf_sample_data *data,
5116                                    struct pt_regs *regs)
5117{
5118        struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5119        struct perf_event *event;
5120        struct hlist_node *node;
5121        struct hlist_head *head;
5122
5123        rcu_read_lock();
5124        head = find_swevent_head_rcu(swhash, type, event_id);
5125        if (!head)
5126                goto end;
5127
5128        hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5129                if (perf_swevent_match(event, type, event_id, data, regs))
5130                        perf_swevent_event(event, nr, data, regs);
5131        }
5132end:
5133        rcu_read_unlock();
5134}
5135
5136int perf_swevent_get_recursion_context(void)
5137{
5138        struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5139
5140        return get_recursion_context(swhash->recursion);
5141}
5142EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5143
5144inline void perf_swevent_put_recursion_context(int rctx)
5145{
5146        struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5147
5148        put_recursion_context(swhash->recursion, rctx);
5149}
5150
5151void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5152{
5153        struct perf_sample_data data;
5154        int rctx;
5155
5156        preempt_disable_notrace();
5157        rctx = perf_swevent_get_recursion_context();
5158        if (rctx < 0)
5159                return;
5160
5161        perf_sample_data_init(&data, addr, 0);
5162
5163        do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5164
5165        perf_swevent_put_recursion_context(rctx);
5166        preempt_enable_notrace();
5167}
5168
5169static void perf_swevent_read(struct perf_event *event)
5170{
5171}
5172
5173static int perf_swevent_add(struct perf_event *event, int flags)
5174{
5175        struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5176        struct hw_perf_event *hwc = &event->hw;
5177        struct hlist_head *head;
5178
5179        if (is_sampling_event(event)) {
5180                hwc->last_period = hwc->sample_period;
5181                perf_swevent_set_period(event);
5182        }
5183
5184        hwc->state = !(flags & PERF_EF_START);
5185
5186        head = find_swevent_head(swhash, event);
5187        if (WARN_ON_ONCE(!head))
5188                return -EINVAL;
5189
5190        hlist_add_head_rcu(&event->hlist_entry, head);
5191
5192        return 0;
5193}
5194
5195static void perf_swevent_del(struct perf_event *event, int flags)
5196{
5197        hlist_del_rcu(&event->hlist_entry);
5198}
5199
5200static void perf_swevent_start(struct perf_event *event, int flags)
5201{
5202        event->hw.state = 0;
5203}
5204
5205static void perf_swevent_stop(struct perf_event *event, int flags)
5206{
5207        event->hw.state = PERF_HES_STOPPED;
5208}
5209
5210/* Deref the hlist from the update side */
5211static inline struct swevent_hlist *
5212swevent_hlist_deref(struct swevent_htable *swhash)
5213{
5214        return rcu_dereference_protected(swhash->swevent_hlist,
5215                                         lockdep_is_held(&swhash->hlist_mutex));
5216}
5217
5218static void swevent_hlist_release(struct swevent_htable *swhash)
5219{
5220        struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5221
5222        if (!hlist)
5223                return;
5224
5225        rcu_assign_pointer(swhash->swevent_hlist, NULL);
5226        kfree_rcu(hlist, rcu_head);
5227}
5228
5229static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5230{
5231        struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5232
5233        mutex_lock(&swhash->hlist_mutex);
5234
5235        if (!--swhash->hlist_refcount)
5236                swevent_hlist_release(swhash);
5237
5238        mutex_unlock(&swhash->hlist_mutex);
5239}
5240
5241static void swevent_hlist_put(struct perf_event *event)
5242{
5243        int cpu;
5244
5245        if (event->cpu != -1) {
5246                swevent_hlist_put_cpu(event, event->cpu);
5247                return;
5248        }
5249
5250        for_each_possible_cpu(cpu)
5251                swevent_hlist_put_cpu(event, cpu);
5252}
5253
5254static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5255{
5256        struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5257        int err = 0;
5258
5259        mutex_lock(&swhash->hlist_mutex);
5260
5261        if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5262                struct swevent_hlist *hlist;
5263
5264                hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5265                if (!hlist) {
5266                        err = -ENOMEM;
5267                        goto exit;
5268                }
5269                rcu_assign_pointer(swhash->swevent_hlist, hlist);
5270        }
5271        swhash->hlist_refcount++;
5272exit:
5273        mutex_unlock(&swhash->hlist_mutex);
5274
5275        return err;
5276}
5277
5278static int swevent_hlist_get(struct perf_event *event)
5279{
5280        int err;
5281        int cpu, failed_cpu;
5282
5283        if (event->cpu != -1)
5284                return swevent_hlist_get_cpu(event, event->cpu);
5285
5286        get_online_cpus();
5287        for_each_possible_cpu(cpu) {
5288                err = swevent_hlist_get_cpu(event, cpu);
5289                if (err) {
5290                        failed_cpu = cpu;
5291                        goto fail;
5292                }
5293        }
5294        put_online_cpus();
5295
5296        return 0;
5297fail:
5298        for_each_possible_cpu(cpu) {
5299                if (cpu == failed_cpu)
5300                        break;
5301                swevent_hlist_put_cpu(event, cpu);
5302        }
5303
5304        put_online_cpus();
5305        return err;
5306}
5307
5308struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5309
5310static void sw_perf_event_destroy(struct perf_event *event)
5311{
5312        u64 event_id = event->attr.config;
5313
5314        WARN_ON(event->parent);
5315
5316        static_key_slow_dec(&perf_swevent_enabled[event_id]);
5317        swevent_hlist_put(event);
5318}
5319
5320static int perf_swevent_init(struct perf_event *event)
5321{
5322        int event_id = event->attr.config;
5323
5324        if (event->attr.type != PERF_TYPE_SOFTWARE)
5325                return -ENOENT;
5326
5327        /*
5328         * no branch sampling for software events
5329         */
5330        if (has_branch_stack(event))
5331                return -EOPNOTSUPP;
5332
5333        switch (event_id) {
5334        case PERF_COUNT_SW_CPU_CLOCK:
5335        case PERF_COUNT_SW_TASK_CLOCK:
5336                return -ENOENT;
5337
5338        default:
5339                break;
5340        }
5341
5342        if (event_id >= PERF_COUNT_SW_MAX)
5343                return -ENOENT;
5344
5345        if (!event->parent) {
5346                int err;
5347
5348                err = swevent_hlist_get(event);
5349                if (err)
5350                        return err;
5351
5352                static_key_slow_inc(&perf_swevent_enabled[event_id]);
5353                event->destroy = sw_perf_event_destroy;
5354        }
5355
5356        return 0;
5357}
5358
5359static int perf_swevent_event_idx(struct perf_event *event)
5360{
5361        return 0;
5362}
5363
5364static struct pmu perf_swevent = {
5365        .task_ctx_nr    = perf_sw_context,
5366
5367        .event_init     = perf_swevent_init,
5368        .add            = perf_swevent_add,
5369        .del            = perf_swevent_del,
5370        .start          = perf_swevent_start,
5371        .stop           = perf_swevent_stop,
5372        .read           = perf_swevent_read,
5373
5374        .event_idx      = perf_swevent_event_idx,
5375};
5376
5377#ifdef CONFIG_EVENT_TRACING
5378
5379static int perf_tp_filter_match(struct perf_event *event,
5380                                struct perf_sample_data *data)
5381{
5382        void *record = data->raw->data;
5383
5384        if (likely(!event->filter) || filter_match_preds(event->filter, record))
5385                return 1;
5386        return 0;
5387}
5388
5389static int perf_tp_event_match(struct perf_event *event,
5390                                struct perf_sample_data *data,
5391                                struct pt_regs *regs)
5392{
5393        if (event->hw.state & PERF_HES_STOPPED)
5394                return 0;
5395        /*
5396         * All tracepoints are from kernel-space.
5397         */
5398        if (event->attr.exclude_kernel)
5399                return 0;
5400
5401        if (!perf_tp_filter_match(event, data))
5402                return 0;
5403
5404        return 1;
5405}
5406
5407void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5408                   struct pt_regs *regs, struct hlist_head *head, int rctx,
5409                   struct task_struct *task)
5410{
5411        struct perf_sample_data data;
5412        struct perf_event *event;
5413        struct hlist_node *node;
5414
5415        struct perf_raw_record raw = {
5416                .size = entry_size,
5417                .data = record,
5418        };
5419
5420        perf_sample_data_init(&data, addr, 0);
5421        data.raw = &raw;
5422
5423        hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5424                if (perf_tp_event_match(event, &data, regs))
5425                        perf_swevent_event(event, count, &data, regs);
5426        }
5427
5428        /*
5429         * If we got specified a target task, also iterate its context and
5430         * deliver this event there too.
5431         */
5432        if (task && task != current) {
5433                struct perf_event_context *ctx;
5434                struct trace_entry *entry = record;
5435
5436                rcu_read_lock();
5437                ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5438                if (!ctx)
5439                        goto unlock;
5440
5441                list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5442                        if (event->attr.type != PERF_TYPE_TRACEPOINT)
5443                                continue;
5444                        if (event->attr.config != entry->type)
5445                                continue;
5446                        if (perf_tp_event_match(event, &data, regs))
5447                                perf_swevent_event(event, count, &data, regs);
5448                }
5449unlock:
5450                rcu_read_unlock();
5451        }
5452
5453        perf_swevent_put_recursion_context(rctx);
5454}
5455EXPORT_SYMBOL_GPL(perf_tp_event);
5456
5457static void tp_perf_event_destroy(struct perf_event *event)
5458{
5459        perf_trace_destroy(event);
5460}
5461
5462static int perf_tp_event_init(struct perf_event *event)
5463{
5464        int err;
5465
5466        if (event->attr.type != PERF_TYPE_TRACEPOINT)
5467                return -ENOENT;
5468
5469        /*
5470         * no branch sampling for tracepoint events
5471         */
5472        if (has_branch_stack(event))
5473                return -EOPNOTSUPP;
5474
5475        err = perf_trace_init(event);
5476        if (err)
5477                return err;
5478
5479        event->destroy = tp_perf_event_destroy;
5480
5481        return 0;
5482}
5483
5484static struct pmu perf_tracepoint = {
5485        .task_ctx_nr    = perf_sw_context,
5486
5487        .event_init     = perf_tp_event_init,
5488        .add            = perf_trace_add,
5489        .del            = perf_trace_del,
5490        .start          = perf_swevent_start,
5491        .stop           = perf_swevent_stop,
5492        .read           = perf_swevent_read,
5493
5494        .event_idx      = perf_swevent_event_idx,
5495};
5496
5497static inline void perf_tp_register(void)
5498{
5499        perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5500}
5501
5502static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5503{
5504        char *filter_str;
5505        int ret;
5506
5507        if (event->attr.type != PERF_TYPE_TRACEPOINT)
5508                return -EINVAL;
5509
5510        filter_str = strndup_user(arg, PAGE_SIZE);
5511        if (IS_ERR(filter_str))
5512                return PTR_ERR(filter_str);
5513
5514        ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5515
5516        kfree(filter_str);
5517        return ret;
5518}
5519
5520static void perf_event_free_filter(struct perf_event *event)
5521{
5522        ftrace_profile_free_filter(event);
5523}
5524
5525#else
5526
5527static inline void perf_tp_register(void)
5528{
5529}
5530
5531static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5532{
5533        return -ENOENT;
5534}
5535
5536static void perf_event_free_filter(struct perf_event *event)
5537{
5538}
5539
5540#endif /* CONFIG_EVENT_TRACING */
5541
5542#ifdef CONFIG_HAVE_HW_BREAKPOINT
5543void perf_bp_event(struct perf_event *bp, void *data)
5544{
5545        struct perf_sample_data sample;
5546        struct pt_regs *regs = data;
5547
5548        perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5549
5550        if (!bp->hw.state && !perf_exclude_event(bp, regs))
5551                perf_swevent_event(bp, 1, &sample, regs);
5552}
5553#endif
5554
5555/*
5556 * hrtimer based swevent callback
5557 */
5558
5559static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5560{
5561        enum hrtimer_restart ret = HRTIMER_RESTART;
5562        struct perf_sample_data data;
5563        struct pt_regs *regs;
5564        struct perf_event *event;
5565        u64 period;
5566
5567        event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5568
5569        if (event->state != PERF_EVENT_STATE_ACTIVE)
5570                return HRTIMER_NORESTART;
5571
5572        event->pmu->read(event);
5573
5574        perf_sample_data_init(&data, 0, event->hw.last_period);
5575        regs = get_irq_regs();
5576
5577        if (regs && !perf_exclude_event(event, regs)) {
5578                if (!(event->attr.exclude_idle && is_idle_task(current)))
5579                        if (__perf_event_overflow(event, 1, &data, regs))
5580                                ret = HRTIMER_NORESTART;
5581        }
5582
5583        period = max_t(u64, 10000, event->hw.sample_period);
5584        hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5585
5586        return ret;
5587}
5588
5589static void perf_swevent_start_hrtimer(struct perf_event *event)
5590{
5591        struct hw_perf_event *hwc = &event->hw;