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