linux/fs/eventpoll.c
<<
>>
Prefs
   1/*
   2 *  fs/eventpoll.c (Efficient event retrieval implementation)
   3 *  Copyright (C) 2001,...,2009  Davide Libenzi
   4 *
   5 *  This program is free software; you can redistribute it and/or modify
   6 *  it under the terms of the GNU General Public License as published by
   7 *  the Free Software Foundation; either version 2 of the License, or
   8 *  (at your option) any later version.
   9 *
  10 *  Davide Libenzi <davidel@xmailserver.org>
  11 *
  12 */
  13
  14#include <linux/init.h>
  15#include <linux/kernel.h>
  16#include <linux/sched.h>
  17#include <linux/fs.h>
  18#include <linux/file.h>
  19#include <linux/signal.h>
  20#include <linux/errno.h>
  21#include <linux/mm.h>
  22#include <linux/slab.h>
  23#include <linux/poll.h>
  24#include <linux/string.h>
  25#include <linux/list.h>
  26#include <linux/hash.h>
  27#include <linux/spinlock.h>
  28#include <linux/syscalls.h>
  29#include <linux/rbtree.h>
  30#include <linux/wait.h>
  31#include <linux/eventpoll.h>
  32#include <linux/mount.h>
  33#include <linux/bitops.h>
  34#include <linux/mutex.h>
  35#include <linux/anon_inodes.h>
  36#include <linux/device.h>
  37#include <asm/uaccess.h>
  38#include <asm/io.h>
  39#include <asm/mman.h>
  40#include <linux/atomic.h>
  41#include <linux/proc_fs.h>
  42#include <linux/seq_file.h>
  43
  44/*
  45 * LOCKING:
  46 * There are three level of locking required by epoll :
  47 *
  48 * 1) epmutex (mutex)
  49 * 2) ep->mtx (mutex)
  50 * 3) ep->lock (spinlock)
  51 *
  52 * The acquire order is the one listed above, from 1 to 3.
  53 * We need a spinlock (ep->lock) because we manipulate objects
  54 * from inside the poll callback, that might be triggered from
  55 * a wake_up() that in turn might be called from IRQ context.
  56 * So we can't sleep inside the poll callback and hence we need
  57 * a spinlock. During the event transfer loop (from kernel to
  58 * user space) we could end up sleeping due a copy_to_user(), so
  59 * we need a lock that will allow us to sleep. This lock is a
  60 * mutex (ep->mtx). It is acquired during the event transfer loop,
  61 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
  62 * Then we also need a global mutex to serialize eventpoll_release_file()
  63 * and ep_free().
  64 * This mutex is acquired by ep_free() during the epoll file
  65 * cleanup path and it is also acquired by eventpoll_release_file()
  66 * if a file has been pushed inside an epoll set and it is then
  67 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
  68 * It is also acquired when inserting an epoll fd onto another epoll
  69 * fd. We do this so that we walk the epoll tree and ensure that this
  70 * insertion does not create a cycle of epoll file descriptors, which
  71 * could lead to deadlock. We need a global mutex to prevent two
  72 * simultaneous inserts (A into B and B into A) from racing and
  73 * constructing a cycle without either insert observing that it is
  74 * going to.
  75 * It is necessary to acquire multiple "ep->mtx"es at once in the
  76 * case when one epoll fd is added to another. In this case, we
  77 * always acquire the locks in the order of nesting (i.e. after
  78 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
  79 * before e2->mtx). Since we disallow cycles of epoll file
  80 * descriptors, this ensures that the mutexes are well-ordered. In
  81 * order to communicate this nesting to lockdep, when walking a tree
  82 * of epoll file descriptors, we use the current recursion depth as
  83 * the lockdep subkey.
  84 * It is possible to drop the "ep->mtx" and to use the global
  85 * mutex "epmutex" (together with "ep->lock") to have it working,
  86 * but having "ep->mtx" will make the interface more scalable.
  87 * Events that require holding "epmutex" are very rare, while for
  88 * normal operations the epoll private "ep->mtx" will guarantee
  89 * a better scalability.
  90 */
  91
  92/* Epoll private bits inside the event mask */
  93#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
  94
  95/* Maximum number of nesting allowed inside epoll sets */
  96#define EP_MAX_NESTS 4
  97
  98#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
  99
 100#define EP_UNACTIVE_PTR ((void *) -1L)
 101
 102#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
 103
 104struct epoll_filefd {
 105        struct file *file;
 106        int fd;
 107};
 108
 109/*
 110 * Structure used to track possible nested calls, for too deep recursions
 111 * and loop cycles.
 112 */
 113struct nested_call_node {
 114        struct list_head llink;
 115        void *cookie;
 116        void *ctx;
 117};
 118
 119/*
 120 * This structure is used as collector for nested calls, to check for
 121 * maximum recursion dept and loop cycles.
 122 */
 123struct nested_calls {
 124        struct list_head tasks_call_list;
 125        spinlock_t lock;
 126};
 127
 128/*
 129 * Each file descriptor added to the eventpoll interface will
 130 * have an entry of this type linked to the "rbr" RB tree.
 131 */
 132struct epitem {
 133        /* RB tree node used to link this structure to the eventpoll RB tree */
 134        struct rb_node rbn;
 135
 136        /* List header used to link this structure to the eventpoll ready list */
 137        struct list_head rdllink;
 138
 139        /*
 140         * Works together "struct eventpoll"->ovflist in keeping the
 141         * single linked chain of items.
 142         */
 143        struct epitem *next;
 144
 145        /* The file descriptor information this item refers to */
 146        struct epoll_filefd ffd;
 147
 148        /* Number of active wait queue attached to poll operations */
 149        int nwait;
 150
 151        /* List containing poll wait queues */
 152        struct list_head pwqlist;
 153
 154        /* The "container" of this item */
 155        struct eventpoll *ep;
 156
 157        /* List header used to link this item to the "struct file" items list */
 158        struct list_head fllink;
 159
 160        /* wakeup_source used when EPOLLWAKEUP is set */
 161        struct wakeup_source *ws;
 162
 163        /* The structure that describe the interested events and the source fd */
 164        struct epoll_event event;
 165};
 166
 167/*
 168 * This structure is stored inside the "private_data" member of the file
 169 * structure and represents the main data structure for the eventpoll
 170 * interface.
 171 */
 172struct eventpoll {
 173        /* Protect the access to this structure */
 174        spinlock_t lock;
 175
 176        /*
 177         * This mutex is used to ensure that files are not removed
 178         * while epoll is using them. This is held during the event
 179         * collection loop, the file cleanup path, the epoll file exit
 180         * code and the ctl operations.
 181         */
 182        struct mutex mtx;
 183
 184        /* Wait queue used by sys_epoll_wait() */
 185        wait_queue_head_t wq;
 186
 187        /* Wait queue used by file->poll() */
 188        wait_queue_head_t poll_wait;
 189
 190        /* List of ready file descriptors */
 191        struct list_head rdllist;
 192
 193        /* RB tree root used to store monitored fd structs */
 194        struct rb_root rbr;
 195
 196        /*
 197         * This is a single linked list that chains all the "struct epitem" that
 198         * happened while transferring ready events to userspace w/out
 199         * holding ->lock.
 200         */
 201        struct epitem *ovflist;
 202
 203        /* wakeup_source used when ep_scan_ready_list is running */
 204        struct wakeup_source *ws;
 205
 206        /* The user that created the eventpoll descriptor */
 207        struct user_struct *user;
 208
 209        struct file *file;
 210
 211        /* used to optimize loop detection check */
 212        int visited;
 213        struct list_head visited_list_link;
 214};
 215
 216/* Wait structure used by the poll hooks */
 217struct eppoll_entry {
 218        /* List header used to link this structure to the "struct epitem" */
 219        struct list_head llink;
 220
 221        /* The "base" pointer is set to the container "struct epitem" */
 222        struct epitem *base;
 223
 224        /*
 225         * Wait queue item that will be linked to the target file wait
 226         * queue head.
 227         */
 228        wait_queue_t wait;
 229
 230        /* The wait queue head that linked the "wait" wait queue item */
 231        wait_queue_head_t *whead;
 232};
 233
 234/* Wrapper struct used by poll queueing */
 235struct ep_pqueue {
 236        poll_table pt;
 237        struct epitem *epi;
 238};
 239
 240/* Used by the ep_send_events() function as callback private data */
 241struct ep_send_events_data {
 242        int maxevents;
 243        struct epoll_event __user *events;
 244};
 245
 246/*
 247 * Configuration options available inside /proc/sys/fs/epoll/
 248 */
 249/* Maximum number of epoll watched descriptors, per user */
 250static long max_user_watches __read_mostly;
 251
 252/*
 253 * This mutex is used to serialize ep_free() and eventpoll_release_file().
 254 */
 255static DEFINE_MUTEX(epmutex);
 256
 257/* Used to check for epoll file descriptor inclusion loops */
 258static struct nested_calls poll_loop_ncalls;
 259
 260/* Used for safe wake up implementation */
 261static struct nested_calls poll_safewake_ncalls;
 262
 263/* Used to call file's f_op->poll() under the nested calls boundaries */
 264static struct nested_calls poll_readywalk_ncalls;
 265
 266/* Slab cache used to allocate "struct epitem" */
 267static struct kmem_cache *epi_cache __read_mostly;
 268
 269/* Slab cache used to allocate "struct eppoll_entry" */
 270static struct kmem_cache *pwq_cache __read_mostly;
 271
 272/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
 273static LIST_HEAD(visited_list);
 274
 275/*
 276 * List of files with newly added links, where we may need to limit the number
 277 * of emanating paths. Protected by the epmutex.
 278 */
 279static LIST_HEAD(tfile_check_list);
 280
 281#ifdef CONFIG_SYSCTL
 282
 283#include <linux/sysctl.h>
 284
 285static long zero;
 286static long long_max = LONG_MAX;
 287
 288ctl_table epoll_table[] = {
 289        {
 290                .procname       = "max_user_watches",
 291                .data           = &max_user_watches,
 292                .maxlen         = sizeof(max_user_watches),
 293                .mode           = 0644,
 294                .proc_handler   = proc_doulongvec_minmax,
 295                .extra1         = &zero,
 296                .extra2         = &long_max,
 297        },
 298        { }
 299};
 300#endif /* CONFIG_SYSCTL */
 301
 302static const struct file_operations eventpoll_fops;
 303
 304static inline int is_file_epoll(struct file *f)
 305{
 306        return f->f_op == &eventpoll_fops;
 307}
 308
 309/* Setup the structure that is used as key for the RB tree */
 310static inline void ep_set_ffd(struct epoll_filefd *ffd,
 311                              struct file *file, int fd)
 312{
 313        ffd->file = file;
 314        ffd->fd = fd;
 315}
 316
 317/* Compare RB tree keys */
 318static inline int ep_cmp_ffd(struct epoll_filefd *p1,
 319                             struct epoll_filefd *p2)
 320{
 321        return (p1->file > p2->file ? +1:
 322                (p1->file < p2->file ? -1 : p1->fd - p2->fd));
 323}
 324
 325/* Tells us if the item is currently linked */
 326static inline int ep_is_linked(struct list_head *p)
 327{
 328        return !list_empty(p);
 329}
 330
 331static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
 332{
 333        return container_of(p, struct eppoll_entry, wait);
 334}
 335
 336/* Get the "struct epitem" from a wait queue pointer */
 337static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
 338{
 339        return container_of(p, struct eppoll_entry, wait)->base;
 340}
 341
 342/* Get the "struct epitem" from an epoll queue wrapper */
 343static inline struct epitem *ep_item_from_epqueue(poll_table *p)
 344{
 345        return container_of(p, struct ep_pqueue, pt)->epi;
 346}
 347
 348/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
 349static inline int ep_op_has_event(int op)
 350{
 351        return op != EPOLL_CTL_DEL;
 352}
 353
 354/* Initialize the poll safe wake up structure */
 355static void ep_nested_calls_init(struct nested_calls *ncalls)
 356{
 357        INIT_LIST_HEAD(&ncalls->tasks_call_list);
 358        spin_lock_init(&ncalls->lock);
 359}
 360
 361/**
 362 * ep_events_available - Checks if ready events might be available.
 363 *
 364 * @ep: Pointer to the eventpoll context.
 365 *
 366 * Returns: Returns a value different than zero if ready events are available,
 367 *          or zero otherwise.
 368 */
 369static inline int ep_events_available(struct eventpoll *ep)
 370{
 371        return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
 372}
 373
 374/**
 375 * ep_call_nested - Perform a bound (possibly) nested call, by checking
 376 *                  that the recursion limit is not exceeded, and that
 377 *                  the same nested call (by the meaning of same cookie) is
 378 *                  no re-entered.
 379 *
 380 * @ncalls: Pointer to the nested_calls structure to be used for this call.
 381 * @max_nests: Maximum number of allowed nesting calls.
 382 * @nproc: Nested call core function pointer.
 383 * @priv: Opaque data to be passed to the @nproc callback.
 384 * @cookie: Cookie to be used to identify this nested call.
 385 * @ctx: This instance context.
 386 *
 387 * Returns: Returns the code returned by the @nproc callback, or -1 if
 388 *          the maximum recursion limit has been exceeded.
 389 */
 390static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
 391                          int (*nproc)(void *, void *, int), void *priv,
 392                          void *cookie, void *ctx)
 393{
 394        int error, call_nests = 0;
 395        unsigned long flags;
 396        struct list_head *lsthead = &ncalls->tasks_call_list;
 397        struct nested_call_node *tncur;
 398        struct nested_call_node tnode;
 399
 400        spin_lock_irqsave(&ncalls->lock, flags);
 401
 402        /*
 403         * Try to see if the current task is already inside this wakeup call.
 404         * We use a list here, since the population inside this set is always
 405         * very much limited.
 406         */
 407        list_for_each_entry(tncur, lsthead, llink) {
 408                if (tncur->ctx == ctx &&
 409                    (tncur->cookie == cookie || ++call_nests > max_nests)) {
 410                        /*
 411                         * Ops ... loop detected or maximum nest level reached.
 412                         * We abort this wake by breaking the cycle itself.
 413                         */
 414                        error = -1;
 415                        goto out_unlock;
 416                }
 417        }
 418
 419        /* Add the current task and cookie to the list */
 420        tnode.ctx = ctx;
 421        tnode.cookie = cookie;
 422        list_add(&tnode.llink, lsthead);
 423
 424        spin_unlock_irqrestore(&ncalls->lock, flags);
 425
 426        /* Call the nested function */
 427        error = (*nproc)(priv, cookie, call_nests);
 428
 429        /* Remove the current task from the list */
 430        spin_lock_irqsave(&ncalls->lock, flags);
 431        list_del(&tnode.llink);
 432out_unlock:
 433        spin_unlock_irqrestore(&ncalls->lock, flags);
 434
 435        return error;
 436}
 437
 438/*
 439 * As described in commit 0ccf831cb lockdep: annotate epoll
 440 * the use of wait queues used by epoll is done in a very controlled
 441 * manner. Wake ups can nest inside each other, but are never done
 442 * with the same locking. For example:
 443 *
 444 *   dfd = socket(...);
 445 *   efd1 = epoll_create();
 446 *   efd2 = epoll_create();
 447 *   epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
 448 *   epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
 449 *
 450 * When a packet arrives to the device underneath "dfd", the net code will
 451 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
 452 * callback wakeup entry on that queue, and the wake_up() performed by the
 453 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
 454 * (efd1) notices that it may have some event ready, so it needs to wake up
 455 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
 456 * that ends up in another wake_up(), after having checked about the
 457 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
 458 * avoid stack blasting.
 459 *
 460 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
 461 * this special case of epoll.
 462 */
 463#ifdef CONFIG_DEBUG_LOCK_ALLOC
 464static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
 465                                     unsigned long events, int subclass)
 466{
 467        unsigned long flags;
 468
 469        spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
 470        wake_up_locked_poll(wqueue, events);
 471        spin_unlock_irqrestore(&wqueue->lock, flags);
 472}
 473#else
 474static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
 475                                     unsigned long events, int subclass)
 476{
 477        wake_up_poll(wqueue, events);
 478}
 479#endif
 480
 481static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
 482{
 483        ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
 484                          1 + call_nests);
 485        return 0;
 486}
 487
 488/*
 489 * Perform a safe wake up of the poll wait list. The problem is that
 490 * with the new callback'd wake up system, it is possible that the
 491 * poll callback is reentered from inside the call to wake_up() done
 492 * on the poll wait queue head. The rule is that we cannot reenter the
 493 * wake up code from the same task more than EP_MAX_NESTS times,
 494 * and we cannot reenter the same wait queue head at all. This will
 495 * enable to have a hierarchy of epoll file descriptor of no more than
 496 * EP_MAX_NESTS deep.
 497 */
 498static void ep_poll_safewake(wait_queue_head_t *wq)
 499{
 500        int this_cpu = get_cpu();
 501
 502        ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
 503                       ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
 504
 505        put_cpu();
 506}
 507
 508static void ep_remove_wait_queue(struct eppoll_entry *pwq)
 509{
 510        wait_queue_head_t *whead;
 511
 512        rcu_read_lock();
 513        /* If it is cleared by POLLFREE, it should be rcu-safe */
 514        whead = rcu_dereference(pwq->whead);
 515        if (whead)
 516                remove_wait_queue(whead, &pwq->wait);
 517        rcu_read_unlock();
 518}
 519
 520/*
 521 * This function unregisters poll callbacks from the associated file
 522 * descriptor.  Must be called with "mtx" held (or "epmutex" if called from
 523 * ep_free).
 524 */
 525static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
 526{
 527        struct list_head *lsthead = &epi->pwqlist;
 528        struct eppoll_entry *pwq;
 529
 530        while (!list_empty(lsthead)) {
 531                pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
 532
 533                list_del(&pwq->llink);
 534                ep_remove_wait_queue(pwq);
 535                kmem_cache_free(pwq_cache, pwq);
 536        }
 537}
 538
 539/**
 540 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
 541 *                      the scan code, to call f_op->poll(). Also allows for
 542 *                      O(NumReady) performance.
 543 *
 544 * @ep: Pointer to the epoll private data structure.
 545 * @sproc: Pointer to the scan callback.
 546 * @priv: Private opaque data passed to the @sproc callback.
 547 * @depth: The current depth of recursive f_op->poll calls.
 548 *
 549 * Returns: The same integer error code returned by the @sproc callback.
 550 */
 551static int ep_scan_ready_list(struct eventpoll *ep,
 552                              int (*sproc)(struct eventpoll *,
 553                                           struct list_head *, void *),
 554                              void *priv,
 555                              int depth)
 556{
 557        int error, pwake = 0;
 558        unsigned long flags;
 559        struct epitem *epi, *nepi;
 560        LIST_HEAD(txlist);
 561
 562        /*
 563         * We need to lock this because we could be hit by
 564         * eventpoll_release_file() and epoll_ctl().
 565         */
 566        mutex_lock_nested(&ep->mtx, depth);
 567
 568        /*
 569         * Steal the ready list, and re-init the original one to the
 570         * empty list. Also, set ep->ovflist to NULL so that events
 571         * happening while looping w/out locks, are not lost. We cannot
 572         * have the poll callback to queue directly on ep->rdllist,
 573         * because we want the "sproc" callback to be able to do it
 574         * in a lockless way.
 575         */
 576        spin_lock_irqsave(&ep->lock, flags);
 577        list_splice_init(&ep->rdllist, &txlist);
 578        ep->ovflist = NULL;
 579        spin_unlock_irqrestore(&ep->lock, flags);
 580
 581        /*
 582         * Now call the callback function.
 583         */
 584        error = (*sproc)(ep, &txlist, priv);
 585
 586        spin_lock_irqsave(&ep->lock, flags);
 587        /*
 588         * During the time we spent inside the "sproc" callback, some
 589         * other events might have been queued by the poll callback.
 590         * We re-insert them inside the main ready-list here.
 591         */
 592        for (nepi = ep->ovflist; (epi = nepi) != NULL;
 593             nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
 594                /*
 595                 * We need to check if the item is already in the list.
 596                 * During the "sproc" callback execution time, items are
 597                 * queued into ->ovflist but the "txlist" might already
 598                 * contain them, and the list_splice() below takes care of them.
 599                 */
 600                if (!ep_is_linked(&epi->rdllink)) {
 601                        list_add_tail(&epi->rdllink, &ep->rdllist);
 602                        __pm_stay_awake(epi->ws);
 603                }
 604        }
 605        /*
 606         * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
 607         * releasing the lock, events will be queued in the normal way inside
 608         * ep->rdllist.
 609         */
 610        ep->ovflist = EP_UNACTIVE_PTR;
 611
 612        /*
 613         * Quickly re-inject items left on "txlist".
 614         */
 615        list_splice(&txlist, &ep->rdllist);
 616        __pm_relax(ep->ws);
 617
 618        if (!list_empty(&ep->rdllist)) {
 619                /*
 620                 * Wake up (if active) both the eventpoll wait list and
 621                 * the ->poll() wait list (delayed after we release the lock).
 622                 */
 623                if (waitqueue_active(&ep->wq))
 624                        wake_up_locked(&ep->wq);
 625                if (waitqueue_active(&ep->poll_wait))
 626                        pwake++;
 627        }
 628        spin_unlock_irqrestore(&ep->lock, flags);
 629
 630        mutex_unlock(&ep->mtx);
 631
 632        /* We have to call this outside the lock */
 633        if (pwake)
 634                ep_poll_safewake(&ep->poll_wait);
 635
 636        return error;
 637}
 638
 639/*
 640 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
 641 * all the associated resources. Must be called with "mtx" held.
 642 */
 643static int ep_remove(struct eventpoll *ep, struct epitem *epi)
 644{
 645        unsigned long flags;
 646        struct file *file = epi->ffd.file;
 647
 648        /*
 649         * Removes poll wait queue hooks. We _have_ to do this without holding
 650         * the "ep->lock" otherwise a deadlock might occur. This because of the
 651         * sequence of the lock acquisition. Here we do "ep->lock" then the wait
 652         * queue head lock when unregistering the wait queue. The wakeup callback
 653         * will run by holding the wait queue head lock and will call our callback
 654         * that will try to get "ep->lock".
 655         */
 656        ep_unregister_pollwait(ep, epi);
 657
 658        /* Remove the current item from the list of epoll hooks */
 659        spin_lock(&file->f_lock);
 660        if (ep_is_linked(&epi->fllink))
 661                list_del_init(&epi->fllink);
 662        spin_unlock(&file->f_lock);
 663
 664        rb_erase(&epi->rbn, &ep->rbr);
 665
 666        spin_lock_irqsave(&ep->lock, flags);
 667        if (ep_is_linked(&epi->rdllink))
 668                list_del_init(&epi->rdllink);
 669        spin_unlock_irqrestore(&ep->lock, flags);
 670
 671        wakeup_source_unregister(epi->ws);
 672
 673        /* At this point it is safe to free the eventpoll item */
 674        kmem_cache_free(epi_cache, epi);
 675
 676        atomic_long_dec(&ep->user->epoll_watches);
 677
 678        return 0;
 679}
 680
 681static void ep_free(struct eventpoll *ep)
 682{
 683        struct rb_node *rbp;
 684        struct epitem *epi;
 685
 686        /* We need to release all tasks waiting for these file */
 687        if (waitqueue_active(&ep->poll_wait))
 688                ep_poll_safewake(&ep->poll_wait);
 689
 690        /*
 691         * We need to lock this because we could be hit by
 692         * eventpoll_release_file() while we're freeing the "struct eventpoll".
 693         * We do not need to hold "ep->mtx" here because the epoll file
 694         * is on the way to be removed and no one has references to it
 695         * anymore. The only hit might come from eventpoll_release_file() but
 696         * holding "epmutex" is sufficient here.
 697         */
 698        mutex_lock(&epmutex);
 699
 700        /*
 701         * Walks through the whole tree by unregistering poll callbacks.
 702         */
 703        for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
 704                epi = rb_entry(rbp, struct epitem, rbn);
 705
 706                ep_unregister_pollwait(ep, epi);
 707        }
 708
 709        /*
 710         * Walks through the whole tree by freeing each "struct epitem". At this
 711         * point we are sure no poll callbacks will be lingering around, and also by
 712         * holding "epmutex" we can be sure that no file cleanup code will hit
 713         * us during this operation. So we can avoid the lock on "ep->lock".
 714         */
 715        while ((rbp = rb_first(&ep->rbr)) != NULL) {
 716                epi = rb_entry(rbp, struct epitem, rbn);
 717                ep_remove(ep, epi);
 718        }
 719
 720        mutex_unlock(&epmutex);
 721        mutex_destroy(&ep->mtx);
 722        free_uid(ep->user);
 723        wakeup_source_unregister(ep->ws);
 724        kfree(ep);
 725}
 726
 727static int ep_eventpoll_release(struct inode *inode, struct file *file)
 728{
 729        struct eventpoll *ep = file->private_data;
 730
 731        if (ep)
 732                ep_free(ep);
 733
 734        return 0;
 735}
 736
 737static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
 738                               void *priv)
 739{
 740        struct epitem *epi, *tmp;
 741        poll_table pt;
 742
 743        init_poll_funcptr(&pt, NULL);
 744        list_for_each_entry_safe(epi, tmp, head, rdllink) {
 745                pt._key = epi->event.events;
 746                if (epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
 747                    epi->event.events)
 748                        return POLLIN | POLLRDNORM;
 749                else {
 750                        /*
 751                         * Item has been dropped into the ready list by the poll
 752                         * callback, but it's not actually ready, as far as
 753                         * caller requested events goes. We can remove it here.
 754                         */
 755                        __pm_relax(epi->ws);
 756                        list_del_init(&epi->rdllink);
 757                }
 758        }
 759
 760        return 0;
 761}
 762
 763static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
 764{
 765        return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1);
 766}
 767
 768static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
 769{
 770        int pollflags;
 771        struct eventpoll *ep = file->private_data;
 772
 773        /* Insert inside our poll wait queue */
 774        poll_wait(file, &ep->poll_wait, wait);
 775
 776        /*
 777         * Proceed to find out if wanted events are really available inside
 778         * the ready list. This need to be done under ep_call_nested()
 779         * supervision, since the call to f_op->poll() done on listed files
 780         * could re-enter here.
 781         */
 782        pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
 783                                   ep_poll_readyevents_proc, ep, ep, current);
 784
 785        return pollflags != -1 ? pollflags : 0;
 786}
 787
 788#ifdef CONFIG_PROC_FS
 789static int ep_show_fdinfo(struct seq_file *m, struct file *f)
 790{
 791        struct eventpoll *ep = f->private_data;
 792        struct rb_node *rbp;
 793        int ret = 0;
 794
 795        mutex_lock(&ep->mtx);
 796        for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
 797                struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
 798
 799                ret = seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
 800                                 epi->ffd.fd, epi->event.events,
 801                                 (long long)epi->event.data);
 802                if (ret)
 803                        break;
 804        }
 805        mutex_unlock(&ep->mtx);
 806
 807        return ret;
 808}
 809#endif
 810
 811/* File callbacks that implement the eventpoll file behaviour */
 812static const struct file_operations eventpoll_fops = {
 813#ifdef CONFIG_PROC_FS
 814        .show_fdinfo    = ep_show_fdinfo,
 815#endif
 816        .release        = ep_eventpoll_release,
 817        .poll           = ep_eventpoll_poll,
 818        .llseek         = noop_llseek,
 819};
 820
 821/*
 822 * This is called from eventpoll_release() to unlink files from the eventpoll
 823 * interface. We need to have this facility to cleanup correctly files that are
 824 * closed without being removed from the eventpoll interface.
 825 */
 826void eventpoll_release_file(struct file *file)
 827{
 828        struct list_head *lsthead = &file->f_ep_links;
 829        struct eventpoll *ep;
 830        struct epitem *epi;
 831
 832        /*
 833         * We don't want to get "file->f_lock" because it is not
 834         * necessary. It is not necessary because we're in the "struct file"
 835         * cleanup path, and this means that no one is using this file anymore.
 836         * So, for example, epoll_ctl() cannot hit here since if we reach this
 837         * point, the file counter already went to zero and fget() would fail.
 838         * The only hit might come from ep_free() but by holding the mutex
 839         * will correctly serialize the operation. We do need to acquire
 840         * "ep->mtx" after "epmutex" because ep_remove() requires it when called
 841         * from anywhere but ep_free().
 842         *
 843         * Besides, ep_remove() acquires the lock, so we can't hold it here.
 844         */
 845        mutex_lock(&epmutex);
 846
 847        while (!list_empty(lsthead)) {
 848                epi = list_first_entry(lsthead, struct epitem, fllink);
 849
 850                ep = epi->ep;
 851                list_del_init(&epi->fllink);
 852                mutex_lock_nested(&ep->mtx, 0);
 853                ep_remove(ep, epi);
 854                mutex_unlock(&ep->mtx);
 855        }
 856
 857        mutex_unlock(&epmutex);
 858}
 859
 860static int ep_alloc(struct eventpoll **pep)
 861{
 862        int error;
 863        struct user_struct *user;
 864        struct eventpoll *ep;
 865
 866        user = get_current_user();
 867        error = -ENOMEM;
 868        ep = kzalloc(sizeof(*ep), GFP_KERNEL);
 869        if (unlikely(!ep))
 870                goto free_uid;
 871
 872        spin_lock_init(&ep->lock);
 873        mutex_init(&ep->mtx);
 874        init_waitqueue_head(&ep->wq);
 875        init_waitqueue_head(&ep->poll_wait);
 876        INIT_LIST_HEAD(&ep->rdllist);
 877        ep->rbr = RB_ROOT;
 878        ep->ovflist = EP_UNACTIVE_PTR;
 879        ep->user = user;
 880
 881        *pep = ep;
 882
 883        return 0;
 884
 885free_uid:
 886        free_uid(user);
 887        return error;
 888}
 889
 890/*
 891 * Search the file inside the eventpoll tree. The RB tree operations
 892 * are protected by the "mtx" mutex, and ep_find() must be called with
 893 * "mtx" held.
 894 */
 895static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
 896{
 897        int kcmp;
 898        struct rb_node *rbp;
 899        struct epitem *epi, *epir = NULL;
 900        struct epoll_filefd ffd;
 901
 902        ep_set_ffd(&ffd, file, fd);
 903        for (rbp = ep->rbr.rb_node; rbp; ) {
 904                epi = rb_entry(rbp, struct epitem, rbn);
 905                kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
 906                if (kcmp > 0)
 907                        rbp = rbp->rb_right;
 908                else if (kcmp < 0)
 909                        rbp = rbp->rb_left;
 910                else {
 911                        epir = epi;
 912                        break;
 913                }
 914        }
 915
 916        return epir;
 917}
 918
 919/*
 920 * This is the callback that is passed to the wait queue wakeup
 921 * mechanism. It is called by the stored file descriptors when they
 922 * have events to report.
 923 */
 924static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
 925{
 926        int pwake = 0;
 927        unsigned long flags;
 928        struct epitem *epi = ep_item_from_wait(wait);
 929        struct eventpoll *ep = epi->ep;
 930
 931        if ((unsigned long)key & POLLFREE) {
 932                ep_pwq_from_wait(wait)->whead = NULL;
 933                /*
 934                 * whead = NULL above can race with ep_remove_wait_queue()
 935                 * which can do another remove_wait_queue() after us, so we
 936                 * can't use __remove_wait_queue(). whead->lock is held by
 937                 * the caller.
 938                 */
 939                list_del_init(&wait->task_list);
 940        }
 941
 942        spin_lock_irqsave(&ep->lock, flags);
 943
 944        /*
 945         * If the event mask does not contain any poll(2) event, we consider the
 946         * descriptor to be disabled. This condition is likely the effect of the
 947         * EPOLLONESHOT bit that disables the descriptor when an event is received,
 948         * until the next EPOLL_CTL_MOD will be issued.
 949         */
 950        if (!(epi->event.events & ~EP_PRIVATE_BITS))
 951                goto out_unlock;
 952
 953        /*
 954         * Check the events coming with the callback. At this stage, not
 955         * every device reports the events in the "key" parameter of the
 956         * callback. We need to be able to handle both cases here, hence the
 957         * test for "key" != NULL before the event match test.
 958         */
 959        if (key && !((unsigned long) key & epi->event.events))
 960                goto out_unlock;
 961
 962        /*
 963         * If we are transferring events to userspace, we can hold no locks
 964         * (because we're accessing user memory, and because of linux f_op->poll()
 965         * semantics). All the events that happen during that period of time are
 966         * chained in ep->ovflist and requeued later on.
 967         */
 968        if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
 969                if (epi->next == EP_UNACTIVE_PTR) {
 970                        epi->next = ep->ovflist;
 971                        ep->ovflist = epi;
 972                        if (epi->ws) {
 973                                /*
 974                                 * Activate ep->ws since epi->ws may get
 975                                 * deactivated at any time.
 976                                 */
 977                                __pm_stay_awake(ep->ws);
 978                        }
 979
 980                }
 981                goto out_unlock;
 982        }
 983
 984        /* If this file is already in the ready list we exit soon */
 985        if (!ep_is_linked(&epi->rdllink)) {
 986                list_add_tail(&epi->rdllink, &ep->rdllist);
 987                __pm_stay_awake(epi->ws);
 988        }
 989
 990        /*
 991         * Wake up ( if active ) both the eventpoll wait list and the ->poll()
 992         * wait list.
 993         */
 994        if (waitqueue_active(&ep->wq))
 995                wake_up_locked(&ep->wq);
 996        if (waitqueue_active(&ep->poll_wait))
 997                pwake++;
 998
 999out_unlock:
1000        spin_unlock_irqrestore(&ep->lock, flags);
1001
1002        /* We have to call this outside the lock */
1003        if (pwake)
1004                ep_poll_safewake(&ep->poll_wait);
1005
1006        return 1;
1007}
1008
1009/*
1010 * This is the callback that is used to add our wait queue to the
1011 * target file wakeup lists.
1012 */
1013static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1014                                 poll_table *pt)
1015{
1016        struct epitem *epi = ep_item_from_epqueue(pt);
1017        struct eppoll_entry *pwq;
1018
1019        if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1020                init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1021                pwq->whead = whead;
1022                pwq->base = epi;
1023                add_wait_queue(whead, &pwq->wait);
1024                list_add_tail(&pwq->llink, &epi->pwqlist);
1025                epi->nwait++;
1026        } else {
1027                /* We have to signal that an error occurred */
1028                epi->nwait = -1;
1029        }
1030}
1031
1032static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1033{
1034        int kcmp;
1035        struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1036        struct epitem *epic;
1037
1038        while (*p) {
1039                parent = *p;
1040                epic = rb_entry(parent, struct epitem, rbn);
1041                kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1042                if (kcmp > 0)
1043                        p = &parent->rb_right;
1044                else
1045                        p = &parent->rb_left;
1046        }
1047        rb_link_node(&epi->rbn, parent, p);
1048        rb_insert_color(&epi->rbn, &ep->rbr);
1049}
1050
1051
1052
1053#define PATH_ARR_SIZE 5
1054/*
1055 * These are the number paths of length 1 to 5, that we are allowing to emanate
1056 * from a single file of interest. For example, we allow 1000 paths of length
1057 * 1, to emanate from each file of interest. This essentially represents the
1058 * potential wakeup paths, which need to be limited in order to avoid massive
1059 * uncontrolled wakeup storms. The common use case should be a single ep which
1060 * is connected to n file sources. In this case each file source has 1 path
1061 * of length 1. Thus, the numbers below should be more than sufficient. These
1062 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1063 * and delete can't add additional paths. Protected by the epmutex.
1064 */
1065static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1066static int path_count[PATH_ARR_SIZE];
1067
1068static int path_count_inc(int nests)
1069{
1070        /* Allow an arbitrary number of depth 1 paths */
1071        if (nests == 0)
1072                return 0;
1073
1074        if (++path_count[nests] > path_limits[nests])
1075                return -1;
1076        return 0;
1077}
1078
1079static void path_count_init(void)
1080{
1081        int i;
1082
1083        for (i = 0; i < PATH_ARR_SIZE; i++)
1084                path_count[i] = 0;
1085}
1086
1087static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1088{
1089        int error = 0;
1090        struct file *file = priv;
1091        struct file *child_file;
1092        struct epitem *epi;
1093
1094        list_for_each_entry(epi, &file->f_ep_links, fllink) {
1095                child_file = epi->ep->file;
1096                if (is_file_epoll(child_file)) {
1097                        if (list_empty(&child_file->f_ep_links)) {
1098                                if (path_count_inc(call_nests)) {
1099                                        error = -1;
1100                                        break;
1101                                }
1102                        } else {
1103                                error = ep_call_nested(&poll_loop_ncalls,
1104                                                        EP_MAX_NESTS,
1105                                                        reverse_path_check_proc,
1106                                                        child_file, child_file,
1107                                                        current);
1108                        }
1109                        if (error != 0)
1110                                break;
1111                } else {
1112                        printk(KERN_ERR "reverse_path_check_proc: "
1113                                "file is not an ep!\n");
1114                }
1115        }
1116        return error;
1117}
1118
1119/**
1120 * reverse_path_check - The tfile_check_list is list of file *, which have
1121 *                      links that are proposed to be newly added. We need to
1122 *                      make sure that those added links don't add too many
1123 *                      paths such that we will spend all our time waking up
1124 *                      eventpoll objects.
1125 *
1126 * Returns: Returns zero if the proposed links don't create too many paths,
1127 *          -1 otherwise.
1128 */
1129static int reverse_path_check(void)
1130{
1131        int error = 0;
1132        struct file *current_file;
1133
1134        /* let's call this for all tfiles */
1135        list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1136                path_count_init();
1137                error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1138                                        reverse_path_check_proc, current_file,
1139                                        current_file, current);
1140                if (error)
1141                        break;
1142        }
1143        return error;
1144}
1145
1146static int ep_create_wakeup_source(struct epitem *epi)
1147{
1148        const char *name;
1149
1150        if (!epi->ep->ws) {
1151                epi->ep->ws = wakeup_source_register("eventpoll");
1152                if (!epi->ep->ws)
1153                        return -ENOMEM;
1154        }
1155
1156        name = epi->ffd.file->f_path.dentry->d_name.name;
1157        epi->ws = wakeup_source_register(name);
1158        if (!epi->ws)
1159                return -ENOMEM;
1160
1161        return 0;
1162}
1163
1164static void ep_destroy_wakeup_source(struct epitem *epi)
1165{
1166        wakeup_source_unregister(epi->ws);
1167        epi->ws = NULL;
1168}
1169
1170/*
1171 * Must be called with "mtx" held.
1172 */
1173static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1174                     struct file *tfile, int fd)
1175{
1176        int error, revents, pwake = 0;
1177        unsigned long flags;
1178        long user_watches;
1179        struct epitem *epi;
1180        struct ep_pqueue epq;
1181
1182        user_watches = atomic_long_read(&ep->user->epoll_watches);
1183        if (unlikely(user_watches >= max_user_watches))
1184                return -ENOSPC;
1185        if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1186                return -ENOMEM;
1187
1188        /* Item initialization follow here ... */
1189        INIT_LIST_HEAD(&epi->rdllink);
1190        INIT_LIST_HEAD(&epi->fllink);
1191        INIT_LIST_HEAD(&epi->pwqlist);
1192        epi->ep = ep;
1193        ep_set_ffd(&epi->ffd, tfile, fd);
1194        epi->event = *event;
1195        epi->nwait = 0;
1196        epi->next = EP_UNACTIVE_PTR;
1197        if (epi->event.events & EPOLLWAKEUP) {
1198                error = ep_create_wakeup_source(epi);
1199                if (error)
1200                        goto error_create_wakeup_source;
1201        } else {
1202                epi->ws = NULL;
1203        }
1204
1205        /* Initialize the poll table using the queue callback */
1206        epq.epi = epi;
1207        init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1208        epq.pt._key = event->events;
1209
1210        /*
1211         * Attach the item to the poll hooks and get current event bits.
1212         * We can safely use the file* here because its usage count has
1213         * been increased by the caller of this function. Note that after
1214         * this operation completes, the poll callback can start hitting
1215         * the new item.
1216         */
1217        revents = tfile->f_op->poll(tfile, &epq.pt);
1218
1219        /*
1220         * We have to check if something went wrong during the poll wait queue
1221         * install process. Namely an allocation for a wait queue failed due
1222         * high memory pressure.
1223         */
1224        error = -ENOMEM;
1225        if (epi->nwait < 0)
1226                goto error_unregister;
1227
1228        /* Add the current item to the list of active epoll hook for this file */
1229        spin_lock(&tfile->f_lock);
1230        list_add_tail(&epi->fllink, &tfile->f_ep_links);
1231        spin_unlock(&tfile->f_lock);
1232
1233        /*
1234         * Add the current item to the RB tree. All RB tree operations are
1235         * protected by "mtx", and ep_insert() is called with "mtx" held.
1236         */
1237        ep_rbtree_insert(ep, epi);
1238
1239        /* now check if we've created too many backpaths */
1240        error = -EINVAL;
1241        if (reverse_path_check())
1242                goto error_remove_epi;
1243
1244        /* We have to drop the new item inside our item list to keep track of it */
1245        spin_lock_irqsave(&ep->lock, flags);
1246
1247        /* If the file is already "ready" we drop it inside the ready list */
1248        if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1249                list_add_tail(&epi->rdllink, &ep->rdllist);
1250                __pm_stay_awake(epi->ws);
1251
1252                /* Notify waiting tasks that events are available */
1253                if (waitqueue_active(&ep->wq))
1254                        wake_up_locked(&ep->wq);
1255                if (waitqueue_active(&ep->poll_wait))
1256                        pwake++;
1257        }
1258
1259        spin_unlock_irqrestore(&ep->lock, flags);
1260
1261        atomic_long_inc(&ep->user->epoll_watches);
1262
1263        /* We have to call this outside the lock */
1264        if (pwake)
1265                ep_poll_safewake(&ep->poll_wait);
1266
1267        return 0;
1268
1269error_remove_epi:
1270        spin_lock(&tfile->f_lock);
1271        if (ep_is_linked(&epi->fllink))
1272                list_del_init(&epi->fllink);
1273        spin_unlock(&tfile->f_lock);
1274
1275        rb_erase(&epi->rbn, &ep->rbr);
1276
1277error_unregister:
1278        ep_unregister_pollwait(ep, epi);
1279
1280        /*
1281         * We need to do this because an event could have been arrived on some
1282         * allocated wait queue. Note that we don't care about the ep->ovflist
1283         * list, since that is used/cleaned only inside a section bound by "mtx".
1284         * And ep_insert() is called with "mtx" held.
1285         */
1286        spin_lock_irqsave(&ep->lock, flags);
1287        if (ep_is_linked(&epi->rdllink))
1288                list_del_init(&epi->rdllink);
1289        spin_unlock_irqrestore(&ep->lock, flags);
1290
1291        wakeup_source_unregister(epi->ws);
1292
1293error_create_wakeup_source:
1294        kmem_cache_free(epi_cache, epi);
1295
1296        return error;
1297}
1298
1299/*
1300 * Modify the interest event mask by dropping an event if the new mask
1301 * has a match in the current file status. Must be called with "mtx" held.
1302 */
1303static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1304{
1305        int pwake = 0;
1306        unsigned int revents;
1307        poll_table pt;
1308
1309        init_poll_funcptr(&pt, NULL);
1310
1311        /*
1312         * Set the new event interest mask before calling f_op->poll();
1313         * otherwise we might miss an event that happens between the
1314         * f_op->poll() call and the new event set registering.
1315         */
1316        epi->event.events = event->events; /* need barrier below */
1317        pt._key = event->events;
1318        epi->event.data = event->data; /* protected by mtx */
1319        if (epi->event.events & EPOLLWAKEUP) {
1320                if (!epi->ws)
1321                        ep_create_wakeup_source(epi);
1322        } else if (epi->ws) {
1323                ep_destroy_wakeup_source(epi);
1324        }
1325
1326        /*
1327         * The following barrier has two effects:
1328         *
1329         * 1) Flush epi changes above to other CPUs.  This ensures
1330         *    we do not miss events from ep_poll_callback if an
1331         *    event occurs immediately after we call f_op->poll().
1332         *    We need this because we did not take ep->lock while
1333         *    changing epi above (but ep_poll_callback does take
1334         *    ep->lock).
1335         *
1336         * 2) We also need to ensure we do not miss _past_ events
1337         *    when calling f_op->poll().  This barrier also
1338         *    pairs with the barrier in wq_has_sleeper (see
1339         *    comments for wq_has_sleeper).
1340         *
1341         * This barrier will now guarantee ep_poll_callback or f_op->poll
1342         * (or both) will notice the readiness of an item.
1343         */
1344        smp_mb();
1345
1346        /*
1347         * Get current event bits. We can safely use the file* here because
1348         * its usage count has been increased by the caller of this function.
1349         */
1350        revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt);
1351
1352        /*
1353         * If the item is "hot" and it is not registered inside the ready
1354         * list, push it inside.
1355         */
1356        if (revents & event->events) {
1357                spin_lock_irq(&ep->lock);
1358                if (!ep_is_linked(&epi->rdllink)) {
1359                        list_add_tail(&epi->rdllink, &ep->rdllist);
1360                        __pm_stay_awake(epi->ws);
1361
1362                        /* Notify waiting tasks that events are available */
1363                        if (waitqueue_active(&ep->wq))
1364                                wake_up_locked(&ep->wq);
1365                        if (waitqueue_active(&ep->poll_wait))
1366                                pwake++;
1367                }
1368                spin_unlock_irq(&ep->lock);
1369        }
1370
1371        /* We have to call this outside the lock */
1372        if (pwake)
1373                ep_poll_safewake(&ep->poll_wait);
1374
1375        return 0;
1376}
1377
1378static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1379                               void *priv)
1380{
1381        struct ep_send_events_data *esed = priv;
1382        int eventcnt;
1383        unsigned int revents;
1384        struct epitem *epi;
1385        struct epoll_event __user *uevent;
1386        poll_table pt;
1387
1388        init_poll_funcptr(&pt, NULL);
1389
1390        /*
1391         * We can loop without lock because we are passed a task private list.
1392         * Items cannot vanish during the loop because ep_scan_ready_list() is
1393         * holding "mtx" during this call.
1394         */
1395        for (eventcnt = 0, uevent = esed->events;
1396             !list_empty(head) && eventcnt < esed->maxevents;) {
1397                epi = list_first_entry(head, struct epitem, rdllink);
1398
1399                /*
1400                 * Activate ep->ws before deactivating epi->ws to prevent
1401                 * triggering auto-suspend here (in case we reactive epi->ws
1402                 * below).
1403                 *
1404                 * This could be rearranged to delay the deactivation of epi->ws
1405                 * instead, but then epi->ws would temporarily be out of sync
1406                 * with ep_is_linked().
1407                 */
1408                if (epi->ws && epi->ws->active)
1409                        __pm_stay_awake(ep->ws);
1410                __pm_relax(epi->ws);
1411                list_del_init(&epi->rdllink);
1412
1413                pt._key = epi->event.events;
1414                revents = epi->ffd.file->f_op->poll(epi->ffd.file, &pt) &
1415                        epi->event.events;
1416
1417                /*
1418                 * If the event mask intersect the caller-requested one,
1419                 * deliver the event to userspace. Again, ep_scan_ready_list()
1420                 * is holding "mtx", so no operations coming from userspace
1421                 * can change the item.
1422                 */
1423                if (revents) {
1424                        if (__put_user(revents, &uevent->events) ||
1425                            __put_user(epi->event.data, &uevent->data)) {
1426                                list_add(&epi->rdllink, head);
1427                                __pm_stay_awake(epi->ws);
1428                                return eventcnt ? eventcnt : -EFAULT;
1429                        }
1430                        eventcnt++;
1431                        uevent++;
1432                        if (epi->event.events & EPOLLONESHOT)
1433                                epi->event.events &= EP_PRIVATE_BITS;
1434                        else if (!(epi->event.events & EPOLLET)) {
1435                                /*
1436                                 * If this file has been added with Level
1437                                 * Trigger mode, we need to insert back inside
1438                                 * the ready list, so that the next call to
1439                                 * epoll_wait() will check again the events
1440                                 * availability. At this point, no one can insert
1441                                 * into ep->rdllist besides us. The epoll_ctl()
1442                                 * callers are locked out by
1443                                 * ep_scan_ready_list() holding "mtx" and the
1444                                 * poll callback will queue them in ep->ovflist.
1445                                 */
1446                                list_add_tail(&epi->rdllink, &ep->rdllist);
1447                                __pm_stay_awake(epi->ws);
1448                        }
1449                }
1450        }
1451
1452        return eventcnt;
1453}
1454
1455static int ep_send_events(struct eventpoll *ep,
1456                          struct epoll_event __user *events, int maxevents)
1457{
1458        struct ep_send_events_data esed;
1459
1460        esed.maxevents = maxevents;
1461        esed.events = events;
1462
1463        return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0);
1464}
1465
1466static inline struct timespec ep_set_mstimeout(long ms)
1467{
1468        struct timespec now, ts = {
1469                .tv_sec = ms / MSEC_PER_SEC,
1470                .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1471        };
1472
1473        ktime_get_ts(&now);
1474        return timespec_add_safe(now, ts);
1475}
1476
1477/**
1478 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1479 *           event buffer.
1480 *
1481 * @ep: Pointer to the eventpoll context.
1482 * @events: Pointer to the userspace buffer where the ready events should be
1483 *          stored.
1484 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1485 * @timeout: Maximum timeout for the ready events fetch operation, in
1486 *           milliseconds. If the @timeout is zero, the function will not block,
1487 *           while if the @timeout is less than zero, the function will block
1488 *           until at least one event has been retrieved (or an error
1489 *           occurred).
1490 *
1491 * Returns: Returns the number of ready events which have been fetched, or an
1492 *          error code, in case of error.
1493 */
1494static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1495                   int maxevents, long timeout)
1496{
1497        int res = 0, eavail, timed_out = 0;
1498        unsigned long flags;
1499        long slack = 0;
1500        wait_queue_t wait;
1501        ktime_t expires, *to = NULL;
1502
1503        if (timeout > 0) {
1504                struct timespec end_time = ep_set_mstimeout(timeout);
1505
1506                slack = select_estimate_accuracy(&end_time);
1507                to = &expires;
1508                *to = timespec_to_ktime(end_time);
1509        } else if (timeout == 0) {
1510                /*
1511                 * Avoid the unnecessary trip to the wait queue loop, if the
1512                 * caller specified a non blocking operation.
1513                 */
1514                timed_out = 1;
1515                spin_lock_irqsave(&ep->lock, flags);
1516                goto check_events;
1517        }
1518
1519fetch_events:
1520        spin_lock_irqsave(&ep->lock, flags);
1521
1522        if (!ep_events_available(ep)) {
1523                /*
1524                 * We don't have any available event to return to the caller.
1525                 * We need to sleep here, and we will be wake up by
1526                 * ep_poll_callback() when events will become available.
1527                 */
1528                init_waitqueue_entry(&wait, current);
1529                __add_wait_queue_exclusive(&ep->wq, &wait);
1530
1531                for (;;) {
1532                        /*
1533                         * We don't want to sleep if the ep_poll_callback() sends us
1534                         * a wakeup in between. That's why we set the task state
1535                         * to TASK_INTERRUPTIBLE before doing the checks.
1536                         */
1537                        set_current_state(TASK_INTERRUPTIBLE);
1538                        if (ep_events_available(ep) || timed_out)
1539                                break;
1540                        if (signal_pending(current)) {
1541                                res = -EINTR;
1542                                break;
1543                        }
1544
1545                        spin_unlock_irqrestore(&ep->lock, flags);
1546                        if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1547                                timed_out = 1;
1548
1549                        spin_lock_irqsave(&ep->lock, flags);
1550                }
1551                __remove_wait_queue(&ep->wq, &wait);
1552
1553                set_current_state(TASK_RUNNING);
1554        }
1555check_events:
1556        /* Is it worth to try to dig for events ? */
1557        eavail = ep_events_available(ep);
1558
1559        spin_unlock_irqrestore(&ep->lock, flags);
1560
1561        /*
1562         * Try to transfer events to user space. In case we get 0 events and
1563         * there's still timeout left over, we go trying again in search of
1564         * more luck.
1565         */
1566        if (!res && eavail &&
1567            !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1568                goto fetch_events;
1569
1570        return res;
1571}
1572
1573/**
1574 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1575 *                      API, to verify that adding an epoll file inside another
1576 *                      epoll structure, does not violate the constraints, in
1577 *                      terms of closed loops, or too deep chains (which can
1578 *                      result in excessive stack usage).
1579 *
1580 * @priv: Pointer to the epoll file to be currently checked.
1581 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1582 *          data structure pointer.
1583 * @call_nests: Current dept of the @ep_call_nested() call stack.
1584 *
1585 * Returns: Returns zero if adding the epoll @file inside current epoll
1586 *          structure @ep does not violate the constraints, or -1 otherwise.
1587 */
1588static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1589{
1590        int error = 0;
1591        struct file *file = priv;
1592        struct eventpoll *ep = file->private_data;
1593        struct eventpoll *ep_tovisit;
1594        struct rb_node *rbp;
1595        struct epitem *epi;
1596
1597        mutex_lock_nested(&ep->mtx, call_nests + 1);
1598        ep->visited = 1;
1599        list_add(&ep->visited_list_link, &visited_list);
1600        for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1601                epi = rb_entry(rbp, struct epitem, rbn);
1602                if (unlikely(is_file_epoll(epi->ffd.file))) {
1603                        ep_tovisit = epi->ffd.file->private_data;
1604                        if (ep_tovisit->visited)
1605                                continue;
1606                        error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1607                                        ep_loop_check_proc, epi->ffd.file,
1608                                        ep_tovisit, current);
1609                        if (error != 0)
1610                                break;
1611                } else {
1612                        /*
1613                         * If we've reached a file that is not associated with
1614                         * an ep, then we need to check if the newly added
1615                         * links are going to add too many wakeup paths. We do
1616                         * this by adding it to the tfile_check_list, if it's
1617                         * not already there, and calling reverse_path_check()
1618                         * during ep_insert().
1619                         */
1620                        if (list_empty(&epi->ffd.file->f_tfile_llink))
1621                                list_add(&epi->ffd.file->f_tfile_llink,
1622                                         &tfile_check_list);
1623                }
1624        }
1625        mutex_unlock(&ep->mtx);
1626
1627        return error;
1628}
1629
1630/**
1631 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1632 *                 another epoll file (represented by @ep) does not create
1633 *                 closed loops or too deep chains.
1634 *
1635 * @ep: Pointer to the epoll private data structure.
1636 * @file: Pointer to the epoll file to be checked.
1637 *
1638 * Returns: Returns zero if adding the epoll @file inside current epoll
1639 *          structure @ep does not violate the constraints, or -1 otherwise.
1640 */
1641static int ep_loop_check(struct eventpoll *ep, struct file *file)
1642{
1643        int ret;
1644        struct eventpoll *ep_cur, *ep_next;
1645
1646        ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1647                              ep_loop_check_proc, file, ep, current);
1648        /* clear visited list */
1649        list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1650                                                        visited_list_link) {
1651                ep_cur->visited = 0;
1652                list_del(&ep_cur->visited_list_link);
1653        }
1654        return ret;
1655}
1656
1657static void clear_tfile_check_list(void)
1658{
1659        struct file *file;
1660
1661        /* first clear the tfile_check_list */
1662        while (!list_empty(&tfile_check_list)) {
1663                file = list_first_entry(&tfile_check_list, struct file,
1664                                        f_tfile_llink);
1665                list_del_init(&file->f_tfile_llink);
1666        }
1667        INIT_LIST_HEAD(&tfile_check_list);
1668}
1669
1670/*
1671 * Open an eventpoll file descriptor.
1672 */
1673SYSCALL_DEFINE1(epoll_create1, int, flags)
1674{
1675        int error, fd;
1676        struct eventpoll *ep = NULL;
1677        struct file *file;
1678
1679        /* Check the EPOLL_* constant for consistency.  */
1680        BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1681
1682        if (flags & ~EPOLL_CLOEXEC)
1683                return -EINVAL;
1684        /*
1685         * Create the internal data structure ("struct eventpoll").
1686         */
1687        error = ep_alloc(&ep);
1688        if (error < 0)
1689                return error;
1690        /*
1691         * Creates all the items needed to setup an eventpoll file. That is,
1692         * a file structure and a free file descriptor.
1693         */
1694        fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1695        if (fd < 0) {
1696                error = fd;
1697                goto out_free_ep;
1698        }
1699        file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1700                                 O_RDWR | (flags & O_CLOEXEC));
1701        if (IS_ERR(file)) {
1702                error = PTR_ERR(file);
1703                goto out_free_fd;
1704        }
1705        ep->file = file;
1706        fd_install(fd, file);
1707        return fd;
1708
1709out_free_fd:
1710        put_unused_fd(fd);
1711out_free_ep:
1712        ep_free(ep);
1713        return error;
1714}
1715
1716SYSCALL_DEFINE1(epoll_create, int, size)
1717{
1718        if (size <= 0)
1719                return -EINVAL;
1720
1721        return sys_epoll_create1(0);
1722}
1723
1724/*
1725 * The following function implements the controller interface for
1726 * the eventpoll file that enables the insertion/removal/change of
1727 * file descriptors inside the interest set.
1728 */
1729SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1730                struct epoll_event __user *, event)
1731{
1732        int error;
1733        int did_lock_epmutex = 0;
1734        struct file *file, *tfile;
1735        struct eventpoll *ep;
1736        struct epitem *epi;
1737        struct epoll_event epds;
1738
1739        error = -EFAULT;
1740        if (ep_op_has_event(op) &&
1741            copy_from_user(&epds, event, sizeof(struct epoll_event)))
1742                goto error_return;
1743
1744        /* Get the "struct file *" for the eventpoll file */
1745        error = -EBADF;
1746        file = fget(epfd);
1747        if (!file)
1748                goto error_return;
1749
1750        /* Get the "struct file *" for the target file */
1751        tfile = fget(fd);
1752        if (!tfile)
1753                goto error_fput;
1754
1755        /* The target file descriptor must support poll */
1756        error = -EPERM;
1757        if (!tfile->f_op || !tfile->f_op->poll)
1758                goto error_tgt_fput;
1759
1760        /* Check if EPOLLWAKEUP is allowed */
1761        if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
1762                epds.events &= ~EPOLLWAKEUP;
1763
1764        /*
1765         * We have to check that the file structure underneath the file descriptor
1766         * the user passed to us _is_ an eventpoll file. And also we do not permit
1767         * adding an epoll file descriptor inside itself.
1768         */
1769        error = -EINVAL;
1770        if (file == tfile || !is_file_epoll(file))
1771                goto error_tgt_fput;
1772
1773        /*
1774         * At this point it is safe to assume that the "private_data" contains
1775         * our own data structure.
1776         */
1777        ep = file->private_data;
1778
1779        /*
1780         * When we insert an epoll file descriptor, inside another epoll file
1781         * descriptor, there is the change of creating closed loops, which are
1782         * better be handled here, than in more critical paths. While we are
1783         * checking for loops we also determine the list of files reachable
1784         * and hang them on the tfile_check_list, so we can check that we
1785         * haven't created too many possible wakeup paths.
1786         *
1787         * We need to hold the epmutex across both ep_insert and ep_remove
1788         * b/c we want to make sure we are looking at a coherent view of
1789         * epoll network.
1790         */
1791        if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) {
1792                mutex_lock(&epmutex);
1793                did_lock_epmutex = 1;
1794        }
1795        if (op == EPOLL_CTL_ADD) {
1796                if (is_file_epoll(tfile)) {
1797                        error = -ELOOP;
1798                        if (ep_loop_check(ep, tfile) != 0) {
1799                                clear_tfile_check_list();
1800                                goto error_tgt_fput;
1801                        }
1802                } else
1803                        list_add(&tfile->f_tfile_llink, &tfile_check_list);
1804        }
1805
1806        mutex_lock_nested(&ep->mtx, 0);
1807
1808        /*
1809         * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1810         * above, we can be sure to be able to use the item looked up by
1811         * ep_find() till we release the mutex.
1812         */
1813        epi = ep_find(ep, tfile, fd);
1814
1815        error = -EINVAL;
1816        switch (op) {
1817        case EPOLL_CTL_ADD:
1818                if (!epi) {
1819                        epds.events |= POLLERR | POLLHUP;
1820                        error = ep_insert(ep, &epds, tfile, fd);
1821                } else
1822                        error = -EEXIST;
1823                clear_tfile_check_list();
1824                break;
1825        case EPOLL_CTL_DEL:
1826                if (epi)
1827                        error = ep_remove(ep, epi);
1828                else
1829                        error = -ENOENT;
1830                break;
1831        case EPOLL_CTL_MOD:
1832                if (epi) {
1833                        epds.events |= POLLERR | POLLHUP;
1834                        error = ep_modify(ep, epi, &epds);
1835                } else
1836                        error = -ENOENT;
1837                break;
1838        }
1839        mutex_unlock(&ep->mtx);
1840
1841error_tgt_fput:
1842        if (did_lock_epmutex)
1843                mutex_unlock(&epmutex);
1844
1845        fput(tfile);
1846error_fput:
1847        fput(file);
1848error_return:
1849
1850        return error;
1851}
1852
1853/*
1854 * Implement the event wait interface for the eventpoll file. It is the kernel
1855 * part of the user space epoll_wait(2).
1856 */
1857SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1858                int, maxevents, int, timeout)
1859{
1860        int error;
1861        struct fd f;
1862        struct eventpoll *ep;
1863
1864        /* The maximum number of event must be greater than zero */
1865        if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1866                return -EINVAL;
1867
1868        /* Verify that the area passed by the user is writeable */
1869        if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1870                return -EFAULT;
1871
1872        /* Get the "struct file *" for the eventpoll file */
1873        f = fdget(epfd);
1874        if (!f.file)
1875                return -EBADF;
1876
1877        /*
1878         * We have to check that the file structure underneath the fd
1879         * the user passed to us _is_ an eventpoll file.
1880         */
1881        error = -EINVAL;
1882        if (!is_file_epoll(f.file))
1883                goto error_fput;
1884
1885        /*
1886         * At this point it is safe to assume that the "private_data" contains
1887         * our own data structure.
1888         */
1889        ep = f.file->private_data;
1890
1891        /* Time to fish for events ... */
1892        error = ep_poll(ep, events, maxevents, timeout);
1893
1894error_fput:
1895        fdput(f);
1896        return error;
1897}
1898
1899/*
1900 * Implement the event wait interface for the eventpoll file. It is the kernel
1901 * part of the user space epoll_pwait(2).
1902 */
1903SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1904                int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1905                size_t, sigsetsize)
1906{
1907        int error;
1908        sigset_t ksigmask, sigsaved;
1909
1910        /*
1911         * If the caller wants a certain signal mask to be set during the wait,
1912         * we apply it here.
1913         */
1914        if (sigmask) {
1915                if (sigsetsize != sizeof(sigset_t))
1916                        return -EINVAL;
1917                if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1918                        return -EFAULT;
1919                sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1920                sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1921        }
1922
1923        error = sys_epoll_wait(epfd, events, maxevents, timeout);
1924
1925        /*
1926         * If we changed the signal mask, we need to restore the original one.
1927         * In case we've got a signal while waiting, we do not restore the
1928         * signal mask yet, and we allow do_signal() to deliver the signal on
1929         * the way back to userspace, before the signal mask is restored.
1930         */
1931        if (sigmask) {
1932                if (error == -EINTR) {
1933                        memcpy(&current->saved_sigmask, &sigsaved,
1934                               sizeof(sigsaved));
1935                        set_restore_sigmask();
1936                } else
1937                        sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1938        }
1939
1940        return error;
1941}
1942
1943static int __init eventpoll_init(void)
1944{
1945        struct sysinfo si;
1946
1947        si_meminfo(&si);
1948        /*
1949         * Allows top 4% of lomem to be allocated for epoll watches (per user).
1950         */
1951        max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1952                EP_ITEM_COST;
1953        BUG_ON(max_user_watches < 0);
1954
1955        /*
1956         * Initialize the structure used to perform epoll file descriptor
1957         * inclusion loops checks.
1958         */
1959        ep_nested_calls_init(&poll_loop_ncalls);
1960
1961        /* Initialize the structure used to perform safe poll wait head wake ups */
1962        ep_nested_calls_init(&poll_safewake_ncalls);
1963
1964        /* Initialize the structure used to perform file's f_op->poll() calls */
1965        ep_nested_calls_init(&poll_readywalk_ncalls);
1966
1967        /* Allocates slab cache used to allocate "struct epitem" items */
1968        epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1969                        0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1970
1971        /* Allocates slab cache used to allocate "struct eppoll_entry" */
1972        pwq_cache = kmem_cache_create("eventpoll_pwq",
1973                        sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1974
1975        return 0;
1976}
1977fs_initcall(eventpoll_init);
1978
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.