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#include <linux/compat.h>
  44#include <linux/rculist.h>
  45
  46/*
  47 * LOCKING:
  48 * There are three level of locking required by epoll :
  49 *
  50 * 1) epmutex (mutex)
  51 * 2) ep->mtx (mutex)
  52 * 3) ep->lock (spinlock)
  53 *
  54 * The acquire order is the one listed above, from 1 to 3.
  55 * We need a spinlock (ep->lock) because we manipulate objects
  56 * from inside the poll callback, that might be triggered from
  57 * a wake_up() that in turn might be called from IRQ context.
  58 * So we can't sleep inside the poll callback and hence we need
  59 * a spinlock. During the event transfer loop (from kernel to
  60 * user space) we could end up sleeping due a copy_to_user(), so
  61 * we need a lock that will allow us to sleep. This lock is a
  62 * mutex (ep->mtx). It is acquired during the event transfer loop,
  63 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
  64 * Then we also need a global mutex to serialize eventpoll_release_file()
  65 * and ep_free().
  66 * This mutex is acquired by ep_free() during the epoll file
  67 * cleanup path and it is also acquired by eventpoll_release_file()
  68 * if a file has been pushed inside an epoll set and it is then
  69 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
  70 * It is also acquired when inserting an epoll fd onto another epoll
  71 * fd. We do this so that we walk the epoll tree and ensure that this
  72 * insertion does not create a cycle of epoll file descriptors, which
  73 * could lead to deadlock. We need a global mutex to prevent two
  74 * simultaneous inserts (A into B and B into A) from racing and
  75 * constructing a cycle without either insert observing that it is
  76 * going to.
  77 * It is necessary to acquire multiple "ep->mtx"es at once in the
  78 * case when one epoll fd is added to another. In this case, we
  79 * always acquire the locks in the order of nesting (i.e. after
  80 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
  81 * before e2->mtx). Since we disallow cycles of epoll file
  82 * descriptors, this ensures that the mutexes are well-ordered. In
  83 * order to communicate this nesting to lockdep, when walking a tree
  84 * of epoll file descriptors, we use the current recursion depth as
  85 * the lockdep subkey.
  86 * It is possible to drop the "ep->mtx" and to use the global
  87 * mutex "epmutex" (together with "ep->lock") to have it working,
  88 * but having "ep->mtx" will make the interface more scalable.
  89 * Events that require holding "epmutex" are very rare, while for
  90 * normal operations the epoll private "ep->mtx" will guarantee
  91 * a better scalability.
  92 */
  93
  94/* Epoll private bits inside the event mask */
  95#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET)
  96
  97/* Maximum number of nesting allowed inside epoll sets */
  98#define EP_MAX_NESTS 4
  99
 100#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
 101
 102#define EP_UNACTIVE_PTR ((void *) -1L)
 103
 104#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
 105
 106struct epoll_filefd {
 107        struct file *file;
 108        int fd;
 109} __packed;
 110
 111/*
 112 * Structure used to track possible nested calls, for too deep recursions
 113 * and loop cycles.
 114 */
 115struct nested_call_node {
 116        struct list_head llink;
 117        void *cookie;
 118        void *ctx;
 119};
 120
 121/*
 122 * This structure is used as collector for nested calls, to check for
 123 * maximum recursion dept and loop cycles.
 124 */
 125struct nested_calls {
 126        struct list_head tasks_call_list;
 127        spinlock_t lock;
 128};
 129
 130/*
 131 * Each file descriptor added to the eventpoll interface will
 132 * have an entry of this type linked to the "rbr" RB tree.
 133 * Avoid increasing the size of this struct, there can be many thousands
 134 * of these on a server and we do not want this to take another cache line.
 135 */
 136struct epitem {
 137        union {
 138                /* RB tree node links this structure to the eventpoll RB tree */
 139                struct rb_node rbn;
 140                /* Used to free the struct epitem */
 141                struct rcu_head rcu;
 142        };
 143
 144        /* List header used to link this structure to the eventpoll ready list */
 145        struct list_head rdllink;
 146
 147        /*
 148         * Works together "struct eventpoll"->ovflist in keeping the
 149         * single linked chain of items.
 150         */
 151        struct epitem *next;
 152
 153        /* The file descriptor information this item refers to */
 154        struct epoll_filefd ffd;
 155
 156        /* Number of active wait queue attached to poll operations */
 157        int nwait;
 158
 159        /* List containing poll wait queues */
 160        struct list_head pwqlist;
 161
 162        /* The "container" of this item */
 163        struct eventpoll *ep;
 164
 165        /* List header used to link this item to the "struct file" items list */
 166        struct list_head fllink;
 167
 168        /* wakeup_source used when EPOLLWAKEUP is set */
 169        struct wakeup_source __rcu *ws;
 170
 171        /* The structure that describe the interested events and the source fd */
 172        struct epoll_event event;
 173};
 174
 175/*
 176 * This structure is stored inside the "private_data" member of the file
 177 * structure and represents the main data structure for the eventpoll
 178 * interface.
 179 */
 180struct eventpoll {
 181        /* Protect the access to this structure */
 182        spinlock_t lock;
 183
 184        /*
 185         * This mutex is used to ensure that files are not removed
 186         * while epoll is using them. This is held during the event
 187         * collection loop, the file cleanup path, the epoll file exit
 188         * code and the ctl operations.
 189         */
 190        struct mutex mtx;
 191
 192        /* Wait queue used by sys_epoll_wait() */
 193        wait_queue_head_t wq;
 194
 195        /* Wait queue used by file->poll() */
 196        wait_queue_head_t poll_wait;
 197
 198        /* List of ready file descriptors */
 199        struct list_head rdllist;
 200
 201        /* RB tree root used to store monitored fd structs */
 202        struct rb_root rbr;
 203
 204        /*
 205         * This is a single linked list that chains all the "struct epitem" that
 206         * happened while transferring ready events to userspace w/out
 207         * holding ->lock.
 208         */
 209        struct epitem *ovflist;
 210
 211        /* wakeup_source used when ep_scan_ready_list is running */
 212        struct wakeup_source *ws;
 213
 214        /* The user that created the eventpoll descriptor */
 215        struct user_struct *user;
 216
 217        struct file *file;
 218
 219        /* used to optimize loop detection check */
 220        int visited;
 221        struct list_head visited_list_link;
 222};
 223
 224/* Wait structure used by the poll hooks */
 225struct eppoll_entry {
 226        /* List header used to link this structure to the "struct epitem" */
 227        struct list_head llink;
 228
 229        /* The "base" pointer is set to the container "struct epitem" */
 230        struct epitem *base;
 231
 232        /*
 233         * Wait queue item that will be linked to the target file wait
 234         * queue head.
 235         */
 236        wait_queue_t wait;
 237
 238        /* The wait queue head that linked the "wait" wait queue item */
 239        wait_queue_head_t *whead;
 240};
 241
 242/* Wrapper struct used by poll queueing */
 243struct ep_pqueue {
 244        poll_table pt;
 245        struct epitem *epi;
 246};
 247
 248/* Used by the ep_send_events() function as callback private data */
 249struct ep_send_events_data {
 250        int maxevents;
 251        struct epoll_event __user *events;
 252};
 253
 254/*
 255 * Configuration options available inside /proc/sys/fs/epoll/
 256 */
 257/* Maximum number of epoll watched descriptors, per user */
 258static long max_user_watches __read_mostly;
 259
 260/*
 261 * This mutex is used to serialize ep_free() and eventpoll_release_file().
 262 */
 263static DEFINE_MUTEX(epmutex);
 264
 265/* Used to check for epoll file descriptor inclusion loops */
 266static struct nested_calls poll_loop_ncalls;
 267
 268/* Used for safe wake up implementation */
 269static struct nested_calls poll_safewake_ncalls;
 270
 271/* Used to call file's f_op->poll() under the nested calls boundaries */
 272static struct nested_calls poll_readywalk_ncalls;
 273
 274/* Slab cache used to allocate "struct epitem" */
 275static struct kmem_cache *epi_cache __read_mostly;
 276
 277/* Slab cache used to allocate "struct eppoll_entry" */
 278static struct kmem_cache *pwq_cache __read_mostly;
 279
 280/* Visited nodes during ep_loop_check(), so we can unset them when we finish */
 281static LIST_HEAD(visited_list);
 282
 283/*
 284 * List of files with newly added links, where we may need to limit the number
 285 * of emanating paths. Protected by the epmutex.
 286 */
 287static LIST_HEAD(tfile_check_list);
 288
 289#ifdef CONFIG_SYSCTL
 290
 291#include <linux/sysctl.h>
 292
 293static long zero;
 294static long long_max = LONG_MAX;
 295
 296struct ctl_table epoll_table[] = {
 297        {
 298                .procname       = "max_user_watches",
 299                .data           = &max_user_watches,
 300                .maxlen         = sizeof(max_user_watches),
 301                .mode           = 0644,
 302                .proc_handler   = proc_doulongvec_minmax,
 303                .extra1         = &zero,
 304                .extra2         = &long_max,
 305        },
 306        { }
 307};
 308#endif /* CONFIG_SYSCTL */
 309
 310static const struct file_operations eventpoll_fops;
 311
 312static inline int is_file_epoll(struct file *f)
 313{
 314        return f->f_op == &eventpoll_fops;
 315}
 316
 317/* Setup the structure that is used as key for the RB tree */
 318static inline void ep_set_ffd(struct epoll_filefd *ffd,
 319                              struct file *file, int fd)
 320{
 321        ffd->file = file;
 322        ffd->fd = fd;
 323}
 324
 325/* Compare RB tree keys */
 326static inline int ep_cmp_ffd(struct epoll_filefd *p1,
 327                             struct epoll_filefd *p2)
 328{
 329        return (p1->file > p2->file ? +1:
 330                (p1->file < p2->file ? -1 : p1->fd - p2->fd));
 331}
 332
 333/* Tells us if the item is currently linked */
 334static inline int ep_is_linked(struct list_head *p)
 335{
 336        return !list_empty(p);
 337}
 338
 339static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p)
 340{
 341        return container_of(p, struct eppoll_entry, wait);
 342}
 343
 344/* Get the "struct epitem" from a wait queue pointer */
 345static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
 346{
 347        return container_of(p, struct eppoll_entry, wait)->base;
 348}
 349
 350/* Get the "struct epitem" from an epoll queue wrapper */
 351static inline struct epitem *ep_item_from_epqueue(poll_table *p)
 352{
 353        return container_of(p, struct ep_pqueue, pt)->epi;
 354}
 355
 356/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
 357static inline int ep_op_has_event(int op)
 358{
 359        return op != EPOLL_CTL_DEL;
 360}
 361
 362/* Initialize the poll safe wake up structure */
 363static void ep_nested_calls_init(struct nested_calls *ncalls)
 364{
 365        INIT_LIST_HEAD(&ncalls->tasks_call_list);
 366        spin_lock_init(&ncalls->lock);
 367}
 368
 369/**
 370 * ep_events_available - Checks if ready events might be available.
 371 *
 372 * @ep: Pointer to the eventpoll context.
 373 *
 374 * Returns: Returns a value different than zero if ready events are available,
 375 *          or zero otherwise.
 376 */
 377static inline int ep_events_available(struct eventpoll *ep)
 378{
 379        return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
 380}
 381
 382/**
 383 * ep_call_nested - Perform a bound (possibly) nested call, by checking
 384 *                  that the recursion limit is not exceeded, and that
 385 *                  the same nested call (by the meaning of same cookie) is
 386 *                  no re-entered.
 387 *
 388 * @ncalls: Pointer to the nested_calls structure to be used for this call.
 389 * @max_nests: Maximum number of allowed nesting calls.
 390 * @nproc: Nested call core function pointer.
 391 * @priv: Opaque data to be passed to the @nproc callback.
 392 * @cookie: Cookie to be used to identify this nested call.
 393 * @ctx: This instance context.
 394 *
 395 * Returns: Returns the code returned by the @nproc callback, or -1 if
 396 *          the maximum recursion limit has been exceeded.
 397 */
 398static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
 399                          int (*nproc)(void *, void *, int), void *priv,
 400                          void *cookie, void *ctx)
 401{
 402        int error, call_nests = 0;
 403        unsigned long flags;
 404        struct list_head *lsthead = &ncalls->tasks_call_list;
 405        struct nested_call_node *tncur;
 406        struct nested_call_node tnode;
 407
 408        spin_lock_irqsave(&ncalls->lock, flags);
 409
 410        /*
 411         * Try to see if the current task is already inside this wakeup call.
 412         * We use a list here, since the population inside this set is always
 413         * very much limited.
 414         */
 415        list_for_each_entry(tncur, lsthead, llink) {
 416                if (tncur->ctx == ctx &&
 417                    (tncur->cookie == cookie || ++call_nests > max_nests)) {
 418                        /*
 419                         * Ops ... loop detected or maximum nest level reached.
 420                         * We abort this wake by breaking the cycle itself.
 421                         */
 422                        error = -1;
 423                        goto out_unlock;
 424                }
 425        }
 426
 427        /* Add the current task and cookie to the list */
 428        tnode.ctx = ctx;
 429        tnode.cookie = cookie;
 430        list_add(&tnode.llink, lsthead);
 431
 432        spin_unlock_irqrestore(&ncalls->lock, flags);
 433
 434        /* Call the nested function */
 435        error = (*nproc)(priv, cookie, call_nests);
 436
 437        /* Remove the current task from the list */
 438        spin_lock_irqsave(&ncalls->lock, flags);
 439        list_del(&tnode.llink);
 440out_unlock:
 441        spin_unlock_irqrestore(&ncalls->lock, flags);
 442
 443        return error;
 444}
 445
 446/*
 447 * As described in commit 0ccf831cb lockdep: annotate epoll
 448 * the use of wait queues used by epoll is done in a very controlled
 449 * manner. Wake ups can nest inside each other, but are never done
 450 * with the same locking. For example:
 451 *
 452 *   dfd = socket(...);
 453 *   efd1 = epoll_create();
 454 *   efd2 = epoll_create();
 455 *   epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
 456 *   epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
 457 *
 458 * When a packet arrives to the device underneath "dfd", the net code will
 459 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
 460 * callback wakeup entry on that queue, and the wake_up() performed by the
 461 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
 462 * (efd1) notices that it may have some event ready, so it needs to wake up
 463 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
 464 * that ends up in another wake_up(), after having checked about the
 465 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
 466 * avoid stack blasting.
 467 *
 468 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
 469 * this special case of epoll.
 470 */
 471#ifdef CONFIG_DEBUG_LOCK_ALLOC
 472static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
 473                                     unsigned long events, int subclass)
 474{
 475        unsigned long flags;
 476
 477        spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
 478        wake_up_locked_poll(wqueue, events);
 479        spin_unlock_irqrestore(&wqueue->lock, flags);
 480}
 481#else
 482static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
 483                                     unsigned long events, int subclass)
 484{
 485        wake_up_poll(wqueue, events);
 486}
 487#endif
 488
 489static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
 490{
 491        ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
 492                          1 + call_nests);
 493        return 0;
 494}
 495
 496/*
 497 * Perform a safe wake up of the poll wait list. The problem is that
 498 * with the new callback'd wake up system, it is possible that the
 499 * poll callback is reentered from inside the call to wake_up() done
 500 * on the poll wait queue head. The rule is that we cannot reenter the
 501 * wake up code from the same task more than EP_MAX_NESTS times,
 502 * and we cannot reenter the same wait queue head at all. This will
 503 * enable to have a hierarchy of epoll file descriptor of no more than
 504 * EP_MAX_NESTS deep.
 505 */
 506static void ep_poll_safewake(wait_queue_head_t *wq)
 507{
 508        int this_cpu = get_cpu();
 509
 510        ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
 511                       ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
 512
 513        put_cpu();
 514}
 515
 516static void ep_remove_wait_queue(struct eppoll_entry *pwq)
 517{
 518        wait_queue_head_t *whead;
 519
 520        rcu_read_lock();
 521        /* If it is cleared by POLLFREE, it should be rcu-safe */
 522        whead = rcu_dereference(pwq->whead);
 523        if (whead)
 524                remove_wait_queue(whead, &pwq->wait);
 525        rcu_read_unlock();
 526}
 527
 528/*
 529 * This function unregisters poll callbacks from the associated file
 530 * descriptor.  Must be called with "mtx" held (or "epmutex" if called from
 531 * ep_free).
 532 */
 533static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
 534{
 535        struct list_head *lsthead = &epi->pwqlist;
 536        struct eppoll_entry *pwq;
 537
 538        while (!list_empty(lsthead)) {
 539                pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
 540
 541                list_del(&pwq->llink);
 542                ep_remove_wait_queue(pwq);
 543                kmem_cache_free(pwq_cache, pwq);
 544        }
 545}
 546
 547/* call only when ep->mtx is held */
 548static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
 549{
 550        return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
 551}
 552
 553/* call only when ep->mtx is held */
 554static inline void ep_pm_stay_awake(struct epitem *epi)
 555{
 556        struct wakeup_source *ws = ep_wakeup_source(epi);
 557
 558        if (ws)
 559                __pm_stay_awake(ws);
 560}
 561
 562static inline bool ep_has_wakeup_source(struct epitem *epi)
 563{
 564        return rcu_access_pointer(epi->ws) ? true : false;
 565}
 566
 567/* call when ep->mtx cannot be held (ep_poll_callback) */
 568static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
 569{
 570        struct wakeup_source *ws;
 571
 572        rcu_read_lock();
 573        ws = rcu_dereference(epi->ws);
 574        if (ws)
 575                __pm_stay_awake(ws);
 576        rcu_read_unlock();
 577}
 578
 579/**
 580 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
 581 *                      the scan code, to call f_op->poll(). Also allows for
 582 *                      O(NumReady) performance.
 583 *
 584 * @ep: Pointer to the epoll private data structure.
 585 * @sproc: Pointer to the scan callback.
 586 * @priv: Private opaque data passed to the @sproc callback.
 587 * @depth: The current depth of recursive f_op->poll calls.
 588 * @ep_locked: caller already holds ep->mtx
 589 *
 590 * Returns: The same integer error code returned by the @sproc callback.
 591 */
 592static int ep_scan_ready_list(struct eventpoll *ep,
 593                              int (*sproc)(struct eventpoll *,
 594                                           struct list_head *, void *),
 595                              void *priv, int depth, bool ep_locked)
 596{
 597        int error, pwake = 0;
 598        unsigned long flags;
 599        struct epitem *epi, *nepi;
 600        LIST_HEAD(txlist);
 601
 602        /*
 603         * We need to lock this because we could be hit by
 604         * eventpoll_release_file() and epoll_ctl().
 605         */
 606
 607        if (!ep_locked)
 608                mutex_lock_nested(&ep->mtx, depth);
 609
 610        /*
 611         * Steal the ready list, and re-init the original one to the
 612         * empty list. Also, set ep->ovflist to NULL so that events
 613         * happening while looping w/out locks, are not lost. We cannot
 614         * have the poll callback to queue directly on ep->rdllist,
 615         * because we want the "sproc" callback to be able to do it
 616         * in a lockless way.
 617         */
 618        spin_lock_irqsave(&ep->lock, flags);
 619        list_splice_init(&ep->rdllist, &txlist);
 620        ep->ovflist = NULL;
 621        spin_unlock_irqrestore(&ep->lock, flags);
 622
 623        /*
 624         * Now call the callback function.
 625         */
 626        error = (*sproc)(ep, &txlist, priv);
 627
 628        spin_lock_irqsave(&ep->lock, flags);
 629        /*
 630         * During the time we spent inside the "sproc" callback, some
 631         * other events might have been queued by the poll callback.
 632         * We re-insert them inside the main ready-list here.
 633         */
 634        for (nepi = ep->ovflist; (epi = nepi) != NULL;
 635             nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
 636                /*
 637                 * We need to check if the item is already in the list.
 638                 * During the "sproc" callback execution time, items are
 639                 * queued into ->ovflist but the "txlist" might already
 640                 * contain them, and the list_splice() below takes care of them.
 641                 */
 642                if (!ep_is_linked(&epi->rdllink)) {
 643                        list_add_tail(&epi->rdllink, &ep->rdllist);
 644                        ep_pm_stay_awake(epi);
 645                }
 646        }
 647        /*
 648         * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
 649         * releasing the lock, events will be queued in the normal way inside
 650         * ep->rdllist.
 651         */
 652        ep->ovflist = EP_UNACTIVE_PTR;
 653
 654        /*
 655         * Quickly re-inject items left on "txlist".
 656         */
 657        list_splice(&txlist, &ep->rdllist);
 658        __pm_relax(ep->ws);
 659
 660        if (!list_empty(&ep->rdllist)) {
 661                /*
 662                 * Wake up (if active) both the eventpoll wait list and
 663                 * the ->poll() wait list (delayed after we release the lock).
 664                 */
 665                if (waitqueue_active(&ep->wq))
 666                        wake_up_locked(&ep->wq);
 667                if (waitqueue_active(&ep->poll_wait))
 668                        pwake++;
 669        }
 670        spin_unlock_irqrestore(&ep->lock, flags);
 671
 672        if (!ep_locked)
 673                mutex_unlock(&ep->mtx);
 674
 675        /* We have to call this outside the lock */
 676        if (pwake)
 677                ep_poll_safewake(&ep->poll_wait);
 678
 679        return error;
 680}
 681
 682static void epi_rcu_free(struct rcu_head *head)
 683{
 684        struct epitem *epi = container_of(head, struct epitem, rcu);
 685        kmem_cache_free(epi_cache, epi);
 686}
 687
 688/*
 689 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
 690 * all the associated resources. Must be called with "mtx" held.
 691 */
 692static int ep_remove(struct eventpoll *ep, struct epitem *epi)
 693{
 694        unsigned long flags;
 695        struct file *file = epi->ffd.file;
 696
 697        /*
 698         * Removes poll wait queue hooks. We _have_ to do this without holding
 699         * the "ep->lock" otherwise a deadlock might occur. This because of the
 700         * sequence of the lock acquisition. Here we do "ep->lock" then the wait
 701         * queue head lock when unregistering the wait queue. The wakeup callback
 702         * will run by holding the wait queue head lock and will call our callback
 703         * that will try to get "ep->lock".
 704         */
 705        ep_unregister_pollwait(ep, epi);
 706
 707        /* Remove the current item from the list of epoll hooks */
 708        spin_lock(&file->f_lock);
 709        list_del_rcu(&epi->fllink);
 710        spin_unlock(&file->f_lock);
 711
 712        rb_erase(&epi->rbn, &ep->rbr);
 713
 714        spin_lock_irqsave(&ep->lock, flags);
 715        if (ep_is_linked(&epi->rdllink))
 716                list_del_init(&epi->rdllink);
 717        spin_unlock_irqrestore(&ep->lock, flags);
 718
 719        wakeup_source_unregister(ep_wakeup_source(epi));
 720        /*
 721         * At this point it is safe to free the eventpoll item. Use the union
 722         * field epi->rcu, since we are trying to minimize the size of
 723         * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
 724         * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
 725         * use of the rbn field.
 726         */
 727        call_rcu(&epi->rcu, epi_rcu_free);
 728
 729        atomic_long_dec(&ep->user->epoll_watches);
 730
 731        return 0;
 732}
 733
 734static void ep_free(struct eventpoll *ep)
 735{
 736        struct rb_node *rbp;
 737        struct epitem *epi;
 738
 739        /* We need to release all tasks waiting for these file */
 740        if (waitqueue_active(&ep->poll_wait))
 741                ep_poll_safewake(&ep->poll_wait);
 742
 743        /*
 744         * We need to lock this because we could be hit by
 745         * eventpoll_release_file() while we're freeing the "struct eventpoll".
 746         * We do not need to hold "ep->mtx" here because the epoll file
 747         * is on the way to be removed and no one has references to it
 748         * anymore. The only hit might come from eventpoll_release_file() but
 749         * holding "epmutex" is sufficient here.
 750         */
 751        mutex_lock(&epmutex);
 752
 753        /*
 754         * Walks through the whole tree by unregistering poll callbacks.
 755         */
 756        for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
 757                epi = rb_entry(rbp, struct epitem, rbn);
 758
 759                ep_unregister_pollwait(ep, epi);
 760                cond_resched();
 761        }
 762
 763        /*
 764         * Walks through the whole tree by freeing each "struct epitem". At this
 765         * point we are sure no poll callbacks will be lingering around, and also by
 766         * holding "epmutex" we can be sure that no file cleanup code will hit
 767         * us during this operation. So we can avoid the lock on "ep->lock".
 768         * We do not need to lock ep->mtx, either, we only do it to prevent
 769         * a lockdep warning.
 770         */
 771        mutex_lock(&ep->mtx);
 772        while ((rbp = rb_first(&ep->rbr)) != NULL) {
 773                epi = rb_entry(rbp, struct epitem, rbn);
 774                ep_remove(ep, epi);
 775                cond_resched();
 776        }
 777        mutex_unlock(&ep->mtx);
 778
 779        mutex_unlock(&epmutex);
 780        mutex_destroy(&ep->mtx);
 781        free_uid(ep->user);
 782        wakeup_source_unregister(ep->ws);
 783        kfree(ep);
 784}
 785
 786static int ep_eventpoll_release(struct inode *inode, struct file *file)
 787{
 788        struct eventpoll *ep = file->private_data;
 789
 790        if (ep)
 791                ep_free(ep);
 792
 793        return 0;
 794}
 795
 796static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt)
 797{
 798        pt->_key = epi->event.events;
 799
 800        return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events;
 801}
 802
 803static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
 804                               void *priv)
 805{
 806        struct epitem *epi, *tmp;
 807        poll_table pt;
 808
 809        init_poll_funcptr(&pt, NULL);
 810
 811        list_for_each_entry_safe(epi, tmp, head, rdllink) {
 812                if (ep_item_poll(epi, &pt))
 813                        return POLLIN | POLLRDNORM;
 814                else {
 815                        /*
 816                         * Item has been dropped into the ready list by the poll
 817                         * callback, but it's not actually ready, as far as
 818                         * caller requested events goes. We can remove it here.
 819                         */
 820                        __pm_relax(ep_wakeup_source(epi));
 821                        list_del_init(&epi->rdllink);
 822                }
 823        }
 824
 825        return 0;
 826}
 827
 828static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
 829                                 poll_table *pt);
 830
 831struct readyevents_arg {
 832        struct eventpoll *ep;
 833        bool locked;
 834};
 835
 836static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
 837{
 838        struct readyevents_arg *arg = priv;
 839
 840        return ep_scan_ready_list(arg->ep, ep_read_events_proc, NULL,
 841                                  call_nests + 1, arg->locked);
 842}
 843
 844static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
 845{
 846        int pollflags;
 847        struct eventpoll *ep = file->private_data;
 848        struct readyevents_arg arg;
 849
 850        /*
 851         * During ep_insert() we already hold the ep->mtx for the tfile.
 852         * Prevent re-aquisition.
 853         */
 854        arg.locked = wait && (wait->_qproc == ep_ptable_queue_proc);
 855        arg.ep = ep;
 856
 857        /* Insert inside our poll wait queue */
 858        poll_wait(file, &ep->poll_wait, wait);
 859
 860        /*
 861         * Proceed to find out if wanted events are really available inside
 862         * the ready list. This need to be done under ep_call_nested()
 863         * supervision, since the call to f_op->poll() done on listed files
 864         * could re-enter here.
 865         */
 866        pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
 867                                   ep_poll_readyevents_proc, &arg, ep, current);
 868
 869        return pollflags != -1 ? pollflags : 0;
 870}
 871
 872#ifdef CONFIG_PROC_FS
 873static void ep_show_fdinfo(struct seq_file *m, struct file *f)
 874{
 875        struct eventpoll *ep = f->private_data;
 876        struct rb_node *rbp;
 877
 878        mutex_lock(&ep->mtx);
 879        for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
 880                struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
 881
 882                seq_printf(m, "tfd: %8d events: %8x data: %16llx\n",
 883                           epi->ffd.fd, epi->event.events,
 884                           (long long)epi->event.data);
 885                if (seq_has_overflowed(m))
 886                        break;
 887        }
 888        mutex_unlock(&ep->mtx);
 889}
 890#endif
 891
 892/* File callbacks that implement the eventpoll file behaviour */
 893static const struct file_operations eventpoll_fops = {
 894#ifdef CONFIG_PROC_FS
 895        .show_fdinfo    = ep_show_fdinfo,
 896#endif
 897        .release        = ep_eventpoll_release,
 898        .poll           = ep_eventpoll_poll,
 899        .llseek         = noop_llseek,
 900};
 901
 902/*
 903 * This is called from eventpoll_release() to unlink files from the eventpoll
 904 * interface. We need to have this facility to cleanup correctly files that are
 905 * closed without being removed from the eventpoll interface.
 906 */
 907void eventpoll_release_file(struct file *file)
 908{
 909        struct eventpoll *ep;
 910        struct epitem *epi, *next;
 911
 912        /*
 913         * We don't want to get "file->f_lock" because it is not
 914         * necessary. It is not necessary because we're in the "struct file"
 915         * cleanup path, and this means that no one is using this file anymore.
 916         * So, for example, epoll_ctl() cannot hit here since if we reach this
 917         * point, the file counter already went to zero and fget() would fail.
 918         * The only hit might come from ep_free() but by holding the mutex
 919         * will correctly serialize the operation. We do need to acquire
 920         * "ep->mtx" after "epmutex" because ep_remove() requires it when called
 921         * from anywhere but ep_free().
 922         *
 923         * Besides, ep_remove() acquires the lock, so we can't hold it here.
 924         */
 925        mutex_lock(&epmutex);
 926        list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
 927                ep = epi->ep;
 928                mutex_lock_nested(&ep->mtx, 0);
 929                ep_remove(ep, epi);
 930                mutex_unlock(&ep->mtx);
 931        }
 932        mutex_unlock(&epmutex);
 933}
 934
 935static int ep_alloc(struct eventpoll **pep)
 936{
 937        int error;
 938        struct user_struct *user;
 939        struct eventpoll *ep;
 940
 941        user = get_current_user();
 942        error = -ENOMEM;
 943        ep = kzalloc(sizeof(*ep), GFP_KERNEL);
 944        if (unlikely(!ep))
 945                goto free_uid;
 946
 947        spin_lock_init(&ep->lock);
 948        mutex_init(&ep->mtx);
 949        init_waitqueue_head(&ep->wq);
 950        init_waitqueue_head(&ep->poll_wait);
 951        INIT_LIST_HEAD(&ep->rdllist);
 952        ep->rbr = RB_ROOT;
 953        ep->ovflist = EP_UNACTIVE_PTR;
 954        ep->user = user;
 955
 956        *pep = ep;
 957
 958        return 0;
 959
 960free_uid:
 961        free_uid(user);
 962        return error;
 963}
 964
 965/*
 966 * Search the file inside the eventpoll tree. The RB tree operations
 967 * are protected by the "mtx" mutex, and ep_find() must be called with
 968 * "mtx" held.
 969 */
 970static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
 971{
 972        int kcmp;
 973        struct rb_node *rbp;
 974        struct epitem *epi, *epir = NULL;
 975        struct epoll_filefd ffd;
 976
 977        ep_set_ffd(&ffd, file, fd);
 978        for (rbp = ep->rbr.rb_node; rbp; ) {
 979                epi = rb_entry(rbp, struct epitem, rbn);
 980                kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
 981                if (kcmp > 0)
 982                        rbp = rbp->rb_right;
 983                else if (kcmp < 0)
 984                        rbp = rbp->rb_left;
 985                else {
 986                        epir = epi;
 987                        break;
 988                }
 989        }
 990
 991        return epir;
 992}
 993
 994/*
 995 * This is the callback that is passed to the wait queue wakeup
 996 * mechanism. It is called by the stored file descriptors when they
 997 * have events to report.
 998 */
 999static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
1000{
1001        int pwake = 0;
1002        unsigned long flags;
1003        struct epitem *epi = ep_item_from_wait(wait);
1004        struct eventpoll *ep = epi->ep;
1005
1006        if ((unsigned long)key & POLLFREE) {
1007                ep_pwq_from_wait(wait)->whead = NULL;
1008                /*
1009                 * whead = NULL above can race with ep_remove_wait_queue()
1010                 * which can do another remove_wait_queue() after us, so we
1011                 * can't use __remove_wait_queue(). whead->lock is held by
1012                 * the caller.
1013                 */
1014                list_del_init(&wait->task_list);
1015        }
1016
1017        spin_lock_irqsave(&ep->lock, flags);
1018
1019        /*
1020         * If the event mask does not contain any poll(2) event, we consider the
1021         * descriptor to be disabled. This condition is likely the effect of the
1022         * EPOLLONESHOT bit that disables the descriptor when an event is received,
1023         * until the next EPOLL_CTL_MOD will be issued.
1024         */
1025        if (!(epi->event.events & ~EP_PRIVATE_BITS))
1026                goto out_unlock;
1027
1028        /*
1029         * Check the events coming with the callback. At this stage, not
1030         * every device reports the events in the "key" parameter of the
1031         * callback. We need to be able to handle both cases here, hence the
1032         * test for "key" != NULL before the event match test.
1033         */
1034        if (key && !((unsigned long) key & epi->event.events))
1035                goto out_unlock;
1036
1037        /*
1038         * If we are transferring events to userspace, we can hold no locks
1039         * (because we're accessing user memory, and because of linux f_op->poll()
1040         * semantics). All the events that happen during that period of time are
1041         * chained in ep->ovflist and requeued later on.
1042         */
1043        if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
1044                if (epi->next == EP_UNACTIVE_PTR) {
1045                        epi->next = ep->ovflist;
1046                        ep->ovflist = epi;
1047                        if (epi->ws) {
1048                                /*
1049                                 * Activate ep->ws since epi->ws may get
1050                                 * deactivated at any time.
1051                                 */
1052                                __pm_stay_awake(ep->ws);
1053                        }
1054
1055                }
1056                goto out_unlock;
1057        }
1058
1059        /* If this file is already in the ready list we exit soon */
1060        if (!ep_is_linked(&epi->rdllink)) {
1061                list_add_tail(&epi->rdllink, &ep->rdllist);
1062                ep_pm_stay_awake_rcu(epi);
1063        }
1064
1065        /*
1066         * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1067         * wait list.
1068         */
1069        if (waitqueue_active(&ep->wq))
1070                wake_up_locked(&ep->wq);
1071        if (waitqueue_active(&ep->poll_wait))
1072                pwake++;
1073
1074out_unlock:
1075        spin_unlock_irqrestore(&ep->lock, flags);
1076
1077        /* We have to call this outside the lock */
1078        if (pwake)
1079                ep_poll_safewake(&ep->poll_wait);
1080
1081        return 1;
1082}
1083
1084/*
1085 * This is the callback that is used to add our wait queue to the
1086 * target file wakeup lists.
1087 */
1088static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1089                                 poll_table *pt)
1090{
1091        struct epitem *epi = ep_item_from_epqueue(pt);
1092        struct eppoll_entry *pwq;
1093
1094        if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1095                init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1096                pwq->whead = whead;
1097                pwq->base = epi;
1098                add_wait_queue(whead, &pwq->wait);
1099                list_add_tail(&pwq->llink, &epi->pwqlist);
1100                epi->nwait++;
1101        } else {
1102                /* We have to signal that an error occurred */
1103                epi->nwait = -1;
1104        }
1105}
1106
1107static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1108{
1109        int kcmp;
1110        struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
1111        struct epitem *epic;
1112
1113        while (*p) {
1114                parent = *p;
1115                epic = rb_entry(parent, struct epitem, rbn);
1116                kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1117                if (kcmp > 0)
1118                        p = &parent->rb_right;
1119                else
1120                        p = &parent->rb_left;
1121        }
1122        rb_link_node(&epi->rbn, parent, p);
1123        rb_insert_color(&epi->rbn, &ep->rbr);
1124}
1125
1126
1127
1128#define PATH_ARR_SIZE 5
1129/*
1130 * These are the number paths of length 1 to 5, that we are allowing to emanate
1131 * from a single file of interest. For example, we allow 1000 paths of length
1132 * 1, to emanate from each file of interest. This essentially represents the
1133 * potential wakeup paths, which need to be limited in order to avoid massive
1134 * uncontrolled wakeup storms. The common use case should be a single ep which
1135 * is connected to n file sources. In this case each file source has 1 path
1136 * of length 1. Thus, the numbers below should be more than sufficient. These
1137 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1138 * and delete can't add additional paths. Protected by the epmutex.
1139 */
1140static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1141static int path_count[PATH_ARR_SIZE];
1142
1143static int path_count_inc(int nests)
1144{
1145        /* Allow an arbitrary number of depth 1 paths */
1146        if (nests == 0)
1147                return 0;
1148
1149        if (++path_count[nests] > path_limits[nests])
1150                return -1;
1151        return 0;
1152}
1153
1154static void path_count_init(void)
1155{
1156        int i;
1157
1158        for (i = 0; i < PATH_ARR_SIZE; i++)
1159                path_count[i] = 0;
1160}
1161
1162static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1163{
1164        int error = 0;
1165        struct file *file = priv;
1166        struct file *child_file;
1167        struct epitem *epi;
1168
1169        /* CTL_DEL can remove links here, but that can't increase our count */
1170        rcu_read_lock();
1171        list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1172                child_file = epi->ep->file;
1173                if (is_file_epoll(child_file)) {
1174                        if (list_empty(&child_file->f_ep_links)) {
1175                                if (path_count_inc(call_nests)) {
1176                                        error = -1;
1177                                        break;
1178                                }
1179                        } else {
1180                                error = ep_call_nested(&poll_loop_ncalls,
1181                                                        EP_MAX_NESTS,
1182                                                        reverse_path_check_proc,
1183                                                        child_file, child_file,
1184                                                        current);
1185                        }
1186                        if (error != 0)
1187                                break;
1188                } else {
1189                        printk(KERN_ERR "reverse_path_check_proc: "
1190                                "file is not an ep!\n");
1191                }
1192        }
1193        rcu_read_unlock();
1194        return error;
1195}
1196
1197/**
1198 * reverse_path_check - The tfile_check_list is list of file *, which have
1199 *                      links that are proposed to be newly added. We need to
1200 *                      make sure that those added links don't add too many
1201 *                      paths such that we will spend all our time waking up
1202 *                      eventpoll objects.
1203 *
1204 * Returns: Returns zero if the proposed links don't create too many paths,
1205 *          -1 otherwise.
1206 */
1207static int reverse_path_check(void)
1208{
1209        int error = 0;
1210        struct file *current_file;
1211
1212        /* let's call this for all tfiles */
1213        list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1214                path_count_init();
1215                error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1216                                        reverse_path_check_proc, current_file,
1217                                        current_file, current);
1218                if (error)
1219                        break;
1220        }
1221        return error;
1222}
1223
1224static int ep_create_wakeup_source(struct epitem *epi)
1225{
1226        const char *name;
1227        struct wakeup_source *ws;
1228
1229        if (!epi->ep->ws) {
1230                epi->ep->ws = wakeup_source_register("eventpoll");
1231                if (!epi->ep->ws)
1232                        return -ENOMEM;
1233        }
1234
1235        name = epi->ffd.file->f_path.dentry->d_name.name;
1236        ws = wakeup_source_register(name);
1237
1238        if (!ws)
1239                return -ENOMEM;
1240        rcu_assign_pointer(epi->ws, ws);
1241
1242        return 0;
1243}
1244
1245/* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
1246static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1247{
1248        struct wakeup_source *ws = ep_wakeup_source(epi);
1249
1250        RCU_INIT_POINTER(epi->ws, NULL);
1251
1252        /*
1253         * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1254         * used internally by wakeup_source_remove, too (called by
1255         * wakeup_source_unregister), so we cannot use call_rcu
1256         */
1257        synchronize_rcu();
1258        wakeup_source_unregister(ws);
1259}
1260
1261/*
1262 * Must be called with "mtx" held.
1263 */
1264static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
1265                     struct file *tfile, int fd, int full_check)
1266{
1267        int error, revents, pwake = 0;
1268        unsigned long flags;
1269        long user_watches;
1270        struct epitem *epi;
1271        struct ep_pqueue epq;
1272
1273        user_watches = atomic_long_read(&ep->user->epoll_watches);
1274        if (unlikely(user_watches >= max_user_watches))
1275                return -ENOSPC;
1276        if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1277                return -ENOMEM;
1278
1279        /* Item initialization follow here ... */
1280        INIT_LIST_HEAD(&epi->rdllink);
1281        INIT_LIST_HEAD(&epi->fllink);
1282        INIT_LIST_HEAD(&epi->pwqlist);
1283        epi->ep = ep;
1284        ep_set_ffd(&epi->ffd, tfile, fd);
1285        epi->event = *event;
1286        epi->nwait = 0;
1287        epi->next = EP_UNACTIVE_PTR;
1288        if (epi->event.events & EPOLLWAKEUP) {
1289                error = ep_create_wakeup_source(epi);
1290                if (error)
1291                        goto error_create_wakeup_source;
1292        } else {
1293                RCU_INIT_POINTER(epi->ws, NULL);
1294        }
1295
1296        /* Initialize the poll table using the queue callback */
1297        epq.epi = epi;
1298        init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1299
1300        /*
1301         * Attach the item to the poll hooks and get current event bits.
1302         * We can safely use the file* here because its usage count has
1303         * been increased by the caller of this function. Note that after
1304         * this operation completes, the poll callback can start hitting
1305         * the new item.
1306         */
1307        revents = ep_item_poll(epi, &epq.pt);
1308
1309        /*
1310         * We have to check if something went wrong during the poll wait queue
1311         * install process. Namely an allocation for a wait queue failed due
1312         * high memory pressure.
1313         */
1314        error = -ENOMEM;
1315        if (epi->nwait < 0)
1316                goto error_unregister;
1317
1318        /* Add the current item to the list of active epoll hook for this file */
1319        spin_lock(&tfile->f_lock);
1320        list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1321        spin_unlock(&tfile->f_lock);
1322
1323        /*
1324         * Add the current item to the RB tree. All RB tree operations are
1325         * protected by "mtx", and ep_insert() is called with "mtx" held.
1326         */
1327        ep_rbtree_insert(ep, epi);
1328
1329        /* now check if we've created too many backpaths */
1330        error = -EINVAL;
1331        if (full_check && reverse_path_check())
1332                goto error_remove_epi;
1333
1334        /* We have to drop the new item inside our item list to keep track of it */
1335        spin_lock_irqsave(&ep->lock, flags);
1336
1337        /* If the file is already "ready" we drop it inside the ready list */
1338        if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
1339                list_add_tail(&epi->rdllink, &ep->rdllist);
1340                ep_pm_stay_awake(epi);
1341
1342                /* Notify waiting tasks that events are available */
1343                if (waitqueue_active(&ep->wq))
1344                        wake_up_locked(&ep->wq);
1345                if (waitqueue_active(&ep->poll_wait))
1346                        pwake++;
1347        }
1348
1349        spin_unlock_irqrestore(&ep->lock, flags);
1350
1351        atomic_long_inc(&ep->user->epoll_watches);
1352
1353        /* We have to call this outside the lock */
1354        if (pwake)
1355                ep_poll_safewake(&ep->poll_wait);
1356
1357        return 0;
1358
1359error_remove_epi:
1360        spin_lock(&tfile->f_lock);
1361        list_del_rcu(&epi->fllink);
1362        spin_unlock(&tfile->f_lock);
1363
1364        rb_erase(&epi->rbn, &ep->rbr);
1365
1366error_unregister:
1367        ep_unregister_pollwait(ep, epi);
1368
1369        /*
1370         * We need to do this because an event could have been arrived on some
1371         * allocated wait queue. Note that we don't care about the ep->ovflist
1372         * list, since that is used/cleaned only inside a section bound by "mtx".
1373         * And ep_insert() is called with "mtx" held.
1374         */
1375        spin_lock_irqsave(&ep->lock, flags);
1376        if (ep_is_linked(&epi->rdllink))
1377                list_del_init(&epi->rdllink);
1378        spin_unlock_irqrestore(&ep->lock, flags);
1379
1380        wakeup_source_unregister(ep_wakeup_source(epi));
1381
1382error_create_wakeup_source:
1383        kmem_cache_free(epi_cache, epi);
1384
1385        return error;
1386}
1387
1388/*
1389 * Modify the interest event mask by dropping an event if the new mask
1390 * has a match in the current file status. Must be called with "mtx" held.
1391 */
1392static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1393{
1394        int pwake = 0;
1395        unsigned int revents;
1396        poll_table pt;
1397
1398        init_poll_funcptr(&pt, NULL);
1399
1400        /*
1401         * Set the new event interest mask before calling f_op->poll();
1402         * otherwise we might miss an event that happens between the
1403         * f_op->poll() call and the new event set registering.
1404         */
1405        epi->event.events = event->events; /* need barrier below */
1406        epi->event.data = event->data; /* protected by mtx */
1407        if (epi->event.events & EPOLLWAKEUP) {
1408                if (!ep_has_wakeup_source(epi))
1409                        ep_create_wakeup_source(epi);
1410        } else if (ep_has_wakeup_source(epi)) {
1411                ep_destroy_wakeup_source(epi);
1412        }
1413
1414        /*
1415         * The following barrier has two effects:
1416         *
1417         * 1) Flush epi changes above to other CPUs.  This ensures
1418         *    we do not miss events from ep_poll_callback if an
1419         *    event occurs immediately after we call f_op->poll().
1420         *    We need this because we did not take ep->lock while
1421         *    changing epi above (but ep_poll_callback does take
1422         *    ep->lock).
1423         *
1424         * 2) We also need to ensure we do not miss _past_ events
1425         *    when calling f_op->poll().  This barrier also
1426         *    pairs with the barrier in wq_has_sleeper (see
1427         *    comments for wq_has_sleeper).
1428         *
1429         * This barrier will now guarantee ep_poll_callback or f_op->poll
1430         * (or both) will notice the readiness of an item.
1431         */
1432        smp_mb();
1433
1434        /*
1435         * Get current event bits. We can safely use the file* here because
1436         * its usage count has been increased by the caller of this function.
1437         */
1438        revents = ep_item_poll(epi, &pt);
1439
1440        /*
1441         * If the item is "hot" and it is not registered inside the ready
1442         * list, push it inside.
1443         */
1444        if (revents & event->events) {
1445                spin_lock_irq(&ep->lock);
1446                if (!ep_is_linked(&epi->rdllink)) {
1447                        list_add_tail(&epi->rdllink, &ep->rdllist);
1448                        ep_pm_stay_awake(epi);
1449
1450                        /* Notify waiting tasks that events are available */
1451                        if (waitqueue_active(&ep->wq))
1452                                wake_up_locked(&ep->wq);
1453                        if (waitqueue_active(&ep->poll_wait))
1454                                pwake++;
1455                }
1456                spin_unlock_irq(&ep->lock);
1457        }
1458
1459        /* We have to call this outside the lock */
1460        if (pwake)
1461                ep_poll_safewake(&ep->poll_wait);
1462
1463        return 0;
1464}
1465
1466static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1467                               void *priv)
1468{
1469        struct ep_send_events_data *esed = priv;
1470        int eventcnt;
1471        unsigned int revents;
1472        struct epitem *epi;
1473        struct epoll_event __user *uevent;
1474        struct wakeup_source *ws;
1475        poll_table pt;
1476
1477        init_poll_funcptr(&pt, NULL);
1478
1479        /*
1480         * We can loop without lock because we are passed a task private list.
1481         * Items cannot vanish during the loop because ep_scan_ready_list() is
1482         * holding "mtx" during this call.
1483         */
1484        for (eventcnt = 0, uevent = esed->events;
1485             !list_empty(head) && eventcnt < esed->maxevents;) {
1486                epi = list_first_entry(head, struct epitem, rdllink);
1487
1488                /*
1489                 * Activate ep->ws before deactivating epi->ws to prevent
1490                 * triggering auto-suspend here (in case we reactive epi->ws
1491                 * below).
1492                 *
1493                 * This could be rearranged to delay the deactivation of epi->ws
1494                 * instead, but then epi->ws would temporarily be out of sync
1495                 * with ep_is_linked().
1496                 */
1497                ws = ep_wakeup_source(epi);
1498                if (ws) {
1499                        if (ws->active)
1500                                __pm_stay_awake(ep->ws);
1501                        __pm_relax(ws);
1502                }
1503
1504                list_del_init(&epi->rdllink);
1505
1506                revents = ep_item_poll(epi, &pt);
1507
1508                /*
1509                 * If the event mask intersect the caller-requested one,
1510                 * deliver the event to userspace. Again, ep_scan_ready_list()
1511                 * is holding "mtx", so no operations coming from userspace
1512                 * can change the item.
1513                 */
1514                if (revents) {
1515                        if (__put_user(revents, &uevent->events) ||
1516                            __put_user(epi->event.data, &uevent->data)) {
1517                                list_add(&epi->rdllink, head);
1518                                ep_pm_stay_awake(epi);
1519                                return eventcnt ? eventcnt : -EFAULT;
1520                        }
1521                        eventcnt++;
1522                        uevent++;
1523                        if (epi->event.events & EPOLLONESHOT)
1524                                epi->event.events &= EP_PRIVATE_BITS;
1525                        else if (!(epi->event.events & EPOLLET)) {
1526                                /*
1527                                 * If this file has been added with Level
1528                                 * Trigger mode, we need to insert back inside
1529                                 * the ready list, so that the next call to
1530                                 * epoll_wait() will check again the events
1531                                 * availability. At this point, no one can insert
1532                                 * into ep->rdllist besides us. The epoll_ctl()
1533                                 * callers are locked out by
1534                                 * ep_scan_ready_list() holding "mtx" and the
1535                                 * poll callback will queue them in ep->ovflist.
1536                                 */
1537                                list_add_tail(&epi->rdllink, &ep->rdllist);
1538                                ep_pm_stay_awake(epi);
1539                        }
1540                }
1541        }
1542
1543        return eventcnt;
1544}
1545
1546static int ep_send_events(struct eventpoll *ep,
1547                          struct epoll_event __user *events, int maxevents)
1548{
1549        struct ep_send_events_data esed;
1550
1551        esed.maxevents = maxevents;
1552        esed.events = events;
1553
1554        return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1555}
1556
1557static inline struct timespec ep_set_mstimeout(long ms)
1558{
1559        struct timespec now, ts = {
1560                .tv_sec = ms / MSEC_PER_SEC,
1561                .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1562        };
1563
1564        ktime_get_ts(&now);
1565        return timespec_add_safe(now, ts);
1566}
1567
1568/**
1569 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1570 *           event buffer.
1571 *
1572 * @ep: Pointer to the eventpoll context.
1573 * @events: Pointer to the userspace buffer where the ready events should be
1574 *          stored.
1575 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1576 * @timeout: Maximum timeout for the ready events fetch operation, in
1577 *           milliseconds. If the @timeout is zero, the function will not block,
1578 *           while if the @timeout is less than zero, the function will block
1579 *           until at least one event has been retrieved (or an error
1580 *           occurred).
1581 *
1582 * Returns: Returns the number of ready events which have been fetched, or an
1583 *          error code, in case of error.
1584 */
1585static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1586                   int maxevents, long timeout)
1587{
1588        int res = 0, eavail, timed_out = 0;
1589        unsigned long flags;
1590        long slack = 0;
1591        wait_queue_t wait;
1592        ktime_t expires, *to = NULL;
1593
1594        if (timeout > 0) {
1595                struct timespec end_time = ep_set_mstimeout(timeout);
1596
1597                slack = select_estimate_accuracy(&end_time);
1598                to = &expires;
1599                *to = timespec_to_ktime(end_time);
1600        } else if (timeout == 0) {
1601                /*
1602                 * Avoid the unnecessary trip to the wait queue loop, if the
1603                 * caller specified a non blocking operation.
1604                 */
1605                timed_out = 1;
1606                spin_lock_irqsave(&ep->lock, flags);
1607                goto check_events;
1608        }
1609
1610fetch_events:
1611        spin_lock_irqsave(&ep->lock, flags);
1612
1613        if (!ep_events_available(ep)) {
1614                /*
1615                 * We don't have any available event to return to the caller.
1616                 * We need to sleep here, and we will be wake up by
1617                 * ep_poll_callback() when events will become available.
1618                 */
1619                init_waitqueue_entry(&wait, current);
1620                __add_wait_queue_exclusive(&ep->wq, &wait);
1621
1622                for (;;) {
1623                        /*
1624                         * We don't want to sleep if the ep_poll_callback() sends us
1625                         * a wakeup in between. That's why we set the task state
1626                         * to TASK_INTERRUPTIBLE before doing the checks.
1627                         */
1628                        set_current_state(TASK_INTERRUPTIBLE);
1629                        if (ep_events_available(ep) || timed_out)
1630                                break;
1631                        if (signal_pending(current)) {
1632                                res = -EINTR;
1633                                break;
1634                        }
1635
1636                        spin_unlock_irqrestore(&ep->lock, flags);
1637                        if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1638                                timed_out = 1;
1639
1640                        spin_lock_irqsave(&ep->lock, flags);
1641                }
1642                __remove_wait_queue(&ep->wq, &wait);
1643
1644                set_current_state(TASK_RUNNING);
1645        }
1646check_events:
1647        /* Is it worth to try to dig for events ? */
1648        eavail = ep_events_available(ep);
1649
1650        spin_unlock_irqrestore(&ep->lock, flags);
1651
1652        /*
1653         * Try to transfer events to user space. In case we get 0 events and
1654         * there's still timeout left over, we go trying again in search of
1655         * more luck.
1656         */
1657        if (!res && eavail &&
1658            !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1659                goto fetch_events;
1660
1661        return res;
1662}
1663
1664/**
1665 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1666 *                      API, to verify that adding an epoll file inside another
1667 *                      epoll structure, does not violate the constraints, in
1668 *                      terms of closed loops, or too deep chains (which can
1669 *                      result in excessive stack usage).
1670 *
1671 * @priv: Pointer to the epoll file to be currently checked.
1672 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1673 *          data structure pointer.
1674 * @call_nests: Current dept of the @ep_call_nested() call stack.
1675 *
1676 * Returns: Returns zero if adding the epoll @file inside current epoll
1677 *          structure @ep does not violate the constraints, or -1 otherwise.
1678 */
1679static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1680{
1681        int error = 0;
1682        struct file *file = priv;
1683        struct eventpoll *ep = file->private_data;
1684        struct eventpoll *ep_tovisit;
1685        struct rb_node *rbp;
1686        struct epitem *epi;
1687
1688        mutex_lock_nested(&ep->mtx, call_nests + 1);
1689        ep->visited = 1;
1690        list_add(&ep->visited_list_link, &visited_list);
1691        for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1692                epi = rb_entry(rbp, struct epitem, rbn);
1693                if (unlikely(is_file_epoll(epi->ffd.file))) {
1694                        ep_tovisit = epi->ffd.file->private_data;
1695                        if (ep_tovisit->visited)
1696                                continue;
1697                        error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1698                                        ep_loop_check_proc, epi->ffd.file,
1699                                        ep_tovisit, current);
1700                        if (error != 0)
1701                                break;
1702                } else {
1703                        /*
1704                         * If we've reached a file that is not associated with
1705                         * an ep, then we need to check if the newly added
1706                         * links are going to add too many wakeup paths. We do
1707                         * this by adding it to the tfile_check_list, if it's
1708                         * not already there, and calling reverse_path_check()
1709                         * during ep_insert().
1710                         */
1711                        if (list_empty(&epi->ffd.file->f_tfile_llink))
1712                                list_add(&epi->ffd.file->f_tfile_llink,
1713                                         &tfile_check_list);
1714                }
1715        }
1716        mutex_unlock(&ep->mtx);
1717
1718        return error;
1719}
1720
1721/**
1722 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1723 *                 another epoll file (represented by @ep) does not create
1724 *                 closed loops or too deep chains.
1725 *
1726 * @ep: Pointer to the epoll private data structure.
1727 * @file: Pointer to the epoll file to be checked.
1728 *
1729 * Returns: Returns zero if adding the epoll @file inside current epoll
1730 *          structure @ep does not violate the constraints, or -1 otherwise.
1731 */
1732static int ep_loop_check(struct eventpoll *ep, struct file *file)
1733{
1734        int ret;
1735        struct eventpoll *ep_cur, *ep_next;
1736
1737        ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1738                              ep_loop_check_proc, file, ep, current);
1739        /* clear visited list */
1740        list_for_each_entry_safe(ep_cur, ep_next, &visited_list,
1741                                                        visited_list_link) {
1742                ep_cur->visited = 0;
1743                list_del(&ep_cur->visited_list_link);
1744        }
1745        return ret;
1746}
1747
1748static void clear_tfile_check_list(void)
1749{
1750        struct file *file;
1751
1752        /* first clear the tfile_check_list */
1753        while (!list_empty(&tfile_check_list)) {
1754                file = list_first_entry(&tfile_check_list, struct file,
1755                                        f_tfile_llink);
1756                list_del_init(&file->f_tfile_llink);
1757        }
1758        INIT_LIST_HEAD(&tfile_check_list);
1759}
1760
1761/*
1762 * Open an eventpoll file descriptor.
1763 */
1764SYSCALL_DEFINE1(epoll_create1, int, flags)
1765{
1766        int error, fd;
1767        struct eventpoll *ep = NULL;
1768        struct file *file;
1769
1770        /* Check the EPOLL_* constant for consistency.  */
1771        BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1772
1773        if (flags & ~EPOLL_CLOEXEC)
1774                return -EINVAL;
1775        /*
1776         * Create the internal data structure ("struct eventpoll").
1777         */
1778        error = ep_alloc(&ep);
1779        if (error < 0)
1780                return error;
1781        /*
1782         * Creates all the items needed to setup an eventpoll file. That is,
1783         * a file structure and a free file descriptor.
1784         */
1785        fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1786        if (fd < 0) {
1787                error = fd;
1788                goto out_free_ep;
1789        }
1790        file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1791                                 O_RDWR | (flags & O_CLOEXEC));
1792        if (IS_ERR(file)) {
1793                error = PTR_ERR(file);
1794                goto out_free_fd;
1795        }
1796        ep->file = file;
1797        fd_install(fd, file);
1798        return fd;
1799
1800out_free_fd:
1801        put_unused_fd(fd);
1802out_free_ep:
1803        ep_free(ep);
1804        return error;
1805}
1806
1807SYSCALL_DEFINE1(epoll_create, int, size)
1808{
1809        if (size <= 0)
1810                return -EINVAL;
1811
1812        return sys_epoll_create1(0);
1813}
1814
1815/*
1816 * The following function implements the controller interface for
1817 * the eventpoll file that enables the insertion/removal/change of
1818 * file descriptors inside the interest set.
1819 */
1820SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1821                struct epoll_event __user *, event)
1822{
1823        int error;
1824        int full_check = 0;
1825        struct fd f, tf;
1826        struct eventpoll *ep;
1827        struct epitem *epi;
1828        struct epoll_event epds;
1829        struct eventpoll *tep = NULL;
1830
1831        error = -EFAULT;
1832        if (ep_op_has_event(op) &&
1833            copy_from_user(&epds, event, sizeof(struct epoll_event)))
1834                goto error_return;
1835
1836        error = -EBADF;
1837        f = fdget(epfd);
1838        if (!f.file)
1839                goto error_return;
1840
1841        /* Get the "struct file *" for the target file */
1842        tf = fdget(fd);
1843        if (!tf.file)
1844                goto error_fput;
1845
1846        /* The target file descriptor must support poll */
1847        error = -EPERM;
1848        if (!tf.file->f_op->poll)
1849                goto error_tgt_fput;
1850
1851        /* Check if EPOLLWAKEUP is allowed */
1852        if (ep_op_has_event(op))
1853                ep_take_care_of_epollwakeup(&epds);
1854
1855        /*
1856         * We have to check that the file structure underneath the file descriptor
1857         * the user passed to us _is_ an eventpoll file. And also we do not permit
1858         * adding an epoll file descriptor inside itself.
1859         */
1860        error = -EINVAL;
1861        if (f.file == tf.file || !is_file_epoll(f.file))
1862                goto error_tgt_fput;
1863
1864        /*
1865         * At this point it is safe to assume that the "private_data" contains
1866         * our own data structure.
1867         */
1868        ep = f.file->private_data;
1869
1870        /*
1871         * When we insert an epoll file descriptor, inside another epoll file
1872         * descriptor, there is the change of creating closed loops, which are
1873         * better be handled here, than in more critical paths. While we are
1874         * checking for loops we also determine the list of files reachable
1875         * and hang them on the tfile_check_list, so we can check that we
1876         * haven't created too many possible wakeup paths.
1877         *
1878         * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
1879         * the epoll file descriptor is attaching directly to a wakeup source,
1880         * unless the epoll file descriptor is nested. The purpose of taking the
1881         * 'epmutex' on add is to prevent complex toplogies such as loops and
1882         * deep wakeup paths from forming in parallel through multiple
1883         * EPOLL_CTL_ADD operations.
1884         */
1885        mutex_lock_nested(&ep->mtx, 0);
1886        if (op == EPOLL_CTL_ADD) {
1887                if (!list_empty(&f.file->f_ep_links) ||
1888                                                is_file_epoll(tf.file)) {
1889                        full_check = 1;
1890                        mutex_unlock(&ep->mtx);
1891                        mutex_lock(&epmutex);
1892                        if (is_file_epoll(tf.file)) {
1893                                error = -ELOOP;
1894                                if (ep_loop_check(ep, tf.file) != 0) {
1895                                        clear_tfile_check_list();
1896                                        goto error_tgt_fput;
1897                                }
1898                        } else
1899                                list_add(&tf.file->f_tfile_llink,
1900                                                        &tfile_check_list);
1901                        mutex_lock_nested(&ep->mtx, 0);
1902                        if (is_file_epoll(tf.file)) {
1903                                tep = tf.file->private_data;
1904                                mutex_lock_nested(&tep->mtx, 1);
1905                        }
1906                }
1907        }
1908
1909        /*
1910         * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1911         * above, we can be sure to be able to use the item looked up by
1912         * ep_find() till we release the mutex.
1913         */
1914        epi = ep_find(ep, tf.file, fd);
1915
1916        error = -EINVAL;
1917        switch (op) {
1918        case EPOLL_CTL_ADD:
1919                if (!epi) {
1920                        epds.events |= POLLERR | POLLHUP;
1921                        error = ep_insert(ep, &epds, tf.file, fd, full_check);
1922                } else
1923                        error = -EEXIST;
1924                if (full_check)
1925                        clear_tfile_check_list();
1926                break;
1927        case EPOLL_CTL_DEL:
1928                if (epi)
1929                        error = ep_remove(ep, epi);
1930                else
1931                        error = -ENOENT;
1932                break;
1933        case EPOLL_CTL_MOD:
1934                if (epi) {
1935                        epds.events |= POLLERR | POLLHUP;
1936                        error = ep_modify(ep, epi, &epds);
1937                } else
1938                        error = -ENOENT;
1939                break;
1940        }
1941        if (tep != NULL)
1942                mutex_unlock(&tep->mtx);
1943        mutex_unlock(&ep->mtx);
1944
1945error_tgt_fput:
1946        if (full_check)
1947                mutex_unlock(&epmutex);
1948
1949        fdput(tf);
1950error_fput:
1951        fdput(f);
1952error_return:
1953
1954        return error;
1955}
1956
1957/*
1958 * Implement the event wait interface for the eventpoll file. It is the kernel
1959 * part of the user space epoll_wait(2).
1960 */
1961SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1962                int, maxevents, int, timeout)
1963{
1964        int error;
1965        struct fd f;
1966        struct eventpoll *ep;
1967
1968        /* The maximum number of event must be greater than zero */
1969        if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1970                return -EINVAL;
1971
1972        /* Verify that the area passed by the user is writeable */
1973        if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
1974                return -EFAULT;
1975
1976        /* Get the "struct file *" for the eventpoll file */
1977        f = fdget(epfd);
1978        if (!f.file)
1979                return -EBADF;
1980
1981        /*
1982         * We have to check that the file structure underneath the fd
1983         * the user passed to us _is_ an eventpoll file.
1984         */
1985        error = -EINVAL;
1986        if (!is_file_epoll(f.file))
1987                goto error_fput;
1988
1989        /*
1990         * At this point it is safe to assume that the "private_data" contains
1991         * our own data structure.
1992         */
1993        ep = f.file->private_data;
1994
1995        /* Time to fish for events ... */
1996        error = ep_poll(ep, events, maxevents, timeout);
1997
1998error_fput:
1999        fdput(f);
2000        return error;
2001}
2002
2003/*
2004 * Implement the event wait interface for the eventpoll file. It is the kernel
2005 * part of the user space epoll_pwait(2).
2006 */
2007SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2008                int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2009                size_t, sigsetsize)
2010{
2011        int error;
2012        sigset_t ksigmask, sigsaved;
2013
2014        /*
2015         * If the caller wants a certain signal mask to be set during the wait,
2016         * we apply it here.
2017         */
2018        if (sigmask) {
2019                if (sigsetsize != sizeof(sigset_t))
2020                        return -EINVAL;
2021                if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2022                        return -EFAULT;
2023                sigsaved = current->blocked;
2024                set_current_blocked(&ksigmask);
2025        }
2026
2027        error = sys_epoll_wait(epfd, events, maxevents, timeout);
2028
2029        /*
2030         * If we changed the signal mask, we need to restore the original one.
2031         * In case we've got a signal while waiting, we do not restore the
2032         * signal mask yet, and we allow do_signal() to deliver the signal on
2033         * the way back to userspace, before the signal mask is restored.
2034         */
2035        if (sigmask) {
2036                if (error == -EINTR) {
2037                        memcpy(&current->saved_sigmask, &sigsaved,
2038                               sizeof(sigsaved));
2039                        set_restore_sigmask();
2040                } else
2041                        set_current_blocked(&sigsaved);
2042        }
2043
2044        return error;
2045}
2046
2047#ifdef CONFIG_COMPAT
2048COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2049                        struct epoll_event __user *, events,
2050                        int, maxevents, int, timeout,
2051                        const compat_sigset_t __user *, sigmask,
2052                        compat_size_t, sigsetsize)
2053{
2054        long err;
2055        compat_sigset_t csigmask;
2056        sigset_t ksigmask, sigsaved;
2057
2058        /*
2059         * If the caller wants a certain signal mask to be set during the wait,
2060         * we apply it here.
2061         */
2062        if (sigmask) {
2063                if (sigsetsize != sizeof(compat_sigset_t))
2064                        return -EINVAL;
2065                if (copy_from_user(&csigmask, sigmask, sizeof(csigmask)))
2066                        return -EFAULT;
2067                sigset_from_compat(&ksigmask, &csigmask);
2068                sigsaved = current->blocked;
2069                set_current_blocked(&ksigmask);
2070        }
2071
2072        err = sys_epoll_wait(epfd, events, maxevents, timeout);
2073
2074        /*
2075         * If we changed the signal mask, we need to restore the original one.
2076         * In case we've got a signal while waiting, we do not restore the
2077         * signal mask yet, and we allow do_signal() to deliver the signal on
2078         * the way back to userspace, before the signal mask is restored.
2079         */
2080        if (sigmask) {
2081                if (err == -EINTR) {
2082                        memcpy(&current->saved_sigmask, &sigsaved,
2083                               sizeof(sigsaved));
2084                        set_restore_sigmask();
2085                } else
2086                        set_current_blocked(&sigsaved);
2087        }
2088
2089        return err;
2090}
2091#endif
2092
2093static int __init eventpoll_init(void)
2094{
2095        struct sysinfo si;
2096
2097        si_meminfo(&si);
2098        /*
2099         * Allows top 4% of lomem to be allocated for epoll watches (per user).
2100         */
2101        max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2102                EP_ITEM_COST;
2103        BUG_ON(max_user_watches < 0);
2104
2105        /*
2106         * Initialize the structure used to perform epoll file descriptor
2107         * inclusion loops checks.
2108         */
2109        ep_nested_calls_init(&poll_loop_ncalls);
2110
2111        /* Initialize the structure used to perform safe poll wait head wake ups */
2112        ep_nested_calls_init(&poll_safewake_ncalls);
2113
2114        /* Initialize the structure used to perform file's f_op->poll() calls */
2115        ep_nested_calls_init(&poll_readywalk_ncalls);
2116
2117        /*
2118         * We can have many thousands of epitems, so prevent this from
2119         * using an extra cache line on 64-bit (and smaller) CPUs
2120         */
2121        BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2122
2123        /* Allocates slab cache used to allocate "struct epitem" items */
2124        epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2125                        0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
2126
2127        /* Allocates slab cache used to allocate "struct eppoll_entry" */
2128        pwq_cache = kmem_cache_create("eventpoll_pwq",
2129                        sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
2130
2131        return 0;
2132}
2133fs_initcall(eventpoll_init);
2134
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.