Root/fs/eventpoll.c

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 <asm/uaccess.h>
37#include <asm/system.h>
38#include <asm/io.h>
39#include <asm/mman.h>
40#include <asm/atomic.h>
41
42/*
43 * LOCKING:
44 * There are three level of locking required by epoll :
45 *
46 * 1) epmutex (mutex)
47 * 2) ep->mtx (mutex)
48 * 3) ep->lock (spinlock)
49 *
50 * The acquire order is the one listed above, from 1 to 3.
51 * We need a spinlock (ep->lock) because we manipulate objects
52 * from inside the poll callback, that might be triggered from
53 * a wake_up() that in turn might be called from IRQ context.
54 * So we can't sleep inside the poll callback and hence we need
55 * a spinlock. During the event transfer loop (from kernel to
56 * user space) we could end up sleeping due a copy_to_user(), so
57 * we need a lock that will allow us to sleep. This lock is a
58 * mutex (ep->mtx). It is acquired during the event transfer loop,
59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
60 * Then we also need a global mutex to serialize eventpoll_release_file()
61 * and ep_free().
62 * This mutex is acquired by ep_free() during the epoll file
63 * cleanup path and it is also acquired by eventpoll_release_file()
64 * if a file has been pushed inside an epoll set and it is then
65 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL).
66 * It is possible to drop the "ep->mtx" and to use the global
67 * mutex "epmutex" (together with "ep->lock") to have it working,
68 * but having "ep->mtx" will make the interface more scalable.
69 * Events that require holding "epmutex" are very rare, while for
70 * normal operations the epoll private "ep->mtx" will guarantee
71 * a better scalability.
72 */
73
74/* Epoll private bits inside the event mask */
75#define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET)
76
77/* Maximum number of nesting allowed inside epoll sets */
78#define EP_MAX_NESTS 4
79
80/* Maximum msec timeout value storeable in a long int */
81#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
82
83#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
84
85#define EP_UNACTIVE_PTR ((void *) -1L)
86
87#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
88
89struct epoll_filefd {
90    struct file *file;
91    int fd;
92};
93
94/*
95 * Structure used to track possible nested calls, for too deep recursions
96 * and loop cycles.
97 */
98struct nested_call_node {
99    struct list_head llink;
100    void *cookie;
101    void *ctx;
102};
103
104/*
105 * This structure is used as collector for nested calls, to check for
106 * maximum recursion dept and loop cycles.
107 */
108struct nested_calls {
109    struct list_head tasks_call_list;
110    spinlock_t lock;
111};
112
113/*
114 * Each file descriptor added to the eventpoll interface will
115 * have an entry of this type linked to the "rbr" RB tree.
116 */
117struct epitem {
118    /* RB tree node used to link this structure to the eventpoll RB tree */
119    struct rb_node rbn;
120
121    /* List header used to link this structure to the eventpoll ready list */
122    struct list_head rdllink;
123
124    /*
125     * Works together "struct eventpoll"->ovflist in keeping the
126     * single linked chain of items.
127     */
128    struct epitem *next;
129
130    /* The file descriptor information this item refers to */
131    struct epoll_filefd ffd;
132
133    /* Number of active wait queue attached to poll operations */
134    int nwait;
135
136    /* List containing poll wait queues */
137    struct list_head pwqlist;
138
139    /* The "container" of this item */
140    struct eventpoll *ep;
141
142    /* List header used to link this item to the "struct file" items list */
143    struct list_head fllink;
144
145    /* The structure that describe the interested events and the source fd */
146    struct epoll_event event;
147};
148
149/*
150 * This structure is stored inside the "private_data" member of the file
151 * structure and rapresent the main data sructure for the eventpoll
152 * interface.
153 */
154struct eventpoll {
155    /* Protect the this structure access */
156    spinlock_t lock;
157
158    /*
159     * This mutex is used to ensure that files are not removed
160     * while epoll is using them. This is held during the event
161     * collection loop, the file cleanup path, the epoll file exit
162     * code and the ctl operations.
163     */
164    struct mutex mtx;
165
166    /* Wait queue used by sys_epoll_wait() */
167    wait_queue_head_t wq;
168
169    /* Wait queue used by file->poll() */
170    wait_queue_head_t poll_wait;
171
172    /* List of ready file descriptors */
173    struct list_head rdllist;
174
175    /* RB tree root used to store monitored fd structs */
176    struct rb_root rbr;
177
178    /*
179     * This is a single linked list that chains all the "struct epitem" that
180     * happened while transfering ready events to userspace w/out
181     * holding ->lock.
182     */
183    struct epitem *ovflist;
184
185    /* The user that created the eventpoll descriptor */
186    struct user_struct *user;
187};
188
189/* Wait structure used by the poll hooks */
190struct eppoll_entry {
191    /* List header used to link this structure to the "struct epitem" */
192    struct list_head llink;
193
194    /* The "base" pointer is set to the container "struct epitem" */
195    struct epitem *base;
196
197    /*
198     * Wait queue item that will be linked to the target file wait
199     * queue head.
200     */
201    wait_queue_t wait;
202
203    /* The wait queue head that linked the "wait" wait queue item */
204    wait_queue_head_t *whead;
205};
206
207/* Wrapper struct used by poll queueing */
208struct ep_pqueue {
209    poll_table pt;
210    struct epitem *epi;
211};
212
213/* Used by the ep_send_events() function as callback private data */
214struct ep_send_events_data {
215    int maxevents;
216    struct epoll_event __user *events;
217};
218
219/*
220 * Configuration options available inside /proc/sys/fs/epoll/
221 */
222/* Maximum number of epoll watched descriptors, per user */
223static int max_user_watches __read_mostly;
224
225/*
226 * This mutex is used to serialize ep_free() and eventpoll_release_file().
227 */
228static DEFINE_MUTEX(epmutex);
229
230/* Used for safe wake up implementation */
231static struct nested_calls poll_safewake_ncalls;
232
233/* Used to call file's f_op->poll() under the nested calls boundaries */
234static struct nested_calls poll_readywalk_ncalls;
235
236/* Slab cache used to allocate "struct epitem" */
237static struct kmem_cache *epi_cache __read_mostly;
238
239/* Slab cache used to allocate "struct eppoll_entry" */
240static struct kmem_cache *pwq_cache __read_mostly;
241
242#ifdef CONFIG_SYSCTL
243
244#include <linux/sysctl.h>
245
246static int zero;
247
248ctl_table epoll_table[] = {
249    {
250        .procname = "max_user_watches",
251        .data = &max_user_watches,
252        .maxlen = sizeof(int),
253        .mode = 0644,
254        .proc_handler = proc_dointvec_minmax,
255        .extra1 = &zero,
256    },
257    { }
258};
259#endif /* CONFIG_SYSCTL */
260
261
262/* Setup the structure that is used as key for the RB tree */
263static inline void ep_set_ffd(struct epoll_filefd *ffd,
264                  struct file *file, int fd)
265{
266    ffd->file = file;
267    ffd->fd = fd;
268}
269
270/* Compare RB tree keys */
271static inline int ep_cmp_ffd(struct epoll_filefd *p1,
272                 struct epoll_filefd *p2)
273{
274    return (p1->file > p2->file ? +1:
275            (p1->file < p2->file ? -1 : p1->fd - p2->fd));
276}
277
278/* Tells us if the item is currently linked */
279static inline int ep_is_linked(struct list_head *p)
280{
281    return !list_empty(p);
282}
283
284/* Get the "struct epitem" from a wait queue pointer */
285static inline struct epitem *ep_item_from_wait(wait_queue_t *p)
286{
287    return container_of(p, struct eppoll_entry, wait)->base;
288}
289
290/* Get the "struct epitem" from an epoll queue wrapper */
291static inline struct epitem *ep_item_from_epqueue(poll_table *p)
292{
293    return container_of(p, struct ep_pqueue, pt)->epi;
294}
295
296/* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
297static inline int ep_op_has_event(int op)
298{
299    return op != EPOLL_CTL_DEL;
300}
301
302/* Initialize the poll safe wake up structure */
303static void ep_nested_calls_init(struct nested_calls *ncalls)
304{
305    INIT_LIST_HEAD(&ncalls->tasks_call_list);
306    spin_lock_init(&ncalls->lock);
307}
308
309/**
310 * ep_call_nested - Perform a bound (possibly) nested call, by checking
311 * that the recursion limit is not exceeded, and that
312 * the same nested call (by the meaning of same cookie) is
313 * no re-entered.
314 *
315 * @ncalls: Pointer to the nested_calls structure to be used for this call.
316 * @max_nests: Maximum number of allowed nesting calls.
317 * @nproc: Nested call core function pointer.
318 * @priv: Opaque data to be passed to the @nproc callback.
319 * @cookie: Cookie to be used to identify this nested call.
320 * @ctx: This instance context.
321 *
322 * Returns: Returns the code returned by the @nproc callback, or -1 if
323 * the maximum recursion limit has been exceeded.
324 */
325static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
326              int (*nproc)(void *, void *, int), void *priv,
327              void *cookie, void *ctx)
328{
329    int error, call_nests = 0;
330    unsigned long flags;
331    struct list_head *lsthead = &ncalls->tasks_call_list;
332    struct nested_call_node *tncur;
333    struct nested_call_node tnode;
334
335    spin_lock_irqsave(&ncalls->lock, flags);
336
337    /*
338     * Try to see if the current task is already inside this wakeup call.
339     * We use a list here, since the population inside this set is always
340     * very much limited.
341     */
342    list_for_each_entry(tncur, lsthead, llink) {
343        if (tncur->ctx == ctx &&
344            (tncur->cookie == cookie || ++call_nests > max_nests)) {
345            /*
346             * Ops ... loop detected or maximum nest level reached.
347             * We abort this wake by breaking the cycle itself.
348             */
349            error = -1;
350            goto out_unlock;
351        }
352    }
353
354    /* Add the current task and cookie to the list */
355    tnode.ctx = ctx;
356    tnode.cookie = cookie;
357    list_add(&tnode.llink, lsthead);
358
359    spin_unlock_irqrestore(&ncalls->lock, flags);
360
361    /* Call the nested function */
362    error = (*nproc)(priv, cookie, call_nests);
363
364    /* Remove the current task from the list */
365    spin_lock_irqsave(&ncalls->lock, flags);
366    list_del(&tnode.llink);
367out_unlock:
368    spin_unlock_irqrestore(&ncalls->lock, flags);
369
370    return error;
371}
372
373#ifdef CONFIG_DEBUG_LOCK_ALLOC
374static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
375                     unsigned long events, int subclass)
376{
377    unsigned long flags;
378
379    spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
380    wake_up_locked_poll(wqueue, events);
381    spin_unlock_irqrestore(&wqueue->lock, flags);
382}
383#else
384static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
385                     unsigned long events, int subclass)
386{
387    wake_up_poll(wqueue, events);
388}
389#endif
390
391static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
392{
393    ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN,
394              1 + call_nests);
395    return 0;
396}
397
398/*
399 * Perform a safe wake up of the poll wait list. The problem is that
400 * with the new callback'd wake up system, it is possible that the
401 * poll callback is reentered from inside the call to wake_up() done
402 * on the poll wait queue head. The rule is that we cannot reenter the
403 * wake up code from the same task more than EP_MAX_NESTS times,
404 * and we cannot reenter the same wait queue head at all. This will
405 * enable to have a hierarchy of epoll file descriptor of no more than
406 * EP_MAX_NESTS deep.
407 */
408static void ep_poll_safewake(wait_queue_head_t *wq)
409{
410    int this_cpu = get_cpu();
411
412    ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
413               ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
414
415    put_cpu();
416}
417
418/*
419 * This function unregisters poll callbacks from the associated file
420 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
421 * ep_free).
422 */
423static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
424{
425    struct list_head *lsthead = &epi->pwqlist;
426    struct eppoll_entry *pwq;
427
428    while (!list_empty(lsthead)) {
429        pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
430
431        list_del(&pwq->llink);
432        remove_wait_queue(pwq->whead, &pwq->wait);
433        kmem_cache_free(pwq_cache, pwq);
434    }
435}
436
437/**
438 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
439 * the scan code, to call f_op->poll(). Also allows for
440 * O(NumReady) performance.
441 *
442 * @ep: Pointer to the epoll private data structure.
443 * @sproc: Pointer to the scan callback.
444 * @priv: Private opaque data passed to the @sproc callback.
445 *
446 * Returns: The same integer error code returned by the @sproc callback.
447 */
448static int ep_scan_ready_list(struct eventpoll *ep,
449                  int (*sproc)(struct eventpoll *,
450                       struct list_head *, void *),
451                  void *priv)
452{
453    int error, pwake = 0;
454    unsigned long flags;
455    struct epitem *epi, *nepi;
456    LIST_HEAD(txlist);
457
458    /*
459     * We need to lock this because we could be hit by
460     * eventpoll_release_file() and epoll_ctl().
461     */
462    mutex_lock(&ep->mtx);
463
464    /*
465     * Steal the ready list, and re-init the original one to the
466     * empty list. Also, set ep->ovflist to NULL so that events
467     * happening while looping w/out locks, are not lost. We cannot
468     * have the poll callback to queue directly on ep->rdllist,
469     * because we want the "sproc" callback to be able to do it
470     * in a lockless way.
471     */
472    spin_lock_irqsave(&ep->lock, flags);
473    list_splice_init(&ep->rdllist, &txlist);
474    ep->ovflist = NULL;
475    spin_unlock_irqrestore(&ep->lock, flags);
476
477    /*
478     * Now call the callback function.
479     */
480    error = (*sproc)(ep, &txlist, priv);
481
482    spin_lock_irqsave(&ep->lock, flags);
483    /*
484     * During the time we spent inside the "sproc" callback, some
485     * other events might have been queued by the poll callback.
486     * We re-insert them inside the main ready-list here.
487     */
488    for (nepi = ep->ovflist; (epi = nepi) != NULL;
489         nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
490        /*
491         * We need to check if the item is already in the list.
492         * During the "sproc" callback execution time, items are
493         * queued into ->ovflist but the "txlist" might already
494         * contain them, and the list_splice() below takes care of them.
495         */
496        if (!ep_is_linked(&epi->rdllink))
497            list_add_tail(&epi->rdllink, &ep->rdllist);
498    }
499    /*
500     * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
501     * releasing the lock, events will be queued in the normal way inside
502     * ep->rdllist.
503     */
504    ep->ovflist = EP_UNACTIVE_PTR;
505
506    /*
507     * Quickly re-inject items left on "txlist".
508     */
509    list_splice(&txlist, &ep->rdllist);
510
511    if (!list_empty(&ep->rdllist)) {
512        /*
513         * Wake up (if active) both the eventpoll wait list and
514         * the ->poll() wait list (delayed after we release the lock).
515         */
516        if (waitqueue_active(&ep->wq))
517            wake_up_locked(&ep->wq);
518        if (waitqueue_active(&ep->poll_wait))
519            pwake++;
520    }
521    spin_unlock_irqrestore(&ep->lock, flags);
522
523    mutex_unlock(&ep->mtx);
524
525    /* We have to call this outside the lock */
526    if (pwake)
527        ep_poll_safewake(&ep->poll_wait);
528
529    return error;
530}
531
532/*
533 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
534 * all the associated resources. Must be called with "mtx" held.
535 */
536static int ep_remove(struct eventpoll *ep, struct epitem *epi)
537{
538    unsigned long flags;
539    struct file *file = epi->ffd.file;
540
541    /*
542     * Removes poll wait queue hooks. We _have_ to do this without holding
543     * the "ep->lock" otherwise a deadlock might occur. This because of the
544     * sequence of the lock acquisition. Here we do "ep->lock" then the wait
545     * queue head lock when unregistering the wait queue. The wakeup callback
546     * will run by holding the wait queue head lock and will call our callback
547     * that will try to get "ep->lock".
548     */
549    ep_unregister_pollwait(ep, epi);
550
551    /* Remove the current item from the list of epoll hooks */
552    spin_lock(&file->f_lock);
553    if (ep_is_linked(&epi->fllink))
554        list_del_init(&epi->fllink);
555    spin_unlock(&file->f_lock);
556
557    rb_erase(&epi->rbn, &ep->rbr);
558
559    spin_lock_irqsave(&ep->lock, flags);
560    if (ep_is_linked(&epi->rdllink))
561        list_del_init(&epi->rdllink);
562    spin_unlock_irqrestore(&ep->lock, flags);
563
564    /* At this point it is safe to free the eventpoll item */
565    kmem_cache_free(epi_cache, epi);
566
567    atomic_dec(&ep->user->epoll_watches);
568
569    return 0;
570}
571
572static void ep_free(struct eventpoll *ep)
573{
574    struct rb_node *rbp;
575    struct epitem *epi;
576
577    /* We need to release all tasks waiting for these file */
578    if (waitqueue_active(&ep->poll_wait))
579        ep_poll_safewake(&ep->poll_wait);
580
581    /*
582     * We need to lock this because we could be hit by
583     * eventpoll_release_file() while we're freeing the "struct eventpoll".
584     * We do not need to hold "ep->mtx" here because the epoll file
585     * is on the way to be removed and no one has references to it
586     * anymore. The only hit might come from eventpoll_release_file() but
587     * holding "epmutex" is sufficent here.
588     */
589    mutex_lock(&epmutex);
590
591    /*
592     * Walks through the whole tree by unregistering poll callbacks.
593     */
594    for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
595        epi = rb_entry(rbp, struct epitem, rbn);
596
597        ep_unregister_pollwait(ep, epi);
598    }
599
600    /*
601     * Walks through the whole tree by freeing each "struct epitem". At this
602     * point we are sure no poll callbacks will be lingering around, and also by
603     * holding "epmutex" we can be sure that no file cleanup code will hit
604     * us during this operation. So we can avoid the lock on "ep->lock".
605     */
606    while ((rbp = rb_first(&ep->rbr)) != NULL) {
607        epi = rb_entry(rbp, struct epitem, rbn);
608        ep_remove(ep, epi);
609    }
610
611    mutex_unlock(&epmutex);
612    mutex_destroy(&ep->mtx);
613    free_uid(ep->user);
614    kfree(ep);
615}
616
617static int ep_eventpoll_release(struct inode *inode, struct file *file)
618{
619    struct eventpoll *ep = file->private_data;
620
621    if (ep)
622        ep_free(ep);
623
624    return 0;
625}
626
627static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
628                   void *priv)
629{
630    struct epitem *epi, *tmp;
631
632    list_for_each_entry_safe(epi, tmp, head, rdllink) {
633        if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
634            epi->event.events)
635            return POLLIN | POLLRDNORM;
636        else {
637            /*
638             * Item has been dropped into the ready list by the poll
639             * callback, but it's not actually ready, as far as
640             * caller requested events goes. We can remove it here.
641             */
642            list_del_init(&epi->rdllink);
643        }
644    }
645
646    return 0;
647}
648
649static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests)
650{
651    return ep_scan_ready_list(priv, ep_read_events_proc, NULL);
652}
653
654static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
655{
656    int pollflags;
657    struct eventpoll *ep = file->private_data;
658
659    /* Insert inside our poll wait queue */
660    poll_wait(file, &ep->poll_wait, wait);
661
662    /*
663     * Proceed to find out if wanted events are really available inside
664     * the ready list. This need to be done under ep_call_nested()
665     * supervision, since the call to f_op->poll() done on listed files
666     * could re-enter here.
667     */
668    pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS,
669                   ep_poll_readyevents_proc, ep, ep, current);
670
671    return pollflags != -1 ? pollflags : 0;
672}
673
674/* File callbacks that implement the eventpoll file behaviour */
675static const struct file_operations eventpoll_fops = {
676    .release = ep_eventpoll_release,
677    .poll = ep_eventpoll_poll
678};
679
680/* Fast test to see if the file is an evenpoll file */
681static inline int is_file_epoll(struct file *f)
682{
683    return f->f_op == &eventpoll_fops;
684}
685
686/*
687 * This is called from eventpoll_release() to unlink files from the eventpoll
688 * interface. We need to have this facility to cleanup correctly files that are
689 * closed without being removed from the eventpoll interface.
690 */
691void eventpoll_release_file(struct file *file)
692{
693    struct list_head *lsthead = &file->f_ep_links;
694    struct eventpoll *ep;
695    struct epitem *epi;
696
697    /*
698     * We don't want to get "file->f_lock" because it is not
699     * necessary. It is not necessary because we're in the "struct file"
700     * cleanup path, and this means that noone is using this file anymore.
701     * So, for example, epoll_ctl() cannot hit here since if we reach this
702     * point, the file counter already went to zero and fget() would fail.
703     * The only hit might come from ep_free() but by holding the mutex
704     * will correctly serialize the operation. We do need to acquire
705     * "ep->mtx" after "epmutex" because ep_remove() requires it when called
706     * from anywhere but ep_free().
707     *
708     * Besides, ep_remove() acquires the lock, so we can't hold it here.
709     */
710    mutex_lock(&epmutex);
711
712    while (!list_empty(lsthead)) {
713        epi = list_first_entry(lsthead, struct epitem, fllink);
714
715        ep = epi->ep;
716        list_del_init(&epi->fllink);
717        mutex_lock(&ep->mtx);
718        ep_remove(ep, epi);
719        mutex_unlock(&ep->mtx);
720    }
721
722    mutex_unlock(&epmutex);
723}
724
725static int ep_alloc(struct eventpoll **pep)
726{
727    int error;
728    struct user_struct *user;
729    struct eventpoll *ep;
730
731    user = get_current_user();
732    error = -ENOMEM;
733    ep = kzalloc(sizeof(*ep), GFP_KERNEL);
734    if (unlikely(!ep))
735        goto free_uid;
736
737    spin_lock_init(&ep->lock);
738    mutex_init(&ep->mtx);
739    init_waitqueue_head(&ep->wq);
740    init_waitqueue_head(&ep->poll_wait);
741    INIT_LIST_HEAD(&ep->rdllist);
742    ep->rbr = RB_ROOT;
743    ep->ovflist = EP_UNACTIVE_PTR;
744    ep->user = user;
745
746    *pep = ep;
747
748    return 0;
749
750free_uid:
751    free_uid(user);
752    return error;
753}
754
755/*
756 * Search the file inside the eventpoll tree. The RB tree operations
757 * are protected by the "mtx" mutex, and ep_find() must be called with
758 * "mtx" held.
759 */
760static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
761{
762    int kcmp;
763    struct rb_node *rbp;
764    struct epitem *epi, *epir = NULL;
765    struct epoll_filefd ffd;
766
767    ep_set_ffd(&ffd, file, fd);
768    for (rbp = ep->rbr.rb_node; rbp; ) {
769        epi = rb_entry(rbp, struct epitem, rbn);
770        kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
771        if (kcmp > 0)
772            rbp = rbp->rb_right;
773        else if (kcmp < 0)
774            rbp = rbp->rb_left;
775        else {
776            epir = epi;
777            break;
778        }
779    }
780
781    return epir;
782}
783
784/*
785 * This is the callback that is passed to the wait queue wakeup
786 * machanism. It is called by the stored file descriptors when they
787 * have events to report.
788 */
789static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key)
790{
791    int pwake = 0;
792    unsigned long flags;
793    struct epitem *epi = ep_item_from_wait(wait);
794    struct eventpoll *ep = epi->ep;
795
796    spin_lock_irqsave(&ep->lock, flags);
797
798    /*
799     * If the event mask does not contain any poll(2) event, we consider the
800     * descriptor to be disabled. This condition is likely the effect of the
801     * EPOLLONESHOT bit that disables the descriptor when an event is received,
802     * until the next EPOLL_CTL_MOD will be issued.
803     */
804    if (!(epi->event.events & ~EP_PRIVATE_BITS))
805        goto out_unlock;
806
807    /*
808     * Check the events coming with the callback. At this stage, not
809     * every device reports the events in the "key" parameter of the
810     * callback. We need to be able to handle both cases here, hence the
811     * test for "key" != NULL before the event match test.
812     */
813    if (key && !((unsigned long) key & epi->event.events))
814        goto out_unlock;
815
816    /*
817     * If we are trasfering events to userspace, we can hold no locks
818     * (because we're accessing user memory, and because of linux f_op->poll()
819     * semantics). All the events that happens during that period of time are
820     * chained in ep->ovflist and requeued later on.
821     */
822    if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) {
823        if (epi->next == EP_UNACTIVE_PTR) {
824            epi->next = ep->ovflist;
825            ep->ovflist = epi;
826        }
827        goto out_unlock;
828    }
829
830    /* If this file is already in the ready list we exit soon */
831    if (!ep_is_linked(&epi->rdllink))
832        list_add_tail(&epi->rdllink, &ep->rdllist);
833
834    /*
835     * Wake up ( if active ) both the eventpoll wait list and the ->poll()
836     * wait list.
837     */
838    if (waitqueue_active(&ep->wq))
839        wake_up_locked(&ep->wq);
840    if (waitqueue_active(&ep->poll_wait))
841        pwake++;
842
843out_unlock:
844    spin_unlock_irqrestore(&ep->lock, flags);
845
846    /* We have to call this outside the lock */
847    if (pwake)
848        ep_poll_safewake(&ep->poll_wait);
849
850    return 1;
851}
852
853/*
854 * This is the callback that is used to add our wait queue to the
855 * target file wakeup lists.
856 */
857static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
858                 poll_table *pt)
859{
860    struct epitem *epi = ep_item_from_epqueue(pt);
861    struct eppoll_entry *pwq;
862
863    if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
864        init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
865        pwq->whead = whead;
866        pwq->base = epi;
867        add_wait_queue(whead, &pwq->wait);
868        list_add_tail(&pwq->llink, &epi->pwqlist);
869        epi->nwait++;
870    } else {
871        /* We have to signal that an error occurred */
872        epi->nwait = -1;
873    }
874}
875
876static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
877{
878    int kcmp;
879    struct rb_node **p = &ep->rbr.rb_node, *parent = NULL;
880    struct epitem *epic;
881
882    while (*p) {
883        parent = *p;
884        epic = rb_entry(parent, struct epitem, rbn);
885        kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
886        if (kcmp > 0)
887            p = &parent->rb_right;
888        else
889            p = &parent->rb_left;
890    }
891    rb_link_node(&epi->rbn, parent, p);
892    rb_insert_color(&epi->rbn, &ep->rbr);
893}
894
895/*
896 * Must be called with "mtx" held.
897 */
898static int ep_insert(struct eventpoll *ep, struct epoll_event *event,
899             struct file *tfile, int fd)
900{
901    int error, revents, pwake = 0;
902    unsigned long flags;
903    struct epitem *epi;
904    struct ep_pqueue epq;
905
906    if (unlikely(atomic_read(&ep->user->epoll_watches) >=
907             max_user_watches))
908        return -ENOSPC;
909    if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
910        return -ENOMEM;
911
912    /* Item initialization follow here ... */
913    INIT_LIST_HEAD(&epi->rdllink);
914    INIT_LIST_HEAD(&epi->fllink);
915    INIT_LIST_HEAD(&epi->pwqlist);
916    epi->ep = ep;
917    ep_set_ffd(&epi->ffd, tfile, fd);
918    epi->event = *event;
919    epi->nwait = 0;
920    epi->next = EP_UNACTIVE_PTR;
921
922    /* Initialize the poll table using the queue callback */
923    epq.epi = epi;
924    init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
925
926    /*
927     * Attach the item to the poll hooks and get current event bits.
928     * We can safely use the file* here because its usage count has
929     * been increased by the caller of this function. Note that after
930     * this operation completes, the poll callback can start hitting
931     * the new item.
932     */
933    revents = tfile->f_op->poll(tfile, &epq.pt);
934
935    /*
936     * We have to check if something went wrong during the poll wait queue
937     * install process. Namely an allocation for a wait queue failed due
938     * high memory pressure.
939     */
940    error = -ENOMEM;
941    if (epi->nwait < 0)
942        goto error_unregister;
943
944    /* Add the current item to the list of active epoll hook for this file */
945    spin_lock(&tfile->f_lock);
946    list_add_tail(&epi->fllink, &tfile->f_ep_links);
947    spin_unlock(&tfile->f_lock);
948
949    /*
950     * Add the current item to the RB tree. All RB tree operations are
951     * protected by "mtx", and ep_insert() is called with "mtx" held.
952     */
953    ep_rbtree_insert(ep, epi);
954
955    /* We have to drop the new item inside our item list to keep track of it */
956    spin_lock_irqsave(&ep->lock, flags);
957
958    /* If the file is already "ready" we drop it inside the ready list */
959    if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) {
960        list_add_tail(&epi->rdllink, &ep->rdllist);
961
962        /* Notify waiting tasks that events are available */
963        if (waitqueue_active(&ep->wq))
964            wake_up_locked(&ep->wq);
965        if (waitqueue_active(&ep->poll_wait))
966            pwake++;
967    }
968
969    spin_unlock_irqrestore(&ep->lock, flags);
970
971    atomic_inc(&ep->user->epoll_watches);
972
973    /* We have to call this outside the lock */
974    if (pwake)
975        ep_poll_safewake(&ep->poll_wait);
976
977    return 0;
978
979error_unregister:
980    ep_unregister_pollwait(ep, epi);
981
982    /*
983     * We need to do this because an event could have been arrived on some
984     * allocated wait queue. Note that we don't care about the ep->ovflist
985     * list, since that is used/cleaned only inside a section bound by "mtx".
986     * And ep_insert() is called with "mtx" held.
987     */
988    spin_lock_irqsave(&ep->lock, flags);
989    if (ep_is_linked(&epi->rdllink))
990        list_del_init(&epi->rdllink);
991    spin_unlock_irqrestore(&ep->lock, flags);
992
993    kmem_cache_free(epi_cache, epi);
994
995    return error;
996}
997
998/*
999 * Modify the interest event mask by dropping an event if the new mask
1000 * has a match in the current file status. Must be called with "mtx" held.
1001 */
1002static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event)
1003{
1004    int pwake = 0;
1005    unsigned int revents;
1006
1007    /*
1008     * Set the new event interest mask before calling f_op->poll();
1009     * otherwise we might miss an event that happens between the
1010     * f_op->poll() call and the new event set registering.
1011     */
1012    epi->event.events = event->events;
1013    epi->event.data = event->data; /* protected by mtx */
1014
1015    /*
1016     * Get current event bits. We can safely use the file* here because
1017     * its usage count has been increased by the caller of this function.
1018     */
1019    revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL);
1020
1021    /*
1022     * If the item is "hot" and it is not registered inside the ready
1023     * list, push it inside.
1024     */
1025    if (revents & event->events) {
1026        spin_lock_irq(&ep->lock);
1027        if (!ep_is_linked(&epi->rdllink)) {
1028            list_add_tail(&epi->rdllink, &ep->rdllist);
1029
1030            /* Notify waiting tasks that events are available */
1031            if (waitqueue_active(&ep->wq))
1032                wake_up_locked(&ep->wq);
1033            if (waitqueue_active(&ep->poll_wait))
1034                pwake++;
1035        }
1036        spin_unlock_irq(&ep->lock);
1037    }
1038
1039    /* We have to call this outside the lock */
1040    if (pwake)
1041        ep_poll_safewake(&ep->poll_wait);
1042
1043    return 0;
1044}
1045
1046static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1047                   void *priv)
1048{
1049    struct ep_send_events_data *esed = priv;
1050    int eventcnt;
1051    unsigned int revents;
1052    struct epitem *epi;
1053    struct epoll_event __user *uevent;
1054
1055    /*
1056     * We can loop without lock because we are passed a task private list.
1057     * Items cannot vanish during the loop because ep_scan_ready_list() is
1058     * holding "mtx" during this call.
1059     */
1060    for (eventcnt = 0, uevent = esed->events;
1061         !list_empty(head) && eventcnt < esed->maxevents;) {
1062        epi = list_first_entry(head, struct epitem, rdllink);
1063
1064        list_del_init(&epi->rdllink);
1065
1066        revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) &
1067            epi->event.events;
1068
1069        /*
1070         * If the event mask intersect the caller-requested one,
1071         * deliver the event to userspace. Again, ep_scan_ready_list()
1072         * is holding "mtx", so no operations coming from userspace
1073         * can change the item.
1074         */
1075        if (revents) {
1076            if (__put_user(revents, &uevent->events) ||
1077                __put_user(epi->event.data, &uevent->data)) {
1078                list_add(&epi->rdllink, head);
1079                return eventcnt ? eventcnt : -EFAULT;
1080            }
1081            eventcnt++;
1082            uevent++;
1083            if (epi->event.events & EPOLLONESHOT)
1084                epi->event.events &= EP_PRIVATE_BITS;
1085            else if (!(epi->event.events & EPOLLET)) {
1086                /*
1087                 * If this file has been added with Level
1088                 * Trigger mode, we need to insert back inside
1089                 * the ready list, so that the next call to
1090                 * epoll_wait() will check again the events
1091                 * availability. At this point, noone can insert
1092                 * into ep->rdllist besides us. The epoll_ctl()
1093                 * callers are locked out by
1094                 * ep_scan_ready_list() holding "mtx" and the
1095                 * poll callback will queue them in ep->ovflist.
1096                 */
1097                list_add_tail(&epi->rdllink, &ep->rdllist);
1098            }
1099        }
1100    }
1101
1102    return eventcnt;
1103}
1104
1105static int ep_send_events(struct eventpoll *ep,
1106              struct epoll_event __user *events, int maxevents)
1107{
1108    struct ep_send_events_data esed;
1109
1110    esed.maxevents = maxevents;
1111    esed.events = events;
1112
1113    return ep_scan_ready_list(ep, ep_send_events_proc, &esed);
1114}
1115
1116static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1117           int maxevents, long timeout)
1118{
1119    int res, eavail;
1120    unsigned long flags;
1121    long jtimeout;
1122    wait_queue_t wait;
1123
1124    /*
1125     * Calculate the timeout by checking for the "infinite" value (-1)
1126     * and the overflow condition. The passed timeout is in milliseconds,
1127     * that why (t * HZ) / 1000.
1128     */
1129    jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
1130        MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
1131
1132retry:
1133    spin_lock_irqsave(&ep->lock, flags);
1134
1135    res = 0;
1136    if (list_empty(&ep->rdllist)) {
1137        /*
1138         * We don't have any available event to return to the caller.
1139         * We need to sleep here, and we will be wake up by
1140         * ep_poll_callback() when events will become available.
1141         */
1142        init_waitqueue_entry(&wait, current);
1143        wait.flags |= WQ_FLAG_EXCLUSIVE;
1144        __add_wait_queue(&ep->wq, &wait);
1145
1146        for (;;) {
1147            /*
1148             * We don't want to sleep if the ep_poll_callback() sends us
1149             * a wakeup in between. That's why we set the task state
1150             * to TASK_INTERRUPTIBLE before doing the checks.
1151             */
1152            set_current_state(TASK_INTERRUPTIBLE);
1153            if (!list_empty(&ep->rdllist) || !jtimeout)
1154                break;
1155            if (signal_pending(current)) {
1156                res = -EINTR;
1157                break;
1158            }
1159
1160            spin_unlock_irqrestore(&ep->lock, flags);
1161            jtimeout = schedule_timeout(jtimeout);
1162            spin_lock_irqsave(&ep->lock, flags);
1163        }
1164        __remove_wait_queue(&ep->wq, &wait);
1165
1166        set_current_state(TASK_RUNNING);
1167    }
1168    /* Is it worth to try to dig for events ? */
1169    eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
1170
1171    spin_unlock_irqrestore(&ep->lock, flags);
1172
1173    /*
1174     * Try to transfer events to user space. In case we get 0 events and
1175     * there's still timeout left over, we go trying again in search of
1176     * more luck.
1177     */
1178    if (!res && eavail &&
1179        !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
1180        goto retry;
1181
1182    return res;
1183}
1184
1185/*
1186 * Open an eventpoll file descriptor.
1187 */
1188SYSCALL_DEFINE1(epoll_create1, int, flags)
1189{
1190    int error;
1191    struct eventpoll *ep = NULL;
1192
1193    /* Check the EPOLL_* constant for consistency. */
1194    BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1195
1196    if (flags & ~EPOLL_CLOEXEC)
1197        return -EINVAL;
1198    /*
1199     * Create the internal data structure ("struct eventpoll").
1200     */
1201    error = ep_alloc(&ep);
1202    if (error < 0)
1203        return error;
1204    /*
1205     * Creates all the items needed to setup an eventpoll file. That is,
1206     * a file structure and a free file descriptor.
1207     */
1208    error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep,
1209                 O_RDWR | (flags & O_CLOEXEC));
1210    if (error < 0)
1211        ep_free(ep);
1212
1213    return error;
1214}
1215
1216SYSCALL_DEFINE1(epoll_create, int, size)
1217{
1218    if (size <= 0)
1219        return -EINVAL;
1220
1221    return sys_epoll_create1(0);
1222}
1223
1224/*
1225 * The following function implements the controller interface for
1226 * the eventpoll file that enables the insertion/removal/change of
1227 * file descriptors inside the interest set.
1228 */
1229SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1230        struct epoll_event __user *, event)
1231{
1232    int error;
1233    struct file *file, *tfile;
1234    struct eventpoll *ep;
1235    struct epitem *epi;
1236    struct epoll_event epds;
1237
1238    error = -EFAULT;
1239    if (ep_op_has_event(op) &&
1240        copy_from_user(&epds, event, sizeof(struct epoll_event)))
1241        goto error_return;
1242
1243    /* Get the "struct file *" for the eventpoll file */
1244    error = -EBADF;
1245    file = fget(epfd);
1246    if (!file)
1247        goto error_return;
1248
1249    /* Get the "struct file *" for the target file */
1250    tfile = fget(fd);
1251    if (!tfile)
1252        goto error_fput;
1253
1254    /* The target file descriptor must support poll */
1255    error = -EPERM;
1256    if (!tfile->f_op || !tfile->f_op->poll)
1257        goto error_tgt_fput;
1258
1259    /*
1260     * We have to check that the file structure underneath the file descriptor
1261     * the user passed to us _is_ an eventpoll file. And also we do not permit
1262     * adding an epoll file descriptor inside itself.
1263     */
1264    error = -EINVAL;
1265    if (file == tfile || !is_file_epoll(file))
1266        goto error_tgt_fput;
1267
1268    /*
1269     * At this point it is safe to assume that the "private_data" contains
1270     * our own data structure.
1271     */
1272    ep = file->private_data;
1273
1274    mutex_lock(&ep->mtx);
1275
1276    /*
1277     * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
1278     * above, we can be sure to be able to use the item looked up by
1279     * ep_find() till we release the mutex.
1280     */
1281    epi = ep_find(ep, tfile, fd);
1282
1283    error = -EINVAL;
1284    switch (op) {
1285    case EPOLL_CTL_ADD:
1286        if (!epi) {
1287            epds.events |= POLLERR | POLLHUP;
1288            error = ep_insert(ep, &epds, tfile, fd);
1289        } else
1290            error = -EEXIST;
1291        break;
1292    case EPOLL_CTL_DEL:
1293        if (epi)
1294            error = ep_remove(ep, epi);
1295        else
1296            error = -ENOENT;
1297        break;
1298    case EPOLL_CTL_MOD:
1299        if (epi) {
1300            epds.events |= POLLERR | POLLHUP;
1301            error = ep_modify(ep, epi, &epds);
1302        } else
1303            error = -ENOENT;
1304        break;
1305    }
1306    mutex_unlock(&ep->mtx);
1307
1308error_tgt_fput:
1309    fput(tfile);
1310error_fput:
1311    fput(file);
1312error_return:
1313
1314    return error;
1315}
1316
1317/*
1318 * Implement the event wait interface for the eventpoll file. It is the kernel
1319 * part of the user space epoll_wait(2).
1320 */
1321SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
1322        int, maxevents, int, timeout)
1323{
1324    int error;
1325    struct file *file;
1326    struct eventpoll *ep;
1327
1328    /* The maximum number of event must be greater than zero */
1329    if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
1330        return -EINVAL;
1331
1332    /* Verify that the area passed by the user is writeable */
1333    if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) {
1334        error = -EFAULT;
1335        goto error_return;
1336    }
1337
1338    /* Get the "struct file *" for the eventpoll file */
1339    error = -EBADF;
1340    file = fget(epfd);
1341    if (!file)
1342        goto error_return;
1343
1344    /*
1345     * We have to check that the file structure underneath the fd
1346     * the user passed to us _is_ an eventpoll file.
1347     */
1348    error = -EINVAL;
1349    if (!is_file_epoll(file))
1350        goto error_fput;
1351
1352    /*
1353     * At this point it is safe to assume that the "private_data" contains
1354     * our own data structure.
1355     */
1356    ep = file->private_data;
1357
1358    /* Time to fish for events ... */
1359    error = ep_poll(ep, events, maxevents, timeout);
1360
1361error_fput:
1362    fput(file);
1363error_return:
1364
1365    return error;
1366}
1367
1368#ifdef HAVE_SET_RESTORE_SIGMASK
1369
1370/*
1371 * Implement the event wait interface for the eventpoll file. It is the kernel
1372 * part of the user space epoll_pwait(2).
1373 */
1374SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
1375        int, maxevents, int, timeout, const sigset_t __user *, sigmask,
1376        size_t, sigsetsize)
1377{
1378    int error;
1379    sigset_t ksigmask, sigsaved;
1380
1381    /*
1382     * If the caller wants a certain signal mask to be set during the wait,
1383     * we apply it here.
1384     */
1385    if (sigmask) {
1386        if (sigsetsize != sizeof(sigset_t))
1387            return -EINVAL;
1388        if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
1389            return -EFAULT;
1390        sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
1391        sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
1392    }
1393
1394    error = sys_epoll_wait(epfd, events, maxevents, timeout);
1395
1396    /*
1397     * If we changed the signal mask, we need to restore the original one.
1398     * In case we've got a signal while waiting, we do not restore the
1399     * signal mask yet, and we allow do_signal() to deliver the signal on
1400     * the way back to userspace, before the signal mask is restored.
1401     */
1402    if (sigmask) {
1403        if (error == -EINTR) {
1404            memcpy(&current->saved_sigmask, &sigsaved,
1405                   sizeof(sigsaved));
1406            set_restore_sigmask();
1407        } else
1408            sigprocmask(SIG_SETMASK, &sigsaved, NULL);
1409    }
1410
1411    return error;
1412}
1413
1414#endif /* HAVE_SET_RESTORE_SIGMASK */
1415
1416static int __init eventpoll_init(void)
1417{
1418    struct sysinfo si;
1419
1420    si_meminfo(&si);
1421    /*
1422     * Allows top 4% of lomem to be allocated for epoll watches (per user).
1423     */
1424    max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
1425        EP_ITEM_COST;
1426
1427    /* Initialize the structure used to perform safe poll wait head wake ups */
1428    ep_nested_calls_init(&poll_safewake_ncalls);
1429
1430    /* Initialize the structure used to perform file's f_op->poll() calls */
1431    ep_nested_calls_init(&poll_readywalk_ncalls);
1432
1433    /* Allocates slab cache used to allocate "struct epitem" items */
1434    epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
1435            0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
1436
1437    /* Allocates slab cache used to allocate "struct eppoll_entry" */
1438    pwq_cache = kmem_cache_create("eventpoll_pwq",
1439            sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
1440
1441    return 0;
1442}
1443fs_initcall(eventpoll_init);
1444

Archive Download this file



interactive