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/signal.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 <linux/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 #include <net/busy_poll.h>
46
47 /*
48 * LOCKING:
49 * There are three level of locking required by epoll :
50 *
51 * 1) epmutex (mutex)
52 * 2) ep->mtx (mutex)
53 * 3) ep->wq.lock (spinlock)
54 *
55 * The acquire order is the one listed above, from 1 to 3.
56 * We need a spinlock (ep->wq.lock) because we manipulate objects
57 * from inside the poll callback, that might be triggered from
58 * a wake_up() that in turn might be called from IRQ context.
59 * So we can't sleep inside the poll callback and hence we need
60 * a spinlock. During the event transfer loop (from kernel to
61 * user space) we could end up sleeping due a copy_to_user(), so
62 * we need a lock that will allow us to sleep. This lock is a
63 * mutex (ep->mtx). It is acquired during the event transfer loop,
64 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().
65 * Then we also need a global mutex to serialize eventpoll_release_file()
66 * and ep_free().
67 * This mutex is acquired by ep_free() during the epoll file
68 * cleanup path and it is also acquired by eventpoll_release_file()
69 * if a file has been pushed inside an epoll set and it is then
70 * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL).
71 * It is also acquired when inserting an epoll fd onto another epoll
72 * fd. We do this so that we walk the epoll tree and ensure that this
73 * insertion does not create a cycle of epoll file descriptors, which
74 * could lead to deadlock. We need a global mutex to prevent two
75 * simultaneous inserts (A into B and B into A) from racing and
76 * constructing a cycle without either insert observing that it is
77 * going to.
78 * It is necessary to acquire multiple "ep->mtx"es at once in the
79 * case when one epoll fd is added to another. In this case, we
80 * always acquire the locks in the order of nesting (i.e. after
81 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired
82 * before e2->mtx). Since we disallow cycles of epoll file
83 * descriptors, this ensures that the mutexes are well-ordered. In
84 * order to communicate this nesting to lockdep, when walking a tree
85 * of epoll file descriptors, we use the current recursion depth as
86 * the lockdep subkey.
87 * It is possible to drop the "ep->mtx" and to use the global
88 * mutex "epmutex" (together with "ep->wq.lock") to have it working,
89 * but having "ep->mtx" will make the interface more scalable.
90 * Events that require holding "epmutex" are very rare, while for
91 * normal operations the epoll private "ep->mtx" will guarantee
92 * a better scalability.
93 */
94
95 /* Epoll private bits inside the event mask */
96 #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
97
98 #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
99
100 #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
101 EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
102
103 /* Maximum number of nesting allowed inside epoll sets */
104 #define EP_MAX_NESTS 4
105
106 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
107
108 #define EP_UNACTIVE_PTR ((void *) -1L)
109
110 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
111
112 struct epoll_filefd {
113 struct file *file;
114 int fd;
115 } __packed;
116
117 /*
118 * Structure used to track possible nested calls, for too deep recursions
119 * and loop cycles.
120 */
121 struct nested_call_node {
122 struct list_head llink;
123 void *cookie;
124 void *ctx;
125 };
126
127 /*
128 * This structure is used as collector for nested calls, to check for
129 * maximum recursion dept and loop cycles.
130 */
131 struct nested_calls {
132 struct list_head tasks_call_list;
133 spinlock_t lock;
134 };
135
136 /*
137 * Each file descriptor added to the eventpoll interface will
138 * have an entry of this type linked to the "rbr" RB tree.
139 * Avoid increasing the size of this struct, there can be many thousands
140 * of these on a server and we do not want this to take another cache line.
141 */
142 struct epitem {
143 union {
144 /* RB tree node links this structure to the eventpoll RB tree */
145 struct rb_node rbn;
146 /* Used to free the struct epitem */
147 struct rcu_head rcu;
148 };
149
150 /* List header used to link this structure to the eventpoll ready list */
151 struct list_head rdllink;
152
153 /*
154 * Works together "struct eventpoll"->ovflist in keeping the
155 * single linked chain of items.
156 */
157 struct epitem *next;
158
159 /* The file descriptor information this item refers to */
160 struct epoll_filefd ffd;
161
162 /* Number of active wait queue attached to poll operations */
163 int nwait;
164
165 /* List containing poll wait queues */
166 struct list_head pwqlist;
167
168 /* The "container" of this item */
169 struct eventpoll *ep;
170
171 /* List header used to link this item to the "struct file" items list */
172 struct list_head fllink;
173
174 /* wakeup_source used when EPOLLWAKEUP is set */
175 struct wakeup_source __rcu *ws;
176
177 /* The structure that describe the interested events and the source fd */
178 struct epoll_event event;
179 };
180
181 /*
182 * This structure is stored inside the "private_data" member of the file
183 * structure and represents the main data structure for the eventpoll
184 * interface.
185 *
186 * Access to it is protected by the lock inside wq.
187 */
188 struct eventpoll {
189 /*
190 * This mutex is used to ensure that files are not removed
191 * while epoll is using them. This is held during the event
192 * collection loop, the file cleanup path, the epoll file exit
193 * code and the ctl operations.
194 */
195 struct mutex mtx;
196
197 /* Wait queue used by sys_epoll_wait() */
198 wait_queue_head_t wq;
199
200 /* Wait queue used by file->poll() */
201 wait_queue_head_t poll_wait;
202
203 /* List of ready file descriptors */
204 struct list_head rdllist;
205
206 /* RB tree root used to store monitored fd structs */
207 struct rb_root_cached rbr;
208
209 /*
210 * This is a single linked list that chains all the "struct epitem" that
211 * happened while transferring ready events to userspace w/out
212 * holding ->wq.lock.
213 */
214 struct epitem *ovflist;
215
216 /* wakeup_source used when ep_scan_ready_list is running */
217 struct wakeup_source *ws;
218
219 /* The user that created the eventpoll descriptor */
220 struct user_struct *user;
221
222 struct file *file;
223
224 /* used to optimize loop detection check */
225 u64 gen;
226
227 #ifdef CONFIG_NET_RX_BUSY_POLL
228 /* used to track busy poll napi_id */
229 unsigned int napi_id;
230 #endif
231 };
232
233 /* Wait structure used by the poll hooks */
234 struct eppoll_entry {
235 /* List header used to link this structure to the "struct epitem" */
236 struct list_head llink;
237
238 /* The "base" pointer is set to the container "struct epitem" */
239 struct epitem *base;
240
241 /*
242 * Wait queue item that will be linked to the target file wait
243 * queue head.
244 */
245 wait_queue_entry_t wait;
246
247 /* The wait queue head that linked the "wait" wait queue item */
248 wait_queue_head_t *whead;
249 };
250
251 /* Wrapper struct used by poll queueing */
252 struct ep_pqueue {
253 poll_table pt;
254 struct epitem *epi;
255 };
256
257 /* Used by the ep_send_events() function as callback private data */
258 struct ep_send_events_data {
259 int maxevents;
260 struct epoll_event __user *events;
261 int res;
262 };
263
264 /*
265 * Configuration options available inside /proc/sys/fs/epoll/
266 */
267 /* Maximum number of epoll watched descriptors, per user */
268 static long max_user_watches __read_mostly;
269
270 /*
271 * This mutex is used to serialize ep_free() and eventpoll_release_file().
272 */
273 static DEFINE_MUTEX(epmutex);
274
275 static u64 loop_check_gen = 0;
276
277 /* Used to check for epoll file descriptor inclusion loops */
278 static struct nested_calls poll_loop_ncalls;
279
280 /* Slab cache used to allocate "struct epitem" */
281 static struct kmem_cache *epi_cache __read_mostly;
282
283 /* Slab cache used to allocate "struct eppoll_entry" */
284 static struct kmem_cache *pwq_cache __read_mostly;
285
286 /*
287 * List of files with newly added links, where we may need to limit the number
288 * of emanating paths. Protected by the epmutex.
289 */
290 static LIST_HEAD(tfile_check_list);
291
292 #ifdef CONFIG_SYSCTL
293
294 #include <linux/sysctl.h>
295
296 static long zero;
297 static long long_max = LONG_MAX;
298
299 struct ctl_table epoll_table[] = {
300 {
301 .procname = "max_user_watches",
302 .data = &max_user_watches,
303 .maxlen = sizeof(max_user_watches),
304 .mode = 0644,
305 .proc_handler = proc_doulongvec_minmax,
306 .extra1 = &zero,
307 .extra2 = &long_max,
308 },
309 { }
310 };
311 #endif /* CONFIG_SYSCTL */
312
313 static const struct file_operations eventpoll_fops;
314
is_file_epoll(struct file * f)315 static inline int is_file_epoll(struct file *f)
316 {
317 return f->f_op == &eventpoll_fops;
318 }
319
320 /* Setup the structure that is used as key for the RB tree */
ep_set_ffd(struct epoll_filefd * ffd,struct file * file,int fd)321 static inline void ep_set_ffd(struct epoll_filefd *ffd,
322 struct file *file, int fd)
323 {
324 ffd->file = file;
325 ffd->fd = fd;
326 }
327
328 /* Compare RB tree keys */
ep_cmp_ffd(struct epoll_filefd * p1,struct epoll_filefd * p2)329 static inline int ep_cmp_ffd(struct epoll_filefd *p1,
330 struct epoll_filefd *p2)
331 {
332 return (p1->file > p2->file ? +1:
333 (p1->file < p2->file ? -1 : p1->fd - p2->fd));
334 }
335
336 /* Tells us if the item is currently linked */
ep_is_linked(struct epitem * epi)337 static inline int ep_is_linked(struct epitem *epi)
338 {
339 return !list_empty(&epi->rdllink);
340 }
341
ep_pwq_from_wait(wait_queue_entry_t * p)342 static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
343 {
344 return container_of(p, struct eppoll_entry, wait);
345 }
346
347 /* Get the "struct epitem" from a wait queue pointer */
ep_item_from_wait(wait_queue_entry_t * p)348 static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
349 {
350 return container_of(p, struct eppoll_entry, wait)->base;
351 }
352
353 /* Get the "struct epitem" from an epoll queue wrapper */
ep_item_from_epqueue(poll_table * p)354 static inline struct epitem *ep_item_from_epqueue(poll_table *p)
355 {
356 return container_of(p, struct ep_pqueue, pt)->epi;
357 }
358
359 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */
ep_op_has_event(int op)360 static inline int ep_op_has_event(int op)
361 {
362 return op != EPOLL_CTL_DEL;
363 }
364
365 /* Initialize the poll safe wake up structure */
ep_nested_calls_init(struct nested_calls * ncalls)366 static void ep_nested_calls_init(struct nested_calls *ncalls)
367 {
368 INIT_LIST_HEAD(&ncalls->tasks_call_list);
369 spin_lock_init(&ncalls->lock);
370 }
371
372 /**
373 * ep_events_available - Checks if ready events might be available.
374 *
375 * @ep: Pointer to the eventpoll context.
376 *
377 * Returns: Returns a value different than zero if ready events are available,
378 * or zero otherwise.
379 */
ep_events_available(struct eventpoll * ep)380 static inline int ep_events_available(struct eventpoll *ep)
381 {
382 return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR;
383 }
384
385 #ifdef CONFIG_NET_RX_BUSY_POLL
ep_busy_loop_end(void * p,unsigned long start_time)386 static bool ep_busy_loop_end(void *p, unsigned long start_time)
387 {
388 struct eventpoll *ep = p;
389
390 return ep_events_available(ep) || busy_loop_timeout(start_time);
391 }
392
393 /*
394 * Busy poll if globally on and supporting sockets found && no events,
395 * busy loop will return if need_resched or ep_events_available.
396 *
397 * we must do our busy polling with irqs enabled
398 */
ep_busy_loop(struct eventpoll * ep,int nonblock)399 static void ep_busy_loop(struct eventpoll *ep, int nonblock)
400 {
401 unsigned int napi_id = READ_ONCE(ep->napi_id);
402
403 if ((napi_id >= MIN_NAPI_ID) && net_busy_loop_on())
404 napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end, ep);
405 }
406
ep_reset_busy_poll_napi_id(struct eventpoll * ep)407 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
408 {
409 if (ep->napi_id)
410 ep->napi_id = 0;
411 }
412
413 /*
414 * Set epoll busy poll NAPI ID from sk.
415 */
ep_set_busy_poll_napi_id(struct epitem * epi)416 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
417 {
418 struct eventpoll *ep;
419 unsigned int napi_id;
420 struct socket *sock;
421 struct sock *sk;
422 int err;
423
424 if (!net_busy_loop_on())
425 return;
426
427 sock = sock_from_file(epi->ffd.file, &err);
428 if (!sock)
429 return;
430
431 sk = sock->sk;
432 if (!sk)
433 return;
434
435 napi_id = READ_ONCE(sk->sk_napi_id);
436 ep = epi->ep;
437
438 /* Non-NAPI IDs can be rejected
439 * or
440 * Nothing to do if we already have this ID
441 */
442 if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)
443 return;
444
445 /* record NAPI ID for use in next busy poll */
446 ep->napi_id = napi_id;
447 }
448
449 #else
450
ep_busy_loop(struct eventpoll * ep,int nonblock)451 static inline void ep_busy_loop(struct eventpoll *ep, int nonblock)
452 {
453 }
454
ep_reset_busy_poll_napi_id(struct eventpoll * ep)455 static inline void ep_reset_busy_poll_napi_id(struct eventpoll *ep)
456 {
457 }
458
ep_set_busy_poll_napi_id(struct epitem * epi)459 static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
460 {
461 }
462
463 #endif /* CONFIG_NET_RX_BUSY_POLL */
464
465 /**
466 * ep_call_nested - Perform a bound (possibly) nested call, by checking
467 * that the recursion limit is not exceeded, and that
468 * the same nested call (by the meaning of same cookie) is
469 * no re-entered.
470 *
471 * @ncalls: Pointer to the nested_calls structure to be used for this call.
472 * @max_nests: Maximum number of allowed nesting calls.
473 * @nproc: Nested call core function pointer.
474 * @priv: Opaque data to be passed to the @nproc callback.
475 * @cookie: Cookie to be used to identify this nested call.
476 * @ctx: This instance context.
477 *
478 * Returns: Returns the code returned by the @nproc callback, or -1 if
479 * the maximum recursion limit has been exceeded.
480 */
ep_call_nested(struct nested_calls * ncalls,int max_nests,int (* nproc)(void *,void *,int),void * priv,void * cookie,void * ctx)481 static int ep_call_nested(struct nested_calls *ncalls, int max_nests,
482 int (*nproc)(void *, void *, int), void *priv,
483 void *cookie, void *ctx)
484 {
485 int error, call_nests = 0;
486 unsigned long flags;
487 struct list_head *lsthead = &ncalls->tasks_call_list;
488 struct nested_call_node *tncur;
489 struct nested_call_node tnode;
490
491 spin_lock_irqsave(&ncalls->lock, flags);
492
493 /*
494 * Try to see if the current task is already inside this wakeup call.
495 * We use a list here, since the population inside this set is always
496 * very much limited.
497 */
498 list_for_each_entry(tncur, lsthead, llink) {
499 if (tncur->ctx == ctx &&
500 (tncur->cookie == cookie || ++call_nests > max_nests)) {
501 /*
502 * Ops ... loop detected or maximum nest level reached.
503 * We abort this wake by breaking the cycle itself.
504 */
505 error = -1;
506 goto out_unlock;
507 }
508 }
509
510 /* Add the current task and cookie to the list */
511 tnode.ctx = ctx;
512 tnode.cookie = cookie;
513 list_add(&tnode.llink, lsthead);
514
515 spin_unlock_irqrestore(&ncalls->lock, flags);
516
517 /* Call the nested function */
518 error = (*nproc)(priv, cookie, call_nests);
519
520 /* Remove the current task from the list */
521 spin_lock_irqsave(&ncalls->lock, flags);
522 list_del(&tnode.llink);
523 out_unlock:
524 spin_unlock_irqrestore(&ncalls->lock, flags);
525
526 return error;
527 }
528
529 /*
530 * As described in commit 0ccf831cb lockdep: annotate epoll
531 * the use of wait queues used by epoll is done in a very controlled
532 * manner. Wake ups can nest inside each other, but are never done
533 * with the same locking. For example:
534 *
535 * dfd = socket(...);
536 * efd1 = epoll_create();
537 * efd2 = epoll_create();
538 * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
539 * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
540 *
541 * When a packet arrives to the device underneath "dfd", the net code will
542 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
543 * callback wakeup entry on that queue, and the wake_up() performed by the
544 * "dfd" net code will end up in ep_poll_callback(). At this point epoll
545 * (efd1) notices that it may have some event ready, so it needs to wake up
546 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
547 * that ends up in another wake_up(), after having checked about the
548 * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to
549 * avoid stack blasting.
550 *
551 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
552 * this special case of epoll.
553 */
554 #ifdef CONFIG_DEBUG_LOCK_ALLOC
555
556 static struct nested_calls poll_safewake_ncalls;
557
ep_poll_wakeup_proc(void * priv,void * cookie,int call_nests)558 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests)
559 {
560 unsigned long flags;
561 wait_queue_head_t *wqueue = (wait_queue_head_t *)cookie;
562
563 spin_lock_irqsave_nested(&wqueue->lock, flags, call_nests + 1);
564 wake_up_locked_poll(wqueue, EPOLLIN);
565 spin_unlock_irqrestore(&wqueue->lock, flags);
566
567 return 0;
568 }
569
ep_poll_safewake(wait_queue_head_t * wq)570 static void ep_poll_safewake(wait_queue_head_t *wq)
571 {
572 int this_cpu = get_cpu();
573
574 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS,
575 ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu);
576
577 put_cpu();
578 }
579
580 #else
581
ep_poll_safewake(wait_queue_head_t * wq)582 static void ep_poll_safewake(wait_queue_head_t *wq)
583 {
584 wake_up_poll(wq, EPOLLIN);
585 }
586
587 #endif
588
ep_remove_wait_queue(struct eppoll_entry * pwq)589 static void ep_remove_wait_queue(struct eppoll_entry *pwq)
590 {
591 wait_queue_head_t *whead;
592
593 rcu_read_lock();
594 /*
595 * If it is cleared by POLLFREE, it should be rcu-safe.
596 * If we read NULL we need a barrier paired with
597 * smp_store_release() in ep_poll_callback(), otherwise
598 * we rely on whead->lock.
599 */
600 whead = smp_load_acquire(&pwq->whead);
601 if (whead)
602 remove_wait_queue(whead, &pwq->wait);
603 rcu_read_unlock();
604 }
605
606 /*
607 * This function unregisters poll callbacks from the associated file
608 * descriptor. Must be called with "mtx" held (or "epmutex" if called from
609 * ep_free).
610 */
ep_unregister_pollwait(struct eventpoll * ep,struct epitem * epi)611 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
612 {
613 struct list_head *lsthead = &epi->pwqlist;
614 struct eppoll_entry *pwq;
615
616 while (!list_empty(lsthead)) {
617 pwq = list_first_entry(lsthead, struct eppoll_entry, llink);
618
619 list_del(&pwq->llink);
620 ep_remove_wait_queue(pwq);
621 kmem_cache_free(pwq_cache, pwq);
622 }
623 }
624
625 /* call only when ep->mtx is held */
ep_wakeup_source(struct epitem * epi)626 static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
627 {
628 return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
629 }
630
631 /* call only when ep->mtx is held */
ep_pm_stay_awake(struct epitem * epi)632 static inline void ep_pm_stay_awake(struct epitem *epi)
633 {
634 struct wakeup_source *ws = ep_wakeup_source(epi);
635
636 if (ws)
637 __pm_stay_awake(ws);
638 }
639
ep_has_wakeup_source(struct epitem * epi)640 static inline bool ep_has_wakeup_source(struct epitem *epi)
641 {
642 return rcu_access_pointer(epi->ws) ? true : false;
643 }
644
645 /* call when ep->mtx cannot be held (ep_poll_callback) */
ep_pm_stay_awake_rcu(struct epitem * epi)646 static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
647 {
648 struct wakeup_source *ws;
649
650 rcu_read_lock();
651 ws = rcu_dereference(epi->ws);
652 if (ws)
653 __pm_stay_awake(ws);
654 rcu_read_unlock();
655 }
656
657 /**
658 * ep_scan_ready_list - Scans the ready list in a way that makes possible for
659 * the scan code, to call f_op->poll(). Also allows for
660 * O(NumReady) performance.
661 *
662 * @ep: Pointer to the epoll private data structure.
663 * @sproc: Pointer to the scan callback.
664 * @priv: Private opaque data passed to the @sproc callback.
665 * @depth: The current depth of recursive f_op->poll calls.
666 * @ep_locked: caller already holds ep->mtx
667 *
668 * Returns: The same integer error code returned by the @sproc callback.
669 */
ep_scan_ready_list(struct eventpoll * ep,__poll_t (* sproc)(struct eventpoll *,struct list_head *,void *),void * priv,int depth,bool ep_locked)670 static __poll_t ep_scan_ready_list(struct eventpoll *ep,
671 __poll_t (*sproc)(struct eventpoll *,
672 struct list_head *, void *),
673 void *priv, int depth, bool ep_locked)
674 {
675 __poll_t res;
676 int pwake = 0;
677 struct epitem *epi, *nepi;
678 LIST_HEAD(txlist);
679
680 lockdep_assert_irqs_enabled();
681
682 /*
683 * We need to lock this because we could be hit by
684 * eventpoll_release_file() and epoll_ctl().
685 */
686
687 if (!ep_locked)
688 mutex_lock_nested(&ep->mtx, depth);
689
690 /*
691 * Steal the ready list, and re-init the original one to the
692 * empty list. Also, set ep->ovflist to NULL so that events
693 * happening while looping w/out locks, are not lost. We cannot
694 * have the poll callback to queue directly on ep->rdllist,
695 * because we want the "sproc" callback to be able to do it
696 * in a lockless way.
697 */
698 spin_lock_irq(&ep->wq.lock);
699 list_splice_init(&ep->rdllist, &txlist);
700 ep->ovflist = NULL;
701 spin_unlock_irq(&ep->wq.lock);
702
703 /*
704 * Now call the callback function.
705 */
706 res = (*sproc)(ep, &txlist, priv);
707
708 spin_lock_irq(&ep->wq.lock);
709 /*
710 * During the time we spent inside the "sproc" callback, some
711 * other events might have been queued by the poll callback.
712 * We re-insert them inside the main ready-list here.
713 */
714 for (nepi = ep->ovflist; (epi = nepi) != NULL;
715 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {
716 /*
717 * We need to check if the item is already in the list.
718 * During the "sproc" callback execution time, items are
719 * queued into ->ovflist but the "txlist" might already
720 * contain them, and the list_splice() below takes care of them.
721 */
722 if (!ep_is_linked(epi)) {
723 list_add_tail(&epi->rdllink, &ep->rdllist);
724 ep_pm_stay_awake(epi);
725 }
726 }
727 /*
728 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after
729 * releasing the lock, events will be queued in the normal way inside
730 * ep->rdllist.
731 */
732 ep->ovflist = EP_UNACTIVE_PTR;
733
734 /*
735 * Quickly re-inject items left on "txlist".
736 */
737 list_splice(&txlist, &ep->rdllist);
738 __pm_relax(ep->ws);
739
740 if (!list_empty(&ep->rdllist)) {
741 /*
742 * Wake up (if active) both the eventpoll wait list and
743 * the ->poll() wait list (delayed after we release the lock).
744 */
745 if (waitqueue_active(&ep->wq))
746 wake_up_locked(&ep->wq);
747 if (waitqueue_active(&ep->poll_wait))
748 pwake++;
749 }
750 spin_unlock_irq(&ep->wq.lock);
751
752 if (!ep_locked)
753 mutex_unlock(&ep->mtx);
754
755 /* We have to call this outside the lock */
756 if (pwake)
757 ep_poll_safewake(&ep->poll_wait);
758
759 return res;
760 }
761
epi_rcu_free(struct rcu_head * head)762 static void epi_rcu_free(struct rcu_head *head)
763 {
764 struct epitem *epi = container_of(head, struct epitem, rcu);
765 kmem_cache_free(epi_cache, epi);
766 }
767
768 /*
769 * Removes a "struct epitem" from the eventpoll RB tree and deallocates
770 * all the associated resources. Must be called with "mtx" held.
771 */
ep_remove(struct eventpoll * ep,struct epitem * epi)772 static int ep_remove(struct eventpoll *ep, struct epitem *epi)
773 {
774 struct file *file = epi->ffd.file;
775
776 lockdep_assert_irqs_enabled();
777
778 /*
779 * Removes poll wait queue hooks.
780 */
781 ep_unregister_pollwait(ep, epi);
782
783 /* Remove the current item from the list of epoll hooks */
784 spin_lock(&file->f_lock);
785 list_del_rcu(&epi->fllink);
786 spin_unlock(&file->f_lock);
787
788 rb_erase_cached(&epi->rbn, &ep->rbr);
789
790 spin_lock_irq(&ep->wq.lock);
791 if (ep_is_linked(epi))
792 list_del_init(&epi->rdllink);
793 spin_unlock_irq(&ep->wq.lock);
794
795 wakeup_source_unregister(ep_wakeup_source(epi));
796 /*
797 * At this point it is safe to free the eventpoll item. Use the union
798 * field epi->rcu, since we are trying to minimize the size of
799 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
800 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
801 * use of the rbn field.
802 */
803 call_rcu(&epi->rcu, epi_rcu_free);
804
805 atomic_long_dec(&ep->user->epoll_watches);
806
807 return 0;
808 }
809
ep_free(struct eventpoll * ep)810 static void ep_free(struct eventpoll *ep)
811 {
812 struct rb_node *rbp;
813 struct epitem *epi;
814
815 /* We need to release all tasks waiting for these file */
816 if (waitqueue_active(&ep->poll_wait))
817 ep_poll_safewake(&ep->poll_wait);
818
819 /*
820 * We need to lock this because we could be hit by
821 * eventpoll_release_file() while we're freeing the "struct eventpoll".
822 * We do not need to hold "ep->mtx" here because the epoll file
823 * is on the way to be removed and no one has references to it
824 * anymore. The only hit might come from eventpoll_release_file() but
825 * holding "epmutex" is sufficient here.
826 */
827 mutex_lock(&epmutex);
828
829 /*
830 * Walks through the whole tree by unregistering poll callbacks.
831 */
832 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
833 epi = rb_entry(rbp, struct epitem, rbn);
834
835 ep_unregister_pollwait(ep, epi);
836 cond_resched();
837 }
838
839 /*
840 * Walks through the whole tree by freeing each "struct epitem". At this
841 * point we are sure no poll callbacks will be lingering around, and also by
842 * holding "epmutex" we can be sure that no file cleanup code will hit
843 * us during this operation. So we can avoid the lock on "ep->wq.lock".
844 * We do not need to lock ep->mtx, either, we only do it to prevent
845 * a lockdep warning.
846 */
847 mutex_lock(&ep->mtx);
848 while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
849 epi = rb_entry(rbp, struct epitem, rbn);
850 ep_remove(ep, epi);
851 cond_resched();
852 }
853 mutex_unlock(&ep->mtx);
854
855 mutex_unlock(&epmutex);
856 mutex_destroy(&ep->mtx);
857 free_uid(ep->user);
858 wakeup_source_unregister(ep->ws);
859 kfree(ep);
860 }
861
ep_eventpoll_release(struct inode * inode,struct file * file)862 static int ep_eventpoll_release(struct inode *inode, struct file *file)
863 {
864 struct eventpoll *ep = file->private_data;
865
866 if (ep)
867 ep_free(ep);
868
869 return 0;
870 }
871
872 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
873 void *priv);
874 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
875 poll_table *pt);
876
877 /*
878 * Differs from ep_eventpoll_poll() in that internal callers already have
879 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
880 * is correctly annotated.
881 */
ep_item_poll(const struct epitem * epi,poll_table * pt,int depth)882 static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
883 int depth)
884 {
885 struct eventpoll *ep;
886 bool locked;
887
888 pt->_key = epi->event.events;
889 if (!is_file_epoll(epi->ffd.file))
890 return vfs_poll(epi->ffd.file, pt) & epi->event.events;
891
892 ep = epi->ffd.file->private_data;
893 poll_wait(epi->ffd.file, &ep->poll_wait, pt);
894 locked = pt && (pt->_qproc == ep_ptable_queue_proc);
895
896 return ep_scan_ready_list(epi->ffd.file->private_data,
897 ep_read_events_proc, &depth, depth,
898 locked) & epi->event.events;
899 }
900
ep_read_events_proc(struct eventpoll * ep,struct list_head * head,void * priv)901 static __poll_t ep_read_events_proc(struct eventpoll *ep, struct list_head *head,
902 void *priv)
903 {
904 struct epitem *epi, *tmp;
905 poll_table pt;
906 int depth = *(int *)priv;
907
908 init_poll_funcptr(&pt, NULL);
909 depth++;
910
911 list_for_each_entry_safe(epi, tmp, head, rdllink) {
912 if (ep_item_poll(epi, &pt, depth)) {
913 return EPOLLIN | EPOLLRDNORM;
914 } else {
915 /*
916 * Item has been dropped into the ready list by the poll
917 * callback, but it's not actually ready, as far as
918 * caller requested events goes. We can remove it here.
919 */
920 __pm_relax(ep_wakeup_source(epi));
921 list_del_init(&epi->rdllink);
922 }
923 }
924
925 return 0;
926 }
927
ep_eventpoll_poll(struct file * file,poll_table * wait)928 static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
929 {
930 struct eventpoll *ep = file->private_data;
931 int depth = 0;
932
933 /* Insert inside our poll wait queue */
934 poll_wait(file, &ep->poll_wait, wait);
935
936 /*
937 * Proceed to find out if wanted events are really available inside
938 * the ready list.
939 */
940 return ep_scan_ready_list(ep, ep_read_events_proc,
941 &depth, depth, false);
942 }
943
944 #ifdef CONFIG_PROC_FS
ep_show_fdinfo(struct seq_file * m,struct file * f)945 static void ep_show_fdinfo(struct seq_file *m, struct file *f)
946 {
947 struct eventpoll *ep = f->private_data;
948 struct rb_node *rbp;
949
950 mutex_lock(&ep->mtx);
951 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
952 struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
953 struct inode *inode = file_inode(epi->ffd.file);
954
955 seq_printf(m, "tfd: %8d events: %8x data: %16llx "
956 " pos:%lli ino:%lx sdev:%x\n",
957 epi->ffd.fd, epi->event.events,
958 (long long)epi->event.data,
959 (long long)epi->ffd.file->f_pos,
960 inode->i_ino, inode->i_sb->s_dev);
961 if (seq_has_overflowed(m))
962 break;
963 }
964 mutex_unlock(&ep->mtx);
965 }
966 #endif
967
968 /* File callbacks that implement the eventpoll file behaviour */
969 static const struct file_operations eventpoll_fops = {
970 #ifdef CONFIG_PROC_FS
971 .show_fdinfo = ep_show_fdinfo,
972 #endif
973 .release = ep_eventpoll_release,
974 .poll = ep_eventpoll_poll,
975 .llseek = noop_llseek,
976 };
977
978 /*
979 * This is called from eventpoll_release() to unlink files from the eventpoll
980 * interface. We need to have this facility to cleanup correctly files that are
981 * closed without being removed from the eventpoll interface.
982 */
eventpoll_release_file(struct file * file)983 void eventpoll_release_file(struct file *file)
984 {
985 struct eventpoll *ep;
986 struct epitem *epi, *next;
987
988 /*
989 * We don't want to get "file->f_lock" because it is not
990 * necessary. It is not necessary because we're in the "struct file"
991 * cleanup path, and this means that no one is using this file anymore.
992 * So, for example, epoll_ctl() cannot hit here since if we reach this
993 * point, the file counter already went to zero and fget() would fail.
994 * The only hit might come from ep_free() but by holding the mutex
995 * will correctly serialize the operation. We do need to acquire
996 * "ep->mtx" after "epmutex" because ep_remove() requires it when called
997 * from anywhere but ep_free().
998 *
999 * Besides, ep_remove() acquires the lock, so we can't hold it here.
1000 */
1001 mutex_lock(&epmutex);
1002 list_for_each_entry_safe(epi, next, &file->f_ep_links, fllink) {
1003 ep = epi->ep;
1004 mutex_lock_nested(&ep->mtx, 0);
1005 ep_remove(ep, epi);
1006 mutex_unlock(&ep->mtx);
1007 }
1008 mutex_unlock(&epmutex);
1009 }
1010
ep_alloc(struct eventpoll ** pep)1011 static int ep_alloc(struct eventpoll **pep)
1012 {
1013 int error;
1014 struct user_struct *user;
1015 struct eventpoll *ep;
1016
1017 user = get_current_user();
1018 error = -ENOMEM;
1019 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
1020 if (unlikely(!ep))
1021 goto free_uid;
1022
1023 mutex_init(&ep->mtx);
1024 init_waitqueue_head(&ep->wq);
1025 init_waitqueue_head(&ep->poll_wait);
1026 INIT_LIST_HEAD(&ep->rdllist);
1027 ep->rbr = RB_ROOT_CACHED;
1028 ep->ovflist = EP_UNACTIVE_PTR;
1029 ep->user = user;
1030
1031 *pep = ep;
1032
1033 return 0;
1034
1035 free_uid:
1036 free_uid(user);
1037 return error;
1038 }
1039
1040 /*
1041 * Search the file inside the eventpoll tree. The RB tree operations
1042 * are protected by the "mtx" mutex, and ep_find() must be called with
1043 * "mtx" held.
1044 */
ep_find(struct eventpoll * ep,struct file * file,int fd)1045 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)
1046 {
1047 int kcmp;
1048 struct rb_node *rbp;
1049 struct epitem *epi, *epir = NULL;
1050 struct epoll_filefd ffd;
1051
1052 ep_set_ffd(&ffd, file, fd);
1053 for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
1054 epi = rb_entry(rbp, struct epitem, rbn);
1055 kcmp = ep_cmp_ffd(&ffd, &epi->ffd);
1056 if (kcmp > 0)
1057 rbp = rbp->rb_right;
1058 else if (kcmp < 0)
1059 rbp = rbp->rb_left;
1060 else {
1061 epir = epi;
1062 break;
1063 }
1064 }
1065
1066 return epir;
1067 }
1068
1069 #ifdef CONFIG_CHECKPOINT_RESTORE
ep_find_tfd(struct eventpoll * ep,int tfd,unsigned long toff)1070 static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
1071 {
1072 struct rb_node *rbp;
1073 struct epitem *epi;
1074
1075 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1076 epi = rb_entry(rbp, struct epitem, rbn);
1077 if (epi->ffd.fd == tfd) {
1078 if (toff == 0)
1079 return epi;
1080 else
1081 toff--;
1082 }
1083 cond_resched();
1084 }
1085
1086 return NULL;
1087 }
1088
get_epoll_tfile_raw_ptr(struct file * file,int tfd,unsigned long toff)1089 struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
1090 unsigned long toff)
1091 {
1092 struct file *file_raw;
1093 struct eventpoll *ep;
1094 struct epitem *epi;
1095
1096 if (!is_file_epoll(file))
1097 return ERR_PTR(-EINVAL);
1098
1099 ep = file->private_data;
1100
1101 mutex_lock(&ep->mtx);
1102 epi = ep_find_tfd(ep, tfd, toff);
1103 if (epi)
1104 file_raw = epi->ffd.file;
1105 else
1106 file_raw = ERR_PTR(-ENOENT);
1107 mutex_unlock(&ep->mtx);
1108
1109 return file_raw;
1110 }
1111 #endif /* CONFIG_CHECKPOINT_RESTORE */
1112
1113 /*
1114 * This is the callback that is passed to the wait queue wakeup
1115 * mechanism. It is called by the stored file descriptors when they
1116 * have events to report.
1117 */
ep_poll_callback(wait_queue_entry_t * wait,unsigned mode,int sync,void * key)1118 static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1119 {
1120 int pwake = 0;
1121 unsigned long flags;
1122 struct epitem *epi = ep_item_from_wait(wait);
1123 struct eventpoll *ep = epi->ep;
1124 __poll_t pollflags = key_to_poll(key);
1125 int ewake = 0;
1126
1127 spin_lock_irqsave(&ep->wq.lock, flags);
1128
1129 ep_set_busy_poll_napi_id(epi);
1130
1131 /*
1132 * If the event mask does not contain any poll(2) event, we consider the
1133 * descriptor to be disabled. This condition is likely the effect of the
1134 * EPOLLONESHOT bit that disables the descriptor when an event is received,
1135 * until the next EPOLL_CTL_MOD will be issued.
1136 */
1137 if (!(epi->event.events & ~EP_PRIVATE_BITS))
1138 goto out_unlock;
1139
1140 /*
1141 * Check the events coming with the callback. At this stage, not
1142 * every device reports the events in the "key" parameter of the
1143 * callback. We need to be able to handle both cases here, hence the
1144 * test for "key" != NULL before the event match test.
1145 */
1146 if (pollflags && !(pollflags & epi->event.events))
1147 goto out_unlock;
1148
1149 /*
1150 * If we are transferring events to userspace, we can hold no locks
1151 * (because we're accessing user memory, and because of linux f_op->poll()
1152 * semantics). All the events that happen during that period of time are
1153 * chained in ep->ovflist and requeued later on.
1154 */
1155 if (ep->ovflist != EP_UNACTIVE_PTR) {
1156 if (epi->next == EP_UNACTIVE_PTR) {
1157 epi->next = ep->ovflist;
1158 ep->ovflist = epi;
1159 if (epi->ws) {
1160 /*
1161 * Activate ep->ws since epi->ws may get
1162 * deactivated at any time.
1163 */
1164 __pm_stay_awake(ep->ws);
1165 }
1166
1167 }
1168 goto out_unlock;
1169 }
1170
1171 /* If this file is already in the ready list we exit soon */
1172 if (!ep_is_linked(epi)) {
1173 list_add_tail(&epi->rdllink, &ep->rdllist);
1174 ep_pm_stay_awake_rcu(epi);
1175 }
1176
1177 /*
1178 * Wake up ( if active ) both the eventpoll wait list and the ->poll()
1179 * wait list.
1180 */
1181 if (waitqueue_active(&ep->wq)) {
1182 if ((epi->event.events & EPOLLEXCLUSIVE) &&
1183 !(pollflags & POLLFREE)) {
1184 switch (pollflags & EPOLLINOUT_BITS) {
1185 case EPOLLIN:
1186 if (epi->event.events & EPOLLIN)
1187 ewake = 1;
1188 break;
1189 case EPOLLOUT:
1190 if (epi->event.events & EPOLLOUT)
1191 ewake = 1;
1192 break;
1193 case 0:
1194 ewake = 1;
1195 break;
1196 }
1197 }
1198 wake_up_locked(&ep->wq);
1199 }
1200 if (waitqueue_active(&ep->poll_wait))
1201 pwake++;
1202
1203 out_unlock:
1204 spin_unlock_irqrestore(&ep->wq.lock, flags);
1205
1206 /* We have to call this outside the lock */
1207 if (pwake)
1208 ep_poll_safewake(&ep->poll_wait);
1209
1210 if (!(epi->event.events & EPOLLEXCLUSIVE))
1211 ewake = 1;
1212
1213 if (pollflags & POLLFREE) {
1214 /*
1215 * If we race with ep_remove_wait_queue() it can miss
1216 * ->whead = NULL and do another remove_wait_queue() after
1217 * us, so we can't use __remove_wait_queue().
1218 */
1219 list_del_init(&wait->entry);
1220 /*
1221 * ->whead != NULL protects us from the race with ep_free()
1222 * or ep_remove(), ep_remove_wait_queue() takes whead->lock
1223 * held by the caller. Once we nullify it, nothing protects
1224 * ep/epi or even wait.
1225 */
1226 smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
1227 }
1228
1229 return ewake;
1230 }
1231
1232 /*
1233 * This is the callback that is used to add our wait queue to the
1234 * target file wakeup lists.
1235 */
ep_ptable_queue_proc(struct file * file,wait_queue_head_t * whead,poll_table * pt)1236 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
1237 poll_table *pt)
1238 {
1239 struct epitem *epi = ep_item_from_epqueue(pt);
1240 struct eppoll_entry *pwq;
1241
1242 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) {
1243 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
1244 pwq->whead = whead;
1245 pwq->base = epi;
1246 if (epi->event.events & EPOLLEXCLUSIVE)
1247 add_wait_queue_exclusive(whead, &pwq->wait);
1248 else
1249 add_wait_queue(whead, &pwq->wait);
1250 list_add_tail(&pwq->llink, &epi->pwqlist);
1251 epi->nwait++;
1252 } else {
1253 /* We have to signal that an error occurred */
1254 epi->nwait = -1;
1255 }
1256 }
1257
ep_rbtree_insert(struct eventpoll * ep,struct epitem * epi)1258 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
1259 {
1260 int kcmp;
1261 struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
1262 struct epitem *epic;
1263 bool leftmost = true;
1264
1265 while (*p) {
1266 parent = *p;
1267 epic = rb_entry(parent, struct epitem, rbn);
1268 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
1269 if (kcmp > 0) {
1270 p = &parent->rb_right;
1271 leftmost = false;
1272 } else
1273 p = &parent->rb_left;
1274 }
1275 rb_link_node(&epi->rbn, parent, p);
1276 rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
1277 }
1278
1279
1280
1281 #define PATH_ARR_SIZE 5
1282 /*
1283 * These are the number paths of length 1 to 5, that we are allowing to emanate
1284 * from a single file of interest. For example, we allow 1000 paths of length
1285 * 1, to emanate from each file of interest. This essentially represents the
1286 * potential wakeup paths, which need to be limited in order to avoid massive
1287 * uncontrolled wakeup storms. The common use case should be a single ep which
1288 * is connected to n file sources. In this case each file source has 1 path
1289 * of length 1. Thus, the numbers below should be more than sufficient. These
1290 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify
1291 * and delete can't add additional paths. Protected by the epmutex.
1292 */
1293 static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
1294 static int path_count[PATH_ARR_SIZE];
1295
path_count_inc(int nests)1296 static int path_count_inc(int nests)
1297 {
1298 /* Allow an arbitrary number of depth 1 paths */
1299 if (nests == 0)
1300 return 0;
1301
1302 if (++path_count[nests] > path_limits[nests])
1303 return -1;
1304 return 0;
1305 }
1306
path_count_init(void)1307 static void path_count_init(void)
1308 {
1309 int i;
1310
1311 for (i = 0; i < PATH_ARR_SIZE; i++)
1312 path_count[i] = 0;
1313 }
1314
reverse_path_check_proc(void * priv,void * cookie,int call_nests)1315 static int reverse_path_check_proc(void *priv, void *cookie, int call_nests)
1316 {
1317 int error = 0;
1318 struct file *file = priv;
1319 struct file *child_file;
1320 struct epitem *epi;
1321
1322 /* CTL_DEL can remove links here, but that can't increase our count */
1323 rcu_read_lock();
1324 list_for_each_entry_rcu(epi, &file->f_ep_links, fllink) {
1325 child_file = epi->ep->file;
1326 if (is_file_epoll(child_file)) {
1327 if (list_empty(&child_file->f_ep_links)) {
1328 if (path_count_inc(call_nests)) {
1329 error = -1;
1330 break;
1331 }
1332 } else {
1333 error = ep_call_nested(&poll_loop_ncalls,
1334 EP_MAX_NESTS,
1335 reverse_path_check_proc,
1336 child_file, child_file,
1337 current);
1338 }
1339 if (error != 0)
1340 break;
1341 } else {
1342 printk(KERN_ERR "reverse_path_check_proc: "
1343 "file is not an ep!\n");
1344 }
1345 }
1346 rcu_read_unlock();
1347 return error;
1348 }
1349
1350 /**
1351 * reverse_path_check - The tfile_check_list is list of file *, which have
1352 * links that are proposed to be newly added. We need to
1353 * make sure that those added links don't add too many
1354 * paths such that we will spend all our time waking up
1355 * eventpoll objects.
1356 *
1357 * Returns: Returns zero if the proposed links don't create too many paths,
1358 * -1 otherwise.
1359 */
reverse_path_check(void)1360 static int reverse_path_check(void)
1361 {
1362 int error = 0;
1363 struct file *current_file;
1364
1365 /* let's call this for all tfiles */
1366 list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) {
1367 path_count_init();
1368 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1369 reverse_path_check_proc, current_file,
1370 current_file, current);
1371 if (error)
1372 break;
1373 }
1374 return error;
1375 }
1376
ep_create_wakeup_source(struct epitem * epi)1377 static int ep_create_wakeup_source(struct epitem *epi)
1378 {
1379 struct name_snapshot n;
1380 struct wakeup_source *ws;
1381
1382 if (!epi->ep->ws) {
1383 epi->ep->ws = wakeup_source_register("eventpoll");
1384 if (!epi->ep->ws)
1385 return -ENOMEM;
1386 }
1387
1388 take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
1389 ws = wakeup_source_register(n.name);
1390 release_dentry_name_snapshot(&n);
1391
1392 if (!ws)
1393 return -ENOMEM;
1394 rcu_assign_pointer(epi->ws, ws);
1395
1396 return 0;
1397 }
1398
1399 /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */
ep_destroy_wakeup_source(struct epitem * epi)1400 static noinline void ep_destroy_wakeup_source(struct epitem *epi)
1401 {
1402 struct wakeup_source *ws = ep_wakeup_source(epi);
1403
1404 RCU_INIT_POINTER(epi->ws, NULL);
1405
1406 /*
1407 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
1408 * used internally by wakeup_source_remove, too (called by
1409 * wakeup_source_unregister), so we cannot use call_rcu
1410 */
1411 synchronize_rcu();
1412 wakeup_source_unregister(ws);
1413 }
1414
1415 /*
1416 * Must be called with "mtx" held.
1417 */
ep_insert(struct eventpoll * ep,const struct epoll_event * event,struct file * tfile,int fd,int full_check)1418 static int ep_insert(struct eventpoll *ep, const struct epoll_event *event,
1419 struct file *tfile, int fd, int full_check)
1420 {
1421 int error, pwake = 0;
1422 __poll_t revents;
1423 long user_watches;
1424 struct epitem *epi;
1425 struct ep_pqueue epq;
1426
1427 lockdep_assert_irqs_enabled();
1428
1429 user_watches = atomic_long_read(&ep->user->epoll_watches);
1430 if (unlikely(user_watches >= max_user_watches))
1431 return -ENOSPC;
1432 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL)))
1433 return -ENOMEM;
1434
1435 /* Item initialization follow here ... */
1436 INIT_LIST_HEAD(&epi->rdllink);
1437 INIT_LIST_HEAD(&epi->fllink);
1438 INIT_LIST_HEAD(&epi->pwqlist);
1439 epi->ep = ep;
1440 ep_set_ffd(&epi->ffd, tfile, fd);
1441 epi->event = *event;
1442 epi->nwait = 0;
1443 epi->next = EP_UNACTIVE_PTR;
1444 if (epi->event.events & EPOLLWAKEUP) {
1445 error = ep_create_wakeup_source(epi);
1446 if (error)
1447 goto error_create_wakeup_source;
1448 } else {
1449 RCU_INIT_POINTER(epi->ws, NULL);
1450 }
1451
1452 /* Add the current item to the list of active epoll hook for this file */
1453 spin_lock(&tfile->f_lock);
1454 list_add_tail_rcu(&epi->fllink, &tfile->f_ep_links);
1455 spin_unlock(&tfile->f_lock);
1456
1457 /*
1458 * Add the current item to the RB tree. All RB tree operations are
1459 * protected by "mtx", and ep_insert() is called with "mtx" held.
1460 */
1461 ep_rbtree_insert(ep, epi);
1462
1463 /* now check if we've created too many backpaths */
1464 error = -EINVAL;
1465 if (full_check && reverse_path_check())
1466 goto error_remove_epi;
1467
1468 /* Initialize the poll table using the queue callback */
1469 epq.epi = epi;
1470 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
1471
1472 /*
1473 * Attach the item to the poll hooks and get current event bits.
1474 * We can safely use the file* here because its usage count has
1475 * been increased by the caller of this function. Note that after
1476 * this operation completes, the poll callback can start hitting
1477 * the new item.
1478 */
1479 revents = ep_item_poll(epi, &epq.pt, 1);
1480
1481 /*
1482 * We have to check if something went wrong during the poll wait queue
1483 * install process. Namely an allocation for a wait queue failed due
1484 * high memory pressure.
1485 */
1486 error = -ENOMEM;
1487 if (epi->nwait < 0)
1488 goto error_unregister;
1489
1490 /* We have to drop the new item inside our item list to keep track of it */
1491 spin_lock_irq(&ep->wq.lock);
1492
1493 /* record NAPI ID of new item if present */
1494 ep_set_busy_poll_napi_id(epi);
1495
1496 /* If the file is already "ready" we drop it inside the ready list */
1497 if (revents && !ep_is_linked(epi)) {
1498 list_add_tail(&epi->rdllink, &ep->rdllist);
1499 ep_pm_stay_awake(epi);
1500
1501 /* Notify waiting tasks that events are available */
1502 if (waitqueue_active(&ep->wq))
1503 wake_up_locked(&ep->wq);
1504 if (waitqueue_active(&ep->poll_wait))
1505 pwake++;
1506 }
1507
1508 spin_unlock_irq(&ep->wq.lock);
1509
1510 atomic_long_inc(&ep->user->epoll_watches);
1511
1512 /* We have to call this outside the lock */
1513 if (pwake)
1514 ep_poll_safewake(&ep->poll_wait);
1515
1516 return 0;
1517
1518 error_unregister:
1519 ep_unregister_pollwait(ep, epi);
1520 error_remove_epi:
1521 spin_lock(&tfile->f_lock);
1522 list_del_rcu(&epi->fllink);
1523 spin_unlock(&tfile->f_lock);
1524
1525 rb_erase_cached(&epi->rbn, &ep->rbr);
1526
1527 /*
1528 * We need to do this because an event could have been arrived on some
1529 * allocated wait queue. Note that we don't care about the ep->ovflist
1530 * list, since that is used/cleaned only inside a section bound by "mtx".
1531 * And ep_insert() is called with "mtx" held.
1532 */
1533 spin_lock_irq(&ep->wq.lock);
1534 if (ep_is_linked(epi))
1535 list_del_init(&epi->rdllink);
1536 spin_unlock_irq(&ep->wq.lock);
1537
1538 wakeup_source_unregister(ep_wakeup_source(epi));
1539
1540 error_create_wakeup_source:
1541 kmem_cache_free(epi_cache, epi);
1542
1543 return error;
1544 }
1545
1546 /*
1547 * Modify the interest event mask by dropping an event if the new mask
1548 * has a match in the current file status. Must be called with "mtx" held.
1549 */
ep_modify(struct eventpoll * ep,struct epitem * epi,const struct epoll_event * event)1550 static int ep_modify(struct eventpoll *ep, struct epitem *epi,
1551 const struct epoll_event *event)
1552 {
1553 int pwake = 0;
1554 poll_table pt;
1555
1556 lockdep_assert_irqs_enabled();
1557
1558 init_poll_funcptr(&pt, NULL);
1559
1560 /*
1561 * Set the new event interest mask before calling f_op->poll();
1562 * otherwise we might miss an event that happens between the
1563 * f_op->poll() call and the new event set registering.
1564 */
1565 epi->event.events = event->events; /* need barrier below */
1566 epi->event.data = event->data; /* protected by mtx */
1567 if (epi->event.events & EPOLLWAKEUP) {
1568 if (!ep_has_wakeup_source(epi))
1569 ep_create_wakeup_source(epi);
1570 } else if (ep_has_wakeup_source(epi)) {
1571 ep_destroy_wakeup_source(epi);
1572 }
1573
1574 /*
1575 * The following barrier has two effects:
1576 *
1577 * 1) Flush epi changes above to other CPUs. This ensures
1578 * we do not miss events from ep_poll_callback if an
1579 * event occurs immediately after we call f_op->poll().
1580 * We need this because we did not take ep->wq.lock while
1581 * changing epi above (but ep_poll_callback does take
1582 * ep->wq.lock).
1583 *
1584 * 2) We also need to ensure we do not miss _past_ events
1585 * when calling f_op->poll(). This barrier also
1586 * pairs with the barrier in wq_has_sleeper (see
1587 * comments for wq_has_sleeper).
1588 *
1589 * This barrier will now guarantee ep_poll_callback or f_op->poll
1590 * (or both) will notice the readiness of an item.
1591 */
1592 smp_mb();
1593
1594 /*
1595 * Get current event bits. We can safely use the file* here because
1596 * its usage count has been increased by the caller of this function.
1597 * If the item is "hot" and it is not registered inside the ready
1598 * list, push it inside.
1599 */
1600 if (ep_item_poll(epi, &pt, 1)) {
1601 spin_lock_irq(&ep->wq.lock);
1602 if (!ep_is_linked(epi)) {
1603 list_add_tail(&epi->rdllink, &ep->rdllist);
1604 ep_pm_stay_awake(epi);
1605
1606 /* Notify waiting tasks that events are available */
1607 if (waitqueue_active(&ep->wq))
1608 wake_up_locked(&ep->wq);
1609 if (waitqueue_active(&ep->poll_wait))
1610 pwake++;
1611 }
1612 spin_unlock_irq(&ep->wq.lock);
1613 }
1614
1615 /* We have to call this outside the lock */
1616 if (pwake)
1617 ep_poll_safewake(&ep->poll_wait);
1618
1619 return 0;
1620 }
1621
ep_send_events_proc(struct eventpoll * ep,struct list_head * head,void * priv)1622 static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head,
1623 void *priv)
1624 {
1625 struct ep_send_events_data *esed = priv;
1626 __poll_t revents;
1627 struct epitem *epi;
1628 struct epoll_event __user *uevent;
1629 struct wakeup_source *ws;
1630 poll_table pt;
1631
1632 init_poll_funcptr(&pt, NULL);
1633
1634 /*
1635 * We can loop without lock because we are passed a task private list.
1636 * Items cannot vanish during the loop because ep_scan_ready_list() is
1637 * holding "mtx" during this call.
1638 */
1639 for (esed->res = 0, uevent = esed->events;
1640 !list_empty(head) && esed->res < esed->maxevents;) {
1641 epi = list_first_entry(head, struct epitem, rdllink);
1642
1643 /*
1644 * Activate ep->ws before deactivating epi->ws to prevent
1645 * triggering auto-suspend here (in case we reactive epi->ws
1646 * below).
1647 *
1648 * This could be rearranged to delay the deactivation of epi->ws
1649 * instead, but then epi->ws would temporarily be out of sync
1650 * with ep_is_linked().
1651 */
1652 ws = ep_wakeup_source(epi);
1653 if (ws) {
1654 if (ws->active)
1655 __pm_stay_awake(ep->ws);
1656 __pm_relax(ws);
1657 }
1658
1659 list_del_init(&epi->rdllink);
1660
1661 revents = ep_item_poll(epi, &pt, 1);
1662
1663 /*
1664 * If the event mask intersect the caller-requested one,
1665 * deliver the event to userspace. Again, ep_scan_ready_list()
1666 * is holding "mtx", so no operations coming from userspace
1667 * can change the item.
1668 */
1669 if (revents) {
1670 if (__put_user(revents, &uevent->events) ||
1671 __put_user(epi->event.data, &uevent->data)) {
1672 list_add(&epi->rdllink, head);
1673 ep_pm_stay_awake(epi);
1674 if (!esed->res)
1675 esed->res = -EFAULT;
1676 return 0;
1677 }
1678 esed->res++;
1679 uevent++;
1680 if (epi->event.events & EPOLLONESHOT)
1681 epi->event.events &= EP_PRIVATE_BITS;
1682 else if (!(epi->event.events & EPOLLET)) {
1683 /*
1684 * If this file has been added with Level
1685 * Trigger mode, we need to insert back inside
1686 * the ready list, so that the next call to
1687 * epoll_wait() will check again the events
1688 * availability. At this point, no one can insert
1689 * into ep->rdllist besides us. The epoll_ctl()
1690 * callers are locked out by
1691 * ep_scan_ready_list() holding "mtx" and the
1692 * poll callback will queue them in ep->ovflist.
1693 */
1694 list_add_tail(&epi->rdllink, &ep->rdllist);
1695 ep_pm_stay_awake(epi);
1696 }
1697 }
1698 }
1699
1700 return 0;
1701 }
1702
ep_send_events(struct eventpoll * ep,struct epoll_event __user * events,int maxevents)1703 static int ep_send_events(struct eventpoll *ep,
1704 struct epoll_event __user *events, int maxevents)
1705 {
1706 struct ep_send_events_data esed;
1707
1708 esed.maxevents = maxevents;
1709 esed.events = events;
1710
1711 ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0, false);
1712 return esed.res;
1713 }
1714
ep_set_mstimeout(long ms)1715 static inline struct timespec64 ep_set_mstimeout(long ms)
1716 {
1717 struct timespec64 now, ts = {
1718 .tv_sec = ms / MSEC_PER_SEC,
1719 .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC),
1720 };
1721
1722 ktime_get_ts64(&now);
1723 return timespec64_add_safe(now, ts);
1724 }
1725
1726 /**
1727 * ep_poll - Retrieves ready events, and delivers them to the caller supplied
1728 * event buffer.
1729 *
1730 * @ep: Pointer to the eventpoll context.
1731 * @events: Pointer to the userspace buffer where the ready events should be
1732 * stored.
1733 * @maxevents: Size (in terms of number of events) of the caller event buffer.
1734 * @timeout: Maximum timeout for the ready events fetch operation, in
1735 * milliseconds. If the @timeout is zero, the function will not block,
1736 * while if the @timeout is less than zero, the function will block
1737 * until at least one event has been retrieved (or an error
1738 * occurred).
1739 *
1740 * Returns: Returns the number of ready events which have been fetched, or an
1741 * error code, in case of error.
1742 */
ep_poll(struct eventpoll * ep,struct epoll_event __user * events,int maxevents,long timeout)1743 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
1744 int maxevents, long timeout)
1745 {
1746 int res = 0, eavail, timed_out = 0;
1747 u64 slack = 0;
1748 wait_queue_entry_t wait;
1749 ktime_t expires, *to = NULL;
1750
1751 lockdep_assert_irqs_enabled();
1752
1753 if (timeout > 0) {
1754 struct timespec64 end_time = ep_set_mstimeout(timeout);
1755
1756 slack = select_estimate_accuracy(&end_time);
1757 to = &expires;
1758 *to = timespec64_to_ktime(end_time);
1759 } else if (timeout == 0) {
1760 /*
1761 * Avoid the unnecessary trip to the wait queue loop, if the
1762 * caller specified a non blocking operation.
1763 */
1764 timed_out = 1;
1765 spin_lock_irq(&ep->wq.lock);
1766 goto check_events;
1767 }
1768
1769 fetch_events:
1770
1771 if (!ep_events_available(ep))
1772 ep_busy_loop(ep, timed_out);
1773
1774 spin_lock_irq(&ep->wq.lock);
1775
1776 if (!ep_events_available(ep)) {
1777 /*
1778 * Busy poll timed out. Drop NAPI ID for now, we can add
1779 * it back in when we have moved a socket with a valid NAPI
1780 * ID onto the ready list.
1781 */
1782 ep_reset_busy_poll_napi_id(ep);
1783
1784 /*
1785 * We don't have any available event to return to the caller.
1786 * We need to sleep here, and we will be wake up by
1787 * ep_poll_callback() when events will become available.
1788 */
1789 init_waitqueue_entry(&wait, current);
1790 __add_wait_queue_exclusive(&ep->wq, &wait);
1791
1792 for (;;) {
1793 /*
1794 * We don't want to sleep if the ep_poll_callback() sends us
1795 * a wakeup in between. That's why we set the task state
1796 * to TASK_INTERRUPTIBLE before doing the checks.
1797 */
1798 set_current_state(TASK_INTERRUPTIBLE);
1799 /*
1800 * Always short-circuit for fatal signals to allow
1801 * threads to make a timely exit without the chance of
1802 * finding more events available and fetching
1803 * repeatedly.
1804 */
1805 if (fatal_signal_pending(current)) {
1806 res = -EINTR;
1807 break;
1808 }
1809 if (ep_events_available(ep) || timed_out)
1810 break;
1811 if (signal_pending(current)) {
1812 res = -EINTR;
1813 break;
1814 }
1815
1816 spin_unlock_irq(&ep->wq.lock);
1817 if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
1818 timed_out = 1;
1819
1820 spin_lock_irq(&ep->wq.lock);
1821 }
1822
1823 __remove_wait_queue(&ep->wq, &wait);
1824 __set_current_state(TASK_RUNNING);
1825 }
1826 check_events:
1827 /* Is it worth to try to dig for events ? */
1828 eavail = ep_events_available(ep);
1829
1830 spin_unlock_irq(&ep->wq.lock);
1831
1832 /*
1833 * Try to transfer events to user space. In case we get 0 events and
1834 * there's still timeout left over, we go trying again in search of
1835 * more luck.
1836 */
1837 if (!res && eavail &&
1838 !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
1839 goto fetch_events;
1840
1841 return res;
1842 }
1843
1844 /**
1845 * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested()
1846 * API, to verify that adding an epoll file inside another
1847 * epoll structure, does not violate the constraints, in
1848 * terms of closed loops, or too deep chains (which can
1849 * result in excessive stack usage).
1850 *
1851 * @priv: Pointer to the epoll file to be currently checked.
1852 * @cookie: Original cookie for this call. This is the top-of-the-chain epoll
1853 * data structure pointer.
1854 * @call_nests: Current dept of the @ep_call_nested() call stack.
1855 *
1856 * Returns: Returns zero if adding the epoll @file inside current epoll
1857 * structure @ep does not violate the constraints, or -1 otherwise.
1858 */
ep_loop_check_proc(void * priv,void * cookie,int call_nests)1859 static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
1860 {
1861 int error = 0;
1862 struct file *file = priv;
1863 struct eventpoll *ep = file->private_data;
1864 struct eventpoll *ep_tovisit;
1865 struct rb_node *rbp;
1866 struct epitem *epi;
1867
1868 mutex_lock_nested(&ep->mtx, call_nests + 1);
1869 ep->gen = loop_check_gen;
1870 for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
1871 epi = rb_entry(rbp, struct epitem, rbn);
1872 if (unlikely(is_file_epoll(epi->ffd.file))) {
1873 ep_tovisit = epi->ffd.file->private_data;
1874 if (ep_tovisit->gen == loop_check_gen)
1875 continue;
1876 error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1877 ep_loop_check_proc, epi->ffd.file,
1878 ep_tovisit, current);
1879 if (error != 0)
1880 break;
1881 } else {
1882 /*
1883 * If we've reached a file that is not associated with
1884 * an ep, then we need to check if the newly added
1885 * links are going to add too many wakeup paths. We do
1886 * this by adding it to the tfile_check_list, if it's
1887 * not already there, and calling reverse_path_check()
1888 * during ep_insert().
1889 */
1890 if (list_empty(&epi->ffd.file->f_tfile_llink)) {
1891 if (get_file_rcu(epi->ffd.file))
1892 list_add(&epi->ffd.file->f_tfile_llink,
1893 &tfile_check_list);
1894 }
1895 }
1896 }
1897 mutex_unlock(&ep->mtx);
1898
1899 return error;
1900 }
1901
1902 /**
1903 * ep_loop_check - Performs a check to verify that adding an epoll file (@file)
1904 * another epoll file (represented by @ep) does not create
1905 * closed loops or too deep chains.
1906 *
1907 * @ep: Pointer to the epoll private data structure.
1908 * @file: Pointer to the epoll file to be checked.
1909 *
1910 * Returns: Returns zero if adding the epoll @file inside current epoll
1911 * structure @ep does not violate the constraints, or -1 otherwise.
1912 */
ep_loop_check(struct eventpoll * ep,struct file * file)1913 static int ep_loop_check(struct eventpoll *ep, struct file *file)
1914 {
1915 return ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS,
1916 ep_loop_check_proc, file, ep, current);
1917 }
1918
clear_tfile_check_list(void)1919 static void clear_tfile_check_list(void)
1920 {
1921 struct file *file;
1922
1923 /* first clear the tfile_check_list */
1924 while (!list_empty(&tfile_check_list)) {
1925 file = list_first_entry(&tfile_check_list, struct file,
1926 f_tfile_llink);
1927 list_del_init(&file->f_tfile_llink);
1928 fput(file);
1929 }
1930 INIT_LIST_HEAD(&tfile_check_list);
1931 }
1932
1933 /*
1934 * Open an eventpoll file descriptor.
1935 */
do_epoll_create(int flags)1936 static int do_epoll_create(int flags)
1937 {
1938 int error, fd;
1939 struct eventpoll *ep = NULL;
1940 struct file *file;
1941
1942 /* Check the EPOLL_* constant for consistency. */
1943 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
1944
1945 if (flags & ~EPOLL_CLOEXEC)
1946 return -EINVAL;
1947 /*
1948 * Create the internal data structure ("struct eventpoll").
1949 */
1950 error = ep_alloc(&ep);
1951 if (error < 0)
1952 return error;
1953 /*
1954 * Creates all the items needed to setup an eventpoll file. That is,
1955 * a file structure and a free file descriptor.
1956 */
1957 fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));
1958 if (fd < 0) {
1959 error = fd;
1960 goto out_free_ep;
1961 }
1962 file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
1963 O_RDWR | (flags & O_CLOEXEC));
1964 if (IS_ERR(file)) {
1965 error = PTR_ERR(file);
1966 goto out_free_fd;
1967 }
1968 ep->file = file;
1969 fd_install(fd, file);
1970 return fd;
1971
1972 out_free_fd:
1973 put_unused_fd(fd);
1974 out_free_ep:
1975 ep_free(ep);
1976 return error;
1977 }
1978
SYSCALL_DEFINE1(epoll_create1,int,flags)1979 SYSCALL_DEFINE1(epoll_create1, int, flags)
1980 {
1981 return do_epoll_create(flags);
1982 }
1983
SYSCALL_DEFINE1(epoll_create,int,size)1984 SYSCALL_DEFINE1(epoll_create, int, size)
1985 {
1986 if (size <= 0)
1987 return -EINVAL;
1988
1989 return do_epoll_create(0);
1990 }
1991
1992 /*
1993 * The following function implements the controller interface for
1994 * the eventpoll file that enables the insertion/removal/change of
1995 * file descriptors inside the interest set.
1996 */
SYSCALL_DEFINE4(epoll_ctl,int,epfd,int,op,int,fd,struct epoll_event __user *,event)1997 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
1998 struct epoll_event __user *, event)
1999 {
2000 int error;
2001 int full_check = 0;
2002 struct fd f, tf;
2003 struct eventpoll *ep;
2004 struct epitem *epi;
2005 struct epoll_event epds;
2006 struct eventpoll *tep = NULL;
2007
2008 error = -EFAULT;
2009 if (ep_op_has_event(op) &&
2010 copy_from_user(&epds, event, sizeof(struct epoll_event)))
2011 goto error_return;
2012
2013 error = -EBADF;
2014 f = fdget(epfd);
2015 if (!f.file)
2016 goto error_return;
2017
2018 /* Get the "struct file *" for the target file */
2019 tf = fdget(fd);
2020 if (!tf.file)
2021 goto error_fput;
2022
2023 /* The target file descriptor must support poll */
2024 error = -EPERM;
2025 if (!file_can_poll(tf.file))
2026 goto error_tgt_fput;
2027
2028 /* Check if EPOLLWAKEUP is allowed */
2029 if (ep_op_has_event(op))
2030 ep_take_care_of_epollwakeup(&epds);
2031
2032 /*
2033 * We have to check that the file structure underneath the file descriptor
2034 * the user passed to us _is_ an eventpoll file. And also we do not permit
2035 * adding an epoll file descriptor inside itself.
2036 */
2037 error = -EINVAL;
2038 if (f.file == tf.file || !is_file_epoll(f.file))
2039 goto error_tgt_fput;
2040
2041 /*
2042 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
2043 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
2044 * Also, we do not currently supported nested exclusive wakeups.
2045 */
2046 if (ep_op_has_event(op) && (epds.events & EPOLLEXCLUSIVE)) {
2047 if (op == EPOLL_CTL_MOD)
2048 goto error_tgt_fput;
2049 if (op == EPOLL_CTL_ADD && (is_file_epoll(tf.file) ||
2050 (epds.events & ~EPOLLEXCLUSIVE_OK_BITS)))
2051 goto error_tgt_fput;
2052 }
2053
2054 /*
2055 * At this point it is safe to assume that the "private_data" contains
2056 * our own data structure.
2057 */
2058 ep = f.file->private_data;
2059
2060 /*
2061 * When we insert an epoll file descriptor, inside another epoll file
2062 * descriptor, there is the change of creating closed loops, which are
2063 * better be handled here, than in more critical paths. While we are
2064 * checking for loops we also determine the list of files reachable
2065 * and hang them on the tfile_check_list, so we can check that we
2066 * haven't created too many possible wakeup paths.
2067 *
2068 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when
2069 * the epoll file descriptor is attaching directly to a wakeup source,
2070 * unless the epoll file descriptor is nested. The purpose of taking the
2071 * 'epmutex' on add is to prevent complex toplogies such as loops and
2072 * deep wakeup paths from forming in parallel through multiple
2073 * EPOLL_CTL_ADD operations.
2074 */
2075 mutex_lock_nested(&ep->mtx, 0);
2076 if (op == EPOLL_CTL_ADD) {
2077 if (!list_empty(&f.file->f_ep_links) ||
2078 ep->gen == loop_check_gen ||
2079 is_file_epoll(tf.file)) {
2080 full_check = 1;
2081 mutex_unlock(&ep->mtx);
2082 mutex_lock(&epmutex);
2083 if (is_file_epoll(tf.file)) {
2084 error = -ELOOP;
2085 if (ep_loop_check(ep, tf.file) != 0)
2086 goto error_tgt_fput;
2087 } else {
2088 get_file(tf.file);
2089 list_add(&tf.file->f_tfile_llink,
2090 &tfile_check_list);
2091 }
2092 mutex_lock_nested(&ep->mtx, 0);
2093 if (is_file_epoll(tf.file)) {
2094 tep = tf.file->private_data;
2095 mutex_lock_nested(&tep->mtx, 1);
2096 }
2097 }
2098 }
2099
2100 /*
2101 * Try to lookup the file inside our RB tree, Since we grabbed "mtx"
2102 * above, we can be sure to be able to use the item looked up by
2103 * ep_find() till we release the mutex.
2104 */
2105 epi = ep_find(ep, tf.file, fd);
2106
2107 error = -EINVAL;
2108 switch (op) {
2109 case EPOLL_CTL_ADD:
2110 if (!epi) {
2111 epds.events |= EPOLLERR | EPOLLHUP;
2112 error = ep_insert(ep, &epds, tf.file, fd, full_check);
2113 } else
2114 error = -EEXIST;
2115 break;
2116 case EPOLL_CTL_DEL:
2117 if (epi)
2118 error = ep_remove(ep, epi);
2119 else
2120 error = -ENOENT;
2121 break;
2122 case EPOLL_CTL_MOD:
2123 if (epi) {
2124 if (!(epi->event.events & EPOLLEXCLUSIVE)) {
2125 epds.events |= EPOLLERR | EPOLLHUP;
2126 error = ep_modify(ep, epi, &epds);
2127 }
2128 } else
2129 error = -ENOENT;
2130 break;
2131 }
2132 if (tep != NULL)
2133 mutex_unlock(&tep->mtx);
2134 mutex_unlock(&ep->mtx);
2135
2136 error_tgt_fput:
2137 if (full_check) {
2138 clear_tfile_check_list();
2139 loop_check_gen++;
2140 mutex_unlock(&epmutex);
2141 }
2142
2143 fdput(tf);
2144 error_fput:
2145 fdput(f);
2146 error_return:
2147
2148 return error;
2149 }
2150
2151 /*
2152 * Implement the event wait interface for the eventpoll file. It is the kernel
2153 * part of the user space epoll_wait(2).
2154 */
do_epoll_wait(int epfd,struct epoll_event __user * events,int maxevents,int timeout)2155 static int do_epoll_wait(int epfd, struct epoll_event __user *events,
2156 int maxevents, int timeout)
2157 {
2158 int error;
2159 struct fd f;
2160 struct eventpoll *ep;
2161
2162 /* The maximum number of event must be greater than zero */
2163 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
2164 return -EINVAL;
2165
2166 /* Verify that the area passed by the user is writeable */
2167 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event)))
2168 return -EFAULT;
2169
2170 /* Get the "struct file *" for the eventpoll file */
2171 f = fdget(epfd);
2172 if (!f.file)
2173 return -EBADF;
2174
2175 /*
2176 * We have to check that the file structure underneath the fd
2177 * the user passed to us _is_ an eventpoll file.
2178 */
2179 error = -EINVAL;
2180 if (!is_file_epoll(f.file))
2181 goto error_fput;
2182
2183 /*
2184 * At this point it is safe to assume that the "private_data" contains
2185 * our own data structure.
2186 */
2187 ep = f.file->private_data;
2188
2189 /* Time to fish for events ... */
2190 error = ep_poll(ep, events, maxevents, timeout);
2191
2192 error_fput:
2193 fdput(f);
2194 return error;
2195 }
2196
SYSCALL_DEFINE4(epoll_wait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout)2197 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
2198 int, maxevents, int, timeout)
2199 {
2200 return do_epoll_wait(epfd, events, maxevents, timeout);
2201 }
2202
2203 /*
2204 * Implement the event wait interface for the eventpoll file. It is the kernel
2205 * part of the user space epoll_pwait(2).
2206 */
SYSCALL_DEFINE6(epoll_pwait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout,const sigset_t __user *,sigmask,size_t,sigsetsize)2207 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
2208 int, maxevents, int, timeout, const sigset_t __user *, sigmask,
2209 size_t, sigsetsize)
2210 {
2211 int error;
2212 sigset_t ksigmask, sigsaved;
2213
2214 /*
2215 * If the caller wants a certain signal mask to be set during the wait,
2216 * we apply it here.
2217 */
2218 if (sigmask) {
2219 if (sigsetsize != sizeof(sigset_t))
2220 return -EINVAL;
2221 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
2222 return -EFAULT;
2223 sigsaved = current->blocked;
2224 set_current_blocked(&ksigmask);
2225 }
2226
2227 error = do_epoll_wait(epfd, events, maxevents, timeout);
2228
2229 /*
2230 * If we changed the signal mask, we need to restore the original one.
2231 * In case we've got a signal while waiting, we do not restore the
2232 * signal mask yet, and we allow do_signal() to deliver the signal on
2233 * the way back to userspace, before the signal mask is restored.
2234 */
2235 if (sigmask) {
2236 if (error == -EINTR) {
2237 memcpy(¤t->saved_sigmask, &sigsaved,
2238 sizeof(sigsaved));
2239 set_restore_sigmask();
2240 } else
2241 set_current_blocked(&sigsaved);
2242 }
2243
2244 return error;
2245 }
2246
2247 #ifdef CONFIG_COMPAT
COMPAT_SYSCALL_DEFINE6(epoll_pwait,int,epfd,struct epoll_event __user *,events,int,maxevents,int,timeout,const compat_sigset_t __user *,sigmask,compat_size_t,sigsetsize)2248 COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
2249 struct epoll_event __user *, events,
2250 int, maxevents, int, timeout,
2251 const compat_sigset_t __user *, sigmask,
2252 compat_size_t, sigsetsize)
2253 {
2254 long err;
2255 sigset_t ksigmask, sigsaved;
2256
2257 /*
2258 * If the caller wants a certain signal mask to be set during the wait,
2259 * we apply it here.
2260 */
2261 if (sigmask) {
2262 if (sigsetsize != sizeof(compat_sigset_t))
2263 return -EINVAL;
2264 if (get_compat_sigset(&ksigmask, sigmask))
2265 return -EFAULT;
2266 sigsaved = current->blocked;
2267 set_current_blocked(&ksigmask);
2268 }
2269
2270 err = do_epoll_wait(epfd, events, maxevents, timeout);
2271
2272 /*
2273 * If we changed the signal mask, we need to restore the original one.
2274 * In case we've got a signal while waiting, we do not restore the
2275 * signal mask yet, and we allow do_signal() to deliver the signal on
2276 * the way back to userspace, before the signal mask is restored.
2277 */
2278 if (sigmask) {
2279 if (err == -EINTR) {
2280 memcpy(¤t->saved_sigmask, &sigsaved,
2281 sizeof(sigsaved));
2282 set_restore_sigmask();
2283 } else
2284 set_current_blocked(&sigsaved);
2285 }
2286
2287 return err;
2288 }
2289 #endif
2290
eventpoll_init(void)2291 static int __init eventpoll_init(void)
2292 {
2293 struct sysinfo si;
2294
2295 si_meminfo(&si);
2296 /*
2297 * Allows top 4% of lomem to be allocated for epoll watches (per user).
2298 */
2299 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
2300 EP_ITEM_COST;
2301 BUG_ON(max_user_watches < 0);
2302
2303 /*
2304 * Initialize the structure used to perform epoll file descriptor
2305 * inclusion loops checks.
2306 */
2307 ep_nested_calls_init(&poll_loop_ncalls);
2308
2309 #ifdef CONFIG_DEBUG_LOCK_ALLOC
2310 /* Initialize the structure used to perform safe poll wait head wake ups */
2311 ep_nested_calls_init(&poll_safewake_ncalls);
2312 #endif
2313
2314 /*
2315 * We can have many thousands of epitems, so prevent this from
2316 * using an extra cache line on 64-bit (and smaller) CPUs
2317 */
2318 BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
2319
2320 /* Allocates slab cache used to allocate "struct epitem" items */
2321 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
2322 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
2323
2324 /* Allocates slab cache used to allocate "struct eppoll_entry" */
2325 pwq_cache = kmem_cache_create("eventpoll_pwq",
2326 sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
2327
2328 return 0;
2329 }
2330 fs_initcall(eventpoll_init);
2331