brintos

brintos / linux-shallow public Read only

0
0
Text · 70.2 KiB · a0c02f8 Raw
2621 lines · c
1// SPDX-License-Identifier: GPL-2.0-or-later2/*3 *  fs/eventpoll.c (Efficient event retrieval implementation)4 *  Copyright (C) 2001,...,2009	 Davide Libenzi5 *6 *  Davide Libenzi <davidel@xmailserver.org>7 */8 9#include <linux/init.h>10#include <linux/kernel.h>11#include <linux/sched/signal.h>12#include <linux/fs.h>13#include <linux/file.h>14#include <linux/signal.h>15#include <linux/errno.h>16#include <linux/mm.h>17#include <linux/slab.h>18#include <linux/poll.h>19#include <linux/string.h>20#include <linux/list.h>21#include <linux/hash.h>22#include <linux/spinlock.h>23#include <linux/syscalls.h>24#include <linux/rbtree.h>25#include <linux/wait.h>26#include <linux/eventpoll.h>27#include <linux/mount.h>28#include <linux/bitops.h>29#include <linux/mutex.h>30#include <linux/anon_inodes.h>31#include <linux/device.h>32#include <linux/uaccess.h>33#include <asm/io.h>34#include <asm/mman.h>35#include <linux/atomic.h>36#include <linux/proc_fs.h>37#include <linux/seq_file.h>38#include <linux/compat.h>39#include <linux/rculist.h>40#include <linux/capability.h>41#include <net/busy_poll.h>42 43/*44 * LOCKING:45 * There are three level of locking required by epoll :46 *47 * 1) epnested_mutex (mutex)48 * 2) ep->mtx (mutex)49 * 3) ep->lock (rwlock)50 *51 * The acquire order is the one listed above, from 1 to 3.52 * We need a rwlock (ep->lock) because we manipulate objects53 * from inside the poll callback, that might be triggered from54 * a wake_up() that in turn might be called from IRQ context.55 * So we can't sleep inside the poll callback and hence we need56 * a spinlock. During the event transfer loop (from kernel to57 * user space) we could end up sleeping due a copy_to_user(), so58 * we need a lock that will allow us to sleep. This lock is a59 * mutex (ep->mtx). It is acquired during the event transfer loop,60 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file().61 * The epnested_mutex is acquired when inserting an epoll fd onto another62 * epoll fd. We do this so that we walk the epoll tree and ensure that this63 * insertion does not create a cycle of epoll file descriptors, which64 * could lead to deadlock. We need a global mutex to prevent two65 * simultaneous inserts (A into B and B into A) from racing and66 * constructing a cycle without either insert observing that it is67 * going to.68 * It is necessary to acquire multiple "ep->mtx"es at once in the69 * case when one epoll fd is added to another. In this case, we70 * always acquire the locks in the order of nesting (i.e. after71 * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired72 * before e2->mtx). Since we disallow cycles of epoll file73 * descriptors, this ensures that the mutexes are well-ordered. In74 * order to communicate this nesting to lockdep, when walking a tree75 * of epoll file descriptors, we use the current recursion depth as76 * the lockdep subkey.77 * It is possible to drop the "ep->mtx" and to use the global78 * mutex "epnested_mutex" (together with "ep->lock") to have it working,79 * but having "ep->mtx" will make the interface more scalable.80 * Events that require holding "epnested_mutex" are very rare, while for81 * normal operations the epoll private "ep->mtx" will guarantee82 * a better scalability.83 */84 85/* Epoll private bits inside the event mask */86#define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)87 88#define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)89 90#define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \91				EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)92 93/* Maximum number of nesting allowed inside epoll sets */94#define EP_MAX_NESTS 495 96#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))97 98#define EP_UNACTIVE_PTR ((void *) -1L)99 100#define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))101 102struct epoll_filefd {103	struct file *file;104	int fd;105} __packed;106 107/* Wait structure used by the poll hooks */108struct eppoll_entry {109	/* List header used to link this structure to the "struct epitem" */110	struct eppoll_entry *next;111 112	/* The "base" pointer is set to the container "struct epitem" */113	struct epitem *base;114 115	/*116	 * Wait queue item that will be linked to the target file wait117	 * queue head.118	 */119	wait_queue_entry_t wait;120 121	/* The wait queue head that linked the "wait" wait queue item */122	wait_queue_head_t *whead;123};124 125/*126 * Each file descriptor added to the eventpoll interface will127 * have an entry of this type linked to the "rbr" RB tree.128 * Avoid increasing the size of this struct, there can be many thousands129 * of these on a server and we do not want this to take another cache line.130 */131struct epitem {132	union {133		/* RB tree node links this structure to the eventpoll RB tree */134		struct rb_node rbn;135		/* Used to free the struct epitem */136		struct rcu_head rcu;137	};138 139	/* List header used to link this structure to the eventpoll ready list */140	struct list_head rdllink;141 142	/*143	 * Works together "struct eventpoll"->ovflist in keeping the144	 * single linked chain of items.145	 */146	struct epitem *next;147 148	/* The file descriptor information this item refers to */149	struct epoll_filefd ffd;150 151	/*152	 * Protected by file->f_lock, true for to-be-released epitem already153	 * removed from the "struct file" items list; together with154	 * eventpoll->refcount orchestrates "struct eventpoll" disposal155	 */156	bool dying;157 158	/* List containing poll wait queues */159	struct eppoll_entry *pwqlist;160 161	/* The "container" of this item */162	struct eventpoll *ep;163 164	/* List header used to link this item to the "struct file" items list */165	struct hlist_node fllink;166 167	/* wakeup_source used when EPOLLWAKEUP is set */168	struct wakeup_source __rcu *ws;169 170	/* The structure that describe the interested events and the source fd */171	struct epoll_event event;172};173 174/*175 * This structure is stored inside the "private_data" member of the file176 * structure and represents the main data structure for the eventpoll177 * interface.178 */179struct eventpoll {180	/*181	 * This mutex is used to ensure that files are not removed182	 * while epoll is using them. This is held during the event183	 * collection loop, the file cleanup path, the epoll file exit184	 * code and the ctl operations.185	 */186	struct mutex mtx;187 188	/* Wait queue used by sys_epoll_wait() */189	wait_queue_head_t wq;190 191	/* Wait queue used by file->poll() */192	wait_queue_head_t poll_wait;193 194	/* List of ready file descriptors */195	struct list_head rdllist;196 197	/* Lock which protects rdllist and ovflist */198	rwlock_t lock;199 200	/* RB tree root used to store monitored fd structs */201	struct rb_root_cached rbr;202 203	/*204	 * This is a single linked list that chains all the "struct epitem" that205	 * happened while transferring ready events to userspace w/out206	 * holding ->lock.207	 */208	struct epitem *ovflist;209 210	/* wakeup_source used when ep_send_events or __ep_eventpoll_poll is running */211	struct wakeup_source *ws;212 213	/* The user that created the eventpoll descriptor */214	struct user_struct *user;215 216	struct file *file;217 218	/* used to optimize loop detection check */219	u64 gen;220	struct hlist_head refs;221 222	/*223	 * usage count, used together with epitem->dying to224	 * orchestrate the disposal of this struct225	 */226	refcount_t refcount;227 228#ifdef CONFIG_NET_RX_BUSY_POLL229	/* used to track busy poll napi_id */230	unsigned int napi_id;231	/* busy poll timeout */232	u32 busy_poll_usecs;233	/* busy poll packet budget */234	u16 busy_poll_budget;235	bool prefer_busy_poll;236#endif237 238#ifdef CONFIG_DEBUG_LOCK_ALLOC239	/* tracks wakeup nests for lockdep validation */240	u8 nests;241#endif242};243 244/* Wrapper struct used by poll queueing */245struct ep_pqueue {246	poll_table pt;247	struct epitem *epi;248};249 250/*251 * Configuration options available inside /proc/sys/fs/epoll/252 */253/* Maximum number of epoll watched descriptors, per user */254static long max_user_watches __read_mostly;255 256/* Used for cycles detection */257static DEFINE_MUTEX(epnested_mutex);258 259static u64 loop_check_gen = 0;260 261/* Used to check for epoll file descriptor inclusion loops */262static struct eventpoll *inserting_into;263 264/* Slab cache used to allocate "struct epitem" */265static struct kmem_cache *epi_cache __ro_after_init;266 267/* Slab cache used to allocate "struct eppoll_entry" */268static struct kmem_cache *pwq_cache __ro_after_init;269 270/*271 * List of files with newly added links, where we may need to limit the number272 * of emanating paths. Protected by the epnested_mutex.273 */274struct epitems_head {275	struct hlist_head epitems;276	struct epitems_head *next;277};278static struct epitems_head *tfile_check_list = EP_UNACTIVE_PTR;279 280static struct kmem_cache *ephead_cache __ro_after_init;281 282static inline void free_ephead(struct epitems_head *head)283{284	if (head)285		kmem_cache_free(ephead_cache, head);286}287 288static void list_file(struct file *file)289{290	struct epitems_head *head;291 292	head = container_of(file->f_ep, struct epitems_head, epitems);293	if (!head->next) {294		head->next = tfile_check_list;295		tfile_check_list = head;296	}297}298 299static void unlist_file(struct epitems_head *head)300{301	struct epitems_head *to_free = head;302	struct hlist_node *p = rcu_dereference(hlist_first_rcu(&head->epitems));303	if (p) {304		struct epitem *epi= container_of(p, struct epitem, fllink);305		spin_lock(&epi->ffd.file->f_lock);306		if (!hlist_empty(&head->epitems))307			to_free = NULL;308		head->next = NULL;309		spin_unlock(&epi->ffd.file->f_lock);310	}311	free_ephead(to_free);312}313 314#ifdef CONFIG_SYSCTL315 316#include <linux/sysctl.h>317 318static long long_zero;319static long long_max = LONG_MAX;320 321static struct ctl_table epoll_table[] = {322	{323		.procname	= "max_user_watches",324		.data		= &max_user_watches,325		.maxlen		= sizeof(max_user_watches),326		.mode		= 0644,327		.proc_handler	= proc_doulongvec_minmax,328		.extra1		= &long_zero,329		.extra2		= &long_max,330	},331};332 333static void __init epoll_sysctls_init(void)334{335	register_sysctl("fs/epoll", epoll_table);336}337#else338#define epoll_sysctls_init() do { } while (0)339#endif /* CONFIG_SYSCTL */340 341static const struct file_operations eventpoll_fops;342 343static inline int is_file_epoll(struct file *f)344{345	return f->f_op == &eventpoll_fops;346}347 348/* Setup the structure that is used as key for the RB tree */349static inline void ep_set_ffd(struct epoll_filefd *ffd,350			      struct file *file, int fd)351{352	ffd->file = file;353	ffd->fd = fd;354}355 356/* Compare RB tree keys */357static inline int ep_cmp_ffd(struct epoll_filefd *p1,358			     struct epoll_filefd *p2)359{360	return (p1->file > p2->file ? +1:361	        (p1->file < p2->file ? -1 : p1->fd - p2->fd));362}363 364/* Tells us if the item is currently linked */365static inline int ep_is_linked(struct epitem *epi)366{367	return !list_empty(&epi->rdllink);368}369 370static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)371{372	return container_of(p, struct eppoll_entry, wait);373}374 375/* Get the "struct epitem" from a wait queue pointer */376static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)377{378	return container_of(p, struct eppoll_entry, wait)->base;379}380 381/**382 * ep_events_available - Checks if ready events might be available.383 *384 * @ep: Pointer to the eventpoll context.385 *386 * Return: a value different than %zero if ready events are available,387 *          or %zero otherwise.388 */389static inline int ep_events_available(struct eventpoll *ep)390{391	return !list_empty_careful(&ep->rdllist) ||392		READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;393}394 395#ifdef CONFIG_NET_RX_BUSY_POLL396/**397 * busy_loop_ep_timeout - check if busy poll has timed out. The timeout value398 * from the epoll instance ep is preferred, but if it is not set fallback to399 * the system-wide global via busy_loop_timeout.400 *401 * @start_time: The start time used to compute the remaining time until timeout.402 * @ep: Pointer to the eventpoll context.403 *404 * Return: true if the timeout has expired, false otherwise.405 */406static bool busy_loop_ep_timeout(unsigned long start_time,407				 struct eventpoll *ep)408{409	unsigned long bp_usec = READ_ONCE(ep->busy_poll_usecs);410 411	if (bp_usec) {412		unsigned long end_time = start_time + bp_usec;413		unsigned long now = busy_loop_current_time();414 415		return time_after(now, end_time);416	} else {417		return busy_loop_timeout(start_time);418	}419}420 421static bool ep_busy_loop_on(struct eventpoll *ep)422{423	return !!READ_ONCE(ep->busy_poll_usecs) || net_busy_loop_on();424}425 426static bool ep_busy_loop_end(void *p, unsigned long start_time)427{428	struct eventpoll *ep = p;429 430	return ep_events_available(ep) || busy_loop_ep_timeout(start_time, ep);431}432 433/*434 * Busy poll if globally on and supporting sockets found && no events,435 * busy loop will return if need_resched or ep_events_available.436 *437 * we must do our busy polling with irqs enabled438 */439static bool ep_busy_loop(struct eventpoll *ep, int nonblock)440{441	unsigned int napi_id = READ_ONCE(ep->napi_id);442	u16 budget = READ_ONCE(ep->busy_poll_budget);443	bool prefer_busy_poll = READ_ONCE(ep->prefer_busy_poll);444 445	if (!budget)446		budget = BUSY_POLL_BUDGET;447 448	if (napi_id >= MIN_NAPI_ID && ep_busy_loop_on(ep)) {449		napi_busy_loop(napi_id, nonblock ? NULL : ep_busy_loop_end,450			       ep, prefer_busy_poll, budget);451		if (ep_events_available(ep))452			return true;453		/*454		 * Busy poll timed out.  Drop NAPI ID for now, we can add455		 * it back in when we have moved a socket with a valid NAPI456		 * ID onto the ready list.457		 */458		ep->napi_id = 0;459		return false;460	}461	return false;462}463 464/*465 * Set epoll busy poll NAPI ID from sk.466 */467static inline void ep_set_busy_poll_napi_id(struct epitem *epi)468{469	struct eventpoll *ep = epi->ep;470	unsigned int napi_id;471	struct socket *sock;472	struct sock *sk;473 474	if (!ep_busy_loop_on(ep))475		return;476 477	sock = sock_from_file(epi->ffd.file);478	if (!sock)479		return;480 481	sk = sock->sk;482	if (!sk)483		return;484 485	napi_id = READ_ONCE(sk->sk_napi_id);486 487	/* Non-NAPI IDs can be rejected488	 *	or489	 * Nothing to do if we already have this ID490	 */491	if (napi_id < MIN_NAPI_ID || napi_id == ep->napi_id)492		return;493 494	/* record NAPI ID for use in next busy poll */495	ep->napi_id = napi_id;496}497 498static long ep_eventpoll_bp_ioctl(struct file *file, unsigned int cmd,499				  unsigned long arg)500{501	struct eventpoll *ep = file->private_data;502	void __user *uarg = (void __user *)arg;503	struct epoll_params epoll_params;504 505	switch (cmd) {506	case EPIOCSPARAMS:507		if (copy_from_user(&epoll_params, uarg, sizeof(epoll_params)))508			return -EFAULT;509 510		/* pad byte must be zero */511		if (epoll_params.__pad)512			return -EINVAL;513 514		if (epoll_params.busy_poll_usecs > S32_MAX)515			return -EINVAL;516 517		if (epoll_params.prefer_busy_poll > 1)518			return -EINVAL;519 520		if (epoll_params.busy_poll_budget > NAPI_POLL_WEIGHT &&521		    !capable(CAP_NET_ADMIN))522			return -EPERM;523 524		WRITE_ONCE(ep->busy_poll_usecs, epoll_params.busy_poll_usecs);525		WRITE_ONCE(ep->busy_poll_budget, epoll_params.busy_poll_budget);526		WRITE_ONCE(ep->prefer_busy_poll, epoll_params.prefer_busy_poll);527		return 0;528	case EPIOCGPARAMS:529		memset(&epoll_params, 0, sizeof(epoll_params));530		epoll_params.busy_poll_usecs = READ_ONCE(ep->busy_poll_usecs);531		epoll_params.busy_poll_budget = READ_ONCE(ep->busy_poll_budget);532		epoll_params.prefer_busy_poll = READ_ONCE(ep->prefer_busy_poll);533		if (copy_to_user(uarg, &epoll_params, sizeof(epoll_params)))534			return -EFAULT;535		return 0;536	default:537		return -ENOIOCTLCMD;538	}539}540 541#else542 543static inline bool ep_busy_loop(struct eventpoll *ep, int nonblock)544{545	return false;546}547 548static inline void ep_set_busy_poll_napi_id(struct epitem *epi)549{550}551 552static long ep_eventpoll_bp_ioctl(struct file *file, unsigned int cmd,553				  unsigned long arg)554{555	return -EOPNOTSUPP;556}557 558#endif /* CONFIG_NET_RX_BUSY_POLL */559 560/*561 * As described in commit 0ccf831cb lockdep: annotate epoll562 * the use of wait queues used by epoll is done in a very controlled563 * manner. Wake ups can nest inside each other, but are never done564 * with the same locking. For example:565 *566 *   dfd = socket(...);567 *   efd1 = epoll_create();568 *   efd2 = epoll_create();569 *   epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);570 *   epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);571 *572 * When a packet arrives to the device underneath "dfd", the net code will573 * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a574 * callback wakeup entry on that queue, and the wake_up() performed by the575 * "dfd" net code will end up in ep_poll_callback(). At this point epoll576 * (efd1) notices that it may have some event ready, so it needs to wake up577 * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()578 * that ends up in another wake_up(), after having checked about the579 * recursion constraints. That are, no more than EP_MAX_NESTS, to avoid580 * stack blasting.581 *582 * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle583 * this special case of epoll.584 */585#ifdef CONFIG_DEBUG_LOCK_ALLOC586 587static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,588			     unsigned pollflags)589{590	struct eventpoll *ep_src;591	unsigned long flags;592	u8 nests = 0;593 594	/*595	 * To set the subclass or nesting level for spin_lock_irqsave_nested()596	 * it might be natural to create a per-cpu nest count. However, since597	 * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can598	 * schedule() in the -rt kernel, the per-cpu variable are no longer599	 * protected. Thus, we are introducing a per eventpoll nest field.600	 * If we are not being call from ep_poll_callback(), epi is NULL and601	 * we are at the first level of nesting, 0. Otherwise, we are being602	 * called from ep_poll_callback() and if a previous wakeup source is603	 * not an epoll file itself, we are at depth 1 since the wakeup source604	 * is depth 0. If the wakeup source is a previous epoll file in the605	 * wakeup chain then we use its nests value and record ours as606	 * nests + 1. The previous epoll file nests value is stable since its607	 * already holding its own poll_wait.lock.608	 */609	if (epi) {610		if ((is_file_epoll(epi->ffd.file))) {611			ep_src = epi->ffd.file->private_data;612			nests = ep_src->nests;613		} else {614			nests = 1;615		}616	}617	spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);618	ep->nests = nests + 1;619	wake_up_locked_poll(&ep->poll_wait, EPOLLIN | pollflags);620	ep->nests = 0;621	spin_unlock_irqrestore(&ep->poll_wait.lock, flags);622}623 624#else625 626static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,627			     __poll_t pollflags)628{629	wake_up_poll(&ep->poll_wait, EPOLLIN | pollflags);630}631 632#endif633 634static void ep_remove_wait_queue(struct eppoll_entry *pwq)635{636	wait_queue_head_t *whead;637 638	rcu_read_lock();639	/*640	 * If it is cleared by POLLFREE, it should be rcu-safe.641	 * If we read NULL we need a barrier paired with642	 * smp_store_release() in ep_poll_callback(), otherwise643	 * we rely on whead->lock.644	 */645	whead = smp_load_acquire(&pwq->whead);646	if (whead)647		remove_wait_queue(whead, &pwq->wait);648	rcu_read_unlock();649}650 651/*652 * This function unregisters poll callbacks from the associated file653 * descriptor.  Must be called with "mtx" held.654 */655static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)656{657	struct eppoll_entry **p = &epi->pwqlist;658	struct eppoll_entry *pwq;659 660	while ((pwq = *p) != NULL) {661		*p = pwq->next;662		ep_remove_wait_queue(pwq);663		kmem_cache_free(pwq_cache, pwq);664	}665}666 667/* call only when ep->mtx is held */668static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)669{670	return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));671}672 673/* call only when ep->mtx is held */674static inline void ep_pm_stay_awake(struct epitem *epi)675{676	struct wakeup_source *ws = ep_wakeup_source(epi);677 678	if (ws)679		__pm_stay_awake(ws);680}681 682static inline bool ep_has_wakeup_source(struct epitem *epi)683{684	return rcu_access_pointer(epi->ws) ? true : false;685}686 687/* call when ep->mtx cannot be held (ep_poll_callback) */688static inline void ep_pm_stay_awake_rcu(struct epitem *epi)689{690	struct wakeup_source *ws;691 692	rcu_read_lock();693	ws = rcu_dereference(epi->ws);694	if (ws)695		__pm_stay_awake(ws);696	rcu_read_unlock();697}698 699 700/*701 * ep->mutex needs to be held because we could be hit by702 * eventpoll_release_file() and epoll_ctl().703 */704static void ep_start_scan(struct eventpoll *ep, struct list_head *txlist)705{706	/*707	 * Steal the ready list, and re-init the original one to the708	 * empty list. Also, set ep->ovflist to NULL so that events709	 * happening while looping w/out locks, are not lost. We cannot710	 * have the poll callback to queue directly on ep->rdllist,711	 * because we want the "sproc" callback to be able to do it712	 * in a lockless way.713	 */714	lockdep_assert_irqs_enabled();715	write_lock_irq(&ep->lock);716	list_splice_init(&ep->rdllist, txlist);717	WRITE_ONCE(ep->ovflist, NULL);718	write_unlock_irq(&ep->lock);719}720 721static HWJS_SUSPENDS void ep_done_scan(struct eventpoll *ep,722			 struct list_head *txlist)723{724	struct epitem *epi, *nepi;725 726	write_lock_irq(&ep->lock);727	/*728	 * During the time we spent inside the "sproc" callback, some729	 * other events might have been queued by the poll callback.730	 * We re-insert them inside the main ready-list here.731	 */732	for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL;733	     nepi = epi->next, epi->next = EP_UNACTIVE_PTR) {734		/*735		 * We need to check if the item is already in the list.736		 * During the "sproc" callback execution time, items are737		 * queued into ->ovflist but the "txlist" might already738		 * contain them, and the list_splice() below takes care of them.739		 */740		if (!ep_is_linked(epi)) {741			/*742			 * ->ovflist is LIFO, so we have to reverse it in order743			 * to keep in FIFO.744			 */745			list_add(&epi->rdllink, &ep->rdllist);746			ep_pm_stay_awake(epi);747		}748	}749	/*750	 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after751	 * releasing the lock, events will be queued in the normal way inside752	 * ep->rdllist.753	 */754	WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);755 756	/*757	 * Quickly re-inject items left on "txlist".758	 */759	list_splice(txlist, &ep->rdllist);760	__pm_relax(ep->ws);761 762	if (!list_empty(&ep->rdllist)) {763		if (waitqueue_active(&ep->wq))764			wake_up(&ep->wq);765	}766 767	write_unlock_irq(&ep->lock);768}769 770static void ep_get(struct eventpoll *ep)771{772	refcount_inc(&ep->refcount);773}774 775/*776 * Returns true if the event poll can be disposed777 */778static bool ep_refcount_dec_and_test(struct eventpoll *ep)779{780	if (!refcount_dec_and_test(&ep->refcount))781		return false;782 783	WARN_ON_ONCE(!RB_EMPTY_ROOT(&ep->rbr.rb_root));784	return true;785}786 787static void ep_free(struct eventpoll *ep)788{789	mutex_destroy(&ep->mtx);790	free_uid(ep->user);791	wakeup_source_unregister(ep->ws);792	kfree(ep);793}794 795/*796 * Removes a "struct epitem" from the eventpoll RB tree and deallocates797 * all the associated resources. Must be called with "mtx" held.798 * If the dying flag is set, do the removal only if force is true.799 * This prevents ep_clear_and_put() from dropping all the ep references800 * while running concurrently with eventpoll_release_file().801 * Returns true if the eventpoll can be disposed.802 */803static HWJS_SUSPENDS bool __ep_remove(struct eventpoll *ep, struct epitem *epi, bool force)804{805	struct file *file = epi->ffd.file;806	struct epitems_head *to_free;807	struct hlist_head *head;808 809	lockdep_assert_irqs_enabled();810 811	/*812	 * Removes poll wait queue hooks.813	 */814	ep_unregister_pollwait(ep, epi);815 816	/* Remove the current item from the list of epoll hooks */817	spin_lock(&file->f_lock);818	if (epi->dying && !force) {819		spin_unlock(&file->f_lock);820		return false;821	}822 823	to_free = NULL;824	head = file->f_ep;825	if (head->first == &epi->fllink && !epi->fllink.next) {826		file->f_ep = NULL;827		if (!is_file_epoll(file)) {828			struct epitems_head *v;829			v = container_of(head, struct epitems_head, epitems);830			if (!smp_load_acquire(&v->next))831				to_free = v;832		}833	}834	hlist_del_rcu(&epi->fllink);835	spin_unlock(&file->f_lock);836	free_ephead(to_free);837 838	rb_erase_cached(&epi->rbn, &ep->rbr);839 840	write_lock_irq(&ep->lock);841	if (ep_is_linked(epi))842		list_del_init(&epi->rdllink);843	write_unlock_irq(&ep->lock);844 845	wakeup_source_unregister(ep_wakeup_source(epi));846	/*847	 * At this point it is safe to free the eventpoll item. Use the union848	 * field epi->rcu, since we are trying to minimize the size of849	 * 'struct epitem'. The 'rbn' field is no longer in use. Protected by850	 * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make851	 * use of the rbn field.852	 */853	kfree_rcu(epi, rcu);854 855	percpu_counter_dec(&ep->user->epoll_watches);856	return ep_refcount_dec_and_test(ep);857}858 859/*860 * ep_remove variant for callers owing an additional reference to the ep861 */862static HWJS_SUSPENDS void ep_remove_safe(struct eventpoll *ep, struct epitem *epi)863{864	WARN_ON_ONCE(__ep_remove(ep, epi, false));865}866 867static HWJS_SUSPENDS void ep_clear_and_put(struct eventpoll *ep)868{869	struct rb_node *rbp, *next;870	struct epitem *epi;871	bool dispose;872 873	/* We need to release all tasks waiting for these file */874	if (waitqueue_active(&ep->poll_wait))875		ep_poll_safewake(ep, NULL, 0);876 877	mutex_lock(&ep->mtx);878 879	/*880	 * Walks through the whole tree by unregistering poll callbacks.881	 */882	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {883		epi = rb_entry(rbp, struct epitem, rbn);884 885		ep_unregister_pollwait(ep, epi);886		cond_resched();887	}888 889	/*890	 * Walks through the whole tree and try to free each "struct epitem".891	 * Note that ep_remove_safe() will not remove the epitem in case of a892	 * racing eventpoll_release_file(); the latter will do the removal.893	 * At this point we are sure no poll callbacks will be lingering around.894	 * Since we still own a reference to the eventpoll struct, the loop can't895	 * dispose it.896	 */897	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = next) {898		next = rb_next(rbp);899		epi = rb_entry(rbp, struct epitem, rbn);900		ep_remove_safe(ep, epi);901		cond_resched();902	}903 904	dispose = ep_refcount_dec_and_test(ep);905	mutex_unlock(&ep->mtx);906 907	if (dispose)908		ep_free(ep);909}910 911static long ep_eventpoll_ioctl(struct file *file, unsigned int cmd,912			       unsigned long arg)913{914	int ret;915 916	if (!is_file_epoll(file))917		return -EINVAL;918 919	switch (cmd) {920	case EPIOCSPARAMS:921	case EPIOCGPARAMS:922		ret = ep_eventpoll_bp_ioctl(file, cmd, arg);923		break;924	default:925		ret = -EINVAL;926		break;927	}928 929	return ret;930}931 932static HWJS_SUSPENDS int ep_eventpoll_release(struct inode *inode, struct file *file)933{934	struct eventpoll *ep = file->private_data;935 936	if (ep)937		ep_clear_and_put(ep);938 939	return 0;940}941 942static HWJS_SUSPENDS __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt, int depth);943 944static HWJS_SUSPENDS __poll_t __ep_eventpoll_poll(struct file *file, poll_table *wait, int depth)945{946	struct eventpoll *ep = file->private_data;947	LIST_HEAD(txlist);948	struct epitem *epi, *tmp;949	poll_table pt;950	__poll_t res = 0;951 952	init_poll_funcptr(&pt, NULL);953 954	/* Insert inside our poll wait queue */955	poll_wait(file, &ep->poll_wait, wait);956 957	/*958	 * Proceed to find out if wanted events are really available inside959	 * the ready list.960	 */961	mutex_lock_nested(&ep->mtx, depth);962	ep_start_scan(ep, &txlist);963	list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {964		if (ep_item_poll(epi, &pt, depth + 1)) {965			res = EPOLLIN | EPOLLRDNORM;966			break;967		} else {968			/*969			 * Item has been dropped into the ready list by the poll970			 * callback, but it's not actually ready, as far as971			 * caller requested events goes. We can remove it here.972			 */973			__pm_relax(ep_wakeup_source(epi));974			list_del_init(&epi->rdllink);975		}976	}977	ep_done_scan(ep, &txlist);978	mutex_unlock(&ep->mtx);979	return res;980}981 982/*983 * The ffd.file pointer may be in the process of being torn down due to984 * being closed, but we may not have finished eventpoll_release() yet.985 *986 * Normally, even with the atomic_long_inc_not_zero, the file may have987 * been free'd and then gotten re-allocated to something else (since988 * files are not RCU-delayed, they are SLAB_TYPESAFE_BY_RCU).989 *990 * But for epoll, users hold the ep->mtx mutex, and as such any file in991 * the process of being free'd will block in eventpoll_release_file()992 * and thus the underlying file allocation will not be free'd, and the993 * file re-use cannot happen.994 *995 * For the same reason we can avoid a rcu_read_lock() around the996 * operation - 'ffd.file' cannot go away even if the refcount has997 * reached zero (but we must still not call out to ->poll() functions998 * etc).999 */1000static struct file *epi_fget(const struct epitem *epi)1001{1002	struct file *file;1003 1004	file = epi->ffd.file;1005	if (!atomic_long_inc_not_zero(&file->f_count))1006		file = NULL;1007	return file;1008}1009 1010/*1011 * Differs from ep_eventpoll_poll() in that internal callers already have1012 * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()1013 * is correctly annotated.1014 */1015static HWJS_SUSPENDS __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,1016				 int depth)1017{1018	struct file *file = epi_fget(epi);1019	__poll_t res;1020 1021	/*1022	 * We could return EPOLLERR | EPOLLHUP or something, but let's1023	 * treat this more as "file doesn't exist, poll didn't happen".1024	 */1025	if (!file)1026		return 0;1027 1028	pt->_key = epi->event.events;1029	if (!is_file_epoll(file))1030		res = vfs_poll(file, pt);1031	else1032		res = __ep_eventpoll_poll(file, pt, depth);1033	fput(file);1034	return res & epi->event.events;1035}1036 1037static HWJS_SUSPENDS __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)1038{1039	return __ep_eventpoll_poll(file, wait, 0);1040}1041 1042#ifdef CONFIG_PROC_FS1043static void ep_show_fdinfo(struct seq_file *m, struct file *f)1044{1045	struct eventpoll *ep = f->private_data;1046	struct rb_node *rbp;1047 1048	mutex_lock(&ep->mtx);1049	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {1050		struct epitem *epi = rb_entry(rbp, struct epitem, rbn);1051		struct inode *inode = file_inode(epi->ffd.file);1052 1053		seq_printf(m, "tfd: %8d events: %8x data: %16llx "1054			   " pos:%lli ino:%lx sdev:%x\n",1055			   epi->ffd.fd, epi->event.events,1056			   (long long)epi->event.data,1057			   (long long)epi->ffd.file->f_pos,1058			   inode->i_ino, inode->i_sb->s_dev);1059		if (seq_has_overflowed(m))1060			break;1061	}1062	mutex_unlock(&ep->mtx);1063}1064#endif1065 1066/* File callbacks that implement the eventpoll file behaviour */1067static const struct file_operations eventpoll_fops = {1068#ifdef CONFIG_PROC_FS1069	.show_fdinfo	= ep_show_fdinfo,1070#endif1071	.release	= ep_eventpoll_release,1072	.poll		= ep_eventpoll_poll,1073	.llseek		= noop_llseek,1074	.unlocked_ioctl	= ep_eventpoll_ioctl,1075	.compat_ioctl   = compat_ptr_ioctl,1076};1077 1078/*1079 * This is called from eventpoll_release() to unlink files from the eventpoll1080 * interface. We need to have this facility to cleanup correctly files that are1081 * closed without being removed from the eventpoll interface.1082 */1083void eventpoll_release_file(struct file *file)1084{1085	struct eventpoll *ep;1086	struct epitem *epi;1087	bool dispose;1088 1089	/*1090	 * Use the 'dying' flag to prevent a concurrent ep_clear_and_put() from1091	 * touching the epitems list before eventpoll_release_file() can access1092	 * the ep->mtx.1093	 */1094again:1095	spin_lock(&file->f_lock);1096	if (file->f_ep && file->f_ep->first) {1097		epi = hlist_entry(file->f_ep->first, struct epitem, fllink);1098		epi->dying = true;1099		spin_unlock(&file->f_lock);1100 1101		/*1102		 * ep access is safe as we still own a reference to the ep1103		 * struct1104		 */1105		ep = epi->ep;1106		mutex_lock(&ep->mtx);1107		dispose = __ep_remove(ep, epi, true);1108		mutex_unlock(&ep->mtx);1109 1110		if (dispose)1111			ep_free(ep);1112		goto again;1113	}1114	spin_unlock(&file->f_lock);1115}1116 1117static int ep_alloc(struct eventpoll **pep)1118{1119	struct eventpoll *ep;1120 1121	ep = kzalloc(sizeof(*ep), GFP_KERNEL);1122	if (unlikely(!ep))1123		return -ENOMEM;1124 1125	mutex_init(&ep->mtx);1126	rwlock_init(&ep->lock);1127	init_waitqueue_head(&ep->wq);1128	init_waitqueue_head(&ep->poll_wait);1129	INIT_LIST_HEAD(&ep->rdllist);1130	ep->rbr = RB_ROOT_CACHED;1131	ep->ovflist = EP_UNACTIVE_PTR;1132	ep->user = get_current_user();1133	refcount_set(&ep->refcount, 1);1134 1135	*pep = ep;1136 1137	return 0;1138}1139 1140/*1141 * Search the file inside the eventpoll tree. The RB tree operations1142 * are protected by the "mtx" mutex, and ep_find() must be called with1143 * "mtx" held.1144 */1145static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd)1146{1147	int kcmp;1148	struct rb_node *rbp;1149	struct epitem *epi, *epir = NULL;1150	struct epoll_filefd ffd;1151 1152	ep_set_ffd(&ffd, file, fd);1153	for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {1154		epi = rb_entry(rbp, struct epitem, rbn);1155		kcmp = ep_cmp_ffd(&ffd, &epi->ffd);1156		if (kcmp > 0)1157			rbp = rbp->rb_right;1158		else if (kcmp < 0)1159			rbp = rbp->rb_left;1160		else {1161			epir = epi;1162			break;1163		}1164	}1165 1166	return epir;1167}1168 1169#ifdef CONFIG_KCMP1170static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)1171{1172	struct rb_node *rbp;1173	struct epitem *epi;1174 1175	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {1176		epi = rb_entry(rbp, struct epitem, rbn);1177		if (epi->ffd.fd == tfd) {1178			if (toff == 0)1179				return epi;1180			else1181				toff--;1182		}1183		cond_resched();1184	}1185 1186	return NULL;1187}1188 1189struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,1190				     unsigned long toff)1191{1192	struct file *file_raw;1193	struct eventpoll *ep;1194	struct epitem *epi;1195 1196	if (!is_file_epoll(file))1197		return ERR_PTR(-EINVAL);1198 1199	ep = file->private_data;1200 1201	mutex_lock(&ep->mtx);1202	epi = ep_find_tfd(ep, tfd, toff);1203	if (epi)1204		file_raw = epi->ffd.file;1205	else1206		file_raw = ERR_PTR(-ENOENT);1207	mutex_unlock(&ep->mtx);1208 1209	return file_raw;1210}1211#endif /* CONFIG_KCMP */1212 1213/*1214 * Adds a new entry to the tail of the list in a lockless way, i.e.1215 * multiple CPUs are allowed to call this function concurrently.1216 *1217 * Beware: it is necessary to prevent any other modifications of the1218 *         existing list until all changes are completed, in other words1219 *         concurrent list_add_tail_lockless() calls should be protected1220 *         with a read lock, where write lock acts as a barrier which1221 *         makes sure all list_add_tail_lockless() calls are fully1222 *         completed.1223 *1224 *        Also an element can be locklessly added to the list only in one1225 *        direction i.e. either to the tail or to the head, otherwise1226 *        concurrent access will corrupt the list.1227 *1228 * Return: %false if element has been already added to the list, %true1229 * otherwise.1230 */1231static inline bool list_add_tail_lockless(struct list_head *new,1232					  struct list_head *head)1233{1234	struct list_head *prev;1235 1236	/*1237	 * This is simple 'new->next = head' operation, but cmpxchg()1238	 * is used in order to detect that same element has been just1239	 * added to the list from another CPU: the winner observes1240	 * new->next == new.1241	 */1242	if (!try_cmpxchg(&new->next, &new, head))1243		return false;1244 1245	/*1246	 * Initially ->next of a new element must be updated with the head1247	 * (we are inserting to the tail) and only then pointers are atomically1248	 * exchanged.  XCHG guarantees memory ordering, thus ->next should be1249	 * updated before pointers are actually swapped and pointers are1250	 * swapped before prev->next is updated.1251	 */1252 1253	prev = xchg(&head->prev, new);1254 1255	/*1256	 * It is safe to modify prev->next and new->prev, because a new element1257	 * is added only to the tail and new->next is updated before XCHG.1258	 */1259 1260	prev->next = new;1261	new->prev = prev;1262 1263	return true;1264}1265 1266/*1267 * Chains a new epi entry to the tail of the ep->ovflist in a lockless way,1268 * i.e. multiple CPUs are allowed to call this function concurrently.1269 *1270 * Return: %false if epi element has been already chained, %true otherwise.1271 */1272static inline bool chain_epi_lockless(struct epitem *epi)1273{1274	struct eventpoll *ep = epi->ep;1275 1276	/* Fast preliminary check */1277	if (epi->next != EP_UNACTIVE_PTR)1278		return false;1279 1280	/* Check that the same epi has not been just chained from another CPU */1281	if (cmpxchg(&epi->next, EP_UNACTIVE_PTR, NULL) != EP_UNACTIVE_PTR)1282		return false;1283 1284	/* Atomically exchange tail */1285	epi->next = xchg(&ep->ovflist, epi);1286 1287	return true;1288}1289 1290/*1291 * This is the callback that is passed to the wait queue wakeup1292 * mechanism. It is called by the stored file descriptors when they1293 * have events to report.1294 *1295 * This callback takes a read lock in order not to contend with concurrent1296 * events from another file descriptor, thus all modifications to ->rdllist1297 * or ->ovflist are lockless.  Read lock is paired with the write lock from1298 * ep_start/done_scan(), which stops all list modifications and guarantees1299 * that lists state is seen correctly.1300 *1301 * Another thing worth to mention is that ep_poll_callback() can be called1302 * concurrently for the same @epi from different CPUs if poll table was inited1303 * with several wait queues entries.  Plural wakeup from different CPUs of a1304 * single wait queue is serialized by wq.lock, but the case when multiple wait1305 * queues are used should be detected accordingly.  This is detected using1306 * cmpxchg() operation.1307 */1308static HWJS_SUSPENDS int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)1309{1310	int pwake = 0;1311	struct epitem *epi = ep_item_from_wait(wait);1312	struct eventpoll *ep = epi->ep;1313	__poll_t pollflags = key_to_poll(key);1314	unsigned long flags;1315	int ewake = 0;1316 1317	read_lock_irqsave(&ep->lock, flags);1318 1319	ep_set_busy_poll_napi_id(epi);1320 1321	/*1322	 * If the event mask does not contain any poll(2) event, we consider the1323	 * descriptor to be disabled. This condition is likely the effect of the1324	 * EPOLLONESHOT bit that disables the descriptor when an event is received,1325	 * until the next EPOLL_CTL_MOD will be issued.1326	 */1327	if (!(epi->event.events & ~EP_PRIVATE_BITS))1328		goto out_unlock;1329 1330	/*1331	 * Check the events coming with the callback. At this stage, not1332	 * every device reports the events in the "key" parameter of the1333	 * callback. We need to be able to handle both cases here, hence the1334	 * test for "key" != NULL before the event match test.1335	 */1336	if (pollflags && !(pollflags & epi->event.events))1337		goto out_unlock;1338 1339	/*1340	 * If we are transferring events to userspace, we can hold no locks1341	 * (because we're accessing user memory, and because of linux f_op->poll()1342	 * semantics). All the events that happen during that period of time are1343	 * chained in ep->ovflist and requeued later on.1344	 */1345	if (READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR) {1346		if (chain_epi_lockless(epi))1347			ep_pm_stay_awake_rcu(epi);1348	} else if (!ep_is_linked(epi)) {1349		/* In the usual case, add event to ready list. */1350		if (list_add_tail_lockless(&epi->rdllink, &ep->rdllist))1351			ep_pm_stay_awake_rcu(epi);1352	}1353 1354	/*1355	 * Wake up ( if active ) both the eventpoll wait list and the ->poll()1356	 * wait list.1357	 */1358	if (waitqueue_active(&ep->wq)) {1359		if ((epi->event.events & EPOLLEXCLUSIVE) &&1360					!(pollflags & POLLFREE)) {1361			switch (pollflags & EPOLLINOUT_BITS) {1362			case EPOLLIN:1363				if (epi->event.events & EPOLLIN)1364					ewake = 1;1365				break;1366			case EPOLLOUT:1367				if (epi->event.events & EPOLLOUT)1368					ewake = 1;1369				break;1370			case 0:1371				ewake = 1;1372				break;1373			}1374		}1375		wake_up(&ep->wq);1376	}1377	if (waitqueue_active(&ep->poll_wait))1378		pwake++;1379 1380out_unlock:1381	read_unlock_irqrestore(&ep->lock, flags);1382 1383	/* We have to call this outside the lock */1384	if (pwake)1385		ep_poll_safewake(ep, epi, pollflags & EPOLL_URING_WAKE);1386 1387	if (!(epi->event.events & EPOLLEXCLUSIVE))1388		ewake = 1;1389 1390	if (pollflags & POLLFREE) {1391		/*1392		 * If we race with ep_remove_wait_queue() it can miss1393		 * ->whead = NULL and do another remove_wait_queue() after1394		 * us, so we can't use __remove_wait_queue().1395		 */1396		list_del_init(&wait->entry);1397		/*1398		 * ->whead != NULL protects us from the race with1399		 * ep_clear_and_put() or ep_remove(), ep_remove_wait_queue()1400		 * takes whead->lock held by the caller. Once we nullify it,1401		 * nothing protects ep/epi or even wait.1402		 */1403		smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);1404	}1405 1406	return ewake;1407}1408 1409/*1410 * This is the callback that is used to add our wait queue to the1411 * target file wakeup lists.1412 */1413static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,1414				 poll_table *pt)1415{1416	struct ep_pqueue *epq = container_of(pt, struct ep_pqueue, pt);1417	struct epitem *epi = epq->epi;1418	struct eppoll_entry *pwq;1419 1420	if (unlikely(!epi))	// an earlier allocation has failed1421		return;1422 1423	pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL);1424	if (unlikely(!pwq)) {1425		epq->epi = NULL;1426		return;1427	}1428 1429	init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);1430	pwq->whead = whead;1431	pwq->base = epi;1432	if (epi->event.events & EPOLLEXCLUSIVE)1433		add_wait_queue_exclusive(whead, &pwq->wait);1434	else1435		add_wait_queue(whead, &pwq->wait);1436	pwq->next = epi->pwqlist;1437	epi->pwqlist = pwq;1438}1439 1440static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)1441{1442	int kcmp;1443	struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;1444	struct epitem *epic;1445	bool leftmost = true;1446 1447	while (*p) {1448		parent = *p;1449		epic = rb_entry(parent, struct epitem, rbn);1450		kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);1451		if (kcmp > 0) {1452			p = &parent->rb_right;1453			leftmost = false;1454		} else1455			p = &parent->rb_left;1456	}1457	rb_link_node(&epi->rbn, parent, p);1458	rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);1459}1460 1461 1462 1463#define PATH_ARR_SIZE 51464/*1465 * These are the number paths of length 1 to 5, that we are allowing to emanate1466 * from a single file of interest. For example, we allow 1000 paths of length1467 * 1, to emanate from each file of interest. This essentially represents the1468 * potential wakeup paths, which need to be limited in order to avoid massive1469 * uncontrolled wakeup storms. The common use case should be a single ep which1470 * is connected to n file sources. In this case each file source has 1 path1471 * of length 1. Thus, the numbers below should be more than sufficient. These1472 * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify1473 * and delete can't add additional paths. Protected by the epnested_mutex.1474 */1475static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };1476static int path_count[PATH_ARR_SIZE];1477 1478static int path_count_inc(int nests)1479{1480	/* Allow an arbitrary number of depth 1 paths */1481	if (nests == 0)1482		return 0;1483 1484	if (++path_count[nests] > path_limits[nests])1485		return -1;1486	return 0;1487}1488 1489static void path_count_init(void)1490{1491	int i;1492 1493	for (i = 0; i < PATH_ARR_SIZE; i++)1494		path_count[i] = 0;1495}1496 1497static int reverse_path_check_proc(struct hlist_head *refs, int depth)1498{1499	int error = 0;1500	struct epitem *epi;1501 1502	if (depth > EP_MAX_NESTS) /* too deep nesting */1503		return -1;1504 1505	/* CTL_DEL can remove links here, but that can't increase our count */1506	hlist_for_each_entry_rcu(epi, refs, fllink) {1507		struct hlist_head *refs = &epi->ep->refs;1508		if (hlist_empty(refs))1509			error = path_count_inc(depth);1510		else1511			error = reverse_path_check_proc(refs, depth + 1);1512		if (error != 0)1513			break;1514	}1515	return error;1516}1517 1518/**1519 * reverse_path_check - The tfile_check_list is list of epitem_head, which have1520 *                      links that are proposed to be newly added. We need to1521 *                      make sure that those added links don't add too many1522 *                      paths such that we will spend all our time waking up1523 *                      eventpoll objects.1524 *1525 * Return: %zero if the proposed links don't create too many paths,1526 *	    %-1 otherwise.1527 */1528static int reverse_path_check(void)1529{1530	struct epitems_head *p;1531 1532	for (p = tfile_check_list; p != EP_UNACTIVE_PTR; p = p->next) {1533		int error;1534		path_count_init();1535		rcu_read_lock();1536		error = reverse_path_check_proc(&p->epitems, 0);1537		rcu_read_unlock();1538		if (error)1539			return error;1540	}1541	return 0;1542}1543 1544static int ep_create_wakeup_source(struct epitem *epi)1545{1546	struct name_snapshot n;1547	struct wakeup_source *ws;1548 1549	if (!epi->ep->ws) {1550		epi->ep->ws = wakeup_source_register(NULL, "eventpoll");1551		if (!epi->ep->ws)1552			return -ENOMEM;1553	}1554 1555	take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);1556	ws = wakeup_source_register(NULL, n.name.name);1557	release_dentry_name_snapshot(&n);1558 1559	if (!ws)1560		return -ENOMEM;1561	rcu_assign_pointer(epi->ws, ws);1562 1563	return 0;1564}1565 1566/* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */1567static HWJS_SUSPENDS noinline void ep_destroy_wakeup_source(struct epitem *epi)1568{1569	struct wakeup_source *ws = ep_wakeup_source(epi);1570 1571	RCU_INIT_POINTER(epi->ws, NULL);1572 1573	/*1574	 * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is1575	 * used internally by wakeup_source_remove, too (called by1576	 * wakeup_source_unregister), so we cannot use call_rcu1577	 */1578	synchronize_rcu();1579	wakeup_source_unregister(ws);1580}1581 1582static int attach_epitem(struct file *file, struct epitem *epi)1583{1584	struct epitems_head *to_free = NULL;1585	struct hlist_head *head = NULL;1586	struct eventpoll *ep = NULL;1587 1588	if (is_file_epoll(file))1589		ep = file->private_data;1590 1591	if (ep) {1592		head = &ep->refs;1593	} else if (!READ_ONCE(file->f_ep)) {1594allocate:1595		to_free = kmem_cache_zalloc(ephead_cache, GFP_KERNEL);1596		if (!to_free)1597			return -ENOMEM;1598		head = &to_free->epitems;1599	}1600	spin_lock(&file->f_lock);1601	if (!file->f_ep) {1602		if (unlikely(!head)) {1603			spin_unlock(&file->f_lock);1604			goto allocate;1605		}1606		file->f_ep = head;1607		to_free = NULL;1608	}1609	hlist_add_head_rcu(&epi->fllink, file->f_ep);1610	spin_unlock(&file->f_lock);1611	free_ephead(to_free);1612	return 0;1613}1614 1615/*1616 * Must be called with "mtx" held.1617 */1618static HWJS_SUSPENDS int ep_insert(struct eventpoll *ep, const struct epoll_event *event,1619		     struct file *tfile, int fd, int full_check)1620{1621	int error, pwake = 0;1622	__poll_t revents;1623	struct epitem *epi;1624	struct ep_pqueue epq;1625	struct eventpoll *tep = NULL;1626 1627	if (is_file_epoll(tfile))1628		tep = tfile->private_data;1629 1630	lockdep_assert_irqs_enabled();1631 1632	if (unlikely(percpu_counter_compare(&ep->user->epoll_watches,1633					    max_user_watches) >= 0))1634		return -ENOSPC;1635	percpu_counter_inc(&ep->user->epoll_watches);1636 1637	if (!(epi = kmem_cache_zalloc(epi_cache, GFP_KERNEL))) {1638		percpu_counter_dec(&ep->user->epoll_watches);1639		return -ENOMEM;1640	}1641 1642	/* Item initialization follow here ... */1643	INIT_LIST_HEAD(&epi->rdllink);1644	epi->ep = ep;1645	ep_set_ffd(&epi->ffd, tfile, fd);1646	epi->event = *event;1647	epi->next = EP_UNACTIVE_PTR;1648 1649	if (tep)1650		mutex_lock_nested(&tep->mtx, 1);1651	/* Add the current item to the list of active epoll hook for this file */1652	if (unlikely(attach_epitem(tfile, epi) < 0)) {1653		if (tep)1654			mutex_unlock(&tep->mtx);1655		kmem_cache_free(epi_cache, epi);1656		percpu_counter_dec(&ep->user->epoll_watches);1657		return -ENOMEM;1658	}1659 1660	if (full_check && !tep)1661		list_file(tfile);1662 1663	/*1664	 * Add the current item to the RB tree. All RB tree operations are1665	 * protected by "mtx", and ep_insert() is called with "mtx" held.1666	 */1667	ep_rbtree_insert(ep, epi);1668	if (tep)1669		mutex_unlock(&tep->mtx);1670 1671	/*1672	 * ep_remove_safe() calls in the later error paths can't lead to1673	 * ep_free() as the ep file itself still holds an ep reference.1674	 */1675	ep_get(ep);1676 1677	/* now check if we've created too many backpaths */1678	if (unlikely(full_check && reverse_path_check())) {1679		ep_remove_safe(ep, epi);1680		return -EINVAL;1681	}1682 1683	if (epi->event.events & EPOLLWAKEUP) {1684		error = ep_create_wakeup_source(epi);1685		if (error) {1686			ep_remove_safe(ep, epi);1687			return error;1688		}1689	}1690 1691	/* Initialize the poll table using the queue callback */1692	epq.epi = epi;1693	init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);1694 1695	/*1696	 * Attach the item to the poll hooks and get current event bits.1697	 * We can safely use the file* here because its usage count has1698	 * been increased by the caller of this function. Note that after1699	 * this operation completes, the poll callback can start hitting1700	 * the new item.1701	 */1702	revents = ep_item_poll(epi, &epq.pt, 1);1703 1704	/*1705	 * We have to check if something went wrong during the poll wait queue1706	 * install process. Namely an allocation for a wait queue failed due1707	 * high memory pressure.1708	 */1709	if (unlikely(!epq.epi)) {1710		ep_remove_safe(ep, epi);1711		return -ENOMEM;1712	}1713 1714	/* We have to drop the new item inside our item list to keep track of it */1715	write_lock_irq(&ep->lock);1716 1717	/* record NAPI ID of new item if present */1718	ep_set_busy_poll_napi_id(epi);1719 1720	/* If the file is already "ready" we drop it inside the ready list */1721	if (revents && !ep_is_linked(epi)) {1722		list_add_tail(&epi->rdllink, &ep->rdllist);1723		ep_pm_stay_awake(epi);1724 1725		/* Notify waiting tasks that events are available */1726		if (waitqueue_active(&ep->wq))1727			wake_up(&ep->wq);1728		if (waitqueue_active(&ep->poll_wait))1729			pwake++;1730	}1731 1732	write_unlock_irq(&ep->lock);1733 1734	/* We have to call this outside the lock */1735	if (pwake)1736		ep_poll_safewake(ep, NULL, 0);1737 1738	return 0;1739}1740 1741/*1742 * Modify the interest event mask by dropping an event if the new mask1743 * has a match in the current file status. Must be called with "mtx" held.1744 */1745static HWJS_SUSPENDS int ep_modify(struct eventpoll *ep, struct epitem *epi,1746		     const struct epoll_event *event)1747{1748	int pwake = 0;1749	poll_table pt;1750 1751	lockdep_assert_irqs_enabled();1752 1753	init_poll_funcptr(&pt, NULL);1754 1755	/*1756	 * Set the new event interest mask before calling f_op->poll();1757	 * otherwise we might miss an event that happens between the1758	 * f_op->poll() call and the new event set registering.1759	 */1760	epi->event.events = event->events; /* need barrier below */1761	epi->event.data = event->data; /* protected by mtx */1762	if (epi->event.events & EPOLLWAKEUP) {1763		if (!ep_has_wakeup_source(epi))1764			ep_create_wakeup_source(epi);1765	} else if (ep_has_wakeup_source(epi)) {1766		ep_destroy_wakeup_source(epi);1767	}1768 1769	/*1770	 * The following barrier has two effects:1771	 *1772	 * 1) Flush epi changes above to other CPUs.  This ensures1773	 *    we do not miss events from ep_poll_callback if an1774	 *    event occurs immediately after we call f_op->poll().1775	 *    We need this because we did not take ep->lock while1776	 *    changing epi above (but ep_poll_callback does take1777	 *    ep->lock).1778	 *1779	 * 2) We also need to ensure we do not miss _past_ events1780	 *    when calling f_op->poll().  This barrier also1781	 *    pairs with the barrier in wq_has_sleeper (see1782	 *    comments for wq_has_sleeper).1783	 *1784	 * This barrier will now guarantee ep_poll_callback or f_op->poll1785	 * (or both) will notice the readiness of an item.1786	 */1787	smp_mb();1788 1789	/*1790	 * Get current event bits. We can safely use the file* here because1791	 * its usage count has been increased by the caller of this function.1792	 * If the item is "hot" and it is not registered inside the ready1793	 * list, push it inside.1794	 */1795	if (ep_item_poll(epi, &pt, 1)) {1796		write_lock_irq(&ep->lock);1797		if (!ep_is_linked(epi)) {1798			list_add_tail(&epi->rdllink, &ep->rdllist);1799			ep_pm_stay_awake(epi);1800 1801			/* Notify waiting tasks that events are available */1802			if (waitqueue_active(&ep->wq))1803				wake_up(&ep->wq);1804			if (waitqueue_active(&ep->poll_wait))1805				pwake++;1806		}1807		write_unlock_irq(&ep->lock);1808	}1809 1810	/* We have to call this outside the lock */1811	if (pwake)1812		ep_poll_safewake(ep, NULL, 0);1813 1814	return 0;1815}1816 1817static HWJS_SUSPENDS int ep_send_events(struct eventpoll *ep,1818			  struct epoll_event __user *events, int maxevents)1819{1820	struct epitem *epi, *tmp;1821	LIST_HEAD(txlist);1822	poll_table pt;1823	int res = 0;1824 1825	/*1826	 * Always short-circuit for fatal signals to allow threads to make a1827	 * timely exit without the chance of finding more events available and1828	 * fetching repeatedly.1829	 */1830	if (fatal_signal_pending(current))1831		return -EINTR;1832 1833	init_poll_funcptr(&pt, NULL);1834 1835	mutex_lock(&ep->mtx);1836	ep_start_scan(ep, &txlist);1837 1838	/*1839	 * We can loop without lock because we are passed a task private list.1840	 * Items cannot vanish during the loop we are holding ep->mtx.1841	 */1842	list_for_each_entry_safe(epi, tmp, &txlist, rdllink) {1843		struct wakeup_source *ws;1844		__poll_t revents;1845 1846		if (res >= maxevents)1847			break;1848 1849		/*1850		 * Activate ep->ws before deactivating epi->ws to prevent1851		 * triggering auto-suspend here (in case we reactive epi->ws1852		 * below).1853		 *1854		 * This could be rearranged to delay the deactivation of epi->ws1855		 * instead, but then epi->ws would temporarily be out of sync1856		 * with ep_is_linked().1857		 */1858		ws = ep_wakeup_source(epi);1859		if (ws) {1860			if (ws->active)1861				__pm_stay_awake(ep->ws);1862			__pm_relax(ws);1863		}1864 1865		list_del_init(&epi->rdllink);1866 1867		/*1868		 * If the event mask intersect the caller-requested one,1869		 * deliver the event to userspace. Again, we are holding ep->mtx,1870		 * so no operations coming from userspace can change the item.1871		 */1872		revents = ep_item_poll(epi, &pt, 1);1873		if (!revents)1874			continue;1875 1876		events = epoll_put_uevent(revents, epi->event.data, events);1877		if (!events) {1878			list_add(&epi->rdllink, &txlist);1879			ep_pm_stay_awake(epi);1880			if (!res)1881				res = -EFAULT;1882			break;1883		}1884		res++;1885		if (epi->event.events & EPOLLONESHOT)1886			epi->event.events &= EP_PRIVATE_BITS;1887		else if (!(epi->event.events & EPOLLET)) {1888			/*1889			 * If this file has been added with Level1890			 * Trigger mode, we need to insert back inside1891			 * the ready list, so that the next call to1892			 * epoll_wait() will check again the events1893			 * availability. At this point, no one can insert1894			 * into ep->rdllist besides us. The epoll_ctl()1895			 * callers are locked out by1896			 * ep_send_events() holding "mtx" and the1897			 * poll callback will queue them in ep->ovflist.1898			 */1899			list_add_tail(&epi->rdllink, &ep->rdllist);1900			ep_pm_stay_awake(epi);1901		}1902	}1903	ep_done_scan(ep, &txlist);1904	mutex_unlock(&ep->mtx);1905 1906	return res;1907}1908 1909static struct timespec64 *ep_timeout_to_timespec(struct timespec64 *to, long ms)1910{1911	struct timespec64 now;1912 1913	if (ms < 0)1914		return NULL;1915 1916	if (!ms) {1917		to->tv_sec = 0;1918		to->tv_nsec = 0;1919		return to;1920	}1921 1922	to->tv_sec = ms / MSEC_PER_SEC;1923	to->tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC);1924 1925	ktime_get_ts64(&now);1926	*to = timespec64_add_safe(now, *to);1927	return to;1928}1929 1930/*1931 * autoremove_wake_function, but remove even on failure to wake up, because we1932 * know that default_wake_function/ttwu will only fail if the thread is already1933 * woken, and in that case the ep_poll loop will remove the entry anyways, not1934 * try to reuse it.1935 */1936static HWJS_SUSPENDS int ep_autoremove_wake_function(struct wait_queue_entry *wq_entry,1937				       unsigned int mode, int sync, void *key)1938{1939	int ret = default_wake_function(wq_entry, mode, sync, key);1940 1941	/*1942	 * Pairs with list_empty_careful in ep_poll, and ensures future loop1943	 * iterations see the cause of this wakeup.1944	 */1945	list_del_init_careful(&wq_entry->entry);1946	return ret;1947}1948 1949/**1950 * ep_poll - Retrieves ready events, and delivers them to the caller-supplied1951 *           event buffer.1952 *1953 * @ep: Pointer to the eventpoll context.1954 * @events: Pointer to the userspace buffer where the ready events should be1955 *          stored.1956 * @maxevents: Size (in terms of number of events) of the caller event buffer.1957 * @timeout: Maximum timeout for the ready events fetch operation, in1958 *           timespec. If the timeout is zero, the function will not block,1959 *           while if the @timeout ptr is NULL, the function will block1960 *           until at least one event has been retrieved (or an error1961 *           occurred).1962 *1963 * Return: the number of ready events which have been fetched, or an1964 *          error code, in case of error.1965 */1966static HWJS_SUSPENDS int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,1967		   int maxevents, struct timespec64 *timeout)1968{1969	int res, eavail, timed_out = 0;1970	u64 slack = 0;1971	wait_queue_entry_t wait;1972	ktime_t expires, *to = NULL;1973 1974	lockdep_assert_irqs_enabled();1975 1976	if (timeout && (timeout->tv_sec | timeout->tv_nsec)) {1977		slack = select_estimate_accuracy(timeout);1978		to = &expires;1979		*to = timespec64_to_ktime(*timeout);1980	} else if (timeout) {1981		/*1982		 * Avoid the unnecessary trip to the wait queue loop, if the1983		 * caller specified a non blocking operation.1984		 */1985		timed_out = 1;1986	}1987 1988	/*1989	 * This call is racy: We may or may not see events that are being added1990	 * to the ready list under the lock (e.g., in IRQ callbacks). For cases1991	 * with a non-zero timeout, this thread will check the ready list under1992	 * lock and will add to the wait queue.  For cases with a zero1993	 * timeout, the user by definition should not care and will have to1994	 * recheck again.1995	 */1996	eavail = ep_events_available(ep);1997 1998	while (1) {1999		if (eavail) {2000			/*2001			 * Try to transfer events to user space. In case we get2002			 * 0 events and there's still timeout left over, we go2003			 * trying again in search of more luck.2004			 */2005			res = ep_send_events(ep, events, maxevents);2006			if (res)2007				return res;2008		}2009 2010		if (timed_out)2011			return 0;2012 2013		eavail = ep_busy_loop(ep, timed_out);2014		if (eavail)2015			continue;2016 2017		if (signal_pending(current))2018			return -EINTR;2019 2020		/*2021		 * Internally init_wait() uses autoremove_wake_function(),2022		 * thus wait entry is removed from the wait queue on each2023		 * wakeup. Why it is important? In case of several waiters2024		 * each new wakeup will hit the next waiter, giving it the2025		 * chance to harvest new event. Otherwise wakeup can be2026		 * lost. This is also good performance-wise, because on2027		 * normal wakeup path no need to call __remove_wait_queue()2028		 * explicitly, thus ep->lock is not taken, which halts the2029		 * event delivery.2030		 *2031		 * In fact, we now use an even more aggressive function that2032		 * unconditionally removes, because we don't reuse the wait2033		 * entry between loop iterations. This lets us also avoid the2034		 * performance issue if a process is killed, causing all of its2035		 * threads to wake up without being removed normally.2036		 */2037		init_wait(&wait);2038		wait.func = ep_autoremove_wake_function;2039 2040		write_lock_irq(&ep->lock);2041		/*2042		 * Barrierless variant, waitqueue_active() is called under2043		 * the same lock on wakeup ep_poll_callback() side, so it2044		 * is safe to avoid an explicit barrier.2045		 */2046		__set_current_state(TASK_INTERRUPTIBLE);2047 2048		/*2049		 * Do the final check under the lock. ep_start/done_scan()2050		 * plays with two lists (->rdllist and ->ovflist) and there2051		 * is always a race when both lists are empty for short2052		 * period of time although events are pending, so lock is2053		 * important.2054		 */2055		eavail = ep_events_available(ep);2056		if (!eavail)2057			__add_wait_queue_exclusive(&ep->wq, &wait);2058 2059		write_unlock_irq(&ep->lock);2060 2061		if (!eavail)2062			timed_out = !schedule_hrtimeout_range(to, slack,2063							      HRTIMER_MODE_ABS);2064		__set_current_state(TASK_RUNNING);2065 2066		/*2067		 * We were woken up, thus go and try to harvest some events.2068		 * If timed out and still on the wait queue, recheck eavail2069		 * carefully under lock, below.2070		 */2071		eavail = 1;2072 2073		if (!list_empty_careful(&wait.entry)) {2074			write_lock_irq(&ep->lock);2075			/*2076			 * If the thread timed out and is not on the wait queue,2077			 * it means that the thread was woken up after its2078			 * timeout expired before it could reacquire the lock.2079			 * Thus, when wait.entry is empty, it needs to harvest2080			 * events.2081			 */2082			if (timed_out)2083				eavail = list_empty(&wait.entry);2084			__remove_wait_queue(&ep->wq, &wait);2085			write_unlock_irq(&ep->lock);2086		}2087	}2088}2089 2090/**2091 * ep_loop_check_proc - verify that adding an epoll file inside another2092 *                      epoll structure does not violate the constraints, in2093 *                      terms of closed loops, or too deep chains (which can2094 *                      result in excessive stack usage).2095 *2096 * @ep: the &struct eventpoll to be currently checked.2097 * @depth: Current depth of the path being checked.2098 *2099 * Return: %zero if adding the epoll @file inside current epoll2100 *          structure @ep does not violate the constraints, or %-1 otherwise.2101 */2102static HWJS_SUSPENDS int ep_loop_check_proc(struct eventpoll *ep, int depth)2103{2104	int error = 0;2105	struct rb_node *rbp;2106	struct epitem *epi;2107 2108	mutex_lock_nested(&ep->mtx, depth + 1);2109	ep->gen = loop_check_gen;2110	for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {2111		epi = rb_entry(rbp, struct epitem, rbn);2112		if (unlikely(is_file_epoll(epi->ffd.file))) {2113			struct eventpoll *ep_tovisit;2114			ep_tovisit = epi->ffd.file->private_data;2115			if (ep_tovisit->gen == loop_check_gen)2116				continue;2117			if (ep_tovisit == inserting_into || depth > EP_MAX_NESTS)2118				error = -1;2119			else2120				error = ep_loop_check_proc(ep_tovisit, depth + 1);2121			if (error != 0)2122				break;2123		} else {2124			/*2125			 * If we've reached a file that is not associated with2126			 * an ep, then we need to check if the newly added2127			 * links are going to add too many wakeup paths. We do2128			 * this by adding it to the tfile_check_list, if it's2129			 * not already there, and calling reverse_path_check()2130			 * during ep_insert().2131			 */2132			list_file(epi->ffd.file);2133		}2134	}2135	mutex_unlock(&ep->mtx);2136 2137	return error;2138}2139 2140/**2141 * ep_loop_check - Performs a check to verify that adding an epoll file (@to)2142 *                 into another epoll file (represented by @ep) does not create2143 *                 closed loops or too deep chains.2144 *2145 * @ep: Pointer to the epoll we are inserting into.2146 * @to: Pointer to the epoll to be inserted.2147 *2148 * Return: %zero if adding the epoll @to inside the epoll @from2149 * does not violate the constraints, or %-1 otherwise.2150 */2151static HWJS_SUSPENDS int ep_loop_check(struct eventpoll *ep, struct eventpoll *to)2152{2153	inserting_into = ep;2154	return ep_loop_check_proc(to, 0);2155}2156 2157static void clear_tfile_check_list(void)2158{2159	rcu_read_lock();2160	while (tfile_check_list != EP_UNACTIVE_PTR) {2161		struct epitems_head *head = tfile_check_list;2162		tfile_check_list = head->next;2163		unlist_file(head);2164	}2165	rcu_read_unlock();2166}2167 2168/*2169 * Open an eventpoll file descriptor.2170 */2171static HWJS_SUSPENDS int do_epoll_create(int flags)2172{2173	int error, fd;2174	struct eventpoll *ep = NULL;2175	struct file *file;2176 2177	/* Check the EPOLL_* constant for consistency.  */2178	BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);2179 2180	if (flags & ~EPOLL_CLOEXEC)2181		return -EINVAL;2182	/*2183	 * Create the internal data structure ("struct eventpoll").2184	 */2185	error = ep_alloc(&ep);2186	if (error < 0)2187		return error;2188	/*2189	 * Creates all the items needed to setup an eventpoll file. That is,2190	 * a file structure and a free file descriptor.2191	 */2192	fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC));2193	if (fd < 0) {2194		error = fd;2195		goto out_free_ep;2196	}2197	file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,2198				 O_RDWR | (flags & O_CLOEXEC));2199	if (IS_ERR(file)) {2200		error = PTR_ERR(file);2201		goto out_free_fd;2202	}2203	ep->file = file;2204	fd_install(fd, file);2205	return fd;2206 2207out_free_fd:2208	put_unused_fd(fd);2209out_free_ep:2210	ep_clear_and_put(ep);2211	return error;2212}2213 2214SYSCALL_DEFINE1(epoll_create1, int, flags)2215{2216	return do_epoll_create(flags);2217}2218 2219SYSCALL_DEFINE1(epoll_create, int, size)2220{2221	if (size <= 0)2222		return -EINVAL;2223 2224	return do_epoll_create(0);2225}2226 2227#ifdef CONFIG_PM_SLEEP2228static inline void ep_take_care_of_epollwakeup(struct epoll_event *epev)2229{2230	if ((epev->events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))2231		epev->events &= ~EPOLLWAKEUP;2232}2233#else2234static inline void ep_take_care_of_epollwakeup(struct epoll_event *epev)2235{2236	epev->events &= ~EPOLLWAKEUP;2237}2238#endif2239 2240static inline int epoll_mutex_lock(struct mutex *mutex, int depth,2241				   bool nonblock)2242{2243	if (!nonblock) {2244		mutex_lock_nested(mutex, depth);2245		return 0;2246	}2247	if (mutex_trylock(mutex))2248		return 0;2249	return -EAGAIN;2250}2251 2252int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,2253		 bool nonblock)2254{2255	int error;2256	int full_check = 0;2257	struct fd f, tf;2258	struct eventpoll *ep;2259	struct epitem *epi;2260	struct eventpoll *tep = NULL;2261 2262	error = -EBADF;2263	f = fdget(epfd);2264	if (!fd_file(f))2265		goto error_return;2266 2267	/* Get the "struct file *" for the target file */2268	tf = fdget(fd);2269	if (!fd_file(tf))2270		goto error_fput;2271 2272	/* The target file descriptor must support poll */2273	error = -EPERM;2274	if (!file_can_poll(fd_file(tf)))2275		goto error_tgt_fput;2276 2277	/* Check if EPOLLWAKEUP is allowed */2278	if (ep_op_has_event(op))2279		ep_take_care_of_epollwakeup(epds);2280 2281	/*2282	 * We have to check that the file structure underneath the file descriptor2283	 * the user passed to us _is_ an eventpoll file. And also we do not permit2284	 * adding an epoll file descriptor inside itself.2285	 */2286	error = -EINVAL;2287	if (fd_file(f) == fd_file(tf) || !is_file_epoll(fd_file(f)))2288		goto error_tgt_fput;2289 2290	/*2291	 * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,2292	 * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.2293	 * Also, we do not currently supported nested exclusive wakeups.2294	 */2295	if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {2296		if (op == EPOLL_CTL_MOD)2297			goto error_tgt_fput;2298		if (op == EPOLL_CTL_ADD && (is_file_epoll(fd_file(tf)) ||2299				(epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))2300			goto error_tgt_fput;2301	}2302 2303	/*2304	 * At this point it is safe to assume that the "private_data" contains2305	 * our own data structure.2306	 */2307	ep = fd_file(f)->private_data;2308 2309	/*2310	 * When we insert an epoll file descriptor inside another epoll file2311	 * descriptor, there is the chance of creating closed loops, which are2312	 * better be handled here, than in more critical paths. While we are2313	 * checking for loops we also determine the list of files reachable2314	 * and hang them on the tfile_check_list, so we can check that we2315	 * haven't created too many possible wakeup paths.2316	 *2317	 * We do not need to take the global 'epumutex' on EPOLL_CTL_ADD when2318	 * the epoll file descriptor is attaching directly to a wakeup source,2319	 * unless the epoll file descriptor is nested. The purpose of taking the2320	 * 'epnested_mutex' on add is to prevent complex toplogies such as loops and2321	 * deep wakeup paths from forming in parallel through multiple2322	 * EPOLL_CTL_ADD operations.2323	 */2324	error = epoll_mutex_lock(&ep->mtx, 0, nonblock);2325	if (error)2326		goto error_tgt_fput;2327	if (op == EPOLL_CTL_ADD) {2328		if (READ_ONCE(fd_file(f)->f_ep) || ep->gen == loop_check_gen ||2329		    is_file_epoll(fd_file(tf))) {2330			mutex_unlock(&ep->mtx);2331			error = epoll_mutex_lock(&epnested_mutex, 0, nonblock);2332			if (error)2333				goto error_tgt_fput;2334			loop_check_gen++;2335			full_check = 1;2336			if (is_file_epoll(fd_file(tf))) {2337				tep = fd_file(tf)->private_data;2338				error = -ELOOP;2339				if (ep_loop_check(ep, tep) != 0)2340					goto error_tgt_fput;2341			}2342			error = epoll_mutex_lock(&ep->mtx, 0, nonblock);2343			if (error)2344				goto error_tgt_fput;2345		}2346	}2347 2348	/*2349	 * Try to lookup the file inside our RB tree. Since we grabbed "mtx"2350	 * above, we can be sure to be able to use the item looked up by2351	 * ep_find() till we release the mutex.2352	 */2353	epi = ep_find(ep, fd_file(tf), fd);2354 2355	error = -EINVAL;2356	switch (op) {2357	case EPOLL_CTL_ADD:2358		if (!epi) {2359			epds->events |= EPOLLERR | EPOLLHUP;2360			error = ep_insert(ep, epds, fd_file(tf), fd, full_check);2361		} else2362			error = -EEXIST;2363		break;2364	case EPOLL_CTL_DEL:2365		if (epi) {2366			/*2367			 * The eventpoll itself is still alive: the refcount2368			 * can't go to zero here.2369			 */2370			ep_remove_safe(ep, epi);2371			error = 0;2372		} else {2373			error = -ENOENT;2374		}2375		break;2376	case EPOLL_CTL_MOD:2377		if (epi) {2378			if (!(epi->event.events & EPOLLEXCLUSIVE)) {2379				epds->events |= EPOLLERR | EPOLLHUP;2380				error = ep_modify(ep, epi, epds);2381			}2382		} else2383			error = -ENOENT;2384		break;2385	}2386	mutex_unlock(&ep->mtx);2387 2388error_tgt_fput:2389	if (full_check) {2390		clear_tfile_check_list();2391		loop_check_gen++;2392		mutex_unlock(&epnested_mutex);2393	}2394 2395	fdput(tf);2396error_fput:2397	fdput(f);2398error_return:2399 2400	return error;2401}2402 2403/*2404 * The following function implements the controller interface for2405 * the eventpoll file that enables the insertion/removal/change of2406 * file descriptors inside the interest set.2407 */2408SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,2409		struct epoll_event __user *, event)2410{2411	struct epoll_event epds;2412 2413	if (ep_op_has_event(op) &&2414	    copy_from_user(&epds, event, sizeof(struct epoll_event)))2415		return -EFAULT;2416 2417	return do_epoll_ctl(epfd, op, fd, &epds, false);2418}2419 2420/*2421 * Implement the event wait interface for the eventpoll file. It is the kernel2422 * part of the user space epoll_wait(2).2423 */2424static HWJS_SUSPENDS int do_epoll_wait(int epfd, struct epoll_event __user *events,2425			 int maxevents, struct timespec64 *to)2426{2427	int error;2428	struct fd f;2429	struct eventpoll *ep;2430 2431	/* The maximum number of event must be greater than zero */2432	if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)2433		return -EINVAL;2434 2435	/* Verify that the area passed by the user is writeable */2436	if (!access_ok(events, maxevents * sizeof(struct epoll_event)))2437		return -EFAULT;2438 2439	/* Get the "struct file *" for the eventpoll file */2440	f = fdget(epfd);2441	if (!fd_file(f))2442		return -EBADF;2443 2444	/*2445	 * We have to check that the file structure underneath the fd2446	 * the user passed to us _is_ an eventpoll file.2447	 */2448	error = -EINVAL;2449	if (!is_file_epoll(fd_file(f)))2450		goto error_fput;2451 2452	/*2453	 * At this point it is safe to assume that the "private_data" contains2454	 * our own data structure.2455	 */2456	ep = fd_file(f)->private_data;2457 2458	/* Time to fish for events ... */2459	error = ep_poll(ep, events, maxevents, to);2460 2461error_fput:2462	fdput(f);2463	return error;2464}2465 2466SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,2467		int, maxevents, int, timeout)2468{2469	struct timespec64 to;2470 2471	return do_epoll_wait(epfd, events, maxevents,2472			     ep_timeout_to_timespec(&to, timeout));2473}2474 2475/*2476 * Implement the event wait interface for the eventpoll file. It is the kernel2477 * part of the user space epoll_pwait(2).2478 */2479static HWJS_SUSPENDS int do_epoll_pwait(int epfd, struct epoll_event __user *events,2480			  int maxevents, struct timespec64 *to,2481			  const sigset_t __user *sigmask, size_t sigsetsize)2482{2483	int error;2484 2485	/*2486	 * If the caller wants a certain signal mask to be set during the wait,2487	 * we apply it here.2488	 */2489	error = set_user_sigmask(sigmask, sigsetsize);2490	if (error)2491		return error;2492 2493	error = do_epoll_wait(epfd, events, maxevents, to);2494 2495	restore_saved_sigmask_unless(error == -EINTR);2496 2497	return error;2498}2499 2500SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,2501		int, maxevents, int, timeout, const sigset_t __user *, sigmask,2502		size_t, sigsetsize)2503{2504	struct timespec64 to;2505 2506	return do_epoll_pwait(epfd, events, maxevents,2507			      ep_timeout_to_timespec(&to, timeout),2508			      sigmask, sigsetsize);2509}2510 2511SYSCALL_DEFINE6(epoll_pwait2, int, epfd, struct epoll_event __user *, events,2512		int, maxevents, const struct __kernel_timespec __user *, timeout,2513		const sigset_t __user *, sigmask, size_t, sigsetsize)2514{2515	struct timespec64 ts, *to = NULL;2516 2517	if (timeout) {2518		if (get_timespec64(&ts, timeout))2519			return -EFAULT;2520		to = &ts;2521		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))2522			return -EINVAL;2523	}2524 2525	return do_epoll_pwait(epfd, events, maxevents, to,2526			      sigmask, sigsetsize);2527}2528 2529#ifdef CONFIG_COMPAT2530static int do_compat_epoll_pwait(int epfd, struct epoll_event __user *events,2531				 int maxevents, struct timespec64 *timeout,2532				 const compat_sigset_t __user *sigmask,2533				 compat_size_t sigsetsize)2534{2535	long err;2536 2537	/*2538	 * If the caller wants a certain signal mask to be set during the wait,2539	 * we apply it here.2540	 */2541	err = set_compat_user_sigmask(sigmask, sigsetsize);2542	if (err)2543		return err;2544 2545	err = do_epoll_wait(epfd, events, maxevents, timeout);2546 2547	restore_saved_sigmask_unless(err == -EINTR);2548 2549	return err;2550}2551 2552COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,2553		       struct epoll_event __user *, events,2554		       int, maxevents, int, timeout,2555		       const compat_sigset_t __user *, sigmask,2556		       compat_size_t, sigsetsize)2557{2558	struct timespec64 to;2559 2560	return do_compat_epoll_pwait(epfd, events, maxevents,2561				     ep_timeout_to_timespec(&to, timeout),2562				     sigmask, sigsetsize);2563}2564 2565COMPAT_SYSCALL_DEFINE6(epoll_pwait2, int, epfd,2566		       struct epoll_event __user *, events,2567		       int, maxevents,2568		       const struct __kernel_timespec __user *, timeout,2569		       const compat_sigset_t __user *, sigmask,2570		       compat_size_t, sigsetsize)2571{2572	struct timespec64 ts, *to = NULL;2573 2574	if (timeout) {2575		if (get_timespec64(&ts, timeout))2576			return -EFAULT;2577		to = &ts;2578		if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))2579			return -EINVAL;2580	}2581 2582	return do_compat_epoll_pwait(epfd, events, maxevents, to,2583				     sigmask, sigsetsize);2584}2585 2586#endif2587 2588static int __init eventpoll_init(void)2589{2590	struct sysinfo si;2591 2592	si_meminfo(&si);2593	/*2594	 * Allows top 4% of lomem to be allocated for epoll watches (per user).2595	 */2596	max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /2597		EP_ITEM_COST;2598	BUG_ON(max_user_watches < 0);2599 2600	/*2601	 * We can have many thousands of epitems, so prevent this from2602	 * using an extra cache line on 64-bit (and smaller) CPUs2603	 */2604	BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);2605 2606	/* Allocates slab cache used to allocate "struct epitem" items */2607	epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),2608			0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);2609 2610	/* Allocates slab cache used to allocate "struct eppoll_entry" */2611	pwq_cache = kmem_cache_create("eventpoll_pwq",2612		sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);2613	epoll_sysctls_init();2614 2615	ephead_cache = kmem_cache_create("ep_head",2616		sizeof(struct epitems_head), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);2617 2618	return 0;2619}2620fs_initcall(eventpoll_init);2621