| 1 | /* |
| 2 | * fs/eventpoll.c (Efficient event retrieval implementation) |
| 3 | * Copyright (C) 2001,...,2009 Davide Libenzi |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or modify |
| 6 | * it under the terms of the GNU General Public License as published by |
| 7 | * the Free Software Foundation; either version 2 of the License, or |
| 8 | * (at your option) any later version. |
| 9 | * |
| 10 | * Davide Libenzi <davidel@xmailserver.org> |
| 11 | * |
| 12 | */ |
| 13 | |
| 14 | #include <linux/init.h> |
| 15 | #include <linux/kernel.h> |
| 16 | #include <linux/sched.h> |
| 17 | #include <linux/fs.h> |
| 18 | #include <linux/file.h> |
| 19 | #include <linux/signal.h> |
| 20 | #include <linux/errno.h> |
| 21 | #include <linux/mm.h> |
| 22 | #include <linux/slab.h> |
| 23 | #include <linux/poll.h> |
| 24 | #include <linux/string.h> |
| 25 | #include <linux/list.h> |
| 26 | #include <linux/hash.h> |
| 27 | #include <linux/spinlock.h> |
| 28 | #include <linux/syscalls.h> |
| 29 | #include <linux/rbtree.h> |
| 30 | #include <linux/wait.h> |
| 31 | #include <linux/eventpoll.h> |
| 32 | #include <linux/mount.h> |
| 33 | #include <linux/bitops.h> |
| 34 | #include <linux/mutex.h> |
| 35 | #include <linux/anon_inodes.h> |
| 36 | #include <linux/device.h> |
| 37 | #include <linux/freezer.h> |
| 38 | #include <asm/uaccess.h> |
| 39 | #include <asm/io.h> |
| 40 | #include <asm/mman.h> |
| 41 | #include <linux/atomic.h> |
| 42 | #include <linux/proc_fs.h> |
| 43 | #include <linux/seq_file.h> |
| 44 | #include <linux/compat.h> |
| 45 | |
| 46 | /* |
| 47 | * LOCKING: |
| 48 | * There are three level of locking required by epoll : |
| 49 | * |
| 50 | * 1) epmutex (mutex) |
| 51 | * 2) ep->mtx (mutex) |
| 52 | * 3) ep->lock (spinlock) |
| 53 | * |
| 54 | * The acquire order is the one listed above, from 1 to 3. |
| 55 | * We need a spinlock (ep->lock) because we manipulate objects |
| 56 | * from inside the poll callback, that might be triggered from |
| 57 | * a wake_up() that in turn might be called from IRQ context. |
| 58 | * So we can't sleep inside the poll callback and hence we need |
| 59 | * a spinlock. During the event transfer loop (from kernel to |
| 60 | * user space) we could end up sleeping due a copy_to_user(), so |
| 61 | * we need a lock that will allow us to sleep. This lock is a |
| 62 | * mutex (ep->mtx). It is acquired during the event transfer loop, |
| 63 | * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file(). |
| 64 | * Then we also need a global mutex to serialize eventpoll_release_file() |
| 65 | * and ep_free(). |
| 66 | * This mutex is acquired by ep_free() during the epoll file |
| 67 | * cleanup path and it is also acquired by eventpoll_release_file() |
| 68 | * if a file has been pushed inside an epoll set and it is then |
| 69 | * close()d without a previous call to epoll_ctl(EPOLL_CTL_DEL). |
| 70 | * It is also acquired when inserting an epoll fd onto another epoll |
| 71 | * fd. We do this so that we walk the epoll tree and ensure that this |
| 72 | * insertion does not create a cycle of epoll file descriptors, which |
| 73 | * could lead to deadlock. We need a global mutex to prevent two |
| 74 | * simultaneous inserts (A into B and B into A) from racing and |
| 75 | * constructing a cycle without either insert observing that it is |
| 76 | * going to. |
| 77 | * It is necessary to acquire multiple "ep->mtx"es at once in the |
| 78 | * case when one epoll fd is added to another. In this case, we |
| 79 | * always acquire the locks in the order of nesting (i.e. after |
| 80 | * epoll_ctl(e1, EPOLL_CTL_ADD, e2), e1->mtx will always be acquired |
| 81 | * before e2->mtx). Since we disallow cycles of epoll file |
| 82 | * descriptors, this ensures that the mutexes are well-ordered. In |
| 83 | * order to communicate this nesting to lockdep, when walking a tree |
| 84 | * of epoll file descriptors, we use the current recursion depth as |
| 85 | * the lockdep subkey. |
| 86 | * It is possible to drop the "ep->mtx" and to use the global |
| 87 | * mutex "epmutex" (together with "ep->lock") to have it working, |
| 88 | * but having "ep->mtx" will make the interface more scalable. |
| 89 | * Events that require holding "epmutex" are very rare, while for |
| 90 | * normal operations the epoll private "ep->mtx" will guarantee |
| 91 | * a better scalability. |
| 92 | */ |
| 93 | |
| 94 | /* Epoll private bits inside the event mask */ |
| 95 | #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET) |
| 96 | |
| 97 | /* Maximum number of nesting allowed inside epoll sets */ |
| 98 | #define EP_MAX_NESTS 4 |
| 99 | |
| 100 | #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event)) |
| 101 | |
| 102 | #define EP_UNACTIVE_PTR ((void *) -1L) |
| 103 | |
| 104 | #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry)) |
| 105 | |
| 106 | struct epoll_filefd { |
| 107 | struct file *file; |
| 108 | int fd; |
| 109 | } __packed; |
| 110 | |
| 111 | /* |
| 112 | * Structure used to track possible nested calls, for too deep recursions |
| 113 | * and loop cycles. |
| 114 | */ |
| 115 | struct nested_call_node { |
| 116 | struct list_head llink; |
| 117 | void *cookie; |
| 118 | void *ctx; |
| 119 | }; |
| 120 | |
| 121 | /* |
| 122 | * This structure is used as collector for nested calls, to check for |
| 123 | * maximum recursion dept and loop cycles. |
| 124 | */ |
| 125 | struct nested_calls { |
| 126 | struct list_head tasks_call_list; |
| 127 | spinlock_t lock; |
| 128 | }; |
| 129 | |
| 130 | /* |
| 131 | * Each file descriptor added to the eventpoll interface will |
| 132 | * have an entry of this type linked to the "rbr" RB tree. |
| 133 | * Avoid increasing the size of this struct, there can be many thousands |
| 134 | * of these on a server and we do not want this to take another cache line. |
| 135 | */ |
| 136 | struct epitem { |
| 137 | /* RB tree node used to link this structure to the eventpoll RB tree */ |
| 138 | struct rb_node rbn; |
| 139 | |
| 140 | /* List header used to link this structure to the eventpoll ready list */ |
| 141 | struct list_head rdllink; |
| 142 | |
| 143 | /* |
| 144 | * Works together "struct eventpoll"->ovflist in keeping the |
| 145 | * single linked chain of items. |
| 146 | */ |
| 147 | struct epitem *next; |
| 148 | |
| 149 | /* The file descriptor information this item refers to */ |
| 150 | struct epoll_filefd ffd; |
| 151 | |
| 152 | /* Number of active wait queue attached to poll operations */ |
| 153 | int nwait; |
| 154 | |
| 155 | /* List containing poll wait queues */ |
| 156 | struct list_head pwqlist; |
| 157 | |
| 158 | /* The "container" of this item */ |
| 159 | struct eventpoll *ep; |
| 160 | |
| 161 | /* List header used to link this item to the "struct file" items list */ |
| 162 | struct list_head fllink; |
| 163 | |
| 164 | /* wakeup_source used when EPOLLWAKEUP is set */ |
| 165 | struct wakeup_source __rcu *ws; |
| 166 | |
| 167 | /* The structure that describe the interested events and the source fd */ |
| 168 | struct epoll_event event; |
| 169 | }; |
| 170 | |
| 171 | /* |
| 172 | * This structure is stored inside the "private_data" member of the file |
| 173 | * structure and represents the main data structure for the eventpoll |
| 174 | * interface. |
| 175 | */ |
| 176 | struct eventpoll { |
| 177 | /* Protect the access to this structure */ |
| 178 | spinlock_t lock; |
| 179 | |
| 180 | /* |
| 181 | * This mutex is used to ensure that files are not removed |
| 182 | * while epoll is using them. This is held during the event |
| 183 | * collection loop, the file cleanup path, the epoll file exit |
| 184 | * 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 | /* RB tree root used to store monitored fd structs */ |
| 198 | struct rb_root rbr; |
| 199 | |
| 200 | /* |
| 201 | * This is a single linked list that chains all the "struct epitem" that |
| 202 | * happened while transferring ready events to userspace w/out |
| 203 | * holding ->lock. |
| 204 | */ |
| 205 | struct epitem *ovflist; |
| 206 | |
| 207 | /* wakeup_source used when ep_scan_ready_list is running */ |
| 208 | struct wakeup_source *ws; |
| 209 | |
| 210 | /* The user that created the eventpoll descriptor */ |
| 211 | struct user_struct *user; |
| 212 | |
| 213 | struct file *file; |
| 214 | |
| 215 | /* used to optimize loop detection check */ |
| 216 | int visited; |
| 217 | struct list_head visited_list_link; |
| 218 | }; |
| 219 | |
| 220 | /* Wait structure used by the poll hooks */ |
| 221 | struct eppoll_entry { |
| 222 | /* List header used to link this structure to the "struct epitem" */ |
| 223 | struct list_head llink; |
| 224 | |
| 225 | /* The "base" pointer is set to the container "struct epitem" */ |
| 226 | struct epitem *base; |
| 227 | |
| 228 | /* |
| 229 | * Wait queue item that will be linked to the target file wait |
| 230 | * queue head. |
| 231 | */ |
| 232 | wait_queue_t wait; |
| 233 | |
| 234 | /* The wait queue head that linked the "wait" wait queue item */ |
| 235 | wait_queue_head_t *whead; |
| 236 | }; |
| 237 | |
| 238 | /* Wrapper struct used by poll queueing */ |
| 239 | struct ep_pqueue { |
| 240 | poll_table pt; |
| 241 | struct epitem *epi; |
| 242 | }; |
| 243 | |
| 244 | /* Used by the ep_send_events() function as callback private data */ |
| 245 | struct ep_send_events_data { |
| 246 | int maxevents; |
| 247 | struct epoll_event __user *events; |
| 248 | }; |
| 249 | |
| 250 | /* |
| 251 | * Configuration options available inside /proc/sys/fs/epoll/ |
| 252 | */ |
| 253 | /* Maximum number of epoll watched descriptors, per user */ |
| 254 | static long max_user_watches __read_mostly; |
| 255 | |
| 256 | /* |
| 257 | * This mutex is used to serialize ep_free() and eventpoll_release_file(). |
| 258 | */ |
| 259 | static DEFINE_MUTEX(epmutex); |
| 260 | |
| 261 | /* Used to check for epoll file descriptor inclusion loops */ |
| 262 | static struct nested_calls poll_loop_ncalls; |
| 263 | |
| 264 | /* Used for safe wake up implementation */ |
| 265 | static struct nested_calls poll_safewake_ncalls; |
| 266 | |
| 267 | /* Used to call file's f_op->poll() under the nested calls boundaries */ |
| 268 | static struct nested_calls poll_readywalk_ncalls; |
| 269 | |
| 270 | /* Slab cache used to allocate "struct epitem" */ |
| 271 | static struct kmem_cache *epi_cache __read_mostly; |
| 272 | |
| 273 | /* Slab cache used to allocate "struct eppoll_entry" */ |
| 274 | static struct kmem_cache *pwq_cache __read_mostly; |
| 275 | |
| 276 | /* Visited nodes during ep_loop_check(), so we can unset them when we finish */ |
| 277 | static LIST_HEAD(visited_list); |
| 278 | |
| 279 | /* |
| 280 | * List of files with newly added links, where we may need to limit the number |
| 281 | * of emanating paths. Protected by the epmutex. |
| 282 | */ |
| 283 | static LIST_HEAD(tfile_check_list); |
| 284 | |
| 285 | #ifdef CONFIG_SYSCTL |
| 286 | |
| 287 | #include <linux/sysctl.h> |
| 288 | |
| 289 | static long zero; |
| 290 | static long long_max = LONG_MAX; |
| 291 | |
| 292 | ctl_table epoll_table[] = { |
| 293 | { |
| 294 | .procname = "max_user_watches", |
| 295 | .data = &max_user_watches, |
| 296 | .maxlen = sizeof(max_user_watches), |
| 297 | .mode = 0644, |
| 298 | .proc_handler = proc_doulongvec_minmax, |
| 299 | .extra1 = &zero, |
| 300 | .extra2 = &long_max, |
| 301 | }, |
| 302 | { } |
| 303 | }; |
| 304 | #endif /* CONFIG_SYSCTL */ |
| 305 | |
| 306 | static const struct file_operations eventpoll_fops; |
| 307 | |
| 308 | static inline int is_file_epoll(struct file *f) |
| 309 | { |
| 310 | return f->f_op == &eventpoll_fops; |
| 311 | } |
| 312 | |
| 313 | /* Setup the structure that is used as key for the RB tree */ |
| 314 | static inline void ep_set_ffd(struct epoll_filefd *ffd, |
| 315 | struct file *file, int fd) |
| 316 | { |
| 317 | ffd->file = file; |
| 318 | ffd->fd = fd; |
| 319 | } |
| 320 | |
| 321 | /* Compare RB tree keys */ |
| 322 | static inline int ep_cmp_ffd(struct epoll_filefd *p1, |
| 323 | struct epoll_filefd *p2) |
| 324 | { |
| 325 | return (p1->file > p2->file ? +1: |
| 326 | (p1->file < p2->file ? -1 : p1->fd - p2->fd)); |
| 327 | } |
| 328 | |
| 329 | /* Tells us if the item is currently linked */ |
| 330 | static inline int ep_is_linked(struct list_head *p) |
| 331 | { |
| 332 | return !list_empty(p); |
| 333 | } |
| 334 | |
| 335 | static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_t *p) |
| 336 | { |
| 337 | return container_of(p, struct eppoll_entry, wait); |
| 338 | } |
| 339 | |
| 340 | /* Get the "struct epitem" from a wait queue pointer */ |
| 341 | static inline struct epitem *ep_item_from_wait(wait_queue_t *p) |
| 342 | { |
| 343 | return container_of(p, struct eppoll_entry, wait)->base; |
| 344 | } |
| 345 | |
| 346 | /* Get the "struct epitem" from an epoll queue wrapper */ |
| 347 | static inline struct epitem *ep_item_from_epqueue(poll_table *p) |
| 348 | { |
| 349 | return container_of(p, struct ep_pqueue, pt)->epi; |
| 350 | } |
| 351 | |
| 352 | /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */ |
| 353 | static inline int ep_op_has_event(int op) |
| 354 | { |
| 355 | return op != EPOLL_CTL_DEL; |
| 356 | } |
| 357 | |
| 358 | /* Initialize the poll safe wake up structure */ |
| 359 | static void ep_nested_calls_init(struct nested_calls *ncalls) |
| 360 | { |
| 361 | INIT_LIST_HEAD(&ncalls->tasks_call_list); |
| 362 | spin_lock_init(&ncalls->lock); |
| 363 | } |
| 364 | |
| 365 | /** |
| 366 | * ep_events_available - Checks if ready events might be available. |
| 367 | * |
| 368 | * @ep: Pointer to the eventpoll context. |
| 369 | * |
| 370 | * Returns: Returns a value different than zero if ready events are available, |
| 371 | * or zero otherwise. |
| 372 | */ |
| 373 | static inline int ep_events_available(struct eventpoll *ep) |
| 374 | { |
| 375 | return !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR; |
| 376 | } |
| 377 | |
| 378 | /** |
| 379 | * ep_call_nested - Perform a bound (possibly) nested call, by checking |
| 380 | * that the recursion limit is not exceeded, and that |
| 381 | * the same nested call (by the meaning of same cookie) is |
| 382 | * no re-entered. |
| 383 | * |
| 384 | * @ncalls: Pointer to the nested_calls structure to be used for this call. |
| 385 | * @max_nests: Maximum number of allowed nesting calls. |
| 386 | * @nproc: Nested call core function pointer. |
| 387 | * @priv: Opaque data to be passed to the @nproc callback. |
| 388 | * @cookie: Cookie to be used to identify this nested call. |
| 389 | * @ctx: This instance context. |
| 390 | * |
| 391 | * Returns: Returns the code returned by the @nproc callback, or -1 if |
| 392 | * the maximum recursion limit has been exceeded. |
| 393 | */ |
| 394 | static int ep_call_nested(struct nested_calls *ncalls, int max_nests, |
| 395 | int (*nproc)(void *, void *, int), void *priv, |
| 396 | void *cookie, void *ctx) |
| 397 | { |
| 398 | int error, call_nests = 0; |
| 399 | unsigned long flags; |
| 400 | struct list_head *lsthead = &ncalls->tasks_call_list; |
| 401 | struct nested_call_node *tncur; |
| 402 | struct nested_call_node tnode; |
| 403 | |
| 404 | spin_lock_irqsave(&ncalls->lock, flags); |
| 405 | |
| 406 | /* |
| 407 | * Try to see if the current task is already inside this wakeup call. |
| 408 | * We use a list here, since the population inside this set is always |
| 409 | * very much limited. |
| 410 | */ |
| 411 | list_for_each_entry(tncur, lsthead, llink) { |
| 412 | if (tncur->ctx == ctx && |
| 413 | (tncur->cookie == cookie || ++call_nests > max_nests)) { |
| 414 | /* |
| 415 | * Ops ... loop detected or maximum nest level reached. |
| 416 | * We abort this wake by breaking the cycle itself. |
| 417 | */ |
| 418 | error = -1; |
| 419 | goto out_unlock; |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | /* Add the current task and cookie to the list */ |
| 424 | tnode.ctx = ctx; |
| 425 | tnode.cookie = cookie; |
| 426 | list_add(&tnode.llink, lsthead); |
| 427 | |
| 428 | spin_unlock_irqrestore(&ncalls->lock, flags); |
| 429 | |
| 430 | /* Call the nested function */ |
| 431 | error = (*nproc)(priv, cookie, call_nests); |
| 432 | |
| 433 | /* Remove the current task from the list */ |
| 434 | spin_lock_irqsave(&ncalls->lock, flags); |
| 435 | list_del(&tnode.llink); |
| 436 | out_unlock: |
| 437 | spin_unlock_irqrestore(&ncalls->lock, flags); |
| 438 | |
| 439 | return error; |
| 440 | } |
| 441 | |
| 442 | /* |
| 443 | * As described in commit 0ccf831cb lockdep: annotate epoll |
| 444 | * the use of wait queues used by epoll is done in a very controlled |
| 445 | * manner. Wake ups can nest inside each other, but are never done |
| 446 | * with the same locking. For example: |
| 447 | * |
| 448 | * dfd = socket(...); |
| 449 | * efd1 = epoll_create(); |
| 450 | * efd2 = epoll_create(); |
| 451 | * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...); |
| 452 | * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...); |
| 453 | * |
| 454 | * When a packet arrives to the device underneath "dfd", the net code will |
| 455 | * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a |
| 456 | * callback wakeup entry on that queue, and the wake_up() performed by the |
| 457 | * "dfd" net code will end up in ep_poll_callback(). At this point epoll |
| 458 | * (efd1) notices that it may have some event ready, so it needs to wake up |
| 459 | * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake() |
| 460 | * that ends up in another wake_up(), after having checked about the |
| 461 | * recursion constraints. That are, no more than EP_MAX_POLLWAKE_NESTS, to |
| 462 | * avoid stack blasting. |
| 463 | * |
| 464 | * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle |
| 465 | * this special case of epoll. |
| 466 | */ |
| 467 | #ifdef CONFIG_DEBUG_LOCK_ALLOC |
| 468 | static inline void ep_wake_up_nested(wait_queue_head_t *wqueue, |
| 469 | unsigned long events, int subclass) |
| 470 | { |
| 471 | unsigned long flags; |
| 472 | |
| 473 | spin_lock_irqsave_nested(&wqueue->lock, flags, subclass); |
| 474 | wake_up_locked_poll(wqueue, events); |
| 475 | spin_unlock_irqrestore(&wqueue->lock, flags); |
| 476 | } |
| 477 | #else |
| 478 | static inline void ep_wake_up_nested(wait_queue_head_t *wqueue, |
| 479 | unsigned long events, int subclass) |
| 480 | { |
| 481 | wake_up_poll(wqueue, events); |
| 482 | } |
| 483 | #endif |
| 484 | |
| 485 | static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests) |
| 486 | { |
| 487 | ep_wake_up_nested((wait_queue_head_t *) cookie, POLLIN, |
| 488 | 1 + call_nests); |
| 489 | return 0; |
| 490 | } |
| 491 | |
| 492 | /* |
| 493 | * Perform a safe wake up of the poll wait list. The problem is that |
| 494 | * with the new callback'd wake up system, it is possible that the |
| 495 | * poll callback is reentered from inside the call to wake_up() done |
| 496 | * on the poll wait queue head. The rule is that we cannot reenter the |
| 497 | * wake up code from the same task more than EP_MAX_NESTS times, |
| 498 | * and we cannot reenter the same wait queue head at all. This will |
| 499 | * enable to have a hierarchy of epoll file descriptor of no more than |
| 500 | * EP_MAX_NESTS deep. |
| 501 | */ |
| 502 | static void ep_poll_safewake(wait_queue_head_t *wq) |
| 503 | { |
| 504 | int this_cpu = get_cpu(); |
| 505 | |
| 506 | ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS, |
| 507 | ep_poll_wakeup_proc, NULL, wq, (void *) (long) this_cpu); |
| 508 | |
| 509 | put_cpu(); |
| 510 | } |
| 511 | |
| 512 | static void ep_remove_wait_queue(struct eppoll_entry *pwq) |
| 513 | { |
| 514 | wait_queue_head_t *whead; |
| 515 | |
| 516 | rcu_read_lock(); |
| 517 | /* If it is cleared by POLLFREE, it should be rcu-safe */ |
| 518 | whead = rcu_dereference(pwq->whead); |
| 519 | if (whead) |
| 520 | remove_wait_queue(whead, &pwq->wait); |
| 521 | rcu_read_unlock(); |
| 522 | } |
| 523 | |
| 524 | /* |
| 525 | * This function unregisters poll callbacks from the associated file |
| 526 | * descriptor. Must be called with "mtx" held (or "epmutex" if called from |
| 527 | * ep_free). |
| 528 | */ |
| 529 | static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi) |
| 530 | { |
| 531 | struct list_head *lsthead = &epi->pwqlist; |
| 532 | struct eppoll_entry *pwq; |
| 533 | |
| 534 | while (!list_empty(lsthead)) { |
| 535 | pwq = list_first_entry(lsthead, struct eppoll_entry, llink); |
| 536 | |
| 537 | list_del(&pwq->llink); |
| 538 | ep_remove_wait_queue(pwq); |
| 539 | kmem_cache_free(pwq_cache, pwq); |
| 540 | } |
| 541 | } |
| 542 | |
| 543 | /* call only when ep->mtx is held */ |
| 544 | static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi) |
| 545 | { |
| 546 | return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx)); |
| 547 | } |
| 548 | |
| 549 | /* call only when ep->mtx is held */ |
| 550 | static inline void ep_pm_stay_awake(struct epitem *epi) |
| 551 | { |
| 552 | struct wakeup_source *ws = ep_wakeup_source(epi); |
| 553 | |
| 554 | if (ws) |
| 555 | __pm_stay_awake(ws); |
| 556 | } |
| 557 | |
| 558 | static inline bool ep_has_wakeup_source(struct epitem *epi) |
| 559 | { |
| 560 | return rcu_access_pointer(epi->ws) ? true : false; |
| 561 | } |
| 562 | |
| 563 | /* call when ep->mtx cannot be held (ep_poll_callback) */ |
| 564 | static inline void ep_pm_stay_awake_rcu(struct epitem *epi) |
| 565 | { |
| 566 | struct wakeup_source *ws; |
| 567 | |
| 568 | rcu_read_lock(); |
| 569 | ws = rcu_dereference(epi->ws); |
| 570 | if (ws) |
| 571 | __pm_stay_awake(ws); |
| 572 | rcu_read_unlock(); |
| 573 | } |
| 574 | |
| 575 | /** |
| 576 | * ep_scan_ready_list - Scans the ready list in a way that makes possible for |
| 577 | * the scan code, to call f_op->poll(). Also allows for |
| 578 | * O(NumReady) performance. |
| 579 | * |
| 580 | * @ep: Pointer to the epoll private data structure. |
| 581 | * @sproc: Pointer to the scan callback. |
| 582 | * @priv: Private opaque data passed to the @sproc callback. |
| 583 | * @depth: The current depth of recursive f_op->poll calls. |
| 584 | * |
| 585 | * Returns: The same integer error code returned by the @sproc callback. |
| 586 | */ |
| 587 | static int ep_scan_ready_list(struct eventpoll *ep, |
| 588 | int (*sproc)(struct eventpoll *, |
| 589 | struct list_head *, void *), |
| 590 | void *priv, |
| 591 | int depth) |
| 592 | { |
| 593 | int error, pwake = 0; |
| 594 | unsigned long flags; |
| 595 | struct epitem *epi, *nepi; |
| 596 | LIST_HEAD(txlist); |
| 597 | |
| 598 | /* |
| 599 | * We need to lock this because we could be hit by |
| 600 | * eventpoll_release_file() and epoll_ctl(). |
| 601 | */ |
| 602 | mutex_lock_nested(&ep->mtx, depth); |
| 603 | |
| 604 | /* |
| 605 | * Steal the ready list, and re-init the original one to the |
| 606 | * empty list. Also, set ep->ovflist to NULL so that events |
| 607 | * happening while looping w/out locks, are not lost. We cannot |
| 608 | * have the poll callback to queue directly on ep->rdllist, |
| 609 | * because we want the "sproc" callback to be able to do it |
| 610 | * in a lockless way. |
| 611 | */ |
| 612 | spin_lock_irqsave(&ep->lock, flags); |
| 613 | list_splice_init(&ep->rdllist, &txlist); |
| 614 | ep->ovflist = NULL; |
| 615 | spin_unlock_irqrestore(&ep->lock, flags); |
| 616 | |
| 617 | /* |
| 618 | * Now call the callback function. |
| 619 | */ |
| 620 | error = (*sproc)(ep, &txlist, priv); |
| 621 | |
| 622 | spin_lock_irqsave(&ep->lock, flags); |
| 623 | /* |
| 624 | * During the time we spent inside the "sproc" callback, some |
| 625 | * other events might have been queued by the poll callback. |
| 626 | * We re-insert them inside the main ready-list here. |
| 627 | */ |
| 628 | for (nepi = ep->ovflist; (epi = nepi) != NULL; |
| 629 | nepi = epi->next, epi->next = EP_UNACTIVE_PTR) { |
| 630 | /* |
| 631 | * We need to check if the item is already in the list. |
| 632 | * During the "sproc" callback execution time, items are |
| 633 | * queued into ->ovflist but the "txlist" might already |
| 634 | * contain them, and the list_splice() below takes care of them. |
| 635 | */ |
| 636 | if (!ep_is_linked(&epi->rdllink)) { |
| 637 | list_add_tail(&epi->rdllink, &ep->rdllist); |
| 638 | ep_pm_stay_awake(epi); |
| 639 | } |
| 640 | } |
| 641 | /* |
| 642 | * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after |
| 643 | * releasing the lock, events will be queued in the normal way inside |
| 644 | * ep->rdllist. |
| 645 | */ |
| 646 | ep->ovflist = EP_UNACTIVE_PTR; |
| 647 | |
| 648 | /* |
| 649 | * Quickly re-inject items left on "txlist". |
| 650 | */ |
| 651 | list_splice(&txlist, &ep->rdllist); |
| 652 | __pm_relax(ep->ws); |
| 653 | |
| 654 | if (!list_empty(&ep->rdllist)) { |
| 655 | /* |
| 656 | * Wake up (if active) both the eventpoll wait list and |
| 657 | * the ->poll() wait list (delayed after we release the lock). |
| 658 | */ |
| 659 | if (waitqueue_active(&ep->wq)) |
| 660 | wake_up_locked(&ep->wq); |
| 661 | if (waitqueue_active(&ep->poll_wait)) |
| 662 | pwake++; |
| 663 | } |
| 664 | spin_unlock_irqrestore(&ep->lock, flags); |
| 665 | |
| 666 | mutex_unlock(&ep->mtx); |
| 667 | |
| 668 | /* We have to call this outside the lock */ |
| 669 | if (pwake) |
| 670 | ep_poll_safewake(&ep->poll_wait); |
| 671 | |
| 672 | return error; |
| 673 | } |
| 674 | |
| 675 | /* |
| 676 | * Removes a "struct epitem" from the eventpoll RB tree and deallocates |
| 677 | * all the associated resources. Must be called with "mtx" held. |
| 678 | */ |
| 679 | static int ep_remove(struct eventpoll *ep, struct epitem *epi) |
| 680 | { |
| 681 | unsigned long flags; |
| 682 | struct file *file = epi->ffd.file; |
| 683 | |
| 684 | /* |
| 685 | * Removes poll wait queue hooks. We _have_ to do this without holding |
| 686 | * the "ep->lock" otherwise a deadlock might occur. This because of the |
| 687 | * sequence of the lock acquisition. Here we do "ep->lock" then the wait |
| 688 | * queue head lock when unregistering the wait queue. The wakeup callback |
| 689 | * will run by holding the wait queue head lock and will call our callback |
| 690 | * that will try to get "ep->lock". |
| 691 | */ |
| 692 | ep_unregister_pollwait(ep, epi); |
| 693 | |
| 694 | /* Remove the current item from the list of epoll hooks */ |
| 695 | spin_lock(&file->f_lock); |
| 696 | if (ep_is_linked(&epi->fllink)) |
| 697 | list_del_init(&epi->fllink); |
| 698 | spin_unlock(&file->f_lock); |
| 699 | |
| 700 | rb_erase(&epi->rbn, &ep->rbr); |
| 701 | |
| 702 | spin_lock_irqsave(&ep->lock, flags); |
| 703 | if (ep_is_linked(&epi->rdllink)) |
| 704 | list_del_init(&epi->rdllink); |
| 705 | spin_unlock_irqrestore(&ep->lock, flags); |
| 706 | |
| 707 | wakeup_source_unregister(ep_wakeup_source(epi)); |
| 708 | |
| 709 | /* At this point it is safe to free the eventpoll item */ |
| 710 | kmem_cache_free(epi_cache, epi); |
| 711 | |
| 712 | atomic_long_dec(&ep->user->epoll_watches); |
| 713 | |
| 714 | return 0; |
| 715 | } |
| 716 | |
| 717 | static void ep_free(struct eventpoll *ep) |
| 718 | { |
| 719 | struct rb_node *rbp; |
| 720 | struct epitem *epi; |
| 721 | |
| 722 | /* We need to release all tasks waiting for these file */ |
| 723 | if (waitqueue_active(&ep->poll_wait)) |
| 724 | ep_poll_safewake(&ep->poll_wait); |
| 725 | |
| 726 | /* |
| 727 | * We need to lock this because we could be hit by |
| 728 | * eventpoll_release_file() while we're freeing the "struct eventpoll". |
| 729 | * We do not need to hold "ep->mtx" here because the epoll file |
| 730 | * is on the way to be removed and no one has references to it |
| 731 | * anymore. The only hit might come from eventpoll_release_file() but |
| 732 | * holding "epmutex" is sufficient here. |
| 733 | */ |
| 734 | mutex_lock(&epmutex); |
| 735 | |
| 736 | /* |
| 737 | * Walks through the whole tree by unregistering poll callbacks. |
| 738 | */ |
| 739 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { |
| 740 | epi = rb_entry(rbp, struct epitem, rbn); |
| 741 | |
| 742 | ep_unregister_pollwait(ep, epi); |
| 743 | } |
| 744 | |
| 745 | /* |
| 746 | * Walks through the whole tree by freeing each "struct epitem". At this |
| 747 | * point we are sure no poll callbacks will be lingering around, and also by |
| 748 | * holding "epmutex" we can be sure that no file cleanup code will hit |
| 749 | * us during this operation. So we can avoid the lock on "ep->lock". |
| 750 | * We do not need to lock ep->mtx, either, we only do it to prevent |
| 751 | * a lockdep warning. |
| 752 | */ |
| 753 | mutex_lock(&ep->mtx); |
| 754 | while ((rbp = rb_first(&ep->rbr)) != NULL) { |
| 755 | epi = rb_entry(rbp, struct epitem, rbn); |
| 756 | ep_remove(ep, epi); |
| 757 | } |
| 758 | mutex_unlock(&ep->mtx); |
| 759 | |
| 760 | mutex_unlock(&epmutex); |
| 761 | mutex_destroy(&ep->mtx); |
| 762 | free_uid(ep->user); |
| 763 | wakeup_source_unregister(ep->ws); |
| 764 | kfree(ep); |
| 765 | } |
| 766 | |
| 767 | static int ep_eventpoll_release(struct inode *inode, struct file *file) |
| 768 | { |
| 769 | struct eventpoll *ep = file->private_data; |
| 770 | |
| 771 | if (ep) |
| 772 | ep_free(ep); |
| 773 | |
| 774 | return 0; |
| 775 | } |
| 776 | |
| 777 | static inline unsigned int ep_item_poll(struct epitem *epi, poll_table *pt) |
| 778 | { |
| 779 | pt->_key = epi->event.events; |
| 780 | |
| 781 | return epi->ffd.file->f_op->poll(epi->ffd.file, pt) & epi->event.events; |
| 782 | } |
| 783 | |
| 784 | static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head, |
| 785 | void *priv) |
| 786 | { |
| 787 | struct epitem *epi, *tmp; |
| 788 | poll_table pt; |
| 789 | |
| 790 | init_poll_funcptr(&pt, NULL); |
| 791 | |
| 792 | list_for_each_entry_safe(epi, tmp, head, rdllink) { |
| 793 | if (ep_item_poll(epi, &pt)) |
| 794 | return POLLIN | POLLRDNORM; |
| 795 | else { |
| 796 | /* |
| 797 | * Item has been dropped into the ready list by the poll |
| 798 | * callback, but it's not actually ready, as far as |
| 799 | * caller requested events goes. We can remove it here. |
| 800 | */ |
| 801 | __pm_relax(ep_wakeup_source(epi)); |
| 802 | list_del_init(&epi->rdllink); |
| 803 | } |
| 804 | } |
| 805 | |
| 806 | return 0; |
| 807 | } |
| 808 | |
| 809 | static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests) |
| 810 | { |
| 811 | return ep_scan_ready_list(priv, ep_read_events_proc, NULL, call_nests + 1); |
| 812 | } |
| 813 | |
| 814 | static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait) |
| 815 | { |
| 816 | int pollflags; |
| 817 | struct eventpoll *ep = file->private_data; |
| 818 | |
| 819 | /* Insert inside our poll wait queue */ |
| 820 | poll_wait(file, &ep->poll_wait, wait); |
| 821 | |
| 822 | /* |
| 823 | * Proceed to find out if wanted events are really available inside |
| 824 | * the ready list. This need to be done under ep_call_nested() |
| 825 | * supervision, since the call to f_op->poll() done on listed files |
| 826 | * could re-enter here. |
| 827 | */ |
| 828 | pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS, |
| 829 | ep_poll_readyevents_proc, ep, ep, current); |
| 830 | |
| 831 | return pollflags != -1 ? pollflags : 0; |
| 832 | } |
| 833 | |
| 834 | #ifdef CONFIG_PROC_FS |
| 835 | static int ep_show_fdinfo(struct seq_file *m, struct file *f) |
| 836 | { |
| 837 | struct eventpoll *ep = f->private_data; |
| 838 | struct rb_node *rbp; |
| 839 | int ret = 0; |
| 840 | |
| 841 | mutex_lock(&ep->mtx); |
| 842 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { |
| 843 | struct epitem *epi = rb_entry(rbp, struct epitem, rbn); |
| 844 | |
| 845 | ret = seq_printf(m, "tfd: %8d events: %8x data: %16llx\n", |
| 846 | epi->ffd.fd, epi->event.events, |
| 847 | (long long)epi->event.data); |
| 848 | if (ret) |
| 849 | break; |
| 850 | } |
| 851 | mutex_unlock(&ep->mtx); |
| 852 | |
| 853 | return ret; |
| 854 | } |
| 855 | #endif |
| 856 | |
| 857 | /* File callbacks that implement the eventpoll file behaviour */ |
| 858 | static const struct file_operations eventpoll_fops = { |
| 859 | #ifdef CONFIG_PROC_FS |
| 860 | .show_fdinfo = ep_show_fdinfo, |
| 861 | #endif |
| 862 | .release = ep_eventpoll_release, |
| 863 | .poll = ep_eventpoll_poll, |
| 864 | .llseek = noop_llseek, |
| 865 | }; |
| 866 | |
| 867 | /* |
| 868 | * This is called from eventpoll_release() to unlink files from the eventpoll |
| 869 | * interface. We need to have this facility to cleanup correctly files that are |
| 870 | * closed without being removed from the eventpoll interface. |
| 871 | */ |
| 872 | void eventpoll_release_file(struct file *file) |
| 873 | { |
| 874 | struct list_head *lsthead = &file->f_ep_links; |
| 875 | struct eventpoll *ep; |
| 876 | struct epitem *epi; |
| 877 | |
| 878 | /* |
| 879 | * We don't want to get "file->f_lock" because it is not |
| 880 | * necessary. It is not necessary because we're in the "struct file" |
| 881 | * cleanup path, and this means that no one is using this file anymore. |
| 882 | * So, for example, epoll_ctl() cannot hit here since if we reach this |
| 883 | * point, the file counter already went to zero and fget() would fail. |
| 884 | * The only hit might come from ep_free() but by holding the mutex |
| 885 | * will correctly serialize the operation. We do need to acquire |
| 886 | * "ep->mtx" after "epmutex" because ep_remove() requires it when called |
| 887 | * from anywhere but ep_free(). |
| 888 | * |
| 889 | * Besides, ep_remove() acquires the lock, so we can't hold it here. |
| 890 | */ |
| 891 | mutex_lock(&epmutex); |
| 892 | |
| 893 | while (!list_empty(lsthead)) { |
| 894 | epi = list_first_entry(lsthead, struct epitem, fllink); |
| 895 | |
| 896 | ep = epi->ep; |
| 897 | list_del_init(&epi->fllink); |
| 898 | mutex_lock_nested(&ep->mtx, 0); |
| 899 | ep_remove(ep, epi); |
| 900 | mutex_unlock(&ep->mtx); |
| 901 | } |
| 902 | |
| 903 | mutex_unlock(&epmutex); |
| 904 | } |
| 905 | |
| 906 | static int ep_alloc(struct eventpoll **pep) |
| 907 | { |
| 908 | int error; |
| 909 | struct user_struct *user; |
| 910 | struct eventpoll *ep; |
| 911 | |
| 912 | user = get_current_user(); |
| 913 | error = -ENOMEM; |
| 914 | ep = kzalloc(sizeof(*ep), GFP_KERNEL); |
| 915 | if (unlikely(!ep)) |
| 916 | goto free_uid; |
| 917 | |
| 918 | spin_lock_init(&ep->lock); |
| 919 | mutex_init(&ep->mtx); |
| 920 | init_waitqueue_head(&ep->wq); |
| 921 | init_waitqueue_head(&ep->poll_wait); |
| 922 | INIT_LIST_HEAD(&ep->rdllist); |
| 923 | ep->rbr = RB_ROOT; |
| 924 | ep->ovflist = EP_UNACTIVE_PTR; |
| 925 | ep->user = user; |
| 926 | |
| 927 | *pep = ep; |
| 928 | |
| 929 | return 0; |
| 930 | |
| 931 | free_uid: |
| 932 | free_uid(user); |
| 933 | return error; |
| 934 | } |
| 935 | |
| 936 | /* |
| 937 | * Search the file inside the eventpoll tree. The RB tree operations |
| 938 | * are protected by the "mtx" mutex, and ep_find() must be called with |
| 939 | * "mtx" held. |
| 940 | */ |
| 941 | static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd) |
| 942 | { |
| 943 | int kcmp; |
| 944 | struct rb_node *rbp; |
| 945 | struct epitem *epi, *epir = NULL; |
| 946 | struct epoll_filefd ffd; |
| 947 | |
| 948 | ep_set_ffd(&ffd, file, fd); |
| 949 | for (rbp = ep->rbr.rb_node; rbp; ) { |
| 950 | epi = rb_entry(rbp, struct epitem, rbn); |
| 951 | kcmp = ep_cmp_ffd(&ffd, &epi->ffd); |
| 952 | if (kcmp > 0) |
| 953 | rbp = rbp->rb_right; |
| 954 | else if (kcmp < 0) |
| 955 | rbp = rbp->rb_left; |
| 956 | else { |
| 957 | epir = epi; |
| 958 | break; |
| 959 | } |
| 960 | } |
| 961 | |
| 962 | return epir; |
| 963 | } |
| 964 | |
| 965 | /* |
| 966 | * This is the callback that is passed to the wait queue wakeup |
| 967 | * mechanism. It is called by the stored file descriptors when they |
| 968 | * have events to report. |
| 969 | */ |
| 970 | static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key) |
| 971 | { |
| 972 | int pwake = 0; |
| 973 | unsigned long flags; |
| 974 | struct epitem *epi = ep_item_from_wait(wait); |
| 975 | struct eventpoll *ep = epi->ep; |
| 976 | |
| 977 | if ((unsigned long)key & POLLFREE) { |
| 978 | ep_pwq_from_wait(wait)->whead = NULL; |
| 979 | /* |
| 980 | * whead = NULL above can race with ep_remove_wait_queue() |
| 981 | * which can do another remove_wait_queue() after us, so we |
| 982 | * can't use __remove_wait_queue(). whead->lock is held by |
| 983 | * the caller. |
| 984 | */ |
| 985 | list_del_init(&wait->task_list); |
| 986 | } |
| 987 | |
| 988 | spin_lock_irqsave(&ep->lock, flags); |
| 989 | |
| 990 | /* |
| 991 | * If the event mask does not contain any poll(2) event, we consider the |
| 992 | * descriptor to be disabled. This condition is likely the effect of the |
| 993 | * EPOLLONESHOT bit that disables the descriptor when an event is received, |
| 994 | * until the next EPOLL_CTL_MOD will be issued. |
| 995 | */ |
| 996 | if (!(epi->event.events & ~EP_PRIVATE_BITS)) |
| 997 | goto out_unlock; |
| 998 | |
| 999 | /* |
| 1000 | * Check the events coming with the callback. At this stage, not |
| 1001 | * every device reports the events in the "key" parameter of the |
| 1002 | * callback. We need to be able to handle both cases here, hence the |
| 1003 | * test for "key" != NULL before the event match test. |
| 1004 | */ |
| 1005 | if (key && !((unsigned long) key & epi->event.events)) |
| 1006 | goto out_unlock; |
| 1007 | |
| 1008 | /* |
| 1009 | * If we are transferring events to userspace, we can hold no locks |
| 1010 | * (because we're accessing user memory, and because of linux f_op->poll() |
| 1011 | * semantics). All the events that happen during that period of time are |
| 1012 | * chained in ep->ovflist and requeued later on. |
| 1013 | */ |
| 1014 | if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) { |
| 1015 | if (epi->next == EP_UNACTIVE_PTR) { |
| 1016 | epi->next = ep->ovflist; |
| 1017 | ep->ovflist = epi; |
| 1018 | if (epi->ws) { |
| 1019 | /* |
| 1020 | * Activate ep->ws since epi->ws may get |
| 1021 | * deactivated at any time. |
| 1022 | */ |
| 1023 | __pm_stay_awake(ep->ws); |
| 1024 | } |
| 1025 | |
| 1026 | } |
| 1027 | goto out_unlock; |
| 1028 | } |
| 1029 | |
| 1030 | /* If this file is already in the ready list we exit soon */ |
| 1031 | if (!ep_is_linked(&epi->rdllink)) { |
| 1032 | list_add_tail(&epi->rdllink, &ep->rdllist); |
| 1033 | ep_pm_stay_awake_rcu(epi); |
| 1034 | } |
| 1035 | |
| 1036 | /* |
| 1037 | * Wake up ( if active ) both the eventpoll wait list and the ->poll() |
| 1038 | * wait list. |
| 1039 | */ |
| 1040 | if (waitqueue_active(&ep->wq)) |
| 1041 | wake_up_locked(&ep->wq); |
| 1042 | if (waitqueue_active(&ep->poll_wait)) |
| 1043 | pwake++; |
| 1044 | |
| 1045 | out_unlock: |
| 1046 | spin_unlock_irqrestore(&ep->lock, flags); |
| 1047 | |
| 1048 | /* We have to call this outside the lock */ |
| 1049 | if (pwake) |
| 1050 | ep_poll_safewake(&ep->poll_wait); |
| 1051 | |
| 1052 | return 1; |
| 1053 | } |
| 1054 | |
| 1055 | /* |
| 1056 | * This is the callback that is used to add our wait queue to the |
| 1057 | * target file wakeup lists. |
| 1058 | */ |
| 1059 | static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead, |
| 1060 | poll_table *pt) |
| 1061 | { |
| 1062 | struct epitem *epi = ep_item_from_epqueue(pt); |
| 1063 | struct eppoll_entry *pwq; |
| 1064 | |
| 1065 | if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) { |
| 1066 | init_waitqueue_func_entry(&pwq->wait, ep_poll_callback); |
| 1067 | pwq->whead = whead; |
| 1068 | pwq->base = epi; |
| 1069 | add_wait_queue(whead, &pwq->wait); |
| 1070 | list_add_tail(&pwq->llink, &epi->pwqlist); |
| 1071 | epi->nwait++; |
| 1072 | } else { |
| 1073 | /* We have to signal that an error occurred */ |
| 1074 | epi->nwait = -1; |
| 1075 | } |
| 1076 | } |
| 1077 | |
| 1078 | static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi) |
| 1079 | { |
| 1080 | int kcmp; |
| 1081 | struct rb_node **p = &ep->rbr.rb_node, *parent = NULL; |
| 1082 | struct epitem *epic; |
| 1083 | |
| 1084 | while (*p) { |
| 1085 | parent = *p; |
| 1086 | epic = rb_entry(parent, struct epitem, rbn); |
| 1087 | kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd); |
| 1088 | if (kcmp > 0) |
| 1089 | p = &parent->rb_right; |
| 1090 | else |
| 1091 | p = &parent->rb_left; |
| 1092 | } |
| 1093 | rb_link_node(&epi->rbn, parent, p); |
| 1094 | rb_insert_color(&epi->rbn, &ep->rbr); |
| 1095 | } |
| 1096 | |
| 1097 | |
| 1098 | |
| 1099 | #define PATH_ARR_SIZE 5 |
| 1100 | /* |
| 1101 | * These are the number paths of length 1 to 5, that we are allowing to emanate |
| 1102 | * from a single file of interest. For example, we allow 1000 paths of length |
| 1103 | * 1, to emanate from each file of interest. This essentially represents the |
| 1104 | * potential wakeup paths, which need to be limited in order to avoid massive |
| 1105 | * uncontrolled wakeup storms. The common use case should be a single ep which |
| 1106 | * is connected to n file sources. In this case each file source has 1 path |
| 1107 | * of length 1. Thus, the numbers below should be more than sufficient. These |
| 1108 | * path limits are enforced during an EPOLL_CTL_ADD operation, since a modify |
| 1109 | * and delete can't add additional paths. Protected by the epmutex. |
| 1110 | */ |
| 1111 | static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 }; |
| 1112 | static int path_count[PATH_ARR_SIZE]; |
| 1113 | |
| 1114 | static int path_count_inc(int nests) |
| 1115 | { |
| 1116 | /* Allow an arbitrary number of depth 1 paths */ |
| 1117 | if (nests == 0) |
| 1118 | return 0; |
| 1119 | |
| 1120 | if (++path_count[nests] > path_limits[nests]) |
| 1121 | return -1; |
| 1122 | return 0; |
| 1123 | } |
| 1124 | |
| 1125 | static void path_count_init(void) |
| 1126 | { |
| 1127 | int i; |
| 1128 | |
| 1129 | for (i = 0; i < PATH_ARR_SIZE; i++) |
| 1130 | path_count[i] = 0; |
| 1131 | } |
| 1132 | |
| 1133 | static int reverse_path_check_proc(void *priv, void *cookie, int call_nests) |
| 1134 | { |
| 1135 | int error = 0; |
| 1136 | struct file *file = priv; |
| 1137 | struct file *child_file; |
| 1138 | struct epitem *epi; |
| 1139 | |
| 1140 | list_for_each_entry(epi, &file->f_ep_links, fllink) { |
| 1141 | child_file = epi->ep->file; |
| 1142 | if (is_file_epoll(child_file)) { |
| 1143 | if (list_empty(&child_file->f_ep_links)) { |
| 1144 | if (path_count_inc(call_nests)) { |
| 1145 | error = -1; |
| 1146 | break; |
| 1147 | } |
| 1148 | } else { |
| 1149 | error = ep_call_nested(&poll_loop_ncalls, |
| 1150 | EP_MAX_NESTS, |
| 1151 | reverse_path_check_proc, |
| 1152 | child_file, child_file, |
| 1153 | current); |
| 1154 | } |
| 1155 | if (error != 0) |
| 1156 | break; |
| 1157 | } else { |
| 1158 | printk(KERN_ERR "reverse_path_check_proc: " |
| 1159 | "file is not an ep!\n"); |
| 1160 | } |
| 1161 | } |
| 1162 | return error; |
| 1163 | } |
| 1164 | |
| 1165 | /** |
| 1166 | * reverse_path_check - The tfile_check_list is list of file *, which have |
| 1167 | * links that are proposed to be newly added. We need to |
| 1168 | * make sure that those added links don't add too many |
| 1169 | * paths such that we will spend all our time waking up |
| 1170 | * eventpoll objects. |
| 1171 | * |
| 1172 | * Returns: Returns zero if the proposed links don't create too many paths, |
| 1173 | * -1 otherwise. |
| 1174 | */ |
| 1175 | static int reverse_path_check(void) |
| 1176 | { |
| 1177 | int error = 0; |
| 1178 | struct file *current_file; |
| 1179 | |
| 1180 | /* let's call this for all tfiles */ |
| 1181 | list_for_each_entry(current_file, &tfile_check_list, f_tfile_llink) { |
| 1182 | path_count_init(); |
| 1183 | error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS, |
| 1184 | reverse_path_check_proc, current_file, |
| 1185 | current_file, current); |
| 1186 | if (error) |
| 1187 | break; |
| 1188 | } |
| 1189 | return error; |
| 1190 | } |
| 1191 | |
| 1192 | static int ep_create_wakeup_source(struct epitem *epi) |
| 1193 | { |
| 1194 | const char *name; |
| 1195 | struct wakeup_source *ws; |
| 1196 | |
| 1197 | if (!epi->ep->ws) { |
| 1198 | epi->ep->ws = wakeup_source_register("eventpoll"); |
| 1199 | if (!epi->ep->ws) |
| 1200 | return -ENOMEM; |
| 1201 | } |
| 1202 | |
| 1203 | name = epi->ffd.file->f_path.dentry->d_name.name; |
| 1204 | ws = wakeup_source_register(name); |
| 1205 | |
| 1206 | if (!ws) |
| 1207 | return -ENOMEM; |
| 1208 | rcu_assign_pointer(epi->ws, ws); |
| 1209 | |
| 1210 | return 0; |
| 1211 | } |
| 1212 | |
| 1213 | /* rare code path, only used when EPOLL_CTL_MOD removes a wakeup source */ |
| 1214 | static noinline void ep_destroy_wakeup_source(struct epitem *epi) |
| 1215 | { |
| 1216 | struct wakeup_source *ws = ep_wakeup_source(epi); |
| 1217 | |
| 1218 | RCU_INIT_POINTER(epi->ws, NULL); |
| 1219 | |
| 1220 | /* |
| 1221 | * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is |
| 1222 | * used internally by wakeup_source_remove, too (called by |
| 1223 | * wakeup_source_unregister), so we cannot use call_rcu |
| 1224 | */ |
| 1225 | synchronize_rcu(); |
| 1226 | wakeup_source_unregister(ws); |
| 1227 | } |
| 1228 | |
| 1229 | /* |
| 1230 | * Must be called with "mtx" held. |
| 1231 | */ |
| 1232 | static int ep_insert(struct eventpoll *ep, struct epoll_event *event, |
| 1233 | struct file *tfile, int fd) |
| 1234 | { |
| 1235 | int error, revents, pwake = 0; |
| 1236 | unsigned long flags; |
| 1237 | long user_watches; |
| 1238 | struct epitem *epi; |
| 1239 | struct ep_pqueue epq; |
| 1240 | |
| 1241 | user_watches = atomic_long_read(&ep->user->epoll_watches); |
| 1242 | if (unlikely(user_watches >= max_user_watches)) |
| 1243 | return -ENOSPC; |
| 1244 | if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL))) |
| 1245 | return -ENOMEM; |
| 1246 | |
| 1247 | /* Item initialization follow here ... */ |
| 1248 | INIT_LIST_HEAD(&epi->rdllink); |
| 1249 | INIT_LIST_HEAD(&epi->fllink); |
| 1250 | INIT_LIST_HEAD(&epi->pwqlist); |
| 1251 | epi->ep = ep; |
| 1252 | ep_set_ffd(&epi->ffd, tfile, fd); |
| 1253 | epi->event = *event; |
| 1254 | epi->nwait = 0; |
| 1255 | epi->next = EP_UNACTIVE_PTR; |
| 1256 | if (epi->event.events & EPOLLWAKEUP) { |
| 1257 | error = ep_create_wakeup_source(epi); |
| 1258 | if (error) |
| 1259 | goto error_create_wakeup_source; |
| 1260 | } else { |
| 1261 | RCU_INIT_POINTER(epi->ws, NULL); |
| 1262 | } |
| 1263 | |
| 1264 | /* Initialize the poll table using the queue callback */ |
| 1265 | epq.epi = epi; |
| 1266 | init_poll_funcptr(&epq.pt, ep_ptable_queue_proc); |
| 1267 | |
| 1268 | /* |
| 1269 | * Attach the item to the poll hooks and get current event bits. |
| 1270 | * We can safely use the file* here because its usage count has |
| 1271 | * been increased by the caller of this function. Note that after |
| 1272 | * this operation completes, the poll callback can start hitting |
| 1273 | * the new item. |
| 1274 | */ |
| 1275 | revents = ep_item_poll(epi, &epq.pt); |
| 1276 | |
| 1277 | /* |
| 1278 | * We have to check if something went wrong during the poll wait queue |
| 1279 | * install process. Namely an allocation for a wait queue failed due |
| 1280 | * high memory pressure. |
| 1281 | */ |
| 1282 | error = -ENOMEM; |
| 1283 | if (epi->nwait < 0) |
| 1284 | goto error_unregister; |
| 1285 | |
| 1286 | /* Add the current item to the list of active epoll hook for this file */ |
| 1287 | spin_lock(&tfile->f_lock); |
| 1288 | list_add_tail(&epi->fllink, &tfile->f_ep_links); |
| 1289 | spin_unlock(&tfile->f_lock); |
| 1290 | |
| 1291 | /* |
| 1292 | * Add the current item to the RB tree. All RB tree operations are |
| 1293 | * protected by "mtx", and ep_insert() is called with "mtx" held. |
| 1294 | */ |
| 1295 | ep_rbtree_insert(ep, epi); |
| 1296 | |
| 1297 | /* now check if we've created too many backpaths */ |
| 1298 | error = -EINVAL; |
| 1299 | if (reverse_path_check()) |
| 1300 | goto error_remove_epi; |
| 1301 | |
| 1302 | /* We have to drop the new item inside our item list to keep track of it */ |
| 1303 | spin_lock_irqsave(&ep->lock, flags); |
| 1304 | |
| 1305 | /* If the file is already "ready" we drop it inside the ready list */ |
| 1306 | if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) { |
| 1307 | list_add_tail(&epi->rdllink, &ep->rdllist); |
| 1308 | ep_pm_stay_awake(epi); |
| 1309 | |
| 1310 | /* Notify waiting tasks that events are available */ |
| 1311 | if (waitqueue_active(&ep->wq)) |
| 1312 | wake_up_locked(&ep->wq); |
| 1313 | if (waitqueue_active(&ep->poll_wait)) |
| 1314 | pwake++; |
| 1315 | } |
| 1316 | |
| 1317 | spin_unlock_irqrestore(&ep->lock, flags); |
| 1318 | |
| 1319 | atomic_long_inc(&ep->user->epoll_watches); |
| 1320 | |
| 1321 | /* We have to call this outside the lock */ |
| 1322 | if (pwake) |
| 1323 | ep_poll_safewake(&ep->poll_wait); |
| 1324 | |
| 1325 | return 0; |
| 1326 | |
| 1327 | error_remove_epi: |
| 1328 | spin_lock(&tfile->f_lock); |
| 1329 | if (ep_is_linked(&epi->fllink)) |
| 1330 | list_del_init(&epi->fllink); |
| 1331 | spin_unlock(&tfile->f_lock); |
| 1332 | |
| 1333 | rb_erase(&epi->rbn, &ep->rbr); |
| 1334 | |
| 1335 | error_unregister: |
| 1336 | ep_unregister_pollwait(ep, epi); |
| 1337 | |
| 1338 | /* |
| 1339 | * We need to do this because an event could have been arrived on some |
| 1340 | * allocated wait queue. Note that we don't care about the ep->ovflist |
| 1341 | * list, since that is used/cleaned only inside a section bound by "mtx". |
| 1342 | * And ep_insert() is called with "mtx" held. |
| 1343 | */ |
| 1344 | spin_lock_irqsave(&ep->lock, flags); |
| 1345 | if (ep_is_linked(&epi->rdllink)) |
| 1346 | list_del_init(&epi->rdllink); |
| 1347 | spin_unlock_irqrestore(&ep->lock, flags); |
| 1348 | |
| 1349 | wakeup_source_unregister(ep_wakeup_source(epi)); |
| 1350 | |
| 1351 | error_create_wakeup_source: |
| 1352 | kmem_cache_free(epi_cache, epi); |
| 1353 | |
| 1354 | return error; |
| 1355 | } |
| 1356 | |
| 1357 | /* |
| 1358 | * Modify the interest event mask by dropping an event if the new mask |
| 1359 | * has a match in the current file status. Must be called with "mtx" held. |
| 1360 | */ |
| 1361 | static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event) |
| 1362 | { |
| 1363 | int pwake = 0; |
| 1364 | unsigned int revents; |
| 1365 | poll_table pt; |
| 1366 | |
| 1367 | init_poll_funcptr(&pt, NULL); |
| 1368 | |
| 1369 | /* |
| 1370 | * Set the new event interest mask before calling f_op->poll(); |
| 1371 | * otherwise we might miss an event that happens between the |
| 1372 | * f_op->poll() call and the new event set registering. |
| 1373 | */ |
| 1374 | epi->event.events = event->events; /* need barrier below */ |
| 1375 | epi->event.data = event->data; /* protected by mtx */ |
| 1376 | if (epi->event.events & EPOLLWAKEUP) { |
| 1377 | if (!ep_has_wakeup_source(epi)) |
| 1378 | ep_create_wakeup_source(epi); |
| 1379 | } else if (ep_has_wakeup_source(epi)) { |
| 1380 | ep_destroy_wakeup_source(epi); |
| 1381 | } |
| 1382 | |
| 1383 | /* |
| 1384 | * The following barrier has two effects: |
| 1385 | * |
| 1386 | * 1) Flush epi changes above to other CPUs. This ensures |
| 1387 | * we do not miss events from ep_poll_callback if an |
| 1388 | * event occurs immediately after we call f_op->poll(). |
| 1389 | * We need this because we did not take ep->lock while |
| 1390 | * changing epi above (but ep_poll_callback does take |
| 1391 | * ep->lock). |
| 1392 | * |
| 1393 | * 2) We also need to ensure we do not miss _past_ events |
| 1394 | * when calling f_op->poll(). This barrier also |
| 1395 | * pairs with the barrier in wq_has_sleeper (see |
| 1396 | * comments for wq_has_sleeper). |
| 1397 | * |
| 1398 | * This barrier will now guarantee ep_poll_callback or f_op->poll |
| 1399 | * (or both) will notice the readiness of an item. |
| 1400 | */ |
| 1401 | smp_mb(); |
| 1402 | |
| 1403 | /* |
| 1404 | * Get current event bits. We can safely use the file* here because |
| 1405 | * its usage count has been increased by the caller of this function. |
| 1406 | */ |
| 1407 | revents = ep_item_poll(epi, &pt); |
| 1408 | |
| 1409 | /* |
| 1410 | * If the item is "hot" and it is not registered inside the ready |
| 1411 | * list, push it inside. |
| 1412 | */ |
| 1413 | if (revents & event->events) { |
| 1414 | spin_lock_irq(&ep->lock); |
| 1415 | if (!ep_is_linked(&epi->rdllink)) { |
| 1416 | list_add_tail(&epi->rdllink, &ep->rdllist); |
| 1417 | ep_pm_stay_awake(epi); |
| 1418 | |
| 1419 | /* Notify waiting tasks that events are available */ |
| 1420 | if (waitqueue_active(&ep->wq)) |
| 1421 | wake_up_locked(&ep->wq); |
| 1422 | if (waitqueue_active(&ep->poll_wait)) |
| 1423 | pwake++; |
| 1424 | } |
| 1425 | spin_unlock_irq(&ep->lock); |
| 1426 | } |
| 1427 | |
| 1428 | /* We have to call this outside the lock */ |
| 1429 | if (pwake) |
| 1430 | ep_poll_safewake(&ep->poll_wait); |
| 1431 | |
| 1432 | return 0; |
| 1433 | } |
| 1434 | |
| 1435 | static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head, |
| 1436 | void *priv) |
| 1437 | { |
| 1438 | struct ep_send_events_data *esed = priv; |
| 1439 | int eventcnt; |
| 1440 | unsigned int revents; |
| 1441 | struct epitem *epi; |
| 1442 | struct epoll_event __user *uevent; |
| 1443 | struct wakeup_source *ws; |
| 1444 | poll_table pt; |
| 1445 | |
| 1446 | init_poll_funcptr(&pt, NULL); |
| 1447 | |
| 1448 | /* |
| 1449 | * We can loop without lock because we are passed a task private list. |
| 1450 | * Items cannot vanish during the loop because ep_scan_ready_list() is |
| 1451 | * holding "mtx" during this call. |
| 1452 | */ |
| 1453 | for (eventcnt = 0, uevent = esed->events; |
| 1454 | !list_empty(head) && eventcnt < esed->maxevents;) { |
| 1455 | epi = list_first_entry(head, struct epitem, rdllink); |
| 1456 | |
| 1457 | /* |
| 1458 | * Activate ep->ws before deactivating epi->ws to prevent |
| 1459 | * triggering auto-suspend here (in case we reactive epi->ws |
| 1460 | * below). |
| 1461 | * |
| 1462 | * This could be rearranged to delay the deactivation of epi->ws |
| 1463 | * instead, but then epi->ws would temporarily be out of sync |
| 1464 | * with ep_is_linked(). |
| 1465 | */ |
| 1466 | ws = ep_wakeup_source(epi); |
| 1467 | if (ws) { |
| 1468 | if (ws->active) |
| 1469 | __pm_stay_awake(ep->ws); |
| 1470 | __pm_relax(ws); |
| 1471 | } |
| 1472 | |
| 1473 | list_del_init(&epi->rdllink); |
| 1474 | |
| 1475 | revents = ep_item_poll(epi, &pt); |
| 1476 | |
| 1477 | /* |
| 1478 | * If the event mask intersect the caller-requested one, |
| 1479 | * deliver the event to userspace. Again, ep_scan_ready_list() |
| 1480 | * is holding "mtx", so no operations coming from userspace |
| 1481 | * can change the item. |
| 1482 | */ |
| 1483 | if (revents) { |
| 1484 | if (__put_user(revents, &uevent->events) || |
| 1485 | __put_user(epi->event.data, &uevent->data)) { |
| 1486 | list_add(&epi->rdllink, head); |
| 1487 | ep_pm_stay_awake(epi); |
| 1488 | return eventcnt ? eventcnt : -EFAULT; |
| 1489 | } |
| 1490 | eventcnt++; |
| 1491 | uevent++; |
| 1492 | if (epi->event.events & EPOLLONESHOT) |
| 1493 | epi->event.events &= EP_PRIVATE_BITS; |
| 1494 | else if (!(epi->event.events & EPOLLET)) { |
| 1495 | /* |
| 1496 | * If this file has been added with Level |
| 1497 | * Trigger mode, we need to insert back inside |
| 1498 | * the ready list, so that the next call to |
| 1499 | * epoll_wait() will check again the events |
| 1500 | * availability. At this point, no one can insert |
| 1501 | * into ep->rdllist besides us. The epoll_ctl() |
| 1502 | * callers are locked out by |
| 1503 | * ep_scan_ready_list() holding "mtx" and the |
| 1504 | * poll callback will queue them in ep->ovflist. |
| 1505 | */ |
| 1506 | list_add_tail(&epi->rdllink, &ep->rdllist); |
| 1507 | ep_pm_stay_awake(epi); |
| 1508 | } |
| 1509 | } |
| 1510 | } |
| 1511 | |
| 1512 | return eventcnt; |
| 1513 | } |
| 1514 | |
| 1515 | static int ep_send_events(struct eventpoll *ep, |
| 1516 | struct epoll_event __user *events, int maxevents) |
| 1517 | { |
| 1518 | struct ep_send_events_data esed; |
| 1519 | |
| 1520 | esed.maxevents = maxevents; |
| 1521 | esed.events = events; |
| 1522 | |
| 1523 | return ep_scan_ready_list(ep, ep_send_events_proc, &esed, 0); |
| 1524 | } |
| 1525 | |
| 1526 | static inline struct timespec ep_set_mstimeout(long ms) |
| 1527 | { |
| 1528 | struct timespec now, ts = { |
| 1529 | .tv_sec = ms / MSEC_PER_SEC, |
| 1530 | .tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC), |
| 1531 | }; |
| 1532 | |
| 1533 | ktime_get_ts(&now); |
| 1534 | return timespec_add_safe(now, ts); |
| 1535 | } |
| 1536 | |
| 1537 | /** |
| 1538 | * ep_poll - Retrieves ready events, and delivers them to the caller supplied |
| 1539 | * event buffer. |
| 1540 | * |
| 1541 | * @ep: Pointer to the eventpoll context. |
| 1542 | * @events: Pointer to the userspace buffer where the ready events should be |
| 1543 | * stored. |
| 1544 | * @maxevents: Size (in terms of number of events) of the caller event buffer. |
| 1545 | * @timeout: Maximum timeout for the ready events fetch operation, in |
| 1546 | * milliseconds. If the @timeout is zero, the function will not block, |
| 1547 | * while if the @timeout is less than zero, the function will block |
| 1548 | * until at least one event has been retrieved (or an error |
| 1549 | * occurred). |
| 1550 | * |
| 1551 | * Returns: Returns the number of ready events which have been fetched, or an |
| 1552 | * error code, in case of error. |
| 1553 | */ |
| 1554 | static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, |
| 1555 | int maxevents, long timeout) |
| 1556 | { |
| 1557 | int res = 0, eavail, timed_out = 0; |
| 1558 | unsigned long flags; |
| 1559 | long slack = 0; |
| 1560 | wait_queue_t wait; |
| 1561 | ktime_t expires, *to = NULL; |
| 1562 | |
| 1563 | if (timeout > 0) { |
| 1564 | struct timespec end_time = ep_set_mstimeout(timeout); |
| 1565 | |
| 1566 | slack = select_estimate_accuracy(&end_time); |
| 1567 | to = &expires; |
| 1568 | *to = timespec_to_ktime(end_time); |
| 1569 | } else if (timeout == 0) { |
| 1570 | /* |
| 1571 | * Avoid the unnecessary trip to the wait queue loop, if the |
| 1572 | * caller specified a non blocking operation. |
| 1573 | */ |
| 1574 | timed_out = 1; |
| 1575 | spin_lock_irqsave(&ep->lock, flags); |
| 1576 | goto check_events; |
| 1577 | } |
| 1578 | |
| 1579 | fetch_events: |
| 1580 | spin_lock_irqsave(&ep->lock, flags); |
| 1581 | |
| 1582 | if (!ep_events_available(ep)) { |
| 1583 | /* |
| 1584 | * We don't have any available event to return to the caller. |
| 1585 | * We need to sleep here, and we will be wake up by |
| 1586 | * ep_poll_callback() when events will become available. |
| 1587 | */ |
| 1588 | init_waitqueue_entry(&wait, current); |
| 1589 | __add_wait_queue_exclusive(&ep->wq, &wait); |
| 1590 | |
| 1591 | for (;;) { |
| 1592 | /* |
| 1593 | * We don't want to sleep if the ep_poll_callback() sends us |
| 1594 | * a wakeup in between. That's why we set the task state |
| 1595 | * to TASK_INTERRUPTIBLE before doing the checks. |
| 1596 | */ |
| 1597 | set_current_state(TASK_INTERRUPTIBLE); |
| 1598 | if (ep_events_available(ep) || timed_out) |
| 1599 | break; |
| 1600 | if (signal_pending(current)) { |
| 1601 | res = -EINTR; |
| 1602 | break; |
| 1603 | } |
| 1604 | |
| 1605 | spin_unlock_irqrestore(&ep->lock, flags); |
| 1606 | if (!freezable_schedule_hrtimeout_range(to, slack, |
| 1607 | HRTIMER_MODE_ABS)) |
| 1608 | timed_out = 1; |
| 1609 | |
| 1610 | spin_lock_irqsave(&ep->lock, flags); |
| 1611 | } |
| 1612 | __remove_wait_queue(&ep->wq, &wait); |
| 1613 | |
| 1614 | set_current_state(TASK_RUNNING); |
| 1615 | } |
| 1616 | check_events: |
| 1617 | /* Is it worth to try to dig for events ? */ |
| 1618 | eavail = ep_events_available(ep); |
| 1619 | |
| 1620 | spin_unlock_irqrestore(&ep->lock, flags); |
| 1621 | |
| 1622 | /* |
| 1623 | * Try to transfer events to user space. In case we get 0 events and |
| 1624 | * there's still timeout left over, we go trying again in search of |
| 1625 | * more luck. |
| 1626 | */ |
| 1627 | if (!res && eavail && |
| 1628 | !(res = ep_send_events(ep, events, maxevents)) && !timed_out) |
| 1629 | goto fetch_events; |
| 1630 | |
| 1631 | return res; |
| 1632 | } |
| 1633 | |
| 1634 | /** |
| 1635 | * ep_loop_check_proc - Callback function to be passed to the @ep_call_nested() |
| 1636 | * API, to verify that adding an epoll file inside another |
| 1637 | * epoll structure, does not violate the constraints, in |
| 1638 | * terms of closed loops, or too deep chains (which can |
| 1639 | * result in excessive stack usage). |
| 1640 | * |
| 1641 | * @priv: Pointer to the epoll file to be currently checked. |
| 1642 | * @cookie: Original cookie for this call. This is the top-of-the-chain epoll |
| 1643 | * data structure pointer. |
| 1644 | * @call_nests: Current dept of the @ep_call_nested() call stack. |
| 1645 | * |
| 1646 | * Returns: Returns zero if adding the epoll @file inside current epoll |
| 1647 | * structure @ep does not violate the constraints, or -1 otherwise. |
| 1648 | */ |
| 1649 | static int ep_loop_check_proc(void *priv, void *cookie, int call_nests) |
| 1650 | { |
| 1651 | int error = 0; |
| 1652 | struct file *file = priv; |
| 1653 | struct eventpoll *ep = file->private_data; |
| 1654 | struct eventpoll *ep_tovisit; |
| 1655 | struct rb_node *rbp; |
| 1656 | struct epitem *epi; |
| 1657 | |
| 1658 | mutex_lock_nested(&ep->mtx, call_nests + 1); |
| 1659 | ep->visited = 1; |
| 1660 | list_add(&ep->visited_list_link, &visited_list); |
| 1661 | for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { |
| 1662 | epi = rb_entry(rbp, struct epitem, rbn); |
| 1663 | if (unlikely(is_file_epoll(epi->ffd.file))) { |
| 1664 | ep_tovisit = epi->ffd.file->private_data; |
| 1665 | if (ep_tovisit->visited) |
| 1666 | continue; |
| 1667 | error = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS, |
| 1668 | ep_loop_check_proc, epi->ffd.file, |
| 1669 | ep_tovisit, current); |
| 1670 | if (error != 0) |
| 1671 | break; |
| 1672 | } else { |
| 1673 | /* |
| 1674 | * If we've reached a file that is not associated with |
| 1675 | * an ep, then we need to check if the newly added |
| 1676 | * links are going to add too many wakeup paths. We do |
| 1677 | * this by adding it to the tfile_check_list, if it's |
| 1678 | * not already there, and calling reverse_path_check() |
| 1679 | * during ep_insert(). |
| 1680 | */ |
| 1681 | if (list_empty(&epi->ffd.file->f_tfile_llink)) |
| 1682 | list_add(&epi->ffd.file->f_tfile_llink, |
| 1683 | &tfile_check_list); |
| 1684 | } |
| 1685 | } |
| 1686 | mutex_unlock(&ep->mtx); |
| 1687 | |
| 1688 | return error; |
| 1689 | } |
| 1690 | |
| 1691 | /** |
| 1692 | * ep_loop_check - Performs a check to verify that adding an epoll file (@file) |
| 1693 | * another epoll file (represented by @ep) does not create |
| 1694 | * closed loops or too deep chains. |
| 1695 | * |
| 1696 | * @ep: Pointer to the epoll private data structure. |
| 1697 | * @file: Pointer to the epoll file to be checked. |
| 1698 | * |
| 1699 | * Returns: Returns zero if adding the epoll @file inside current epoll |
| 1700 | * structure @ep does not violate the constraints, or -1 otherwise. |
| 1701 | */ |
| 1702 | static int ep_loop_check(struct eventpoll *ep, struct file *file) |
| 1703 | { |
| 1704 | int ret; |
| 1705 | struct eventpoll *ep_cur, *ep_next; |
| 1706 | |
| 1707 | ret = ep_call_nested(&poll_loop_ncalls, EP_MAX_NESTS, |
| 1708 | ep_loop_check_proc, file, ep, current); |
| 1709 | /* clear visited list */ |
| 1710 | list_for_each_entry_safe(ep_cur, ep_next, &visited_list, |
| 1711 | visited_list_link) { |
| 1712 | ep_cur->visited = 0; |
| 1713 | list_del(&ep_cur->visited_list_link); |
| 1714 | } |
| 1715 | return ret; |
| 1716 | } |
| 1717 | |
| 1718 | static void clear_tfile_check_list(void) |
| 1719 | { |
| 1720 | struct file *file; |
| 1721 | |
| 1722 | /* first clear the tfile_check_list */ |
| 1723 | while (!list_empty(&tfile_check_list)) { |
| 1724 | file = list_first_entry(&tfile_check_list, struct file, |
| 1725 | f_tfile_llink); |
| 1726 | list_del_init(&file->f_tfile_llink); |
| 1727 | } |
| 1728 | INIT_LIST_HEAD(&tfile_check_list); |
| 1729 | } |
| 1730 | |
| 1731 | /* |
| 1732 | * Open an eventpoll file descriptor. |
| 1733 | */ |
| 1734 | SYSCALL_DEFINE1(epoll_create1, int, flags) |
| 1735 | { |
| 1736 | int error, fd; |
| 1737 | struct eventpoll *ep = NULL; |
| 1738 | struct file *file; |
| 1739 | |
| 1740 | /* Check the EPOLL_* constant for consistency. */ |
| 1741 | BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC); |
| 1742 | |
| 1743 | if (flags & ~EPOLL_CLOEXEC) |
| 1744 | return -EINVAL; |
| 1745 | /* |
| 1746 | * Create the internal data structure ("struct eventpoll"). |
| 1747 | */ |
| 1748 | error = ep_alloc(&ep); |
| 1749 | if (error < 0) |
| 1750 | return error; |
| 1751 | /* |
| 1752 | * Creates all the items needed to setup an eventpoll file. That is, |
| 1753 | * a file structure and a free file descriptor. |
| 1754 | */ |
| 1755 | fd = get_unused_fd_flags(O_RDWR | (flags & O_CLOEXEC)); |
| 1756 | if (fd < 0) { |
| 1757 | error = fd; |
| 1758 | goto out_free_ep; |
| 1759 | } |
| 1760 | file = anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep, |
| 1761 | O_RDWR | (flags & O_CLOEXEC)); |
| 1762 | if (IS_ERR(file)) { |
| 1763 | error = PTR_ERR(file); |
| 1764 | goto out_free_fd; |
| 1765 | } |
| 1766 | ep->file = file; |
| 1767 | fd_install(fd, file); |
| 1768 | return fd; |
| 1769 | |
| 1770 | out_free_fd: |
| 1771 | put_unused_fd(fd); |
| 1772 | out_free_ep: |
| 1773 | ep_free(ep); |
| 1774 | return error; |
| 1775 | } |
| 1776 | |
| 1777 | SYSCALL_DEFINE1(epoll_create, int, size) |
| 1778 | { |
| 1779 | if (size <= 0) |
| 1780 | return -EINVAL; |
| 1781 | |
| 1782 | return sys_epoll_create1(0); |
| 1783 | } |
| 1784 | |
| 1785 | /* |
| 1786 | * The following function implements the controller interface for |
| 1787 | * the eventpoll file that enables the insertion/removal/change of |
| 1788 | * file descriptors inside the interest set. |
| 1789 | */ |
| 1790 | SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd, |
| 1791 | struct epoll_event __user *, event) |
| 1792 | { |
| 1793 | int error; |
| 1794 | int did_lock_epmutex = 0; |
| 1795 | struct file *file, *tfile; |
| 1796 | struct eventpoll *ep; |
| 1797 | struct epitem *epi; |
| 1798 | struct epoll_event epds; |
| 1799 | |
| 1800 | error = -EFAULT; |
| 1801 | if (ep_op_has_event(op) && |
| 1802 | copy_from_user(&epds, event, sizeof(struct epoll_event))) |
| 1803 | goto error_return; |
| 1804 | |
| 1805 | /* Get the "struct file *" for the eventpoll file */ |
| 1806 | error = -EBADF; |
| 1807 | file = fget(epfd); |
| 1808 | if (!file) |
| 1809 | goto error_return; |
| 1810 | |
| 1811 | /* Get the "struct file *" for the target file */ |
| 1812 | tfile = fget(fd); |
| 1813 | if (!tfile) |
| 1814 | goto error_fput; |
| 1815 | |
| 1816 | /* The target file descriptor must support poll */ |
| 1817 | error = -EPERM; |
| 1818 | if (!tfile->f_op || !tfile->f_op->poll) |
| 1819 | goto error_tgt_fput; |
| 1820 | |
| 1821 | /* Check if EPOLLWAKEUP is allowed */ |
| 1822 | if ((epds.events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND)) |
| 1823 | epds.events &= ~EPOLLWAKEUP; |
| 1824 | |
| 1825 | /* |
| 1826 | * We have to check that the file structure underneath the file descriptor |
| 1827 | * the user passed to us _is_ an eventpoll file. And also we do not permit |
| 1828 | * adding an epoll file descriptor inside itself. |
| 1829 | */ |
| 1830 | error = -EINVAL; |
| 1831 | if (file == tfile || !is_file_epoll(file)) |
| 1832 | goto error_tgt_fput; |
| 1833 | |
| 1834 | /* |
| 1835 | * At this point it is safe to assume that the "private_data" contains |
| 1836 | * our own data structure. |
| 1837 | */ |
| 1838 | ep = file->private_data; |
| 1839 | |
| 1840 | /* |
| 1841 | * When we insert an epoll file descriptor, inside another epoll file |
| 1842 | * descriptor, there is the change of creating closed loops, which are |
| 1843 | * better be handled here, than in more critical paths. While we are |
| 1844 | * checking for loops we also determine the list of files reachable |
| 1845 | * and hang them on the tfile_check_list, so we can check that we |
| 1846 | * haven't created too many possible wakeup paths. |
| 1847 | * |
| 1848 | * We need to hold the epmutex across both ep_insert and ep_remove |
| 1849 | * b/c we want to make sure we are looking at a coherent view of |
| 1850 | * epoll network. |
| 1851 | */ |
| 1852 | if (op == EPOLL_CTL_ADD || op == EPOLL_CTL_DEL) { |
| 1853 | mutex_lock(&epmutex); |
| 1854 | did_lock_epmutex = 1; |
| 1855 | } |
| 1856 | if (op == EPOLL_CTL_ADD) { |
| 1857 | if (is_file_epoll(tfile)) { |
| 1858 | error = -ELOOP; |
| 1859 | if (ep_loop_check(ep, tfile) != 0) { |
| 1860 | clear_tfile_check_list(); |
| 1861 | goto error_tgt_fput; |
| 1862 | } |
| 1863 | } else |
| 1864 | list_add(&tfile->f_tfile_llink, &tfile_check_list); |
| 1865 | } |
| 1866 | |
| 1867 | mutex_lock_nested(&ep->mtx, 0); |
| 1868 | |
| 1869 | /* |
| 1870 | * Try to lookup the file inside our RB tree, Since we grabbed "mtx" |
| 1871 | * above, we can be sure to be able to use the item looked up by |
| 1872 | * ep_find() till we release the mutex. |
| 1873 | */ |
| 1874 | epi = ep_find(ep, tfile, fd); |
| 1875 | |
| 1876 | error = -EINVAL; |
| 1877 | switch (op) { |
| 1878 | case EPOLL_CTL_ADD: |
| 1879 | if (!epi) { |
| 1880 | epds.events |= POLLERR | POLLHUP; |
| 1881 | error = ep_insert(ep, &epds, tfile, fd); |
| 1882 | } else |
| 1883 | error = -EEXIST; |
| 1884 | clear_tfile_check_list(); |
| 1885 | break; |
| 1886 | case EPOLL_CTL_DEL: |
| 1887 | if (epi) |
| 1888 | error = ep_remove(ep, epi); |
| 1889 | else |
| 1890 | error = -ENOENT; |
| 1891 | break; |
| 1892 | case EPOLL_CTL_MOD: |
| 1893 | if (epi) { |
| 1894 | epds.events |= POLLERR | POLLHUP; |
| 1895 | error = ep_modify(ep, epi, &epds); |
| 1896 | } else |
| 1897 | error = -ENOENT; |
| 1898 | break; |
| 1899 | } |
| 1900 | mutex_unlock(&ep->mtx); |
| 1901 | |
| 1902 | error_tgt_fput: |
| 1903 | if (did_lock_epmutex) |
| 1904 | mutex_unlock(&epmutex); |
| 1905 | |
| 1906 | fput(tfile); |
| 1907 | error_fput: |
| 1908 | fput(file); |
| 1909 | error_return: |
| 1910 | |
| 1911 | return error; |
| 1912 | } |
| 1913 | |
| 1914 | /* |
| 1915 | * Implement the event wait interface for the eventpoll file. It is the kernel |
| 1916 | * part of the user space epoll_wait(2). |
| 1917 | */ |
| 1918 | SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events, |
| 1919 | int, maxevents, int, timeout) |
| 1920 | { |
| 1921 | int error; |
| 1922 | struct fd f; |
| 1923 | struct eventpoll *ep; |
| 1924 | |
| 1925 | /* The maximum number of event must be greater than zero */ |
| 1926 | if (maxevents <= 0 || maxevents > EP_MAX_EVENTS) |
| 1927 | return -EINVAL; |
| 1928 | |
| 1929 | /* Verify that the area passed by the user is writeable */ |
| 1930 | if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) |
| 1931 | return -EFAULT; |
| 1932 | |
| 1933 | /* Get the "struct file *" for the eventpoll file */ |
| 1934 | f = fdget(epfd); |
| 1935 | if (!f.file) |
| 1936 | return -EBADF; |
| 1937 | |
| 1938 | /* |
| 1939 | * We have to check that the file structure underneath the fd |
| 1940 | * the user passed to us _is_ an eventpoll file. |
| 1941 | */ |
| 1942 | error = -EINVAL; |
| 1943 | if (!is_file_epoll(f.file)) |
| 1944 | goto error_fput; |
| 1945 | |
| 1946 | /* |
| 1947 | * At this point it is safe to assume that the "private_data" contains |
| 1948 | * our own data structure. |
| 1949 | */ |
| 1950 | ep = f.file->private_data; |
| 1951 | |
| 1952 | /* Time to fish for events ... */ |
| 1953 | error = ep_poll(ep, events, maxevents, timeout); |
| 1954 | |
| 1955 | error_fput: |
| 1956 | fdput(f); |
| 1957 | return error; |
| 1958 | } |
| 1959 | |
| 1960 | /* |
| 1961 | * Implement the event wait interface for the eventpoll file. It is the kernel |
| 1962 | * part of the user space epoll_pwait(2). |
| 1963 | */ |
| 1964 | SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events, |
| 1965 | int, maxevents, int, timeout, const sigset_t __user *, sigmask, |
| 1966 | size_t, sigsetsize) |
| 1967 | { |
| 1968 | int error; |
| 1969 | sigset_t ksigmask, sigsaved; |
| 1970 | |
| 1971 | /* |
| 1972 | * If the caller wants a certain signal mask to be set during the wait, |
| 1973 | * we apply it here. |
| 1974 | */ |
| 1975 | if (sigmask) { |
| 1976 | if (sigsetsize != sizeof(sigset_t)) |
| 1977 | return -EINVAL; |
| 1978 | if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) |
| 1979 | return -EFAULT; |
| 1980 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); |
| 1981 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); |
| 1982 | } |
| 1983 | |
| 1984 | error = sys_epoll_wait(epfd, events, maxevents, timeout); |
| 1985 | |
| 1986 | /* |
| 1987 | * If we changed the signal mask, we need to restore the original one. |
| 1988 | * In case we've got a signal while waiting, we do not restore the |
| 1989 | * signal mask yet, and we allow do_signal() to deliver the signal on |
| 1990 | * the way back to userspace, before the signal mask is restored. |
| 1991 | */ |
| 1992 | if (sigmask) { |
| 1993 | if (error == -EINTR) { |
| 1994 | memcpy(¤t->saved_sigmask, &sigsaved, |
| 1995 | sizeof(sigsaved)); |
| 1996 | set_restore_sigmask(); |
| 1997 | } else |
| 1998 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); |
| 1999 | } |
| 2000 | |
| 2001 | return error; |
| 2002 | } |
| 2003 | |
| 2004 | #ifdef CONFIG_COMPAT |
| 2005 | COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd, |
| 2006 | struct epoll_event __user *, events, |
| 2007 | int, maxevents, int, timeout, |
| 2008 | const compat_sigset_t __user *, sigmask, |
| 2009 | compat_size_t, sigsetsize) |
| 2010 | { |
| 2011 | long err; |
| 2012 | compat_sigset_t csigmask; |
| 2013 | sigset_t ksigmask, sigsaved; |
| 2014 | |
| 2015 | /* |
| 2016 | * If the caller wants a certain signal mask to be set during the wait, |
| 2017 | * we apply it here. |
| 2018 | */ |
| 2019 | if (sigmask) { |
| 2020 | if (sigsetsize != sizeof(compat_sigset_t)) |
| 2021 | return -EINVAL; |
| 2022 | if (copy_from_user(&csigmask, sigmask, sizeof(csigmask))) |
| 2023 | return -EFAULT; |
| 2024 | sigset_from_compat(&ksigmask, &csigmask); |
| 2025 | sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); |
| 2026 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); |
| 2027 | } |
| 2028 | |
| 2029 | err = sys_epoll_wait(epfd, events, maxevents, timeout); |
| 2030 | |
| 2031 | /* |
| 2032 | * If we changed the signal mask, we need to restore the original one. |
| 2033 | * In case we've got a signal while waiting, we do not restore the |
| 2034 | * signal mask yet, and we allow do_signal() to deliver the signal on |
| 2035 | * the way back to userspace, before the signal mask is restored. |
| 2036 | */ |
| 2037 | if (sigmask) { |
| 2038 | if (err == -EINTR) { |
| 2039 | memcpy(¤t->saved_sigmask, &sigsaved, |
| 2040 | sizeof(sigsaved)); |
| 2041 | set_restore_sigmask(); |
| 2042 | } else |
| 2043 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); |
| 2044 | } |
| 2045 | |
| 2046 | return err; |
| 2047 | } |
| 2048 | #endif |
| 2049 | |
| 2050 | static int __init eventpoll_init(void) |
| 2051 | { |
| 2052 | struct sysinfo si; |
| 2053 | |
| 2054 | si_meminfo(&si); |
| 2055 | /* |
| 2056 | * Allows top 4% of lomem to be allocated for epoll watches (per user). |
| 2057 | */ |
| 2058 | max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) / |
| 2059 | EP_ITEM_COST; |
| 2060 | BUG_ON(max_user_watches < 0); |
| 2061 | |
| 2062 | /* |
| 2063 | * Initialize the structure used to perform epoll file descriptor |
| 2064 | * inclusion loops checks. |
| 2065 | */ |
| 2066 | ep_nested_calls_init(&poll_loop_ncalls); |
| 2067 | |
| 2068 | /* Initialize the structure used to perform safe poll wait head wake ups */ |
| 2069 | ep_nested_calls_init(&poll_safewake_ncalls); |
| 2070 | |
| 2071 | /* Initialize the structure used to perform file's f_op->poll() calls */ |
| 2072 | ep_nested_calls_init(&poll_readywalk_ncalls); |
| 2073 | |
| 2074 | /* |
| 2075 | * We can have many thousands of epitems, so prevent this from |
| 2076 | * using an extra cache line on 64-bit (and smaller) CPUs |
| 2077 | */ |
| 2078 | BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128); |
| 2079 | |
| 2080 | /* Allocates slab cache used to allocate "struct epitem" items */ |
| 2081 | epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem), |
| 2082 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); |
| 2083 | |
| 2084 | /* Allocates slab cache used to allocate "struct eppoll_entry" */ |
| 2085 | pwq_cache = kmem_cache_create("eventpoll_pwq", |
| 2086 | sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL); |
| 2087 | |
| 2088 | return 0; |
| 2089 | } |
| 2090 | fs_initcall(eventpoll_init); |