1 /* 2 * fs/eventpoll.c (Efficient event retrieval implementation) 3 * Copyright (C) 2001,...,2009 Davide Libenzi 4 * 5 * This program is free software; you can redistribute it and/or modify 6 * it under the terms of the GNU General Public License as published by 7 * the Free Software Foundation; either version 2 of the License, or 8 * (at your option) any later version. 9 * 10 * Davide Libenzi <davidel@xmailserver.org> 11 * 12 */ 13 14 #include <linux/init.h> 15 #include <linux/kernel.h> 16 #include <linux/sched.h> 17 #include <linux/fs.h> 18 #include <linux/file.h> 19 #include <linux/signal.h> 20 #include <linux/errno.h> 21 #include <linux/mm.h> 22 #include <linux/slab.h> 23 #include <linux/poll.h> 24 #include <linux/string.h> 25 #include <linux/list.h> 26 #include <linux/hash.h> 27 #include <linux/spinlock.h> 28 #include <linux/syscalls.h> 29 #include <linux/rbtree.h> 30 #include <linux/wait.h> 31 #include <linux/eventpoll.h> 32 #include <linux/mount.h> 33 #include <linux/bitops.h> 34 #include <linux/mutex.h> 35 #include <linux/anon_inodes.h> 36 #include <asm/uaccess.h> 37 #include <asm/system.h> 38 #include <asm/io.h> 39 #include <asm/mman.h> 40 #include <asm/atomic.h> 41 42 /* 43 * LOCKING: 44 * There are three level of locking required by epoll : 45 * 46 * 1) epmutex (mutex) 47 * 2) ep->mtx (mutex) 48 * 3) ep->lock (spinlock) 49 * 50 * The acquire order is the one listed above, from 1 to 3. 51 * We need a spinlock (ep->lock) because we manipulate objects 52 * from inside the poll callback, that might be triggered from 53 * a wake_up() that in turn might be called from IRQ context. 54 * So we can't sleep inside the poll callback and hence we need 55 * a spinlock. During the event transfer loop (from kernel to 56 * user space) we could end up sleeping due a copy_to_user(), so 57 * we need a lock that will allow us to sleep. This lock is a 58 * mutex (ep->mtx). It is acquired during the event transfer loop, 59 * during epoll_ctl(EPOLL_CTL_DEL) and during eventpoll_release_file(). 60 * Then we also need a global mutex to serialize eventpoll_release_file() 61 * and ep_free(). 62 * This mutex is acquired by ep_free() during the epoll file 63 * cleanup path and it is also acquired by eventpoll_release_file() 64 * if a file has been pushed inside an epoll set and it is then 65 * close()d without a previous call toepoll_ctl(EPOLL_CTL_DEL). 66 * It is possible to drop the "ep->mtx" and to use the global 67 * mutex "epmutex" (together with "ep->lock") to have it working, 68 * but having "ep->mtx" will make the interface more scalable. 69 * Events that require holding "epmutex" are very rare, while for 70 * normal operations the epoll private "ep->mtx" will guarantee 71 * a better scalability. 72 */ 73 74 /* Epoll private bits inside the event mask */ 75 #define EP_PRIVATE_BITS (EPOLLONESHOT | EPOLLET) 76 77 /* Maximum number of nesting allowed inside epoll sets */ 78 #define EP_MAX_NESTS 4 79 80 /* Maximum msec timeout value storeable in a long int */ 81 #define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ) 82 83 #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event)) 84 85 #define EP_UNACTIVE_PTR ((void *) -1L) 86 87 #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry)) 88 89 struct epoll_filefd { 90 struct file *file; 91 int fd; 92 }; 93 94 /* 95 * Structure used to track possible nested calls, for too deep recursions 96 * and loop cycles. 97 */ 98 struct nested_call_node { 99 struct list_head llink; 100 void *cookie; 101 int cpu; 102 }; 103 104 /* 105 * This structure is used as collector for nested calls, to check for 106 * maximum recursion dept and loop cycles. 107 */ 108 struct nested_calls { 109 struct list_head tasks_call_list; 110 spinlock_t lock; 111 }; 112 113 /* 114 * Each file descriptor added to the eventpoll interface will 115 * have an entry of this type linked to the "rbr" RB tree. 116 */ 117 struct epitem { 118 /* RB tree node used to link this structure to the eventpoll RB tree */ 119 struct rb_node rbn; 120 121 /* List header used to link this structure to the eventpoll ready list */ 122 struct list_head rdllink; 123 124 /* 125 * Works together "struct eventpoll"->ovflist in keeping the 126 * single linked chain of items. 127 */ 128 struct epitem *next; 129 130 /* The file descriptor information this item refers to */ 131 struct epoll_filefd ffd; 132 133 /* Number of active wait queue attached to poll operations */ 134 int nwait; 135 136 /* List containing poll wait queues */ 137 struct list_head pwqlist; 138 139 /* The "container" of this item */ 140 struct eventpoll *ep; 141 142 /* List header used to link this item to the "struct file" items list */ 143 struct list_head fllink; 144 145 /* The structure that describe the interested events and the source fd */ 146 struct epoll_event event; 147 }; 148 149 /* 150 * This structure is stored inside the "private_data" member of the file 151 * structure and rapresent the main data sructure for the eventpoll 152 * interface. 153 */ 154 struct eventpoll { 155 /* Protect the this structure access */ 156 spinlock_t lock; 157 158 /* 159 * This mutex is used to ensure that files are not removed 160 * while epoll is using them. This is held during the event 161 * collection loop, the file cleanup path, the epoll file exit 162 * code and the ctl operations. 163 */ 164 struct mutex mtx; 165 166 /* Wait queue used by sys_epoll_wait() */ 167 wait_queue_head_t wq; 168 169 /* Wait queue used by file->poll() */ 170 wait_queue_head_t poll_wait; 171 172 /* List of ready file descriptors */ 173 struct list_head rdllist; 174 175 /* RB tree root used to store monitored fd structs */ 176 struct rb_root rbr; 177 178 /* 179 * This is a single linked list that chains all the "struct epitem" that 180 * happened while transfering ready events to userspace w/out 181 * holding ->lock. 182 */ 183 struct epitem *ovflist; 184 185 /* The user that created the eventpoll descriptor */ 186 struct user_struct *user; 187 }; 188 189 /* Wait structure used by the poll hooks */ 190 struct eppoll_entry { 191 /* List header used to link this structure to the "struct epitem" */ 192 struct list_head llink; 193 194 /* The "base" pointer is set to the container "struct epitem" */ 195 void *base; 196 197 /* 198 * Wait queue item that will be linked to the target file wait 199 * queue head. 200 */ 201 wait_queue_t wait; 202 203 /* The wait queue head that linked the "wait" wait queue item */ 204 wait_queue_head_t *whead; 205 }; 206 207 /* Wrapper struct used by poll queueing */ 208 struct ep_pqueue { 209 poll_table pt; 210 struct epitem *epi; 211 }; 212 213 /* Used by the ep_send_events() function as callback private data */ 214 struct ep_send_events_data { 215 int maxevents; 216 struct epoll_event __user *events; 217 }; 218 219 /* 220 * Configuration options available inside /proc/sys/fs/epoll/ 221 */ 222 /* Maximum number of epoll watched descriptors, per user */ 223 static int max_user_watches __read_mostly; 224 225 /* 226 * This mutex is used to serialize ep_free() and eventpoll_release_file(). 227 */ 228 static DEFINE_MUTEX(epmutex); 229 230 /* Used for safe wake up implementation */ 231 static struct nested_calls poll_safewake_ncalls; 232 233 /* Used to call file's f_op->poll() under the nested calls boundaries */ 234 static struct nested_calls poll_readywalk_ncalls; 235 236 /* Slab cache used to allocate "struct epitem" */ 237 static struct kmem_cache *epi_cache __read_mostly; 238 239 /* Slab cache used to allocate "struct eppoll_entry" */ 240 static struct kmem_cache *pwq_cache __read_mostly; 241 242 #ifdef CONFIG_SYSCTL 243 244 #include <linux/sysctl.h> 245 246 static int zero; 247 248 ctl_table epoll_table[] = { 249 { 250 .procname = "max_user_watches", 251 .data = &max_user_watches, 252 .maxlen = sizeof(int), 253 .mode = 0644, 254 .proc_handler = &proc_dointvec_minmax, 255 .extra1 = &zero, 256 }, 257 { .ctl_name = 0 } 258 }; 259 #endif /* CONFIG_SYSCTL */ 260 261 262 /* Setup the structure that is used as key for the RB tree */ 263 static inline void ep_set_ffd(struct epoll_filefd *ffd, 264 struct file *file, int fd) 265 { 266 ffd->file = file; 267 ffd->fd = fd; 268 } 269 270 /* Compare RB tree keys */ 271 static inline int ep_cmp_ffd(struct epoll_filefd *p1, 272 struct epoll_filefd *p2) 273 { 274 return (p1->file > p2->file ? +1: 275 (p1->file < p2->file ? -1 : p1->fd - p2->fd)); 276 } 277 278 /* Tells us if the item is currently linked */ 279 static inline int ep_is_linked(struct list_head *p) 280 { 281 return !list_empty(p); 282 } 283 284 /* Get the "struct epitem" from a wait queue pointer */ 285 static inline struct epitem *ep_item_from_wait(wait_queue_t *p) 286 { 287 return container_of(p, struct eppoll_entry, wait)->base; 288 } 289 290 /* Get the "struct epitem" from an epoll queue wrapper */ 291 static inline struct epitem *ep_item_from_epqueue(poll_table *p) 292 { 293 return container_of(p, struct ep_pqueue, pt)->epi; 294 } 295 296 /* Tells if the epoll_ctl(2) operation needs an event copy from userspace */ 297 static inline int ep_op_has_event(int op) 298 { 299 return op != EPOLL_CTL_DEL; 300 } 301 302 /* Initialize the poll safe wake up structure */ 303 static void ep_nested_calls_init(struct nested_calls *ncalls) 304 { 305 INIT_LIST_HEAD(&ncalls->tasks_call_list); 306 spin_lock_init(&ncalls->lock); 307 } 308 309 /** 310 * ep_call_nested - Perform a bound (possibly) nested call, by checking 311 * that the recursion limit is not exceeded, and that 312 * the same nested call (by the meaning of same cookie) is 313 * no re-entered. 314 * 315 * @ncalls: Pointer to the nested_calls structure to be used for this call. 316 * @max_nests: Maximum number of allowed nesting calls. 317 * @nproc: Nested call core function pointer. 318 * @priv: Opaque data to be passed to the @nproc callback. 319 * @cookie: Cookie to be used to identify this nested call. 320 * 321 * Returns: Returns the code returned by the @nproc callback, or -1 if 322 * the maximum recursion limit has been exceeded. 323 */ 324 static int ep_call_nested(struct nested_calls *ncalls, int max_nests, 325 int (*nproc)(void *, void *, int), void *priv, 326 void *cookie) 327 { 328 int error, call_nests = 0; 329 unsigned long flags; 330 int this_cpu = get_cpu(); 331 struct list_head *lsthead = &ncalls->tasks_call_list; 332 struct nested_call_node *tncur; 333 struct nested_call_node tnode; 334 335 spin_lock_irqsave(&ncalls->lock, flags); 336 337 /* 338 * Try to see if the current task is already inside this wakeup call. 339 * We use a list here, since the population inside this set is always 340 * very much limited. 341 */ 342 list_for_each_entry(tncur, lsthead, llink) { 343 if (tncur->cpu == this_cpu && 344 (tncur->cookie == cookie || ++call_nests > max_nests)) { 345 /* 346 * Ops ... loop detected or maximum nest level reached. 347 * We abort this wake by breaking the cycle itself. 348 */ 349 error = -1; 350 goto out_unlock; 351 } 352 } 353 354 /* Add the current task and cookie to the list */ 355 tnode.cpu = this_cpu; 356 tnode.cookie = cookie; 357 list_add(&tnode.llink, lsthead); 358 359 spin_unlock_irqrestore(&ncalls->lock, flags); 360 361 /* Call the nested function */ 362 error = (*nproc)(priv, cookie, call_nests); 363 364 /* Remove the current task from the list */ 365 spin_lock_irqsave(&ncalls->lock, flags); 366 list_del(&tnode.llink); 367 out_unlock: 368 spin_unlock_irqrestore(&ncalls->lock, flags); 369 370 put_cpu(); 371 return error; 372 } 373 374 static int ep_poll_wakeup_proc(void *priv, void *cookie, int call_nests) 375 { 376 wake_up_nested((wait_queue_head_t *) cookie, 1 + call_nests); 377 return 0; 378 } 379 380 /* 381 * Perform a safe wake up of the poll wait list. The problem is that 382 * with the new callback'd wake up system, it is possible that the 383 * poll callback is reentered from inside the call to wake_up() done 384 * on the poll wait queue head. The rule is that we cannot reenter the 385 * wake up code from the same task more than EP_MAX_NESTS times, 386 * and we cannot reenter the same wait queue head at all. This will 387 * enable to have a hierarchy of epoll file descriptor of no more than 388 * EP_MAX_NESTS deep. 389 */ 390 static void ep_poll_safewake(wait_queue_head_t *wq) 391 { 392 ep_call_nested(&poll_safewake_ncalls, EP_MAX_NESTS, 393 ep_poll_wakeup_proc, NULL, wq); 394 } 395 396 /* 397 * This function unregisters poll callbacks from the associated file 398 * descriptor. Must be called with "mtx" held (or "epmutex" if called from 399 * ep_free). 400 */ 401 static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi) 402 { 403 struct list_head *lsthead = &epi->pwqlist; 404 struct eppoll_entry *pwq; 405 406 while (!list_empty(lsthead)) { 407 pwq = list_first_entry(lsthead, struct eppoll_entry, llink); 408 409 list_del(&pwq->llink); 410 remove_wait_queue(pwq->whead, &pwq->wait); 411 kmem_cache_free(pwq_cache, pwq); 412 } 413 } 414 415 /** 416 * ep_scan_ready_list - Scans the ready list in a way that makes possible for 417 * the scan code, to call f_op->poll(). Also allows for 418 * O(NumReady) performance. 419 * 420 * @ep: Pointer to the epoll private data structure. 421 * @sproc: Pointer to the scan callback. 422 * @priv: Private opaque data passed to the @sproc callback. 423 * 424 * Returns: The same integer error code returned by the @sproc callback. 425 */ 426 static int ep_scan_ready_list(struct eventpoll *ep, 427 int (*sproc)(struct eventpoll *, 428 struct list_head *, void *), 429 void *priv) 430 { 431 int error, pwake = 0; 432 unsigned long flags; 433 struct epitem *epi, *nepi; 434 LIST_HEAD(txlist); 435 436 /* 437 * We need to lock this because we could be hit by 438 * eventpoll_release_file() and epoll_ctl(). 439 */ 440 mutex_lock(&ep->mtx); 441 442 /* 443 * Steal the ready list, and re-init the original one to the 444 * empty list. Also, set ep->ovflist to NULL so that events 445 * happening while looping w/out locks, are not lost. We cannot 446 * have the poll callback to queue directly on ep->rdllist, 447 * because we want the "sproc" callback to be able to do it 448 * in a lockless way. 449 */ 450 spin_lock_irqsave(&ep->lock, flags); 451 list_splice_init(&ep->rdllist, &txlist); 452 ep->ovflist = NULL; 453 spin_unlock_irqrestore(&ep->lock, flags); 454 455 /* 456 * Now call the callback function. 457 */ 458 error = (*sproc)(ep, &txlist, priv); 459 460 spin_lock_irqsave(&ep->lock, flags); 461 /* 462 * During the time we spent inside the "sproc" callback, some 463 * other events might have been queued by the poll callback. 464 * We re-insert them inside the main ready-list here. 465 */ 466 for (nepi = ep->ovflist; (epi = nepi) != NULL; 467 nepi = epi->next, epi->next = EP_UNACTIVE_PTR) { 468 /* 469 * We need to check if the item is already in the list. 470 * During the "sproc" callback execution time, items are 471 * queued into ->ovflist but the "txlist" might already 472 * contain them, and the list_splice() below takes care of them. 473 */ 474 if (!ep_is_linked(&epi->rdllink)) 475 list_add_tail(&epi->rdllink, &ep->rdllist); 476 } 477 /* 478 * We need to set back ep->ovflist to EP_UNACTIVE_PTR, so that after 479 * releasing the lock, events will be queued in the normal way inside 480 * ep->rdllist. 481 */ 482 ep->ovflist = EP_UNACTIVE_PTR; 483 484 /* 485 * Quickly re-inject items left on "txlist". 486 */ 487 list_splice(&txlist, &ep->rdllist); 488 489 if (!list_empty(&ep->rdllist)) { 490 /* 491 * Wake up (if active) both the eventpoll wait list and 492 * the ->poll() wait list (delayed after we release the lock). 493 */ 494 if (waitqueue_active(&ep->wq)) 495 wake_up_locked(&ep->wq); 496 if (waitqueue_active(&ep->poll_wait)) 497 pwake++; 498 } 499 spin_unlock_irqrestore(&ep->lock, flags); 500 501 mutex_unlock(&ep->mtx); 502 503 /* We have to call this outside the lock */ 504 if (pwake) 505 ep_poll_safewake(&ep->poll_wait); 506 507 return error; 508 } 509 510 /* 511 * Removes a "struct epitem" from the eventpoll RB tree and deallocates 512 * all the associated resources. Must be called with "mtx" held. 513 */ 514 static int ep_remove(struct eventpoll *ep, struct epitem *epi) 515 { 516 unsigned long flags; 517 struct file *file = epi->ffd.file; 518 519 /* 520 * Removes poll wait queue hooks. We _have_ to do this without holding 521 * the "ep->lock" otherwise a deadlock might occur. This because of the 522 * sequence of the lock acquisition. Here we do "ep->lock" then the wait 523 * queue head lock when unregistering the wait queue. The wakeup callback 524 * will run by holding the wait queue head lock and will call our callback 525 * that will try to get "ep->lock". 526 */ 527 ep_unregister_pollwait(ep, epi); 528 529 /* Remove the current item from the list of epoll hooks */ 530 spin_lock(&file->f_lock); 531 if (ep_is_linked(&epi->fllink)) 532 list_del_init(&epi->fllink); 533 spin_unlock(&file->f_lock); 534 535 rb_erase(&epi->rbn, &ep->rbr); 536 537 spin_lock_irqsave(&ep->lock, flags); 538 if (ep_is_linked(&epi->rdllink)) 539 list_del_init(&epi->rdllink); 540 spin_unlock_irqrestore(&ep->lock, flags); 541 542 /* At this point it is safe to free the eventpoll item */ 543 kmem_cache_free(epi_cache, epi); 544 545 atomic_dec(&ep->user->epoll_watches); 546 547 return 0; 548 } 549 550 static void ep_free(struct eventpoll *ep) 551 { 552 struct rb_node *rbp; 553 struct epitem *epi; 554 555 /* We need to release all tasks waiting for these file */ 556 if (waitqueue_active(&ep->poll_wait)) 557 ep_poll_safewake(&ep->poll_wait); 558 559 /* 560 * We need to lock this because we could be hit by 561 * eventpoll_release_file() while we're freeing the "struct eventpoll". 562 * We do not need to hold "ep->mtx" here because the epoll file 563 * is on the way to be removed and no one has references to it 564 * anymore. The only hit might come from eventpoll_release_file() but 565 * holding "epmutex" is sufficent here. 566 */ 567 mutex_lock(&epmutex); 568 569 /* 570 * Walks through the whole tree by unregistering poll callbacks. 571 */ 572 for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) { 573 epi = rb_entry(rbp, struct epitem, rbn); 574 575 ep_unregister_pollwait(ep, epi); 576 } 577 578 /* 579 * Walks through the whole tree by freeing each "struct epitem". At this 580 * point we are sure no poll callbacks will be lingering around, and also by 581 * holding "epmutex" we can be sure that no file cleanup code will hit 582 * us during this operation. So we can avoid the lock on "ep->lock". 583 */ 584 while ((rbp = rb_first(&ep->rbr)) != NULL) { 585 epi = rb_entry(rbp, struct epitem, rbn); 586 ep_remove(ep, epi); 587 } 588 589 mutex_unlock(&epmutex); 590 mutex_destroy(&ep->mtx); 591 free_uid(ep->user); 592 kfree(ep); 593 } 594 595 static int ep_eventpoll_release(struct inode *inode, struct file *file) 596 { 597 struct eventpoll *ep = file->private_data; 598 599 if (ep) 600 ep_free(ep); 601 602 return 0; 603 } 604 605 static int ep_read_events_proc(struct eventpoll *ep, struct list_head *head, 606 void *priv) 607 { 608 struct epitem *epi, *tmp; 609 610 list_for_each_entry_safe(epi, tmp, head, rdllink) { 611 if (epi->ffd.file->f_op->poll(epi->ffd.file, NULL) & 612 epi->event.events) 613 return POLLIN | POLLRDNORM; 614 else { 615 /* 616 * Item has been dropped into the ready list by the poll 617 * callback, but it's not actually ready, as far as 618 * caller requested events goes. We can remove it here. 619 */ 620 list_del_init(&epi->rdllink); 621 } 622 } 623 624 return 0; 625 } 626 627 static int ep_poll_readyevents_proc(void *priv, void *cookie, int call_nests) 628 { 629 return ep_scan_ready_list(priv, ep_read_events_proc, NULL); 630 } 631 632 static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait) 633 { 634 int pollflags; 635 struct eventpoll *ep = file->private_data; 636 637 /* Insert inside our poll wait queue */ 638 poll_wait(file, &ep->poll_wait, wait); 639 640 /* 641 * Proceed to find out if wanted events are really available inside 642 * the ready list. This need to be done under ep_call_nested() 643 * supervision, since the call to f_op->poll() done on listed files 644 * could re-enter here. 645 */ 646 pollflags = ep_call_nested(&poll_readywalk_ncalls, EP_MAX_NESTS, 647 ep_poll_readyevents_proc, ep, ep); 648 649 return pollflags != -1 ? pollflags : 0; 650 } 651 652 /* File callbacks that implement the eventpoll file behaviour */ 653 static const struct file_operations eventpoll_fops = { 654 .release = ep_eventpoll_release, 655 .poll = ep_eventpoll_poll 656 }; 657 658 /* Fast test to see if the file is an evenpoll file */ 659 static inline int is_file_epoll(struct file *f) 660 { 661 return f->f_op == &eventpoll_fops; 662 } 663 664 /* 665 * This is called from eventpoll_release() to unlink files from the eventpoll 666 * interface. We need to have this facility to cleanup correctly files that are 667 * closed without being removed from the eventpoll interface. 668 */ 669 void eventpoll_release_file(struct file *file) 670 { 671 struct list_head *lsthead = &file->f_ep_links; 672 struct eventpoll *ep; 673 struct epitem *epi; 674 675 /* 676 * We don't want to get "file->f_lock" because it is not 677 * necessary. It is not necessary because we're in the "struct file" 678 * cleanup path, and this means that noone is using this file anymore. 679 * So, for example, epoll_ctl() cannot hit here since if we reach this 680 * point, the file counter already went to zero and fget() would fail. 681 * The only hit might come from ep_free() but by holding the mutex 682 * will correctly serialize the operation. We do need to acquire 683 * "ep->mtx" after "epmutex" because ep_remove() requires it when called 684 * from anywhere but ep_free(). 685 * 686 * Besides, ep_remove() acquires the lock, so we can't hold it here. 687 */ 688 mutex_lock(&epmutex); 689 690 while (!list_empty(lsthead)) { 691 epi = list_first_entry(lsthead, struct epitem, fllink); 692 693 ep = epi->ep; 694 list_del_init(&epi->fllink); 695 mutex_lock(&ep->mtx); 696 ep_remove(ep, epi); 697 mutex_unlock(&ep->mtx); 698 } 699 700 mutex_unlock(&epmutex); 701 } 702 703 static int ep_alloc(struct eventpoll **pep) 704 { 705 int error; 706 struct user_struct *user; 707 struct eventpoll *ep; 708 709 user = get_current_user(); 710 error = -ENOMEM; 711 ep = kzalloc(sizeof(*ep), GFP_KERNEL); 712 if (unlikely(!ep)) 713 goto free_uid; 714 715 spin_lock_init(&ep->lock); 716 mutex_init(&ep->mtx); 717 init_waitqueue_head(&ep->wq); 718 init_waitqueue_head(&ep->poll_wait); 719 INIT_LIST_HEAD(&ep->rdllist); 720 ep->rbr = RB_ROOT; 721 ep->ovflist = EP_UNACTIVE_PTR; 722 ep->user = user; 723 724 *pep = ep; 725 726 return 0; 727 728 free_uid: 729 free_uid(user); 730 return error; 731 } 732 733 /* 734 * Search the file inside the eventpoll tree. The RB tree operations 735 * are protected by the "mtx" mutex, and ep_find() must be called with 736 * "mtx" held. 737 */ 738 static struct epitem *ep_find(struct eventpoll *ep, struct file *file, int fd) 739 { 740 int kcmp; 741 struct rb_node *rbp; 742 struct epitem *epi, *epir = NULL; 743 struct epoll_filefd ffd; 744 745 ep_set_ffd(&ffd, file, fd); 746 for (rbp = ep->rbr.rb_node; rbp; ) { 747 epi = rb_entry(rbp, struct epitem, rbn); 748 kcmp = ep_cmp_ffd(&ffd, &epi->ffd); 749 if (kcmp > 0) 750 rbp = rbp->rb_right; 751 else if (kcmp < 0) 752 rbp = rbp->rb_left; 753 else { 754 epir = epi; 755 break; 756 } 757 } 758 759 return epir; 760 } 761 762 /* 763 * This is the callback that is passed to the wait queue wakeup 764 * machanism. It is called by the stored file descriptors when they 765 * have events to report. 766 */ 767 static int ep_poll_callback(wait_queue_t *wait, unsigned mode, int sync, void *key) 768 { 769 int pwake = 0; 770 unsigned long flags; 771 struct epitem *epi = ep_item_from_wait(wait); 772 struct eventpoll *ep = epi->ep; 773 774 spin_lock_irqsave(&ep->lock, flags); 775 776 /* 777 * If the event mask does not contain any poll(2) event, we consider the 778 * descriptor to be disabled. This condition is likely the effect of the 779 * EPOLLONESHOT bit that disables the descriptor when an event is received, 780 * until the next EPOLL_CTL_MOD will be issued. 781 */ 782 if (!(epi->event.events & ~EP_PRIVATE_BITS)) 783 goto out_unlock; 784 785 /* 786 * If we are trasfering events to userspace, we can hold no locks 787 * (because we're accessing user memory, and because of linux f_op->poll() 788 * semantics). All the events that happens during that period of time are 789 * chained in ep->ovflist and requeued later on. 790 */ 791 if (unlikely(ep->ovflist != EP_UNACTIVE_PTR)) { 792 if (epi->next == EP_UNACTIVE_PTR) { 793 epi->next = ep->ovflist; 794 ep->ovflist = epi; 795 } 796 goto out_unlock; 797 } 798 799 /* If this file is already in the ready list we exit soon */ 800 if (!ep_is_linked(&epi->rdllink)) 801 list_add_tail(&epi->rdllink, &ep->rdllist); 802 803 /* 804 * Wake up ( if active ) both the eventpoll wait list and the ->poll() 805 * wait list. 806 */ 807 if (waitqueue_active(&ep->wq)) 808 wake_up_locked(&ep->wq); 809 if (waitqueue_active(&ep->poll_wait)) 810 pwake++; 811 812 out_unlock: 813 spin_unlock_irqrestore(&ep->lock, flags); 814 815 /* We have to call this outside the lock */ 816 if (pwake) 817 ep_poll_safewake(&ep->poll_wait); 818 819 return 1; 820 } 821 822 /* 823 * This is the callback that is used to add our wait queue to the 824 * target file wakeup lists. 825 */ 826 static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead, 827 poll_table *pt) 828 { 829 struct epitem *epi = ep_item_from_epqueue(pt); 830 struct eppoll_entry *pwq; 831 832 if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) { 833 init_waitqueue_func_entry(&pwq->wait, ep_poll_callback); 834 pwq->whead = whead; 835 pwq->base = epi; 836 add_wait_queue(whead, &pwq->wait); 837 list_add_tail(&pwq->llink, &epi->pwqlist); 838 epi->nwait++; 839 } else { 840 /* We have to signal that an error occurred */ 841 epi->nwait = -1; 842 } 843 } 844 845 static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi) 846 { 847 int kcmp; 848 struct rb_node **p = &ep->rbr.rb_node, *parent = NULL; 849 struct epitem *epic; 850 851 while (*p) { 852 parent = *p; 853 epic = rb_entry(parent, struct epitem, rbn); 854 kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd); 855 if (kcmp > 0) 856 p = &parent->rb_right; 857 else 858 p = &parent->rb_left; 859 } 860 rb_link_node(&epi->rbn, parent, p); 861 rb_insert_color(&epi->rbn, &ep->rbr); 862 } 863 864 /* 865 * Must be called with "mtx" held. 866 */ 867 static int ep_insert(struct eventpoll *ep, struct epoll_event *event, 868 struct file *tfile, int fd) 869 { 870 int error, revents, pwake = 0; 871 unsigned long flags; 872 struct epitem *epi; 873 struct ep_pqueue epq; 874 875 if (unlikely(atomic_read(&ep->user->epoll_watches) >= 876 max_user_watches)) 877 return -ENOSPC; 878 if (!(epi = kmem_cache_alloc(epi_cache, GFP_KERNEL))) 879 return -ENOMEM; 880 881 /* Item initialization follow here ... */ 882 INIT_LIST_HEAD(&epi->rdllink); 883 INIT_LIST_HEAD(&epi->fllink); 884 INIT_LIST_HEAD(&epi->pwqlist); 885 epi->ep = ep; 886 ep_set_ffd(&epi->ffd, tfile, fd); 887 epi->event = *event; 888 epi->nwait = 0; 889 epi->next = EP_UNACTIVE_PTR; 890 891 /* Initialize the poll table using the queue callback */ 892 epq.epi = epi; 893 init_poll_funcptr(&epq.pt, ep_ptable_queue_proc); 894 895 /* 896 * Attach the item to the poll hooks and get current event bits. 897 * We can safely use the file* here because its usage count has 898 * been increased by the caller of this function. Note that after 899 * this operation completes, the poll callback can start hitting 900 * the new item. 901 */ 902 revents = tfile->f_op->poll(tfile, &epq.pt); 903 904 /* 905 * We have to check if something went wrong during the poll wait queue 906 * install process. Namely an allocation for a wait queue failed due 907 * high memory pressure. 908 */ 909 error = -ENOMEM; 910 if (epi->nwait < 0) 911 goto error_unregister; 912 913 /* Add the current item to the list of active epoll hook for this file */ 914 spin_lock(&tfile->f_lock); 915 list_add_tail(&epi->fllink, &tfile->f_ep_links); 916 spin_unlock(&tfile->f_lock); 917 918 /* 919 * Add the current item to the RB tree. All RB tree operations are 920 * protected by "mtx", and ep_insert() is called with "mtx" held. 921 */ 922 ep_rbtree_insert(ep, epi); 923 924 /* We have to drop the new item inside our item list to keep track of it */ 925 spin_lock_irqsave(&ep->lock, flags); 926 927 /* If the file is already "ready" we drop it inside the ready list */ 928 if ((revents & event->events) && !ep_is_linked(&epi->rdllink)) { 929 list_add_tail(&epi->rdllink, &ep->rdllist); 930 931 /* Notify waiting tasks that events are available */ 932 if (waitqueue_active(&ep->wq)) 933 wake_up_locked(&ep->wq); 934 if (waitqueue_active(&ep->poll_wait)) 935 pwake++; 936 } 937 938 spin_unlock_irqrestore(&ep->lock, flags); 939 940 atomic_inc(&ep->user->epoll_watches); 941 942 /* We have to call this outside the lock */ 943 if (pwake) 944 ep_poll_safewake(&ep->poll_wait); 945 946 return 0; 947 948 error_unregister: 949 ep_unregister_pollwait(ep, epi); 950 951 /* 952 * We need to do this because an event could have been arrived on some 953 * allocated wait queue. Note that we don't care about the ep->ovflist 954 * list, since that is used/cleaned only inside a section bound by "mtx". 955 * And ep_insert() is called with "mtx" held. 956 */ 957 spin_lock_irqsave(&ep->lock, flags); 958 if (ep_is_linked(&epi->rdllink)) 959 list_del_init(&epi->rdllink); 960 spin_unlock_irqrestore(&ep->lock, flags); 961 962 kmem_cache_free(epi_cache, epi); 963 964 return error; 965 } 966 967 /* 968 * Modify the interest event mask by dropping an event if the new mask 969 * has a match in the current file status. Must be called with "mtx" held. 970 */ 971 static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_event *event) 972 { 973 int pwake = 0; 974 unsigned int revents; 975 976 /* 977 * Set the new event interest mask before calling f_op->poll(); 978 * otherwise we might miss an event that happens between the 979 * f_op->poll() call and the new event set registering. 980 */ 981 epi->event.events = event->events; 982 epi->event.data = event->data; /* protected by mtx */ 983 984 /* 985 * Get current event bits. We can safely use the file* here because 986 * its usage count has been increased by the caller of this function. 987 */ 988 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL); 989 990 /* 991 * If the item is "hot" and it is not registered inside the ready 992 * list, push it inside. 993 */ 994 if (revents & event->events) { 995 spin_lock_irq(&ep->lock); 996 if (!ep_is_linked(&epi->rdllink)) { 997 list_add_tail(&epi->rdllink, &ep->rdllist); 998 999 /* Notify waiting tasks that events are available */ 1000 if (waitqueue_active(&ep->wq)) 1001 wake_up_locked(&ep->wq); 1002 if (waitqueue_active(&ep->poll_wait)) 1003 pwake++; 1004 } 1005 spin_unlock_irq(&ep->lock); 1006 } 1007 1008 /* We have to call this outside the lock */ 1009 if (pwake) 1010 ep_poll_safewake(&ep->poll_wait); 1011 1012 return 0; 1013 } 1014 1015 static int ep_send_events_proc(struct eventpoll *ep, struct list_head *head, 1016 void *priv) 1017 { 1018 struct ep_send_events_data *esed = priv; 1019 int eventcnt; 1020 unsigned int revents; 1021 struct epitem *epi; 1022 struct epoll_event __user *uevent; 1023 1024 /* 1025 * We can loop without lock because we are passed a task private list. 1026 * Items cannot vanish during the loop because ep_scan_ready_list() is 1027 * holding "mtx" during this call. 1028 */ 1029 for (eventcnt = 0, uevent = esed->events; 1030 !list_empty(head) && eventcnt < esed->maxevents;) { 1031 epi = list_first_entry(head, struct epitem, rdllink); 1032 1033 list_del_init(&epi->rdllink); 1034 1035 revents = epi->ffd.file->f_op->poll(epi->ffd.file, NULL) & 1036 epi->event.events; 1037 1038 /* 1039 * If the event mask intersect the caller-requested one, 1040 * deliver the event to userspace. Again, ep_scan_ready_list() 1041 * is holding "mtx", so no operations coming from userspace 1042 * can change the item. 1043 */ 1044 if (revents) { 1045 if (__put_user(revents, &uevent->events) || 1046 __put_user(epi->event.data, &uevent->data)) { 1047 list_add(&epi->rdllink, head); 1048 return eventcnt ? eventcnt : -EFAULT; 1049 } 1050 eventcnt++; 1051 uevent++; 1052 if (epi->event.events & EPOLLONESHOT) 1053 epi->event.events &= EP_PRIVATE_BITS; 1054 else if (!(epi->event.events & EPOLLET)) { 1055 /* 1056 * If this file has been added with Level 1057 * Trigger mode, we need to insert back inside 1058 * the ready list, so that the next call to 1059 * epoll_wait() will check again the events 1060 * availability. At this point, noone can insert 1061 * into ep->rdllist besides us. The epoll_ctl() 1062 * callers are locked out by 1063 * ep_scan_ready_list() holding "mtx" and the 1064 * poll callback will queue them in ep->ovflist. 1065 */ 1066 list_add_tail(&epi->rdllink, &ep->rdllist); 1067 } 1068 } 1069 } 1070 1071 return eventcnt; 1072 } 1073 1074 static int ep_send_events(struct eventpoll *ep, 1075 struct epoll_event __user *events, int maxevents) 1076 { 1077 struct ep_send_events_data esed; 1078 1079 esed.maxevents = maxevents; 1080 esed.events = events; 1081 1082 return ep_scan_ready_list(ep, ep_send_events_proc, &esed); 1083 } 1084 1085 static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events, 1086 int maxevents, long timeout) 1087 { 1088 int res, eavail; 1089 unsigned long flags; 1090 long jtimeout; 1091 wait_queue_t wait; 1092 1093 /* 1094 * Calculate the timeout by checking for the "infinite" value (-1) 1095 * and the overflow condition. The passed timeout is in milliseconds, 1096 * that why (t * HZ) / 1000. 1097 */ 1098 jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ? 1099 MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000; 1100 1101 retry: 1102 spin_lock_irqsave(&ep->lock, flags); 1103 1104 res = 0; 1105 if (list_empty(&ep->rdllist)) { 1106 /* 1107 * We don't have any available event to return to the caller. 1108 * We need to sleep here, and we will be wake up by 1109 * ep_poll_callback() when events will become available. 1110 */ 1111 init_waitqueue_entry(&wait, current); 1112 wait.flags |= WQ_FLAG_EXCLUSIVE; 1113 __add_wait_queue(&ep->wq, &wait); 1114 1115 for (;;) { 1116 /* 1117 * We don't want to sleep if the ep_poll_callback() sends us 1118 * a wakeup in between. That's why we set the task state 1119 * to TASK_INTERRUPTIBLE before doing the checks. 1120 */ 1121 set_current_state(TASK_INTERRUPTIBLE); 1122 if (!list_empty(&ep->rdllist) || !jtimeout) 1123 break; 1124 if (signal_pending(current)) { 1125 res = -EINTR; 1126 break; 1127 } 1128 1129 spin_unlock_irqrestore(&ep->lock, flags); 1130 jtimeout = schedule_timeout(jtimeout); 1131 spin_lock_irqsave(&ep->lock, flags); 1132 } 1133 __remove_wait_queue(&ep->wq, &wait); 1134 1135 set_current_state(TASK_RUNNING); 1136 } 1137 /* Is it worth to try to dig for events ? */ 1138 eavail = !list_empty(&ep->rdllist) || ep->ovflist != EP_UNACTIVE_PTR; 1139 1140 spin_unlock_irqrestore(&ep->lock, flags); 1141 1142 /* 1143 * Try to transfer events to user space. In case we get 0 events and 1144 * there's still timeout left over, we go trying again in search of 1145 * more luck. 1146 */ 1147 if (!res && eavail && 1148 !(res = ep_send_events(ep, events, maxevents)) && jtimeout) 1149 goto retry; 1150 1151 return res; 1152 } 1153 1154 /* 1155 * Open an eventpoll file descriptor. 1156 */ 1157 SYSCALL_DEFINE1(epoll_create1, int, flags) 1158 { 1159 int error; 1160 struct eventpoll *ep = NULL; 1161 1162 /* Check the EPOLL_* constant for consistency. */ 1163 BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC); 1164 1165 if (flags & ~EPOLL_CLOEXEC) 1166 return -EINVAL; 1167 /* 1168 * Create the internal data structure ("struct eventpoll"). 1169 */ 1170 error = ep_alloc(&ep); 1171 if (error < 0) 1172 return error; 1173 /* 1174 * Creates all the items needed to setup an eventpoll file. That is, 1175 * a file structure and a free file descriptor. 1176 */ 1177 error = anon_inode_getfd("[eventpoll]", &eventpoll_fops, ep, 1178 flags & O_CLOEXEC); 1179 if (error < 0) 1180 ep_free(ep); 1181 1182 return error; 1183 } 1184 1185 SYSCALL_DEFINE1(epoll_create, int, size) 1186 { 1187 if (size < 0) 1188 return -EINVAL; 1189 1190 return sys_epoll_create1(0); 1191 } 1192 1193 /* 1194 * The following function implements the controller interface for 1195 * the eventpoll file that enables the insertion/removal/change of 1196 * file descriptors inside the interest set. 1197 */ 1198 SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd, 1199 struct epoll_event __user *, event) 1200 { 1201 int error; 1202 struct file *file, *tfile; 1203 struct eventpoll *ep; 1204 struct epitem *epi; 1205 struct epoll_event epds; 1206 1207 error = -EFAULT; 1208 if (ep_op_has_event(op) && 1209 copy_from_user(&epds, event, sizeof(struct epoll_event))) 1210 goto error_return; 1211 1212 /* Get the "struct file *" for the eventpoll file */ 1213 error = -EBADF; 1214 file = fget(epfd); 1215 if (!file) 1216 goto error_return; 1217 1218 /* Get the "struct file *" for the target file */ 1219 tfile = fget(fd); 1220 if (!tfile) 1221 goto error_fput; 1222 1223 /* The target file descriptor must support poll */ 1224 error = -EPERM; 1225 if (!tfile->f_op || !tfile->f_op->poll) 1226 goto error_tgt_fput; 1227 1228 /* 1229 * We have to check that the file structure underneath the file descriptor 1230 * the user passed to us _is_ an eventpoll file. And also we do not permit 1231 * adding an epoll file descriptor inside itself. 1232 */ 1233 error = -EINVAL; 1234 if (file == tfile || !is_file_epoll(file)) 1235 goto error_tgt_fput; 1236 1237 /* 1238 * At this point it is safe to assume that the "private_data" contains 1239 * our own data structure. 1240 */ 1241 ep = file->private_data; 1242 1243 mutex_lock(&ep->mtx); 1244 1245 /* 1246 * Try to lookup the file inside our RB tree, Since we grabbed "mtx" 1247 * above, we can be sure to be able to use the item looked up by 1248 * ep_find() till we release the mutex. 1249 */ 1250 epi = ep_find(ep, tfile, fd); 1251 1252 error = -EINVAL; 1253 switch (op) { 1254 case EPOLL_CTL_ADD: 1255 if (!epi) { 1256 epds.events |= POLLERR | POLLHUP; 1257 1258 error = ep_insert(ep, &epds, tfile, fd); 1259 } else 1260 error = -EEXIST; 1261 break; 1262 case EPOLL_CTL_DEL: 1263 if (epi) 1264 error = ep_remove(ep, epi); 1265 else 1266 error = -ENOENT; 1267 break; 1268 case EPOLL_CTL_MOD: 1269 if (epi) { 1270 epds.events |= POLLERR | POLLHUP; 1271 error = ep_modify(ep, epi, &epds); 1272 } else 1273 error = -ENOENT; 1274 break; 1275 } 1276 mutex_unlock(&ep->mtx); 1277 1278 error_tgt_fput: 1279 fput(tfile); 1280 error_fput: 1281 fput(file); 1282 error_return: 1283 1284 return error; 1285 } 1286 1287 /* 1288 * Implement the event wait interface for the eventpoll file. It is the kernel 1289 * part of the user space epoll_wait(2). 1290 */ 1291 SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events, 1292 int, maxevents, int, timeout) 1293 { 1294 int error; 1295 struct file *file; 1296 struct eventpoll *ep; 1297 1298 /* The maximum number of event must be greater than zero */ 1299 if (maxevents <= 0 || maxevents > EP_MAX_EVENTS) 1300 return -EINVAL; 1301 1302 /* Verify that the area passed by the user is writeable */ 1303 if (!access_ok(VERIFY_WRITE, events, maxevents * sizeof(struct epoll_event))) { 1304 error = -EFAULT; 1305 goto error_return; 1306 } 1307 1308 /* Get the "struct file *" for the eventpoll file */ 1309 error = -EBADF; 1310 file = fget(epfd); 1311 if (!file) 1312 goto error_return; 1313 1314 /* 1315 * We have to check that the file structure underneath the fd 1316 * the user passed to us _is_ an eventpoll file. 1317 */ 1318 error = -EINVAL; 1319 if (!is_file_epoll(file)) 1320 goto error_fput; 1321 1322 /* 1323 * At this point it is safe to assume that the "private_data" contains 1324 * our own data structure. 1325 */ 1326 ep = file->private_data; 1327 1328 /* Time to fish for events ... */ 1329 error = ep_poll(ep, events, maxevents, timeout); 1330 1331 error_fput: 1332 fput(file); 1333 error_return: 1334 1335 return error; 1336 } 1337 1338 #ifdef HAVE_SET_RESTORE_SIGMASK 1339 1340 /* 1341 * Implement the event wait interface for the eventpoll file. It is the kernel 1342 * part of the user space epoll_pwait(2). 1343 */ 1344 SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events, 1345 int, maxevents, int, timeout, const sigset_t __user *, sigmask, 1346 size_t, sigsetsize) 1347 { 1348 int error; 1349 sigset_t ksigmask, sigsaved; 1350 1351 /* 1352 * If the caller wants a certain signal mask to be set during the wait, 1353 * we apply it here. 1354 */ 1355 if (sigmask) { 1356 if (sigsetsize != sizeof(sigset_t)) 1357 return -EINVAL; 1358 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) 1359 return -EFAULT; 1360 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP)); 1361 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); 1362 } 1363 1364 error = sys_epoll_wait(epfd, events, maxevents, timeout); 1365 1366 /* 1367 * If we changed the signal mask, we need to restore the original one. 1368 * In case we've got a signal while waiting, we do not restore the 1369 * signal mask yet, and we allow do_signal() to deliver the signal on 1370 * the way back to userspace, before the signal mask is restored. 1371 */ 1372 if (sigmask) { 1373 if (error == -EINTR) { 1374 memcpy(¤t->saved_sigmask, &sigsaved, 1375 sizeof(sigsaved)); 1376 set_restore_sigmask(); 1377 } else 1378 sigprocmask(SIG_SETMASK, &sigsaved, NULL); 1379 } 1380 1381 return error; 1382 } 1383 1384 #endif /* HAVE_SET_RESTORE_SIGMASK */ 1385 1386 static int __init eventpoll_init(void) 1387 { 1388 struct sysinfo si; 1389 1390 si_meminfo(&si); 1391 /* 1392 * Allows top 4% of lomem to be allocated for epoll watches (per user). 1393 */ 1394 max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) / 1395 EP_ITEM_COST; 1396 1397 /* Initialize the structure used to perform safe poll wait head wake ups */ 1398 ep_nested_calls_init(&poll_safewake_ncalls); 1399 1400 /* Initialize the structure used to perform file's f_op->poll() calls */ 1401 ep_nested_calls_init(&poll_readywalk_ncalls); 1402 1403 /* Allocates slab cache used to allocate "struct epitem" items */ 1404 epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem), 1405 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 1406 1407 /* Allocates slab cache used to allocate "struct eppoll_entry" */ 1408 pwq_cache = kmem_cache_create("eventpoll_pwq", 1409 sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL); 1410 1411 return 0; 1412 } 1413 fs_initcall(eventpoll_init); 1414