| 1 | |
| 2 | |
| 3 | * fs/eventpoll.c (Efficient event retrieval implementation)
|
| 4 | * Copyright (C) 2001,...,2009 Davide Libenzi
|
| 5 | *
|
| 6 | * Davide Libenzi <[email protected]>
|
| 7 |
|
| 8 |
|
| 9 | #include <linux/init.h>
|
| 10 | #include <linux/kernel.h>
|
| 11 | #include <linux/sched/signal.h>
|
| 12 | #include <linux/fs.h>
|
| 13 | #include <linux/file.h>
|
| 14 | #include <linux/signal.h>
|
| 15 | #include <linux/errno.h>
|
| 16 | #include <linux/mm.h>
|
| 17 | #include <linux/slab.h>
|
| 18 | #include <linux/poll.h>
|
| 19 | #include <linux/string.h>
|
| 20 | #include <linux/list.h>
|
| 21 | #include <linux/hash.h>
|
| 22 | #include <linux/spinlock.h>
|
| 23 | #include <linux/syscalls.h>
|
| 24 | #include <linux/rbtree.h>
|
| 25 | #include <linux/wait.h>
|
| 26 | #include <linux/eventpoll.h>
|
| 27 | #include <linux/mount.h>
|
| 28 | #include <linux/bitops.h>
|
| 29 | #include <linux/mutex.h>
|
| 30 | #include <linux/anon_inodes.h>
|
| 31 | #include <linux/device.h>
|
| 32 | #include <linux/uaccess.h>
|
| 33 | #include <asm/io.h>
|
| 34 | #include <asm/mman.h>
|
| 35 | #include <linux/atomic.h>
|
| 36 | #include <linux/proc_fs.h>
|
| 37 | #include <linux/seq_file.h>
|
| 38 | #include <linux/compat.h>
|
| 39 | #include <linux/rculist.h>
|
| 40 | #include <linux/capability.h>
|
| 41 | #include <linux/seqlock.h>
|
| 42 | #include <net/busy_poll.h>
|
| 43 |
|
| 44 | |
| 45 | * fs/eventpoll.c - Efficient event polling ("epoll") kernel implementation.
|
| 46 | *
|
| 47 | *
|
| 48 | * Overview
|
| 49 | * --------
|
| 50 | *
|
| 51 | * Each epoll_create(2) returns an anonymous [eventpoll] file whose
|
| 52 | * ->private_data is a struct eventpoll. Each EPOLL_CTL_ADD installs
|
| 53 | * a struct epitem linking one (watched file, fd) pair back to that
|
| 54 | * eventpoll via the watched file's f_op->poll() wait queue(s). When
|
| 55 | * the watched file signals readiness, ep_poll_callback() fires and
|
| 56 | * marks the epitem ready. epoll_wait(2) drains the ready list under
|
| 57 | * ep->mtx, re-queueing items in level-triggered mode.
|
| 58 | *
|
| 59 | * epoll instances can watch other epoll instances up to EP_MAX_NESTS
|
| 60 | * deep; cycles are forbidden and detected at EPOLL_CTL_ADD time.
|
| 61 | *
|
| 62 | *
|
| 63 | * Locking
|
| 64 | * -------
|
| 65 | *
|
| 66 | * Three levels, acquired from outer to inner:
|
| 67 | *
|
| 68 | * epnested_mutex (global; rare; taken only for EPOLL_CTL_ADD
|
| 69 | * loop / path checks)
|
| 70 | * > ep->mtx (per-eventpoll; sleepable; serializes most ops)
|
| 71 | * > ep->lock (per-eventpoll; IRQ-safe spinlock)
|
| 72 | *
|
| 73 | * file->f_lock (per-file; NOT IRQ-safe; guards f_ep hlist ops;
|
| 74 | * nested inside ep->mtx, outside ep->lock)
|
| 75 | *
|
| 76 | * Rationale:
|
| 77 | * - ep->lock is a spinlock because ep_poll_callback() is called from
|
| 78 | * wake_up() which may run in hard-IRQ context. All ep->lock
|
| 79 | * critical sections use spin_lock_irqsave().
|
| 80 | * - ep->mtx is a sleepable mutex because the event delivery loop
|
| 81 | * calls copy_to_user(), and ep_insert() may sleep in
|
| 82 | * kmem_cache_alloc() and f_op->poll().
|
| 83 | * - epnested_mutex is global because cycle detection needs a global
|
| 84 | * view of the epoll topology; a per-object scheme would let two
|
| 85 | * concurrent inserts (A into B, B into A) construct a cycle
|
| 86 | * without either observer seeing it.
|
| 87 | * - Per-ep ep->mtx is preferred for scalability elsewhere. Events
|
| 88 | * that require epnested_mutex are rare.
|
| 89 | *
|
| 90 | * When EPOLL_CTL_ADD nests one eventpoll inside another we acquire
|
| 91 | * ep->mtx on both: outer first, target second. Since cycles are
|
| 92 | * forbidden the set of live ep->mtx holds is always a strict chain,
|
| 93 | * communicated to lockdep via mutex_lock_nested() subclasses derived
|
| 94 | * from the current recursion depth.
|
| 95 | *
|
| 96 | *
|
| 97 | * Field protection
|
| 98 | * ----------------
|
| 99 | *
|
| 100 | * struct eventpoll:
|
| 101 | * mtx - self
|
| 102 | * rbr - ep->mtx
|
| 103 | * ovflist, rdllist - ep->lock (IRQ-safe)
|
| 104 | * wq - ep->lock for queue mutation
|
| 105 | * poll_wait - internal waitqueue spinlock
|
| 106 | * refs - file->f_lock for adds; ep->mtx for removes;
|
| 107 | * RCU for readers (hlist_del_rcu + kfree_rcu(ep))
|
| 108 | * ws - ep->mtx
|
| 109 | * gen, loop_check_depth - epnested_mutex
|
| 110 | * file, user - immutable after setup
|
| 111 | * refcount - atomic (refcount_t)
|
| 112 | * napi_* - READ_ONCE / WRITE_ONCE
|
| 113 | *
|
| 114 | * struct epitem:
|
| 115 | * rbn / rcu union - rbn: ep->mtx (while epi is linked in ep->rbr).
|
| 116 | * rcu: written only by kfree_rcu(epi) on the free
|
| 117 | * path; otherwise untouched by epoll code.
|
| 118 | * rdllink, next - ep->lock
|
| 119 | * ffd, ep - immutable after ep_insert()
|
| 120 | * pwqlist - ep->mtx for writes; POLLFREE clears pwq->whead
|
| 121 | * via smp_store_release(), see below
|
| 122 | * fllink - file->f_lock for mutation; hlist_del_rcu +
|
| 123 | * kfree_rcu(epi) for safe RCU readers
|
| 124 | * ws - RCU (rcu_assign_pointer /
|
| 125 | * rcu_dereference_check(mtx))
|
| 126 | * event - ep->mtx for writes; lockless read in
|
| 127 | * ep_poll_callback pairs with smp_mb() in
|
| 128 | * ep_modify()
|
| 129 | *
|
| 130 | *
|
| 131 | * Ready-list state machine
|
| 132 | * ------------------------
|
| 133 | *
|
| 134 | * Readiness is tracked in two lists under ep->lock:
|
| 135 | *
|
| 136 | * rdllist - doubly-linked FIFO; the "current" ready list.
|
| 137 | * ovflist - singly-linked LIFO; used during a scan to catch
|
| 138 | * events that arrive while rdllist is being iterated
|
| 139 | * without ep->lock.
|
| 140 | *
|
| 141 | * Encoded in ep->ovflist:
|
| 142 | * EP_UNACTIVE_PTR - no scan active; callback appends to rdllist.
|
| 143 | * NULL - scan active, no spill yet.
|
| 144 | * pointer to epi - scan active with spilled items (LIFO).
|
| 145 | *
|
| 146 | * Encoded in epi->ovflist_next:
|
| 147 | * EP_UNACTIVE_PTR - epi is not on ovflist.
|
| 148 | * otherwise - next epi on ovflist (NULL at tail).
|
| 149 | *
|
| 150 | * ep_start_scan() flips "not scanning" to "scanning" and splices
|
| 151 | * rdllist into a caller-local scan_batch. ep_done_scan() drains ovflist
|
| 152 | * back to rdllist (list_add head-insert reverses LIFO to FIFO),
|
| 153 | * flips back to "not scanning", and re-splices any items the caller
|
| 154 | * left in scan_batch (e.g., level-triggered re-queues).
|
| 155 | *
|
| 156 | *
|
| 157 | * Removal paths
|
| 158 | * -------------
|
| 159 | *
|
| 160 | * Three paths dispose of epitems and/or eventpolls:
|
| 161 | *
|
| 162 | * A. ep_remove() - EPOLL_CTL_DEL and ep_insert()
|
| 163 | * rollback. Caller holds ep->mtx.
|
| 164 | * B. ep_clear_and_put() - close of the epoll fd itself
|
| 165 | * (ep_eventpoll_release).
|
| 166 | * C. eventpoll_release_file() - close of a watched file, invoked
|
| 167 | * from __fput().
|
| 168 | *
|
| 169 | * Coordination:
|
| 170 | * A and C exclude each other via the watched file's refcount.
|
| 171 | * A pins the file with epi_fget() before touching file->f_ep or
|
| 172 | * file->f_lock; if the pin fails, __fput() is in flight and C
|
| 173 | * will clean this epi up. See the epi_fget() block comment.
|
| 174 | * A and B both hold ep->mtx serially. B walks the rbtree with
|
| 175 | * rb_next() captured before ep_remove() erases the current node.
|
| 176 | * B and C both take ep->mtx; the loser sees fewer entries or an
|
| 177 | * empty file->f_ep.
|
| 178 | *
|
| 179 | * Within every path the internal order is strict:
|
| 180 | * ep_unregister_pollwait() - drain pwqlist; synchronizes with any
|
| 181 | * in-flight ep_poll_callback via the
|
| 182 | * watched wait-queue head's lock.
|
| 183 | * ep_remove_file() - hlist_del_rcu of epi->fllink and,
|
| 184 | * if last watcher, clear file->f_ep,
|
| 185 | * under file->f_lock.
|
| 186 | * ep_remove_epi() - rb_erase, rdllist unlink (ep->lock),
|
| 187 | * wakeup_source_unregister,
|
| 188 | * kfree_rcu(epi).
|
| 189 | *
|
| 190 | * kfree_rcu(epi) defers the free past RCU readers in
|
| 191 | * reverse_path_check_proc(); kfree_rcu(ep) defers past readers in
|
| 192 | * ep_get_upwards_depth_proc().
|
| 193 | *
|
| 194 | *
|
| 195 | * POLLFREE handshake
|
| 196 | * ------------------
|
| 197 | *
|
| 198 | * When a subsystem tears down a wait-queue head that an epitem is
|
| 199 | * registered on (binder, signalfd, ...), it wakes the callback with
|
| 200 | * POLLFREE and must RCU-defer the head's free. The store/load pair:
|
| 201 | *
|
| 202 | * ep_poll_callback() POLLFREE branch:
|
| 203 | * smp_store_release(&pwq->whead, NULL)
|
| 204 | *
|
| 205 | * ep_remove_wait_queue():
|
| 206 | * smp_load_acquire(&pwq->whead)
|
| 207 | *
|
| 208 | * See those sites for the full argument.
|
| 209 |
|
| 210 |
|
| 211 |
|
| 212 | #define EP_PRIVATE_BITS (EPOLLWAKEUP | EPOLLONESHOT | EPOLLET | EPOLLEXCLUSIVE)
|
| 213 |
|
| 214 | #define EPOLLINOUT_BITS (EPOLLIN | EPOLLOUT)
|
| 215 |
|
| 216 | #define EPOLLEXCLUSIVE_OK_BITS (EPOLLINOUT_BITS | EPOLLERR | EPOLLHUP | \
|
| 217 | EPOLLWAKEUP | EPOLLET | EPOLLEXCLUSIVE)
|
| 218 |
|
| 219 |
|
| 220 | #define EP_MAX_NESTS 4
|
| 221 |
|
| 222 | #define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
|
| 223 |
|
| 224 | #define EP_UNACTIVE_PTR ((void *) -1L)
|
| 225 |
|
| 226 | #define EP_ITEM_COST (sizeof(struct epitem) + sizeof(struct eppoll_entry))
|
| 227 |
|
| 228 |
|
| 229 | struct eppoll_entry {
|
| 230 |
|
| 231 | struct eppoll_entry *next;
|
| 232 |
|
| 233 |
|
| 234 | struct epitem *base;
|
| 235 |
|
| 236 | |
| 237 | * Wait queue item that will be linked to the target file wait
|
| 238 | * queue head.
|
| 239 |
|
| 240 | wait_queue_entry_t wait;
|
| 241 |
|
| 242 |
|
| 243 | wait_queue_head_t *whead;
|
| 244 | };
|
| 245 |
|
| 246 | |
| 247 | * Each file descriptor added to the eventpoll interface will
|
| 248 | * have an entry of this type linked to the "rbr" RB tree.
|
| 249 | * Avoid increasing the size of this struct, there can be many thousands
|
| 250 | * of these on a server and we do not want this to take another cache line.
|
| 251 |
|
| 252 | struct epitem {
|
| 253 | union {
|
| 254 |
|
| 255 | struct rb_node rbn;
|
| 256 |
|
| 257 | struct rcu_head rcu;
|
| 258 | };
|
| 259 |
|
| 260 |
|
| 261 | struct list_head rdllink;
|
| 262 |
|
| 263 | |
| 264 | * Link on the owning eventpoll's scan-overflow list (ep->ovflist),
|
| 265 | * EP_UNACTIVE_PTR when not linked. See epi_on_ovflist() /
|
| 266 | * epi_clear_ovflist() and the "Ready-list state machine" section
|
| 267 | * in the top-of-file banner.
|
| 268 |
|
| 269 | struct epitem *ovflist_next;
|
| 270 |
|
| 271 |
|
| 272 | struct epoll_key ffd;
|
| 273 |
|
| 274 |
|
| 275 | struct eppoll_entry *pwqlist;
|
| 276 |
|
| 277 |
|
| 278 | struct eventpoll *ep;
|
| 279 |
|
| 280 |
|
| 281 | struct hlist_node fllink;
|
| 282 |
|
| 283 |
|
| 284 | struct wakeup_source __rcu *ws;
|
| 285 |
|
| 286 |
|
| 287 | struct epoll_event event;
|
| 288 | };
|
| 289 |
|
| 290 | |
| 291 | * This structure is stored inside the "private_data" member of the file
|
| 292 | * structure and represents the main data structure for the eventpoll
|
| 293 | * interface.
|
| 294 |
|
| 295 | struct eventpoll {
|
| 296 | |
| 297 | * This mutex is used to ensure that files are not removed
|
| 298 | * while epoll is using them. This is held during the event
|
| 299 | * collection loop, the file cleanup path, the epoll file exit
|
| 300 | * code and the ctl operations.
|
| 301 |
|
| 302 | struct mutex mtx;
|
| 303 |
|
| 304 |
|
| 305 | wait_queue_head_t wq;
|
| 306 |
|
| 307 |
|
| 308 | wait_queue_head_t poll_wait;
|
| 309 |
|
| 310 |
|
| 311 | struct list_head rdllist;
|
| 312 |
|
| 313 |
|
| 314 | spinlock_t lock;
|
| 315 |
|
| 316 |
|
| 317 | seqcount_spinlock_t seq;
|
| 318 |
|
| 319 |
|
| 320 | struct rb_root_cached rbr;
|
| 321 |
|
| 322 | |
| 323 | * This is a single linked list that chains all the "struct epitem" that
|
| 324 | * happened while transferring ready events to userspace w/out
|
| 325 | * holding ->lock.
|
| 326 |
|
| 327 | struct epitem *ovflist;
|
| 328 |
|
| 329 |
|
| 330 | struct wakeup_source *ws;
|
| 331 |
|
| 332 |
|
| 333 | struct user_struct *user;
|
| 334 |
|
| 335 | struct file *file;
|
| 336 |
|
| 337 |
|
| 338 | u64 gen;
|
| 339 | struct hlist_head refs;
|
| 340 | u8 loop_check_depth;
|
| 341 |
|
| 342 |
|
| 343 | refcount_t refcount;
|
| 344 |
|
| 345 |
|
| 346 | struct rcu_head rcu;
|
| 347 |
|
| 348 | #ifdef CONFIG_NET_RX_BUSY_POLL
|
| 349 |
|
| 350 | unsigned int napi_id;
|
| 351 |
|
| 352 | u32 busy_poll_usecs;
|
| 353 |
|
| 354 | u16 busy_poll_budget;
|
| 355 | bool prefer_busy_poll;
|
| 356 | #endif
|
| 357 |
|
| 358 | #ifdef CONFIG_DEBUG_LOCK_ALLOC
|
| 359 |
|
| 360 | u8 nests;
|
| 361 | #endif
|
| 362 | };
|
| 363 |
|
| 364 |
|
| 365 | struct ep_pqueue {
|
| 366 | poll_table pt;
|
| 367 | struct epitem *epi;
|
| 368 | };
|
| 369 |
|
| 370 | |
| 371 | * Configuration options available inside /proc/sys/fs/epoll/
|
| 372 |
|
| 373 |
|
| 374 | static long max_user_watches __read_mostly;
|
| 375 |
|
| 376 | |
| 377 | * Cycle and path-length checks at EPOLL_CTL_ADD
|
| 378 | * ---------------------------------------------
|
| 379 | *
|
| 380 | * When EPOLL_CTL_ADD creates a link that either targets an eventpoll
|
| 381 | * file or extends an existing chain of eventpolls, two checks run:
|
| 382 | *
|
| 383 | * 1. no cycle is being formed -- ep_loop_check() walks downward
|
| 384 | * from the candidate target, and ep_get_upwards_depth_proc()
|
| 385 | * walks upward from the outer ep, both bounded by EP_MAX_NESTS.
|
| 386 | * 2. no file accumulates more than path_limits[depth] wakeup paths
|
| 387 | * of a given length -- reverse_path_check().
|
| 388 | *
|
| 389 | * Both need a global view of the epoll topology and must be atomic
|
| 390 | * with the insertion, so the check is serialized by epnested_mutex
|
| 391 | * and carries its scratch state on a stack-allocated struct
|
| 392 | * ep_ctl_ctx scoped to one do_epoll_ctl() call. Non-nested inserts
|
| 393 | * skip this machinery entirely and take only ep->mtx.
|
| 394 | *
|
| 395 | * epnested_mutex Serializes the whole check.
|
| 396 | * loop_check_gen Global monotonic stamp, bumped at the start of
|
| 397 | * a check and again at the end. ep->gen caches
|
| 398 | * the value under which ep was last visited by
|
| 399 | * ep_loop_check_proc() or
|
| 400 | * ep_get_upwards_depth_proc(); the post-check
|
| 401 | * bump ensures those cached stamps can no longer
|
| 402 | * equal loop_check_gen, so the
|
| 403 | * "ep->gen == loop_check_gen" trigger in
|
| 404 | * ep_ctl_lock() only fires while another check
|
| 405 | * is in flight.
|
| 406 | *
|
| 407 | * struct ep_ctl_ctx carries the rest (inserting_into, tfile_check_list,
|
| 408 | * path_count[]) through the walk; see its declaration below.
|
| 409 | *
|
| 410 | * Commits fdcfce93073d ("eventpoll: Fix integer overflow in
|
| 411 | * ep_loop_check_proc()") and f2e467a48287 ("eventpoll: Fix
|
| 412 | * semi-unbounded recursion") hardened the walk; any refactor must
|
| 413 | * preserve both bail-outs.
|
| 414 |
|
| 415 | static DEFINE_MUTEX(epnested_mutex);
|
| 416 | static u64 loop_check_gen = 0;
|
| 417 |
|
| 418 | #define PATH_ARR_SIZE 5
|
| 419 |
|
| 420 | |
| 421 | * Per-do_epoll_ctl() scratch for the loop / path checks. Allocated on
|
| 422 | * the caller's stack; populated by ep_ctl_lock() and the downward
|
| 423 | * walk; consumed by reverse_path_check(); released by ep_ctl_unlock().
|
| 424 | * Only valid while the caller holds epnested_mutex.
|
| 425 |
|
| 426 | struct ep_ctl_ctx {
|
| 427 | |
| 428 | * Outer eventpoll for one ep_loop_check(); if the downward walk
|
| 429 | * reaches it the insert would form a cycle.
|
| 430 |
|
| 431 | struct eventpoll *inserting_into;
|
| 432 |
|
| 433 | |
| 434 | * Singly-linked list of epitems_head objects collected during
|
| 435 | * ep_loop_check_proc(), then walked by reverse_path_check().
|
| 436 | * Terminated by EP_UNACTIVE_PTR, not NULL: epitems_head->next
|
| 437 | * doubles as a membership flag (a NULL ->next means "not on this
|
| 438 | * list", see ep_remove_file()), so the list uses a non-NULL
|
| 439 | * sentinel to keep the tail head distinguishable from an unlisted
|
| 440 | * one.
|
| 441 |
|
| 442 | struct epitems_head *tfile_check_list;
|
| 443 |
|
| 444 | |
| 445 | * Per-depth wakeup-path tally used by reverse_path_check_proc();
|
| 446 | * reinitialized to zero at the start of each reverse_path_check()
|
| 447 | * iteration.
|
| 448 |
|
| 449 | int path_count[PATH_ARR_SIZE];
|
| 450 | };
|
| 451 |
|
| 452 |
|
| 453 | static struct kmem_cache *epi_cache __ro_after_init;
|
| 454 |
|
| 455 |
|
| 456 | static struct kmem_cache *pwq_cache __ro_after_init;
|
| 457 |
|
| 458 | |
| 459 | * Wrapper anchor for file->f_ep when the watched file is not itself an
|
| 460 | * eventpoll; for the epoll-watches-epoll case, file->f_ep points at
|
| 461 | * &watched_ep->refs directly. The ->next field threads
|
| 462 | * ctx->tfile_check_list during one EPOLL_CTL_ADD path check. The ->file
|
| 463 | * field holds a reference to the associated file while the head is on
|
| 464 | * the list.
|
| 465 |
|
| 466 | struct epitems_head {
|
| 467 | struct hlist_head epitems;
|
| 468 | struct epitems_head *next;
|
| 469 | struct file *file;
|
| 470 | };
|
| 471 |
|
| 472 | static struct kmem_cache *ephead_cache __ro_after_init;
|
| 473 |
|
| 474 | static inline void free_ephead(struct epitems_head *head)
|
| 475 | {
|
| 476 | if (head)
|
| 477 | kmem_cache_free(ephead_cache, head);
|
| 478 | }
|
| 479 |
|
| 480 | static void list_file(struct file *file, struct ep_ctl_ctx *ctx)
|
| 481 | {
|
| 482 | struct epitems_head *head;
|
| 483 |
|
| 484 | head = container_of(file->f_ep, struct epitems_head, epitems);
|
| 485 | if (!head->next) {
|
| 486 | |
| 487 | * The caller owns a reference to @file or holds the ep->mtx for the
|
| 488 | * epitem that led here. The latter blocks eventpoll_release_file()
|
| 489 | * before the file allocation can be freed and reused. A dying leaf
|
| 490 | * can be skipped since removing links cannot increase the reverse
|
| 491 | * path count.
|
| 492 |
|
| 493 | if (!file_ref_get(&file->f_ref))
|
| 494 | return;
|
| 495 | head->file = file;
|
| 496 | head->next = ctx->tfile_check_list;
|
| 497 | ctx->tfile_check_list = head;
|
| 498 | }
|
| 499 | }
|
| 500 |
|
| 501 | static void unlist_file(struct epitems_head *head)
|
| 502 | {
|
| 503 | struct epitems_head *to_free = head;
|
| 504 | struct hlist_node *p = rcu_dereference(hlist_first_rcu(&head->epitems));
|
| 505 | struct file *file = head->file;
|
| 506 | if (p) {
|
| 507 | struct epitem *epi= container_of(p, struct epitem, fllink);
|
| 508 | spin_lock(&epi->ffd.file->f_lock);
|
| 509 | if (!hlist_empty(&head->epitems))
|
| 510 | to_free = NULL;
|
| 511 | head->next = NULL;
|
| 512 | head->file = NULL;
|
| 513 | spin_unlock(&epi->ffd.file->f_lock);
|
| 514 | }
|
| 515 | free_ephead(to_free);
|
| 516 | fput(file);
|
| 517 | }
|
| 518 |
|
| 519 | #ifdef CONFIG_SYSCTL
|
| 520 |
|
| 521 | #include <linux/sysctl.h>
|
| 522 |
|
| 523 | static long long_zero;
|
| 524 | static long long_max = LONG_MAX;
|
| 525 |
|
| 526 | static const struct ctl_table epoll_table[] = {
|
| 527 | {
|
| 528 | .procname = "max_user_watches",
|
| 529 | .data = &max_user_watches,
|
| 530 | .maxlen = sizeof(max_user_watches),
|
| 531 | .mode = 0644,
|
| 532 | .proc_handler = proc_doulongvec_minmax,
|
| 533 | .extra1 = &long_zero,
|
| 534 | .extra2 = &long_max,
|
| 535 | },
|
| 536 | };
|
| 537 |
|
| 538 | static void __init epoll_sysctls_init(void)
|
| 539 | {
|
| 540 | register_sysctl("fs/epoll", epoll_table);
|
| 541 | }
|
| 542 | #else
|
| 543 | #define epoll_sysctls_init() do { } while (0)
|
| 544 | #endif
|
| 545 |
|
| 546 | static const struct file_operations eventpoll_fops;
|
| 547 |
|
| 548 | bool is_file_epoll(struct file *f)
|
| 549 | {
|
| 550 | return f->f_op == &eventpoll_fops;
|
| 551 | }
|
| 552 |
|
| 553 |
|
| 554 | static inline int ep_cmp_ffd(struct epoll_key *p1, struct epoll_key *p2)
|
| 555 | {
|
| 556 | return (p1->file > p2->file ? +1:
|
| 557 | (p1->file < p2->file ? -1 : p1->fd - p2->fd));
|
| 558 | }
|
| 559 |
|
| 560 |
|
| 561 | static inline bool ep_is_linked(struct epitem *epi)
|
| 562 | {
|
| 563 | return !list_empty(&epi->rdllink);
|
| 564 | }
|
| 565 |
|
| 566 | static inline struct eppoll_entry *ep_pwq_from_wait(wait_queue_entry_t *p)
|
| 567 | {
|
| 568 | return container_of(p, struct eppoll_entry, wait);
|
| 569 | }
|
| 570 |
|
| 571 |
|
| 572 | static inline struct epitem *ep_item_from_wait(wait_queue_entry_t *p)
|
| 573 | {
|
| 574 | return container_of(p, struct eppoll_entry, wait)->base;
|
| 575 | }
|
| 576 |
|
| 577 | |
| 578 | * Ready-list / ovflist state (see "Ready-list state machine" in the
|
| 579 | * top-of-file banner for the full state machine). EP_UNACTIVE_PTR is
|
| 580 | * the sentinel; these wrappers name each transition and each test so
|
| 581 | * call sites do not need to know the sentinel's value.
|
| 582 |
|
| 583 |
|
| 584 |
|
| 585 | static inline bool ep_is_scanning(struct eventpoll *ep)
|
| 586 | {
|
| 587 | return READ_ONCE(ep->ovflist) != EP_UNACTIVE_PTR;
|
| 588 | }
|
| 589 |
|
| 590 |
|
| 591 | static inline void ep_enter_scan(struct eventpoll *ep)
|
| 592 | {
|
| 593 | WRITE_ONCE(ep->ovflist, NULL);
|
| 594 | }
|
| 595 |
|
| 596 |
|
| 597 | static inline void ep_exit_scan(struct eventpoll *ep)
|
| 598 | {
|
| 599 | WRITE_ONCE(ep->ovflist, EP_UNACTIVE_PTR);
|
| 600 | }
|
| 601 |
|
| 602 |
|
| 603 | static inline bool epi_on_ovflist(const struct epitem *epi)
|
| 604 | {
|
| 605 | return epi->ovflist_next != EP_UNACTIVE_PTR;
|
| 606 | }
|
| 607 |
|
| 608 |
|
| 609 | static inline void epi_clear_ovflist(struct epitem *epi)
|
| 610 | {
|
| 611 | epi->ovflist_next = EP_UNACTIVE_PTR;
|
| 612 | }
|
| 613 |
|
| 614 |
|
| 615 | static inline bool ep_events_available(struct eventpoll *ep)
|
| 616 | {
|
| 617 | unsigned int seq = read_seqcount_begin(&ep->seq);
|
| 618 |
|
| 619 | return !list_empty_careful(&ep->rdllist) || ep_is_scanning(ep) ||
|
| 620 | read_seqcount_retry(&ep->seq, seq);
|
| 621 | }
|
| 622 |
|
| 623 | #ifdef CONFIG_NET_RX_BUSY_POLL
|
| 624 | |
| 625 | * busy_loop_ep_timeout - check if busy poll has timed out. The timeout value
|
| 626 | * from the epoll instance ep is preferred, but if it is not set fallback to
|
| 627 | * the system-wide global via busy_loop_timeout.
|
| 628 | *
|
| 629 | * @start_time: The start time used to compute the remaining time until timeout.
|
| 630 | * @ep: Pointer to the eventpoll context.
|
| 631 | *
|
| 632 | * Return: true if the timeout has expired, false otherwise.
|
| 633 |
|
| 634 | static bool busy_loop_ep_timeout(unsigned long start_time,
|
| 635 | struct eventpoll *ep)
|
| 636 | {
|
| 637 | unsigned long bp_usec = READ_ONCE(ep->busy_poll_usecs);
|
| 638 |
|
| 639 | if (bp_usec) {
|
| 640 | unsigned long end_time = start_time + bp_usec;
|
| 641 | unsigned long now = busy_loop_current_time();
|
| 642 |
|
| 643 | return time_after(now, end_time);
|
| 644 | } else {
|
| 645 | return busy_loop_timeout(start_time);
|
| 646 | }
|
| 647 | }
|
| 648 |
|
| 649 | static bool ep_busy_loop_on(struct eventpoll *ep)
|
| 650 | {
|
| 651 | return !!READ_ONCE(ep->busy_poll_usecs) ||
|
| 652 | READ_ONCE(ep->prefer_busy_poll) ||
|
| 653 | net_busy_loop_on();
|
| 654 | }
|
| 655 |
|
| 656 | static bool ep_busy_loop_end(void *p, unsigned long start_time)
|
| 657 | {
|
| 658 | struct eventpoll *ep = p;
|
| 659 |
|
| 660 | return ep_events_available(ep) || busy_loop_ep_timeout(start_time, ep);
|
| 661 | }
|
| 662 |
|
| 663 | |
| 664 | * Busy poll if globally on and supporting sockets found && no events,
|
| 665 | * busy loop will return if need_resched or ep_events_available.
|
| 666 | *
|
| 667 | * we must do our busy polling with irqs enabled
|
| 668 |
|
| 669 | static bool ep_busy_loop(struct eventpoll *ep)
|
| 670 | {
|
| 671 | unsigned int napi_id = READ_ONCE(ep->napi_id);
|
| 672 | u16 budget = READ_ONCE(ep->busy_poll_budget);
|
| 673 | bool prefer_busy_poll = READ_ONCE(ep->prefer_busy_poll);
|
| 674 |
|
| 675 | if (!budget)
|
| 676 | budget = BUSY_POLL_BUDGET;
|
| 677 |
|
| 678 | if (napi_id_valid(napi_id) && ep_busy_loop_on(ep)) {
|
| 679 | napi_busy_loop(napi_id, ep_busy_loop_end,
|
| 680 | ep, prefer_busy_poll, budget);
|
| 681 | if (ep_events_available(ep))
|
| 682 | return true;
|
| 683 | |
| 684 | * Busy poll timed out. Drop NAPI ID for now, we can add
|
| 685 | * it back in when we have moved a socket with a valid NAPI
|
| 686 | * ID onto the ready list.
|
| 687 |
|
| 688 | if (prefer_busy_poll)
|
| 689 | napi_resume_irqs(napi_id);
|
| 690 | ep->napi_id = 0;
|
| 691 | return false;
|
| 692 | }
|
| 693 | return false;
|
| 694 | }
|
| 695 |
|
| 696 | |
| 697 | * Set epoll busy poll NAPI ID from sk.
|
| 698 |
|
| 699 | static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
|
| 700 | {
|
| 701 | struct eventpoll *ep = epi->ep;
|
| 702 | unsigned int napi_id;
|
| 703 | struct socket *sock;
|
| 704 | struct sock *sk;
|
| 705 |
|
| 706 | if (!ep_busy_loop_on(ep))
|
| 707 | return;
|
| 708 |
|
| 709 | sock = sock_from_file(epi->ffd.file);
|
| 710 | if (!sock)
|
| 711 | return;
|
| 712 |
|
| 713 | sk = sock->sk;
|
| 714 | if (!sk)
|
| 715 | return;
|
| 716 |
|
| 717 | napi_id = READ_ONCE(sk->sk_napi_id);
|
| 718 |
|
| 719 | |
| 720 | * or
|
| 721 | * Nothing to do if we already have this ID
|
| 722 |
|
| 723 | if (!napi_id_valid(napi_id) || napi_id == ep->napi_id)
|
| 724 | return;
|
| 725 |
|
| 726 |
|
| 727 | ep->napi_id = napi_id;
|
| 728 | }
|
| 729 |
|
| 730 | static long ep_eventpoll_bp_ioctl(struct file *file, unsigned int cmd,
|
| 731 | unsigned long arg)
|
| 732 | {
|
| 733 | struct eventpoll *ep = file->private_data;
|
| 734 | void __user *uarg = (void __user *)arg;
|
| 735 | struct epoll_params epoll_params;
|
| 736 |
|
| 737 | switch (cmd) {
|
| 738 | case EPIOCSPARAMS:
|
| 739 | if (copy_from_user(&epoll_params, uarg, sizeof(epoll_params)))
|
| 740 | return -EFAULT;
|
| 741 |
|
| 742 |
|
| 743 | if (epoll_params.__pad)
|
| 744 | return -EINVAL;
|
| 745 |
|
| 746 | if (epoll_params.busy_poll_usecs > S32_MAX)
|
| 747 | return -EINVAL;
|
| 748 |
|
| 749 | if (epoll_params.prefer_busy_poll > 1)
|
| 750 | return -EINVAL;
|
| 751 |
|
| 752 | if (epoll_params.busy_poll_budget > NAPI_POLL_WEIGHT &&
|
| 753 | !capable(CAP_NET_ADMIN))
|
| 754 | return -EPERM;
|
| 755 |
|
| 756 | WRITE_ONCE(ep->busy_poll_usecs, epoll_params.busy_poll_usecs);
|
| 757 | WRITE_ONCE(ep->busy_poll_budget, epoll_params.busy_poll_budget);
|
| 758 | WRITE_ONCE(ep->prefer_busy_poll, epoll_params.prefer_busy_poll);
|
| 759 | return 0;
|
| 760 | case EPIOCGPARAMS:
|
| 761 | memset(&epoll_params, 0, sizeof(epoll_params));
|
| 762 | epoll_params.busy_poll_usecs = READ_ONCE(ep->busy_poll_usecs);
|
| 763 | epoll_params.busy_poll_budget = READ_ONCE(ep->busy_poll_budget);
|
| 764 | epoll_params.prefer_busy_poll = READ_ONCE(ep->prefer_busy_poll);
|
| 765 | if (copy_to_user(uarg, &epoll_params, sizeof(epoll_params)))
|
| 766 | return -EFAULT;
|
| 767 | return 0;
|
| 768 | default:
|
| 769 | return -ENOIOCTLCMD;
|
| 770 | }
|
| 771 | }
|
| 772 |
|
| 773 | static void ep_suspend_napi_irqs(struct eventpoll *ep)
|
| 774 | {
|
| 775 | unsigned int napi_id = READ_ONCE(ep->napi_id);
|
| 776 |
|
| 777 | if (napi_id_valid(napi_id) && READ_ONCE(ep->prefer_busy_poll))
|
| 778 | napi_suspend_irqs(napi_id);
|
| 779 | }
|
| 780 |
|
| 781 | static void ep_resume_napi_irqs(struct eventpoll *ep)
|
| 782 | {
|
| 783 | unsigned int napi_id = READ_ONCE(ep->napi_id);
|
| 784 |
|
| 785 | if (napi_id_valid(napi_id) && READ_ONCE(ep->prefer_busy_poll))
|
| 786 | napi_resume_irqs(napi_id);
|
| 787 | }
|
| 788 |
|
| 789 | #else
|
| 790 |
|
| 791 | static inline bool ep_busy_loop(struct eventpoll *ep)
|
| 792 | {
|
| 793 | return false;
|
| 794 | }
|
| 795 |
|
| 796 | static inline void ep_set_busy_poll_napi_id(struct epitem *epi)
|
| 797 | {
|
| 798 | }
|
| 799 |
|
| 800 | static long ep_eventpoll_bp_ioctl(struct file *file, unsigned int cmd,
|
| 801 | unsigned long arg)
|
| 802 | {
|
| 803 | return -EOPNOTSUPP;
|
| 804 | }
|
| 805 |
|
| 806 | static void ep_suspend_napi_irqs(struct eventpoll *ep)
|
| 807 | {
|
| 808 | }
|
| 809 |
|
| 810 | static void ep_resume_napi_irqs(struct eventpoll *ep)
|
| 811 | {
|
| 812 | }
|
| 813 |
|
| 814 | #endif
|
| 815 |
|
| 816 | |
| 817 | * As described in commit 0ccf831cb lockdep: annotate epoll
|
| 818 | * the use of wait queues used by epoll is done in a very controlled
|
| 819 | * manner. Wake ups can nest inside each other, but are never done
|
| 820 | * with the same locking. For example:
|
| 821 | *
|
| 822 | * dfd = socket(...);
|
| 823 | * efd1 = epoll_create();
|
| 824 | * efd2 = epoll_create();
|
| 825 | * epoll_ctl(efd1, EPOLL_CTL_ADD, dfd, ...);
|
| 826 | * epoll_ctl(efd2, EPOLL_CTL_ADD, efd1, ...);
|
| 827 | *
|
| 828 | * When a packet arrives to the device underneath "dfd", the net code will
|
| 829 | * issue a wake_up() on its poll wake list. Epoll (efd1) has installed a
|
| 830 | * callback wakeup entry on that queue, and the wake_up() performed by the
|
| 831 | * "dfd" net code will end up in ep_poll_callback(). At this point epoll
|
| 832 | * (efd1) notices that it may have some event ready, so it needs to wake up
|
| 833 | * the waiters on its poll wait list (efd2). So it calls ep_poll_safewake()
|
| 834 | * that ends up in another wake_up(), after having checked about the
|
| 835 | * recursion constraints. That are, no more than EP_MAX_NESTS, to avoid
|
| 836 | * stack blasting.
|
| 837 | *
|
| 838 | * When CONFIG_DEBUG_LOCK_ALLOC is enabled, make sure lockdep can handle
|
| 839 | * this special case of epoll.
|
| 840 |
|
| 841 | #ifdef CONFIG_DEBUG_LOCK_ALLOC
|
| 842 |
|
| 843 | static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
|
| 844 | unsigned pollflags)
|
| 845 | {
|
| 846 | struct eventpoll *ep_src;
|
| 847 | unsigned long flags;
|
| 848 | u8 nests = 0;
|
| 849 |
|
| 850 | |
| 851 | * To set the subclass or nesting level for spin_lock_irqsave_nested()
|
| 852 | * it might be natural to create a per-cpu nest count. However, since
|
| 853 | * we can recurse on ep->poll_wait.lock, and a non-raw spinlock can
|
| 854 | * schedule() in the -rt kernel, the per-cpu variable are no longer
|
| 855 | * protected. Thus, we are introducing a per eventpoll nest field.
|
| 856 | * If we are not being call from ep_poll_callback(), epi is NULL and
|
| 857 | * we are at the first level of nesting, 0. Otherwise, we are being
|
| 858 | * called from ep_poll_callback() and if a previous wakeup source is
|
| 859 | * not an epoll file itself, we are at depth 1 since the wakeup source
|
| 860 | * is depth 0. If the wakeup source is a previous epoll file in the
|
| 861 | * wakeup chain then we use its nests value and record ours as
|
| 862 | * nests + 1. The previous epoll file nests value is stable since its
|
| 863 | * already holding its own poll_wait.lock.
|
| 864 |
|
| 865 | if (epi) {
|
| 866 | if ((is_file_epoll(epi->ffd.file))) {
|
| 867 | ep_src = epi->ffd.file->private_data;
|
| 868 | nests = ep_src->nests;
|
| 869 | } else {
|
| 870 | nests = 1;
|
| 871 | }
|
| 872 | }
|
| 873 | spin_lock_irqsave_nested(&ep->poll_wait.lock, flags, nests);
|
| 874 | ep->nests = nests + 1;
|
| 875 | wake_up_locked_poll(&ep->poll_wait, EPOLLIN | pollflags);
|
| 876 | ep->nests = 0;
|
| 877 | spin_unlock_irqrestore(&ep->poll_wait.lock, flags);
|
| 878 | }
|
| 879 |
|
| 880 | #else
|
| 881 |
|
| 882 | static void ep_poll_safewake(struct eventpoll *ep, struct epitem *epi,
|
| 883 | __poll_t pollflags)
|
| 884 | {
|
| 885 | wake_up_poll(&ep->poll_wait, EPOLLIN | pollflags);
|
| 886 | }
|
| 887 |
|
| 888 | #endif
|
| 889 |
|
| 890 | static void ep_remove_wait_queue(struct eppoll_entry *pwq)
|
| 891 | {
|
| 892 | wait_queue_head_t *whead;
|
| 893 |
|
| 894 | rcu_read_lock();
|
| 895 | |
| 896 | * POLLFREE handshake, acquire side; see "POLLFREE handshake"
|
| 897 | * at the top of this file.
|
| 898 | *
|
| 899 | * A NULL load is paired with the smp_store_release(&whead, NULL)
|
| 900 | * in ep_poll_callback()'s POLLFREE branch: the teardown is
|
| 901 | * complete and we must not touch whead again. On a non-NULL load
|
| 902 | * rcu_read_lock() keeps the waitqueue memory alive (POLLFREE
|
| 903 | * firers RCU-defer the free) and whead->lock inside
|
| 904 | * remove_wait_queue() serializes us against the store side.
|
| 905 |
|
| 906 | whead = smp_load_acquire(&pwq->whead);
|
| 907 | if (whead)
|
| 908 | remove_wait_queue(whead, &pwq->wait);
|
| 909 | rcu_read_unlock();
|
| 910 | }
|
| 911 |
|
| 912 | |
| 913 | * This function unregisters poll callbacks from the associated file
|
| 914 | * descriptor. Must be called with "mtx" held.
|
| 915 |
|
| 916 | static void ep_unregister_pollwait(struct eventpoll *ep, struct epitem *epi)
|
| 917 | {
|
| 918 | struct eppoll_entry **p = &epi->pwqlist;
|
| 919 | struct eppoll_entry *pwq;
|
| 920 |
|
| 921 | while ((pwq = *p) != NULL) {
|
| 922 | *p = pwq->next;
|
| 923 | ep_remove_wait_queue(pwq);
|
| 924 | kmem_cache_free(pwq_cache, pwq);
|
| 925 | }
|
| 926 | }
|
| 927 |
|
| 928 |
|
| 929 | static inline struct wakeup_source *ep_wakeup_source(struct epitem *epi)
|
| 930 | {
|
| 931 | return rcu_dereference_check(epi->ws, lockdep_is_held(&epi->ep->mtx));
|
| 932 | }
|
| 933 |
|
| 934 |
|
| 935 | static inline void ep_pm_stay_awake(struct epitem *epi)
|
| 936 | {
|
| 937 | struct wakeup_source *ws = ep_wakeup_source(epi);
|
| 938 |
|
| 939 | if (ws)
|
| 940 | __pm_stay_awake(ws);
|
| 941 | }
|
| 942 |
|
| 943 | static inline bool ep_has_wakeup_source(struct epitem *epi)
|
| 944 | {
|
| 945 | return rcu_access_pointer(epi->ws) ? true : false;
|
| 946 | }
|
| 947 |
|
| 948 |
|
| 949 | static inline void ep_pm_stay_awake_rcu(struct epitem *epi)
|
| 950 | {
|
| 951 | struct wakeup_source *ws;
|
| 952 |
|
| 953 | rcu_read_lock();
|
| 954 | ws = rcu_dereference(epi->ws);
|
| 955 | if (ws)
|
| 956 | __pm_stay_awake(ws);
|
| 957 | rcu_read_unlock();
|
| 958 | }
|
| 959 |
|
| 960 |
|
| 961 | |
| 962 | * ep->mutex needs to be held because we could be hit by
|
| 963 | * eventpoll_release_file() and epoll_ctl().
|
| 964 |
|
| 965 | static void ep_start_scan(struct eventpoll *ep, struct list_head *scan_batch)
|
| 966 | {
|
| 967 | |
| 968 | * Steal the ready list, and re-init the original one to the
|
| 969 | * empty list. Also, set ep->ovflist to NULL so that events
|
| 970 | * happening while looping w/out locks, are not lost. We cannot
|
| 971 | * have the poll callback to queue directly on ep->rdllist,
|
| 972 | * because we want the "sproc" callback to be able to do it
|
| 973 | * in a lockless way.
|
| 974 |
|
| 975 | lockdep_assert_irqs_enabled();
|
| 976 | spin_lock_irq(&ep->lock);
|
| 977 | write_seqcount_begin(&ep->seq);
|
| 978 |
|
| 979 | list_splice_init(&ep->rdllist, scan_batch);
|
| 980 | ep_enter_scan(ep);
|
| 981 |
|
| 982 | write_seqcount_end(&ep->seq);
|
| 983 | spin_unlock_irq(&ep->lock);
|
| 984 | }
|
| 985 |
|
| 986 | static void ep_done_scan(struct eventpoll *ep,
|
| 987 | struct list_head *scan_batch)
|
| 988 | {
|
| 989 | struct epitem *epi, *nepi;
|
| 990 |
|
| 991 | spin_lock_irq(&ep->lock);
|
| 992 | |
| 993 | * During the time we spent inside the "sproc" callback, some
|
| 994 | * other events might have been queued by the poll callback.
|
| 995 | * We re-insert them inside the main ready-list here.
|
| 996 |
|
| 997 | for (nepi = READ_ONCE(ep->ovflist); (epi = nepi) != NULL; ) {
|
| 998 | nepi = epi->ovflist_next;
|
| 999 | epi_clear_ovflist(epi);
|
| 1000 | |
| 1001 | * Skip items that the caller already returned via @scan_batch
|
| 1002 | * -- the list_splice() below takes care of those.
|
| 1003 |
|
| 1004 | if (!ep_is_linked(epi)) {
|
| 1005 | |
| 1006 | * ovflist is LIFO; list_add() head-insert here
|
| 1007 | * reverses the iteration order into FIFO.
|
| 1008 |
|
| 1009 | list_add(&epi->rdllink, &ep->rdllist);
|
| 1010 | ep_pm_stay_awake(epi);
|
| 1011 | }
|
| 1012 | }
|
| 1013 |
|
| 1014 | write_seqcount_begin(&ep->seq);
|
| 1015 |
|
| 1016 |
|
| 1017 | ep_exit_scan(ep);
|
| 1018 |
|
| 1019 | |
| 1020 | * Quickly re-inject items left on "scan_batch".
|
| 1021 |
|
| 1022 | list_splice(scan_batch, &ep->rdllist);
|
| 1023 |
|
| 1024 | write_seqcount_end(&ep->seq);
|
| 1025 |
|
| 1026 | __pm_relax(ep->ws);
|
| 1027 |
|
| 1028 | if (!list_empty(&ep->rdllist)) {
|
| 1029 | if (waitqueue_active(&ep->wq))
|
| 1030 | wake_up(&ep->wq);
|
| 1031 | }
|
| 1032 |
|
| 1033 | spin_unlock_irq(&ep->lock);
|
| 1034 | }
|
| 1035 |
|
| 1036 | static void ep_get(struct eventpoll *ep)
|
| 1037 | {
|
| 1038 | refcount_inc(&ep->refcount);
|
| 1039 | }
|
| 1040 |
|
| 1041 | |
| 1042 | * Drop a reference to @ep; returns true iff it was the last, in which
|
| 1043 | * case the caller is responsible for ep_free().
|
| 1044 |
|
| 1045 | static bool ep_put(struct eventpoll *ep)
|
| 1046 | {
|
| 1047 | if (!refcount_dec_and_test(&ep->refcount))
|
| 1048 | return false;
|
| 1049 |
|
| 1050 | WARN_ON_ONCE(!RB_EMPTY_ROOT(&ep->rbr.rb_root));
|
| 1051 | return true;
|
| 1052 | }
|
| 1053 |
|
| 1054 | static void ep_free(struct eventpoll *ep)
|
| 1055 | {
|
| 1056 | ep_resume_napi_irqs(ep);
|
| 1057 | mutex_destroy(&ep->mtx);
|
| 1058 | free_uid(ep->user);
|
| 1059 | wakeup_source_unregister(ep->ws);
|
| 1060 |
|
| 1061 | kfree_rcu(ep, rcu);
|
| 1062 | }
|
| 1063 |
|
| 1064 | |
| 1065 | * Pin @epi->ffd.file for operations that require both safe dereference
|
| 1066 | * and exclusion from __fput().
|
| 1067 | *
|
| 1068 | * struct file uses SLAB_TYPESAFE_BY_RCU, so a freed slot can be
|
| 1069 | * reassigned at any time. The bare load of epi->ffd.file is safe here
|
| 1070 | * because the caller holds ep->mtx and eventpoll_release_file() blocks
|
| 1071 | * on that mutex while tearing down the epi, so the backing file
|
| 1072 | * allocation cannot be freed and reused under us. An rcu_read_lock()
|
| 1073 | * is therefore unnecessary for the load.
|
| 1074 | *
|
| 1075 | * A successful file_ref_get() additionally blocks __fput() from
|
| 1076 | * starting on this file: once the refcount has reached zero it cannot
|
| 1077 | * come back. ep_remove() relies on that to touch file->f_lock and
|
| 1078 | * file->f_ep without racing eventpoll_release_file() (see commit
|
| 1079 | * a6dc643c6931). A NULL return means __fput() is already in flight;
|
| 1080 | * the caller must bail without touching the file, and
|
| 1081 | * eventpoll_release_file() will clean the epi up from its side.
|
| 1082 |
|
| 1083 | static struct file *epi_fget(const struct epitem *epi)
|
| 1084 | {
|
| 1085 | struct file *file;
|
| 1086 |
|
| 1087 | file = epi->ffd.file;
|
| 1088 | if (!file_ref_get(&file->f_ref))
|
| 1089 | file = NULL;
|
| 1090 | return file;
|
| 1091 | }
|
| 1092 |
|
| 1093 | |
| 1094 | * Takes &file->f_lock; returns with it released.
|
| 1095 |
|
| 1096 | static void ep_remove_file(struct eventpoll *ep, struct epitem *epi,
|
| 1097 | struct file *file)
|
| 1098 | {
|
| 1099 | struct epitems_head *to_free = NULL;
|
| 1100 | struct hlist_head *head;
|
| 1101 |
|
| 1102 | lockdep_assert_held(&ep->mtx);
|
| 1103 |
|
| 1104 | spin_lock(&file->f_lock);
|
| 1105 | head = file->f_ep;
|
| 1106 | if (hlist_is_singular_node(&epi->fllink, head)) {
|
| 1107 | |
| 1108 | * Last watcher: publish NULL so the eventpoll_release()
|
| 1109 | * fastpath in include/linux/eventpoll.h can skip the slow
|
| 1110 | * path on a future __fput(). Safe because every f_ep writer
|
| 1111 | * either holds a pin on @file via epi_fget() or is __fput()
|
| 1112 | * itself -- see the comment in eventpoll_release().
|
| 1113 |
|
| 1114 | WRITE_ONCE(file->f_ep, NULL);
|
| 1115 | if (!is_file_epoll(file)) {
|
| 1116 | struct epitems_head *v;
|
| 1117 | v = container_of(head, struct epitems_head, epitems);
|
| 1118 | if (!smp_load_acquire(&v->next))
|
| 1119 | to_free = v;
|
| 1120 | }
|
| 1121 | }
|
| 1122 | hlist_del_rcu(&epi->fllink);
|
| 1123 | spin_unlock(&file->f_lock);
|
| 1124 | free_ephead(to_free);
|
| 1125 | }
|
| 1126 |
|
| 1127 | static void ep_remove_epi(struct eventpoll *ep, struct epitem *epi)
|
| 1128 | {
|
| 1129 | lockdep_assert_held(&ep->mtx);
|
| 1130 |
|
| 1131 | rb_erase_cached(&epi->rbn, &ep->rbr);
|
| 1132 |
|
| 1133 | spin_lock_irq(&ep->lock);
|
| 1134 | if (ep_is_linked(epi))
|
| 1135 | list_del_init(&epi->rdllink);
|
| 1136 | spin_unlock_irq(&ep->lock);
|
| 1137 |
|
| 1138 | wakeup_source_unregister(ep_wakeup_source(epi));
|
| 1139 | |
| 1140 | * At this point it is safe to free the eventpoll item. Use the union
|
| 1141 | * field epi->rcu, since we are trying to minimize the size of
|
| 1142 | * 'struct epitem'. The 'rbn' field is no longer in use. Protected by
|
| 1143 | * ep->mtx. The rcu read side, reverse_path_check_proc(), does not make
|
| 1144 | * use of the rbn field.
|
| 1145 |
|
| 1146 | kfree_rcu(epi, rcu);
|
| 1147 |
|
| 1148 | percpu_counter_dec(&ep->user->epoll_watches);
|
| 1149 | }
|
| 1150 |
|
| 1151 | |
| 1152 | * ep_remove variant for callers owing an additional reference to the ep
|
| 1153 |
|
| 1154 | static void ep_remove(struct eventpoll *ep, struct epitem *epi)
|
| 1155 | {
|
| 1156 | struct file *file __free(fput) = NULL;
|
| 1157 |
|
| 1158 | lockdep_assert_irqs_enabled();
|
| 1159 | lockdep_assert_held(&ep->mtx);
|
| 1160 |
|
| 1161 | ep_unregister_pollwait(ep, epi);
|
| 1162 |
|
| 1163 | |
| 1164 | * If we manage to grab a reference it means we're not in
|
| 1165 | * eventpoll_release_file() and aren't going to be: once @file's
|
| 1166 | * refcount has reached zero, file_ref_get() cannot bring it back.
|
| 1167 |
|
| 1168 | file = epi_fget(epi);
|
| 1169 | if (!file)
|
| 1170 | return;
|
| 1171 |
|
| 1172 | ep_remove_file(ep, epi, file);
|
| 1173 | ep_remove_epi(ep, epi);
|
| 1174 | WARN_ON_ONCE(ep_put(ep));
|
| 1175 | }
|
| 1176 |
|
| 1177 | |
| 1178 | * Pass 1 of ep_clear_and_put(): drain every epi's pwqlist.
|
| 1179 | * ep_unregister_pollwait() takes each watched wait-queue head's lock,
|
| 1180 | * which synchronizes with any in-flight ep_poll_callback(); after
|
| 1181 | * this returns no callback can still be about to dereference an epi
|
| 1182 | * on this ep. Must strictly precede ep_drain_tree() -- fusing the
|
| 1183 | * two walks would let a callback queued on epi_i still fire after
|
| 1184 | * epi_{i+k} had already been freed.
|
| 1185 |
|
| 1186 | static void ep_drain_pollwaits(struct eventpoll *ep)
|
| 1187 | {
|
| 1188 | struct rb_node *rbp;
|
| 1189 | struct epitem *epi;
|
| 1190 |
|
| 1191 | lockdep_assert_held(&ep->mtx);
|
| 1192 |
|
| 1193 | for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
|
| 1194 | epi = rb_entry(rbp, struct epitem, rbn);
|
| 1195 |
|
| 1196 | ep_unregister_pollwait(ep, epi);
|
| 1197 | cond_resched();
|
| 1198 | }
|
| 1199 | }
|
| 1200 |
|
| 1201 | |
| 1202 | * Pass 2 of ep_clear_and_put(): ep_remove() every epi. The per-epi
|
| 1203 | * pwqlist is already empty (ep_drain_pollwaits ran), but the rest of
|
| 1204 | * ep_remove() still runs: epi_fget() pin, f_ep clear under f_lock,
|
| 1205 | * rbtree erase, rdllist unlink, kfree_rcu(epi). rb_next() is captured
|
| 1206 | * before each erase so the iteration is stable.
|
| 1207 | *
|
| 1208 | * A concurrent eventpoll_release_file() (removal path C) on a watched
|
| 1209 | * file serializes with us via ep->mtx; ep_remove() transparently
|
| 1210 | * hands off any epi whose file is in __fput() by bailing when
|
| 1211 | * epi_fget() returns NULL, and path C will clean that epi up.
|
| 1212 |
|
| 1213 | static void ep_drain_tree(struct eventpoll *ep)
|
| 1214 | {
|
| 1215 | struct rb_node *rbp, *next;
|
| 1216 | struct epitem *epi;
|
| 1217 |
|
| 1218 | lockdep_assert_held(&ep->mtx);
|
| 1219 |
|
| 1220 | for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = next) {
|
| 1221 | next = rb_next(rbp);
|
| 1222 | epi = rb_entry(rbp, struct epitem, rbn);
|
| 1223 | ep_remove(ep, epi);
|
| 1224 | cond_resched();
|
| 1225 | }
|
| 1226 | }
|
| 1227 |
|
| 1228 | |
| 1229 | * Removal path B (see "Removal paths" in the top-of-file banner):
|
| 1230 | * close of the epoll fd itself, reached via ep_eventpoll_release().
|
| 1231 | *
|
| 1232 | * Two passes under ep->mtx: first ep_drain_pollwaits() quiesces
|
| 1233 | * in-flight callbacks, then ep_drain_tree() frees the epis. The
|
| 1234 | * ep->refcount is kept > 0 across the walk by the ep file's own
|
| 1235 | * share, which we drop below; ep_free() runs iff we were the last
|
| 1236 | * holder after the tree drained.
|
| 1237 |
|
| 1238 | static void ep_clear_and_put(struct eventpoll *ep)
|
| 1239 | {
|
| 1240 |
|
| 1241 | if (waitqueue_active(&ep->poll_wait))
|
| 1242 | ep_poll_safewake(ep, NULL, 0);
|
| 1243 |
|
| 1244 | mutex_lock(&ep->mtx);
|
| 1245 | ep_drain_pollwaits(ep);
|
| 1246 | ep_drain_tree(ep);
|
| 1247 | mutex_unlock(&ep->mtx);
|
| 1248 |
|
| 1249 | if (ep_put(ep))
|
| 1250 | ep_free(ep);
|
| 1251 | }
|
| 1252 |
|
| 1253 | static long ep_eventpoll_ioctl(struct file *file, unsigned int cmd,
|
| 1254 | unsigned long arg)
|
| 1255 | {
|
| 1256 | int ret;
|
| 1257 |
|
| 1258 | if (!is_file_epoll(file))
|
| 1259 | return -EINVAL;
|
| 1260 |
|
| 1261 | switch (cmd) {
|
| 1262 | case EPIOCSPARAMS:
|
| 1263 | case EPIOCGPARAMS:
|
| 1264 | ret = ep_eventpoll_bp_ioctl(file, cmd, arg);
|
| 1265 | break;
|
| 1266 | default:
|
| 1267 | ret = -EINVAL;
|
| 1268 | break;
|
| 1269 | }
|
| 1270 |
|
| 1271 | return ret;
|
| 1272 | }
|
| 1273 |
|
| 1274 | static int ep_eventpoll_release(struct inode *inode, struct file *file)
|
| 1275 | {
|
| 1276 | struct eventpoll *ep = file->private_data;
|
| 1277 |
|
| 1278 | if (ep)
|
| 1279 | ep_clear_and_put(ep);
|
| 1280 |
|
| 1281 | return 0;
|
| 1282 | }
|
| 1283 |
|
| 1284 | static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt, int depth);
|
| 1285 |
|
| 1286 | static __poll_t __ep_eventpoll_poll(struct file *file, poll_table *wait, int depth)
|
| 1287 | {
|
| 1288 | struct eventpoll *ep = file->private_data;
|
| 1289 | LIST_HEAD(scan_batch);
|
| 1290 | struct epitem *epi, *tmp;
|
| 1291 | poll_table pt;
|
| 1292 | __poll_t res = 0;
|
| 1293 |
|
| 1294 | init_poll_funcptr(&pt, NULL);
|
| 1295 |
|
| 1296 |
|
| 1297 | poll_wait(file, &ep->poll_wait, wait);
|
| 1298 |
|
| 1299 | |
| 1300 | * Proceed to find out if wanted events are really available inside
|
| 1301 | * the ready list.
|
| 1302 |
|
| 1303 | mutex_lock_nested(&ep->mtx, depth);
|
| 1304 | ep_start_scan(ep, &scan_batch);
|
| 1305 | list_for_each_entry_safe(epi, tmp, &scan_batch, rdllink) {
|
| 1306 | if (ep_item_poll(epi, &pt, depth + 1)) {
|
| 1307 | res = EPOLLIN | EPOLLRDNORM;
|
| 1308 | break;
|
| 1309 | } else {
|
| 1310 | |
| 1311 | * Item has been dropped into the ready list by the poll
|
| 1312 | * callback, but it's not actually ready, as far as
|
| 1313 | * caller requested events goes. We can remove it here.
|
| 1314 |
|
| 1315 | __pm_relax(ep_wakeup_source(epi));
|
| 1316 | list_del_init(&epi->rdllink);
|
| 1317 | }
|
| 1318 | }
|
| 1319 | ep_done_scan(ep, &scan_batch);
|
| 1320 | mutex_unlock(&ep->mtx);
|
| 1321 | return res;
|
| 1322 | }
|
| 1323 |
|
| 1324 | |
| 1325 | * Differs from ep_eventpoll_poll() in that internal callers already have
|
| 1326 | * the ep->mtx so we need to start from depth=1, such that mutex_lock_nested()
|
| 1327 | * is correctly annotated.
|
| 1328 |
|
| 1329 | static __poll_t ep_item_poll(const struct epitem *epi, poll_table *pt,
|
| 1330 | int depth)
|
| 1331 | {
|
| 1332 | struct file *file = epi_fget(epi);
|
| 1333 | __poll_t res;
|
| 1334 |
|
| 1335 | |
| 1336 | * We could return EPOLLERR | EPOLLHUP or something, but let's
|
| 1337 | * treat this more as "file doesn't exist, poll didn't happen".
|
| 1338 |
|
| 1339 | if (!file)
|
| 1340 | return 0;
|
| 1341 |
|
| 1342 | pt->_key = epi->event.events;
|
| 1343 | if (!is_file_epoll(file))
|
| 1344 | res = vfs_poll(file, pt);
|
| 1345 | else
|
| 1346 | res = __ep_eventpoll_poll(file, pt, depth);
|
| 1347 | fput(file);
|
| 1348 | return res & epi->event.events;
|
| 1349 | }
|
| 1350 |
|
| 1351 | static __poll_t ep_eventpoll_poll(struct file *file, poll_table *wait)
|
| 1352 | {
|
| 1353 | return __ep_eventpoll_poll(file, wait, 0);
|
| 1354 | }
|
| 1355 |
|
| 1356 | #ifdef CONFIG_PROC_FS
|
| 1357 | static void ep_show_fdinfo(struct seq_file *m, struct file *f)
|
| 1358 | {
|
| 1359 | struct eventpoll *ep = f->private_data;
|
| 1360 | struct rb_node *rbp;
|
| 1361 |
|
| 1362 | mutex_lock(&ep->mtx);
|
| 1363 | for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
|
| 1364 | struct epitem *epi = rb_entry(rbp, struct epitem, rbn);
|
| 1365 | struct inode *inode = file_inode(epi->ffd.file);
|
| 1366 |
|
| 1367 | seq_printf(m, "tfd: %8d events: %8x data: %16llx "
|
| 1368 | " pos:%lli ino:%llx sdev:%x\n",
|
| 1369 | epi->ffd.fd, epi->event.events,
|
| 1370 | (long long)epi->event.data,
|
| 1371 | (long long)epi->ffd.file->f_pos,
|
| 1372 | inode->i_ino, inode->i_sb->s_dev);
|
| 1373 | if (seq_has_overflowed(m))
|
| 1374 | break;
|
| 1375 | }
|
| 1376 | mutex_unlock(&ep->mtx);
|
| 1377 | }
|
| 1378 | #endif
|
| 1379 |
|
| 1380 |
|
| 1381 | static const struct file_operations eventpoll_fops = {
|
| 1382 | #ifdef CONFIG_PROC_FS
|
| 1383 | .show_fdinfo = ep_show_fdinfo,
|
| 1384 | #endif
|
| 1385 | .release = ep_eventpoll_release,
|
| 1386 | .poll = ep_eventpoll_poll,
|
| 1387 | .llseek = noop_llseek,
|
| 1388 | .unlocked_ioctl = ep_eventpoll_ioctl,
|
| 1389 | .compat_ioctl = compat_ptr_ioctl,
|
| 1390 | };
|
| 1391 |
|
| 1392 | |
| 1393 | * This is called from eventpoll_release() to unlink files from the eventpoll
|
| 1394 | * interface. We need to have this facility to cleanup correctly files that are
|
| 1395 | * closed without being removed from the eventpoll interface.
|
| 1396 |
|
| 1397 | void eventpoll_release_file(struct file *file)
|
| 1398 | {
|
| 1399 | struct eventpoll *ep;
|
| 1400 | struct epitem *epi;
|
| 1401 |
|
| 1402 | |
| 1403 | * A concurrent ep_remove() cannot outrace us: it pins @file via
|
| 1404 | * epi_fget(), which fails once __fput() has dropped the refcount
|
| 1405 | * to zero -- the path we're on. So any racing ep_remove() bails
|
| 1406 | * and leaves the epi for us to clean up here.
|
| 1407 |
|
| 1408 | again:
|
| 1409 | spin_lock(&file->f_lock);
|
| 1410 | if (file->f_ep && file->f_ep->first) {
|
| 1411 | epi = hlist_entry(file->f_ep->first, struct epitem, fllink);
|
| 1412 | spin_unlock(&file->f_lock);
|
| 1413 |
|
| 1414 | |
| 1415 | * ep access is safe as we still own a reference to the ep
|
| 1416 | * struct
|
| 1417 |
|
| 1418 | ep = epi->ep;
|
| 1419 | mutex_lock(&ep->mtx);
|
| 1420 |
|
| 1421 | ep_unregister_pollwait(ep, epi);
|
| 1422 |
|
| 1423 | ep_remove_file(ep, epi, file);
|
| 1424 | ep_remove_epi(ep, epi);
|
| 1425 |
|
| 1426 | mutex_unlock(&ep->mtx);
|
| 1427 |
|
| 1428 | if (ep_put(ep))
|
| 1429 | ep_free(ep);
|
| 1430 | goto again;
|
| 1431 | }
|
| 1432 | spin_unlock(&file->f_lock);
|
| 1433 | }
|
| 1434 |
|
| 1435 | static int ep_alloc(struct eventpoll **pep)
|
| 1436 | {
|
| 1437 | struct eventpoll *ep;
|
| 1438 |
|
| 1439 | ep = kzalloc_obj(*ep);
|
| 1440 | if (unlikely(!ep))
|
| 1441 | return -ENOMEM;
|
| 1442 |
|
| 1443 | mutex_init(&ep->mtx);
|
| 1444 | spin_lock_init(&ep->lock);
|
| 1445 | seqcount_spinlock_init(&ep->seq, &ep->lock);
|
| 1446 | init_waitqueue_head(&ep->wq);
|
| 1447 | init_waitqueue_head(&ep->poll_wait);
|
| 1448 | INIT_LIST_HEAD(&ep->rdllist);
|
| 1449 | ep->rbr = RB_ROOT_CACHED;
|
| 1450 | ep->ovflist = EP_UNACTIVE_PTR;
|
| 1451 | ep->user = get_current_user();
|
| 1452 | refcount_set(&ep->refcount, 1);
|
| 1453 |
|
| 1454 | *pep = ep;
|
| 1455 |
|
| 1456 | return 0;
|
| 1457 | }
|
| 1458 |
|
| 1459 | |
| 1460 | * Search the file inside the eventpoll tree. The RB tree operations
|
| 1461 | * are protected by the "mtx" mutex, and ep_find() must be called with
|
| 1462 | * "mtx" held.
|
| 1463 |
|
| 1464 | static struct epitem *ep_find(struct eventpoll *ep, struct epoll_key *tf)
|
| 1465 | {
|
| 1466 | int kcmp;
|
| 1467 | struct rb_node *rbp;
|
| 1468 | struct epitem *epi, *epir = NULL;
|
| 1469 |
|
| 1470 | for (rbp = ep->rbr.rb_root.rb_node; rbp; ) {
|
| 1471 | epi = rb_entry(rbp, struct epitem, rbn);
|
| 1472 | kcmp = ep_cmp_ffd(tf, &epi->ffd);
|
| 1473 | if (kcmp > 0)
|
| 1474 | rbp = rbp->rb_right;
|
| 1475 | else if (kcmp < 0)
|
| 1476 | rbp = rbp->rb_left;
|
| 1477 | else {
|
| 1478 | epir = epi;
|
| 1479 | break;
|
| 1480 | }
|
| 1481 | }
|
| 1482 |
|
| 1483 | return epir;
|
| 1484 | }
|
| 1485 |
|
| 1486 | |
| 1487 | * This is the callback that is passed to the wait queue wakeup
|
| 1488 | * mechanism. It is called by the stored file descriptors when they
|
| 1489 | * have events to report.
|
| 1490 |
|
| 1491 | static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
|
| 1492 | {
|
| 1493 | int pwake = 0;
|
| 1494 | struct epitem *epi = ep_item_from_wait(wait);
|
| 1495 | struct eventpoll *ep = epi->ep;
|
| 1496 | __poll_t pollflags = key_to_poll(key);
|
| 1497 | unsigned long flags;
|
| 1498 | int ewake = 0;
|
| 1499 |
|
| 1500 | spin_lock_irqsave(&ep->lock, flags);
|
| 1501 |
|
| 1502 | ep_set_busy_poll_napi_id(epi);
|
| 1503 |
|
| 1504 | |
| 1505 | * If the event mask does not contain any poll(2) event, we consider the
|
| 1506 | * descriptor to be disabled. This condition is likely the effect of the
|
| 1507 | * EPOLLONESHOT bit that disables the descriptor when an event is received,
|
| 1508 | * until the next EPOLL_CTL_MOD will be issued.
|
| 1509 |
|
| 1510 | if (!(epi->event.events & ~EP_PRIVATE_BITS))
|
| 1511 | goto out_unlock;
|
| 1512 |
|
| 1513 | |
| 1514 | * Check the events coming with the callback. At this stage, not
|
| 1515 | * every device reports the events in the "key" parameter of the
|
| 1516 | * callback. We need to be able to handle both cases here, hence the
|
| 1517 | * test for "key" != NULL before the event match test.
|
| 1518 |
|
| 1519 | if (pollflags && !(pollflags & epi->event.events))
|
| 1520 | goto out_unlock;
|
| 1521 |
|
| 1522 | |
| 1523 | * If we are transferring events to userspace, we can hold no locks
|
| 1524 | * (because we're accessing user memory, and because of linux f_op->poll()
|
| 1525 | * semantics). All the events that happen during that period of time are
|
| 1526 | * chained in ep->ovflist and requeued later on.
|
| 1527 |
|
| 1528 | if (ep_is_scanning(ep)) {
|
| 1529 | if (!epi_on_ovflist(epi)) {
|
| 1530 | epi->ovflist_next = READ_ONCE(ep->ovflist);
|
| 1531 | WRITE_ONCE(ep->ovflist, epi);
|
| 1532 | ep_pm_stay_awake_rcu(epi);
|
| 1533 | }
|
| 1534 | } else if (!ep_is_linked(epi)) {
|
| 1535 |
|
| 1536 | list_add_tail(&epi->rdllink, &ep->rdllist);
|
| 1537 | ep_pm_stay_awake_rcu(epi);
|
| 1538 | }
|
| 1539 |
|
| 1540 | |
| 1541 | * Wake up ( if active ) both the eventpoll wait list and the ->poll()
|
| 1542 | * wait list.
|
| 1543 |
|
| 1544 | if (waitqueue_active(&ep->wq)) {
|
| 1545 | if ((epi->event.events & EPOLLEXCLUSIVE) &&
|
| 1546 | !(pollflags & POLLFREE)) {
|
| 1547 | switch (pollflags & EPOLLINOUT_BITS) {
|
| 1548 | case EPOLLIN:
|
| 1549 | if (epi->event.events & EPOLLIN)
|
| 1550 | ewake = 1;
|
| 1551 | break;
|
| 1552 | case EPOLLOUT:
|
| 1553 | if (epi->event.events & EPOLLOUT)
|
| 1554 | ewake = 1;
|
| 1555 | break;
|
| 1556 | case 0:
|
| 1557 | ewake = 1;
|
| 1558 | break;
|
| 1559 | }
|
| 1560 | }
|
| 1561 | if (sync)
|
| 1562 | wake_up_sync(&ep->wq);
|
| 1563 | else
|
| 1564 | wake_up(&ep->wq);
|
| 1565 | }
|
| 1566 | if (waitqueue_active(&ep->poll_wait))
|
| 1567 | pwake++;
|
| 1568 |
|
| 1569 | out_unlock:
|
| 1570 | spin_unlock_irqrestore(&ep->lock, flags);
|
| 1571 |
|
| 1572 |
|
| 1573 | if (pwake)
|
| 1574 | ep_poll_safewake(ep, epi, pollflags & EPOLL_URING_WAKE);
|
| 1575 |
|
| 1576 | if (!(epi->event.events & EPOLLEXCLUSIVE))
|
| 1577 | ewake = 1;
|
| 1578 |
|
| 1579 | if (pollflags & POLLFREE) {
|
| 1580 | |
| 1581 | * POLLFREE handshake, release side; see "POLLFREE handshake"
|
| 1582 | * at the top of this file.
|
| 1583 | *
|
| 1584 | * Unlink our wait entry with list_del_init rather than
|
| 1585 | * __remove_wait_queue: a concurrent ep_remove_wait_queue()
|
| 1586 | * that already loaded a non-NULL whead may still call
|
| 1587 | * remove_wait_queue() after us, and list_del_init() tolerates
|
| 1588 | * the second delete.
|
| 1589 | *
|
| 1590 | * smp_store_release(&whead, NULL) publishes the teardown to
|
| 1591 | * ep_remove_wait_queue()'s smp_load_acquire(). Before this
|
| 1592 | * store, a racing ep_clear_and_put() / ep_remove() reaches
|
| 1593 | * ep_remove_wait_queue() which sees whead != NULL and takes
|
| 1594 | * whead->lock -- the same lock held by our caller, so it
|
| 1595 | * serializes behind us. Once whead is zeroed, nothing else
|
| 1596 | * protects ep / epi / wait.
|
| 1597 |
|
| 1598 | list_del_init(&wait->entry);
|
| 1599 | smp_store_release(&ep_pwq_from_wait(wait)->whead, NULL);
|
| 1600 | }
|
| 1601 |
|
| 1602 | return ewake;
|
| 1603 | }
|
| 1604 |
|
| 1605 | |
| 1606 | * This is the callback that is used to add our wait queue to the
|
| 1607 | * target file wakeup lists.
|
| 1608 |
|
| 1609 | static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead,
|
| 1610 | poll_table *pt)
|
| 1611 | {
|
| 1612 | struct ep_pqueue *epq = container_of(pt, struct ep_pqueue, pt);
|
| 1613 | struct epitem *epi = epq->epi;
|
| 1614 | struct eppoll_entry *pwq;
|
| 1615 |
|
| 1616 | if (unlikely(!epi)) |
| 1617 | return;
|
| 1618 |
|
| 1619 | pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL);
|
| 1620 | if (unlikely(!pwq)) {
|
| 1621 | epq->epi = NULL;
|
| 1622 | return;
|
| 1623 | }
|
| 1624 |
|
| 1625 | init_waitqueue_func_entry(&pwq->wait, ep_poll_callback);
|
| 1626 | pwq->whead = whead;
|
| 1627 | pwq->base = epi;
|
| 1628 | if (epi->event.events & EPOLLEXCLUSIVE)
|
| 1629 | add_wait_queue_exclusive(whead, &pwq->wait);
|
| 1630 | else
|
| 1631 | add_wait_queue(whead, &pwq->wait);
|
| 1632 | pwq->next = epi->pwqlist;
|
| 1633 | epi->pwqlist = pwq;
|
| 1634 | }
|
| 1635 |
|
| 1636 | static void ep_rbtree_insert(struct eventpoll *ep, struct epitem *epi)
|
| 1637 | {
|
| 1638 | int kcmp;
|
| 1639 | struct rb_node **p = &ep->rbr.rb_root.rb_node, *parent = NULL;
|
| 1640 | struct epitem *epic;
|
| 1641 | bool leftmost = true;
|
| 1642 |
|
| 1643 | while (*p) {
|
| 1644 | parent = *p;
|
| 1645 | epic = rb_entry(parent, struct epitem, rbn);
|
| 1646 | kcmp = ep_cmp_ffd(&epi->ffd, &epic->ffd);
|
| 1647 | if (kcmp > 0) {
|
| 1648 | p = &parent->rb_right;
|
| 1649 | leftmost = false;
|
| 1650 | } else
|
| 1651 | p = &parent->rb_left;
|
| 1652 | }
|
| 1653 | rb_link_node(&epi->rbn, parent, p);
|
| 1654 | rb_insert_color_cached(&epi->rbn, &ep->rbr, leftmost);
|
| 1655 | }
|
| 1656 |
|
| 1657 |
|
| 1658 |
|
| 1659 | |
| 1660 | * Upper bound on wakeup paths emanating from any one watched file,
|
| 1661 | * indexed by path depth (1..PATH_ARR_SIZE). For example, we allow
|
| 1662 | * 1000 paths of length 1 from each watched file. These caps limit
|
| 1663 | * the wakeup amplification that can be built from epoll-watches-
|
| 1664 | * epoll topologies without rejecting reasonable usage.
|
| 1665 | *
|
| 1666 | * Enforced at EPOLL_CTL_ADD; CTL_MOD and CTL_DEL cannot add paths.
|
| 1667 | * The running tallies live in ctx->path_count[] and are protected by
|
| 1668 | * epnested_mutex.
|
| 1669 |
|
| 1670 | static const int path_limits[PATH_ARR_SIZE] = { 1000, 500, 100, 50, 10 };
|
| 1671 |
|
| 1672 | static int path_count_inc(struct ep_ctl_ctx *ctx, int nests)
|
| 1673 | {
|
| 1674 |
|
| 1675 | if (nests == 0)
|
| 1676 | return 0;
|
| 1677 |
|
| 1678 | if (++ctx->path_count[nests] > path_limits[nests])
|
| 1679 | return -1;
|
| 1680 | return 0;
|
| 1681 | }
|
| 1682 |
|
| 1683 | static void path_count_init(struct ep_ctl_ctx *ctx)
|
| 1684 | {
|
| 1685 | int i;
|
| 1686 |
|
| 1687 | for (i = 0; i < PATH_ARR_SIZE; i++)
|
| 1688 | ctx->path_count[i] = 0;
|
| 1689 | }
|
| 1690 |
|
| 1691 | static int reverse_path_check_proc(struct ep_ctl_ctx *ctx,
|
| 1692 | struct hlist_head *refs, int depth)
|
| 1693 | {
|
| 1694 | int error = 0;
|
| 1695 | struct epitem *epi;
|
| 1696 |
|
| 1697 | if (depth > EP_MAX_NESTS)
|
| 1698 | return -1;
|
| 1699 |
|
| 1700 |
|
| 1701 | hlist_for_each_entry_rcu(epi, refs, fllink) {
|
| 1702 | struct hlist_head *refs = &epi->ep->refs;
|
| 1703 | if (hlist_empty(refs))
|
| 1704 | error = path_count_inc(ctx, depth);
|
| 1705 | else
|
| 1706 | error = reverse_path_check_proc(ctx, refs, depth + 1);
|
| 1707 | if (error != 0)
|
| 1708 | break;
|
| 1709 | }
|
| 1710 | return error;
|
| 1711 | }
|
| 1712 |
|
| 1713 | |
| 1714 | * reverse_path_check - ctx->tfile_check_list is a list of epitems_head
|
| 1715 | * anchoring files with newly proposed links; make
|
| 1716 | * sure those links don't push any path-length bucket
|
| 1717 | * over its limit in path_limits[].
|
| 1718 | * @ctx: Per-do_epoll_ctl() scratch for the loop / path checks.
|
| 1719 | *
|
| 1720 | * Return: %zero if the proposed links don't create too many paths,
|
| 1721 | * %-1 otherwise.
|
| 1722 |
|
| 1723 | static int reverse_path_check(struct ep_ctl_ctx *ctx)
|
| 1724 | {
|
| 1725 | struct epitems_head *p;
|
| 1726 |
|
| 1727 | for (p = ctx->tfile_check_list; p != EP_UNACTIVE_PTR; p = p->next) {
|
| 1728 | int error;
|
| 1729 | path_count_init(ctx);
|
| 1730 | rcu_read_lock();
|
| 1731 | error = reverse_path_check_proc(ctx, &p->epitems, 0);
|
| 1732 | rcu_read_unlock();
|
| 1733 | if (error)
|
| 1734 | return error;
|
| 1735 | }
|
| 1736 | return 0;
|
| 1737 | }
|
| 1738 |
|
| 1739 | static int ep_create_wakeup_source(struct epitem *epi)
|
| 1740 | {
|
| 1741 | struct name_snapshot n;
|
| 1742 | struct wakeup_source *ws;
|
| 1743 |
|
| 1744 | if (!epi->ep->ws) {
|
| 1745 | epi->ep->ws = wakeup_source_register(NULL, "eventpoll");
|
| 1746 | if (!epi->ep->ws)
|
| 1747 | return -ENOMEM;
|
| 1748 | }
|
| 1749 |
|
| 1750 | take_dentry_name_snapshot(&n, epi->ffd.file->f_path.dentry);
|
| 1751 | ws = wakeup_source_register(NULL, n.name.name);
|
| 1752 | release_dentry_name_snapshot(&n);
|
| 1753 |
|
| 1754 | if (!ws)
|
| 1755 | return -ENOMEM;
|
| 1756 | rcu_assign_pointer(epi->ws, ws);
|
| 1757 |
|
| 1758 | return 0;
|
| 1759 | }
|
| 1760 |
|
| 1761 |
|
| 1762 | static noinline void ep_destroy_wakeup_source(struct epitem *epi)
|
| 1763 | {
|
| 1764 | struct wakeup_source *ws = ep_wakeup_source(epi);
|
| 1765 |
|
| 1766 | RCU_INIT_POINTER(epi->ws, NULL);
|
| 1767 |
|
| 1768 | |
| 1769 | * wait for ep_pm_stay_awake_rcu to finish, synchronize_rcu is
|
| 1770 | * used internally by wakeup_source_remove, too (called by
|
| 1771 | * wakeup_source_unregister), so we cannot use call_rcu
|
| 1772 |
|
| 1773 | synchronize_rcu();
|
| 1774 | wakeup_source_unregister(ws);
|
| 1775 | }
|
| 1776 |
|
| 1777 | static int ep_attach_file(struct file *file, struct epitem *epi)
|
| 1778 | {
|
| 1779 | struct epitems_head *to_free = NULL;
|
| 1780 | struct hlist_head *head = NULL;
|
| 1781 | struct eventpoll *ep = NULL;
|
| 1782 |
|
| 1783 | if (is_file_epoll(file))
|
| 1784 | ep = file->private_data;
|
| 1785 |
|
| 1786 | if (ep) {
|
| 1787 | head = &ep->refs;
|
| 1788 | } else if (!READ_ONCE(file->f_ep)) {
|
| 1789 | allocate:
|
| 1790 | to_free = kmem_cache_zalloc(ephead_cache, GFP_KERNEL);
|
| 1791 | if (!to_free)
|
| 1792 | return -ENOMEM;
|
| 1793 | head = &to_free->epitems;
|
| 1794 | }
|
| 1795 | spin_lock(&file->f_lock);
|
| 1796 | if (!file->f_ep) {
|
| 1797 | if (unlikely(!head)) {
|
| 1798 | spin_unlock(&file->f_lock);
|
| 1799 | goto allocate;
|
| 1800 | }
|
| 1801 |
|
| 1802 | WRITE_ONCE(file->f_ep, head);
|
| 1803 | to_free = NULL;
|
| 1804 | }
|
| 1805 | hlist_add_head_rcu(&epi->fllink, file->f_ep);
|
| 1806 | spin_unlock(&file->f_lock);
|
| 1807 | free_ephead(to_free);
|
| 1808 | return 0;
|
| 1809 | }
|
| 1810 |
|
| 1811 | |
| 1812 | * Charge the user's epoll_watches quota, allocate a fresh epitem for
|
| 1813 | * @tf, and initialize its fields. The returned item is not yet linked
|
| 1814 | * into any data structure; the caller must install it via
|
| 1815 | * ep_register_epitem() (which takes over on success) or kmem_cache_free()
|
| 1816 | * it and decrement epoll_watches on its own.
|
| 1817 | *
|
| 1818 | * Returns ERR_PTR(-ENOSPC) if the quota is exceeded, ERR_PTR(-ENOMEM)
|
| 1819 | * if the slab allocation fails.
|
| 1820 |
|
| 1821 | static struct epitem *ep_alloc_epitem(struct eventpoll *ep,
|
| 1822 | const struct epoll_event *event,
|
| 1823 | struct epoll_key *tf)
|
| 1824 | {
|
| 1825 | struct epitem *epi;
|
| 1826 |
|
| 1827 | if (unlikely(percpu_counter_compare(&ep->user->epoll_watches,
|
| 1828 | max_user_watches) >= 0))
|
| 1829 | return ERR_PTR(-ENOSPC);
|
| 1830 | percpu_counter_inc(&ep->user->epoll_watches);
|
| 1831 |
|
| 1832 | epi = kmem_cache_zalloc(epi_cache, GFP_KERNEL);
|
| 1833 | if (unlikely(!epi)) {
|
| 1834 | percpu_counter_dec(&ep->user->epoll_watches);
|
| 1835 | return ERR_PTR(-ENOMEM);
|
| 1836 | }
|
| 1837 |
|
| 1838 | INIT_LIST_HEAD(&epi->rdllink);
|
| 1839 | epi->ep = ep;
|
| 1840 | epi->ffd = *tf;
|
| 1841 | epi->event = *event;
|
| 1842 | epi_clear_ovflist(epi);
|
| 1843 |
|
| 1844 | return epi;
|
| 1845 | }
|
| 1846 |
|
| 1847 | |
| 1848 | * Install @epi into its target file's f_ep hlist and into @ep's rbtree,
|
| 1849 | * taking one additional reference on @ep for the lifetime of the item.
|
| 1850 | *
|
| 1851 | * If @tep is non-NULL, the target file is itself an eventpoll; we hold
|
| 1852 | * tep->mtx at subclass 1 across the attach + rbtree insert to serialize
|
| 1853 | * with the target side. RB tree ops are protected by @ep->mtx, which
|
| 1854 | * the caller already holds.
|
| 1855 | *
|
| 1856 | * On failure the epi is freed and the epoll_watches counter decremented,
|
| 1857 | * matching ep_alloc_epitem()'s allocation. After this returns
|
| 1858 | * successfully, ep_insert()'s later error paths use ep_remove() for
|
| 1859 | * unwind; that cannot drop @ep's refcount to zero because the ep file
|
| 1860 | * itself still holds the original reference.
|
| 1861 |
|
| 1862 | static int ep_register_epitem(struct ep_ctl_ctx *ctx, struct eventpoll *ep,
|
| 1863 | struct epitem *epi, struct eventpoll *tep,
|
| 1864 | int full_check)
|
| 1865 | {
|
| 1866 | struct file *tfile = epi->ffd.file;
|
| 1867 | int error;
|
| 1868 |
|
| 1869 | if (tep)
|
| 1870 | mutex_lock_nested(&tep->mtx, 1);
|
| 1871 |
|
| 1872 | error = ep_attach_file(tfile, epi);
|
| 1873 | if (unlikely(error)) {
|
| 1874 | if (tep)
|
| 1875 | mutex_unlock(&tep->mtx);
|
| 1876 | kmem_cache_free(epi_cache, epi);
|
| 1877 | percpu_counter_dec(&ep->user->epoll_watches);
|
| 1878 | return error;
|
| 1879 | }
|
| 1880 |
|
| 1881 | if (full_check && !tep)
|
| 1882 | list_file(tfile, ctx);
|
| 1883 |
|
| 1884 | ep_rbtree_insert(ep, epi);
|
| 1885 |
|
| 1886 | if (tep)
|
| 1887 | mutex_unlock(&tep->mtx);
|
| 1888 |
|
| 1889 | ep_get(ep);
|
| 1890 | return 0;
|
| 1891 | }
|
| 1892 |
|
| 1893 | |
| 1894 | * Must be called with "mtx" held.
|
| 1895 |
|
| 1896 | static int ep_insert(struct ep_ctl_ctx *ctx, struct eventpoll *ep,
|
| 1897 | const struct epoll_event *event, struct epoll_key *tf,
|
| 1898 | int full_check)
|
| 1899 | {
|
| 1900 | int error, pwake = 0;
|
| 1901 | __poll_t revents;
|
| 1902 | struct epitem *epi;
|
| 1903 | struct ep_pqueue epq;
|
| 1904 | struct eventpoll *tep = NULL;
|
| 1905 |
|
| 1906 | if (is_file_epoll(tf->file))
|
| 1907 | tep = tf->file->private_data;
|
| 1908 |
|
| 1909 | lockdep_assert_irqs_enabled();
|
| 1910 |
|
| 1911 | epi = ep_alloc_epitem(ep, event, tf);
|
| 1912 | if (IS_ERR(epi))
|
| 1913 | return PTR_ERR(epi);
|
| 1914 |
|
| 1915 | error = ep_register_epitem(ctx, ep, epi, tep, full_check);
|
| 1916 | if (error)
|
| 1917 | return error;
|
| 1918 |
|
| 1919 |
|
| 1920 | if (unlikely(full_check && reverse_path_check(ctx))) {
|
| 1921 | ep_remove(ep, epi);
|
| 1922 | return -EINVAL;
|
| 1923 | }
|
| 1924 |
|
| 1925 | if (epi->event.events & EPOLLWAKEUP) {
|
| 1926 | error = ep_create_wakeup_source(epi);
|
| 1927 | if (error) {
|
| 1928 | ep_remove(ep, epi);
|
| 1929 | return error;
|
| 1930 | }
|
| 1931 | }
|
| 1932 |
|
| 1933 |
|
| 1934 | epq.epi = epi;
|
| 1935 | init_poll_funcptr(&epq.pt, ep_ptable_queue_proc);
|
| 1936 |
|
| 1937 | |
| 1938 | * Attach the item to the poll hooks and get current event bits.
|
| 1939 | * We can safely use the file* here because its usage count has
|
| 1940 | * been increased by the caller of this function. Note that after
|
| 1941 | * this operation completes, the poll callback can start hitting
|
| 1942 | * the new item.
|
| 1943 |
|
| 1944 | revents = ep_item_poll(epi, &epq.pt, 1);
|
| 1945 |
|
| 1946 |
|
| 1947 | if (unlikely(!epq.epi)) {
|
| 1948 | ep_remove(ep, epi);
|
| 1949 | return -ENOMEM;
|
| 1950 | }
|
| 1951 |
|
| 1952 |
|
| 1953 | spin_lock_irq(&ep->lock);
|
| 1954 |
|
| 1955 | ep_set_busy_poll_napi_id(epi);
|
| 1956 |
|
| 1957 | if (revents && !ep_is_linked(epi)) {
|
| 1958 | list_add_tail(&epi->rdllink, &ep->rdllist);
|
| 1959 | ep_pm_stay_awake(epi);
|
| 1960 |
|
| 1961 | if (waitqueue_active(&ep->wq))
|
| 1962 | wake_up(&ep->wq);
|
| 1963 | if (waitqueue_active(&ep->poll_wait))
|
| 1964 | pwake++;
|
| 1965 | }
|
| 1966 |
|
| 1967 | spin_unlock_irq(&ep->lock);
|
| 1968 |
|
| 1969 |
|
| 1970 | if (pwake)
|
| 1971 | ep_poll_safewake(ep, NULL, 0);
|
| 1972 |
|
| 1973 | return 0;
|
| 1974 | }
|
| 1975 |
|
| 1976 | |
| 1977 | * Modify the interest event mask by dropping an event if the new mask
|
| 1978 | * has a match in the current file status. Must be called with "mtx" held.
|
| 1979 |
|
| 1980 | static int ep_modify(struct eventpoll *ep, struct epitem *epi,
|
| 1981 | const struct epoll_event *event)
|
| 1982 | {
|
| 1983 | int pwake = 0;
|
| 1984 | poll_table pt;
|
| 1985 |
|
| 1986 | lockdep_assert_irqs_enabled();
|
| 1987 |
|
| 1988 | init_poll_funcptr(&pt, NULL);
|
| 1989 |
|
| 1990 | |
| 1991 | * Set the new event interest mask before calling f_op->poll();
|
| 1992 | * otherwise we might miss an event that happens between the
|
| 1993 | * f_op->poll() call and the new event set registering.
|
| 1994 |
|
| 1995 | epi->event.events = event->events;
|
| 1996 | epi->event.data = event->data;
|
| 1997 | if (epi->event.events & EPOLLWAKEUP) {
|
| 1998 | if (!ep_has_wakeup_source(epi))
|
| 1999 | ep_create_wakeup_source(epi);
|
| 2000 | } else if (ep_has_wakeup_source(epi)) {
|
| 2001 | ep_destroy_wakeup_source(epi);
|
| 2002 | }
|
| 2003 |
|
| 2004 | |
| 2005 | * The following barrier has two effects:
|
| 2006 | *
|
| 2007 | * 1) Flush epi changes above to other CPUs. This ensures
|
| 2008 | * we do not miss events from ep_poll_callback if an
|
| 2009 | * event occurs immediately after we call f_op->poll().
|
| 2010 | * We need this because we did not take ep->lock while
|
| 2011 | * changing epi above (but ep_poll_callback does take
|
| 2012 | * ep->lock).
|
| 2013 | *
|
| 2014 | * 2) We also need to ensure we do not miss _past_ events
|
| 2015 | * when calling f_op->poll(). This barrier also
|
| 2016 | * pairs with the barrier in wq_has_sleeper (see
|
| 2017 | * comments for wq_has_sleeper).
|
| 2018 | *
|
| 2019 | * This barrier will now guarantee ep_poll_callback or f_op->poll
|
| 2020 | * (or both) will notice the readiness of an item.
|
| 2021 |
|
| 2022 | smp_mb();
|
| 2023 |
|
| 2024 | |
| 2025 | * Get current event bits. We can safely use the file* here because
|
| 2026 | * its usage count has been increased by the caller of this function.
|
| 2027 | * If the item is "hot" and it is not registered inside the ready
|
| 2028 | * list, push it inside.
|
| 2029 |
|
| 2030 | if (ep_item_poll(epi, &pt, 1)) {
|
| 2031 | spin_lock_irq(&ep->lock);
|
| 2032 | if (!ep_is_linked(epi)) {
|
| 2033 | list_add_tail(&epi->rdllink, &ep->rdllist);
|
| 2034 | ep_pm_stay_awake(epi);
|
| 2035 |
|
| 2036 |
|
| 2037 | if (waitqueue_active(&ep->wq))
|
| 2038 | wake_up(&ep->wq);
|
| 2039 | if (waitqueue_active(&ep->poll_wait))
|
| 2040 | pwake++;
|
| 2041 | }
|
| 2042 | spin_unlock_irq(&ep->lock);
|
| 2043 | }
|
| 2044 |
|
| 2045 |
|
| 2046 | if (pwake)
|
| 2047 | ep_poll_safewake(ep, NULL, 0);
|
| 2048 |
|
| 2049 | return 0;
|
| 2050 | }
|
| 2051 |
|
| 2052 | |
| 2053 | * Attempt to deliver one event for @epi into @*uevents.
|
| 2054 | *
|
| 2055 | * Returns 1 if an event was delivered (with *uevents advanced to the
|
| 2056 | * next slot), 0 if the re-poll reported no caller-requested events
|
| 2057 | * (@epi drops out of the ready list; a future callback will re-add
|
| 2058 | * it), or -EFAULT if copy_to_user() faulted (in which case @epi is
|
| 2059 | * re-inserted at the head of @scan_batch so ep_done_scan() merges it
|
| 2060 | * back to rdllist for the next attempt).
|
| 2061 | *
|
| 2062 | * PM bookkeeping and level-triggered re-queue are handled here.
|
| 2063 | * Caller holds ep->mtx and the scan is active.
|
| 2064 |
|
| 2065 | static int ep_deliver_event(struct eventpoll *ep, struct epitem *epi,
|
| 2066 | poll_table *pt,
|
| 2067 | struct epoll_event __user **uevents,
|
| 2068 | struct list_head *scan_batch)
|
| 2069 | {
|
| 2070 | struct epoll_event __user *next;
|
| 2071 | struct wakeup_source *ws;
|
| 2072 | __poll_t revents;
|
| 2073 |
|
| 2074 | |
| 2075 | * Activate ep->ws before deactivating epi->ws to prevent
|
| 2076 | * triggering auto-suspend here (in case we reactivate epi->ws
|
| 2077 | * below). Rearranging to delay the deactivation would let
|
| 2078 | * epi->ws drift out of sync with ep_is_linked().
|
| 2079 |
|
| 2080 | ws = ep_wakeup_source(epi);
|
| 2081 | if (ws) {
|
| 2082 | if (ws->active)
|
| 2083 | __pm_stay_awake(ep->ws);
|
| 2084 | __pm_relax(ws);
|
| 2085 | }
|
| 2086 |
|
| 2087 | list_del_init(&epi->rdllink);
|
| 2088 |
|
| 2089 | |
| 2090 | * Re-poll under ep->mtx so userspace cannot change the item
|
| 2091 | * out from under us. If no caller-requested events remain,
|
| 2092 | * @epi stays off the ready list; the poll callback will
|
| 2093 | * re-queue it when events next appear.
|
| 2094 |
|
| 2095 | revents = ep_item_poll(epi, pt, 1);
|
| 2096 | if (!revents)
|
| 2097 | return 0;
|
| 2098 |
|
| 2099 | next = epoll_put_uevent(revents, epi->event.data, *uevents);
|
| 2100 | if (!next) {
|
| 2101 | |
| 2102 | * copy_to_user() faulted: put the item back so
|
| 2103 | * ep_done_scan() splices it onto rdllist for the next
|
| 2104 | * attempt.
|
| 2105 |
|
| 2106 | list_add(&epi->rdllink, scan_batch);
|
| 2107 | ep_pm_stay_awake(epi);
|
| 2108 | return -EFAULT;
|
| 2109 | }
|
| 2110 | *uevents = next;
|
| 2111 |
|
| 2112 | if (epi->event.events & EPOLLONESHOT) {
|
| 2113 | epi->event.events &= EP_PRIVATE_BITS;
|
| 2114 | } else if (!(epi->event.events & EPOLLET)) {
|
| 2115 | |
| 2116 | * Level-triggered: re-queue so the next epoll_wait()
|
| 2117 | * rechecks availability. We are the sole writer to
|
| 2118 | * rdllist here -- epoll_ctl() callers are locked out
|
| 2119 | * by ep->mtx, and the poll callback queues to ovflist
|
| 2120 | * during scans.
|
| 2121 |
|
| 2122 | list_add_tail(&epi->rdllink, &ep->rdllist);
|
| 2123 | ep_pm_stay_awake(epi);
|
| 2124 | }
|
| 2125 | return 1;
|
| 2126 | }
|
| 2127 |
|
| 2128 | static int ep_send_events(struct eventpoll *ep,
|
| 2129 | struct epoll_event __user *events, int maxevents)
|
| 2130 | {
|
| 2131 | struct epitem *epi, *tmp;
|
| 2132 | LIST_HEAD(scan_batch);
|
| 2133 | poll_table pt;
|
| 2134 | int res = 0;
|
| 2135 |
|
| 2136 | |
| 2137 | * Always short-circuit for fatal signals to allow threads to make a
|
| 2138 | * timely exit without the chance of finding more events available and
|
| 2139 | * fetching repeatedly.
|
| 2140 |
|
| 2141 | if (fatal_signal_pending(current))
|
| 2142 | return -EINTR;
|
| 2143 |
|
| 2144 | init_poll_funcptr(&pt, NULL);
|
| 2145 |
|
| 2146 | mutex_lock(&ep->mtx);
|
| 2147 | ep_start_scan(ep, &scan_batch);
|
| 2148 |
|
| 2149 | |
| 2150 | * We can loop without lock because we are passed a task-private
|
| 2151 | * scan_batch; items cannot vanish while we hold ep->mtx.
|
| 2152 |
|
| 2153 | list_for_each_entry_safe(epi, tmp, &scan_batch, rdllink) {
|
| 2154 | int delivered;
|
| 2155 |
|
| 2156 | if (res >= maxevents)
|
| 2157 | break;
|
| 2158 |
|
| 2159 | delivered = ep_deliver_event(ep, epi, &pt, &events, &scan_batch);
|
| 2160 | if (delivered < 0) {
|
| 2161 | if (!res)
|
| 2162 | res = delivered;
|
| 2163 | break;
|
| 2164 | }
|
| 2165 | res += delivered;
|
| 2166 | }
|
| 2167 |
|
| 2168 | ep_done_scan(ep, &scan_batch);
|
| 2169 | mutex_unlock(&ep->mtx);
|
| 2170 |
|
| 2171 | return res;
|
| 2172 | }
|
| 2173 |
|
| 2174 | static struct timespec64 *ep_timeout_to_timespec(struct timespec64 *to, long ms)
|
| 2175 | {
|
| 2176 | struct timespec64 now;
|
| 2177 |
|
| 2178 | if (ms < 0)
|
| 2179 | return NULL;
|
| 2180 |
|
| 2181 | if (!ms) {
|
| 2182 | to->tv_sec = 0;
|
| 2183 | to->tv_nsec = 0;
|
| 2184 | return to;
|
| 2185 | }
|
| 2186 |
|
| 2187 | to->tv_sec = ms / MSEC_PER_SEC;
|
| 2188 | to->tv_nsec = NSEC_PER_MSEC * (ms % MSEC_PER_SEC);
|
| 2189 |
|
| 2190 | ktime_get_ts64(&now);
|
| 2191 | *to = timespec64_add_safe(now, *to);
|
| 2192 | return to;
|
| 2193 | }
|
| 2194 |
|
| 2195 | |
| 2196 | * autoremove_wake_function, but remove even on failure to wake up, because we
|
| 2197 | * know that default_wake_function/ttwu will only fail if the thread is already
|
| 2198 | * woken, and in that case the ep_poll loop will remove the entry anyways, not
|
| 2199 | * try to reuse it.
|
| 2200 |
|
| 2201 | static int ep_autoremove_wake_function(struct wait_queue_entry *wq_entry,
|
| 2202 | unsigned int mode, int sync, void *key)
|
| 2203 | {
|
| 2204 | int ret = default_wake_function(wq_entry, mode, sync, key);
|
| 2205 |
|
| 2206 | |
| 2207 | * Pairs with list_empty_careful in ep_poll, and ensures future loop
|
| 2208 | * iterations see the cause of this wakeup.
|
| 2209 |
|
| 2210 | list_del_init_careful(&wq_entry->entry);
|
| 2211 | return ret;
|
| 2212 | }
|
| 2213 |
|
| 2214 | static int ep_try_send_events(struct eventpoll *ep,
|
| 2215 | struct epoll_event __user *events, int maxevents)
|
| 2216 | {
|
| 2217 | int res;
|
| 2218 |
|
| 2219 | |
| 2220 | * Try to transfer events to user space. In case we get 0 events and
|
| 2221 | * there's still timeout left over, we go trying again in search of
|
| 2222 | * more luck.
|
| 2223 |
|
| 2224 | res = ep_send_events(ep, events, maxevents);
|
| 2225 | if (res > 0)
|
| 2226 | ep_suspend_napi_irqs(ep);
|
| 2227 | return res;
|
| 2228 | }
|
| 2229 |
|
| 2230 | static int ep_schedule_timeout(ktime_t *to)
|
| 2231 | {
|
| 2232 | if (to)
|
| 2233 | return ktime_after(*to, ktime_get());
|
| 2234 | else
|
| 2235 | return 1;
|
| 2236 | }
|
| 2237 |
|
| 2238 | |
| 2239 | * ep_poll - Retrieves ready events, and delivers them to the caller-supplied
|
| 2240 | * event buffer.
|
| 2241 | *
|
| 2242 | * @ep: Pointer to the eventpoll context.
|
| 2243 | * @events: Pointer to the userspace buffer where the ready events should be
|
| 2244 | * stored.
|
| 2245 | * @maxevents: Size (in terms of number of events) of the caller event buffer.
|
| 2246 | * @timeout: Maximum timeout for the ready events fetch operation, in
|
| 2247 | * timespec. If the timeout is zero, the function will not block,
|
| 2248 | * while if the @timeout ptr is NULL, the function will block
|
| 2249 | * until at least one event has been retrieved (or an error
|
| 2250 | * occurred).
|
| 2251 | *
|
| 2252 | * Return: the number of ready events which have been fetched, or an
|
| 2253 | * error code, in case of error.
|
| 2254 |
|
| 2255 | static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
|
| 2256 | int maxevents, struct timespec64 *timeout)
|
| 2257 | {
|
| 2258 | int res, timed_out = 0;
|
| 2259 | bool eavail;
|
| 2260 | u64 slack = 0;
|
| 2261 | wait_queue_entry_t wait;
|
| 2262 | ktime_t expires, *to = NULL;
|
| 2263 |
|
| 2264 | lockdep_assert_irqs_enabled();
|
| 2265 |
|
| 2266 | if (timeout && (timeout->tv_sec | timeout->tv_nsec)) {
|
| 2267 | to = &expires;
|
| 2268 | *to = timespec64_to_ktime(*timeout);
|
| 2269 | } else if (timeout) {
|
| 2270 | |
| 2271 | * Avoid the unnecessary trip to the wait queue loop, if the
|
| 2272 | * caller specified a non blocking operation.
|
| 2273 |
|
| 2274 | timed_out = 1;
|
| 2275 | }
|
| 2276 |
|
| 2277 | |
| 2278 | * This call is racy: We may or may not see events that are being added
|
| 2279 | * to the ready list under the lock (e.g., in IRQ callbacks). For cases
|
| 2280 | * with a non-zero timeout, this thread will check the ready list under
|
| 2281 | * lock and will add to the wait queue. For cases with a zero
|
| 2282 | * timeout, the user by definition should not care and will have to
|
| 2283 | * recheck again.
|
| 2284 |
|
| 2285 | eavail = ep_events_available(ep);
|
| 2286 |
|
| 2287 | while (1) {
|
| 2288 | if (eavail) {
|
| 2289 | res = ep_try_send_events(ep, events, maxevents);
|
| 2290 | if (res)
|
| 2291 | return res;
|
| 2292 | }
|
| 2293 |
|
| 2294 | if (timed_out)
|
| 2295 | return 0;
|
| 2296 |
|
| 2297 | eavail = ep_busy_loop(ep);
|
| 2298 | if (eavail)
|
| 2299 | continue;
|
| 2300 |
|
| 2301 | if (signal_pending(current))
|
| 2302 | return -EINTR;
|
| 2303 |
|
| 2304 | |
| 2305 | * Internally init_wait() uses autoremove_wake_function(),
|
| 2306 | * thus wait entry is removed from the wait queue on each
|
| 2307 | * wakeup. Why it is important? In case of several waiters
|
| 2308 | * each new wakeup will hit the next waiter, giving it the
|
| 2309 | * chance to harvest new event. Otherwise wakeup can be
|
| 2310 | * lost. This is also good performance-wise, because on
|
| 2311 | * normal wakeup path no need to call __remove_wait_queue()
|
| 2312 | * explicitly, thus ep->lock is not taken, which halts the
|
| 2313 | * event delivery.
|
| 2314 | *
|
| 2315 | * In fact, we now use an even more aggressive function that
|
| 2316 | * unconditionally removes, because we don't reuse the wait
|
| 2317 | * entry between loop iterations. This lets us also avoid the
|
| 2318 | * performance issue if a process is killed, causing all of its
|
| 2319 | * threads to wake up without being removed normally.
|
| 2320 |
|
| 2321 | init_wait(&wait);
|
| 2322 | wait.func = ep_autoremove_wake_function;
|
| 2323 |
|
| 2324 | spin_lock_irq(&ep->lock);
|
| 2325 | |
| 2326 | * Barrierless variant, waitqueue_active() is called under
|
| 2327 | * the same lock on wakeup ep_poll_callback() side, so it
|
| 2328 | * is safe to avoid an explicit barrier.
|
| 2329 |
|
| 2330 | __set_current_state(TASK_INTERRUPTIBLE);
|
| 2331 |
|
| 2332 | |
| 2333 | * Do the final check under the lock. ep_start/done_scan()
|
| 2334 | * plays with two lists (->rdllist and ->ovflist) and there
|
| 2335 | * is always a race when both lists are empty for short
|
| 2336 | * period of time although events are pending, so lock is
|
| 2337 | * important.
|
| 2338 |
|
| 2339 | eavail = ep_events_available(ep);
|
| 2340 | if (!eavail)
|
| 2341 | __add_wait_queue_exclusive(&ep->wq, &wait);
|
| 2342 |
|
| 2343 | spin_unlock_irq(&ep->lock);
|
| 2344 |
|
| 2345 | if (!eavail) {
|
| 2346 | if (to)
|
| 2347 | slack = select_estimate_accuracy(timeout);
|
| 2348 | timed_out = !ep_schedule_timeout(to) ||
|
| 2349 | !schedule_hrtimeout_range(to, slack,
|
| 2350 | HRTIMER_MODE_ABS);
|
| 2351 | }
|
| 2352 | __set_current_state(TASK_RUNNING);
|
| 2353 |
|
| 2354 | |
| 2355 | * We were woken up, thus go and try to harvest some events.
|
| 2356 | * If timed out and still on the wait queue, recheck eavail
|
| 2357 | * carefully under lock, below.
|
| 2358 |
|
| 2359 | eavail = true;
|
| 2360 |
|
| 2361 | if (!list_empty_careful(&wait.entry)) {
|
| 2362 | spin_lock_irq(&ep->lock);
|
| 2363 | |
| 2364 | * If the thread timed out and is not on the wait queue,
|
| 2365 | * it means that the thread was woken up after its
|
| 2366 | * timeout expired before it could reacquire the lock.
|
| 2367 | * Thus, when wait.entry is empty, it needs to harvest
|
| 2368 | * events.
|
| 2369 |
|
| 2370 | if (timed_out)
|
| 2371 | eavail = list_empty(&wait.entry);
|
| 2372 | __remove_wait_queue(&ep->wq, &wait);
|
| 2373 | spin_unlock_irq(&ep->lock);
|
| 2374 | }
|
| 2375 | }
|
| 2376 | }
|
| 2377 |
|
| 2378 | |
| 2379 | * ep_loop_check_proc - verify that adding an epoll file @ep inside another
|
| 2380 | * epoll file does not create closed loops, and
|
| 2381 | * determine the depth of the subtree starting at @ep
|
| 2382 | *
|
| 2383 | * @ctx: Per-do_epoll_ctl() scratch for the loop / path checks.
|
| 2384 | * @ep: the &struct eventpoll to be currently checked.
|
| 2385 | * @depth: Current depth of the path being checked.
|
| 2386 | *
|
| 2387 | * Return: depth of the subtree, or a value bigger than EP_MAX_NESTS if we found
|
| 2388 | * a loop or went too deep.
|
| 2389 |
|
| 2390 | static int ep_loop_check_proc(struct ep_ctl_ctx *ctx,
|
| 2391 | struct eventpoll *ep, int depth)
|
| 2392 | {
|
| 2393 | int result = 0;
|
| 2394 | struct rb_node *rbp;
|
| 2395 | struct epitem *epi;
|
| 2396 |
|
| 2397 | if (ep->gen == loop_check_gen)
|
| 2398 | return ep->loop_check_depth;
|
| 2399 |
|
| 2400 | mutex_lock_nested(&ep->mtx, depth + 1);
|
| 2401 | ep->gen = loop_check_gen;
|
| 2402 | for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
|
| 2403 | epi = rb_entry(rbp, struct epitem, rbn);
|
| 2404 | if (unlikely(is_file_epoll(epi->ffd.file))) {
|
| 2405 | struct eventpoll *ep_tovisit;
|
| 2406 | ep_tovisit = epi->ffd.file->private_data;
|
| 2407 | if (ep_tovisit == ctx->inserting_into ||
|
| 2408 | depth > EP_MAX_NESTS)
|
| 2409 | result = EP_MAX_NESTS+1;
|
| 2410 | else
|
| 2411 | result = max(result,
|
| 2412 | ep_loop_check_proc(ctx, ep_tovisit,
|
| 2413 | depth + 1) + 1);
|
| 2414 | if (result > EP_MAX_NESTS)
|
| 2415 | break;
|
| 2416 | } else {
|
| 2417 | |
| 2418 | * A non-epoll leaf. Queue it for the companion
|
| 2419 | * reverse_path_check() that runs after this walk so
|
| 2420 | * any new links we propose don't add too many wakeup
|
| 2421 | * paths.
|
| 2422 |
|
| 2423 | list_file(epi->ffd.file, ctx);
|
| 2424 | }
|
| 2425 | }
|
| 2426 | ep->loop_check_depth = result;
|
| 2427 | mutex_unlock(&ep->mtx);
|
| 2428 |
|
| 2429 | return result;
|
| 2430 | }
|
| 2431 |
|
| 2432 |
|
| 2433 | static int ep_get_upwards_depth_proc(struct eventpoll *ep, int depth)
|
| 2434 | {
|
| 2435 | int result = 0;
|
| 2436 | struct epitem *epi;
|
| 2437 |
|
| 2438 | if (ep->gen == loop_check_gen)
|
| 2439 | return ep->loop_check_depth;
|
| 2440 | hlist_for_each_entry_rcu(epi, &ep->refs, fllink)
|
| 2441 | result = max(result, ep_get_upwards_depth_proc(epi->ep, depth + 1) + 1);
|
| 2442 | ep->gen = loop_check_gen;
|
| 2443 | ep->loop_check_depth = result;
|
| 2444 | return result;
|
| 2445 | }
|
| 2446 |
|
| 2447 | |
| 2448 | * ep_loop_check - Performs a check to verify that adding an epoll file (@to)
|
| 2449 | * into another epoll file (represented by @ep) does not create
|
| 2450 | * closed loops or too deep chains.
|
| 2451 | *
|
| 2452 | * @ctx: Per-CTL_ADD scratch context.
|
| 2453 | * @ep: Pointer to the epoll we are inserting into.
|
| 2454 | * @to: Pointer to the epoll to be inserted.
|
| 2455 | *
|
| 2456 | * Return: %zero if adding the epoll @to inside the epoll @from
|
| 2457 | * does not violate the constraints, or %-1 otherwise.
|
| 2458 |
|
| 2459 | static int ep_loop_check(struct ep_ctl_ctx *ctx, struct eventpoll *ep,
|
| 2460 | struct eventpoll *to)
|
| 2461 | {
|
| 2462 | int depth, upwards_depth;
|
| 2463 |
|
| 2464 | ctx->inserting_into = ep;
|
| 2465 | |
| 2466 | * Check how deep down we can get from @to, and whether it is possible
|
| 2467 | * to loop up to @ep.
|
| 2468 |
|
| 2469 | depth = ep_loop_check_proc(ctx, to, 0);
|
| 2470 | if (depth > EP_MAX_NESTS)
|
| 2471 | return -1;
|
| 2472 |
|
| 2473 | rcu_read_lock();
|
| 2474 | upwards_depth = ep_get_upwards_depth_proc(ep, 0);
|
| 2475 | rcu_read_unlock();
|
| 2476 |
|
| 2477 | return (depth+1+upwards_depth > EP_MAX_NESTS) ? -1 : 0;
|
| 2478 | }
|
| 2479 |
|
| 2480 | static void clear_tfile_check_list(struct ep_ctl_ctx *ctx)
|
| 2481 | {
|
| 2482 | rcu_read_lock();
|
| 2483 | while (ctx->tfile_check_list != EP_UNACTIVE_PTR) {
|
| 2484 | struct epitems_head *head = ctx->tfile_check_list;
|
| 2485 | ctx->tfile_check_list = head->next;
|
| 2486 | unlist_file(head);
|
| 2487 | }
|
| 2488 | rcu_read_unlock();
|
| 2489 | }
|
| 2490 |
|
| 2491 | |
| 2492 | * Open an eventpoll file descriptor.
|
| 2493 |
|
| 2494 | static int do_epoll_create(int flags)
|
| 2495 | {
|
| 2496 | int error;
|
| 2497 | struct eventpoll *ep;
|
| 2498 |
|
| 2499 |
|
| 2500 | BUILD_BUG_ON(EPOLL_CLOEXEC != O_CLOEXEC);
|
| 2501 |
|
| 2502 | if (flags & ~EPOLL_CLOEXEC)
|
| 2503 | return -EINVAL;
|
| 2504 | |
| 2505 | * Create the internal data structure ("struct eventpoll").
|
| 2506 |
|
| 2507 | error = ep_alloc(&ep);
|
| 2508 | if (error < 0)
|
| 2509 | return error;
|
| 2510 | |
| 2511 | * Creates all the items needed to setup an eventpoll file. That is,
|
| 2512 | * a file structure and a free file descriptor.
|
| 2513 |
|
| 2514 | FD_PREPARE(fdf, O_RDWR | (flags & O_CLOEXEC),
|
| 2515 | anon_inode_getfile("[eventpoll]", &eventpoll_fops, ep,
|
| 2516 | O_RDWR | (flags & O_CLOEXEC)));
|
| 2517 | if (fdf->fd < 0) {
|
| 2518 | ep_clear_and_put(ep);
|
| 2519 | return fdf->fd;
|
| 2520 | }
|
| 2521 | ep->file = fdf->file;
|
| 2522 | return fd_publish(fdf);
|
| 2523 | }
|
| 2524 |
|
| 2525 | SYSCALL_DEFINE1(epoll_create1, int, flags)
|
| 2526 | {
|
| 2527 | return do_epoll_create(flags);
|
| 2528 | }
|
| 2529 |
|
| 2530 | SYSCALL_DEFINE1(epoll_create, int, size)
|
| 2531 | {
|
| 2532 | if (size <= 0)
|
| 2533 | return -EINVAL;
|
| 2534 |
|
| 2535 | return do_epoll_create(0);
|
| 2536 | }
|
| 2537 |
|
| 2538 | #ifdef CONFIG_PM_SLEEP
|
| 2539 | static inline void ep_take_care_of_epollwakeup(struct epoll_event *epev)
|
| 2540 | {
|
| 2541 | if ((epev->events & EPOLLWAKEUP) && !capable(CAP_BLOCK_SUSPEND))
|
| 2542 | epev->events &= ~EPOLLWAKEUP;
|
| 2543 | }
|
| 2544 | #else
|
| 2545 | static inline void ep_take_care_of_epollwakeup(struct epoll_event *epev)
|
| 2546 | {
|
| 2547 | epev->events &= ~EPOLLWAKEUP;
|
| 2548 | }
|
| 2549 | #endif
|
| 2550 |
|
| 2551 | static inline int epoll_mutex_lock(struct mutex *mutex, bool nonblock)
|
| 2552 | {
|
| 2553 | if (!nonblock) {
|
| 2554 | mutex_lock(mutex);
|
| 2555 | return 0;
|
| 2556 | }
|
| 2557 | return mutex_trylock(mutex) ? 0 : -EAGAIN;
|
| 2558 | }
|
| 2559 |
|
| 2560 | |
| 2561 | * Acquire the locks required for do_epoll_ctl() on @ep for @op.
|
| 2562 | *
|
| 2563 | * Always takes ep->mtx. For EPOLL_CTL_ADD, additionally runs the
|
| 2564 | * loop / path check under epnested_mutex when the topology can
|
| 2565 | * change: @ep is already watched (epfile->f_ep non-NULL), @ep was
|
| 2566 | * recently loop-checked (ep->gen == loop_check_gen), or @tfile is
|
| 2567 | * itself an eventpoll.
|
| 2568 | *
|
| 2569 | * Return value encodes both outcome and lock state:
|
| 2570 | *
|
| 2571 | * 0 success; ep->mtx held.
|
| 2572 | * 1 success; ep->mtx held AND the full check ran under
|
| 2573 | * epnested_mutex (which is also still held). The value
|
| 2574 | * doubles as the @full_check argument to ep_insert().
|
| 2575 | * -errno failure; no locks held.
|
| 2576 | *
|
| 2577 | * The caller releases what was taken with ep_ctl_unlock(ep, ret).
|
| 2578 | *
|
| 2579 | * Holding epnested_mutex on add is what prevents two racing
|
| 2580 | * EPOLL_CTL_ADDs on different eps from building a cycle without
|
| 2581 | * either walker observing it.
|
| 2582 |
|
| 2583 | static int ep_ctl_lock(struct ep_ctl_ctx *ctx, struct eventpoll *ep, int op,
|
| 2584 | struct file *epfile, struct file *tfile, bool nonblock)
|
| 2585 | {
|
| 2586 | struct eventpoll *tep;
|
| 2587 | int error;
|
| 2588 |
|
| 2589 | error = epoll_mutex_lock(&ep->mtx, nonblock);
|
| 2590 | if (error)
|
| 2591 | return error;
|
| 2592 |
|
| 2593 | if (op != EPOLL_CTL_ADD)
|
| 2594 | return 0;
|
| 2595 | if (!READ_ONCE(epfile->f_ep) && ep->gen != loop_check_gen &&
|
| 2596 | !is_file_epoll(tfile))
|
| 2597 | return 0;
|
| 2598 |
|
| 2599 |
|
| 2600 | mutex_unlock(&ep->mtx);
|
| 2601 | error = epoll_mutex_lock(&epnested_mutex, nonblock);
|
| 2602 | if (error)
|
| 2603 | return error;
|
| 2604 |
|
| 2605 | loop_check_gen++;
|
| 2606 |
|
| 2607 | if (is_file_epoll(tfile)) {
|
| 2608 | tep = tfile->private_data;
|
| 2609 | if (ep_loop_check(ctx, ep, tep) != 0) {
|
| 2610 | error = -ELOOP;
|
| 2611 | goto err_unlock_nested;
|
| 2612 | }
|
| 2613 | }
|
| 2614 |
|
| 2615 | error = epoll_mutex_lock(&ep->mtx, nonblock);
|
| 2616 | if (error)
|
| 2617 | goto err_unlock_nested;
|
| 2618 |
|
| 2619 | return 1;
|
| 2620 |
|
| 2621 | err_unlock_nested:
|
| 2622 | clear_tfile_check_list(ctx);
|
| 2623 | loop_check_gen++;
|
| 2624 | mutex_unlock(&epnested_mutex);
|
| 2625 | return error;
|
| 2626 | }
|
| 2627 |
|
| 2628 | static void ep_ctl_unlock(struct ep_ctl_ctx *ctx, struct eventpoll *ep,
|
| 2629 | int full_check)
|
| 2630 | {
|
| 2631 | mutex_unlock(&ep->mtx);
|
| 2632 | if (full_check) {
|
| 2633 | clear_tfile_check_list(ctx);
|
| 2634 | loop_check_gen++;
|
| 2635 | mutex_unlock(&epnested_mutex);
|
| 2636 | }
|
| 2637 | }
|
| 2638 |
|
| 2639 | int do_epoll_ctl_file(struct file *f, int op, struct epoll_key *tf,
|
| 2640 | struct epoll_event *epds, bool nonblock)
|
| 2641 | {
|
| 2642 | int error;
|
| 2643 | int full_check;
|
| 2644 | struct eventpoll *ep;
|
| 2645 | struct epitem *epi;
|
| 2646 | struct ep_ctl_ctx ctx = {
|
| 2647 | .tfile_check_list = EP_UNACTIVE_PTR,
|
| 2648 | };
|
| 2649 |
|
| 2650 |
|
| 2651 | if (!file_can_poll(tf->file))
|
| 2652 | return -EPERM;
|
| 2653 |
|
| 2654 |
|
| 2655 | if (ep_op_has_event(op))
|
| 2656 | ep_take_care_of_epollwakeup(epds);
|
| 2657 |
|
| 2658 | |
| 2659 | * The @f file must itself be an eventpoll, and we do not permit
|
| 2660 | * adding an epoll file descriptor inside itself.
|
| 2661 |
|
| 2662 | if (f == tf->file || !is_file_epoll(f))
|
| 2663 | return -EINVAL;
|
| 2664 |
|
| 2665 | |
| 2666 | * epoll adds to the wakeup queue at EPOLL_CTL_ADD time only,
|
| 2667 | * so EPOLLEXCLUSIVE is not allowed for a EPOLL_CTL_MOD operation.
|
| 2668 | * Also, nested exclusive wakeups are not supported.
|
| 2669 |
|
| 2670 | if (ep_op_has_event(op) && (epds->events & EPOLLEXCLUSIVE)) {
|
| 2671 | if (op == EPOLL_CTL_MOD)
|
| 2672 | return -EINVAL;
|
| 2673 | if (op == EPOLL_CTL_ADD && (is_file_epoll(tf->file) ||
|
| 2674 | (epds->events & ~EPOLLEXCLUSIVE_OK_BITS)))
|
| 2675 | return -EINVAL;
|
| 2676 | }
|
| 2677 |
|
| 2678 | ep = f->private_data;
|
| 2679 |
|
| 2680 | full_check = ep_ctl_lock(&ctx, ep, op, f, tf->file, nonblock);
|
| 2681 | if (full_check < 0)
|
| 2682 | return full_check;
|
| 2683 |
|
| 2684 | |
| 2685 | * Look the target up in ep's RB tree. We hold ep->mtx, so the
|
| 2686 | * item stays valid until we release.
|
| 2687 |
|
| 2688 | epi = ep_find(ep, tf);
|
| 2689 |
|
| 2690 | error = -EINVAL;
|
| 2691 | switch (op) {
|
| 2692 | case EPOLL_CTL_ADD:
|
| 2693 | if (!epi) {
|
| 2694 | epds->events |= EPOLLERR | EPOLLHUP;
|
| 2695 | error = ep_insert(&ctx, ep, epds, tf, full_check);
|
| 2696 | } else
|
| 2697 | error = -EEXIST;
|
| 2698 | break;
|
| 2699 | case EPOLL_CTL_DEL:
|
| 2700 | if (epi) {
|
| 2701 | |
| 2702 | * The eventpoll itself is still alive: the refcount
|
| 2703 | * can't go to zero here.
|
| 2704 |
|
| 2705 | ep_remove(ep, epi);
|
| 2706 | error = 0;
|
| 2707 | } else {
|
| 2708 | error = -ENOENT;
|
| 2709 | }
|
| 2710 | break;
|
| 2711 | case EPOLL_CTL_MOD:
|
| 2712 | if (epi) {
|
| 2713 | if (!(epi->event.events & EPOLLEXCLUSIVE)) {
|
| 2714 | epds->events |= EPOLLERR | EPOLLHUP;
|
| 2715 | error = ep_modify(ep, epi, epds);
|
| 2716 | }
|
| 2717 | } else
|
| 2718 | error = -ENOENT;
|
| 2719 | break;
|
| 2720 | }
|
| 2721 |
|
| 2722 | ep_ctl_unlock(&ctx, ep, full_check);
|
| 2723 | return error;
|
| 2724 | }
|
| 2725 |
|
| 2726 | int do_epoll_ctl(int epfd, int op, int fd, struct epoll_event *epds,
|
| 2727 | bool nonblock)
|
| 2728 | {
|
| 2729 | struct epoll_key efd;
|
| 2730 |
|
| 2731 | CLASS(fd, f)(epfd);
|
| 2732 | if (fd_empty(f))
|
| 2733 | return -EBADF;
|
| 2734 |
|
| 2735 |
|
| 2736 | CLASS(fd, tf)(fd);
|
| 2737 | if (fd_empty(tf))
|
| 2738 | return -EBADF;
|
| 2739 |
|
| 2740 | efd.file = fd_file(tf);
|
| 2741 | efd.fd = fd;
|
| 2742 | return do_epoll_ctl_file(fd_file(f), op, &efd, epds, nonblock);
|
| 2743 | }
|
| 2744 |
|
| 2745 | |
| 2746 | * The following function implements the controller interface for
|
| 2747 | * the eventpoll file that enables the insertion/removal/change of
|
| 2748 | * file descriptors inside the interest set.
|
| 2749 |
|
| 2750 | SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
|
| 2751 | struct epoll_event __user *, event)
|
| 2752 | {
|
| 2753 | struct epoll_event epds;
|
| 2754 |
|
| 2755 | if (ep_op_has_event(op) &&
|
| 2756 | copy_from_user(&epds, event, sizeof(struct epoll_event)))
|
| 2757 | return -EFAULT;
|
| 2758 |
|
| 2759 | return do_epoll_ctl(epfd, op, fd, &epds, false);
|
| 2760 | }
|
| 2761 |
|
| 2762 | static int ep_check_params(struct file *file, struct epoll_event __user *evs,
|
| 2763 | int maxevents)
|
| 2764 | {
|
| 2765 |
|
| 2766 | if (maxevents <= 0 || maxevents > EP_MAX_EVENTS)
|
| 2767 | return -EINVAL;
|
| 2768 |
|
| 2769 |
|
| 2770 | if (!access_ok(evs, maxevents * sizeof(struct epoll_event)))
|
| 2771 | return -EFAULT;
|
| 2772 |
|
| 2773 | |
| 2774 | * We have to check that the file structure underneath the fd
|
| 2775 | * the user passed to us _is_ an eventpoll file.
|
| 2776 |
|
| 2777 | if (!is_file_epoll(file))
|
| 2778 | return -EINVAL;
|
| 2779 |
|
| 2780 | return 0;
|
| 2781 | }
|
| 2782 |
|
| 2783 | int epoll_sendevents(struct file *file, struct epoll_event __user *events,
|
| 2784 | int maxevents)
|
| 2785 | {
|
| 2786 | struct eventpoll *ep;
|
| 2787 | int ret;
|
| 2788 |
|
| 2789 | ret = ep_check_params(file, events, maxevents);
|
| 2790 | if (unlikely(ret))
|
| 2791 | return ret;
|
| 2792 |
|
| 2793 | ep = file->private_data;
|
| 2794 | |
| 2795 | * Racy call, but that's ok - it should get retried based on
|
| 2796 | * poll readiness anyway.
|
| 2797 |
|
| 2798 | if (ep_events_available(ep))
|
| 2799 | return ep_try_send_events(ep, events, maxevents);
|
| 2800 | return 0;
|
| 2801 | }
|
| 2802 |
|
| 2803 | |
| 2804 | * Implement the event wait interface for the eventpoll file. It is the kernel
|
| 2805 | * part of the user space epoll_wait(2).
|
| 2806 |
|
| 2807 | static int do_epoll_wait(int epfd, struct epoll_event __user *events,
|
| 2808 | int maxevents, struct timespec64 *to)
|
| 2809 | {
|
| 2810 | struct eventpoll *ep;
|
| 2811 | int ret;
|
| 2812 |
|
| 2813 |
|
| 2814 | CLASS(fd, f)(epfd);
|
| 2815 | if (fd_empty(f))
|
| 2816 | return -EBADF;
|
| 2817 |
|
| 2818 | ret = ep_check_params(fd_file(f), events, maxevents);
|
| 2819 | if (unlikely(ret))
|
| 2820 | return ret;
|
| 2821 |
|
| 2822 | |
| 2823 | * At this point it is safe to assume that the "private_data" contains
|
| 2824 | * our own data structure.
|
| 2825 |
|
| 2826 | ep = fd_file(f)->private_data;
|
| 2827 |
|
| 2828 |
|
| 2829 | return ep_poll(ep, events, maxevents, to);
|
| 2830 | }
|
| 2831 |
|
| 2832 | SYSCALL_DEFINE4(epoll_wait, int, epfd, struct epoll_event __user *, events,
|
| 2833 | int, maxevents, int, timeout)
|
| 2834 | {
|
| 2835 | struct timespec64 to;
|
| 2836 |
|
| 2837 | return do_epoll_wait(epfd, events, maxevents,
|
| 2838 | ep_timeout_to_timespec(&to, timeout));
|
| 2839 | }
|
| 2840 |
|
| 2841 | |
| 2842 | * Implement the event wait interface for the eventpoll file. It is the kernel
|
| 2843 | * part of the user space epoll_pwait(2).
|
| 2844 |
|
| 2845 | static int do_epoll_pwait(int epfd, struct epoll_event __user *events,
|
| 2846 | int maxevents, struct timespec64 *to,
|
| 2847 | const sigset_t __user *sigmask, size_t sigsetsize)
|
| 2848 | {
|
| 2849 | int error;
|
| 2850 |
|
| 2851 | |
| 2852 | * If the caller wants a certain signal mask to be set during the wait,
|
| 2853 | * we apply it here.
|
| 2854 |
|
| 2855 | error = set_user_sigmask(sigmask, sigsetsize);
|
| 2856 | if (error)
|
| 2857 | return error;
|
| 2858 |
|
| 2859 | error = do_epoll_wait(epfd, events, maxevents, to);
|
| 2860 |
|
| 2861 | restore_saved_sigmask_unless(error == -EINTR);
|
| 2862 |
|
| 2863 | return error;
|
| 2864 | }
|
| 2865 |
|
| 2866 | SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
|
| 2867 | int, maxevents, int, timeout, const sigset_t __user *, sigmask,
|
| 2868 | size_t, sigsetsize)
|
| 2869 | {
|
| 2870 | struct timespec64 to;
|
| 2871 |
|
| 2872 | return do_epoll_pwait(epfd, events, maxevents,
|
| 2873 | ep_timeout_to_timespec(&to, timeout),
|
| 2874 | sigmask, sigsetsize);
|
| 2875 | }
|
| 2876 |
|
| 2877 | SYSCALL_DEFINE6(epoll_pwait2, int, epfd, struct epoll_event __user *, events,
|
| 2878 | int, maxevents, const struct __kernel_timespec __user *, timeout,
|
| 2879 | const sigset_t __user *, sigmask, size_t, sigsetsize)
|
| 2880 | {
|
| 2881 | struct timespec64 ts, *to = NULL;
|
| 2882 |
|
| 2883 | if (timeout) {
|
| 2884 | if (get_timespec64(&ts, timeout))
|
| 2885 | return -EFAULT;
|
| 2886 | to = &ts;
|
| 2887 | if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
|
| 2888 | return -EINVAL;
|
| 2889 | }
|
| 2890 |
|
| 2891 | return do_epoll_pwait(epfd, events, maxevents, to,
|
| 2892 | sigmask, sigsetsize);
|
| 2893 | }
|
| 2894 |
|
| 2895 | #ifdef CONFIG_KCMP
|
| 2896 | static struct epitem *ep_find_tfd(struct eventpoll *ep, int tfd, unsigned long toff)
|
| 2897 | {
|
| 2898 | struct rb_node *rbp;
|
| 2899 | struct epitem *epi;
|
| 2900 |
|
| 2901 | for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = rb_next(rbp)) {
|
| 2902 | epi = rb_entry(rbp, struct epitem, rbn);
|
| 2903 | if (epi->ffd.fd == tfd) {
|
| 2904 | if (toff == 0)
|
| 2905 | return epi;
|
| 2906 | else
|
| 2907 | toff--;
|
| 2908 | }
|
| 2909 | cond_resched();
|
| 2910 | }
|
| 2911 |
|
| 2912 | return NULL;
|
| 2913 | }
|
| 2914 |
|
| 2915 | struct file *get_epoll_tfile_raw_ptr(struct file *file, int tfd,
|
| 2916 | unsigned long toff)
|
| 2917 | {
|
| 2918 | struct file *file_raw;
|
| 2919 | struct eventpoll *ep;
|
| 2920 | struct epitem *epi;
|
| 2921 |
|
| 2922 | if (!is_file_epoll(file))
|
| 2923 | return ERR_PTR(-EINVAL);
|
| 2924 |
|
| 2925 | ep = file->private_data;
|
| 2926 |
|
| 2927 | mutex_lock(&ep->mtx);
|
| 2928 | epi = ep_find_tfd(ep, tfd, toff);
|
| 2929 | if (epi)
|
| 2930 | file_raw = epi->ffd.file;
|
| 2931 | else
|
| 2932 | file_raw = ERR_PTR(-ENOENT);
|
| 2933 | mutex_unlock(&ep->mtx);
|
| 2934 |
|
| 2935 | return file_raw;
|
| 2936 | }
|
| 2937 | #endif
|
| 2938 |
|
| 2939 | #ifdef CONFIG_COMPAT
|
| 2940 | static int do_compat_epoll_pwait(int epfd, struct epoll_event __user *events,
|
| 2941 | int maxevents, struct timespec64 *timeout,
|
| 2942 | const compat_sigset_t __user *sigmask,
|
| 2943 | compat_size_t sigsetsize)
|
| 2944 | {
|
| 2945 | long err;
|
| 2946 |
|
| 2947 | |
| 2948 | * If the caller wants a certain signal mask to be set during the wait,
|
| 2949 | * we apply it here.
|
| 2950 |
|
| 2951 | err = set_compat_user_sigmask(sigmask, sigsetsize);
|
| 2952 | if (err)
|
| 2953 | return err;
|
| 2954 |
|
| 2955 | err = do_epoll_wait(epfd, events, maxevents, timeout);
|
| 2956 |
|
| 2957 | restore_saved_sigmask_unless(err == -EINTR);
|
| 2958 |
|
| 2959 | return err;
|
| 2960 | }
|
| 2961 |
|
| 2962 | COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
|
| 2963 | struct epoll_event __user *, events,
|
| 2964 | int, maxevents, int, timeout,
|
| 2965 | const compat_sigset_t __user *, sigmask,
|
| 2966 | compat_size_t, sigsetsize)
|
| 2967 | {
|
| 2968 | struct timespec64 to;
|
| 2969 |
|
| 2970 | return do_compat_epoll_pwait(epfd, events, maxevents,
|
| 2971 | ep_timeout_to_timespec(&to, timeout),
|
| 2972 | sigmask, sigsetsize);
|
| 2973 | }
|
| 2974 |
|
| 2975 | COMPAT_SYSCALL_DEFINE6(epoll_pwait2, int, epfd,
|
| 2976 | struct epoll_event __user *, events,
|
| 2977 | int, maxevents,
|
| 2978 | const struct __kernel_timespec __user *, timeout,
|
| 2979 | const compat_sigset_t __user *, sigmask,
|
| 2980 | compat_size_t, sigsetsize)
|
| 2981 | {
|
| 2982 | struct timespec64 ts, *to = NULL;
|
| 2983 |
|
| 2984 | if (timeout) {
|
| 2985 | if (get_timespec64(&ts, timeout))
|
| 2986 | return -EFAULT;
|
| 2987 | to = &ts;
|
| 2988 | if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec))
|
| 2989 | return -EINVAL;
|
| 2990 | }
|
| 2991 |
|
| 2992 | return do_compat_epoll_pwait(epfd, events, maxevents, to,
|
| 2993 | sigmask, sigsetsize);
|
| 2994 | }
|
| 2995 |
|
| 2996 | #endif
|
| 2997 |
|
| 2998 | static int __init eventpoll_init(void)
|
| 2999 | {
|
| 3000 | struct sysinfo si;
|
| 3001 |
|
| 3002 | si_meminfo(&si);
|
| 3003 | |
| 3004 | * Allows top 4% of lomem to be allocated for epoll watches (per user).
|
| 3005 |
|
| 3006 | max_user_watches = (((si.totalram - si.totalhigh) / 25) << PAGE_SHIFT) /
|
| 3007 | EP_ITEM_COST;
|
| 3008 | BUG_ON(max_user_watches < 0);
|
| 3009 |
|
| 3010 | |
| 3011 | * We can have many thousands of epitems, so prevent this from
|
| 3012 | * using an extra cache line on 64-bit (and smaller) CPUs
|
| 3013 |
|
| 3014 | BUILD_BUG_ON(sizeof(void *) <= 8 && sizeof(struct epitem) > 128);
|
| 3015 |
|
| 3016 |
|
| 3017 | epi_cache = kmem_cache_create("eventpoll_epi", sizeof(struct epitem),
|
| 3018 | 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC|SLAB_ACCOUNT, NULL);
|
| 3019 |
|
| 3020 |
|
| 3021 | pwq_cache = kmem_cache_create("eventpoll_pwq",
|
| 3022 | sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
|
| 3023 | epoll_sysctls_init();
|
| 3024 |
|
| 3025 | ephead_cache = kmem_cache_create("ep_head",
|
| 3026 | sizeof(struct epitems_head), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL);
|
| 3027 |
|
| 3028 | return 0;
|
| 3029 | }
|
| 3030 | fs_initcall(eventpoll_init);
|