kernel/sched/wait.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Generic waiting primitives.
   4  *
   5  * (C) 2004 Nadia Yvette Chambers, Oracle
   6  */
   7
   8 void __init_waitqueue_head(struct wait_queue_head *wq_head, const char *name, struct lock_class_key *key)
   9 {
  10         spin_lock_init(&wq_head->lock);
  11         lockdep_set_class_and_name(&wq_head->lock, key, name);
  12         INIT_LIST_HEAD(&wq_head->head);
  13 }
  14
  15 EXPORT_SYMBOL(__init_waitqueue_head);
  16
  17 void add_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  18 {
  19         unsigned long flags;
  20
  21         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
  22         spin_lock_irqsave(&wq_head->lock, flags);
  23         __add_wait_queue(wq_head, wq_entry);
  24         spin_unlock_irqrestore(&wq_head->lock, flags);
  25 }
  26 EXPORT_SYMBOL(add_wait_queue);
  27
  28 void add_wait_queue_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  29 {
  30         unsigned long flags;
  31
  32         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
  33         spin_lock_irqsave(&wq_head->lock, flags);
  34         __add_wait_queue_entry_tail(wq_head, wq_entry);
  35         spin_unlock_irqrestore(&wq_head->lock, flags);
  36 }
  37 EXPORT_SYMBOL(add_wait_queue_exclusive);
  38
  39 void add_wait_queue_priority(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  40 {
  41         unsigned long flags;
  42
  43         wq_entry->flags |= WQ_FLAG_EXCLUSIVE | WQ_FLAG_PRIORITY;
  44         spin_lock_irqsave(&wq_head->lock, flags);
  45         __add_wait_queue(wq_head, wq_entry);
  46         spin_unlock_irqrestore(&wq_head->lock, flags);
  47 }
  48 EXPORT_SYMBOL_GPL(add_wait_queue_priority);
  49
  50 void remove_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  51 {
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&wq_head->lock, flags);
  55         __remove_wait_queue(wq_head, wq_entry);
  56         spin_unlock_irqrestore(&wq_head->lock, flags);
  57 }
  58 EXPORT_SYMBOL(remove_wait_queue);
  59
  60 /*
  61  * Scan threshold to break wait queue walk.
  62  * This allows a waker to take a break from holding the
  63  * wait queue lock during the wait queue walk.
  64  */
  65 #define WAITQUEUE_WALK_BREAK_CNT 64
  66
  67 /*
  68  * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
  69  * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
  70  * number) then we wake that number of exclusive tasks, and potentially all
  71  * the non-exclusive tasks. Normally, exclusive tasks will be at the end of
  72  * the list and any non-exclusive tasks will be woken first. A priority task
  73  * may be at the head of the list, and can consume the event without any other
  74  * tasks being woken.
  75  *
  76  * There are circumstances in which we can try to wake a task which has already
  77  * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
  78  * zero in this (rare) case, and we handle it by continuing to scan the queue.
  79  */
  80 static int __wake_up_common(struct wait_queue_head *wq_head, unsigned int mode,
  81                         int nr_exclusive, int wake_flags, void *key,
  82                         wait_queue_entry_t *bookmark)
  83 {
  84         wait_queue_entry_t *curr, *next;
  85         int cnt = 0;
  86
  87         lockdep_assert_held(&wq_head->lock);
  88
  89         if (bookmark && (bookmark->flags & WQ_FLAG_BOOKMARK)) {
  90                 curr = list_next_entry(bookmark, entry);
  91
  92                 list_del(&bookmark->entry);
  93                 bookmark->flags = 0;
  94         } else
  95                 curr = list_first_entry(&wq_head->head, wait_queue_entry_t, entry);
  96
  97         if (&curr->entry == &wq_head->head)
  98                 return nr_exclusive;
  99
 100         list_for_each_entry_safe_from(curr, next, &wq_head->head, entry) {
 101                 unsigned flags = curr->flags;
 102                 int ret;
 103
 104                 if (flags & WQ_FLAG_BOOKMARK)
 105                         continue;
 106
 107                 ret = curr->func(curr, mode, wake_flags, key);
 108                 if (ret < 0)
 109                         break;
 110                 if (ret && (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
 111                         break;
 112
 113                 if (bookmark && (++cnt > WAITQUEUE_WALK_BREAK_CNT) &&
 114                                 (&next->entry != &wq_head->head)) {
 115                         bookmark->flags = WQ_FLAG_BOOKMARK;
 116                         list_add_tail(&bookmark->entry, &next->entry);
 117                         break;
 118                 }
 119         }
 120
 121         return nr_exclusive;
 122 }
 123
 124 static int __wake_up_common_lock(struct wait_queue_head *wq_head, unsigned int mode,
 125                         int nr_exclusive, int wake_flags, void *key)
 126 {
 127         unsigned long flags;
 128         wait_queue_entry_t bookmark;
 129         int remaining = nr_exclusive;
 130
 131         bookmark.flags = 0;
 132         bookmark.private = NULL;
 133         bookmark.func = NULL;
 134         INIT_LIST_HEAD(&bookmark.entry);
 135
 136         do {
 137                 spin_lock_irqsave(&wq_head->lock, flags);
 138                 remaining = __wake_up_common(wq_head, mode, remaining,
 139                                                 wake_flags, key, &bookmark);
 140                 spin_unlock_irqrestore(&wq_head->lock, flags);
 141         } while (bookmark.flags & WQ_FLAG_BOOKMARK);
 142
 143         return nr_exclusive - remaining;
 144 }
 145
 146 /**
 147  * __wake_up - wake up threads blocked on a waitqueue.
 148  * @wq_head: the waitqueue
 149  * @mode: which threads
 150  * @nr_exclusive: how many wake-one or wake-many threads to wake up
 151  * @key: is directly passed to the wakeup function
 152  *
 153  * If this function wakes up a task, it executes a full memory barrier
 154  * before accessing the task state.  Returns the number of exclusive
 155  * tasks that were awaken.
 156  */
 157 int __wake_up(struct wait_queue_head *wq_head, unsigned int mode,
 158               int nr_exclusive, void *key)
 159 {
 160         return __wake_up_common_lock(wq_head, mode, nr_exclusive, 0, key);
 161 }
 162 EXPORT_SYMBOL(__wake_up);
 163
 164 /*
 165  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
 166  */
 167 void __wake_up_locked(struct wait_queue_head *wq_head, unsigned int mode, int nr)
 168 {
 169         __wake_up_common(wq_head, mode, nr, 0, NULL, NULL);
 170 }
 171 EXPORT_SYMBOL_GPL(__wake_up_locked);
 172
 173 void __wake_up_locked_key(struct wait_queue_head *wq_head, unsigned int mode, void *key)
 174 {
 175         __wake_up_common(wq_head, mode, 1, 0, key, NULL);
 176 }
 177 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
 178
 179 void __wake_up_locked_key_bookmark(struct wait_queue_head *wq_head,
 180                 unsigned int mode, void *key, wait_queue_entry_t *bookmark)
 181 {
 182         __wake_up_common(wq_head, mode, 1, 0, key, bookmark);
 183 }
 184 EXPORT_SYMBOL_GPL(__wake_up_locked_key_bookmark);
 185
 186 /**
 187  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
 188  * @wq_head: the waitqueue
 189  * @mode: which threads
 190  * @key: opaque value to be passed to wakeup targets
 191  *
 192  * The sync wakeup differs that the waker knows that it will schedule
 193  * away soon, so while the target thread will be woken up, it will not
 194  * be migrated to another CPU - ie. the two threads are 'synchronized'
 195  * with each other. This can prevent needless bouncing between CPUs.
 196  *
 197  * On UP it can prevent extra preemption.
 198  *
 199  * If this function wakes up a task, it executes a full memory barrier before
 200  * accessing the task state.
 201  */
 202 void __wake_up_sync_key(struct wait_queue_head *wq_head, unsigned int mode,
 203                         void *key)
 204 {
 205         if (unlikely(!wq_head))
 206                 return;
 207
 208         __wake_up_common_lock(wq_head, mode, 1, WF_SYNC, key);
 209 }
 210 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
 211
 212 /**
 213  * __wake_up_locked_sync_key - wake up a thread blocked on a locked waitqueue.
 214  * @wq_head: the waitqueue
 215  * @mode: which threads
 216  * @key: opaque value to be passed to wakeup targets
 217  *
 218  * The sync wakeup differs in that the waker knows that it will schedule
 219  * away soon, so while the target thread will be woken up, it will not
 220  * be migrated to another CPU - ie. the two threads are 'synchronized'
 221  * with each other. This can prevent needless bouncing between CPUs.
 222  *
 223  * On UP it can prevent extra preemption.
 224  *
 225  * If this function wakes up a task, it executes a full memory barrier before
 226  * accessing the task state.
 227  */
 228 void __wake_up_locked_sync_key(struct wait_queue_head *wq_head,
 229                                unsigned int mode, void *key)
 230 {
 231         __wake_up_common(wq_head, mode, 1, WF_SYNC, key, NULL);
 232 }
 233 EXPORT_SYMBOL_GPL(__wake_up_locked_sync_key);
 234
 235 /*
 236  * __wake_up_sync - see __wake_up_sync_key()
 237  */
 238 void __wake_up_sync(struct wait_queue_head *wq_head, unsigned int mode)
 239 {
 240         __wake_up_sync_key(wq_head, mode, NULL);
 241 }
 242 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
 243
 244 void __wake_up_pollfree(struct wait_queue_head *wq_head)
 245 {
 246         __wake_up(wq_head, TASK_NORMAL, 0, poll_to_key(EPOLLHUP | POLLFREE));
 247         /* POLLFREE must have cleared the queue. */
 248         WARN_ON_ONCE(waitqueue_active(wq_head));
 249 }
 250
 251 /*
 252  * Note: we use "set_current_state()" _after_ the wait-queue add,
 253  * because we need a memory barrier there on SMP, so that any
 254  * wake-function that tests for the wait-queue being active
 255  * will be guaranteed to see waitqueue addition _or_ subsequent
 256  * tests in this thread will see the wakeup having taken place.
 257  *
 258  * The spin_unlock() itself is semi-permeable and only protects
 259  * one way (it only protects stuff inside the critical region and
 260  * stops them from bleeding out - it would still allow subsequent
 261  * loads to move into the critical region).
 262  */
 263 void
 264 prepare_to_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 265 {
 266         unsigned long flags;
 267
 268         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
 269         spin_lock_irqsave(&wq_head->lock, flags);
 270         if (list_empty(&wq_entry->entry))
 271                 __add_wait_queue(wq_head, wq_entry);
 272         set_current_state(state);
 273         spin_unlock_irqrestore(&wq_head->lock, flags);
 274 }
 275 EXPORT_SYMBOL(prepare_to_wait);
 276
 277 /* Returns true if we are the first waiter in the queue, false otherwise. */
 278 bool
 279 prepare_to_wait_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 280 {
 281         unsigned long flags;
 282         bool was_empty = false;
 283
 284         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
 285         spin_lock_irqsave(&wq_head->lock, flags);
 286         if (list_empty(&wq_entry->entry)) {
 287                 was_empty = list_empty(&wq_head->head);
 288                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 289         }
 290         set_current_state(state);
 291         spin_unlock_irqrestore(&wq_head->lock, flags);
 292         return was_empty;
 293 }
 294 EXPORT_SYMBOL(prepare_to_wait_exclusive);
 295
 296 void init_wait_entry(struct wait_queue_entry *wq_entry, int flags)
 297 {
 298         wq_entry->flags = flags;
 299         wq_entry->private = current;
 300         wq_entry->func = autoremove_wake_function;
 301         INIT_LIST_HEAD(&wq_entry->entry);
 302 }
 303 EXPORT_SYMBOL(init_wait_entry);
 304
 305 long prepare_to_wait_event(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 306 {
 307         unsigned long flags;
 308         long ret = 0;
 309
 310         spin_lock_irqsave(&wq_head->lock, flags);
 311         if (signal_pending_state(state, current)) {
 312                 /*
 313                  * Exclusive waiter must not fail if it was selected by wakeup,
 314                  * it should "consume" the condition we were waiting for.
 315                  *
 316                  * The caller will recheck the condition and return success if
 317                  * we were already woken up, we can not miss the event because
 318                  * wakeup locks/unlocks the same wq_head->lock.
 319                  *
 320                  * But we need to ensure that set-condition + wakeup after that
 321                  * can't see us, it should wake up another exclusive waiter if
 322                  * we fail.
 323                  */
 324                 list_del_init(&wq_entry->entry);
 325                 ret = -ERESTARTSYS;
 326         } else {
 327                 if (list_empty(&wq_entry->entry)) {
 328                         if (wq_entry->flags & WQ_FLAG_EXCLUSIVE)
 329                                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 330                         else
 331                                 __add_wait_queue(wq_head, wq_entry);
 332                 }
 333                 set_current_state(state);
 334         }
 335         spin_unlock_irqrestore(&wq_head->lock, flags);
 336
 337         return ret;
 338 }
 339 EXPORT_SYMBOL(prepare_to_wait_event);
 340
 341 /*
 342  * Note! These two wait functions are entered with the
 343  * wait-queue lock held (and interrupts off in the _irq
 344  * case), so there is no race with testing the wakeup
 345  * condition in the caller before they add the wait
 346  * entry to the wake queue.
 347  */
 348 int do_wait_intr(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 349 {
 350         if (likely(list_empty(&wait->entry)))
 351                 __add_wait_queue_entry_tail(wq, wait);
 352
 353         set_current_state(TASK_INTERRUPTIBLE);
 354         if (signal_pending(current))
 355                 return -ERESTARTSYS;
 356
 357         spin_unlock(&wq->lock);
 358         schedule();
 359         spin_lock(&wq->lock);
 360
 361         return 0;
 362 }
 363 EXPORT_SYMBOL(do_wait_intr);
 364
 365 int do_wait_intr_irq(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 366 {
 367         if (likely(list_empty(&wait->entry)))
 368                 __add_wait_queue_entry_tail(wq, wait);
 369
 370         set_current_state(TASK_INTERRUPTIBLE);
 371         if (signal_pending(current))
 372                 return -ERESTARTSYS;
 373
 374         spin_unlock_irq(&wq->lock);
 375         schedule();
 376         spin_lock_irq(&wq->lock);
 377
 378         return 0;
 379 }
 380 EXPORT_SYMBOL(do_wait_intr_irq);
 381
 382 /**
 383  * finish_wait - clean up after waiting in a queue
 384  * @wq_head: waitqueue waited on
 385  * @wq_entry: wait descriptor
 386  *
 387  * Sets current thread back to running state and removes
 388  * the wait descriptor from the given waitqueue if still
 389  * queued.
 390  */
 391 void finish_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
 392 {
 393         unsigned long flags;
 394
 395         __set_current_state(TASK_RUNNING);
 396         /*
 397          * We can check for list emptiness outside the lock
 398          * IFF:
 399          *  - we use the "careful" check that verifies both
 400          *    the next and prev pointers, so that there cannot
 401          *    be any half-pending updates in progress on other
 402          *    CPU's that we haven't seen yet (and that might
 403          *    still change the stack area.
 404          * and
 405          *  - all other users take the lock (ie we can only
 406          *    have _one_ other CPU that looks at or modifies
 407          *    the list).
 408          */
 409         if (!list_empty_careful(&wq_entry->entry)) {
 410                 spin_lock_irqsave(&wq_head->lock, flags);
 411                 list_del_init(&wq_entry->entry);
 412                 spin_unlock_irqrestore(&wq_head->lock, flags);
 413         }
 414 }
 415 EXPORT_SYMBOL(finish_wait);
 416
 417 int autoremove_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 418 {
 419         int ret = default_wake_function(wq_entry, mode, sync, key);
 420
 421         if (ret)
 422                 list_del_init_careful(&wq_entry->entry);
 423
 424         return ret;
 425 }
 426 EXPORT_SYMBOL(autoremove_wake_function);
 427
 428 static inline bool is_kthread_should_stop(void)
 429 {
 430         return (current->flags & PF_KTHREAD) && kthread_should_stop();
 431 }
 432
 433 /*
 434  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
 435  *
 436  * add_wait_queue(&wq_head, &wait);
 437  * for (;;) {
 438  *     if (condition)
 439  *         break;
 440  *
 441  *     // in wait_woken()                       // in woken_wake_function()
 442  *
 443  *     p->state = mode;                         wq_entry->flags |= WQ_FLAG_WOKEN;
 444  *     smp_mb(); // A                           try_to_wake_up():
 445  *     if (!(wq_entry->flags & WQ_FLAG_WOKEN))     <full barrier>
 446  *         schedule()                              if (p->state & mode)
 447  *     p->state = TASK_RUNNING;                       p->state = TASK_RUNNING;
 448  *     wq_entry->flags &= ~WQ_FLAG_WOKEN;       ~~~~~~~~~~~~~~~~~~
 449  *     smp_mb(); // B                           condition = true;
 450  * }                                            smp_mb(); // C
 451  * remove_wait_queue(&wq_head, &wait);          wq_entry->flags |= WQ_FLAG_WOKEN;
 452  */
 453 long wait_woken(struct wait_queue_entry *wq_entry, unsigned mode, long timeout)
 454 {
 455         /*
 456          * The below executes an smp_mb(), which matches with the full barrier
 457          * executed by the try_to_wake_up() in woken_wake_function() such that
 458          * either we see the store to wq_entry->flags in woken_wake_function()
 459          * or woken_wake_function() sees our store to current->state.
 460          */
 461         set_current_state(mode); /* A */
 462         if (!(wq_entry->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
 463                 timeout = schedule_timeout(timeout);
 464         __set_current_state(TASK_RUNNING);
 465
 466         /*
 467          * The below executes an smp_mb(), which matches with the smp_mb() (C)
 468          * in woken_wake_function() such that either we see the wait condition
 469          * being true or the store to wq_entry->flags in woken_wake_function()
 470          * follows ours in the coherence order.
 471          */
 472         smp_store_mb(wq_entry->flags, wq_entry->flags & ~WQ_FLAG_WOKEN); /* B */
 473
 474         return timeout;
 475 }
 476 EXPORT_SYMBOL(wait_woken);
 477
 478 int woken_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 479 {
 480         /* Pairs with the smp_store_mb() in wait_woken(). */
 481         smp_mb(); /* C */
 482         wq_entry->flags |= WQ_FLAG_WOKEN;
 483
 484         return default_wake_function(wq_entry, mode, sync, key);
 485 }
 486 EXPORT_SYMBOL(woken_wake_function);