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 void __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
 130         bookmark.flags = 0;
 131         bookmark.private = NULL;
 132         bookmark.func = NULL;
 133         INIT_LIST_HEAD(&bookmark.entry);
 134
 135         do {
 136                 spin_lock_irqsave(&wq_head->lock, flags);
 137                 nr_exclusive = __wake_up_common(wq_head, mode, nr_exclusive,
 138                                                 wake_flags, key, &bookmark);
 139                 spin_unlock_irqrestore(&wq_head->lock, flags);
 140         } while (bookmark.flags & WQ_FLAG_BOOKMARK);
 141 }
 142
 143 /**
 144  * __wake_up - wake up threads blocked on a waitqueue.
 145  * @wq_head: the waitqueue
 146  * @mode: which threads
 147  * @nr_exclusive: how many wake-one or wake-many threads to wake up
 148  * @key: is directly passed to the wakeup function
 149  *
 150  * If this function wakes up a task, it executes a full memory barrier before
 151  * accessing the task state.
 152  */
 153 void __wake_up(struct wait_queue_head *wq_head, unsigned int mode,
 154                         int nr_exclusive, void *key)
 155 {
 156         __wake_up_common_lock(wq_head, mode, nr_exclusive, 0, key);
 157 }
 158 EXPORT_SYMBOL(__wake_up);
 159
 160 /*
 161  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
 162  */
 163 void __wake_up_locked(struct wait_queue_head *wq_head, unsigned int mode, int nr)
 164 {
 165         __wake_up_common(wq_head, mode, nr, 0, NULL, NULL);
 166 }
 167 EXPORT_SYMBOL_GPL(__wake_up_locked);
 168
 169 void __wake_up_locked_key(struct wait_queue_head *wq_head, unsigned int mode, void *key)
 170 {
 171         __wake_up_common(wq_head, mode, 1, 0, key, NULL);
 172 }
 173 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
 174
 175 void __wake_up_locked_key_bookmark(struct wait_queue_head *wq_head,
 176                 unsigned int mode, void *key, wait_queue_entry_t *bookmark)
 177 {
 178         __wake_up_common(wq_head, mode, 1, 0, key, bookmark);
 179 }
 180 EXPORT_SYMBOL_GPL(__wake_up_locked_key_bookmark);
 181
 182 /**
 183  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
 184  * @wq_head: the waitqueue
 185  * @mode: which threads
 186  * @key: opaque value to be passed to wakeup targets
 187  *
 188  * The sync wakeup differs that the waker knows that it will schedule
 189  * away soon, so while the target thread will be woken up, it will not
 190  * be migrated to another CPU - ie. the two threads are 'synchronized'
 191  * with each other. This can prevent needless bouncing between CPUs.
 192  *
 193  * On UP it can prevent extra preemption.
 194  *
 195  * If this function wakes up a task, it executes a full memory barrier before
 196  * accessing the task state.
 197  */
 198 void __wake_up_sync_key(struct wait_queue_head *wq_head, unsigned int mode,
 199                         void *key)
 200 {
 201         if (unlikely(!wq_head))
 202                 return;
 203
 204         __wake_up_common_lock(wq_head, mode, 1, WF_SYNC, key);
 205 }
 206 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
 207
 208 /**
 209  * __wake_up_locked_sync_key - wake up a thread blocked on a locked waitqueue.
 210  * @wq_head: the waitqueue
 211  * @mode: which threads
 212  * @key: opaque value to be passed to wakeup targets
 213  *
 214  * The sync wakeup differs in that the waker knows that it will schedule
 215  * away soon, so while the target thread will be woken up, it will not
 216  * be migrated to another CPU - ie. the two threads are 'synchronized'
 217  * with each other. This can prevent needless bouncing between CPUs.
 218  *
 219  * On UP it can prevent extra preemption.
 220  *
 221  * If this function wakes up a task, it executes a full memory barrier before
 222  * accessing the task state.
 223  */
 224 void __wake_up_locked_sync_key(struct wait_queue_head *wq_head,
 225                                unsigned int mode, void *key)
 226 {
 227         __wake_up_common(wq_head, mode, 1, WF_SYNC, key, NULL);
 228 }
 229 EXPORT_SYMBOL_GPL(__wake_up_locked_sync_key);
 230
 231 /*
 232  * __wake_up_sync - see __wake_up_sync_key()
 233  */
 234 void __wake_up_sync(struct wait_queue_head *wq_head, unsigned int mode)
 235 {
 236         __wake_up_sync_key(wq_head, mode, NULL);
 237 }
 238 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
 239
 240 void __wake_up_pollfree(struct wait_queue_head *wq_head)
 241 {
 242         __wake_up(wq_head, TASK_NORMAL, 0, poll_to_key(EPOLLHUP | POLLFREE));
 243         /* POLLFREE must have cleared the queue. */
 244         WARN_ON_ONCE(waitqueue_active(wq_head));
 245 }
 246
 247 /*
 248  * Note: we use "set_current_state()" _after_ the wait-queue add,
 249  * because we need a memory barrier there on SMP, so that any
 250  * wake-function that tests for the wait-queue being active
 251  * will be guaranteed to see waitqueue addition _or_ subsequent
 252  * tests in this thread will see the wakeup having taken place.
 253  *
 254  * The spin_unlock() itself is semi-permeable and only protects
 255  * one way (it only protects stuff inside the critical region and
 256  * stops them from bleeding out - it would still allow subsequent
 257  * loads to move into the critical region).
 258  */
 259 void
 260 prepare_to_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 261 {
 262         unsigned long flags;
 263
 264         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
 265         spin_lock_irqsave(&wq_head->lock, flags);
 266         if (list_empty(&wq_entry->entry))
 267                 __add_wait_queue(wq_head, wq_entry);
 268         set_current_state(state);
 269         spin_unlock_irqrestore(&wq_head->lock, flags);
 270 }
 271 EXPORT_SYMBOL(prepare_to_wait);
 272
 273 /* Returns true if we are the first waiter in the queue, false otherwise. */
 274 bool
 275 prepare_to_wait_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 276 {
 277         unsigned long flags;
 278         bool was_empty = false;
 279
 280         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
 281         spin_lock_irqsave(&wq_head->lock, flags);
 282         if (list_empty(&wq_entry->entry)) {
 283                 was_empty = list_empty(&wq_head->head);
 284                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 285         }
 286         set_current_state(state);
 287         spin_unlock_irqrestore(&wq_head->lock, flags);
 288         return was_empty;
 289 }
 290 EXPORT_SYMBOL(prepare_to_wait_exclusive);
 291
 292 void init_wait_entry(struct wait_queue_entry *wq_entry, int flags)
 293 {
 294         wq_entry->flags = flags;
 295         wq_entry->private = current;
 296         wq_entry->func = autoremove_wake_function;
 297         INIT_LIST_HEAD(&wq_entry->entry);
 298 }
 299 EXPORT_SYMBOL(init_wait_entry);
 300
 301 long prepare_to_wait_event(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 302 {
 303         unsigned long flags;
 304         long ret = 0;
 305
 306         spin_lock_irqsave(&wq_head->lock, flags);
 307         if (signal_pending_state(state, current)) {
 308                 /*
 309                  * Exclusive waiter must not fail if it was selected by wakeup,
 310                  * it should "consume" the condition we were waiting for.
 311                  *
 312                  * The caller will recheck the condition and return success if
 313                  * we were already woken up, we can not miss the event because
 314                  * wakeup locks/unlocks the same wq_head->lock.
 315                  *
 316                  * But we need to ensure that set-condition + wakeup after that
 317                  * can't see us, it should wake up another exclusive waiter if
 318                  * we fail.
 319                  */
 320                 list_del_init(&wq_entry->entry);
 321                 ret = -ERESTARTSYS;
 322         } else {
 323                 if (list_empty(&wq_entry->entry)) {
 324                         if (wq_entry->flags & WQ_FLAG_EXCLUSIVE)
 325                                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 326                         else
 327                                 __add_wait_queue(wq_head, wq_entry);
 328                 }
 329                 set_current_state(state);
 330         }
 331         spin_unlock_irqrestore(&wq_head->lock, flags);
 332
 333         return ret;
 334 }
 335 EXPORT_SYMBOL(prepare_to_wait_event);
 336
 337 /*
 338  * Note! These two wait functions are entered with the
 339  * wait-queue lock held (and interrupts off in the _irq
 340  * case), so there is no race with testing the wakeup
 341  * condition in the caller before they add the wait
 342  * entry to the wake queue.
 343  */
 344 int do_wait_intr(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 345 {
 346         if (likely(list_empty(&wait->entry)))
 347                 __add_wait_queue_entry_tail(wq, wait);
 348
 349         set_current_state(TASK_INTERRUPTIBLE);
 350         if (signal_pending(current))
 351                 return -ERESTARTSYS;
 352
 353         spin_unlock(&wq->lock);
 354         schedule();
 355         spin_lock(&wq->lock);
 356
 357         return 0;
 358 }
 359 EXPORT_SYMBOL(do_wait_intr);
 360
 361 int do_wait_intr_irq(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 362 {
 363         if (likely(list_empty(&wait->entry)))
 364                 __add_wait_queue_entry_tail(wq, wait);
 365
 366         set_current_state(TASK_INTERRUPTIBLE);
 367         if (signal_pending(current))
 368                 return -ERESTARTSYS;
 369
 370         spin_unlock_irq(&wq->lock);
 371         schedule();
 372         spin_lock_irq(&wq->lock);
 373
 374         return 0;
 375 }
 376 EXPORT_SYMBOL(do_wait_intr_irq);
 377
 378 /**
 379  * finish_wait - clean up after waiting in a queue
 380  * @wq_head: waitqueue waited on
 381  * @wq_entry: wait descriptor
 382  *
 383  * Sets current thread back to running state and removes
 384  * the wait descriptor from the given waitqueue if still
 385  * queued.
 386  */
 387 void finish_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
 388 {
 389         unsigned long flags;
 390
 391         __set_current_state(TASK_RUNNING);
 392         /*
 393          * We can check for list emptiness outside the lock
 394          * IFF:
 395          *  - we use the "careful" check that verifies both
 396          *    the next and prev pointers, so that there cannot
 397          *    be any half-pending updates in progress on other
 398          *    CPU's that we haven't seen yet (and that might
 399          *    still change the stack area.
 400          * and
 401          *  - all other users take the lock (ie we can only
 402          *    have _one_ other CPU that looks at or modifies
 403          *    the list).
 404          */
 405         if (!list_empty_careful(&wq_entry->entry)) {
 406                 spin_lock_irqsave(&wq_head->lock, flags);
 407                 list_del_init(&wq_entry->entry);
 408                 spin_unlock_irqrestore(&wq_head->lock, flags);
 409         }
 410 }
 411 EXPORT_SYMBOL(finish_wait);
 412
 413 int autoremove_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 414 {
 415         int ret = default_wake_function(wq_entry, mode, sync, key);
 416
 417         if (ret)
 418                 list_del_init_careful(&wq_entry->entry);
 419
 420         return ret;
 421 }
 422 EXPORT_SYMBOL(autoremove_wake_function);
 423
 424 static inline bool is_kthread_should_stop(void)
 425 {
 426         return (current->flags & PF_KTHREAD) && kthread_should_stop();
 427 }
 428
 429 /*
 430  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
 431  *
 432  * add_wait_queue(&wq_head, &wait);
 433  * for (;;) {
 434  *     if (condition)
 435  *         break;
 436  *
 437  *     // in wait_woken()                       // in woken_wake_function()
 438  *
 439  *     p->state = mode;                         wq_entry->flags |= WQ_FLAG_WOKEN;
 440  *     smp_mb(); // A                           try_to_wake_up():
 441  *     if (!(wq_entry->flags & WQ_FLAG_WOKEN))     <full barrier>
 442  *         schedule()                              if (p->state & mode)
 443  *     p->state = TASK_RUNNING;                       p->state = TASK_RUNNING;
 444  *     wq_entry->flags &= ~WQ_FLAG_WOKEN;       ~~~~~~~~~~~~~~~~~~
 445  *     smp_mb(); // B                           condition = true;
 446  * }                                            smp_mb(); // C
 447  * remove_wait_queue(&wq_head, &wait);          wq_entry->flags |= WQ_FLAG_WOKEN;
 448  */
 449 long wait_woken(struct wait_queue_entry *wq_entry, unsigned mode, long timeout)
 450 {
 451         /*
 452          * The below executes an smp_mb(), which matches with the full barrier
 453          * executed by the try_to_wake_up() in woken_wake_function() such that
 454          * either we see the store to wq_entry->flags in woken_wake_function()
 455          * or woken_wake_function() sees our store to current->state.
 456          */
 457         set_current_state(mode); /* A */
 458         if (!(wq_entry->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
 459                 timeout = schedule_timeout(timeout);
 460         __set_current_state(TASK_RUNNING);
 461
 462         /*
 463          * The below executes an smp_mb(), which matches with the smp_mb() (C)
 464          * in woken_wake_function() such that either we see the wait condition
 465          * being true or the store to wq_entry->flags in woken_wake_function()
 466          * follows ours in the coherence order.
 467          */
 468         smp_store_mb(wq_entry->flags, wq_entry->flags & ~WQ_FLAG_WOKEN); /* B */
 469
 470         return timeout;
 471 }
 472 EXPORT_SYMBOL(wait_woken);
 473
 474 int woken_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 475 {
 476         /* Pairs with the smp_store_mb() in wait_woken(). */
 477         smp_mb(); /* C */
 478         wq_entry->flags |= WQ_FLAG_WOKEN;
 479
 480         return default_wake_function(wq_entry, mode, sync, key);
 481 }
 482 EXPORT_SYMBOL(woken_wake_function);