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