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