kernel/sched/wait.c

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