kernel/sched/wait.c

   1 /*
   2  * Generic waiting primitives.
   3  *
   4  * (C) 2004 Nadia Yvette Chambers, Oracle
   5  */
   6 #include <linux/init.h>
   7 #include <linux/export.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/sched/debug.h>
  10 #include <linux/mm.h>
  11 #include <linux/wait.h>
  12 #include <linux/hash.h>
  13 #include <linux/kthread.h>
  14
  15 void __init_waitqueue_head(struct wait_queue_head *wq_head, const char *name, struct lock_class_key *key)
  16 {
  17         spin_lock_init(&wq_head->lock);
  18         lockdep_set_class_and_name(&wq_head->lock, key, name);
  19         INIT_LIST_HEAD(&wq_head->head);
  20 }
  21
  22 EXPORT_SYMBOL(__init_waitqueue_head);
  23
  24 void add_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  25 {
  26         unsigned long flags;
  27
  28         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
  29         spin_lock_irqsave(&wq_head->lock, flags);
  30         __add_wait_queue_entry_tail(wq_head, wq_entry);
  31         spin_unlock_irqrestore(&wq_head->lock, flags);
  32 }
  33 EXPORT_SYMBOL(add_wait_queue);
  34
  35 void add_wait_queue_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  36 {
  37         unsigned long flags;
  38
  39         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
  40         spin_lock_irqsave(&wq_head->lock, flags);
  41         __add_wait_queue_entry_tail(wq_head, wq_entry);
  42         spin_unlock_irqrestore(&wq_head->lock, flags);
  43 }
  44 EXPORT_SYMBOL(add_wait_queue_exclusive);
  45
  46 void remove_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  47 {
  48         unsigned long flags;
  49
  50         spin_lock_irqsave(&wq_head->lock, flags);
  51         __remove_wait_queue(wq_head, wq_entry);
  52         spin_unlock_irqrestore(&wq_head->lock, flags);
  53 }
  54 EXPORT_SYMBOL(remove_wait_queue);
  55
  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 void __wake_up_common(struct wait_queue_head *wq_head, unsigned int mode,
  67                         int nr_exclusive, int wake_flags, void *key)
  68 {
  69         wait_queue_entry_t *curr, *next;
  70
  71         list_for_each_entry_safe(curr, next, &wq_head->head, entry) {
  72                 unsigned flags = curr->flags;
  73
  74                 if (curr->func(curr, mode, wake_flags, key) &&
  75                                 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
  76                         break;
  77         }
  78 }
  79
  80 /**
  81  * __wake_up - wake up threads blocked on a waitqueue.
  82  * @wq_head: the waitqueue
  83  * @mode: which threads
  84  * @nr_exclusive: how many wake-one or wake-many threads to wake up
  85  * @key: is directly passed to the wakeup function
  86  *
  87  * It may be assumed that this function implies a write memory barrier before
  88  * changing the task state if and only if any tasks are woken up.
  89  */
  90 void __wake_up(struct wait_queue_head *wq_head, unsigned int mode,
  91                         int nr_exclusive, void *key)
  92 {
  93         unsigned long flags;
  94
  95         spin_lock_irqsave(&wq_head->lock, flags);
  96         __wake_up_common(wq_head, mode, nr_exclusive, 0, key);
  97         spin_unlock_irqrestore(&wq_head->lock, flags);
  98 }
  99 EXPORT_SYMBOL(__wake_up);
 100
 101 /*
 102  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
 103  */
 104 void __wake_up_locked(struct wait_queue_head *wq_head, unsigned int mode, int nr)
 105 {
 106         __wake_up_common(wq_head, mode, nr, 0, NULL);
 107 }
 108 EXPORT_SYMBOL_GPL(__wake_up_locked);
 109
 110 void __wake_up_locked_key(struct wait_queue_head *wq_head, unsigned int mode, void *key)
 111 {
 112         __wake_up_common(wq_head, mode, 1, 0, key);
 113 }
 114 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
 115
 116 /**
 117  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
 118  * @wq_head: the waitqueue
 119  * @mode: which threads
 120  * @nr_exclusive: how many wake-one or wake-many threads to wake up
 121  * @key: opaque value to be passed to wakeup targets
 122  *
 123  * The sync wakeup differs that the waker knows that it will schedule
 124  * away soon, so while the target thread will be woken up, it will not
 125  * be migrated to another CPU - ie. the two threads are 'synchronized'
 126  * with each other. This can prevent needless bouncing between CPUs.
 127  *
 128  * On UP it can prevent extra preemption.
 129  *
 130  * It may be assumed that this function implies a write memory barrier before
 131  * changing the task state if and only if any tasks are woken up.
 132  */
 133 void __wake_up_sync_key(struct wait_queue_head *wq_head, unsigned int mode,
 134                         int nr_exclusive, void *key)
 135 {
 136         unsigned long flags;
 137         int wake_flags = 1; /* XXX WF_SYNC */
 138
 139         if (unlikely(!wq_head))
 140                 return;
 141
 142         if (unlikely(nr_exclusive != 1))
 143                 wake_flags = 0;
 144
 145         spin_lock_irqsave(&wq_head->lock, flags);
 146         __wake_up_common(wq_head, mode, nr_exclusive, wake_flags, key);
 147         spin_unlock_irqrestore(&wq_head->lock, flags);
 148 }
 149 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
 150
 151 /*
 152  * __wake_up_sync - see __wake_up_sync_key()
 153  */
 154 void __wake_up_sync(struct wait_queue_head *wq_head, unsigned int mode, int nr_exclusive)
 155 {
 156         __wake_up_sync_key(wq_head, mode, nr_exclusive, NULL);
 157 }
 158 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
 159
 160 /*
 161  * Note: we use "set_current_state()" _after_ the wait-queue add,
 162  * because we need a memory barrier there on SMP, so that any
 163  * wake-function that tests for the wait-queue being active
 164  * will be guaranteed to see waitqueue addition _or_ subsequent
 165  * tests in this thread will see the wakeup having taken place.
 166  *
 167  * The spin_unlock() itself is semi-permeable and only protects
 168  * one way (it only protects stuff inside the critical region and
 169  * stops them from bleeding out - it would still allow subsequent
 170  * loads to move into the critical region).
 171  */
 172 void
 173 prepare_to_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 174 {
 175         unsigned long flags;
 176
 177         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
 178         spin_lock_irqsave(&wq_head->lock, flags);
 179         if (list_empty(&wq_entry->entry))
 180                 __add_wait_queue(wq_head, wq_entry);
 181         set_current_state(state);
 182         spin_unlock_irqrestore(&wq_head->lock, flags);
 183 }
 184 EXPORT_SYMBOL(prepare_to_wait);
 185
 186 void
 187 prepare_to_wait_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 188 {
 189         unsigned long flags;
 190
 191         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
 192         spin_lock_irqsave(&wq_head->lock, flags);
 193         if (list_empty(&wq_entry->entry))
 194                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 195         set_current_state(state);
 196         spin_unlock_irqrestore(&wq_head->lock, flags);
 197 }
 198 EXPORT_SYMBOL(prepare_to_wait_exclusive);
 199
 200 void init_wait_entry(struct wait_queue_entry *wq_entry, int flags)
 201 {
 202         wq_entry->flags = flags;
 203         wq_entry->private = current;
 204         wq_entry->func = autoremove_wake_function;
 205         INIT_LIST_HEAD(&wq_entry->entry);
 206 }
 207 EXPORT_SYMBOL(init_wait_entry);
 208
 209 long prepare_to_wait_event(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 210 {
 211         unsigned long flags;
 212         long ret = 0;
 213
 214         spin_lock_irqsave(&wq_head->lock, flags);
 215         if (unlikely(signal_pending_state(state, current))) {
 216                 /*
 217                  * Exclusive waiter must not fail if it was selected by wakeup,
 218                  * it should "consume" the condition we were waiting for.
 219                  *
 220                  * The caller will recheck the condition and return success if
 221                  * we were already woken up, we can not miss the event because
 222                  * wakeup locks/unlocks the same wq_head->lock.
 223                  *
 224                  * But we need to ensure that set-condition + wakeup after that
 225                  * can't see us, it should wake up another exclusive waiter if
 226                  * we fail.
 227                  */
 228                 list_del_init(&wq_entry->entry);
 229                 ret = -ERESTARTSYS;
 230         } else {
 231                 if (list_empty(&wq_entry->entry)) {
 232                         if (wq_entry->flags & WQ_FLAG_EXCLUSIVE)
 233                                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 234                         else
 235                                 __add_wait_queue(wq_head, wq_entry);
 236                 }
 237                 set_current_state(state);
 238         }
 239         spin_unlock_irqrestore(&wq_head->lock, flags);
 240
 241         return ret;
 242 }
 243 EXPORT_SYMBOL(prepare_to_wait_event);
 244
 245 /*
 246  * Note! These two wait functions are entered with the
 247  * wait-queue lock held (and interrupts off in the _irq
 248  * case), so there is no race with testing the wakeup
 249  * condition in the caller before they add the wait
 250  * entry to the wake queue.
 251  */
 252 int do_wait_intr(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 253 {
 254         if (likely(list_empty(&wait->entry)))
 255                 __add_wait_queue_entry_tail(wq, wait);
 256
 257         set_current_state(TASK_INTERRUPTIBLE);
 258         if (signal_pending(current))
 259                 return -ERESTARTSYS;
 260
 261         spin_unlock(&wq->lock);
 262         schedule();
 263         spin_lock(&wq->lock);
 264         return 0;
 265 }
 266 EXPORT_SYMBOL(do_wait_intr);
 267
 268 int do_wait_intr_irq(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 269 {
 270         if (likely(list_empty(&wait->entry)))
 271                 __add_wait_queue_entry_tail(wq, wait);
 272
 273         set_current_state(TASK_INTERRUPTIBLE);
 274         if (signal_pending(current))
 275                 return -ERESTARTSYS;
 276
 277         spin_unlock_irq(&wq->lock);
 278         schedule();
 279         spin_lock_irq(&wq->lock);
 280         return 0;
 281 }
 282 EXPORT_SYMBOL(do_wait_intr_irq);
 283
 284 /**
 285  * finish_wait - clean up after waiting in a queue
 286  * @wq_head: waitqueue waited on
 287  * @wq_entry: wait descriptor
 288  *
 289  * Sets current thread back to running state and removes
 290  * the wait descriptor from the given waitqueue if still
 291  * queued.
 292  */
 293 void finish_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
 294 {
 295         unsigned long flags;
 296
 297         __set_current_state(TASK_RUNNING);
 298         /*
 299          * We can check for list emptiness outside the lock
 300          * IFF:
 301          *  - we use the "careful" check that verifies both
 302          *    the next and prev pointers, so that there cannot
 303          *    be any half-pending updates in progress on other
 304          *    CPU's that we haven't seen yet (and that might
 305          *    still change the stack area.
 306          * and
 307          *  - all other users take the lock (ie we can only
 308          *    have _one_ other CPU that looks at or modifies
 309          *    the list).
 310          */
 311         if (!list_empty_careful(&wq_entry->entry)) {
 312                 spin_lock_irqsave(&wq_head->lock, flags);
 313                 list_del_init(&wq_entry->entry);
 314                 spin_unlock_irqrestore(&wq_head->lock, flags);
 315         }
 316 }
 317 EXPORT_SYMBOL(finish_wait);
 318
 319 int autoremove_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 320 {
 321         int ret = default_wake_function(wq_entry, mode, sync, key);
 322
 323         if (ret)
 324                 list_del_init(&wq_entry->entry);
 325         return ret;
 326 }
 327 EXPORT_SYMBOL(autoremove_wake_function);
 328
 329 static inline bool is_kthread_should_stop(void)
 330 {
 331         return (current->flags & PF_KTHREAD) && kthread_should_stop();
 332 }
 333
 334 /*
 335  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
 336  *
 337  * add_wait_queue(&wq_head, &wait);
 338  * for (;;) {
 339  *     if (condition)
 340  *         break;
 341  *
 342  *     p->state = mode;                         condition = true;
 343  *     smp_mb(); // A                           smp_wmb(); // C
 344  *     if (!wq_entry->flags & WQ_FLAG_WOKEN)    wq_entry->flags |= WQ_FLAG_WOKEN;
 345  *         schedule()                           try_to_wake_up();
 346  *     p->state = TASK_RUNNING;             ~~~~~~~~~~~~~~~~~~
 347  *     wq_entry->flags &= ~WQ_FLAG_WOKEN;               condition = true;
 348  *     smp_mb() // B                            smp_wmb(); // C
 349  *                                              wq_entry->flags |= WQ_FLAG_WOKEN;
 350  * }
 351  * remove_wait_queue(&wq_head, &wait);
 352  *
 353  */
 354 long wait_woken(struct wait_queue_entry *wq_entry, unsigned mode, long timeout)
 355 {
 356         set_current_state(mode); /* A */
 357         /*
 358          * The above implies an smp_mb(), which matches with the smp_wmb() from
 359          * woken_wake_function() such that if we observe WQ_FLAG_WOKEN we must
 360          * also observe all state before the wakeup.
 361          */
 362         if (!(wq_entry->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
 363                 timeout = schedule_timeout(timeout);
 364         __set_current_state(TASK_RUNNING);
 365
 366         /*
 367          * The below implies an smp_mb(), it too pairs with the smp_wmb() from
 368          * woken_wake_function() such that we must either observe the wait
 369          * condition being true _OR_ WQ_FLAG_WOKEN such that we will not miss
 370          * an event.
 371          */
 372         smp_store_mb(wq_entry->flags, wq_entry->flags & ~WQ_FLAG_WOKEN); /* B */
 373
 374         return timeout;
 375 }
 376 EXPORT_SYMBOL(wait_woken);
 377
 378 int woken_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 379 {
 380         /*
 381          * Although this function is called under waitqueue lock, LOCK
 382          * doesn't imply write barrier and the users expects write
 383          * barrier semantics on wakeup functions.  The following
 384          * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up()
 385          * and is paired with smp_store_mb() in wait_woken().
 386          */
 387         smp_wmb(); /* C */
 388         wq_entry->flags |= WQ_FLAG_WOKEN;
 389
 390         return default_wake_function(wq_entry, mode, sync, key);
 391 }
 392 EXPORT_SYMBOL(woken_wake_function);