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