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 #include <linux/poll.h>
  15
  16 void __init_waitqueue_head(struct wait_queue_head *wq_head, const char *name, struct lock_class_key *key)
  17 {
  18         spin_lock_init(&wq_head->lock);
  19         lockdep_set_class_and_name(&wq_head->lock, key, name);
  20         INIT_LIST_HEAD(&wq_head->head);
  21 }
  22
  23 EXPORT_SYMBOL(__init_waitqueue_head);
  24
  25 void add_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  26 {
  27         unsigned long flags;
  28
  29         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
  30         spin_lock_irqsave(&wq_head->lock, flags);
  31         __add_wait_queue(wq_head, wq_entry);
  32         spin_unlock_irqrestore(&wq_head->lock, flags);
  33 }
  34 EXPORT_SYMBOL(add_wait_queue);
  35
  36 void add_wait_queue_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  37 {
  38         unsigned long flags;
  39
  40         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
  41         spin_lock_irqsave(&wq_head->lock, flags);
  42         __add_wait_queue_entry_tail(wq_head, wq_entry);
  43         spin_unlock_irqrestore(&wq_head->lock, flags);
  44 }
  45 EXPORT_SYMBOL(add_wait_queue_exclusive);
  46
  47 void remove_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  48 {
  49         unsigned long flags;
  50
  51         spin_lock_irqsave(&wq_head->lock, flags);
  52         __remove_wait_queue(wq_head, wq_entry);
  53         spin_unlock_irqrestore(&wq_head->lock, flags);
  54 }
  55 EXPORT_SYMBOL(remove_wait_queue);
  56
  57 /*
  58  * Scan threshold to break wait queue walk.
  59  * This allows a waker to take a break from holding the
  60  * wait queue lock during the wait queue walk.
  61  */
  62 #define WAITQUEUE_WALK_BREAK_CNT 64
  63
  64 /*
  65  * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
  66  * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
  67  * number) then we wake all the non-exclusive tasks and one exclusive task.
  68  *
  69  * There are circumstances in which we can try to wake a task which has already
  70  * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
  71  * zero in this (rare) case, and we handle it by continuing to scan the queue.
  72  */
  73 static int __wake_up_common(struct wait_queue_head *wq_head, unsigned int mode,
  74                         int nr_exclusive, int wake_flags, void *key,
  75                         wait_queue_entry_t *bookmark)
  76 {
  77         wait_queue_entry_t *curr, *next;
  78         int cnt = 0;
  79
  80         if (bookmark && (bookmark->flags & WQ_FLAG_BOOKMARK)) {
  81                 curr = list_next_entry(bookmark, entry);
  82
  83                 list_del(&bookmark->entry);
  84                 bookmark->flags = 0;
  85         } else
  86                 curr = list_first_entry(&wq_head->head, wait_queue_entry_t, entry);
  87
  88         if (&curr->entry == &wq_head->head)
  89                 return nr_exclusive;
  90
  91         list_for_each_entry_safe_from(curr, next, &wq_head->head, entry) {
  92                 unsigned flags = curr->flags;
  93                 int ret;
  94
  95                 if (flags & WQ_FLAG_BOOKMARK)
  96                         continue;
  97
  98                 ret = curr->func(curr, mode, wake_flags, key);
  99                 if (ret < 0)
 100                         break;
 101                 if (ret && (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
 102                         break;
 103
 104                 if (bookmark && (++cnt > WAITQUEUE_WALK_BREAK_CNT) &&
 105                                 (&next->entry != &wq_head->head)) {
 106                         bookmark->flags = WQ_FLAG_BOOKMARK;
 107                         list_add_tail(&bookmark->entry, &next->entry);
 108                         break;
 109                 }
 110         }
 111         return nr_exclusive;
 112 }
 113
 114 static void __wake_up_common_lock(struct wait_queue_head *wq_head, unsigned int mode,
 115                         int nr_exclusive, int wake_flags, void *key)
 116 {
 117         unsigned long flags;
 118         wait_queue_entry_t bookmark;
 119
 120         bookmark.flags = 0;
 121         bookmark.private = NULL;
 122         bookmark.func = NULL;
 123         INIT_LIST_HEAD(&bookmark.entry);
 124
 125         spin_lock_irqsave(&wq_head->lock, flags);
 126         nr_exclusive = __wake_up_common(wq_head, mode, nr_exclusive, wake_flags, key, &bookmark);
 127         spin_unlock_irqrestore(&wq_head->lock, flags);
 128
 129         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
 130                 spin_lock_irqsave(&wq_head->lock, flags);
 131                 nr_exclusive = __wake_up_common(wq_head, mode, nr_exclusive,
 132                                                 wake_flags, key, &bookmark);
 133                 spin_unlock_irqrestore(&wq_head->lock, flags);
 134         }
 135 }
 136
 137 /**
 138  * __wake_up - wake up threads blocked on a waitqueue.
 139  * @wq_head: the waitqueue
 140  * @mode: which threads
 141  * @nr_exclusive: how many wake-one or wake-many threads to wake up
 142  * @key: is directly passed to the wakeup function
 143  *
 144  * It may be assumed that this function implies a write memory barrier before
 145  * changing the task state if and only if any tasks are woken up.
 146  */
 147 void __wake_up(struct wait_queue_head *wq_head, unsigned int mode,
 148                         int nr_exclusive, void *key)
 149 {
 150         __wake_up_common_lock(wq_head, mode, nr_exclusive, 0, key);
 151 }
 152 EXPORT_SYMBOL(__wake_up);
 153
 154 /*
 155  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
 156  */
 157 void __wake_up_locked(struct wait_queue_head *wq_head, unsigned int mode, int nr)
 158 {
 159         __wake_up_common(wq_head, mode, nr, 0, NULL, NULL);
 160 }
 161 EXPORT_SYMBOL_GPL(__wake_up_locked);
 162
 163 void __wake_up_locked_key(struct wait_queue_head *wq_head, unsigned int mode, void *key)
 164 {
 165         __wake_up_common(wq_head, mode, 1, 0, key, NULL);
 166 }
 167 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
 168
 169 void __wake_up_locked_key_bookmark(struct wait_queue_head *wq_head,
 170                 unsigned int mode, void *key, wait_queue_entry_t *bookmark)
 171 {
 172         __wake_up_common(wq_head, mode, 1, 0, key, bookmark);
 173 }
 174 EXPORT_SYMBOL_GPL(__wake_up_locked_key_bookmark);
 175
 176 /**
 177  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
 178  * @wq_head: the waitqueue
 179  * @mode: which threads
 180  * @nr_exclusive: how many wake-one or wake-many threads to wake up
 181  * @key: opaque value to be passed to wakeup targets
 182  *
 183  * The sync wakeup differs that the waker knows that it will schedule
 184  * away soon, so while the target thread will be woken up, it will not
 185  * be migrated to another CPU - ie. the two threads are 'synchronized'
 186  * with each other. This can prevent needless bouncing between CPUs.
 187  *
 188  * On UP it can prevent extra preemption.
 189  *
 190  * It may be assumed that this function implies a write memory barrier before
 191  * changing the task state if and only if any tasks are woken up.
 192  */
 193 void __wake_up_sync_key(struct wait_queue_head *wq_head, unsigned int mode,
 194                         int nr_exclusive, void *key)
 195 {
 196         int wake_flags = 1; /* XXX WF_SYNC */
 197
 198         if (unlikely(!wq_head))
 199                 return;
 200
 201         if (unlikely(nr_exclusive != 1))
 202                 wake_flags = 0;
 203
 204         __wake_up_common_lock(wq_head, mode, nr_exclusive, wake_flags, key);
 205 }
 206 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
 207
 208 /*
 209  * __wake_up_sync - see __wake_up_sync_key()
 210  */
 211 void __wake_up_sync(struct wait_queue_head *wq_head, unsigned int mode, int nr_exclusive)
 212 {
 213         __wake_up_sync_key(wq_head, mode, nr_exclusive, NULL);
 214 }
 215 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
 216
 217 void __wake_up_pollfree(struct wait_queue_head *wq_head)
 218 {
 219         __wake_up(wq_head, TASK_NORMAL, 0, (void *)(POLLHUP | POLLFREE));
 220         /* POLLFREE must have cleared the queue. */
 221         WARN_ON_ONCE(waitqueue_active(wq_head));
 222 }
 223
 224 /*
 225  * Note: we use "set_current_state()" _after_ the wait-queue add,
 226  * because we need a memory barrier there on SMP, so that any
 227  * wake-function that tests for the wait-queue being active
 228  * will be guaranteed to see waitqueue addition _or_ subsequent
 229  * tests in this thread will see the wakeup having taken place.
 230  *
 231  * The spin_unlock() itself is semi-permeable and only protects
 232  * one way (it only protects stuff inside the critical region and
 233  * stops them from bleeding out - it would still allow subsequent
 234  * loads to move into the critical region).
 235  */
 236 void
 237 prepare_to_wait(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(wq_head, wq_entry);
 245         set_current_state(state);
 246         spin_unlock_irqrestore(&wq_head->lock, flags);
 247 }
 248 EXPORT_SYMBOL(prepare_to_wait);
 249
 250 void
 251 prepare_to_wait_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 252 {
 253         unsigned long flags;
 254
 255         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
 256         spin_lock_irqsave(&wq_head->lock, flags);
 257         if (list_empty(&wq_entry->entry))
 258                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 259         set_current_state(state);
 260         spin_unlock_irqrestore(&wq_head->lock, flags);
 261 }
 262 EXPORT_SYMBOL(prepare_to_wait_exclusive);
 263
 264 void init_wait_entry(struct wait_queue_entry *wq_entry, int flags)
 265 {
 266         wq_entry->flags = flags;
 267         wq_entry->private = current;
 268         wq_entry->func = autoremove_wake_function;
 269         INIT_LIST_HEAD(&wq_entry->entry);
 270 }
 271 EXPORT_SYMBOL(init_wait_entry);
 272
 273 long prepare_to_wait_event(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 274 {
 275         unsigned long flags;
 276         long ret = 0;
 277
 278         spin_lock_irqsave(&wq_head->lock, flags);
 279         if (unlikely(signal_pending_state(state, current))) {
 280                 /*
 281                  * Exclusive waiter must not fail if it was selected by wakeup,
 282                  * it should "consume" the condition we were waiting for.
 283                  *
 284                  * The caller will recheck the condition and return success if
 285                  * we were already woken up, we can not miss the event because
 286                  * wakeup locks/unlocks the same wq_head->lock.
 287                  *
 288                  * But we need to ensure that set-condition + wakeup after that
 289                  * can't see us, it should wake up another exclusive waiter if
 290                  * we fail.
 291                  */
 292                 list_del_init(&wq_entry->entry);
 293                 ret = -ERESTARTSYS;
 294         } else {
 295                 if (list_empty(&wq_entry->entry)) {
 296                         if (wq_entry->flags & WQ_FLAG_EXCLUSIVE)
 297                                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 298                         else
 299                                 __add_wait_queue(wq_head, wq_entry);
 300                 }
 301                 set_current_state(state);
 302         }
 303         spin_unlock_irqrestore(&wq_head->lock, flags);
 304
 305         return ret;
 306 }
 307 EXPORT_SYMBOL(prepare_to_wait_event);
 308
 309 /*
 310  * Note! These two wait functions are entered with the
 311  * wait-queue lock held (and interrupts off in the _irq
 312  * case), so there is no race with testing the wakeup
 313  * condition in the caller before they add the wait
 314  * entry to the wake queue.
 315  */
 316 int do_wait_intr(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 317 {
 318         if (likely(list_empty(&wait->entry)))
 319                 __add_wait_queue_entry_tail(wq, wait);
 320
 321         set_current_state(TASK_INTERRUPTIBLE);
 322         if (signal_pending(current))
 323                 return -ERESTARTSYS;
 324
 325         spin_unlock(&wq->lock);
 326         schedule();
 327         spin_lock(&wq->lock);
 328         return 0;
 329 }
 330 EXPORT_SYMBOL(do_wait_intr);
 331
 332 int do_wait_intr_irq(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 333 {
 334         if (likely(list_empty(&wait->entry)))
 335                 __add_wait_queue_entry_tail(wq, wait);
 336
 337         set_current_state(TASK_INTERRUPTIBLE);
 338         if (signal_pending(current))
 339                 return -ERESTARTSYS;
 340
 341         spin_unlock_irq(&wq->lock);
 342         schedule();
 343         spin_lock_irq(&wq->lock);
 344         return 0;
 345 }
 346 EXPORT_SYMBOL(do_wait_intr_irq);
 347
 348 /**
 349  * finish_wait - clean up after waiting in a queue
 350  * @wq_head: waitqueue waited on
 351  * @wq_entry: wait descriptor
 352  *
 353  * Sets current thread back to running state and removes
 354  * the wait descriptor from the given waitqueue if still
 355  * queued.
 356  */
 357 void finish_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
 358 {
 359         unsigned long flags;
 360
 361         __set_current_state(TASK_RUNNING);
 362         /*
 363          * We can check for list emptiness outside the lock
 364          * IFF:
 365          *  - we use the "careful" check that verifies both
 366          *    the next and prev pointers, so that there cannot
 367          *    be any half-pending updates in progress on other
 368          *    CPU's that we haven't seen yet (and that might
 369          *    still change the stack area.
 370          * and
 371          *  - all other users take the lock (ie we can only
 372          *    have _one_ other CPU that looks at or modifies
 373          *    the list).
 374          */
 375         if (!list_empty_careful(&wq_entry->entry)) {
 376                 spin_lock_irqsave(&wq_head->lock, flags);
 377                 list_del_init(&wq_entry->entry);
 378                 spin_unlock_irqrestore(&wq_head->lock, flags);
 379         }
 380 }
 381 EXPORT_SYMBOL(finish_wait);
 382
 383 int autoremove_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 384 {
 385         int ret = default_wake_function(wq_entry, mode, sync, key);
 386
 387         if (ret)
 388                 list_del_init(&wq_entry->entry);
 389         return ret;
 390 }
 391 EXPORT_SYMBOL(autoremove_wake_function);
 392
 393 static inline bool is_kthread_should_stop(void)
 394 {
 395         return (current->flags & PF_KTHREAD) && kthread_should_stop();
 396 }
 397
 398 /*
 399  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
 400  *
 401  * add_wait_queue(&wq_head, &wait);
 402  * for (;;) {
 403  *     if (condition)
 404  *         break;
 405  *
 406  *     // in wait_woken()                       // in woken_wake_function()
 407  *
 408  *     p->state = mode;                         wq_entry->flags |= WQ_FLAG_WOKEN;
 409  *     smp_mb(); // A                           try_to_wake_up():
 410  *     if (!(wq_entry->flags & WQ_FLAG_WOKEN))     <full barrier>
 411  *         schedule()                              if (p->state & mode)
 412  *     p->state = TASK_RUNNING;                       p->state = TASK_RUNNING;
 413  *     wq_entry->flags &= ~WQ_FLAG_WOKEN;       ~~~~~~~~~~~~~~~~~~
 414  *     smp_mb(); // B                           condition = true;
 415  * }                                            smp_mb(); // C
 416  * remove_wait_queue(&wq_head, &wait);          wq_entry->flags |= WQ_FLAG_WOKEN;
 417  */
 418 long wait_woken(struct wait_queue_entry *wq_entry, unsigned mode, long timeout)
 419 {
 420         /*
 421          * The below executes an smp_mb(), which matches with the full barrier
 422          * executed by the try_to_wake_up() in woken_wake_function() such that
 423          * either we see the store to wq_entry->flags in woken_wake_function()
 424          * or woken_wake_function() sees our store to current->state.
 425          */
 426         set_current_state(mode); /* A */
 427         if (!(wq_entry->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
 428                 timeout = schedule_timeout(timeout);
 429         __set_current_state(TASK_RUNNING);
 430
 431         /*
 432          * The below executes an smp_mb(), which matches with the smp_mb() (C)
 433          * in woken_wake_function() such that either we see the wait condition
 434          * being true or the store to wq_entry->flags in woken_wake_function()
 435          * follows ours in the coherence order.
 436          */
 437         smp_store_mb(wq_entry->flags, wq_entry->flags & ~WQ_FLAG_WOKEN); /* B */
 438
 439         return timeout;
 440 }
 441 EXPORT_SYMBOL(wait_woken);
 442
 443 int woken_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 444 {
 445         /* Pairs with the smp_store_mb() in wait_woken(). */
 446         smp_mb(); /* C */
 447         wq_entry->flags |= WQ_FLAG_WOKEN;
 448
 449         return default_wake_function(wq_entry, mode, sync, key);
 450 }
 451 EXPORT_SYMBOL(woken_wake_function);