kernel/wait.c

   1 /*
   2  * Generic waiting primitives.
   3  *
   4  * (C) 2004 William Irwin, Oracle
   5  */
   6 #include <linux/init.h>
   7 #include <linux/module.h>
   8 #include <linux/sched.h>
   9 #include <linux/mm.h>
  10 #include <linux/wait.h>
  11 #include <linux/hash.h>
  12
  13 void __init_waitqueue_head(wait_queue_head_t *q, struct lock_class_key *key)
  14 {
  15         spin_lock_init(&q->lock);
  16         lockdep_set_class(&q->lock, key);
  17         INIT_LIST_HEAD(&q->task_list);
  18 }
  19
  20 EXPORT_SYMBOL(__init_waitqueue_head);
  21
  22 void add_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
  23 {
  24         unsigned long flags;
  25
  26         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
  27         spin_lock_irqsave(&q->lock, flags);
  28         __add_wait_queue(q, wait);
  29         spin_unlock_irqrestore(&q->lock, flags);
  30 }
  31 EXPORT_SYMBOL(add_wait_queue);
  32
  33 void add_wait_queue_exclusive(wait_queue_head_t *q, wait_queue_t *wait)
  34 {
  35         unsigned long flags;
  36
  37         wait->flags |= WQ_FLAG_EXCLUSIVE;
  38         spin_lock_irqsave(&q->lock, flags);
  39         __add_wait_queue_tail(q, wait);
  40         spin_unlock_irqrestore(&q->lock, flags);
  41 }
  42 EXPORT_SYMBOL(add_wait_queue_exclusive);
  43
  44 void remove_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
  45 {
  46         unsigned long flags;
  47
  48         spin_lock_irqsave(&q->lock, flags);
  49         __remove_wait_queue(q, wait);
  50         spin_unlock_irqrestore(&q->lock, flags);
  51 }
  52 EXPORT_SYMBOL(remove_wait_queue);
  53
  54
  55 /*
  56  * Note: we use "set_current_state()" _after_ the wait-queue add,
  57  * because we need a memory barrier there on SMP, so that any
  58  * wake-function that tests for the wait-queue being active
  59  * will be guaranteed to see waitqueue addition _or_ subsequent
  60  * tests in this thread will see the wakeup having taken place.
  61  *
  62  * The spin_unlock() itself is semi-permeable and only protects
  63  * one way (it only protects stuff inside the critical region and
  64  * stops them from bleeding out - it would still allow subsequent
  65  * loads to move into the critical region).
  66  */
  67 void
  68 prepare_to_wait(wait_queue_head_t *q, wait_queue_t *wait, int state)
  69 {
  70         unsigned long flags;
  71
  72         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
  73         spin_lock_irqsave(&q->lock, flags);
  74         if (list_empty(&wait->task_list))
  75                 __add_wait_queue(q, wait);
  76         if (is_sync_wait(wait))
  77                 set_current_state(state);
  78         spin_unlock_irqrestore(&q->lock, flags);
  79 }
  80 EXPORT_SYMBOL(prepare_to_wait);
  81
  82 void
  83 prepare_to_wait_exclusive(wait_queue_head_t *q, wait_queue_t *wait, int state)
  84 {
  85         unsigned long flags;
  86
  87         wait->flags |= WQ_FLAG_EXCLUSIVE;
  88         spin_lock_irqsave(&q->lock, flags);
  89         if (list_empty(&wait->task_list))
  90                 __add_wait_queue_tail(q, wait);
  91         if (is_sync_wait(wait))
  92                 set_current_state(state);
  93         spin_unlock_irqrestore(&q->lock, flags);
  94 }
  95 EXPORT_SYMBOL(prepare_to_wait_exclusive);
  96
  97 /*
  98  * finish_wait - clean up after waiting in a queue
  99  * @q: waitqueue waited on
 100  * @wait: wait descriptor
 101  *
 102  * Sets current thread back to running state and removes
 103  * the wait descriptor from the given waitqueue if still
 104  * queued.
 105  */
 106 void finish_wait(wait_queue_head_t *q, wait_queue_t *wait)
 107 {
 108         unsigned long flags;
 109
 110         __set_current_state(TASK_RUNNING);
 111         /*
 112          * We can check for list emptiness outside the lock
 113          * IFF:
 114          *  - we use the "careful" check that verifies both
 115          *    the next and prev pointers, so that there cannot
 116          *    be any half-pending updates in progress on other
 117          *    CPU's that we haven't seen yet (and that might
 118          *    still change the stack area.
 119          * and
 120          *  - all other users take the lock (ie we can only
 121          *    have _one_ other CPU that looks at or modifies
 122          *    the list).
 123          */
 124         if (!list_empty_careful(&wait->task_list)) {
 125                 spin_lock_irqsave(&q->lock, flags);
 126                 list_del_init(&wait->task_list);
 127                 spin_unlock_irqrestore(&q->lock, flags);
 128         }
 129 }
 130 EXPORT_SYMBOL(finish_wait);
 131
 132 /*
 133  * abort_exclusive_wait - abort exclusive waiting in a queue
 134  * @q: waitqueue waited on
 135  * @wait: wait descriptor
 136  * @state: runstate of the waiter to be woken
 137  * @key: key to identify a wait bit queue or %NULL
 138  *
 139  * Sets current thread back to running state and removes
 140  * the wait descriptor from the given waitqueue if still
 141  * queued.
 142  *
 143  * Wakes up the next waiter if the caller is concurrently
 144  * woken up through the queue.
 145  *
 146  * This prevents waiter starvation where an exclusive waiter
 147  * aborts and is woken up concurrently and noone wakes up
 148  * the next waiter.
 149  */
 150 void abort_exclusive_wait(wait_queue_head_t *q, wait_queue_t *wait,
 151                         unsigned int mode, void *key)
 152 {
 153         unsigned long flags;
 154
 155         __set_current_state(TASK_RUNNING);
 156         spin_lock_irqsave(&q->lock, flags);
 157         if (!list_empty(&wait->task_list))
 158                 list_del_init(&wait->task_list);
 159         else if (waitqueue_active(q))
 160                 __wake_up_locked_key(q, mode, key);
 161         spin_unlock_irqrestore(&q->lock, flags);
 162 }
 163 EXPORT_SYMBOL(abort_exclusive_wait);
 164
 165 int autoremove_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
 166 {
 167         int ret = default_wake_function(wait, mode, sync, key);
 168
 169         if (ret)
 170                 list_del_init(&wait->task_list);
 171         return ret;
 172 }
 173 EXPORT_SYMBOL(autoremove_wake_function);
 174
 175 int wake_bit_function(wait_queue_t *wait, unsigned mode, int sync, void *arg)
 176 {
 177         struct wait_bit_key *key = arg;
 178         struct wait_bit_queue *wait_bit
 179                 = container_of(wait, struct wait_bit_queue, wait);
 180
 181         if (!wait_bit_cleared(wait_bit, key))
 182                 return 0;
 183
 184         return autoremove_wake_function(wait, mode, sync, key);
 185 }
 186 EXPORT_SYMBOL(wake_bit_function);
 187
 188 /*
 189  * To allow interruptible waiting and asynchronous (i.e. nonblocking)
 190  * waiting, the actions of __wait_on_bit() and __wait_on_bit_lock() are
 191  * permitted return codes. Nonzero return codes halt waiting and return.
 192  */
 193 int __sched
 194 __wait_on_bit(wait_queue_head_t *wq, struct wait_bit_queue *q,
 195                         int (*action)(void *), unsigned mode)
 196 {
 197         int ret = 0;
 198
 199         do {
 200                 prepare_to_wait(wq, &q->wait, mode);
 201                 if (test_bit(q->key.bit_nr, q->key.flags)) {
 202                         ret = (*action)(q->key.flags);
 203                         if (ret)
 204                                 break;
 205                 }
 206                 if (!is_sync_wait(&q->wait))
 207                         ret = -EIOCBRETRY;
 208         } while (test_bit(q->key.bit_nr, q->key.flags) && !ret);
 209         finish_wait(wq, &q->wait);
 210         return ret;
 211 }
 212 EXPORT_SYMBOL(__wait_on_bit);
 213
 214 int __sched out_of_line_wait_on_bit(void *word, int bit,
 215                                         int (*action)(void *), unsigned mode)
 216 {
 217         wait_queue_head_t *wq = bit_waitqueue(word, bit);
 218         DEFINE_WAIT_BIT(wait, word, bit);
 219
 220         return __wait_on_bit(wq, &wait, action, mode);
 221 }
 222 EXPORT_SYMBOL(out_of_line_wait_on_bit);
 223
 224 int __sched
 225 __wait_on_bit_lock(wait_queue_head_t *wq, struct wait_bit_queue *q,
 226                         int (*action)(void *), unsigned mode)
 227 {
 228         do {
 229                 int ret;
 230
 231                 prepare_to_wait_exclusive(wq, &q->wait, mode);
 232                 if (!test_bit(q->key.bit_nr, q->key.flags))
 233                         continue;
 234                 ret = action(q->key.flags);
 235                 if (!ret) {
 236                         if (!is_sync_wait(&q->wait))
 237                                 ret = -EIOCBRETRY;
 238                         else
 239                                 continue;
 240                 }
 241                 abort_exclusive_wait(wq, &q->wait, mode, &q->key);
 242                 return ret;
 243         } while (test_and_set_bit(q->key.bit_nr, q->key.flags));
 244         finish_wait(wq, &q->wait);
 245         return 0;
 246 }
 247 EXPORT_SYMBOL(__wait_on_bit_lock);
 248
 249 int __sched out_of_line_wait_on_bit_lock(void *word, int bit,
 250                                         int (*action)(void *), unsigned mode)
 251 {
 252         wait_queue_head_t *wq = bit_waitqueue(word, bit);
 253         DEFINE_WAIT_BIT(wait, word, bit);
 254
 255         return __wait_on_bit_lock(wq, &wait, action, mode);
 256 }
 257 EXPORT_SYMBOL(out_of_line_wait_on_bit_lock);
 258
 259 void __wake_up_bit(wait_queue_head_t *wq, void *word, int bit)
 260 {
 261         struct wait_bit_key key = __WAIT_BIT_KEY_INITIALIZER(word, bit);
 262         if (waitqueue_active(wq))
 263                 __wake_up(wq, TASK_NORMAL, 1, &key);
 264 }
 265 EXPORT_SYMBOL(__wake_up_bit);
 266
 267 /**
 268  * wake_up_bit - wake up a waiter on a bit
 269  * @word: the word being waited on, a kernel virtual address
 270  * @bit: the bit of the word being waited on
 271  *
 272  * There is a standard hashed waitqueue table for generic use. This
 273  * is the part of the hashtable's accessor API that wakes up waiters
 274  * on a bit. For instance, if one were to have waiters on a bitflag,
 275  * one would call wake_up_bit() after clearing the bit.
 276  *
 277  * In order for this to function properly, as it uses waitqueue_active()
 278  * internally, some kind of memory barrier must be done prior to calling
 279  * this. Typically, this will be smp_mb__after_clear_bit(), but in some
 280  * cases where bitflags are manipulated non-atomically under a lock, one
 281  * may need to use a less regular barrier, such fs/inode.c's smp_mb(),
 282  * because spin_unlock() does not guarantee a memory barrier.
 283  */
 284 void wake_up_bit(void *word, int bit)
 285 {
 286         __wake_up_bit(bit_waitqueue(word, bit), word, bit);
 287 }
 288 EXPORT_SYMBOL(wake_up_bit);
 289
 290 wait_queue_head_t *bit_waitqueue(void *word, int bit)
 291 {
 292         const int shift = BITS_PER_LONG == 32 ? 5 : 6;
 293         const struct zone *zone = page_zone(virt_to_page(word));
 294         unsigned long val = (unsigned long)word << shift | bit;
 295
 296         return &zone->wait_table[hash_long(val, zone->wait_table_bits)];
 297 }
 298 EXPORT_SYMBOL(bit_waitqueue);