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