lib/rwsem.c

   1 /* rwsem.c: R/W semaphores: contention handling functions
   2  *
   3  * Written by David Howells (dhowells@redhat.com).
   4  * Derived from arch/i386/kernel/semaphore.c
   5  *
   6  * Writer lock-stealing by Alex Shi <alex.shi@intel.com>
   7  */
   8 #include <linux/rwsem.h>
   9 #include <linux/sched.h>
  10 #include <linux/init.h>
  11 #include <linux/export.h>
  12
  13 /*
  14  * Initialize an rwsem:
  15  */
  16 void __init_rwsem(struct rw_semaphore *sem, const char *name,
  17                   struct lock_class_key *key)
  18 {
  19 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  20         /*
  21          * Make sure we are not reinitializing a held semaphore:
  22          */
  23         debug_check_no_locks_freed((void *)sem, sizeof(*sem));
  24         lockdep_init_map(&sem->dep_map, name, key, 0);
  25 #endif
  26         sem->count = RWSEM_UNLOCKED_VALUE;
  27         raw_spin_lock_init(&sem->wait_lock);
  28         INIT_LIST_HEAD(&sem->wait_list);
  29 }
  30
  31 EXPORT_SYMBOL(__init_rwsem);
  32
  33 enum rwsem_waiter_type {
  34         RWSEM_WAITING_FOR_WRITE,
  35         RWSEM_WAITING_FOR_READ
  36 };
  37
  38 struct rwsem_waiter {
  39         struct list_head list;
  40         struct task_struct *task;
  41         enum rwsem_waiter_type type;
  42 };
  43
  44 /* Wake types for __rwsem_do_wake().  Note that RWSEM_WAKE_NO_ACTIVE and
  45  * RWSEM_WAKE_READ_OWNED imply that the spinlock must have been kept held
  46  * since the rwsem value was observed.
  47  */
  48 #define RWSEM_WAKE_ANY        0 /* Wake whatever's at head of wait list */
  49 #define RWSEM_WAKE_NO_ACTIVE  1 /* rwsem was observed with no active thread */
  50 #define RWSEM_WAKE_READ_OWNED 2 /* rwsem was observed to be read owned */
  51
  52 /*
  53  * handle the lock release when processes blocked on it that can now run
  54  * - if we come here from up_xxxx(), then:
  55  *   - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed)
  56  *   - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so)
  57  * - there must be someone on the queue
  58  * - the spinlock must be held by the caller
  59  * - woken process blocks are discarded from the list after having task zeroed
  60  * - writers are only woken if downgrading is false
  61  */
  62 static struct rw_semaphore *
  63 __rwsem_do_wake(struct rw_semaphore *sem, int wake_type)
  64 {
  65         struct rwsem_waiter *waiter;
  66         struct task_struct *tsk;
  67         struct list_head *next;
  68         signed long woken, loop, adjustment;
  69
  70         waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
  71         if (waiter->type != RWSEM_WAITING_FOR_WRITE)
  72                 goto readers_only;
  73
  74         if (wake_type == RWSEM_WAKE_READ_OWNED)
  75                 /* Another active reader was observed, so wakeup is not
  76                  * likely to succeed. Save the atomic op.
  77                  */
  78                 goto out;
  79
  80         /* Wake up the writing waiter and let the task grab the sem: */
  81         wake_up_process(waiter->task);
  82         goto out;
  83
  84  readers_only:
  85         /* If we come here from up_xxxx(), another thread might have reached
  86          * rwsem_down_failed_common() before we acquired the spinlock and
  87          * woken up a waiter, making it now active.  We prefer to check for
  88          * this first in order to not spend too much time with the spinlock
  89          * held if we're not going to be able to wake up readers in the end.
  90          *
  91          * Note that we do not need to update the rwsem count: any writer
  92          * trying to acquire rwsem will run rwsem_down_write_failed() due
  93          * to the waiting threads and block trying to acquire the spinlock.
  94          *
  95          * We use a dummy atomic update in order to acquire the cache line
  96          * exclusively since we expect to succeed and run the final rwsem
  97          * count adjustment pretty soon.
  98          */
  99         if (wake_type == RWSEM_WAKE_ANY &&
 100             rwsem_atomic_update(0, sem) < RWSEM_WAITING_BIAS)
 101                 /* Someone grabbed the sem for write already */
 102                 goto out;
 103
 104         /* Grant an infinite number of read locks to the readers at the front
 105          * of the queue.  Note we increment the 'active part' of the count by
 106          * the number of readers before waking any processes up.
 107          */
 108         woken = 0;
 109         do {
 110                 woken++;
 111
 112                 if (waiter->list.next == &sem->wait_list)
 113                         break;
 114
 115                 waiter = list_entry(waiter->list.next,
 116                                         struct rwsem_waiter, list);
 117
 118         } while (waiter->type != RWSEM_WAITING_FOR_WRITE);
 119
 120         adjustment = woken * RWSEM_ACTIVE_READ_BIAS;
 121         if (waiter->type != RWSEM_WAITING_FOR_WRITE)
 122                 /* hit end of list above */
 123                 adjustment -= RWSEM_WAITING_BIAS;
 124
 125         rwsem_atomic_add(adjustment, sem);
 126
 127         next = sem->wait_list.next;
 128         for (loop = woken; loop > 0; loop--) {
 129                 waiter = list_entry(next, struct rwsem_waiter, list);
 130                 next = waiter->list.next;
 131                 tsk = waiter->task;
 132                 smp_mb();
 133                 waiter->task = NULL;
 134                 wake_up_process(tsk);
 135                 put_task_struct(tsk);
 136         }
 137
 138         sem->wait_list.next = next;
 139         next->prev = &sem->wait_list;
 140
 141  out:
 142         return sem;
 143 }
 144
 145 /*
 146  * wait for the read lock to be granted
 147  */
 148 struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
 149 {
 150         signed long adjustment = -RWSEM_ACTIVE_READ_BIAS;
 151         struct rwsem_waiter waiter;
 152         struct task_struct *tsk = current;
 153         signed long count;
 154
 155         /* set up my own style of waitqueue */
 156         waiter.task = tsk;
 157         waiter.type = RWSEM_WAITING_FOR_READ;
 158         get_task_struct(tsk);
 159
 160         raw_spin_lock_irq(&sem->wait_lock);
 161         if (list_empty(&sem->wait_list))
 162                 adjustment += RWSEM_WAITING_BIAS;
 163         list_add_tail(&waiter.list, &sem->wait_list);
 164
 165         /* we're now waiting on the lock, but no longer actively locking */
 166         count = rwsem_atomic_update(adjustment, sem);
 167
 168         /* If there are no active locks, wake the front queued process(es). */
 169         if (count == RWSEM_WAITING_BIAS)
 170                 sem = __rwsem_do_wake(sem, RWSEM_WAKE_NO_ACTIVE);
 171
 172         raw_spin_unlock_irq(&sem->wait_lock);
 173
 174         /* wait to be given the lock */
 175         while (true) {
 176                 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 177                 if (!waiter.task)
 178                         break;
 179                 schedule();
 180         }
 181
 182         tsk->state = TASK_RUNNING;
 183
 184         return sem;
 185 }
 186
 187 /*
 188  * wait until we successfully acquire the write lock
 189  */
 190 struct rw_semaphore __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
 191 {
 192         signed long adjustment = -RWSEM_ACTIVE_WRITE_BIAS;
 193         struct rwsem_waiter waiter;
 194         struct task_struct *tsk = current;
 195         signed long count;
 196
 197         /* set up my own style of waitqueue */
 198         waiter.task = tsk;
 199         waiter.type = RWSEM_WAITING_FOR_WRITE;
 200
 201         raw_spin_lock_irq(&sem->wait_lock);
 202         if (list_empty(&sem->wait_list))
 203                 adjustment += RWSEM_WAITING_BIAS;
 204         list_add_tail(&waiter.list, &sem->wait_list);
 205
 206         /* we're now waiting on the lock, but no longer actively locking */
 207         count = rwsem_atomic_update(adjustment, sem);
 208
 209         /* If there were already threads queued before us and there are no
 210          * active writers, the lock must be read owned; so we try to wake
 211          * any read locks that were queued ahead of us. */
 212         if (count > RWSEM_WAITING_BIAS &&
 213             adjustment == -RWSEM_ACTIVE_WRITE_BIAS)
 214                 sem = __rwsem_do_wake(sem, RWSEM_WAKE_READ_OWNED);
 215
 216         /* wait until we successfully acquire the lock */
 217         set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 218         while (true) {
 219
 220                 /* Try acquiring the write lock. */
 221                 count = RWSEM_ACTIVE_WRITE_BIAS;
 222                 if (!list_is_singular(&sem->wait_list))
 223                         count += RWSEM_WAITING_BIAS;
 224                 if (cmpxchg(&sem->count, RWSEM_WAITING_BIAS, count) ==
 225                                                         RWSEM_WAITING_BIAS)
 226                         break;
 227
 228                 raw_spin_unlock_irq(&sem->wait_lock);
 229
 230                 /* Block until there are no active lockers. */
 231                 do {
 232                         schedule();
 233                         set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 234                 } while (sem->count & RWSEM_ACTIVE_MASK);
 235
 236                 raw_spin_lock_irq(&sem->wait_lock);
 237         }
 238
 239         list_del(&waiter.list);
 240         raw_spin_unlock_irq(&sem->wait_lock);
 241         tsk->state = TASK_RUNNING;
 242
 243         return sem;
 244 }
 245
 246 /*
 247  * handle waking up a waiter on the semaphore
 248  * - up_read/up_write has decremented the active part of count if we come here
 249  */
 250 struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
 251 {
 252         unsigned long flags;
 253
 254         raw_spin_lock_irqsave(&sem->wait_lock, flags);
 255
 256         /* do nothing if list empty */
 257         if (!list_empty(&sem->wait_list))
 258                 sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
 259
 260         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
 261
 262         return sem;
 263 }
 264
 265 /*
 266  * downgrade a write lock into a read lock
 267  * - caller incremented waiting part of count and discovered it still negative
 268  * - just wake up any readers at the front of the queue
 269  */
 270 struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
 271 {
 272         unsigned long flags;
 273
 274         raw_spin_lock_irqsave(&sem->wait_lock, flags);
 275
 276         /* do nothing if list empty */
 277         if (!list_empty(&sem->wait_list))
 278                 sem = __rwsem_do_wake(sem, RWSEM_WAKE_READ_OWNED);
 279
 280         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
 281
 282         return sem;
 283 }
 284
 285 EXPORT_SYMBOL(rwsem_down_read_failed);
 286 EXPORT_SYMBOL(rwsem_down_write_failed);
 287 EXPORT_SYMBOL(rwsem_wake);
 288 EXPORT_SYMBOL(rwsem_downgrade_wake);