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 #include <linux/rwsem.h>
   7 #include <linux/sched.h>
   8 #include <linux/mm.h>
   9 #include <linux/module.h>
  10
  11 struct rwsem_waiter {
  12         struct list_head        list;
  13         struct task_struct      *task;
  14         unsigned int            flags;
  15 #define RWSEM_WAITING_FOR_READ  0x00000001
  16 #define RWSEM_WAITING_FOR_WRITE 0x00000002
  17 };
  18
  19 #if RWSEM_DEBUG
  20 #undef rwsemtrace
  21 void rwsemtrace(struct rw_semaphore *sem, const char *str)
  22 {
  23         printk("sem=%p\n",sem);
  24         printk("(sem)=%08lx\n",sem->count);
  25         if (sem->debug)
  26                 printk("[%d] %s({%08lx})\n",current->pid,str,sem->count);
  27 }
  28 #endif
  29
  30 /*
  31  * handle the lock being released whilst there are processes blocked on it that can now run
  32  * - if we come here, then:
  33  *   - the 'active part' of the count (&0x0000ffff) reached zero but has been re-incremented
  34  *   - the 'waiting part' of the count (&0xffff0000) is negative (and will still be so)
  35  *   - there must be someone on the queue
  36  * - the spinlock must be held by the caller
  37  * - woken process blocks are discarded from the list after having flags zeroised
  38  */
  39 static inline struct rw_semaphore *__rwsem_do_wake(struct rw_semaphore *sem)
  40 {
  41         struct rwsem_waiter *waiter;
  42         struct task_struct *tsk;
  43         struct list_head *next;
  44         signed long oldcount, woken, loop;
  45
  46         rwsemtrace(sem,"Entering __rwsem_do_wake");
  47
  48         /* only wake someone up if we can transition the active part of the count from 0 -> 1 */
  49  try_again:
  50         oldcount = rwsem_atomic_update(RWSEM_ACTIVE_BIAS,sem) - RWSEM_ACTIVE_BIAS;
  51         if (oldcount & RWSEM_ACTIVE_MASK)
  52                 goto undo;
  53
  54         waiter = list_entry(sem->wait_list.next,struct rwsem_waiter,list);
  55
  56         /* try to grant a single write lock if there's a writer at the front of the queue
  57          * - note we leave the 'active part' of the count incremented by 1 and the waiting part
  58          *   incremented by 0x00010000
  59          */
  60         if (!(waiter->flags & RWSEM_WAITING_FOR_WRITE))
  61                 goto readers_only;
  62
  63         list_del(&waiter->list);
  64         tsk = waiter->task;
  65         mb();
  66         waiter->task = NULL;
  67         wake_up_process(tsk);
  68         free_task_struct(tsk);
  69         goto out;
  70
  71         /* grant an infinite number of read locks to the readers at the front of the queue
  72          * - note we increment the 'active part' of the count by the number of readers (less one
  73          *   for the activity decrement we've already done) before waking any processes up
  74          */
  75  readers_only:
  76         woken = 0;
  77         do {
  78                 woken++;
  79
  80                 if (waiter->list.next==&sem->wait_list)
  81                         break;
  82
  83                 waiter = list_entry(waiter->list.next,struct rwsem_waiter,list);
  84
  85         } while (waiter->flags & RWSEM_WAITING_FOR_READ);
  86
  87         loop = woken;
  88         woken *= RWSEM_ACTIVE_BIAS-RWSEM_WAITING_BIAS;
  89         woken -= RWSEM_ACTIVE_BIAS;
  90         rwsem_atomic_add(woken,sem);
  91
  92         next = sem->wait_list.next;
  93         for (; loop>0; loop--) {
  94                 waiter = list_entry(next,struct rwsem_waiter,list);
  95                 next = waiter->list.next;
  96                 tsk = waiter->task;
  97                 mb();
  98                 waiter->task = NULL;
  99                 wake_up_process(tsk);
 100                 free_task_struct(tsk);
 101         }
 102
 103         sem->wait_list.next = next;
 104         next->prev = &sem->wait_list;
 105
 106  out:
 107         rwsemtrace(sem,"Leaving __rwsem_do_wake");
 108         return sem;
 109
 110         /* undo the change to count, but check for a transition 1->0 */
 111  undo:
 112         if (rwsem_atomic_update(-RWSEM_ACTIVE_BIAS,sem)!=0)
 113                 goto out;
 114         goto try_again;
 115 }
 116
 117 /*
 118  * wait for a lock to be granted
 119  */
 120 static inline struct rw_semaphore *rwsem_down_failed_common(struct rw_semaphore *sem,
 121                                                                  struct rwsem_waiter *waiter,
 122                                                                  signed long adjustment)
 123 {
 124         struct task_struct *tsk = current;
 125         signed long count;
 126
 127         set_task_state(tsk,TASK_UNINTERRUPTIBLE);
 128
 129         /* set up my own style of waitqueue */
 130         spin_lock_irq(&sem->wait_lock);
 131         waiter->task = tsk;
 132         get_task_struct(tsk);
 133
 134         list_add_tail(&waiter->list,&sem->wait_list);
 135
 136         /* note that we're now waiting on the lock, but no longer actively read-locking */
 137         count = rwsem_atomic_update(adjustment,sem);
 138
 139         /* if there are no longer active locks, wake the front queued process(es) up
 140          * - it might even be this process, since the waker takes a more active part
 141          */
 142         if (!(count & RWSEM_ACTIVE_MASK))
 143                 sem = __rwsem_do_wake(sem);
 144
 145         spin_unlock_irq(&sem->wait_lock);
 146
 147         /* wait to be given the lock */
 148         for (;;) {
 149                 if (!waiter->task)
 150                         break;
 151                 schedule();
 152                 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 153         }
 154
 155         tsk->state = TASK_RUNNING;
 156
 157         return sem;
 158 }
 159
 160 /*
 161  * wait for the read lock to be granted
 162  */
 163 struct rw_semaphore fastcall *rwsem_down_read_failed(struct rw_semaphore *sem)
 164 {
 165         struct rwsem_waiter waiter;
 166
 167         rwsemtrace(sem,"Entering rwsem_down_read_failed");
 168
 169         waiter.flags = RWSEM_WAITING_FOR_READ;
 170         rwsem_down_failed_common(sem,&waiter,RWSEM_WAITING_BIAS-RWSEM_ACTIVE_BIAS);
 171
 172         rwsemtrace(sem,"Leaving rwsem_down_read_failed");
 173         return sem;
 174 }
 175
 176 /*
 177  * wait for the write lock to be granted
 178  */
 179 struct rw_semaphore fastcall *rwsem_down_write_failed(struct rw_semaphore *sem)
 180 {
 181         struct rwsem_waiter waiter;
 182
 183         rwsemtrace(sem,"Entering rwsem_down_write_failed");
 184
 185         waiter.flags = RWSEM_WAITING_FOR_WRITE;
 186         rwsem_down_failed_common(sem,&waiter,-RWSEM_ACTIVE_BIAS);
 187
 188         rwsemtrace(sem,"Leaving rwsem_down_write_failed");
 189         return sem;
 190 }
 191
 192 /*
 193  * handle waking up a waiter on the semaphore
 194  * - up_read has decremented the active part of the count if we come here
 195  */
 196 struct rw_semaphore fastcall *rwsem_wake(struct rw_semaphore *sem)
 197 {
 198         unsigned long flags;
 199         rwsemtrace(sem,"Entering rwsem_wake");
 200
 201         spin_lock_irqsave(&sem->wait_lock, flags);
 202
 203         /* do nothing if list empty */
 204         if (!list_empty(&sem->wait_list))
 205                 sem = __rwsem_do_wake(sem);
 206
 207         spin_unlock_irqrestore(&sem->wait_lock, flags);
 208
 209         rwsemtrace(sem,"Leaving rwsem_wake");
 210
 211         return sem;
 212 }
 213
 214 EXPORT_SYMBOL_NOVERS(rwsem_down_read_failed);
 215 EXPORT_SYMBOL_NOVERS(rwsem_down_write_failed);
 216 EXPORT_SYMBOL_NOVERS(rwsem_wake);
 217 #if RWSEM_DEBUG
 218 EXPORT_SYMBOL(rwsemtrace);
 219 #endif