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/init.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  * - writers are only woken if wakewrite is non-zero
  39  */
  40 static inline struct rw_semaphore *__rwsem_do_wake(struct rw_semaphore *sem, int wakewrite)
  41 {
  42         struct rwsem_waiter *waiter;
  43         struct list_head *next;
  44         signed long oldcount;
  45         int woken, loop;
  46
  47         rwsemtrace(sem,"Entering __rwsem_do_wake");
  48
  49         if (!wakewrite)
  50                 goto dont_wake_writers;
  51
  52         /* only wake someone up if we can transition the active part of the count from 0 -> 1 */
  53  try_again:
  54         oldcount = rwsem_atomic_update(RWSEM_ACTIVE_BIAS,sem) - RWSEM_ACTIVE_BIAS;
  55         if (oldcount & RWSEM_ACTIVE_MASK)
  56                 goto undo;
  57
  58         waiter = list_entry(sem->wait_list.next,struct rwsem_waiter,list);
  59
  60         /* try to grant a single write lock if there's a writer at the front of the queue
  61          * - note we leave the 'active part' of the count incremented by 1 and the waiting part
  62          *   incremented by 0x00010000
  63          */
  64         if (!(waiter->flags & RWSEM_WAITING_FOR_WRITE))
  65                 goto readers_only;
  66
  67         list_del(&waiter->list);
  68         waiter->flags = 0;
  69         wake_up_process(waiter->task);
  70         goto out;
  71
  72         /* don't want to wake any writers */
  73  dont_wake_writers:
  74         waiter = list_entry(sem->wait_list.next,struct rwsem_waiter,list);
  75         if (waiter->flags & RWSEM_WAITING_FOR_WRITE)
  76                 goto out;
  77
  78         /* grant an infinite number of read locks to the readers at the front of the queue
  79          * - note we increment the 'active part' of the count by the number of readers (less one
  80          *   for the activity decrement we've already done) before waking any processes up
  81          */
  82  readers_only:
  83         woken = 0;
  84         do {
  85                 woken++;
  86
  87                 if (waiter->list.next==&sem->wait_list)
  88                         break;
  89
  90                 waiter = list_entry(waiter->list.next,struct rwsem_waiter,list);
  91
  92         } while (waiter->flags & RWSEM_WAITING_FOR_READ);
  93
  94         loop = woken;
  95         woken *= RWSEM_ACTIVE_BIAS-RWSEM_WAITING_BIAS;
  96         woken -= RWSEM_ACTIVE_BIAS;
  97         rwsem_atomic_add(woken,sem);
  98
  99         next = sem->wait_list.next;
 100         for (; loop>0; loop--) {
 101                 waiter = list_entry(next,struct rwsem_waiter,list);
 102                 next = waiter->list.next;
 103                 waiter->flags = 0;
 104                 wake_up_process(waiter->task);
 105         }
 106
 107         sem->wait_list.next = next;
 108         next->prev = &sem->wait_list;
 109
 110  out:
 111         rwsemtrace(sem,"Leaving __rwsem_do_wake");
 112         return sem;
 113
 114         /* undo the change to count, but check for a transition 1->0 */
 115  undo:
 116         if (rwsem_atomic_update(-RWSEM_ACTIVE_BIAS,sem)!=0)
 117                 goto out;
 118         goto try_again;
 119 }
 120
 121 /*
 122  * wait for a lock to be granted
 123  */
 124 static inline struct rw_semaphore *rwsem_down_failed_common(struct rw_semaphore *sem,
 125                                                                  struct rwsem_waiter *waiter,
 126                                                                  signed long adjustment)
 127 {
 128         struct task_struct *tsk = current;
 129         signed long count;
 130
 131         set_task_state(tsk,TASK_UNINTERRUPTIBLE);
 132
 133         /* set up my own style of waitqueue */
 134         spin_lock(&sem->wait_lock);
 135         waiter->task = tsk;
 136
 137         list_add_tail(&waiter->list,&sem->wait_list);
 138
 139         /* note that we're now waiting on the lock, but no longer actively read-locking */
 140         count = rwsem_atomic_update(adjustment,sem);
 141
 142         /* if there are no longer active locks, wake the front queued process(es) up
 143          * - it might even be this process, since the waker takes a more active part
 144          */
 145         if (!(count & RWSEM_ACTIVE_MASK))
 146                 sem = __rwsem_do_wake(sem,1);
 147
 148         spin_unlock(&sem->wait_lock);
 149
 150         /* wait to be given the lock */
 151         for (;;) {
 152                 if (!waiter->flags)
 153                         break;
 154                 schedule();
 155                 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 156         }
 157
 158         tsk->state = TASK_RUNNING;
 159
 160         return sem;
 161 }
 162
 163 /*
 164  * wait for the read lock to be granted
 165  */
 166 struct rw_semaphore fastcall __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
 167 {
 168         struct rwsem_waiter waiter;
 169
 170         rwsemtrace(sem,"Entering rwsem_down_read_failed");
 171
 172         waiter.flags = RWSEM_WAITING_FOR_READ;
 173         rwsem_down_failed_common(sem,&waiter,RWSEM_WAITING_BIAS-RWSEM_ACTIVE_BIAS);
 174
 175         rwsemtrace(sem,"Leaving rwsem_down_read_failed");
 176         return sem;
 177 }
 178
 179 /*
 180  * wait for the write lock to be granted
 181  */
 182 struct rw_semaphore fastcall __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
 183 {
 184         struct rwsem_waiter waiter;
 185
 186         rwsemtrace(sem,"Entering rwsem_down_write_failed");
 187
 188         waiter.flags = RWSEM_WAITING_FOR_WRITE;
 189         rwsem_down_failed_common(sem,&waiter,-RWSEM_ACTIVE_BIAS);
 190
 191         rwsemtrace(sem,"Leaving rwsem_down_write_failed");
 192         return sem;
 193 }
 194
 195 /*
 196  * handle waking up a waiter on the semaphore
 197  * - up_read has decremented the active part of the count if we come here
 198  */
 199 struct rw_semaphore fastcall *rwsem_wake(struct rw_semaphore *sem)
 200 {
 201         rwsemtrace(sem,"Entering rwsem_wake");
 202
 203         spin_lock(&sem->wait_lock);
 204
 205         /* do nothing if list empty */
 206         if (!list_empty(&sem->wait_list))
 207                 sem = __rwsem_do_wake(sem,1);
 208
 209         spin_unlock(&sem->wait_lock);
 210
 211         rwsemtrace(sem,"Leaving rwsem_wake");
 212
 213         return sem;
 214 }
 215
 216 /*
 217  * downgrade a write lock into a read lock
 218  * - caller incremented waiting part of count, and discovered it to be still negative
 219  * - just wake up any readers at the front of the queue
 220  */
 221 struct rw_semaphore fastcall *rwsem_downgrade_wake(struct rw_semaphore *sem)
 222 {
 223         rwsemtrace(sem,"Entering rwsem_downgrade_wake");
 224
 225         spin_lock(&sem->wait_lock);
 226
 227         /* do nothing if list empty */
 228         if (!list_empty(&sem->wait_list))
 229                 sem = __rwsem_do_wake(sem,0);
 230
 231         spin_unlock(&sem->wait_lock);
 232
 233         rwsemtrace(sem,"Leaving rwsem_downgrade_wake");
 234         return sem;
 235 }
 236
 237 EXPORT_SYMBOL_NOVERS(rwsem_down_read_failed);
 238 EXPORT_SYMBOL_NOVERS(rwsem_down_write_failed);
 239 EXPORT_SYMBOL_NOVERS(rwsem_wake);
 240 EXPORT_SYMBOL_NOVERS(rwsem_downgrade_wake);
 241 #if RWSEM_DEBUG
 242 EXPORT_SYMBOL(rwsemtrace);
 243 #endif