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