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 from up_xxxx(), then:
  33  *   - the 'active part' of the count (&0x0000ffff) had reached zero (but may have changed)
  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 task zeroed
  38  * - writers are only woken if downgrading is false
  39  */
  40 static inline struct rw_semaphore *__rwsem_do_wake(struct rw_semaphore *sem, int downgrading)
  41 {
  42         struct rwsem_waiter *waiter;
  43         struct task_struct *tsk;
  44         struct list_head *next;
  45         signed long oldcount, woken, loop;
  46
  47         rwsemtrace(sem,"Entering __rwsem_do_wake");
  48
  49         if (downgrading)
  50                 goto dont_wake_writers;
  51
  52         /* if we came through an up_xxxx() call, we only only wake someone up
  53          * if we can transition the active part of the count from 0 -> 1
  54          */
  55  try_again:
  56         oldcount = rwsem_atomic_update(RWSEM_ACTIVE_BIAS,sem) - RWSEM_ACTIVE_BIAS;
  57         if (oldcount & RWSEM_ACTIVE_MASK)
  58                 goto undo;
  59
  60         waiter = list_entry(sem->wait_list.next,struct rwsem_waiter,list);
  61
  62         /* try to grant a single write lock if there's a writer at the front of the queue
  63          * - note we leave the 'active part' of the count incremented by 1 and the waiting part
  64          *   incremented by 0x00010000
  65          */
  66         if (!(waiter->flags & RWSEM_WAITING_FOR_WRITE))
  67                 goto readers_only;
  68
  69         list_del(&waiter->list);
  70         tsk = waiter->task;
  71         mb();
  72         waiter->task = NULL;
  73         wake_up_process(tsk);
  74         put_task_struct(tsk);
  75         goto out;
  76
  77         /* don't want to wake any writers */
  78  dont_wake_writers:
  79         waiter = list_entry(sem->wait_list.next,struct rwsem_waiter,list);
  80         if (waiter->flags & RWSEM_WAITING_FOR_WRITE)
  81                 goto out;
  82
  83         /* grant an infinite number of read locks to the readers at the front
  84          * of the queue
  85          * - note we increment the 'active part' of the count by the number of
  86          *   readers before waking any processes up
  87          */
  88  readers_only:
  89         woken = 0;
  90         do {
  91                 woken++;
  92
  93                 if (waiter->list.next==&sem->wait_list)
  94                         break;
  95
  96                 waiter = list_entry(waiter->list.next,struct rwsem_waiter,list);
  97
  98         } while (waiter->flags & RWSEM_WAITING_FOR_READ);
  99
 100         loop = woken;
 101         woken *= RWSEM_ACTIVE_BIAS - RWSEM_WAITING_BIAS;
 102         if (!downgrading)
 103                 woken -= RWSEM_ACTIVE_BIAS; /* we'd already done one increment
 104                                              * earlier */
 105         rwsem_atomic_add(woken,sem);
 106
 107         next = sem->wait_list.next;
 108         for (; loop>0; loop--) {
 109                 waiter = list_entry(next,struct rwsem_waiter,list);
 110                 next = waiter->list.next;
 111                 tsk = waiter->task;
 112                 mb();
 113                 waiter->task = NULL;
 114                 wake_up_process(tsk);
 115                 put_task_struct(tsk);
 116         }
 117
 118         sem->wait_list.next = next;
 119         next->prev = &sem->wait_list;
 120
 121  out:
 122         rwsemtrace(sem,"Leaving __rwsem_do_wake");
 123         return sem;
 124
 125         /* undo the change to count, but check for a transition 1->0 */
 126  undo:
 127         if (rwsem_atomic_update(-RWSEM_ACTIVE_BIAS,sem)!=0)
 128                 goto out;
 129         goto try_again;
 130 }
 131
 132 /*
 133  * wait for a lock to be granted
 134  */
 135 static inline struct rw_semaphore *rwsem_down_failed_common(struct rw_semaphore *sem,
 136                                                                  struct rwsem_waiter *waiter,
 137                                                                  signed long adjustment)
 138 {
 139         struct task_struct *tsk = current;
 140         signed long count;
 141
 142         set_task_state(tsk,TASK_UNINTERRUPTIBLE);
 143
 144         /* set up my own style of waitqueue */
 145         spin_lock(&sem->wait_lock);
 146         waiter->task = tsk;
 147         get_task_struct(tsk);
 148
 149         list_add_tail(&waiter->list,&sem->wait_list);
 150
 151         /* note that we're now waiting on the lock, but no longer actively read-locking */
 152         count = rwsem_atomic_update(adjustment,sem);
 153
 154         /* if there are no longer active locks, wake the front queued process(es) up
 155          * - it might even be this process, since the waker takes a more active part
 156          */
 157         if (!(count & RWSEM_ACTIVE_MASK))
 158                 sem = __rwsem_do_wake(sem, 0);
 159
 160         spin_unlock(&sem->wait_lock);
 161
 162         /* wait to be given the lock */
 163         for (;;) {
 164                 if (!waiter->task)
 165                         break;
 166                 schedule();
 167                 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 168         }
 169
 170         tsk->state = TASK_RUNNING;
 171
 172         return sem;
 173 }
 174
 175 /*
 176  * wait for the read lock to be granted
 177  */
 178 struct rw_semaphore fastcall __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
 179 {
 180         struct rwsem_waiter waiter;
 181
 182         rwsemtrace(sem,"Entering rwsem_down_read_failed");
 183
 184         waiter.flags = RWSEM_WAITING_FOR_READ;
 185         rwsem_down_failed_common(sem,&waiter,RWSEM_WAITING_BIAS-RWSEM_ACTIVE_BIAS);
 186
 187         rwsemtrace(sem,"Leaving rwsem_down_read_failed");
 188         return sem;
 189 }
 190
 191 /*
 192  * wait for the write lock to be granted
 193  */
 194 struct rw_semaphore fastcall __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
 195 {
 196         struct rwsem_waiter waiter;
 197
 198         rwsemtrace(sem,"Entering rwsem_down_write_failed");
 199
 200         waiter.flags = RWSEM_WAITING_FOR_WRITE;
 201         rwsem_down_failed_common(sem,&waiter,-RWSEM_ACTIVE_BIAS);
 202
 203         rwsemtrace(sem,"Leaving rwsem_down_write_failed");
 204         return sem;
 205 }
 206
 207 /*
 208  * handle waking up a waiter on the semaphore
 209  * - up_read/up_write has decremented the active part of the count if we come here
 210  */
 211 struct rw_semaphore fastcall *rwsem_wake(struct rw_semaphore *sem)
 212 {
 213         rwsemtrace(sem,"Entering rwsem_wake");
 214
 215         spin_lock(&sem->wait_lock);
 216
 217         /* do nothing if list empty */
 218         if (!list_empty(&sem->wait_list))
 219                 sem = __rwsem_do_wake(sem, 0);
 220
 221         spin_unlock(&sem->wait_lock);
 222
 223         rwsemtrace(sem,"Leaving rwsem_wake");
 224
 225         return sem;
 226 }
 227
 228 /*
 229  * downgrade a write lock into a read lock
 230  * - caller incremented waiting part of count, and discovered it to be still negative
 231  * - just wake up any readers at the front of the queue
 232  */
 233 struct rw_semaphore fastcall *rwsem_downgrade_wake(struct rw_semaphore *sem)
 234 {
 235         rwsemtrace(sem,"Entering rwsem_downgrade_wake");
 236
 237         spin_lock(&sem->wait_lock);
 238
 239         /* do nothing if list empty */
 240         if (!list_empty(&sem->wait_list))
 241                 sem = __rwsem_do_wake(sem, 1);
 242
 243         spin_unlock(&sem->wait_lock);
 244
 245         rwsemtrace(sem,"Leaving rwsem_downgrade_wake");
 246         return sem;
 247 }
 248
 249 EXPORT_SYMBOL_NOVERS(rwsem_down_read_failed);
 250 EXPORT_SYMBOL_NOVERS(rwsem_down_write_failed);
 251 EXPORT_SYMBOL_NOVERS(rwsem_wake);
 252 EXPORT_SYMBOL_NOVERS(rwsem_downgrade_wake);
 253 #if RWSEM_DEBUG
 254 EXPORT_SYMBOL(rwsemtrace);
 255 #endif