Merge "Merge remote-tracking branch 'origin/dev/dev-kernel-3.18-lz4-zram' into dev...
[linux-3.10.git] / lib / rwsem.c
index 8337e1b..19c5fa9 100644 (file)
@@ -2,6 +2,9 @@
  *
  * Written by David Howells (dhowells@redhat.com).
  * Derived from arch/i386/kernel/semaphore.c
+ *
+ * Writer lock-stealing by Alex Shi <alex.shi@intel.com>
+ * and Michel Lespinasse <walken@google.com>
  */
 #include <linux/rwsem.h>
 #include <linux/sched.h>
@@ -28,21 +31,22 @@ void __init_rwsem(struct rw_semaphore *sem, const char *name,
 
 EXPORT_SYMBOL(__init_rwsem);
 
+enum rwsem_waiter_type {
+       RWSEM_WAITING_FOR_WRITE,
+       RWSEM_WAITING_FOR_READ
+};
+
 struct rwsem_waiter {
        struct list_head list;
        struct task_struct *task;
-       unsigned int flags;
-#define RWSEM_WAITING_FOR_READ 0x00000001
-#define RWSEM_WAITING_FOR_WRITE        0x00000002
+       enum rwsem_waiter_type type;
 };
 
-/* Wake types for __rwsem_do_wake().  Note that RWSEM_WAKE_NO_ACTIVE and
- * RWSEM_WAKE_READ_OWNED imply that the spinlock must have been kept held
- * since the rwsem value was observed.
- */
-#define RWSEM_WAKE_ANY        0 /* Wake whatever's at head of wait list */
-#define RWSEM_WAKE_NO_ACTIVE  1 /* rwsem was observed with no active thread */
-#define RWSEM_WAKE_READ_OWNED 2 /* rwsem was observed to be read owned */
+enum rwsem_wake_type {
+       RWSEM_WAKE_ANY,         /* Wake whatever's at head of wait list */
+       RWSEM_WAKE_READERS,     /* Wake readers only */
+       RWSEM_WAKE_READ_OWNED   /* Waker thread holds the read lock */
+};
 
 /*
  * handle the lock release when processes blocked on it that can now run
@@ -55,68 +59,43 @@ struct rwsem_waiter {
  * - writers are only woken if downgrading is false
  */
 static struct rw_semaphore *
-__rwsem_do_wake(struct rw_semaphore *sem, int wake_type)
+__rwsem_do_wake(struct rw_semaphore *sem, enum rwsem_wake_type wake_type)
 {
        struct rwsem_waiter *waiter;
        struct task_struct *tsk;
        struct list_head *next;
-       signed long oldcount, woken, loop, adjustment;
+       long oldcount, woken, loop, adjustment;
 
        waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
-       if (!(waiter->flags & RWSEM_WAITING_FOR_WRITE))
-               goto readers_only;
-
-       if (wake_type == RWSEM_WAKE_READ_OWNED)
-               /* Another active reader was observed, so wakeup is not
-                * likely to succeed. Save the atomic op.
-                */
+       if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
+               if (wake_type == RWSEM_WAKE_ANY)
+                       /* Wake writer at the front of the queue, but do not
+                        * grant it the lock yet as we want other writers
+                        * to be able to steal it.  Readers, on the other hand,
+                        * will block as they will notice the queued writer.
+                        */
+                       wake_up_process(waiter->task);
                goto out;
+       }
 
-       /* There's a writer at the front of the queue - try to grant it the
-        * write lock.  However, we only wake this writer if we can transition
-        * the active part of the count from 0 -> 1
-        */
-       adjustment = RWSEM_ACTIVE_WRITE_BIAS;
-       if (waiter->list.next == &sem->wait_list)
-               adjustment -= RWSEM_WAITING_BIAS;
-
- try_again_write:
-       oldcount = rwsem_atomic_update(adjustment, sem) - adjustment;
-       if (oldcount & RWSEM_ACTIVE_MASK)
-               /* Someone grabbed the sem already */
-               goto undo_write;
-
-       /* We must be careful not to touch 'waiter' after we set ->task = NULL.
-        * It is an allocated on the waiter's stack and may become invalid at
-        * any time after that point (due to a wakeup from another source).
+       /* Writers might steal the lock before we grant it to the next reader.
+        * We prefer to do the first reader grant before counting readers
+        * so we can bail out early if a writer stole the lock.
         */
-       list_del(&waiter->list);
-       tsk = waiter->task;
-       smp_mb();
-       waiter->task = NULL;
-       wake_up_process(tsk);
-       put_task_struct(tsk);
-       goto out;
-
- readers_only:
-       /* If we come here from up_xxxx(), another thread might have reached
-        * rwsem_down_failed_common() before we acquired the spinlock and
-        * woken up a waiter, making it now active.  We prefer to check for
-        * this first in order to not spend too much time with the spinlock
-        * held if we're not going to be able to wake up readers in the end.
-        *
-        * Note that we do not need to update the rwsem count: any writer
-        * trying to acquire rwsem will run rwsem_down_write_failed() due
-        * to the waiting threads and block trying to acquire the spinlock.
-        *
-        * We use a dummy atomic update in order to acquire the cache line
-        * exclusively since we expect to succeed and run the final rwsem
-        * count adjustment pretty soon.
-        */
-       if (wake_type == RWSEM_WAKE_ANY &&
-           rwsem_atomic_update(0, sem) < RWSEM_WAITING_BIAS)
-               /* Someone grabbed the sem for write already */
-               goto out;
+       adjustment = 0;
+       if (wake_type != RWSEM_WAKE_READ_OWNED) {
+               adjustment = RWSEM_ACTIVE_READ_BIAS;
+ try_reader_grant:
+               oldcount = rwsem_atomic_update(adjustment, sem) - adjustment;
+               if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
+                       /* A writer stole the lock. Undo our reader grant. */
+                       if (rwsem_atomic_update(-adjustment, sem) &
+                                               RWSEM_ACTIVE_MASK)
+                               goto out;
+                       /* Last active locker left. Retry waking readers. */
+                       goto try_reader_grant;
+               }
+       }
 
        /* Grant an infinite number of read locks to the readers at the front
         * of the queue.  Note we increment the 'active part' of the count by
@@ -132,17 +111,19 @@ __rwsem_do_wake(struct rw_semaphore *sem, int wake_type)
                waiter = list_entry(waiter->list.next,
                                        struct rwsem_waiter, list);
 
-       } while (waiter->flags & RWSEM_WAITING_FOR_READ);
+       } while (waiter->type != RWSEM_WAITING_FOR_WRITE);
 
-       adjustment = woken * RWSEM_ACTIVE_READ_BIAS;
-       if (waiter->flags & RWSEM_WAITING_FOR_READ)
+       adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
+       if (waiter->type != RWSEM_WAITING_FOR_WRITE)
                /* hit end of list above */
                adjustment -= RWSEM_WAITING_BIAS;
 
-       rwsem_atomic_add(adjustment, sem);
+       if (adjustment)
+               rwsem_atomic_add(adjustment, sem);
 
        next = sem->wait_list.next;
-       for (loop = woken; loop > 0; loop--) {
+       loop = woken;
+       do {
                waiter = list_entry(next, struct rwsem_waiter, list);
                next = waiter->list.next;
                tsk = waiter->task;
@@ -150,41 +131,30 @@ __rwsem_do_wake(struct rw_semaphore *sem, int wake_type)
                waiter->task = NULL;
                wake_up_process(tsk);
                put_task_struct(tsk);
-       }
+       } while (--loop);
 
        sem->wait_list.next = next;
        next->prev = &sem->wait_list;
 
  out:
        return sem;
-
-       /* undo the change to the active count, but check for a transition
-        * 1->0 */
- undo_write:
-       if (rwsem_atomic_update(-adjustment, sem) & RWSEM_ACTIVE_MASK)
-               goto out;
-       goto try_again_write;
 }
 
 /*
- * wait for a lock to be granted
+ * wait for the read lock to be granted
  */
-static struct rw_semaphore __sched *
-rwsem_down_failed_common(struct rw_semaphore *sem,
-                        unsigned int flags, signed long adjustment)
+struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
 {
+       long count, adjustment = -RWSEM_ACTIVE_READ_BIAS;
        struct rwsem_waiter waiter;
        struct task_struct *tsk = current;
-       signed long count;
-
-       set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 
        /* set up my own style of waitqueue */
-       raw_spin_lock_irq(&sem->wait_lock);
        waiter.task = tsk;
-       waiter.flags = flags;
+       waiter.type = RWSEM_WAITING_FOR_READ;
        get_task_struct(tsk);
 
+       raw_spin_lock_irq(&sem->wait_lock);
        if (list_empty(&sem->wait_list))
                adjustment += RWSEM_WAITING_BIAS;
        list_add_tail(&waiter.list, &sem->wait_list);
@@ -192,26 +162,24 @@ rwsem_down_failed_common(struct rw_semaphore *sem,
        /* we're now waiting on the lock, but no longer actively locking */
        count = rwsem_atomic_update(adjustment, sem);
 
-       /* If there are no active locks, wake the front queued process(es) up.
+       /* If there are no active locks, wake the front queued process(es).
         *
-        * Alternatively, if we're called from a failed down_write(), there
-        * were already threads queued before us and there are no active
-        * writers, the lock must be read owned; so we try to wake any read
-        * locks that were queued ahead of us. */
-       if (count == RWSEM_WAITING_BIAS)
-               sem = __rwsem_do_wake(sem, RWSEM_WAKE_NO_ACTIVE);
-       else if (count > RWSEM_WAITING_BIAS &&
-                adjustment == -RWSEM_ACTIVE_WRITE_BIAS)
-               sem = __rwsem_do_wake(sem, RWSEM_WAKE_READ_OWNED);
+        * If there are no writers and we are first in the queue,
+        * wake our own waiter to join the existing active readers !
+        */
+       if (count == RWSEM_WAITING_BIAS ||
+           (count > RWSEM_WAITING_BIAS &&
+            adjustment != -RWSEM_ACTIVE_READ_BIAS))
+               sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
 
        raw_spin_unlock_irq(&sem->wait_lock);
 
        /* wait to be given the lock */
-       for (;;) {
+       while (true) {
+               set_task_state(tsk, TASK_UNINTERRUPTIBLE);
                if (!waiter.task)
                        break;
                schedule();
-               set_task_state(tsk, TASK_UNINTERRUPTIBLE);
        }
 
        tsk->state = TASK_RUNNING;
@@ -220,21 +188,64 @@ rwsem_down_failed_common(struct rw_semaphore *sem,
 }
 
 /*
- * wait for the read lock to be granted
- */
-struct rw_semaphore __sched *rwsem_down_read_failed(struct rw_semaphore *sem)
-{
-       return rwsem_down_failed_common(sem, RWSEM_WAITING_FOR_READ,
-                                       -RWSEM_ACTIVE_READ_BIAS);
-}
-
-/*
- * wait for the write lock to be granted
+ * wait until we successfully acquire the write lock
  */
 struct rw_semaphore __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
 {
-       return rwsem_down_failed_common(sem, RWSEM_WAITING_FOR_WRITE,
-                                       -RWSEM_ACTIVE_WRITE_BIAS);
+       long count, adjustment = -RWSEM_ACTIVE_WRITE_BIAS;
+       struct rwsem_waiter waiter;
+       struct task_struct *tsk = current;
+
+       /* set up my own style of waitqueue */
+       waiter.task = tsk;
+       waiter.type = RWSEM_WAITING_FOR_WRITE;
+
+       raw_spin_lock_irq(&sem->wait_lock);
+       if (list_empty(&sem->wait_list))
+               adjustment += RWSEM_WAITING_BIAS;
+       list_add_tail(&waiter.list, &sem->wait_list);
+
+       /* we're now waiting on the lock, but no longer actively locking */
+       count = rwsem_atomic_update(adjustment, sem);
+
+       /* If there were already threads queued before us and there are no
+        * active writers, the lock must be read owned; so we try to wake
+        * any read locks that were queued ahead of us. */
+       if (count > RWSEM_WAITING_BIAS &&
+           adjustment == -RWSEM_ACTIVE_WRITE_BIAS)
+               sem = __rwsem_do_wake(sem, RWSEM_WAKE_READERS);
+
+       /* wait until we successfully acquire the lock */
+       set_task_state(tsk, TASK_UNINTERRUPTIBLE);
+       while (true) {
+               if (!(count & RWSEM_ACTIVE_MASK)) {
+                       /* Try acquiring the write lock. */
+                       count = RWSEM_ACTIVE_WRITE_BIAS;
+                       if (!list_is_singular(&sem->wait_list))
+                               count += RWSEM_WAITING_BIAS;
+
+                       if (sem->count == RWSEM_WAITING_BIAS &&
+                           cmpxchg(&sem->count, RWSEM_WAITING_BIAS, count) ==
+                                                       RWSEM_WAITING_BIAS)
+                               break;
+               }
+
+               raw_spin_unlock_irq(&sem->wait_lock);
+
+               /* Block until there are no active lockers. */
+               do {
+                       schedule();
+                       set_task_state(tsk, TASK_UNINTERRUPTIBLE);
+               } while ((count = sem->count) & RWSEM_ACTIVE_MASK);
+
+               raw_spin_lock_irq(&sem->wait_lock);
+       }
+
+       list_del(&waiter.list);
+       raw_spin_unlock_irq(&sem->wait_lock);
+       tsk->state = TASK_RUNNING;
+
+       return sem;
 }
 
 /*