Merge "Merge remote-tracking branch 'origin/dev/dev-kernel-3.18-lz4-zram' into dev...
[linux-3.10.git] / lib / rwsem.c
index 901d0e7..19c5fa9 100644 (file)
@@ -2,11 +2,14 @@
  *
  * 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>
 #include <linux/init.h>
-#include <linux/module.h>
+#include <linux/export.h>
 
 /*
  * Initialize an rwsem:
@@ -19,21 +22,30 @@ void __init_rwsem(struct rw_semaphore *sem, const char *name,
         * Make sure we are not reinitializing a held semaphore:
         */
        debug_check_no_locks_freed((void *)sem, sizeof(*sem));
-       lockdep_init_map(&sem->dep_map, name, key);
+       lockdep_init_map(&sem->dep_map, name, key, 0);
 #endif
        sem->count = RWSEM_UNLOCKED_VALUE;
-       spin_lock_init(&sem->wait_lock);
+       raw_spin_lock_init(&sem->wait_lock);
        INIT_LIST_HEAD(&sem->wait_list);
 }
 
 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;
+};
+
+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 */
 };
 
 /*
@@ -41,64 +53,54 @@ struct rwsem_waiter {
  * - if we come here from up_xxxx(), then:
  *   - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed)
  *   - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so)
- *   - there must be someone on the queue
+ * - there must be someone on the queue
  * - the spinlock must be held by the caller
  * - woken process blocks are discarded from the list after having task zeroed
  * - writers are only woken if downgrading is false
  */
-static inline struct rw_semaphore *
-__rwsem_do_wake(struct rw_semaphore *sem, int downgrading)
+static struct rw_semaphore *
+__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;
-
-       if (downgrading)
-               goto dont_wake_writers;
-
-       /* if we came through an up_xxxx() call, we only only wake someone up
-        * if we can transition the active part of the count from 0 -> 1
-        */
- try_again:
-       oldcount = rwsem_atomic_update(RWSEM_ACTIVE_BIAS, sem)
-                                               - RWSEM_ACTIVE_BIAS;
-       if (oldcount & RWSEM_ACTIVE_MASK)
-               goto undo;
+       long oldcount, woken, loop, adjustment;
 
        waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
+       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;
+       }
 
-       /* try to grant a single write lock if there's a writer at the front
-        * of the queue - note we leave the 'active part' of the count
-        * incremented by 1 and the waiting part incremented by 0x00010000
-        */
-       if (!(waiter->flags & RWSEM_WAITING_FOR_WRITE))
-               goto readers_only;
-
-       /* 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;
-
-       /* don't want to wake any writers */
- dont_wake_writers:
-       waiter = list_entry(sem->wait_list.next, struct rwsem_waiter, list);
-       if (waiter->flags & RWSEM_WAITING_FOR_WRITE)
-               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 the number of
-        *   readers before waking any processes up
+       /* 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
+        * the number of readers before waking any processes up.
         */
- readers_only:
        woken = 0;
        do {
                woken++;
@@ -109,18 +111,19 @@ __rwsem_do_wake(struct rw_semaphore *sem, int downgrading)
                waiter = list_entry(waiter->list.next,
                                        struct rwsem_waiter, list);
 
-       } while (waiter->flags & RWSEM_WAITING_FOR_READ);
+       } while (waiter->type != RWSEM_WAITING_FOR_WRITE);
 
-       loop = woken;
-       woken *= RWSEM_ACTIVE_BIAS - RWSEM_WAITING_BIAS;
-       if (!downgrading)
-               /* we'd already done one increment earlier */
-               woken -= RWSEM_ACTIVE_BIAS;
+       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(woken, sem);
+       if (adjustment)
+               rwsem_atomic_add(adjustment, sem);
 
        next = sem->wait_list.next;
-       for (; loop > 0; loop--) {
+       loop = woken;
+       do {
                waiter = list_entry(next, struct rwsem_waiter, list);
                next = waiter->list.next;
                tsk = waiter->task;
@@ -128,55 +131,55 @@ __rwsem_do_wake(struct rw_semaphore *sem, int downgrading)
                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 count, but check for a transition 1->0 */
- undo:
-       if (rwsem_atomic_update(-RWSEM_ACTIVE_BIAS, sem) != 0)
-               goto out;
-       goto try_again;
 }
 
 /*
- * wait for a lock to be granted
+ * wait for the read lock to be granted
  */
-static struct rw_semaphore *
-rwsem_down_failed_common(struct rw_semaphore *sem,
-                       struct rwsem_waiter *waiter, 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 */
-       spin_lock_irq(&sem->wait_lock);
-       waiter->task = tsk;
+       waiter.task = tsk;
+       waiter.type = RWSEM_WAITING_FOR_READ;
        get_task_struct(tsk);
 
-       list_add_tail(&waiter->list, &sem->wait_list);
+       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 read-locking */
+       /* 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 (!(count & RWSEM_ACTIVE_MASK))
-               sem = __rwsem_do_wake(sem, 0);
+       /* If there are no active locks, wake the front queued process(es).
+        *
+        * 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);
 
-       spin_unlock_irq(&sem->wait_lock);
+       raw_spin_unlock_irq(&sem->wait_lock);
 
        /* wait to be given the lock */
-       for (;;) {
-               if (!waiter->task)
+       while (true) {
+               set_task_state(tsk, TASK_UNINTERRUPTIBLE);
+               if (!waiter.task)
                        break;
                schedule();
-               set_task_state(tsk, TASK_UNINTERRUPTIBLE);
        }
 
        tsk->state = TASK_RUNNING;
@@ -185,29 +188,62 @@ rwsem_down_failed_common(struct rw_semaphore *sem,
 }
 
 /*
- * wait for the read lock to be granted
+ * wait until we successfully acquire the write lock
  */
-struct rw_semaphore fastcall __sched *
-rwsem_down_read_failed(struct rw_semaphore *sem)
+struct rw_semaphore __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
 {
+       long count, adjustment = -RWSEM_ACTIVE_WRITE_BIAS;
        struct rwsem_waiter waiter;
+       struct task_struct *tsk = current;
 
-       waiter.flags = RWSEM_WAITING_FOR_READ;
-       rwsem_down_failed_common(sem, &waiter,
-                               RWSEM_WAITING_BIAS - RWSEM_ACTIVE_BIAS);
-       return sem;
-}
+       /* set up my own style of waitqueue */
+       waiter.task = tsk;
+       waiter.type = RWSEM_WAITING_FOR_WRITE;
 
-/*
- * wait for the write lock to be granted
- */
-struct rw_semaphore fastcall __sched *
-rwsem_down_write_failed(struct rw_semaphore *sem)
-{
-       struct rwsem_waiter waiter;
+       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);
 
-       waiter.flags = RWSEM_WAITING_FOR_WRITE;
-       rwsem_down_failed_common(sem, &waiter, -RWSEM_ACTIVE_BIAS);
+       /* 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;
 }
@@ -216,17 +252,17 @@ rwsem_down_write_failed(struct rw_semaphore *sem)
  * handle waking up a waiter on the semaphore
  * - up_read/up_write has decremented the active part of count if we come here
  */
-struct rw_semaphore fastcall *rwsem_wake(struct rw_semaphore *sem)
+struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
 {
        unsigned long flags;
 
-       spin_lock_irqsave(&sem->wait_lock, flags);
+       raw_spin_lock_irqsave(&sem->wait_lock, flags);
 
        /* do nothing if list empty */
        if (!list_empty(&sem->wait_list))
-               sem = __rwsem_do_wake(sem, 0);
+               sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
 
-       spin_unlock_irqrestore(&sem->wait_lock, flags);
+       raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
 
        return sem;
 }
@@ -236,17 +272,17 @@ struct rw_semaphore fastcall *rwsem_wake(struct rw_semaphore *sem)
  * - caller incremented waiting part of count and discovered it still negative
  * - just wake up any readers at the front of the queue
  */
-struct rw_semaphore fastcall *rwsem_downgrade_wake(struct rw_semaphore *sem)
+struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
 {
        unsigned long flags;
 
-       spin_lock_irqsave(&sem->wait_lock, flags);
+       raw_spin_lock_irqsave(&sem->wait_lock, flags);
 
        /* do nothing if list empty */
        if (!list_empty(&sem->wait_list))
-               sem = __rwsem_do_wake(sem, 1);
+               sem = __rwsem_do_wake(sem, RWSEM_WAKE_READ_OWNED);
 
-       spin_unlock_irqrestore(&sem->wait_lock, flags);
+       raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
 
        return sem;
 }