microblaze: uaccess: fix put_user and get_user macros
[linux-2.6.git] / kernel / futex.c
index 381125a..e7a35f1 100644 (file)
@@ -89,33 +89,36 @@ struct futex_pi_state {
        union futex_key key;
 };
 
-/*
- * We use this hashed waitqueue instead of a normal wait_queue_t, so
+/**
+ * struct futex_q - The hashed futex queue entry, one per waiting task
+ * @task:              the task waiting on the futex
+ * @lock_ptr:          the hash bucket lock
+ * @key:               the key the futex is hashed on
+ * @pi_state:          optional priority inheritance state
+ * @rt_waiter:         rt_waiter storage for use with requeue_pi
+ * @requeue_pi_key:    the requeue_pi target futex key
+ * @bitset:            bitset for the optional bitmasked wakeup
+ *
+ * We use this hashed waitqueue, instead of a normal wait_queue_t, so
  * we can wake only the relevant ones (hashed queues may be shared).
  *
  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
  * The order of wakup is always to make the first condition true, then
- * wake up q->waiter, then make the second condition true.
+ * the second.
+ *
+ * PI futexes are typically woken before they are removed from the hash list via
+ * the rt_mutex code. See unqueue_me_pi().
  */
 struct futex_q {
        struct plist_node list;
-       /* Waiter reference */
-       struct task_struct *task;
 
-       /* Which hash list lock to use: */
+       struct task_struct *task;
        spinlock_t *lock_ptr;
-
-       /* Key which the futex is hashed on: */
        union futex_key key;
-
-       /* Optional priority inheritance state: */
        struct futex_pi_state *pi_state;
-
-       /* rt_waiter storage for requeue_pi: */
        struct rt_mutex_waiter *rt_waiter;
-
-       /* Bitset for the optional bitmasked wakeup */
+       union futex_key *requeue_pi_key;
        u32 bitset;
 };
 
@@ -147,7 +150,8 @@ static struct futex_hash_bucket *hash_futex(union futex_key *key)
  */
 static inline int match_futex(union futex_key *key1, union futex_key *key2)
 {
-       return (key1->both.word == key2->both.word
+       return (key1 && key2
+               && key1->both.word == key2->both.word
                && key1->both.ptr == key2->both.ptr
                && key1->both.offset == key2->both.offset);
 }
@@ -195,11 +199,10 @@ static void drop_futex_key_refs(union futex_key *key)
 }
 
 /**
- * get_futex_key - Get parameters which are the keys for a futex.
- * @uaddr: virtual address of the futex
- * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
- * @key: address where result is stored.
- * @rw: mapping needs to be read/write (values: VERIFY_READ, VERIFY_WRITE)
+ * get_futex_key() - Get parameters which are the keys for a futex
+ * @uaddr:     virtual address of the futex
+ * @fshared:   0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
+ * @key:       address where result is stored.
  *
  * Returns a negative error code or 0
  * The key words are stored in *key on success.
@@ -211,7 +214,7 @@ static void drop_futex_key_refs(union futex_key *key)
  * lock_page() might sleep, the caller should not hold a spinlock.
  */
 static int
-get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
+get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
 {
        unsigned long address = (unsigned long)uaddr;
        struct mm_struct *mm = current->mm;
@@ -234,7 +237,7 @@ get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
         *        but access_ok() should be faster than find_vma()
         */
        if (!fshared) {
-               if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
+               if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
                        return -EFAULT;
                key->private.mm = mm;
                key->private.address = address;
@@ -243,10 +246,11 @@ get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
        }
 
 again:
-       err = get_user_pages_fast(address, 1, rw == VERIFY_WRITE, &page);
+       err = get_user_pages_fast(address, 1, 1, &page);
        if (err < 0)
                return err;
 
+       page = compound_head(page);
        lock_page(page);
        if (!page->mapping) {
                unlock_page(page);
@@ -285,9 +289,34 @@ void put_futex_key(int fshared, union futex_key *key)
 }
 
 /**
+ * fault_in_user_writeable() - Fault in user address and verify RW access
+ * @uaddr:     pointer to faulting user space address
+ *
+ * Slow path to fixup the fault we just took in the atomic write
+ * access to @uaddr.
+ *
+ * We have no generic implementation of a non destructive write to the
+ * user address. We know that we faulted in the atomic pagefault
+ * disabled section so we can as well avoid the #PF overhead by
+ * calling get_user_pages() right away.
+ */
+static int fault_in_user_writeable(u32 __user *uaddr)
+{
+       struct mm_struct *mm = current->mm;
+       int ret;
+
+       down_read(&mm->mmap_sem);
+       ret = get_user_pages(current, mm, (unsigned long)uaddr,
+                            1, 1, 0, NULL, NULL);
+       up_read(&mm->mmap_sem);
+
+       return ret < 0 ? ret : 0;
+}
+
+/**
  * futex_top_waiter() - Return the highest priority waiter on a futex
- * @hb:     the hash bucket the futex_q's reside in
- * @key:    the futex key (to distinguish it from other futex futex_q's)
+ * @hb:                the hash bucket the futex_q's reside in
+ * @key:       the futex key (to distinguish it from other futex futex_q's)
  *
  * Must be called with the hb lock held.
  */
@@ -372,9 +401,9 @@ static void free_pi_state(struct futex_pi_state *pi_state)
         * and has cleaned up the pi_state already
         */
        if (pi_state->owner) {
-               spin_lock_irq(&pi_state->owner->pi_lock);
+               raw_spin_lock_irq(&pi_state->owner->pi_lock);
                list_del_init(&pi_state->list);
-               spin_unlock_irq(&pi_state->owner->pi_lock);
+               raw_spin_unlock_irq(&pi_state->owner->pi_lock);
 
                rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
        }
@@ -439,18 +468,18 @@ void exit_pi_state_list(struct task_struct *curr)
         * pi_state_list anymore, but we have to be careful
         * versus waiters unqueueing themselves:
         */
-       spin_lock_irq(&curr->pi_lock);
+       raw_spin_lock_irq(&curr->pi_lock);
        while (!list_empty(head)) {
 
                next = head->next;
                pi_state = list_entry(next, struct futex_pi_state, list);
                key = pi_state->key;
                hb = hash_futex(&key);
-               spin_unlock_irq(&curr->pi_lock);
+               raw_spin_unlock_irq(&curr->pi_lock);
 
                spin_lock(&hb->lock);
 
-               spin_lock_irq(&curr->pi_lock);
+               raw_spin_lock_irq(&curr->pi_lock);
                /*
                 * We dropped the pi-lock, so re-check whether this
                 * task still owns the PI-state:
@@ -464,15 +493,15 @@ void exit_pi_state_list(struct task_struct *curr)
                WARN_ON(list_empty(&pi_state->list));
                list_del_init(&pi_state->list);
                pi_state->owner = NULL;
-               spin_unlock_irq(&curr->pi_lock);
+               raw_spin_unlock_irq(&curr->pi_lock);
 
                rt_mutex_unlock(&pi_state->pi_mutex);
 
                spin_unlock(&hb->lock);
 
-               spin_lock_irq(&curr->pi_lock);
+               raw_spin_lock_irq(&curr->pi_lock);
        }
-       spin_unlock_irq(&curr->pi_lock);
+       raw_spin_unlock_irq(&curr->pi_lock);
 }
 
 static int
@@ -501,8 +530,25 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
                                return -EINVAL;
 
                        WARN_ON(!atomic_read(&pi_state->refcount));
-                       WARN_ON(pid && pi_state->owner &&
-                               pi_state->owner->pid != pid);
+
+                       /*
+                        * When pi_state->owner is NULL then the owner died
+                        * and another waiter is on the fly. pi_state->owner
+                        * is fixed up by the task which acquires
+                        * pi_state->rt_mutex.
+                        *
+                        * We do not check for pid == 0 which can happen when
+                        * the owner died and robust_list_exit() cleared the
+                        * TID.
+                        */
+                       if (pid && pi_state->owner) {
+                               /*
+                                * Bail out if user space manipulated the
+                                * futex value.
+                                */
+                               if (pid != task_pid_vnr(pi_state->owner))
+                                       return -EINVAL;
+                       }
 
                        atomic_inc(&pi_state->refcount);
                        *ps = pi_state;
@@ -527,7 +573,7 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
         * change of the task flags, we do this protected by
         * p->pi_lock:
         */
-       spin_lock_irq(&p->pi_lock);
+       raw_spin_lock_irq(&p->pi_lock);
        if (unlikely(p->flags & PF_EXITING)) {
                /*
                 * The task is on the way out. When PF_EXITPIDONE is
@@ -536,7 +582,7 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
                 */
                int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
 
-               spin_unlock_irq(&p->pi_lock);
+               raw_spin_unlock_irq(&p->pi_lock);
                put_task_struct(p);
                return ret;
        }
@@ -555,7 +601,7 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
        WARN_ON(!list_empty(&pi_state->list));
        list_add(&pi_state->list, &p->pi_state_list);
        pi_state->owner = p;
-       spin_unlock_irq(&p->pi_lock);
+       raw_spin_unlock_irq(&p->pi_lock);
 
        put_task_struct(p);
 
@@ -565,7 +611,7 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
 }
 
 /**
- * futex_lock_pi_atomic() - atomic work required to acquire a pi aware futex
+ * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
  * @uaddr:             the pi futex user address
  * @hb:                        the pi futex hash bucket
  * @key:               the futex key associated with uaddr and hb
@@ -729,7 +775,14 @@ static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
        if (!pi_state)
                return -EINVAL;
 
-       spin_lock(&pi_state->pi_mutex.wait_lock);
+       /*
+        * If current does not own the pi_state then the futex is
+        * inconsistent and user space fiddled with the futex value.
+        */
+       if (pi_state->owner != current)
+               return -EINVAL;
+
+       raw_spin_lock(&pi_state->pi_mutex.wait_lock);
        new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
 
        /*
@@ -758,23 +811,23 @@ static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
                else if (curval != uval)
                        ret = -EINVAL;
                if (ret) {
-                       spin_unlock(&pi_state->pi_mutex.wait_lock);
+                       raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
                        return ret;
                }
        }
 
-       spin_lock_irq(&pi_state->owner->pi_lock);
+       raw_spin_lock_irq(&pi_state->owner->pi_lock);
        WARN_ON(list_empty(&pi_state->list));
        list_del_init(&pi_state->list);
-       spin_unlock_irq(&pi_state->owner->pi_lock);
+       raw_spin_unlock_irq(&pi_state->owner->pi_lock);
 
-       spin_lock_irq(&new_owner->pi_lock);
+       raw_spin_lock_irq(&new_owner->pi_lock);
        WARN_ON(!list_empty(&pi_state->list));
        list_add(&pi_state->list, &new_owner->pi_state_list);
        pi_state->owner = new_owner;
-       spin_unlock_irq(&new_owner->pi_lock);
+       raw_spin_unlock_irq(&new_owner->pi_lock);
 
-       spin_unlock(&pi_state->pi_mutex.wait_lock);
+       raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
        rt_mutex_unlock(&pi_state->pi_mutex);
 
        return 0;
@@ -836,7 +889,7 @@ static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
        if (!bitset)
                return -EINVAL;
 
-       ret = get_futex_key(uaddr, fshared, &key, VERIFY_READ);
+       ret = get_futex_key(uaddr, fshared, &key);
        if (unlikely(ret != 0))
                goto out;
 
@@ -882,21 +935,20 @@ futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
        int ret, op_ret;
 
 retry:
-       ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
+       ret = get_futex_key(uaddr1, fshared, &key1);
        if (unlikely(ret != 0))
                goto out;
-       ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out_put_key1;
 
        hb1 = hash_futex(&key1);
        hb2 = hash_futex(&key2);
 
-       double_lock_hb(hb1, hb2);
 retry_private:
+       double_lock_hb(hb1, hb2);
        op_ret = futex_atomic_op_inuser(op, uaddr2);
        if (unlikely(op_ret < 0)) {
-               u32 dummy;
 
                double_unlock_hb(hb1, hb2);
 
@@ -914,7 +966,7 @@ retry_private:
                        goto out_put_keys;
                }
 
-               ret = get_user(dummy, uaddr2);
+               ret = fault_in_user_writeable(uaddr2);
                if (ret)
                        goto out_put_keys;
 
@@ -980,7 +1032,7 @@ void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
                plist_add(&q->list, &hb2->chain);
                q->lock_ptr = &hb2->lock;
 #ifdef CONFIG_DEBUG_PI_LIST
-               q->list.plist.lock = &hb2->lock;
+               q->list.plist.spinlock = &hb2->lock;
 #endif
        }
        get_futex_key_refs(key2);
@@ -989,19 +1041,22 @@ void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
 
 /**
  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
- * q:  the futex_q
- * key:        the key of the requeue target futex
+ * @q:         the futex_q
+ * @key:       the key of the requeue target futex
+ * @hb:                the hash_bucket of the requeue target futex
  *
  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
  * to the requeue target futex so the waiter can detect the wakeup on the right
  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
- * atomic lock acquisition.  Must be called with the q->lock_ptr held.
+ * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
+ * to protect access to the pi_state to fixup the owner later.  Must be called
+ * with both q->lock_ptr and hb->lock held.
  */
 static inline
-void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key)
+void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
+                          struct futex_hash_bucket *hb)
 {
-       drop_futex_key_refs(&q->key);
        get_futex_key_refs(key);
        q->key = *key;
 
@@ -1011,6 +1066,11 @@ void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key)
        WARN_ON(!q->rt_waiter);
        q->rt_waiter = NULL;
 
+       q->lock_ptr = &hb->lock;
+#ifdef CONFIG_DEBUG_PI_LIST
+       q->list.plist.spinlock = &hb->lock;
+#endif
+
        wake_up_state(q->task, TASK_NORMAL);
 }
 
@@ -1061,6 +1121,10 @@ static int futex_proxy_trylock_atomic(u32 __user *pifutex,
        if (!top_waiter)
                return 0;
 
+       /* Ensure we requeue to the expected futex. */
+       if (!match_futex(top_waiter->requeue_pi_key, key2))
+               return -EINVAL;
+
        /*
         * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
         * the contended case or if set_waiters is 1.  The pi_state is returned
@@ -1069,7 +1133,7 @@ static int futex_proxy_trylock_atomic(u32 __user *pifutex,
        ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
                                   set_waiters);
        if (ret == 1)
-               requeue_pi_wake_futex(top_waiter, key2);
+               requeue_pi_wake_futex(top_waiter, key2, hb2);
 
        return ret;
 }
@@ -1133,11 +1197,10 @@ retry:
                pi_state = NULL;
        }
 
-       ret = get_futex_key(uaddr1, fshared, &key1, VERIFY_READ);
+       ret = get_futex_key(uaddr1, fshared, &key1);
        if (unlikely(ret != 0))
                goto out;
-       ret = get_futex_key(uaddr2, fshared, &key2,
-                           requeue_pi ? VERIFY_WRITE : VERIFY_READ);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out_put_key1;
 
@@ -1190,6 +1253,7 @@ retry_private:
                 */
                if (ret == 1) {
                        WARN_ON(pi_state);
+                       drop_count++;
                        task_count++;
                        ret = get_futex_value_locked(&curval2, uaddr2);
                        if (!ret)
@@ -1204,7 +1268,7 @@ retry_private:
                        double_unlock_hb(hb1, hb2);
                        put_futex_key(fshared, &key2);
                        put_futex_key(fshared, &key1);
-                       ret = get_user(curval2, uaddr2);
+                       ret = fault_in_user_writeable(uaddr2);
                        if (!ret)
                                goto retry;
                        goto out;
@@ -1228,8 +1292,15 @@ retry_private:
                if (!match_futex(&this->key, &key1))
                        continue;
 
-               WARN_ON(!requeue_pi && this->rt_waiter);
-               WARN_ON(requeue_pi && !this->rt_waiter);
+               /*
+                * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
+                * be paired with each other and no other futex ops.
+                */
+               if ((requeue_pi && !this->rt_waiter) ||
+                   (!requeue_pi && this->rt_waiter)) {
+                       ret = -EINVAL;
+                       break;
+               }
 
                /*
                 * Wake nr_wake waiters.  For requeue_pi, if we acquired the
@@ -1241,6 +1312,12 @@ retry_private:
                        continue;
                }
 
+               /* Ensure we requeue to the expected futex for requeue_pi. */
+               if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
+                       ret = -EINVAL;
+                       break;
+               }
+
                /*
                 * Requeue nr_requeue waiters and possibly one more in the case
                 * of requeue_pi if we couldn't acquire the lock atomically.
@@ -1254,7 +1331,8 @@ retry_private:
                                                        this->task, 1);
                        if (ret == 1) {
                                /* We got the lock. */
-                               requeue_pi_wake_futex(this, &key2);
+                               requeue_pi_wake_futex(this, &key2, hb2);
+                               drop_count++;
                                continue;
                        } else if (ret) {
                                /* -EDEADLK */
@@ -1302,6 +1380,25 @@ static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
        return hb;
 }
 
+static inline void
+queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
+{
+       spin_unlock(&hb->lock);
+       drop_futex_key_refs(&q->key);
+}
+
+/**
+ * queue_me() - Enqueue the futex_q on the futex_hash_bucket
+ * @q: The futex_q to enqueue
+ * @hb:        The destination hash bucket
+ *
+ * The hb->lock must be held by the caller, and is released here. A call to
+ * queue_me() is typically paired with exactly one call to unqueue_me().  The
+ * exceptions involve the PI related operations, which may use unqueue_me_pi()
+ * or nothing if the unqueue is done as part of the wake process and the unqueue
+ * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
+ * an example).
+ */
 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
 {
        int prio;
@@ -1318,26 +1415,24 @@ static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
 
        plist_node_init(&q->list, prio);
 #ifdef CONFIG_DEBUG_PI_LIST
-       q->list.plist.lock = &hb->lock;
+       q->list.plist.spinlock = &hb->lock;
 #endif
        plist_add(&q->list, &hb->chain);
        q->task = current;
        spin_unlock(&hb->lock);
 }
 
-static inline void
-queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
-{
-       spin_unlock(&hb->lock);
-       drop_futex_key_refs(&q->key);
-}
-
-/*
- * queue_me and unqueue_me must be called as a pair, each
- * exactly once.  They are called with the hashed spinlock held.
+/**
+ * unqueue_me() - Remove the futex_q from its futex_hash_bucket
+ * @q: The futex_q to unqueue
+ *
+ * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
+ * be paired with exactly one earlier call to queue_me().
+ *
+ * Returns:
+ *   1 - if the futex_q was still queued (and we removed unqueued it)
+ *   0 - if the futex_q was already removed by the waking thread
  */
-
-/* Return 1 if we were still queued (ie. 0 means we were woken) */
 static int unqueue_me(struct futex_q *q)
 {
        spinlock_t *lock_ptr;
@@ -1455,18 +1550,18 @@ retry:
         * itself.
         */
        if (pi_state->owner != NULL) {
-               spin_lock_irq(&pi_state->owner->pi_lock);
+               raw_spin_lock_irq(&pi_state->owner->pi_lock);
                WARN_ON(list_empty(&pi_state->list));
                list_del_init(&pi_state->list);
-               spin_unlock_irq(&pi_state->owner->pi_lock);
+               raw_spin_unlock_irq(&pi_state->owner->pi_lock);
        }
 
        pi_state->owner = newowner;
 
-       spin_lock_irq(&newowner->pi_lock);
+       raw_spin_lock_irq(&newowner->pi_lock);
        WARN_ON(!list_empty(&pi_state->list));
        list_add(&pi_state->list, &newowner->pi_state_list);
-       spin_unlock_irq(&newowner->pi_lock);
+       raw_spin_unlock_irq(&newowner->pi_lock);
        return 0;
 
        /*
@@ -1482,7 +1577,7 @@ retry:
 handle_fault:
        spin_unlock(q->lock_ptr);
 
-       ret = get_user(uval, uaddr);
+       ret = fault_in_user_writeable(uaddr);
 
        spin_lock(q->lock_ptr);
 
@@ -1507,7 +1602,6 @@ handle_fault:
 #define FLAGS_HAS_TIMEOUT      0x04
 
 static long futex_wait_restart(struct restart_block *restart);
-static long futex_lock_pi_restart(struct restart_block *restart);
 
 /**
  * fixup_owner() - Post lock pi_state and corner case management
@@ -1591,17 +1685,14 @@ out:
 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
                                struct hrtimer_sleeper *timeout)
 {
-       queue_me(q, hb);
-
        /*
-        * There might have been scheduling since the queue_me(), as we
-        * cannot hold a spinlock across the get_user() in case it
-        * faults, and we cannot just set TASK_INTERRUPTIBLE state when
-        * queueing ourselves into the futex hash. This code thus has to
-        * rely on the futex_wake() code removing us from hash when it
-        * wakes us up.
+        * The task state is guaranteed to be set before another task can
+        * wake it. set_current_state() is implemented using set_mb() and
+        * queue_me() calls spin_unlock() upon completion, both serializing
+        * access to the hash list and forcing another memory barrier.
         */
        set_current_state(TASK_INTERRUPTIBLE);
+       queue_me(q, hb);
 
        /* Arm the timer */
        if (timeout) {
@@ -1611,8 +1702,8 @@ static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
        }
 
        /*
-        * !plist_node_empty() is safe here without any lock.
-        * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
+        * If we have been removed from the hash list, then another task
+        * has tried to wake us, and we can skip the call to schedule().
         */
        if (likely(!plist_node_empty(&q->list))) {
                /*
@@ -1668,7 +1759,7 @@ static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
         */
 retry:
        q->key = FUTEX_KEY_INIT;
-       ret = get_futex_key(uaddr, fshared, &q->key, VERIFY_READ);
+       ret = get_futex_key(uaddr, fshared, &q->key);
        if (unlikely(ret != 0))
                return ret;
 
@@ -1717,6 +1808,7 @@ static int futex_wait(u32 __user *uaddr, int fshared,
        q.pi_state = NULL;
        q.bitset = bitset;
        q.rt_waiter = NULL;
+       q.requeue_pi_key = NULL;
 
        if (abs_time) {
                to = &timeout;
@@ -1728,6 +1820,7 @@ static int futex_wait(u32 __user *uaddr, int fshared,
                                             current->timer_slack_ns);
        }
 
+retry:
        /* Prepare to wait on uaddr. */
        ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
        if (ret)
@@ -1745,9 +1838,14 @@ static int futex_wait(u32 __user *uaddr, int fshared,
                goto out_put_key;
 
        /*
-        * We expect signal_pending(current), but another thread may
-        * have handled it for us already.
+        * We expect signal_pending(current), but we might be the
+        * victim of a spurious wakeup as well.
         */
+       if (!signal_pending(current)) {
+               put_futex_key(fshared, &q.key);
+               goto retry;
+       }
+
        ret = -ERESTARTSYS;
        if (!abs_time)
                goto out_put_key;
@@ -1808,7 +1906,6 @@ static int futex_lock_pi(u32 __user *uaddr, int fshared,
 {
        struct hrtimer_sleeper timeout, *to = NULL;
        struct futex_hash_bucket *hb;
-       u32 uval;
        struct futex_q q;
        int res, ret;
 
@@ -1825,9 +1922,10 @@ static int futex_lock_pi(u32 __user *uaddr, int fshared,
 
        q.pi_state = NULL;
        q.rt_waiter = NULL;
+       q.requeue_pi_key = NULL;
 retry:
        q.key = FUTEX_KEY_INIT;
-       ret = get_futex_key(uaddr, fshared, &q.key, VERIFY_WRITE);
+       ret = get_futex_key(uaddr, fshared, &q.key);
        if (unlikely(ret != 0))
                goto out;
 
@@ -1897,7 +1995,7 @@ retry_private:
        /* Unqueue and drop the lock */
        unqueue_me_pi(&q);
 
-       goto out;
+       goto out_put_key;
 
 out_unlock_put_key:
        queue_unlock(&q, hb);
@@ -1910,16 +2008,9 @@ out:
        return ret != -EINTR ? ret : -ERESTARTNOINTR;
 
 uaddr_faulted:
-       /*
-        * We have to r/w  *(int __user *)uaddr, and we have to modify it
-        * atomically.  Therefore, if we continue to fault after get_user()
-        * below, we need to handle the fault ourselves, while still holding
-        * the mmap_sem.  This can occur if the uaddr is under contention as
-        * we have to drop the mmap_sem in order to call get_user().
-        */
        queue_unlock(&q, hb);
 
-       ret = get_user(uval, uaddr);
+       ret = fault_in_user_writeable(uaddr);
        if (ret)
                goto out_put_key;
 
@@ -1930,21 +2021,6 @@ uaddr_faulted:
        goto retry;
 }
 
-static long futex_lock_pi_restart(struct restart_block *restart)
-{
-       u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
-       ktime_t t, *tp = NULL;
-       int fshared = restart->futex.flags & FLAGS_SHARED;
-
-       if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
-               t.tv64 = restart->futex.time;
-               tp = &t;
-       }
-       restart->fn = do_no_restart_syscall;
-
-       return (long)futex_lock_pi(uaddr, fshared, restart->futex.val, tp, 0);
-}
-
 /*
  * Userspace attempted a TID -> 0 atomic transition, and failed.
  * This is the in-kernel slowpath: we look up the PI state (if any),
@@ -1968,7 +2044,7 @@ retry:
        if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
                return -EPERM;
 
-       ret = get_futex_key(uaddr, fshared, &key, VERIFY_WRITE);
+       ret = get_futex_key(uaddr, fshared, &key);
        if (unlikely(ret != 0))
                goto out;
 
@@ -2029,17 +2105,10 @@ out:
        return ret;
 
 pi_faulted:
-       /*
-        * We have to r/w  *(int __user *)uaddr, and we have to modify it
-        * atomically.  Therefore, if we continue to fault after get_user()
-        * below, we need to handle the fault ourselves, while still holding
-        * the mmap_sem.  This can occur if the uaddr is under contention as
-        * we have to drop the mmap_sem in order to call get_user().
-        */
        spin_unlock(&hb->lock);
        put_futex_key(fshared, &key);
 
-       ret = get_user(uval, uaddr);
+       ret = fault_in_user_writeable(uaddr);
        if (!ret)
                goto retry;
 
@@ -2060,7 +2129,7 @@ pi_faulted:
  *
  * Returns
  *  0 - no early wakeup detected
- * <0 - -ETIMEDOUT or -ERESTARTSYS (FIXME: or ERESTARTNOINTR?)
+ * <0 - -ETIMEDOUT or -ERESTARTNOINTR
  */
 static inline
 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
@@ -2083,31 +2152,25 @@ int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
                 * Unqueue the futex_q and determine which it was.
                 */
                plist_del(&q->list, &q->list.plist);
-               drop_futex_key_refs(&q->key);
 
+               /* Handle spurious wakeups gracefully */
+               ret = -EWOULDBLOCK;
                if (timeout && !timeout->task)
                        ret = -ETIMEDOUT;
-               else {
-                       /*
-                        * We expect signal_pending(current), but another
-                        * thread may have handled it for us already.
-                        */
-                       /* FIXME: ERESTARTSYS or ERESTARTNOINTR?  Do we care if
-                        * the user specified SA_RESTART or not? */
-                       ret = -ERESTARTSYS;
-               }
+               else if (signal_pending(current))
+                       ret = -ERESTARTNOINTR;
        }
        return ret;
 }
 
 /**
  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
- * @uaddr:     the futex we initialyl wait on (non-pi)
+ * @uaddr:     the futex we initially wait on (non-pi)
  * @fshared:   whether the futexes are shared (1) or not (0).  They must be
  *             the same type, no requeueing from private to shared, etc.
  * @val:       the expected value of uaddr
  * @abs_time:  absolute timeout
- * @bitset:    32 bit wakeup bitset set by userspace, defaults to all.
+ * @bitset:    32 bit wakeup bitset set by userspace, defaults to all
  * @clockrt:   whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
  * @uaddr2:    the pi futex we will take prior to returning to user-space
  *
@@ -2121,11 +2184,11 @@ int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
  * We call schedule in futex_wait_queue_me() when we enqueue and return there
  * via the following:
  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
- * 2) wakeup on uaddr2 after a requeue and subsequent unlock
- * 3) signal (before or after requeue)
- * 4) timeout (before or after requeue)
+ * 2) wakeup on uaddr2 after a requeue
+ * 3) signal
+ * 4) timeout
  *
- * If 3, we setup a restart_block with futex_wait_requeue_pi() as the function.
+ * If 3, cleanup and return -ERESTARTNOINTR.
  *
  * If 2, we may then block on trying to take the rt_mutex and return via:
  * 5) successful lock
@@ -2133,7 +2196,7 @@ int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
  * 7) timeout
  * 8) other lock acquisition failure
  *
- * If 6, we setup a restart_block with futex_lock_pi() as the function.
+ * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
  *
  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
  *
@@ -2148,12 +2211,10 @@ static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
        struct hrtimer_sleeper timeout, *to = NULL;
        struct rt_mutex_waiter rt_waiter;
        struct rt_mutex *pi_mutex = NULL;
-       struct restart_block *restart;
        struct futex_hash_bucket *hb;
        union futex_key key2;
        struct futex_q q;
        int res, ret;
-       u32 uval;
 
        if (!bitset)
                return -EINVAL;
@@ -2174,15 +2235,16 @@ static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
        debug_rt_mutex_init_waiter(&rt_waiter);
        rt_waiter.task = NULL;
 
-       q.pi_state = NULL;
-       q.bitset = bitset;
-       q.rt_waiter = &rt_waiter;
-
        key2 = FUTEX_KEY_INIT;
-       ret = get_futex_key(uaddr2, fshared, &key2, VERIFY_WRITE);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out;
 
+       q.pi_state = NULL;
+       q.bitset = bitset;
+       q.rt_waiter = &rt_waiter;
+       q.requeue_pi_key = &key2;
+
        /* Prepare to wait on uaddr. */
        ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
        if (ret)
@@ -2235,7 +2297,7 @@ static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
                res = fixup_owner(uaddr2, fshared, &q, !ret);
                /*
                 * If fixup_owner() returned an error, proprogate that.  If it
-                * acquired the lock, clear our -ETIMEDOUT or -EINTR.
+                * acquired the lock, clear -ETIMEDOUT or -EINTR.
                 */
                if (res)
                        ret = (res < 0) ? res : 0;
@@ -2252,30 +2314,14 @@ static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
                if (rt_mutex_owner(pi_mutex) == current)
                        rt_mutex_unlock(pi_mutex);
        } else if (ret == -EINTR) {
-               ret = -EFAULT;
-               if (get_user(uval, uaddr2))
-                       goto out_put_keys;
-
                /*
-                * We've already been requeued, so restart by calling
-                * futex_lock_pi() directly, rather then returning to this
-                * function.
+                * We've already been requeued, but cannot restart by calling
+                * futex_lock_pi() directly. We could restart this syscall, but
+                * it would detect that the user space "val" changed and return
+                * -EWOULDBLOCK.  Save the overhead of the restart and return
+                * -EWOULDBLOCK directly.
                 */
-               ret = -ERESTART_RESTARTBLOCK;
-               restart = &current_thread_info()->restart_block;
-               restart->fn = futex_lock_pi_restart;
-               restart->futex.uaddr = (u32 *)uaddr2;
-               restart->futex.val = uval;
-               restart->futex.flags = 0;
-               if (abs_time) {
-                       restart->futex.flags |= FLAGS_HAS_TIMEOUT;
-                       restart->futex.time = abs_time->tv64;
-               }
-
-               if (fshared)
-                       restart->futex.flags |= FLAGS_SHARED;
-               if (clockrt)
-                       restart->futex.flags |= FLAGS_CLOCKRT;
+               ret = -EWOULDBLOCK;
        }
 
 out_put_keys:
@@ -2307,9 +2353,9 @@ out:
  */
 
 /**
- * sys_set_robust_list - set the robust-futex list head of a task
- * @head: pointer to the list-head
- * @len: length of the list-head, as userspace expects
+ * sys_set_robust_list() - Set the robust-futex list head of a task
+ * @head:      pointer to the list-head
+ * @len:       length of the list-head, as userspace expects
  */
 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
                size_t, len)
@@ -2328,10 +2374,10 @@ SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
 }
 
 /**
- * sys_get_robust_list - get the robust-futex list head of a task
- * @pid: pid of the process [zero for current task]
- * @head_ptr: pointer to a list-head pointer, the kernel fills it in
- * @len_ptr: pointer to a length field, the kernel fills in the header size
+ * sys_get_robust_list() - Get the robust-futex list head of a task
+ * @pid:       pid of the process [zero for current task]
+ * @head_ptr:  pointer to a list-head pointer, the kernel fills it in
+ * @len_ptr:   pointer to a length field, the kernel fills in the header size
  */
 SYSCALL_DEFINE3(get_robust_list, int, pid,
                struct robust_list_head __user * __user *, head_ptr,