Memory controller: OOM handling
[linux-2.6.git] / kernel / futex.c
index cf0c8e2..a6baaec 100644 (file)
@@ -16,6 +16,9 @@
  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
  *
+ *  PRIVATE futexes by Eric Dumazet
+ *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
+ *
  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
  *  enough at me, Linus for the original (flawed) idea, Matthew
  *  Kirkwood for proof-of-concept implementation.
 #include <linux/pagemap.h>
 #include <linux/syscalls.h>
 #include <linux/signal.h>
+#include <linux/module.h>
+#include <linux/magic.h>
+#include <linux/pid.h>
+#include <linux/nsproxy.h>
+
 #include <asm/futex.h>
 
 #include "rtmutex_common.h"
 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
 
 /*
- * Futexes are matched on equal values of this key.
- * The key type depends on whether it's a shared or private mapping.
- * Don't rearrange members without looking at hash_futex().
- *
- * offset is aligned to a multiple of sizeof(u32) (== 4) by definition.
- * We set bit 0 to indicate if it's an inode-based key.
- */
-union futex_key {
-       struct {
-               unsigned long pgoff;
-               struct inode *inode;
-               int offset;
-       } shared;
-       struct {
-               unsigned long address;
-               struct mm_struct *mm;
-               int offset;
-       } private;
-       struct {
-               unsigned long word;
-               void *ptr;
-               int offset;
-       } both;
-};
-
-/*
  * Priority Inheritance state:
  */
 struct futex_pi_state {
@@ -106,12 +88,12 @@ struct futex_pi_state {
  * 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 list_empty(&q->list) || q->lock_ptr == 0.
+ * 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->waiters, then make the second condition true.
  */
 struct futex_q {
-       struct list_head list;
+       struct plist_node list;
        wait_queue_head_t waiters;
 
        /* Which hash list lock to use: */
@@ -127,14 +109,17 @@ struct futex_q {
        /* Optional priority inheritance state: */
        struct futex_pi_state *pi_state;
        struct task_struct *task;
+
+       /* Bitset for the optional bitmasked wakeup */
+       u32 bitset;
 };
 
 /*
  * Split the global futex_lock into every hash list lock.
  */
 struct futex_hash_bucket {
-       spinlock_t              lock;
-       struct list_head       chain;
+       spinlock_t lock;
+       struct plist_head chain;
 };
 
 static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
@@ -143,6 +128,24 @@ static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
 static struct vfsmount *futex_mnt;
 
 /*
+ * Take mm->mmap_sem, when futex is shared
+ */
+static inline void futex_lock_mm(struct rw_semaphore *fshared)
+{
+       if (fshared)
+               down_read(fshared);
+}
+
+/*
+ * Release mm->mmap_sem, when the futex is shared
+ */
+static inline void futex_unlock_mm(struct rw_semaphore *fshared)
+{
+       if (fshared)
+               up_read(fshared);
+}
+
+/*
  * We hash on the keys returned from get_futex_key (see below).
  */
 static struct futex_hash_bucket *hash_futex(union futex_key *key)
@@ -163,19 +166,26 @@ static inline int match_futex(union futex_key *key1, union futex_key *key2)
                && key1->both.offset == key2->both.offset);
 }
 
-/*
- * Get parameters which are the keys for a futex.
+/**
+ * get_futex_key - Get parameters which are the keys for a futex.
+ * @uaddr: virtual address of the futex
+ * @shared: NULL for a PROCESS_PRIVATE futex,
+ *     &current->mm->mmap_sem for a PROCESS_SHARED futex
+ * @key: address where result is stored.
  *
- * For shared mappings, it's (page->index, vma->vm_file->f_dentry->d_inode,
+ * Returns a negative error code or 0
+ * The key words are stored in *key on success.
+ *
+ * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
  * We can usually work out the index without swapping in the page.
  *
- * Returns: 0, or negative error code.
- * The key words are stored in *key on success.
- *
- * Should be called with &current->mm->mmap_sem but NOT any spinlocks.
+ * fshared is NULL for PROCESS_PRIVATE futexes
+ * For other futexes, it points to &current->mm->mmap_sem and
+ * caller must have taken the reader lock. but NOT any spinlocks.
  */
-static int get_futex_key(u32 __user *uaddr, union futex_key *key)
+static int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
+                        union futex_key *key)
 {
        unsigned long address = (unsigned long)uaddr;
        struct mm_struct *mm = current->mm;
@@ -187,11 +197,25 @@ static int get_futex_key(u32 __user *uaddr, union futex_key *key)
         * The futex address must be "naturally" aligned.
         */
        key->both.offset = address % PAGE_SIZE;
-       if (unlikely((key->both.offset % sizeof(u32)) != 0))
+       if (unlikely((address % sizeof(u32)) != 0))
                return -EINVAL;
        address -= key->both.offset;
 
        /*
+        * PROCESS_PRIVATE futexes are fast.
+        * As the mm cannot disappear under us and the 'key' only needs
+        * virtual address, we dont even have to find the underlying vma.
+        * Note : We do have to check 'uaddr' is a valid user address,
+        *        but access_ok() should be faster than find_vma()
+        */
+       if (!fshared) {
+               if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
+                       return -EFAULT;
+               key->private.mm = mm;
+               key->private.address = address;
+               return 0;
+       }
+       /*
         * The futex is hashed differently depending on whether
         * it's in a shared or private mapping.  So check vma first.
         */
@@ -215,6 +239,7 @@ static int get_futex_key(u32 __user *uaddr, union futex_key *key)
         * mappings of _writable_ handles.
         */
        if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
+               key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
                key->private.mm = mm;
                key->private.address = address;
                return 0;
@@ -223,8 +248,8 @@ static int get_futex_key(u32 __user *uaddr, union futex_key *key)
        /*
         * Linear file mappings are also simple.
         */
-       key->shared.inode = vma->vm_file->f_dentry->d_inode;
-       key->both.offset++; /* Bit 0 of offset indicates inode-based key. */
+       key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
+       key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
        if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
                key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
                                     + vma->vm_pgoff);
@@ -251,16 +276,18 @@ static int get_futex_key(u32 __user *uaddr, union futex_key *key)
  * Take a reference to the resource addressed by a key.
  * Can be called while holding spinlocks.
  *
- * NOTE: mmap_sem MUST be held between get_futex_key() and calling this
- * function, if it is called at all.  mmap_sem keeps key->shared.inode valid.
  */
-static inline void get_key_refs(union futex_key *key)
+static void get_futex_key_refs(union futex_key *key)
 {
-       if (key->both.ptr != 0) {
-               if (key->both.offset & 1)
+       if (key->both.ptr == 0)
+               return;
+       switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
+               case FUT_OFF_INODE:
                        atomic_inc(&key->shared.inode->i_count);
-               else
+                       break;
+               case FUT_OFF_MMSHARED:
                        atomic_inc(&key->private.mm->mm_count);
+                       break;
        }
 }
 
@@ -268,50 +295,80 @@ static inline void get_key_refs(union futex_key *key)
  * Drop a reference to the resource addressed by a key.
  * The hash bucket spinlock must not be held.
  */
-static void drop_key_refs(union futex_key *key)
+static void drop_futex_key_refs(union futex_key *key)
 {
-       if (key->both.ptr != 0) {
-               if (key->both.offset & 1)
+       if (!key->both.ptr)
+               return;
+       switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
+               case FUT_OFF_INODE:
                        iput(key->shared.inode);
-               else
+                       break;
+               case FUT_OFF_MMSHARED:
                        mmdrop(key->private.mm);
+                       break;
        }
 }
 
-static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
+static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
+{
+       u32 curval;
+
+       pagefault_disable();
+       curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
+       pagefault_enable();
+
+       return curval;
+}
+
+static int get_futex_value_locked(u32 *dest, u32 __user *from)
 {
        int ret;
 
-       inc_preempt_count();
+       pagefault_disable();
        ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
-       dec_preempt_count();
+       pagefault_enable();
 
        return ret ? -EFAULT : 0;
 }
 
 /*
- * Fault handling. Called with current->mm->mmap_sem held.
+ * Fault handling.
+ * if fshared is non NULL, current->mm->mmap_sem is already held
  */
-static int futex_handle_fault(unsigned long address, int attempt)
+static int futex_handle_fault(unsigned long address,
+                             struct rw_semaphore *fshared, int attempt)
 {
        struct vm_area_struct * vma;
        struct mm_struct *mm = current->mm;
+       int ret = -EFAULT;
 
-       if (attempt >= 2 || !(vma = find_vma(mm, address)) ||
-           vma->vm_start > address || !(vma->vm_flags & VM_WRITE))
-               return -EFAULT;
+       if (attempt > 2)
+               return ret;
 
-       switch (handle_mm_fault(mm, vma, address, 1)) {
-       case VM_FAULT_MINOR:
-               current->min_flt++;
-               break;
-       case VM_FAULT_MAJOR:
-               current->maj_flt++;
-               break;
-       default:
-               return -EFAULT;
+       if (!fshared)
+               down_read(&mm->mmap_sem);
+       vma = find_vma(mm, address);
+       if (vma && address >= vma->vm_start &&
+           (vma->vm_flags & VM_WRITE)) {
+               int fault;
+               fault = handle_mm_fault(mm, vma, address, 1);
+               if (unlikely((fault & VM_FAULT_ERROR))) {
+#if 0
+                       /* XXX: let's do this when we verify it is OK */
+                       if (ret & VM_FAULT_OOM)
+                               ret = -ENOMEM;
+#endif
+               } else {
+                       ret = 0;
+                       if (fault & VM_FAULT_MAJOR)
+                               current->maj_flt++;
+                       else
+                               current->min_flt++;
+               }
        }
-       return 0;
+       if (!fshared)
+               up_read(&mm->mmap_sem);
+       return ret;
 }
 
 /*
@@ -324,12 +381,11 @@ static int refill_pi_state_cache(void)
        if (likely(current->pi_state_cache))
                return 0;
 
-       pi_state = kmalloc(sizeof(*pi_state), GFP_KERNEL);
+       pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
 
        if (!pi_state)
                return -ENOMEM;
 
-       memset(pi_state, 0, sizeof(*pi_state));
        INIT_LIST_HEAD(&pi_state->list);
        /* pi_mutex gets initialized later */
        pi_state->owner = NULL;
@@ -389,21 +445,14 @@ static struct task_struct * futex_find_get_task(pid_t pid)
 {
        struct task_struct *p;
 
-       read_lock(&tasklist_lock);
-       p = find_task_by_pid(pid);
-       if (!p)
-               goto out_unlock;
-       if ((current->euid != p->euid) && (current->euid != p->uid)) {
-               p = NULL;
-               goto out_unlock;
-       }
-       if (p->state == EXIT_ZOMBIE || p->exit_state == EXIT_ZOMBIE) {
-               p = NULL;
-               goto out_unlock;
-       }
-       get_task_struct(p);
-out_unlock:
-       read_unlock(&tasklist_lock);
+       rcu_read_lock();
+       p = find_task_by_vpid(pid);
+       if (!p || ((current->euid != p->euid) && (current->euid != p->uid)))
+               p = ERR_PTR(-ESRCH);
+       else
+               get_task_struct(p);
+
+       rcu_read_unlock();
 
        return p;
 }
@@ -415,15 +464,15 @@ out_unlock:
  */
 void exit_pi_state_list(struct task_struct *curr)
 {
-       struct futex_hash_bucket *hb;
        struct list_head *next, *head = &curr->pi_state_list;
        struct futex_pi_state *pi_state;
+       struct futex_hash_bucket *hb;
        union futex_key key;
 
        /*
         * We are a ZOMBIE and nobody can enqueue itself on
         * pi_state_list anymore, but we have to be careful
-        * versus waiters unqueueing themselfs
+        * versus waiters unqueueing themselves:
         */
        spin_lock_irq(&curr->pi_lock);
        while (!list_empty(head)) {
@@ -431,21 +480,24 @@ void exit_pi_state_list(struct task_struct *curr)
                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);
 
-               hb = hash_futex(&key);
                spin_lock(&hb->lock);
 
                spin_lock_irq(&curr->pi_lock);
+               /*
+                * We dropped the pi-lock, so re-check whether this
+                * task still owns the PI-state:
+                */
                if (head->next != next) {
                        spin_unlock(&hb->lock);
                        continue;
                }
 
-               list_del_init(&pi_state->list);
-
                WARN_ON(pi_state->owner != curr);
-
+               WARN_ON(list_empty(&pi_state->list));
+               list_del_init(&pi_state->list);
                pi_state->owner = NULL;
                spin_unlock_irq(&curr->pi_lock);
 
@@ -459,18 +511,19 @@ void exit_pi_state_list(struct task_struct *curr)
 }
 
 static int
-lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
+lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
+               union futex_key *key, struct futex_pi_state **ps)
 {
        struct futex_pi_state *pi_state = NULL;
        struct futex_q *this, *next;
-       struct list_head *head;
+       struct plist_head *head;
        struct task_struct *p;
-       pid_t pid;
+       pid_t pid = uval & FUTEX_TID_MASK;
 
        head = &hb->chain;
 
-       list_for_each_entry_safe(this, next, head, list) {
-               if (match_futex (&this->key, &me->key)) {
+       plist_for_each_entry_safe(this, next, head, list) {
+               if (match_futex(&this->key, key)) {
                        /*
                         * Another waiter already exists - bump up
                         * the refcount and return its pi_state:
@@ -482,21 +535,46 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
                        if (unlikely(!pi_state))
                                return -EINVAL;
 
+                       WARN_ON(!atomic_read(&pi_state->refcount));
+                       WARN_ON(pid && pi_state->owner &&
+                               pi_state->owner->pid != pid);
+
                        atomic_inc(&pi_state->refcount);
-                       me->pi_state = pi_state;
+                       *ps = pi_state;
 
                        return 0;
                }
        }
 
        /*
-        * We are the first waiter - try to look up the real owner and
-        * attach the new pi_state to it:
+        * We are the first waiter - try to look up the real owner and attach
+        * the new pi_state to it, but bail out when TID = 0
         */
-       pid = uval & FUTEX_TID_MASK;
-       p = futex_find_get_task(pid);
-       if (!p)
+       if (!pid)
                return -ESRCH;
+       p = futex_find_get_task(pid);
+       if (IS_ERR(p))
+               return PTR_ERR(p);
+
+       /*
+        * We need to look at the task state flags to figure out,
+        * whether the task is exiting. To protect against the do_exit
+        * change of the task flags, we do this protected by
+        * p->pi_lock:
+        */
+       spin_lock_irq(&p->pi_lock);
+       if (unlikely(p->flags & PF_EXITING)) {
+               /*
+                * The task is on the way out. When PF_EXITPIDONE is
+                * set, we know that the task has finished the
+                * cleanup:
+                */
+               int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
+
+               spin_unlock_irq(&p->pi_lock);
+               put_task_struct(p);
+               return ret;
+       }
 
        pi_state = alloc_pi_state();
 
@@ -507,16 +585,16 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
        rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
 
        /* Store the key for possible exit cleanups: */
-       pi_state->key = me->key;
+       pi_state->key = *key;
 
-       spin_lock_irq(&p->pi_lock);
+       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);
 
        put_task_struct(p);
 
-       me->pi_state = pi_state;
+       *ps = pi_state;
 
        return 0;
 }
@@ -527,12 +605,12 @@ lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
  */
 static void wake_futex(struct futex_q *q)
 {
-       list_del_init(&q->list);
+       plist_del(&q->list, &q->list.plist);
        if (q->filp)
                send_sigio(&q->filp->f_owner, q->fd, POLL_IN);
        /*
         * The lock in wake_up_all() is a crucial memory barrier after the
-        * list_del_init() and also before assigning to q->lock_ptr.
+        * plist_del() and also before assigning to q->lock_ptr.
         */
        wake_up_all(&q->waiters);
        /*
@@ -544,7 +622,7 @@ static void wake_futex(struct futex_q *q)
         * at the end of wake_up_all() does not prevent this store from
         * moving.
         */
-       wmb();
+       smp_wmb();
        q->lock_ptr = NULL;
 }
 
@@ -557,6 +635,7 @@ 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);
        new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
 
        /*
@@ -573,20 +652,35 @@ static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
         * kept enabled while there is PI state around. We must also
         * preserve the owner died bit.)
         */
-       newval = (uval & FUTEX_OWNER_DIED) | FUTEX_WAITERS | new_owner->pid;
+       if (!(uval & FUTEX_OWNER_DIED)) {
+               int ret = 0;
 
-       inc_preempt_count();
-       curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
-       dec_preempt_count();
+               newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
 
-       if (curval == -EFAULT)
-               return -EFAULT;
-       if (curval != uval)
-               return -EINVAL;
+               curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
 
-       list_del_init(&pi_state->owner->pi_state_list);
+               if (curval == -EFAULT)
+                       ret = -EFAULT;
+               else if (curval != uval)
+                       ret = -EINVAL;
+               if (ret) {
+                       spin_unlock(&pi_state->pi_mutex.wait_lock);
+                       return ret;
+               }
+       }
+
+       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);
+
+       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);
+
+       spin_unlock(&pi_state->pi_mutex.wait_lock);
        rt_mutex_unlock(&pi_state->pi_mutex);
 
        return 0;
@@ -600,9 +694,7 @@ static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
         * There is no waiter, so we unlock the futex. The owner died
         * bit has not to be preserved here. We are the owner:
         */
-       inc_preempt_count();
-       oldval = futex_atomic_cmpxchg_inatomic(uaddr, uval, 0);
-       dec_preempt_count();
+       oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
 
        if (oldval == -EFAULT)
                return oldval;
@@ -632,17 +724,21 @@ double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
  * Wake up all waiters hashed on the physical page that is mapped
  * to this virtual address:
  */
-static int futex_wake(u32 __user *uaddr, int nr_wake)
+static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
+                     int nr_wake, u32 bitset)
 {
        struct futex_hash_bucket *hb;
        struct futex_q *this, *next;
-       struct list_head *head;
+       struct plist_head *head;
        union futex_key key;
        int ret;
 
-       down_read(&current->mm->mmap_sem);
+       if (!bitset)
+               return -EINVAL;
 
-       ret = get_futex_key(uaddr, &key);
+       futex_lock_mm(fshared);
+
+       ret = get_futex_key(uaddr, fshared, &key);
        if (unlikely(ret != 0))
                goto out;
 
@@ -650,12 +746,17 @@ static int futex_wake(u32 __user *uaddr, int nr_wake)
        spin_lock(&hb->lock);
        head = &hb->chain;
 
-       list_for_each_entry_safe(this, next, head, list) {
+       plist_for_each_entry_safe(this, next, head, list) {
                if (match_futex (&this->key, &key)) {
                        if (this->pi_state) {
                                ret = -EINVAL;
                                break;
                        }
+
+                       /* Check if one of the bits is set in both bitsets */
+                       if (!(this->bitset & bitset))
+                               continue;
+
                        wake_futex(this);
                        if (++ret >= nr_wake)
                                break;
@@ -664,7 +765,7 @@ static int futex_wake(u32 __user *uaddr, int nr_wake)
 
        spin_unlock(&hb->lock);
 out:
-       up_read(&current->mm->mmap_sem);
+       futex_unlock_mm(fshared);
        return ret;
 }
 
@@ -673,22 +774,23 @@ out:
  * to this virtual address:
  */
 static int
-futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
+futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
+             u32 __user *uaddr2,
              int nr_wake, int nr_wake2, int op)
 {
        union futex_key key1, key2;
        struct futex_hash_bucket *hb1, *hb2;
-       struct list_head *head;
+       struct plist_head *head;
        struct futex_q *this, *next;
        int ret, op_ret, attempt = 0;
 
 retryfull:
-       down_read(&current->mm->mmap_sem);
+       futex_lock_mm(fshared);
 
-       ret = get_futex_key(uaddr1, &key1);
+       ret = get_futex_key(uaddr1, fshared, &key1);
        if (unlikely(ret != 0))
                goto out;
-       ret = get_futex_key(uaddr2, &key2);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out;
 
@@ -728,8 +830,9 @@ retry:
                 * still holding the mmap_sem.
                 */
                if (attempt++) {
-                       if (futex_handle_fault((unsigned long)uaddr2,
-                                              attempt))
+                       ret = futex_handle_fault((unsigned long)uaddr2,
+                                                fshared, attempt);
+                       if (ret)
                                goto out;
                        goto retry;
                }
@@ -738,7 +841,7 @@ retry:
                 * If we would have faulted, release mmap_sem,
                 * fault it in and start all over again.
                 */
-               up_read(&current->mm->mmap_sem);
+               futex_unlock_mm(fshared);
 
                ret = get_user(dummy, uaddr2);
                if (ret)
@@ -749,7 +852,7 @@ retry:
 
        head = &hb1->chain;
 
-       list_for_each_entry_safe(this, next, head, list) {
+       plist_for_each_entry_safe(this, next, head, list) {
                if (match_futex (&this->key, &key1)) {
                        wake_futex(this);
                        if (++ret >= nr_wake)
@@ -761,7 +864,7 @@ retry:
                head = &hb2->chain;
 
                op_ret = 0;
-               list_for_each_entry_safe(this, next, head, list) {
+               plist_for_each_entry_safe(this, next, head, list) {
                        if (match_futex (&this->key, &key2)) {
                                wake_futex(this);
                                if (++op_ret >= nr_wake2)
@@ -775,7 +878,8 @@ retry:
        if (hb1 != hb2)
                spin_unlock(&hb2->lock);
 out:
-       up_read(&current->mm->mmap_sem);
+       futex_unlock_mm(fshared);
+
        return ret;
 }
 
@@ -783,22 +887,23 @@ out:
  * Requeue all waiters hashed on one physical page to another
  * physical page.
  */
-static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
+static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
+                        u32 __user *uaddr2,
                         int nr_wake, int nr_requeue, u32 *cmpval)
 {
        union futex_key key1, key2;
        struct futex_hash_bucket *hb1, *hb2;
-       struct list_head *head1;
+       struct plist_head *head1;
        struct futex_q *this, *next;
        int ret, drop_count = 0;
 
  retry:
-       down_read(&current->mm->mmap_sem);
+       futex_lock_mm(fshared);
 
-       ret = get_futex_key(uaddr1, &key1);
+       ret = get_futex_key(uaddr1, fshared, &key1);
        if (unlikely(ret != 0))
                goto out;
-       ret = get_futex_key(uaddr2, &key2);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out;
 
@@ -821,7 +926,7 @@ static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
                         * If we would have faulted, release mmap_sem, fault
                         * it in and start all over again.
                         */
-                       up_read(&current->mm->mmap_sem);
+                       futex_unlock_mm(fshared);
 
                        ret = get_user(curval, uaddr1);
 
@@ -837,7 +942,7 @@ static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
        }
 
        head1 = &hb1->chain;
-       list_for_each_entry_safe(this, next, head1, list) {
+       plist_for_each_entry_safe(this, next, head1, list) {
                if (!match_futex (&this->key, &key1))
                        continue;
                if (++ret <= nr_wake) {
@@ -848,11 +953,15 @@ static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
                         * requeue.
                         */
                        if (likely(head1 != &hb2->chain)) {
-                               list_move_tail(&this->list, &hb2->chain);
+                               plist_del(&this->list, &hb1->chain);
+                               plist_add(&this->list, &hb2->chain);
                                this->lock_ptr = &hb2->lock;
+#ifdef CONFIG_DEBUG_PI_LIST
+                               this->list.plist.lock = &hb2->lock;
+#endif
                        }
                        this->key = key2;
-                       get_key_refs(&key2);
+                       get_futex_key_refs(&key2);
                        drop_count++;
 
                        if (ret - nr_wake >= nr_requeue)
@@ -865,12 +974,12 @@ out_unlock:
        if (hb1 != hb2)
                spin_unlock(&hb2->lock);
 
-       /* drop_key_refs() must be called outside the spinlocks. */
+       /* drop_futex_key_refs() must be called outside the spinlocks. */
        while (--drop_count >= 0)
-               drop_key_refs(&key1);
+               drop_futex_key_refs(&key1);
 
 out:
-       up_read(&current->mm->mmap_sem);
+       futex_unlock_mm(fshared);
        return ret;
 }
 
@@ -885,7 +994,7 @@ queue_lock(struct futex_q *q, int fd, struct file *filp)
 
        init_waitqueue_head(&q->waiters);
 
-       get_key_refs(&q->key);
+       get_futex_key_refs(&q->key);
        hb = hash_futex(&q->key);
        q->lock_ptr = &hb->lock;
 
@@ -895,7 +1004,23 @@ queue_lock(struct futex_q *q, int fd, struct file *filp)
 
 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
 {
-       list_add_tail(&q->list, &hb->chain);
+       int prio;
+
+       /*
+        * The priority used to register this element is
+        * - either the real thread-priority for the real-time threads
+        * (i.e. threads with a priority lower than MAX_RT_PRIO)
+        * - or MAX_RT_PRIO for non-RT threads.
+        * Thus, all RT-threads are woken first in priority order, and
+        * the others are woken last, in FIFO order.
+        */
+       prio = min(current->normal_prio, MAX_RT_PRIO);
+
+       plist_node_init(&q->list, prio);
+#ifdef CONFIG_DEBUG_PI_LIST
+       q->list.plist.lock = &hb->lock;
+#endif
+       plist_add(&q->list, &hb->chain);
        q->task = current;
        spin_unlock(&hb->lock);
 }
@@ -904,7 +1029,7 @@ static inline void
 queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
 {
        spin_unlock(&hb->lock);
-       drop_key_refs(&q->key);
+       drop_futex_key_refs(&q->key);
 }
 
 /*
@@ -930,7 +1055,8 @@ static int unqueue_me(struct futex_q *q)
        /* In the common case we don't take the spinlock, which is nice. */
  retry:
        lock_ptr = q->lock_ptr;
-       if (lock_ptr != 0) {
+       barrier();
+       if (lock_ptr != NULL) {
                spin_lock(lock_ptr);
                /*
                 * q->lock_ptr can change between reading it and
@@ -949,8 +1075,8 @@ static int unqueue_me(struct futex_q *q)
                        spin_unlock(lock_ptr);
                        goto retry;
                }
-               WARN_ON(list_empty(&q->list));
-               list_del(&q->list);
+               WARN_ON(plist_node_empty(&q->list));
+               plist_del(&q->list, &q->list.plist);
 
                BUG_ON(q->pi_state);
 
@@ -958,29 +1084,90 @@ static int unqueue_me(struct futex_q *q)
                ret = 1;
        }
 
-       drop_key_refs(&q->key);
+       drop_futex_key_refs(&q->key);
        return ret;
 }
 
 /*
  * PI futexes can not be requeued and must remove themself from the
- * hash bucket. The hash bucket lock is held on entry and dropped here.
+ * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
+ * and dropped here.
  */
-static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
+static void unqueue_me_pi(struct futex_q *q)
 {
-       WARN_ON(list_empty(&q->list));
-       list_del(&q->list);
+       WARN_ON(plist_node_empty(&q->list));
+       plist_del(&q->list, &q->list.plist);
 
        BUG_ON(!q->pi_state);
        free_pi_state(q->pi_state);
        q->pi_state = NULL;
 
-       spin_unlock(&hb->lock);
+       spin_unlock(q->lock_ptr);
 
-       drop_key_refs(&q->key);
+       drop_futex_key_refs(&q->key);
 }
 
-static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
+/*
+ * Fixup the pi_state owner with the new owner.
+ *
+ * Must be called with hash bucket lock held and mm->sem held for non
+ * private futexes.
+ */
+static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
+                               struct task_struct *newowner)
+{
+       u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
+       struct futex_pi_state *pi_state = q->pi_state;
+       u32 uval, curval, newval;
+       int ret;
+
+       /* Owner died? */
+       if (pi_state->owner != NULL) {
+               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);
+       } else
+               newtid |= FUTEX_OWNER_DIED;
+
+       pi_state->owner = newowner;
+
+       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);
+
+       /*
+        * We own it, so we have to replace the pending owner
+        * TID. This must be atomic as we have preserve the
+        * owner died bit here.
+        */
+       ret = get_futex_value_locked(&uval, uaddr);
+
+       while (!ret) {
+               newval = (uval & FUTEX_OWNER_DIED) | newtid;
+
+               curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
+
+               if (curval == -EFAULT)
+                       ret = -EFAULT;
+               if (curval == uval)
+                       break;
+               uval = curval;
+       }
+       return ret;
+}
+
+/*
+ * In case we must use restart_block to restart a futex_wait,
+ * we encode in the 'flags' shared capability
+ */
+#define FLAGS_SHARED  1
+
+static long futex_wait_restart(struct restart_block *restart);
+
+static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
+                     u32 val, ktime_t *abs_time, u32 bitset)
 {
        struct task_struct *curr = current;
        DECLARE_WAITQUEUE(wait, curr);
@@ -988,12 +1175,18 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
        struct futex_q q;
        u32 uval;
        int ret;
+       struct hrtimer_sleeper t;
+       int rem = 0;
+
+       if (!bitset)
+               return -EINVAL;
 
        q.pi_state = NULL;
+       q.bitset = bitset;
  retry:
-       down_read(&curr->mm->mmap_sem);
+       futex_lock_mm(fshared);
 
-       ret = get_futex_key(uaddr, &q.key);
+       ret = get_futex_key(uaddr, fshared, &q.key);
        if (unlikely(ret != 0))
                goto out_release_sem;
 
@@ -1016,8 +1209,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
         * a wakeup when *uaddr != val on entry to the syscall.  This is
         * rare, but normal.
         *
-        * We hold the mmap semaphore, so the mapping cannot have changed
-        * since we looked it up in get_futex_key.
+        * for shared futexes, we hold the mmap semaphore, so the mapping
+        * cannot have changed since we looked it up in get_futex_key.
         */
        ret = get_futex_value_locked(&uval, uaddr);
 
@@ -1028,7 +1221,7 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
                 * If we would have faulted, release mmap_sem, fault it in and
                 * start all over again.
                 */
-               up_read(&curr->mm->mmap_sem);
+               futex_unlock_mm(fshared);
 
                ret = get_user(uval, uaddr);
 
@@ -1047,7 +1240,7 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
         * Now the futex is queued and we have checked the data, we
         * don't want to hold mmap_sem while we sleep.
         */
-       up_read(&curr->mm->mmap_sem);
+       futex_unlock_mm(fshared);
 
        /*
         * There might have been scheduling since the queue_me(), as we
@@ -1062,11 +1255,35 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
        __set_current_state(TASK_INTERRUPTIBLE);
        add_wait_queue(&q.waiters, &wait);
        /*
-        * !list_empty() is safe here without any lock.
+        * !plist_node_empty() is safe here without any lock.
         * q.lock_ptr != 0 is not safe, because of ordering against wakeup.
         */
-       if (likely(!list_empty(&q.list)))
-               time = schedule_timeout(time);
+       if (likely(!plist_node_empty(&q.list))) {
+               if (!abs_time)
+                       schedule();
+               else {
+                       hrtimer_init(&t.timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
+                       hrtimer_init_sleeper(&t, current);
+                       t.timer.expires = *abs_time;
+
+                       hrtimer_start(&t.timer, t.timer.expires, HRTIMER_MODE_ABS);
+                       if (!hrtimer_active(&t.timer))
+                               t.task = NULL;
+
+                       /*
+                        * the timer could have already expired, in which
+                        * case current would be flagged for rescheduling.
+                        * Don't bother calling schedule.
+                        */
+                       if (likely(t.task))
+                               schedule();
+
+                       hrtimer_cancel(&t.timer);
+
+                       /* Flag if a timeout occured */
+                       rem = (t.task == NULL);
+               }
+       }
        __set_current_state(TASK_RUNNING);
 
        /*
@@ -1077,86 +1294,145 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
        /* If we were woken (and unqueued), we succeeded, whatever. */
        if (!unqueue_me(&q))
                return 0;
-       if (time == 0)
+       if (rem)
                return -ETIMEDOUT;
+
        /*
         * We expect signal_pending(current), but another thread may
         * have handled it for us already.
         */
-       return -EINTR;
+       if (!abs_time)
+               return -ERESTARTSYS;
+       else {
+               struct restart_block *restart;
+               restart = &current_thread_info()->restart_block;
+               restart->fn = futex_wait_restart;
+               restart->futex.uaddr = (u32 *)uaddr;
+               restart->futex.val = val;
+               restart->futex.time = abs_time->tv64;
+               restart->futex.bitset = bitset;
+               restart->futex.flags = 0;
+
+               if (fshared)
+                       restart->futex.flags |= FLAGS_SHARED;
+               return -ERESTART_RESTARTBLOCK;
+       }
 
  out_unlock_release_sem:
        queue_unlock(&q, hb);
 
  out_release_sem:
-       up_read(&curr->mm->mmap_sem);
+       futex_unlock_mm(fshared);
        return ret;
 }
 
+
+static long futex_wait_restart(struct restart_block *restart)
+{
+       u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
+       struct rw_semaphore *fshared = NULL;
+       ktime_t t;
+
+       t.tv64 = restart->futex.time;
+       restart->fn = do_no_restart_syscall;
+       if (restart->futex.flags & FLAGS_SHARED)
+               fshared = &current->mm->mmap_sem;
+       return (long)futex_wait(uaddr, fshared, restart->futex.val, &t,
+                               restart->futex.bitset);
+}
+
+
 /*
  * Userspace tried a 0 -> TID atomic transition of the futex value
  * and failed. The kernel side here does the whole locking operation:
  * if there are waiters then it will block, it does PI, etc. (Due to
  * races the kernel might see a 0 value of the futex too.)
  */
-static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
-                           struct hrtimer_sleeper *to)
+static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
+                        int detect, ktime_t *time, int trylock)
 {
+       struct hrtimer_sleeper timeout, *to = NULL;
        struct task_struct *curr = current;
        struct futex_hash_bucket *hb;
        u32 uval, newval, curval;
        struct futex_q q;
-       int ret, attempt = 0;
+       int ret, lock_taken, ownerdied = 0, attempt = 0;
 
        if (refill_pi_state_cache())
                return -ENOMEM;
 
+       if (time) {
+               to = &timeout;
+               hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
+               hrtimer_init_sleeper(to, current);
+               to->timer.expires = *time;
+       }
+
        q.pi_state = NULL;
  retry:
-       down_read(&curr->mm->mmap_sem);
+       futex_lock_mm(fshared);
 
-       ret = get_futex_key(uaddr, &q.key);
+       ret = get_futex_key(uaddr, fshared, &q.key);
        if (unlikely(ret != 0))
                goto out_release_sem;
 
+ retry_unlocked:
        hb = queue_lock(&q, -1, NULL);
 
  retry_locked:
+       ret = lock_taken = 0;
+
        /*
         * To avoid races, we attempt to take the lock here again
         * (by doing a 0 -> TID atomic cmpxchg), while holding all
         * the locks. It will most likely not succeed.
         */
-       newval = current->pid;
+       newval = task_pid_vnr(current);
 
-       inc_preempt_count();
-       curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
-       dec_preempt_count();
+       curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
 
        if (unlikely(curval == -EFAULT))
                goto uaddr_faulted;
 
-       /* We own the lock already */
-       if (unlikely((curval & FUTEX_TID_MASK) == current->pid)) {
-               if (!detect && 0)
-                       force_sig(SIGKILL, current);
+       /*
+        * Detect deadlocks. In case of REQUEUE_PI this is a valid
+        * situation and we return success to user space.
+        */
+       if (unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(current))) {
                ret = -EDEADLK;
                goto out_unlock_release_sem;
        }
 
        /*
-        * Surprise - we got the lock. Just return
-        * to userspace:
+        * Surprise - we got the lock. Just return to userspace:
         */
        if (unlikely(!curval))
                goto out_unlock_release_sem;
 
        uval = curval;
-       newval = uval | FUTEX_WAITERS;
 
-       inc_preempt_count();
-       curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
-       dec_preempt_count();
+       /*
+        * Set the WAITERS flag, so the owner will know it has someone
+        * to wake at next unlock
+        */
+       newval = curval | FUTEX_WAITERS;
+
+       /*
+        * There are two cases, where a futex might have no owner (the
+        * owner TID is 0): OWNER_DIED. We take over the futex in this
+        * case. We also do an unconditional take over, when the owner
+        * of the futex died.
+        *
+        * This is safe as we are protected by the hash bucket lock !
+        */
+       if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
+               /* Keep the OWNER_DIED bit */
+               newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(current);
+               ownerdied = 0;
+               lock_taken = 1;
+       }
+
+       curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
 
        if (unlikely(curval == -EFAULT))
                goto uaddr_faulted;
@@ -1164,40 +1440,51 @@ static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
                goto retry_locked;
 
        /*
+        * We took the lock due to owner died take over.
+        */
+       if (unlikely(lock_taken))
+               goto out_unlock_release_sem;
+
+       /*
         * We dont have the lock. Look up the PI state (or create it if
         * we are the first waiter):
         */
-       ret = lookup_pi_state(uval, hb, &q);
+       ret = lookup_pi_state(uval, hb, &q.key, &q.pi_state);
 
        if (unlikely(ret)) {
-               /*
-                * There were no waiters and the owner task lookup
-                * failed. When the OWNER_DIED bit is set, then we
-                * know that this is a robust futex and we actually
-                * take the lock. This is safe as we are protected by
-                * the hash bucket lock. We also set the waiters bit
-                * unconditionally here, to simplify glibc handling of
-                * multiple tasks racing to acquire the lock and
-                * cleanup the problems which were left by the dead
-                * owner.
-                */
-               if (curval & FUTEX_OWNER_DIED) {
-                       uval = newval;
-                       newval = current->pid |
-                               FUTEX_OWNER_DIED | FUTEX_WAITERS;
+               switch (ret) {
 
-                       inc_preempt_count();
-                       curval = futex_atomic_cmpxchg_inatomic(uaddr,
-                                                              uval, newval);
-                       dec_preempt_count();
+               case -EAGAIN:
+                       /*
+                        * Task is exiting and we just wait for the
+                        * exit to complete.
+                        */
+                       queue_unlock(&q, hb);
+                       futex_unlock_mm(fshared);
+                       cond_resched();
+                       goto retry;
 
-                       if (unlikely(curval == -EFAULT))
+               case -ESRCH:
+                       /*
+                        * No owner found for this futex. Check if the
+                        * OWNER_DIED bit is set to figure out whether
+                        * this is a robust futex or not.
+                        */
+                       if (get_futex_value_locked(&curval, uaddr))
                                goto uaddr_faulted;
-                       if (unlikely(curval != uval))
+
+                       /*
+                        * We simply start over in case of a robust
+                        * futex. The code above will take the futex
+                        * and return happy.
+                        */
+                       if (curval & FUTEX_OWNER_DIED) {
+                               ownerdied = 1;
                                goto retry_locked;
-                       ret = 0;
+                       }
+               default:
+                       goto out_unlock_release_sem;
                }
-               goto out_unlock_release_sem;
        }
 
        /*
@@ -1209,7 +1496,7 @@ static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
         * Now the futex is queued and we have checked the data, we
         * don't want to hold mmap_sem while we sleep.
         */
-       up_read(&curr->mm->mmap_sem);
+       futex_unlock_mm(fshared);
 
        WARN_ON(!q.pi_state);
        /*
@@ -1223,74 +1510,80 @@ static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
                ret = ret ? 0 : -EWOULDBLOCK;
        }
 
-       down_read(&curr->mm->mmap_sem);
+       futex_lock_mm(fshared);
        spin_lock(q.lock_ptr);
 
-       /*
-        * Got the lock. We might not be the anticipated owner if we
-        * did a lock-steal - fix up the PI-state in that case.
-        */
-       if (!ret && q.pi_state->owner != curr) {
-               u32 newtid = current->pid | FUTEX_WAITERS;
-
-               /* Owner died? */
-               if (q.pi_state->owner != NULL) {
-                       spin_lock_irq(&q.pi_state->owner->pi_lock);
-                       list_del_init(&q.pi_state->list);
-                       spin_unlock_irq(&q.pi_state->owner->pi_lock);
-               } else
-                       newtid |= FUTEX_OWNER_DIED;
-
-               q.pi_state->owner = current;
-
-               spin_lock_irq(&current->pi_lock);
-               list_add(&q.pi_state->list, &current->pi_state_list);
-               spin_unlock_irq(&current->pi_lock);
-
-               /* Unqueue and drop the lock */
-               unqueue_me_pi(&q, hb);
-               up_read(&curr->mm->mmap_sem);
+       if (!ret) {
                /*
-                * We own it, so we have to replace the pending owner
-                * TID. This must be atomic as we have preserve the
-                * owner died bit here.
+                * Got the lock. We might not be the anticipated owner
+                * if we did a lock-steal - fix up the PI-state in
+                * that case:
                 */
-               ret = get_user(uval, uaddr);
-               while (!ret) {
-                       newval = (uval & FUTEX_OWNER_DIED) | newtid;
-                       curval = futex_atomic_cmpxchg_inatomic(uaddr,
-                                                              uval, newval);
-                       if (curval == -EFAULT)
-                               ret = -EFAULT;
-                       if (curval == uval)
-                               break;
-                       uval = curval;
-               }
+               if (q.pi_state->owner != curr)
+                       ret = fixup_pi_state_owner(uaddr, &q, curr);
        } else {
                /*
                 * Catch the rare case, where the lock was released
-                * when we were on the way back before we locked
-                * the hash bucket.
+                * when we were on the way back before we locked the
+                * hash bucket.
                 */
-               if (ret && q.pi_state->owner == curr) {
+               if (q.pi_state->owner == curr) {
+                       /*
+                        * Try to get the rt_mutex now. This might
+                        * fail as some other task acquired the
+                        * rt_mutex after we removed ourself from the
+                        * rt_mutex waiters list.
+                        */
                        if (rt_mutex_trylock(&q.pi_state->pi_mutex))
                                ret = 0;
+                       else {
+                               /*
+                                * pi_state is incorrect, some other
+                                * task did a lock steal and we
+                                * returned due to timeout or signal
+                                * without taking the rt_mutex. Too
+                                * late. We can access the
+                                * rt_mutex_owner without locking, as
+                                * the other task is now blocked on
+                                * the hash bucket lock. Fix the state
+                                * up.
+                                */
+                               struct task_struct *owner;
+                               int res;
+
+                               owner = rt_mutex_owner(&q.pi_state->pi_mutex);
+                               res = fixup_pi_state_owner(uaddr, &q, owner);
+
+                               /* propagate -EFAULT, if the fixup failed */
+                               if (res)
+                                       ret = res;
+                       }
+               } else {
+                       /*
+                        * Paranoia check. If we did not take the lock
+                        * in the trylock above, then we should not be
+                        * the owner of the rtmutex, neither the real
+                        * nor the pending one:
+                        */
+                       if (rt_mutex_owner(&q.pi_state->pi_mutex) == curr)
+                               printk(KERN_ERR "futex_lock_pi: ret = %d "
+                                      "pi-mutex: %p pi-state %p\n", ret,
+                                      q.pi_state->pi_mutex.owner,
+                                      q.pi_state->owner);
                }
-               /* Unqueue and drop the lock */
-               unqueue_me_pi(&q, hb);
-               up_read(&curr->mm->mmap_sem);
        }
 
-       if (!detect && ret == -EDEADLK && 0)
-               force_sig(SIGKILL, current);
+       /* Unqueue and drop the lock */
+       unqueue_me_pi(&q);
+       futex_unlock_mm(fshared);
 
-       return ret;
+       return ret != -EINTR ? ret : -ERESTARTNOINTR;
 
  out_unlock_release_sem:
        queue_unlock(&q, hb);
 
  out_release_sem:
-       up_read(&curr->mm->mmap_sem);
+       futex_unlock_mm(fshared);
        return ret;
 
  uaddr_faulted:
@@ -1299,16 +1592,20 @@ static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
         * non-atomically.  Therefore, if get_user below is not
         * enough, we need to handle the fault ourselves, while
         * still holding the mmap_sem.
+        *
+        * ... and hb->lock. :-) --ANK
         */
-       if (attempt++) {
-               if (futex_handle_fault((unsigned long)uaddr, attempt))
-                       goto out_unlock_release_sem;
+       queue_unlock(&q, hb);
 
-               goto retry_locked;
+       if (attempt++) {
+               ret = futex_handle_fault((unsigned long)uaddr, fshared,
+                                        attempt);
+               if (ret)
+                       goto out_release_sem;
+               goto retry_unlocked;
        }
 
-       queue_unlock(&q, hb);
-       up_read(&curr->mm->mmap_sem);
+       futex_unlock_mm(fshared);
 
        ret = get_user(uval, uaddr);
        if (!ret && (uval != -EFAULT))
@@ -1318,86 +1615,16 @@ static int do_futex_lock_pi(u32 __user *uaddr, int detect, int trylock,
 }
 
 /*
- * Restart handler
- */
-static long futex_lock_pi_restart(struct restart_block *restart)
-{
-       struct hrtimer_sleeper timeout, *to = NULL;
-       int ret;
-
-       restart->fn = do_no_restart_syscall;
-
-       if (restart->arg2 || restart->arg3) {
-               to = &timeout;
-               hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
-               hrtimer_init_sleeper(to, current);
-               to->timer.expires.tv64 = ((u64)restart->arg1 << 32) |
-                       (u64) restart->arg0;
-       }
-
-       pr_debug("lock_pi restart: %p, %d (%d)\n",
-                (u32 __user *)restart->arg0, current->pid);
-
-       ret = do_futex_lock_pi((u32 __user *)restart->arg0, restart->arg1,
-                              0, to);
-
-       if (ret != -EINTR)
-               return ret;
-
-       restart->fn = futex_lock_pi_restart;
-
-       /* The other values are filled in */
-       return -ERESTART_RESTARTBLOCK;
-}
-
-/*
- * Called from the syscall entry below.
- */
-static int futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
-                        long nsec, int trylock)
-{
-       struct hrtimer_sleeper timeout, *to = NULL;
-       struct restart_block *restart;
-       int ret;
-
-       if (sec != MAX_SCHEDULE_TIMEOUT) {
-               to = &timeout;
-               hrtimer_init(&to->timer, CLOCK_REALTIME, HRTIMER_ABS);
-               hrtimer_init_sleeper(to, current);
-               to->timer.expires = ktime_set(sec, nsec);
-       }
-
-       ret = do_futex_lock_pi(uaddr, detect, trylock, to);
-
-       if (ret != -EINTR)
-               return ret;
-
-       pr_debug("lock_pi interrupted: %p, %d (%d)\n", uaddr, current->pid);
-
-       restart = &current_thread_info()->restart_block;
-       restart->fn = futex_lock_pi_restart;
-       restart->arg0 = (unsigned long) uaddr;
-       restart->arg1 = detect;
-       if (to) {
-               restart->arg2 = to->timer.expires.tv64 & 0xFFFFFFFF;
-               restart->arg3 = to->timer.expires.tv64 >> 32;
-       } else
-               restart->arg2 = restart->arg3 = 0;
-
-       return -ERESTART_RESTARTBLOCK;
-}
-
-/*
  * Userspace attempted a TID -> 0 atomic transition, and failed.
  * This is the in-kernel slowpath: we look up the PI state (if any),
  * and do the rt-mutex unlock.
  */
-static int futex_unlock_pi(u32 __user *uaddr)
+static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
 {
        struct futex_hash_bucket *hb;
        struct futex_q *this, *next;
        u32 uval;
-       struct list_head *head;
+       struct plist_head *head;
        union futex_key key;
        int ret, attempt = 0;
 
@@ -1407,29 +1634,29 @@ retry:
        /*
         * We release only a lock we actually own:
         */
-       if ((uval & FUTEX_TID_MASK) != current->pid)
+       if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
                return -EPERM;
        /*
         * First take all the futex related locks:
         */
-       down_read(&current->mm->mmap_sem);
+       futex_lock_mm(fshared);
 
-       ret = get_futex_key(uaddr, &key);
+       ret = get_futex_key(uaddr, fshared, &key);
        if (unlikely(ret != 0))
                goto out;
 
        hb = hash_futex(&key);
+retry_unlocked:
        spin_lock(&hb->lock);
 
-retry_locked:
        /*
         * To avoid races, try to do the TID -> 0 atomic transition
         * again. If it succeeds then we can return without waking
         * anyone else up:
         */
-       inc_preempt_count();
-       uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
-       dec_preempt_count();
+       if (!(uval & FUTEX_OWNER_DIED))
+               uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
+
 
        if (unlikely(uval == -EFAULT))
                goto pi_faulted;
@@ -1437,7 +1664,7 @@ retry_locked:
         * Rare case: we managed to release the lock atomically,
         * no need to wake anyone else up:
         */
-       if (unlikely(uval == current->pid))
+       if (unlikely(uval == task_pid_vnr(current)))
                goto out_unlock;
 
        /*
@@ -1446,7 +1673,7 @@ retry_locked:
         */
        head = &hb->chain;
 
-       list_for_each_entry_safe(this, next, head, list) {
+       plist_for_each_entry_safe(this, next, head, list) {
                if (!match_futex (&this->key, &key))
                        continue;
                ret = wake_futex_pi(uaddr, uval, this);
@@ -1462,14 +1689,16 @@ retry_locked:
        /*
         * No waiters - kernel unlocks the futex:
         */
-       ret = unlock_futex_pi(uaddr, uval);
-       if (ret == -EFAULT)
-               goto pi_faulted;
+       if (!(uval & FUTEX_OWNER_DIED)) {
+               ret = unlock_futex_pi(uaddr, uval);
+               if (ret == -EFAULT)
+                       goto pi_faulted;
+       }
 
 out_unlock:
        spin_unlock(&hb->lock);
 out:
-       up_read(&current->mm->mmap_sem);
+       futex_unlock_mm(fshared);
 
        return ret;
 
@@ -1479,16 +1708,21 @@ pi_faulted:
         * non-atomically.  Therefore, if get_user below is not
         * enough, we need to handle the fault ourselves, while
         * still holding the mmap_sem.
+        *
+        * ... and hb->lock. --ANK
         */
-       if (attempt++) {
-               if (futex_handle_fault((unsigned long)uaddr, attempt))
-                       goto out_unlock;
+       spin_unlock(&hb->lock);
 
-               goto retry_locked;
+       if (attempt++) {
+               ret = futex_handle_fault((unsigned long)uaddr, fshared,
+                                        attempt);
+               if (ret)
+                       goto out;
+               uval = 0;
+               goto retry_unlocked;
        }
 
-       spin_unlock(&hb->lock);
-       up_read(&current->mm->mmap_sem);
+       futex_unlock_mm(fshared);
 
        ret = get_user(uval, uaddr);
        if (!ret && (uval != -EFAULT))
@@ -1517,16 +1751,16 @@ static unsigned int futex_poll(struct file *filp,
        poll_wait(filp, &q->waiters, wait);
 
        /*
-        * list_empty() is safe here without any lock.
+        * plist_node_empty() is safe here without any lock.
         * q->lock_ptr != 0 is not safe, because of ordering against wakeup.
         */
-       if (list_empty(&q->list))
+       if (plist_node_empty(&q->list))
                ret = POLLIN | POLLRDNORM;
 
        return ret;
 }
 
-static struct file_operations futex_fops = {
+static const struct file_operations futex_fops = {
        .release        = futex_close,
        .poll           = futex_poll,
 };
@@ -1540,6 +1774,14 @@ static int futex_fd(u32 __user *uaddr, int signal)
        struct futex_q *q;
        struct file *filp;
        int ret, err;
+       struct rw_semaphore *fshared;
+       static unsigned long printk_interval;
+
+       if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
+               printk(KERN_WARNING "Process `%s' used FUTEX_FD, which "
+                      "will be removed from the kernel in June 2007\n",
+                      current->comm);
+       }
 
        ret = -EINVAL;
        if (!valid_signal(signal))
@@ -1555,12 +1797,12 @@ static int futex_fd(u32 __user *uaddr, int signal)
                goto out;
        }
        filp->f_op = &futex_fops;
-       filp->f_vfsmnt = mntget(futex_mnt);
-       filp->f_dentry = dget(futex_mnt->mnt_root);
-       filp->f_mapping = filp->f_dentry->d_inode->i_mapping;
+       filp->f_path.mnt = mntget(futex_mnt);
+       filp->f_path.dentry = dget(futex_mnt->mnt_root);
+       filp->f_mapping = filp->f_path.dentry->d_inode->i_mapping;
 
        if (signal) {
-               err = f_setown(filp, current->pid, 1);
+               err = __f_setown(filp, task_pid(current), PIDTYPE_PID, 1);
                if (err < 0) {
                        goto error;
                }
@@ -1574,11 +1816,12 @@ static int futex_fd(u32 __user *uaddr, int signal)
        }
        q->pi_state = NULL;
 
-       down_read(&current->mm->mmap_sem);
-       err = get_futex_key(uaddr, &q->key);
+       fshared = &current->mm->mmap_sem;
+       down_read(fshared);
+       err = get_futex_key(uaddr, fshared, &q->key);
 
        if (unlikely(err != 0)) {
-               up_read(&current->mm->mmap_sem);
+               up_read(fshared);
                kfree(q);
                goto error;
        }
@@ -1590,7 +1833,7 @@ static int futex_fd(u32 __user *uaddr, int signal)
        filp->private_data = q;
 
        queue_me(q, ret, filp);
-       up_read(&current->mm->mmap_sem);
+       up_read(fshared);
 
        /* Now we map fd to filp, so userspace can access it */
        fd_install(ret, filp);
@@ -1645,10 +1888,10 @@ sys_set_robust_list(struct robust_list_head __user *head,
  * @len_ptr: pointer to a length field, the kernel fills in the header size
  */
 asmlinkage long
-sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
+sys_get_robust_list(int pid, struct robust_list_head __user * __user *head_ptr,
                    size_t __user *len_ptr)
 {
-       struct robust_list_head *head;
+       struct robust_list_head __user *head;
        unsigned long ret;
 
        if (!pid)
@@ -1657,8 +1900,8 @@ sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
                struct task_struct *p;
 
                ret = -ESRCH;
-               read_lock(&tasklist_lock);
-               p = find_task_by_pid(pid);
+               rcu_read_lock();
+               p = find_task_by_vpid(pid);
                if (!p)
                        goto err_unlock;
                ret = -EPERM;
@@ -1666,7 +1909,7 @@ sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
                                !capable(CAP_SYS_PTRACE))
                        goto err_unlock;
                head = p->robust_list;
-               read_unlock(&tasklist_lock);
+               rcu_read_unlock();
        }
 
        if (put_user(sizeof(*head), len_ptr))
@@ -1674,7 +1917,7 @@ sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
        return put_user(head, head_ptr);
 
 err_unlock:
-       read_unlock(&tasklist_lock);
+       rcu_read_unlock();
 
        return ret;
 }
@@ -1683,15 +1926,15 @@ err_unlock:
  * Process a futex-list entry, check whether it's owned by the
  * dying task, and do notification if so:
  */
-int handle_futex_death(u32 __user *uaddr, struct task_struct *curr)
+int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
 {
-       u32 uval, nval;
+       u32 uval, nval, mval;
 
 retry:
        if (get_user(uval, uaddr))
                return -1;
 
-       if ((uval & FUTEX_TID_MASK) == curr->pid) {
+       if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
                /*
                 * Ok, this dying thread is truly holding a futex
                 * of interest. Set the OWNER_DIED bit atomically
@@ -1702,21 +1945,45 @@ retry:
                 * thread-death.) The rest of the cleanup is done in
                 * userspace.
                 */
-               nval = futex_atomic_cmpxchg_inatomic(uaddr, uval,
-                                                    uval | FUTEX_OWNER_DIED);
+               mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
+               nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
+
                if (nval == -EFAULT)
                        return -1;
 
                if (nval != uval)
                        goto retry;
 
-               if (uval & FUTEX_WAITERS)
-                       futex_wake(uaddr, 1);
+               /*
+                * Wake robust non-PI futexes here. The wakeup of
+                * PI futexes happens in exit_pi_state():
+                */
+               if (!pi && (uval & FUTEX_WAITERS))
+                       futex_wake(uaddr, &curr->mm->mmap_sem, 1,
+                                  FUTEX_BITSET_MATCH_ANY);
        }
        return 0;
 }
 
 /*
+ * Fetch a robust-list pointer. Bit 0 signals PI futexes:
+ */
+static inline int fetch_robust_entry(struct robust_list __user **entry,
+                                    struct robust_list __user * __user *head,
+                                    int *pi)
+{
+       unsigned long uentry;
+
+       if (get_user(uentry, (unsigned long __user *)head))
+               return -EFAULT;
+
+       *entry = (void __user *)(uentry & ~1UL);
+       *pi = uentry & 1;
+
+       return 0;
+}
+
+/*
  * Walk curr->robust_list (very carefully, it's a userspace list!)
  * and mark any locks found there dead, and notify any waiters.
  *
@@ -1725,15 +1992,16 @@ retry:
 void exit_robust_list(struct task_struct *curr)
 {
        struct robust_list_head __user *head = curr->robust_list;
-       struct robust_list __user *entry, *pending;
-       unsigned int limit = ROBUST_LIST_LIMIT;
+       struct robust_list __user *entry, *next_entry, *pending;
+       unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
        unsigned long futex_offset;
+       int rc;
 
        /*
         * Fetch the list head (which was registered earlier, via
         * sys_set_robust_list()):
         */
-       if (get_user(entry, &head->list.next))
+       if (fetch_robust_entry(&entry, &head->list.next, &pi))
                return;
        /*
         * Fetch the relative futex offset:
@@ -1744,25 +2012,28 @@ void exit_robust_list(struct task_struct *curr)
         * Fetch any possibly pending lock-add first, and handle it
         * if it exists:
         */
-       if (get_user(pending, &head->list_op_pending))
+       if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
                return;
-       if (pending)
-               handle_futex_death((void *)pending + futex_offset, curr);
 
+       next_entry = NULL;      /* avoid warning with gcc */
        while (entry != &head->list) {
                /*
+                * Fetch the next entry in the list before calling
+                * handle_futex_death:
+                */
+               rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
+               /*
                 * A pending lock might already be on the list, so
                 * don't process it twice:
                 */
                if (entry != pending)
-                       if (handle_futex_death((void *)entry + futex_offset,
-                                               curr))
+                       if (handle_futex_death((void __user *)entry + futex_offset,
+                                               curr, pi))
                                return;
-               /*
-                * Fetch the next entry in the list:
-                */
-               if (get_user(entry, &entry->next))
+               if (rc)
                        return;
+               entry = next_entry;
+               pi = next_pi;
                /*
                 * Avoid excessively long or circular lists:
                 */
@@ -1771,41 +2042,54 @@ void exit_robust_list(struct task_struct *curr)
 
                cond_resched();
        }
+
+       if (pending)
+               handle_futex_death((void __user *)pending + futex_offset,
+                                  curr, pip);
 }
 
-long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
+long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
                u32 __user *uaddr2, u32 val2, u32 val3)
 {
        int ret;
+       int cmd = op & FUTEX_CMD_MASK;
+       struct rw_semaphore *fshared = NULL;
 
-       switch (op) {
+       if (!(op & FUTEX_PRIVATE_FLAG))
+               fshared = &current->mm->mmap_sem;
+
+       switch (cmd) {
        case FUTEX_WAIT:
-               ret = futex_wait(uaddr, val, timeout);
+               val3 = FUTEX_BITSET_MATCH_ANY;
+       case FUTEX_WAIT_BITSET:
+               ret = futex_wait(uaddr, fshared, val, timeout, val3);
                break;
        case FUTEX_WAKE:
-               ret = futex_wake(uaddr, val);
+               val3 = FUTEX_BITSET_MATCH_ANY;
+       case FUTEX_WAKE_BITSET:
+               ret = futex_wake(uaddr, fshared, val, val3);
                break;
        case FUTEX_FD:
                /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
                ret = futex_fd(uaddr, val);
                break;
        case FUTEX_REQUEUE:
-               ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
+               ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
                break;
        case FUTEX_CMP_REQUEUE:
-               ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
+               ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
                break;
        case FUTEX_WAKE_OP:
-               ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
+               ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
                break;
        case FUTEX_LOCK_PI:
-               ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
+               ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
                break;
        case FUTEX_UNLOCK_PI:
-               ret = futex_unlock_pi(uaddr);
+               ret = futex_unlock_pi(uaddr, fshared);
                break;
        case FUTEX_TRYLOCK_PI:
-               ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
+               ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
                break;
        default:
                ret = -ENOSYS;
@@ -1818,36 +2102,39 @@ asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
                          struct timespec __user *utime, u32 __user *uaddr2,
                          u32 val3)
 {
-       struct timespec t;
-       unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
+       struct timespec ts;
+       ktime_t t, *tp = NULL;
        u32 val2 = 0;
+       int cmd = op & FUTEX_CMD_MASK;
 
-       if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
-               if (copy_from_user(&t, utime, sizeof(t)) != 0)
+       if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
+                     cmd == FUTEX_WAIT_BITSET)) {
+               if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
                        return -EFAULT;
-               if (!timespec_valid(&t))
+               if (!timespec_valid(&ts))
                        return -EINVAL;
-               if (op == FUTEX_WAIT)
-                       timeout = timespec_to_jiffies(&t) + 1;
-               else {
-                       timeout = t.tv_sec;
-                       val2 = t.tv_nsec;
-               }
+
+               t = timespec_to_ktime(ts);
+               if (cmd == FUTEX_WAIT)
+                       t = ktime_add(ktime_get(), t);
+               tp = &t;
        }
        /*
-        * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
+        * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
+        * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
         */
-       if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
+       if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
+           cmd == FUTEX_WAKE_OP)
                val2 = (u32) (unsigned long) utime;
 
-       return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);
+       return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
 }
 
 static int futexfs_get_sb(struct file_system_type *fs_type,
                          int flags, const char *dev_name, void *data,
                          struct vfsmount *mnt)
 {
-       return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt);
+       return get_sb_pseudo(fs_type, "futex", NULL, FUTEXFS_SUPER_MAGIC, mnt);
 }
 
 static struct file_system_type futex_fs_type = {
@@ -1858,13 +2145,19 @@ static struct file_system_type futex_fs_type = {
 
 static int __init init(void)
 {
-       unsigned int i;
+       int i = register_filesystem(&futex_fs_type);
+
+       if (i)
+               return i;
 
-       register_filesystem(&futex_fs_type);
        futex_mnt = kern_mount(&futex_fs_type);
+       if (IS_ERR(futex_mnt)) {
+               unregister_filesystem(&futex_fs_type);
+               return PTR_ERR(futex_mnt);
+       }
 
        for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
-               INIT_LIST_HEAD(&futex_queues[i].chain);
+               plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
                spin_lock_init(&futex_queues[i].lock);
        }
        return 0;