[PATCH] Flush MMIO writes in reset sequence
[linux-2.6.git] / kernel / futex.c
index 50356fb..9d260e8 100644 (file)
  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
  *
+ *  PI-futex support started by Ingo Molnar and Thomas Gleixner
+ *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
+ *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.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.
@@ -46,6 +50,8 @@
 #include <linux/signal.h>
 #include <asm/futex.h>
 
+#include "rtmutex_common.h"
+
 #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
 
 /*
@@ -75,6 +81,27 @@ union futex_key {
 };
 
 /*
+ * Priority Inheritance state:
+ */
+struct futex_pi_state {
+       /*
+        * list of 'owned' pi_state instances - these have to be
+        * cleaned up in do_exit() if the task exits prematurely:
+        */
+       struct list_head list;
+
+       /*
+        * The PI object:
+        */
+       struct rt_mutex pi_mutex;
+
+       struct task_struct *owner;
+       atomic_t refcount;
+
+       union futex_key key;
+};
+
+/*
  * 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).
  *
@@ -96,6 +123,10 @@ struct futex_q {
        /* For fd, sigio sent using these: */
        int fd;
        struct file *filp;
+
+       /* Optional priority inheritance state: */
+       struct futex_pi_state *pi_state;
+       struct task_struct *task;
 };
 
 /*
@@ -259,6 +290,247 @@ static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
 }
 
 /*
+ * Fault handling. Called with current->mm->mmap_sem held.
+ */
+static int futex_handle_fault(unsigned long address, int attempt)
+{
+       struct vm_area_struct * vma;
+       struct mm_struct *mm = current->mm;
+
+       if (attempt > 2 || !(vma = find_vma(mm, address)) ||
+           vma->vm_start > address || !(vma->vm_flags & VM_WRITE))
+               return -EFAULT;
+
+       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;
+       }
+       return 0;
+}
+
+/*
+ * PI code:
+ */
+static int refill_pi_state_cache(void)
+{
+       struct futex_pi_state *pi_state;
+
+       if (likely(current->pi_state_cache))
+               return 0;
+
+       pi_state = kmalloc(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;
+       atomic_set(&pi_state->refcount, 1);
+
+       current->pi_state_cache = pi_state;
+
+       return 0;
+}
+
+static struct futex_pi_state * alloc_pi_state(void)
+{
+       struct futex_pi_state *pi_state = current->pi_state_cache;
+
+       WARN_ON(!pi_state);
+       current->pi_state_cache = NULL;
+
+       return pi_state;
+}
+
+static void free_pi_state(struct futex_pi_state *pi_state)
+{
+       if (!atomic_dec_and_test(&pi_state->refcount))
+               return;
+
+       /*
+        * If pi_state->owner is NULL, the owner is most probably dying
+        * and has cleaned up the pi_state already
+        */
+       if (pi_state->owner) {
+               spin_lock_irq(&pi_state->owner->pi_lock);
+               list_del_init(&pi_state->list);
+               spin_unlock_irq(&pi_state->owner->pi_lock);
+
+               rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
+       }
+
+       if (current->pi_state_cache)
+               kfree(pi_state);
+       else {
+               /*
+                * pi_state->list is already empty.
+                * clear pi_state->owner.
+                * refcount is at 0 - put it back to 1.
+                */
+               pi_state->owner = NULL;
+               atomic_set(&pi_state->refcount, 1);
+               current->pi_state_cache = pi_state;
+       }
+}
+
+/*
+ * Look up the task based on what TID userspace gave us.
+ * We dont trust it.
+ */
+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->exit_state != 0) {
+               p = NULL;
+               goto out_unlock;
+       }
+       get_task_struct(p);
+out_unlock:
+       read_unlock(&tasklist_lock);
+
+       return p;
+}
+
+/*
+ * This task is holding PI mutexes at exit time => bad.
+ * Kernel cleans up PI-state, but userspace is likely hosed.
+ * (Robust-futex cleanup is separate and might save the day for userspace.)
+ */
+void exit_pi_state_list(struct task_struct *curr)
+{
+       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 themselves:
+        */
+       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);
+
+               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;
+               }
+
+               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);
+
+               rt_mutex_unlock(&pi_state->pi_mutex);
+
+               spin_unlock(&hb->lock);
+
+               spin_lock_irq(&curr->pi_lock);
+       }
+       spin_unlock_irq(&curr->pi_lock);
+}
+
+static int
+lookup_pi_state(u32 uval, struct futex_hash_bucket *hb, struct futex_q *me)
+{
+       struct futex_pi_state *pi_state = NULL;
+       struct futex_q *this, *next;
+       struct list_head *head;
+       struct task_struct *p;
+       pid_t pid;
+
+       head = &hb->chain;
+
+       list_for_each_entry_safe(this, next, head, list) {
+               if (match_futex(&this->key, &me->key)) {
+                       /*
+                        * Another waiter already exists - bump up
+                        * the refcount and return its pi_state:
+                        */
+                       pi_state = this->pi_state;
+                       /*
+                        * Userspace might have messed up non PI and PI futexes
+                        */
+                       if (unlikely(!pi_state))
+                               return -EINVAL;
+
+                       WARN_ON(!atomic_read(&pi_state->refcount));
+
+                       atomic_inc(&pi_state->refcount);
+                       me->pi_state = pi_state;
+
+                       return 0;
+               }
+       }
+
+       /*
+        * We are the first waiter - try to look up the real owner and attach
+        * the new pi_state to it, but bail out when the owner died bit is set
+        * and TID = 0:
+        */
+       pid = uval & FUTEX_TID_MASK;
+       if (!pid && (uval & FUTEX_OWNER_DIED))
+               return -ESRCH;
+       p = futex_find_get_task(pid);
+       if (!p)
+               return -ESRCH;
+
+       pi_state = alloc_pi_state();
+
+       /*
+        * Initialize the pi_mutex in locked state and make 'p'
+        * the owner of it:
+        */
+       rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
+
+       /* Store the key for possible exit cleanups: */
+       pi_state->key = me->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;
+
+       return 0;
+}
+
+/*
  * The hash bucket lock must be held when this is called.
  * Afterwards, the futex_q must not be accessed.
  */
@@ -285,6 +557,95 @@ static void wake_futex(struct futex_q *q)
        q->lock_ptr = NULL;
 }
 
+static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
+{
+       struct task_struct *new_owner;
+       struct futex_pi_state *pi_state = this->pi_state;
+       u32 curval, newval;
+
+       if (!pi_state)
+               return -EINVAL;
+
+       new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
+
+       /*
+        * This happens when we have stolen the lock and the original
+        * pending owner did not enqueue itself back on the rt_mutex.
+        * Thats not a tragedy. We know that way, that a lock waiter
+        * is on the fly. We make the futex_q waiter the pending owner.
+        */
+       if (!new_owner)
+               new_owner = this->task;
+
+       /*
+        * We pass it to the next owner. (The WAITERS bit is always
+        * kept enabled while there is PI state around. We must also
+        * preserve the owner died bit.)
+        */
+       if (!(uval & FUTEX_OWNER_DIED)) {
+               newval = FUTEX_WAITERS | new_owner->pid;
+
+               inc_preempt_count();
+               curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
+               dec_preempt_count();
+               if (curval == -EFAULT)
+                       return -EFAULT;
+               if (curval != uval)
+                       return -EINVAL;
+       }
+
+       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);
+
+       rt_mutex_unlock(&pi_state->pi_mutex);
+
+       return 0;
+}
+
+static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
+{
+       u32 oldval;
+
+       /*
+        * 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();
+
+       if (oldval == -EFAULT)
+               return oldval;
+       if (oldval != uval)
+               return -EAGAIN;
+
+       return 0;
+}
+
+/*
+ * Express the locking dependencies for lockdep:
+ */
+static inline void
+double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
+{
+       if (hb1 <= hb2) {
+               spin_lock(&hb1->lock);
+               if (hb1 < hb2)
+                       spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
+       } else { /* hb1 > hb2 */
+               spin_lock(&hb2->lock);
+               spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
+       }
+}
+
 /*
  * Wake up all waiters hashed on the physical page that is mapped
  * to this virtual address:
@@ -309,6 +670,10 @@ static int futex_wake(u32 __user *uaddr, int nr_wake)
 
        list_for_each_entry_safe(this, next, head, list) {
                if (match_futex (&this->key, &key)) {
+                       if (this->pi_state) {
+                               ret = -EINVAL;
+                               break;
+                       }
                        wake_futex(this);
                        if (++ret >= nr_wake)
                                break;
@@ -349,11 +714,7 @@ retryfull:
        hb2 = hash_futex(&key2);
 
 retry:
-       if (hb1 < hb2)
-               spin_lock(&hb1->lock);
-       spin_lock(&hb2->lock);
-       if (hb1 > hb2)
-               spin_lock(&hb1->lock);
+       double_lock_hb(hb1, hb2);
 
        op_ret = futex_atomic_op_inuser(op, uaddr2);
        if (unlikely(op_ret < 0)) {
@@ -385,25 +746,9 @@ retry:
                 * still holding the mmap_sem.
                 */
                if (attempt++) {
-                       struct vm_area_struct * vma;
-                       struct mm_struct *mm = current->mm;
-                       unsigned long address = (unsigned long)uaddr2;
-
-                       ret = -EFAULT;
-                       if (attempt >= 2 ||
-                           !(vma = find_vma(mm, address)) ||
-                           vma->vm_start > address ||
-                           !(vma->vm_flags & VM_WRITE))
-                               goto out;
-
-                       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:
+                       if (futex_handle_fault((unsigned long)uaddr2,
+                                               attempt)) {
+                               ret = -EFAULT;
                                goto out;
                        }
                        goto retry;
@@ -480,11 +825,7 @@ static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
        hb1 = hash_futex(&key1);
        hb2 = hash_futex(&key2);
 
-       if (hb1 < hb2)
-               spin_lock(&hb1->lock);
-       spin_lock(&hb2->lock);
-       if (hb1 > hb2)
-               spin_lock(&hb1->lock);
+       double_lock_hb(hb1, hb2);
 
        if (likely(cmpval != NULL)) {
                u32 curval;
@@ -522,17 +863,20 @@ static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
                if (++ret <= nr_wake) {
                        wake_futex(this);
                } else {
-                       list_move_tail(&this->list, &hb2->chain);
-                       this->lock_ptr = &hb2->lock;
+                       /*
+                        * If key1 and key2 hash to the same bucket, no need to
+                        * requeue.
+                        */
+                       if (likely(head1 != &hb2->chain)) {
+                               list_move_tail(&this->list, &hb2->chain);
+                               this->lock_ptr = &hb2->lock;
+                       }
                        this->key = key2;
                        get_key_refs(&key2);
                        drop_count++;
 
                        if (ret - nr_wake >= nr_requeue)
                                break;
-                       /* Make sure to stop if key1 == key2: */
-                       if (head1 == &hb2->chain && head1 != &next->list)
-                               head1 = &this->list;
                }
        }
 
@@ -572,6 +916,7 @@ 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);
+       q->task = current;
        spin_unlock(&hb->lock);
 }
 
@@ -605,6 +950,7 @@ 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;
+       barrier();
        if (lock_ptr != 0) {
                spin_lock(lock_ptr);
                /*
@@ -626,6 +972,9 @@ static int unqueue_me(struct futex_q *q)
                }
                WARN_ON(list_empty(&q->list));
                list_del(&q->list);
+
+               BUG_ON(q->pi_state);
+
                spin_unlock(lock_ptr);
                ret = 1;
        }
@@ -634,16 +983,36 @@ static int unqueue_me(struct futex_q *q)
        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.
+ */
+static void unqueue_me_pi(struct futex_q *q, struct futex_hash_bucket *hb)
+{
+       WARN_ON(list_empty(&q->list));
+       list_del(&q->list);
+
+       BUG_ON(!q->pi_state);
+       free_pi_state(q->pi_state);
+       q->pi_state = NULL;
+
+       spin_unlock(&hb->lock);
+
+       drop_key_refs(&q->key);
+}
+
 static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
 {
-       DECLARE_WAITQUEUE(wait, current);
+       struct task_struct *curr = current;
+       DECLARE_WAITQUEUE(wait, curr);
        struct futex_hash_bucket *hb;
        struct futex_q q;
        u32 uval;
        int ret;
 
+       q.pi_state = NULL;
  retry:
-       down_read(&current->mm->mmap_sem);
+       down_read(&curr->mm->mmap_sem);
 
        ret = get_futex_key(uaddr, &q.key);
        if (unlikely(ret != 0))
@@ -680,7 +1049,7 @@ 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(&current->mm->mmap_sem);
+               up_read(&curr->mm->mmap_sem);
 
                ret = get_user(uval, uaddr);
 
@@ -688,11 +1057,9 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
                        goto retry;
                return ret;
        }
-       if (uval != val) {
-               ret = -EWOULDBLOCK;
-               queue_unlock(&q, hb);
-               goto out_release_sem;
-       }
+       ret = -EWOULDBLOCK;
+       if (uval != val)
+               goto out_unlock_release_sem;
 
        /* Only actually queue if *uaddr contained val.  */
        __queue_me(&q, hb);
@@ -700,8 +1067,8 @@ 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(&current->mm->mmap_sem);
+        */
+       up_read(&curr->mm->mmap_sem);
 
        /*
         * There might have been scheduling since the queue_me(), as we
@@ -739,8 +1106,361 @@ static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time)
         */
        return -EINTR;
 
+ out_unlock_release_sem:
+       queue_unlock(&q, hb);
+
+ out_release_sem:
+       up_read(&curr->mm->mmap_sem);
+       return ret;
+}
+
+/*
+ * 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 futex_lock_pi(u32 __user *uaddr, int detect, unsigned long sec,
+                        long nsec, 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;
+
+       if (refill_pi_state_cache())
+               return -ENOMEM;
+
+       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);
+       }
+
+       q.pi_state = NULL;
+ retry:
+       down_read(&curr->mm->mmap_sem);
+
+       ret = get_futex_key(uaddr, &q.key);
+       if (unlikely(ret != 0))
+               goto out_release_sem;
+
+       hb = queue_lock(&q, -1, NULL);
+
+ retry_locked:
+       /*
+        * 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;
+
+       inc_preempt_count();
+       curval = futex_atomic_cmpxchg_inatomic(uaddr, 0, newval);
+       dec_preempt_count();
+
+       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);
+               ret = -EDEADLK;
+               goto out_unlock_release_sem;
+       }
+
+       /*
+        * 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();
+
+       if (unlikely(curval == -EFAULT))
+               goto uaddr_faulted;
+       if (unlikely(curval != uval))
+               goto retry_locked;
+
+       /*
+        * 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);
+
+       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;
+
+                       inc_preempt_count();
+                       curval = futex_atomic_cmpxchg_inatomic(uaddr,
+                                                              uval, newval);
+                       dec_preempt_count();
+
+                       if (unlikely(curval == -EFAULT))
+                               goto uaddr_faulted;
+                       if (unlikely(curval != uval))
+                               goto retry_locked;
+                       ret = 0;
+               }
+               goto out_unlock_release_sem;
+       }
+
+       /*
+        * Only actually queue now that the atomic ops are done:
+        */
+       __queue_me(&q, hb);
+
+       /*
+        * 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);
+
+       WARN_ON(!q.pi_state);
+       /*
+        * Block on the PI mutex:
+        */
+       if (!trylock)
+               ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
+       else {
+               ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
+               /* Fixup the trylock return value: */
+               ret = ret ? 0 : -EWOULDBLOCK;
+       }
+
+       down_read(&curr->mm->mmap_sem);
+       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);
+                       WARN_ON(list_empty(&q.pi_state->list));
+                       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);
+               WARN_ON(!list_empty(&q.pi_state->list));
+               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);
+               /*
+                * 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_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;
+               }
+       } else {
+               /*
+                * Catch the rare case, where the lock was released
+                * when we were on the way back before we locked
+                * the hash bucket.
+                */
+               if (ret && q.pi_state->owner == curr) {
+                       if (rt_mutex_trylock(&q.pi_state->pi_mutex))
+                               ret = 0;
+               }
+               /* 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);
+
+       return ret != -EINTR ? ret : -ERESTARTNOINTR;
+
+ out_unlock_release_sem:
+       queue_unlock(&q, hb);
+
  out_release_sem:
+       up_read(&curr->mm->mmap_sem);
+       return ret;
+
+ uaddr_faulted:
+       /*
+        * We have to r/w  *(int __user *)uaddr, but we can't modify it
+        * non-atomically.  Therefore, if get_user below is not
+        * enough, we need to handle the fault ourselves, while
+        * still holding the mmap_sem.
+        */
+       if (attempt++) {
+               if (futex_handle_fault((unsigned long)uaddr, attempt)) {
+                       ret = -EFAULT;
+                       goto out_unlock_release_sem;
+               }
+               goto retry_locked;
+       }
+
+       queue_unlock(&q, hb);
+       up_read(&curr->mm->mmap_sem);
+
+       ret = get_user(uval, uaddr);
+       if (!ret && (uval != -EFAULT))
+               goto retry;
+
+       return ret;
+}
+
+/*
+ * 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)
+{
+       struct futex_hash_bucket *hb;
+       struct futex_q *this, *next;
+       u32 uval;
+       struct list_head *head;
+       union futex_key key;
+       int ret, attempt = 0;
+
+retry:
+       if (get_user(uval, uaddr))
+               return -EFAULT;
+       /*
+        * We release only a lock we actually own:
+        */
+       if ((uval & FUTEX_TID_MASK) != current->pid)
+               return -EPERM;
+       /*
+        * First take all the futex related locks:
+        */
+       down_read(&current->mm->mmap_sem);
+
+       ret = get_futex_key(uaddr, &key);
+       if (unlikely(ret != 0))
+               goto out;
+
+       hb = hash_futex(&key);
+       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:
+        */
+       if (!(uval & FUTEX_OWNER_DIED)) {
+               inc_preempt_count();
+               uval = futex_atomic_cmpxchg_inatomic(uaddr, current->pid, 0);
+               dec_preempt_count();
+       }
+
+       if (unlikely(uval == -EFAULT))
+               goto pi_faulted;
+       /*
+        * Rare case: we managed to release the lock atomically,
+        * no need to wake anyone else up:
+        */
+       if (unlikely(uval == current->pid))
+               goto out_unlock;
+
+       /*
+        * Ok, other tasks may need to be woken up - check waiters
+        * and do the wakeup if necessary:
+        */
+       head = &hb->chain;
+
+       list_for_each_entry_safe(this, next, head, list) {
+               if (!match_futex (&this->key, &key))
+                       continue;
+               ret = wake_futex_pi(uaddr, uval, this);
+               /*
+                * The atomic access to the futex value
+                * generated a pagefault, so retry the
+                * user-access and the wakeup:
+                */
+               if (ret == -EFAULT)
+                       goto pi_faulted;
+               goto out_unlock;
+       }
+       /*
+        * No waiters - kernel unlocks the futex:
+        */
+       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);
+
+       return ret;
+
+pi_faulted:
+       /*
+        * We have to r/w  *(int __user *)uaddr, but we can't modify it
+        * non-atomically.  Therefore, if get_user below is not
+        * enough, we need to handle the fault ourselves, while
+        * still holding the mmap_sem.
+        */
+       if (attempt++) {
+               if (futex_handle_fault((unsigned long)uaddr, attempt)) {
+                       ret = -EFAULT;
+                       goto out_unlock;
+               }
+               goto retry_locked;
+       }
+
+       spin_unlock(&hb->lock);
        up_read(&current->mm->mmap_sem);
+
+       ret = get_user(uval, uaddr);
+       if (!ret && (uval != -EFAULT))
+               goto retry;
+
        return ret;
 }
 
@@ -819,6 +1539,7 @@ static int futex_fd(u32 __user *uaddr, int signal)
                err = -ENOMEM;
                goto error;
        }
+       q->pi_state = NULL;
 
        down_read(&current->mm->mmap_sem);
        err = get_futex_key(uaddr, &q->key);
@@ -856,7 +1577,7 @@ error:
  * Implementation: user-space maintains a per-thread list of locks it
  * is holding. Upon do_exit(), the kernel carefully walks this list,
  * and marks all locks that are owned by this thread with the
- * FUTEX_OWNER_DEAD bit, and wakes up a waiter (if any). The list is
+ * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
  * always manipulated with the lock held, so the list is private and
  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
  * field, to allow the kernel to clean up if the thread dies after
@@ -929,9 +1650,9 @@ 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;
+       u32 uval, nval, mval;
 
 retry:
        if (get_user(uval, uaddr))
@@ -948,17 +1669,45 @@ retry:
                 * thread-death.) The rest of the cleanup is done in
                 * userspace.
                 */
-               if (futex_atomic_cmpxchg_inatomic(uaddr, uval,
-                                        uval | FUTEX_OWNER_DIED) != uval)
+               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) {
+                       if (uval & FUTEX_WAITERS)
+                               futex_wake(uaddr, 1);
+               }
        }
        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 **head, int *pi)
+{
+       unsigned long uentry;
+
+       if (get_user(uentry, (unsigned long *)head))
+               return -EFAULT;
+
+       *entry = (void *)(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.
  *
@@ -968,14 +1717,14 @@ 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;
+       unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
        unsigned long futex_offset;
 
        /*
         * 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:
@@ -986,24 +1735,25 @@ 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);
+               handle_futex_death((void *)pending + futex_offset, curr, pip);
 
        while (entry != &head->list) {
                /*
                 * A pending lock might already be on the list, so
-                * dont process it twice:
+                * don't process it twice:
                 */
                if (entry != pending)
                        if (handle_futex_death((void *)entry + futex_offset,
-                                               curr))
+                                               curr, pi))
                                return;
                /*
                 * Fetch the next entry in the list:
                 */
-               if (get_user(entry, &entry->next))
+               if (fetch_robust_entry(&entry, &entry->next, &pi))
                        return;
                /*
                 * Avoid excessively long or circular lists:
@@ -1040,6 +1790,15 @@ long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout,
        case FUTEX_WAKE_OP:
                ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
                break;
+       case FUTEX_LOCK_PI:
+               ret = futex_lock_pi(uaddr, val, timeout, val2, 0);
+               break;
+       case FUTEX_UNLOCK_PI:
+               ret = futex_unlock_pi(uaddr);
+               break;
+       case FUTEX_TRYLOCK_PI:
+               ret = futex_lock_pi(uaddr, 0, timeout, val2, 1);
+               break;
        default:
                ret = -ENOSYS;
        }
@@ -1055,17 +1814,22 @@ asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
        unsigned long timeout = MAX_SCHEDULE_TIMEOUT;
        u32 val2 = 0;
 
-       if (utime && (op == FUTEX_WAIT)) {
+       if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
                if (copy_from_user(&t, utime, sizeof(t)) != 0)
                        return -EFAULT;
                if (!timespec_valid(&t))
                        return -EINVAL;
-               timeout = timespec_to_jiffies(&t) + 1;
+               if (op == FUTEX_WAIT)
+                       timeout = timespec_to_jiffies(&t) + 1;
+               else {
+                       timeout = t.tv_sec;
+                       val2 = t.tv_nsec;
+               }
        }
        /*
         * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
         */
-       if (op >= FUTEX_REQUEUE)
+       if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE)
                val2 = (u32) (unsigned long) utime;
 
        return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3);