perf: Fix double start/stop in x86_pmu_start()
[linux-2.6.git] / kernel / rtmutex.c
index 180978c..a242e69 100644 (file)
  *  See Documentation/rt-mutex-design.txt for details.
  */
 #include <linux/spinlock.h>
-#include <linux/module.h>
+#include <linux/export.h>
 #include <linux/sched.h>
 #include <linux/timer.h>
 
 #include "rtmutex_common.h"
 
-#ifdef CONFIG_DEBUG_RT_MUTEXES
-# include "rtmutex-debug.h"
-#else
-# include "rtmutex.h"
-#endif
-
 /*
  * lock->owner state tracking:
  *
- * lock->owner holds the task_struct pointer of the owner. Bit 0 and 1
- * are used to keep track of the "owner is pending" and "lock has
- * waiters" state.
- *
- * owner       bit1    bit0
- * NULL                0       0       lock is free (fast acquire possible)
- * NULL                0       1       invalid state
- * NULL                1       0       Transitional State*
- * NULL                1       1       invalid state
- * taskpointer 0       0       lock is held (fast release possible)
- * taskpointer 0       1       task is pending owner
- * taskpointer 1       0       lock is held and has waiters
- * taskpointer 1       1       task is pending owner and lock has more waiters
- *
- * Pending ownership is assigned to the top (highest priority)
- * waiter of the lock, when the lock is released. The thread is woken
- * up and can now take the lock. Until the lock is taken (bit 0
- * cleared) a competing higher priority thread can steal the lock
- * which puts the woken up thread back on the waiters list.
+ * lock->owner holds the task_struct pointer of the owner. Bit 0
+ * is used to keep track of the "lock has waiters" state.
+ *
+ * owner       bit0
+ * NULL                0       lock is free (fast acquire possible)
+ * NULL                1       lock is free and has waiters and the top waiter
+ *                             is going to take the lock*
+ * taskpointer 0       lock is held (fast release possible)
+ * taskpointer 1       lock is held and has waiters**
  *
  * The fast atomic compare exchange based acquire and release is only
- * possible when bit 0 and 1 of lock->owner are 0.
+ * possible when bit 0 of lock->owner is 0.
+ *
+ * (*) It also can be a transitional state when grabbing the lock
+ * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
+ * we need to set the bit0 before looking at the lock, and the owner may be
+ * NULL in this small time, hence this can be a transitional state.
  *
- * (*) There's a small time where the owner can be NULL and the
- * "lock has waiters" bit is set.  This can happen when grabbing the lock.
- * To prevent a cmpxchg of the owner releasing the lock, we need to set this
- * bit before looking at the lock, hence the reason this is a transitional
- * state.
+ * (**) There is a small time when bit 0 is set but there are no
+ * waiters. This can happen when grabbing the lock in the slow path.
+ * To prevent a cmpxchg of the owner releasing the lock, we need to
+ * set this bit before looking at the lock.
  */
 
 static void
-rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner,
-                  unsigned long mask)
+rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
 {
-       unsigned long val = (unsigned long)owner | mask;
+       unsigned long val = (unsigned long)owner;
 
        if (rt_mutex_has_waiters(lock))
                val |= RT_MUTEX_HAS_WAITERS;
@@ -144,9 +131,9 @@ static void rt_mutex_adjust_prio(struct task_struct *task)
 {
        unsigned long flags;
 
-       spin_lock_irqsave(&task->pi_lock, flags);
+       raw_spin_lock_irqsave(&task->pi_lock, flags);
        __rt_mutex_adjust_prio(task);
-       spin_unlock_irqrestore(&task->pi_lock, flags);
+       raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 }
 
 /*
@@ -191,7 +178,7 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
                        prev_max = max_lock_depth;
                        printk(KERN_WARNING "Maximum lock depth %d reached "
                               "task: %s (%d)\n", max_lock_depth,
-                              top_task->comm, top_task->pid);
+                              top_task->comm, task_pid_nr(top_task));
                }
                put_task_struct(task);
 
@@ -201,7 +188,7 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
        /*
         * Task can not go away as we did a get_task() before !
         */
-       spin_lock_irqsave(&task->pi_lock, flags);
+       raw_spin_lock_irqsave(&task->pi_lock, flags);
 
        waiter = task->pi_blocked_on;
        /*
@@ -209,9 +196,21 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
         * reached or the state of the chain has changed while we
         * dropped the locks.
         */
-       if (!waiter || !waiter->task)
+       if (!waiter)
                goto out_unlock_pi;
 
+       /*
+        * Check the orig_waiter state. After we dropped the locks,
+        * the previous owner of the lock might have released the lock.
+        */
+       if (orig_waiter && !rt_mutex_owner(orig_lock))
+               goto out_unlock_pi;
+
+       /*
+        * Drop out, when the task has no waiters. Note,
+        * top_waiter can be NULL, when we are in the deboosting
+        * mode!
+        */
        if (top_waiter && (!task_has_pi_waiters(task) ||
                           top_waiter != task_top_pi_waiter(task)))
                goto out_unlock_pi;
@@ -224,8 +223,8 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
                goto out_unlock_pi;
 
        lock = waiter->lock;
-       if (!spin_trylock(&lock->wait_lock)) {
-               spin_unlock_irqrestore(&task->pi_lock, flags);
+       if (!raw_spin_trylock(&lock->wait_lock)) {
+               raw_spin_unlock_irqrestore(&task->pi_lock, flags);
                cpu_relax();
                goto retry;
        }
@@ -233,7 +232,7 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
        /* Deadlock detection */
        if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
                debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
-               spin_unlock(&lock->wait_lock);
+               raw_spin_unlock(&lock->wait_lock);
                ret = deadlock_detect ? -EDEADLK : 0;
                goto out_unlock_pi;
        }
@@ -246,13 +245,24 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
        plist_add(&waiter->list_entry, &lock->wait_list);
 
        /* Release the task */
-       spin_unlock_irqrestore(&task->pi_lock, flags);
+       raw_spin_unlock_irqrestore(&task->pi_lock, flags);
+       if (!rt_mutex_owner(lock)) {
+               /*
+                * If the requeue above changed the top waiter, then we need
+                * to wake the new top waiter up to try to get the lock.
+                */
+
+               if (top_waiter != rt_mutex_top_waiter(lock))
+                       wake_up_process(rt_mutex_top_waiter(lock)->task);
+               raw_spin_unlock(&lock->wait_lock);
+               goto out_put_task;
+       }
        put_task_struct(task);
 
        /* Grab the next task */
        task = rt_mutex_owner(lock);
        get_task_struct(task);
-       spin_lock_irqsave(&task->pi_lock, flags);
+       raw_spin_lock_irqsave(&task->pi_lock, flags);
 
        if (waiter == rt_mutex_top_waiter(lock)) {
                /* Boost the owner */
@@ -270,10 +280,10 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
                __rt_mutex_adjust_prio(task);
        }
 
-       spin_unlock_irqrestore(&task->pi_lock, flags);
+       raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 
        top_waiter = rt_mutex_top_waiter(lock);
-       spin_unlock(&lock->wait_lock);
+       raw_spin_unlock(&lock->wait_lock);
 
        if (!detect_deadlock && waiter != top_waiter)
                goto out_put_task;
@@ -281,7 +291,7 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
        goto again;
 
  out_unlock_pi:
-       spin_unlock_irqrestore(&task->pi_lock, flags);
+       raw_spin_unlock_irqrestore(&task->pi_lock, flags);
  out_put_task:
        put_task_struct(task);
 
@@ -289,77 +299,16 @@ static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 }
 
 /*
- * Optimization: check if we can steal the lock from the
- * assigned pending owner [which might not have taken the
- * lock yet]:
- */
-static inline int try_to_steal_lock(struct rt_mutex *lock)
-{
-       struct task_struct *pendowner = rt_mutex_owner(lock);
-       struct rt_mutex_waiter *next;
-       unsigned long flags;
-
-       if (!rt_mutex_owner_pending(lock))
-               return 0;
-
-       if (pendowner == current)
-               return 1;
-
-       spin_lock_irqsave(&pendowner->pi_lock, flags);
-       if (current->prio >= pendowner->prio) {
-               spin_unlock_irqrestore(&pendowner->pi_lock, flags);
-               return 0;
-       }
-
-       /*
-        * Check if a waiter is enqueued on the pending owners
-        * pi_waiters list. Remove it and readjust pending owners
-        * priority.
-        */
-       if (likely(!rt_mutex_has_waiters(lock))) {
-               spin_unlock_irqrestore(&pendowner->pi_lock, flags);
-               return 1;
-       }
-
-       /* No chain handling, pending owner is not blocked on anything: */
-       next = rt_mutex_top_waiter(lock);
-       plist_del(&next->pi_list_entry, &pendowner->pi_waiters);
-       __rt_mutex_adjust_prio(pendowner);
-       spin_unlock_irqrestore(&pendowner->pi_lock, flags);
-
-       /*
-        * We are going to steal the lock and a waiter was
-        * enqueued on the pending owners pi_waiters queue. So
-        * we have to enqueue this waiter into
-        * current->pi_waiters list. This covers the case,
-        * where current is boosted because it holds another
-        * lock and gets unboosted because the booster is
-        * interrupted, so we would delay a waiter with higher
-        * priority as current->normal_prio.
-        *
-        * Note: in the rare case of a SCHED_OTHER task changing
-        * its priority and thus stealing the lock, next->task
-        * might be current:
-        */
-       if (likely(next->task != current)) {
-               spin_lock_irqsave(&current->pi_lock, flags);
-               plist_add(&next->pi_list_entry, &current->pi_waiters);
-               __rt_mutex_adjust_prio(current);
-               spin_unlock_irqrestore(&current->pi_lock, flags);
-       }
-       return 1;
-}
-
-/*
  * Try to take an rt-mutex
  *
- * This fails
- * - when the lock has a real owner
- * - when a different pending owner exists and has higher priority than current
- *
  * Must be called with lock->wait_lock held.
+ *
+ * @lock:   the lock to be acquired.
+ * @task:   the task which wants to acquire the lock
+ * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
  */
-static int try_to_take_rt_mutex(struct rt_mutex *lock)
+static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
+               struct rt_mutex_waiter *waiter)
 {
        /*
         * We have to be careful here if the atomic speedups are
@@ -382,15 +331,52 @@ static int try_to_take_rt_mutex(struct rt_mutex *lock)
         */
        mark_rt_mutex_waiters(lock);
 
-       if (rt_mutex_owner(lock) && !try_to_steal_lock(lock))
+       if (rt_mutex_owner(lock))
                return 0;
 
+       /*
+        * It will get the lock because of one of these conditions:
+        * 1) there is no waiter
+        * 2) higher priority than waiters
+        * 3) it is top waiter
+        */
+       if (rt_mutex_has_waiters(lock)) {
+               if (task->prio >= rt_mutex_top_waiter(lock)->list_entry.prio) {
+                       if (!waiter || waiter != rt_mutex_top_waiter(lock))
+                               return 0;
+               }
+       }
+
+       if (waiter || rt_mutex_has_waiters(lock)) {
+               unsigned long flags;
+               struct rt_mutex_waiter *top;
+
+               raw_spin_lock_irqsave(&task->pi_lock, flags);
+
+               /* remove the queued waiter. */
+               if (waiter) {
+                       plist_del(&waiter->list_entry, &lock->wait_list);
+                       task->pi_blocked_on = NULL;
+               }
+
+               /*
+                * We have to enqueue the top waiter(if it exists) into
+                * task->pi_waiters list.
+                */
+               if (rt_mutex_has_waiters(lock)) {
+                       top = rt_mutex_top_waiter(lock);
+                       top->pi_list_entry.prio = top->list_entry.prio;
+                       plist_add(&top->pi_list_entry, &task->pi_waiters);
+               }
+               raw_spin_unlock_irqrestore(&task->pi_lock, flags);
+       }
+
        /* We got the lock. */
        debug_rt_mutex_lock(lock);
 
-       rt_mutex_set_owner(lock, current, 0);
+       rt_mutex_set_owner(lock, task);
 
-       rt_mutex_deadlock_account_lock(lock, current);
+       rt_mutex_deadlock_account_lock(lock, task);
 
        return 1;
 }
@@ -404,6 +390,7 @@ static int try_to_take_rt_mutex(struct rt_mutex *lock)
  */
 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
                                   struct rt_mutex_waiter *waiter,
+                                  struct task_struct *task,
                                   int detect_deadlock)
 {
        struct task_struct *owner = rt_mutex_owner(lock);
@@ -411,31 +398,34 @@ static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
        unsigned long flags;
        int chain_walk = 0, res;
 
-       spin_lock_irqsave(&current->pi_lock, flags);
-       __rt_mutex_adjust_prio(current);
-       waiter->task = current;
+       raw_spin_lock_irqsave(&task->pi_lock, flags);
+       __rt_mutex_adjust_prio(task);
+       waiter->task = task;
        waiter->lock = lock;
-       plist_node_init(&waiter->list_entry, current->prio);
-       plist_node_init(&waiter->pi_list_entry, current->prio);
+       plist_node_init(&waiter->list_entry, task->prio);
+       plist_node_init(&waiter->pi_list_entry, task->prio);
 
        /* Get the top priority waiter on the lock */
        if (rt_mutex_has_waiters(lock))
                top_waiter = rt_mutex_top_waiter(lock);
        plist_add(&waiter->list_entry, &lock->wait_list);
 
-       current->pi_blocked_on = waiter;
+       task->pi_blocked_on = waiter;
+
+       raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 
-       spin_unlock_irqrestore(&current->pi_lock, flags);
+       if (!owner)
+               return 0;
 
        if (waiter == rt_mutex_top_waiter(lock)) {
-               spin_lock_irqsave(&owner->pi_lock, flags);
+               raw_spin_lock_irqsave(&owner->pi_lock, flags);
                plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
                plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 
                __rt_mutex_adjust_prio(owner);
                if (owner->pi_blocked_on)
                        chain_walk = 1;
-               spin_unlock_irqrestore(&owner->pi_lock, flags);
+               raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
        }
        else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
                chain_walk = 1;
@@ -450,12 +440,12 @@ static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
         */
        get_task_struct(owner);
 
-       spin_unlock(&lock->wait_lock);
+       raw_spin_unlock(&lock->wait_lock);
 
        res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
-                                        current);
+                                        task);
 
-       spin_lock(&lock->wait_lock);
+       raw_spin_lock(&lock->wait_lock);
 
        return res;
 }
@@ -463,21 +453,18 @@ static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 /*
  * Wake up the next waiter on the lock.
  *
- * Remove the top waiter from the current tasks waiter list and from
- * the lock waiter list. Set it as pending owner. Then wake it up.
+ * Remove the top waiter from the current tasks waiter list and wake it up.
  *
  * Called with lock->wait_lock held.
  */
 static void wakeup_next_waiter(struct rt_mutex *lock)
 {
        struct rt_mutex_waiter *waiter;
-       struct task_struct *pendowner;
        unsigned long flags;
 
-       spin_lock_irqsave(&current->pi_lock, flags);
+       raw_spin_lock_irqsave(&current->pi_lock, flags);
 
        waiter = rt_mutex_top_waiter(lock);
-       plist_del(&waiter->list_entry, &lock->wait_list);
 
        /*
         * Remove it from current->pi_waiters. We do not adjust a
@@ -486,43 +473,19 @@ static void wakeup_next_waiter(struct rt_mutex *lock)
         * lock->wait_lock.
         */
        plist_del(&waiter->pi_list_entry, &current->pi_waiters);
-       pendowner = waiter->task;
-       waiter->task = NULL;
-
-       rt_mutex_set_owner(lock, pendowner, RT_MUTEX_OWNER_PENDING);
-
-       spin_unlock_irqrestore(&current->pi_lock, flags);
 
-       /*
-        * Clear the pi_blocked_on variable and enqueue a possible
-        * waiter into the pi_waiters list of the pending owner. This
-        * prevents that in case the pending owner gets unboosted a
-        * waiter with higher priority than pending-owner->normal_prio
-        * is blocked on the unboosted (pending) owner.
-        */
-       spin_lock_irqsave(&pendowner->pi_lock, flags);
+       rt_mutex_set_owner(lock, NULL);
 
-       WARN_ON(!pendowner->pi_blocked_on);
-       WARN_ON(pendowner->pi_blocked_on != waiter);
-       WARN_ON(pendowner->pi_blocked_on->lock != lock);
+       raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 
-       pendowner->pi_blocked_on = NULL;
-
-       if (rt_mutex_has_waiters(lock)) {
-               struct rt_mutex_waiter *next;
-
-               next = rt_mutex_top_waiter(lock);
-               plist_add(&next->pi_list_entry, &pendowner->pi_waiters);
-       }
-       spin_unlock_irqrestore(&pendowner->pi_lock, flags);
-
-       wake_up_process(pendowner);
+       wake_up_process(waiter->task);
 }
 
 /*
- * Remove a waiter from a lock
+ * Remove a waiter from a lock and give up
  *
- * Must be called with lock->wait_lock held
+ * Must be called with lock->wait_lock held and
+ * have just failed to try_to_take_rt_mutex().
  */
 static void remove_waiter(struct rt_mutex *lock,
                          struct rt_mutex_waiter *waiter)
@@ -532,15 +495,17 @@ static void remove_waiter(struct rt_mutex *lock,
        unsigned long flags;
        int chain_walk = 0;
 
-       spin_lock_irqsave(&current->pi_lock, flags);
+       raw_spin_lock_irqsave(&current->pi_lock, flags);
        plist_del(&waiter->list_entry, &lock->wait_list);
-       waiter->task = NULL;
        current->pi_blocked_on = NULL;
-       spin_unlock_irqrestore(&current->pi_lock, flags);
+       raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 
-       if (first && owner != current) {
+       if (!owner)
+               return;
 
-               spin_lock_irqsave(&owner->pi_lock, flags);
+       if (first) {
+
+               raw_spin_lock_irqsave(&owner->pi_lock, flags);
 
                plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 
@@ -555,7 +520,7 @@ static void remove_waiter(struct rt_mutex *lock,
                if (owner->pi_blocked_on)
                        chain_walk = 1;
 
-               spin_unlock_irqrestore(&owner->pi_lock, flags);
+               raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
        }
 
        WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
@@ -566,11 +531,11 @@ static void remove_waiter(struct rt_mutex *lock,
        /* gets dropped in rt_mutex_adjust_prio_chain()! */
        get_task_struct(owner);
 
-       spin_unlock(&lock->wait_lock);
+       raw_spin_unlock(&lock->wait_lock);
 
        rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 
-       spin_lock(&lock->wait_lock);
+       raw_spin_lock(&lock->wait_lock);
 }
 
 /*
@@ -583,53 +548,41 @@ void rt_mutex_adjust_pi(struct task_struct *task)
        struct rt_mutex_waiter *waiter;
        unsigned long flags;
 
-       spin_lock_irqsave(&task->pi_lock, flags);
+       raw_spin_lock_irqsave(&task->pi_lock, flags);
 
        waiter = task->pi_blocked_on;
        if (!waiter || waiter->list_entry.prio == task->prio) {
-               spin_unlock_irqrestore(&task->pi_lock, flags);
+               raw_spin_unlock_irqrestore(&task->pi_lock, flags);
                return;
        }
 
-       spin_unlock_irqrestore(&task->pi_lock, flags);
+       raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 
        /* gets dropped in rt_mutex_adjust_prio_chain()! */
        get_task_struct(task);
        rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 }
 
-/*
- * Slow path lock function:
+/**
+ * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
+ * @lock:               the rt_mutex to take
+ * @state:              the state the task should block in (TASK_INTERRUPTIBLE
+ *                      or TASK_UNINTERRUPTIBLE)
+ * @timeout:            the pre-initialized and started timer, or NULL for none
+ * @waiter:             the pre-initialized rt_mutex_waiter
+ *
+ * lock->wait_lock must be held by the caller.
  */
 static int __sched
-rt_mutex_slowlock(struct rt_mutex *lock, int state,
-                 struct hrtimer_sleeper *timeout,
-                 int detect_deadlock)
+__rt_mutex_slowlock(struct rt_mutex *lock, int state,
+                   struct hrtimer_sleeper *timeout,
+                   struct rt_mutex_waiter *waiter)
 {
-       struct rt_mutex_waiter waiter;
        int ret = 0;
 
-       debug_rt_mutex_init_waiter(&waiter);
-       waiter.task = NULL;
-
-       spin_lock(&lock->wait_lock);
-
-       /* Try to acquire the lock again: */
-       if (try_to_take_rt_mutex(lock)) {
-               spin_unlock(&lock->wait_lock);
-               return 0;
-       }
-
-       set_current_state(state);
-
-       /* Setup the timer, when timeout != NULL */
-       if (unlikely(timeout))
-               hrtimer_start(&timeout->timer, timeout->timer.expires,
-                             HRTIMER_MODE_ABS);
-
        for (;;) {
                /* Try to acquire the lock: */
-               if (try_to_take_rt_mutex(lock))
+               if (try_to_take_rt_mutex(lock, current, waiter))
                        break;
 
                /*
@@ -646,40 +599,57 @@ rt_mutex_slowlock(struct rt_mutex *lock, int state,
                                break;
                }
 
-               /*
-                * waiter.task is NULL the first time we come here and
-                * when we have been woken up by the previous owner
-                * but the lock got stolen by a higher prio task.
-                */
-               if (!waiter.task) {
-                       ret = task_blocks_on_rt_mutex(lock, &waiter,
-                                                     detect_deadlock);
-                       /*
-                        * If we got woken up by the owner then start loop
-                        * all over without going into schedule to try
-                        * to get the lock now:
-                        */
-                       if (unlikely(!waiter.task))
-                               continue;
-
-                       if (unlikely(ret))
-                               break;
-               }
-
-               spin_unlock(&lock->wait_lock);
+               raw_spin_unlock(&lock->wait_lock);
 
-               debug_rt_mutex_print_deadlock(&waiter);
+               debug_rt_mutex_print_deadlock(waiter);
 
-               if (waiter.task)
-                       schedule_rt_mutex(lock);
+               schedule_rt_mutex(lock);
 
-               spin_lock(&lock->wait_lock);
+               raw_spin_lock(&lock->wait_lock);
                set_current_state(state);
        }
 
+       return ret;
+}
+
+/*
+ * Slow path lock function:
+ */
+static int __sched
+rt_mutex_slowlock(struct rt_mutex *lock, int state,
+                 struct hrtimer_sleeper *timeout,
+                 int detect_deadlock)
+{
+       struct rt_mutex_waiter waiter;
+       int ret = 0;
+
+       debug_rt_mutex_init_waiter(&waiter);
+
+       raw_spin_lock(&lock->wait_lock);
+
+       /* Try to acquire the lock again: */
+       if (try_to_take_rt_mutex(lock, current, NULL)) {
+               raw_spin_unlock(&lock->wait_lock);
+               return 0;
+       }
+
+       set_current_state(state);
+
+       /* Setup the timer, when timeout != NULL */
+       if (unlikely(timeout)) {
+               hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
+               if (!hrtimer_active(&timeout->timer))
+                       timeout->task = NULL;
+       }
+
+       ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
+
+       if (likely(!ret))
+               ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
+
        set_current_state(TASK_RUNNING);
 
-       if (unlikely(waiter.task))
+       if (unlikely(ret))
                remove_waiter(lock, &waiter);
 
        /*
@@ -688,20 +658,12 @@ rt_mutex_slowlock(struct rt_mutex *lock, int state,
         */
        fixup_rt_mutex_waiters(lock);
 
-       spin_unlock(&lock->wait_lock);
+       raw_spin_unlock(&lock->wait_lock);
 
        /* Remove pending timer: */
        if (unlikely(timeout))
                hrtimer_cancel(&timeout->timer);
 
-       /*
-        * Readjust priority, when we did not get the lock. We might
-        * have been the pending owner and boosted. Since we did not
-        * take the lock, the PI boost has to go.
-        */
-       if (unlikely(ret))
-               rt_mutex_adjust_prio(current);
-
        debug_rt_mutex_free_waiter(&waiter);
 
        return ret;
@@ -715,11 +677,11 @@ rt_mutex_slowtrylock(struct rt_mutex *lock)
 {
        int ret = 0;
 
-       spin_lock(&lock->wait_lock);
+       raw_spin_lock(&lock->wait_lock);
 
        if (likely(rt_mutex_owner(lock) != current)) {
 
-               ret = try_to_take_rt_mutex(lock);
+               ret = try_to_take_rt_mutex(lock, current, NULL);
                /*
                 * try_to_take_rt_mutex() sets the lock waiters
                 * bit unconditionally. Clean this up.
@@ -727,7 +689,7 @@ rt_mutex_slowtrylock(struct rt_mutex *lock)
                fixup_rt_mutex_waiters(lock);
        }
 
-       spin_unlock(&lock->wait_lock);
+       raw_spin_unlock(&lock->wait_lock);
 
        return ret;
 }
@@ -738,7 +700,7 @@ rt_mutex_slowtrylock(struct rt_mutex *lock)
 static void __sched
 rt_mutex_slowunlock(struct rt_mutex *lock)
 {
-       spin_lock(&lock->wait_lock);
+       raw_spin_lock(&lock->wait_lock);
 
        debug_rt_mutex_unlock(lock);
 
@@ -746,13 +708,13 @@ rt_mutex_slowunlock(struct rt_mutex *lock)
 
        if (!rt_mutex_has_waiters(lock)) {
                lock->owner = NULL;
-               spin_unlock(&lock->wait_lock);
+               raw_spin_unlock(&lock->wait_lock);
                return;
        }
 
        wakeup_next_waiter(lock);
 
-       spin_unlock(&lock->wait_lock);
+       raw_spin_unlock(&lock->wait_lock);
 
        /* Undo pi boosting if necessary: */
        rt_mutex_adjust_prio(current);
@@ -848,9 +810,9 @@ int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 
 /**
- * rt_mutex_lock_interruptible_ktime - lock a rt_mutex interruptible
- *                                    the timeout structure is provided
- *                                    by the caller
+ * rt_mutex_timed_lock - lock a rt_mutex interruptible
+ *                     the timeout structure is provided
+ *                     by the caller
  *
  * @lock:              the rt_mutex to be locked
  * @timeout:           timeout structure or NULL (no timeout)
@@ -859,7 +821,7 @@ EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
  * Returns:
  *  0          on success
  * -EINTR      when interrupted by a signal
- * -ETIMEOUT   when the timeout expired
+ * -ETIMEDOUT  when the timeout expired
  * -EDEADLK    when the lock would deadlock (when deadlock detection is on)
  */
 int
@@ -897,7 +859,7 @@ void __sched rt_mutex_unlock(struct rt_mutex *lock)
 }
 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 
-/***
+/**
  * rt_mutex_destroy - mark a mutex unusable
  * @lock: the mutex to be destroyed
  *
@@ -927,8 +889,8 @@ EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 {
        lock->owner = NULL;
-       spin_lock_init(&lock->wait_lock);
-       plist_head_init(&lock->wait_list, &lock->wait_lock);
+       raw_spin_lock_init(&lock->wait_lock);
+       plist_head_init(&lock->wait_list);
 
        debug_rt_mutex_init(lock, name);
 }
@@ -949,7 +911,7 @@ void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 {
        __rt_mutex_init(lock, NULL);
        debug_rt_mutex_proxy_lock(lock, proxy_owner);
-       rt_mutex_set_owner(lock, proxy_owner, 0);
+       rt_mutex_set_owner(lock, proxy_owner);
        rt_mutex_deadlock_account_lock(lock, proxy_owner);
 }
 
@@ -965,11 +927,60 @@ void rt_mutex_proxy_unlock(struct rt_mutex *lock,
                           struct task_struct *proxy_owner)
 {
        debug_rt_mutex_proxy_unlock(lock);
-       rt_mutex_set_owner(lock, NULL, 0);
+       rt_mutex_set_owner(lock, NULL);
        rt_mutex_deadlock_account_unlock(proxy_owner);
 }
 
 /**
+ * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
+ * @lock:              the rt_mutex to take
+ * @waiter:            the pre-initialized rt_mutex_waiter
+ * @task:              the task to prepare
+ * @detect_deadlock:   perform deadlock detection (1) or not (0)
+ *
+ * Returns:
+ *  0 - task blocked on lock
+ *  1 - acquired the lock for task, caller should wake it up
+ * <0 - error
+ *
+ * Special API call for FUTEX_REQUEUE_PI support.
+ */
+int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
+                             struct rt_mutex_waiter *waiter,
+                             struct task_struct *task, int detect_deadlock)
+{
+       int ret;
+
+       raw_spin_lock(&lock->wait_lock);
+
+       if (try_to_take_rt_mutex(lock, task, NULL)) {
+               raw_spin_unlock(&lock->wait_lock);
+               return 1;
+       }
+
+       ret = task_blocks_on_rt_mutex(lock, waiter, task, detect_deadlock);
+
+       if (ret && !rt_mutex_owner(lock)) {
+               /*
+                * Reset the return value. We might have
+                * returned with -EDEADLK and the owner
+                * released the lock while we were walking the
+                * pi chain.  Let the waiter sort it out.
+                */
+               ret = 0;
+       }
+
+       if (unlikely(ret))
+               remove_waiter(lock, waiter);
+
+       raw_spin_unlock(&lock->wait_lock);
+
+       debug_rt_mutex_print_deadlock(waiter);
+
+       return ret;
+}
+
+/**
  * rt_mutex_next_owner - return the next owner of the lock
  *
  * @lock: the rt lock query
@@ -988,3 +999,48 @@ struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
 
        return rt_mutex_top_waiter(lock)->task;
 }
+
+/**
+ * rt_mutex_finish_proxy_lock() - Complete lock acquisition
+ * @lock:              the rt_mutex we were woken on
+ * @to:                        the timeout, null if none. hrtimer should already have
+ *                     been started.
+ * @waiter:            the pre-initialized rt_mutex_waiter
+ * @detect_deadlock:   perform deadlock detection (1) or not (0)
+ *
+ * Complete the lock acquisition started our behalf by another thread.
+ *
+ * Returns:
+ *  0 - success
+ * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
+ *
+ * Special API call for PI-futex requeue support
+ */
+int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
+                              struct hrtimer_sleeper *to,
+                              struct rt_mutex_waiter *waiter,
+                              int detect_deadlock)
+{
+       int ret;
+
+       raw_spin_lock(&lock->wait_lock);
+
+       set_current_state(TASK_INTERRUPTIBLE);
+
+       ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
+
+       set_current_state(TASK_RUNNING);
+
+       if (unlikely(ret))
+               remove_waiter(lock, waiter);
+
+       /*
+        * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
+        * have to fix that up.
+        */
+       fixup_rt_mutex_waiters(lock);
+
+       raw_spin_unlock(&lock->wait_lock);
+
+       return ret;
+}