Merge tag 'perf-tools-for-v5.15-2021-09-11' of git://git.kernel.org/pub/scm/linux...
[linux-2.6-microblaze.git] / kernel / locking / rtmutex.c
index ad0db32..6bb116c 100644 (file)
@@ -8,20 +8,58 @@
  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
  *  Copyright (C) 2006 Esben Nielsen
+ * Adaptive Spinlocks:
+ *  Copyright (C) 2008 Novell, Inc., Gregory Haskins, Sven Dietrich,
+ *                                  and Peter Morreale,
+ * Adaptive Spinlocks simplification:
+ *  Copyright (C) 2008 Red Hat, Inc., Steven Rostedt <srostedt@redhat.com>
  *
  *  See Documentation/locking/rt-mutex-design.rst for details.
  */
-#include <linux/spinlock.h>
-#include <linux/export.h>
+#include <linux/sched.h>
+#include <linux/sched/debug.h>
+#include <linux/sched/deadline.h>
 #include <linux/sched/signal.h>
 #include <linux/sched/rt.h>
-#include <linux/sched/deadline.h>
 #include <linux/sched/wake_q.h>
-#include <linux/sched/debug.h>
-#include <linux/timer.h>
+#include <linux/ww_mutex.h>
 
 #include "rtmutex_common.h"
 
+#ifndef WW_RT
+# define build_ww_mutex()      (false)
+# define ww_container_of(rtm)  NULL
+
+static inline int __ww_mutex_add_waiter(struct rt_mutex_waiter *waiter,
+                                       struct rt_mutex *lock,
+                                       struct ww_acquire_ctx *ww_ctx)
+{
+       return 0;
+}
+
+static inline void __ww_mutex_check_waiters(struct rt_mutex *lock,
+                                           struct ww_acquire_ctx *ww_ctx)
+{
+}
+
+static inline void ww_mutex_lock_acquired(struct ww_mutex *lock,
+                                         struct ww_acquire_ctx *ww_ctx)
+{
+}
+
+static inline int __ww_mutex_check_kill(struct rt_mutex *lock,
+                                       struct rt_mutex_waiter *waiter,
+                                       struct ww_acquire_ctx *ww_ctx)
+{
+       return 0;
+}
+
+#else
+# define build_ww_mutex()      (true)
+# define ww_container_of(rtm)  container_of(rtm, struct ww_mutex, base)
+# include "ww_mutex.h"
+#endif
+
 /*
  * lock->owner state tracking:
  *
@@ -50,7 +88,7 @@
  */
 
 static __always_inline void
-rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
+rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
 {
        unsigned long val = (unsigned long)owner;
 
@@ -60,13 +98,13 @@ rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
        WRITE_ONCE(lock->owner, (struct task_struct *)val);
 }
 
-static __always_inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
+static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
 {
        lock->owner = (struct task_struct *)
                        ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
 }
 
-static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex *lock)
+static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex_base *lock)
 {
        unsigned long owner, *p = (unsigned long *) &lock->owner;
 
@@ -141,15 +179,26 @@ static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  * set up.
  */
 #ifndef CONFIG_DEBUG_RT_MUTEXES
-# define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
-# define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
+static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
+                                                    struct task_struct *old,
+                                                    struct task_struct *new)
+{
+       return try_cmpxchg_acquire(&lock->owner, &old, new);
+}
+
+static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
+                                                    struct task_struct *old,
+                                                    struct task_struct *new)
+{
+       return try_cmpxchg_release(&lock->owner, &old, new);
+}
 
 /*
  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
  * relaxed semantics suffice.
  */
-static __always_inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
+static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
 {
        unsigned long owner, *p = (unsigned long *) &lock->owner;
 
@@ -165,7 +214,7 @@ static __always_inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  * 2) Drop lock->wait_lock
  * 3) Try to unlock the lock with cmpxchg
  */
-static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
+static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
                                                 unsigned long flags)
        __releases(lock->wait_lock)
 {
@@ -201,10 +250,22 @@ static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
 }
 
 #else
-# define rt_mutex_cmpxchg_acquire(l,c,n)       (0)
-# define rt_mutex_cmpxchg_release(l,c,n)       (0)
+static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
+                                                    struct task_struct *old,
+                                                    struct task_struct *new)
+{
+       return false;
+
+}
+
+static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
+                                                    struct task_struct *old,
+                                                    struct task_struct *new)
+{
+       return false;
+}
 
-static __always_inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
+static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
 {
        lock->owner = (struct task_struct *)
                        ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
@@ -213,7 +274,7 @@ static __always_inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 /*
  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
  */
-static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
+static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
                                                 unsigned long flags)
        __releases(lock->wait_lock)
 {
@@ -223,11 +284,28 @@ static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
 }
 #endif
 
+static __always_inline int __waiter_prio(struct task_struct *task)
+{
+       int prio = task->prio;
+
+       if (!rt_prio(prio))
+               return DEFAULT_PRIO;
+
+       return prio;
+}
+
+static __always_inline void
+waiter_update_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
+{
+       waiter->prio = __waiter_prio(task);
+       waiter->deadline = task->dl.deadline;
+}
+
 /*
  * Only use with rt_mutex_waiter_{less,equal}()
  */
 #define task_to_waiter(p)      \
-       &(struct rt_mutex_waiter){ .prio = (p)->prio, .deadline = (p)->dl.deadline }
+       &(struct rt_mutex_waiter){ .prio = __waiter_prio(p), .deadline = (p)->dl.deadline }
 
 static __always_inline int rt_mutex_waiter_less(struct rt_mutex_waiter *left,
                                                struct rt_mutex_waiter *right)
@@ -265,22 +343,63 @@ static __always_inline int rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
        return 1;
 }
 
+static inline bool rt_mutex_steal(struct rt_mutex_waiter *waiter,
+                                 struct rt_mutex_waiter *top_waiter)
+{
+       if (rt_mutex_waiter_less(waiter, top_waiter))
+               return true;
+
+#ifdef RT_MUTEX_BUILD_SPINLOCKS
+       /*
+        * Note that RT tasks are excluded from same priority (lateral)
+        * steals to prevent the introduction of an unbounded latency.
+        */
+       if (rt_prio(waiter->prio) || dl_prio(waiter->prio))
+               return false;
+
+       return rt_mutex_waiter_equal(waiter, top_waiter);
+#else
+       return false;
+#endif
+}
+
 #define __node_2_waiter(node) \
        rb_entry((node), struct rt_mutex_waiter, tree_entry)
 
 static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
 {
-       return rt_mutex_waiter_less(__node_2_waiter(a), __node_2_waiter(b));
+       struct rt_mutex_waiter *aw = __node_2_waiter(a);
+       struct rt_mutex_waiter *bw = __node_2_waiter(b);
+
+       if (rt_mutex_waiter_less(aw, bw))
+               return 1;
+
+       if (!build_ww_mutex())
+               return 0;
+
+       if (rt_mutex_waiter_less(bw, aw))
+               return 0;
+
+       /* NOTE: relies on waiter->ww_ctx being set before insertion */
+       if (aw->ww_ctx) {
+               if (!bw->ww_ctx)
+                       return 1;
+
+               return (signed long)(aw->ww_ctx->stamp -
+                                    bw->ww_ctx->stamp) < 0;
+       }
+
+       return 0;
 }
 
 static __always_inline void
-rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
+rt_mutex_enqueue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
 {
        rb_add_cached(&waiter->tree_entry, &lock->waiters, __waiter_less);
 }
 
 static __always_inline void
-rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
+rt_mutex_dequeue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
 {
        if (RB_EMPTY_NODE(&waiter->tree_entry))
                return;
@@ -326,6 +445,35 @@ static __always_inline void rt_mutex_adjust_prio(struct task_struct *p)
        rt_mutex_setprio(p, pi_task);
 }
 
+/* RT mutex specific wake_q wrappers */
+static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
+                                               struct rt_mutex_waiter *w)
+{
+       if (IS_ENABLED(CONFIG_PREEMPT_RT) && w->wake_state != TASK_NORMAL) {
+               if (IS_ENABLED(CONFIG_PROVE_LOCKING))
+                       WARN_ON_ONCE(wqh->rtlock_task);
+               get_task_struct(w->task);
+               wqh->rtlock_task = w->task;
+       } else {
+               wake_q_add(&wqh->head, w->task);
+       }
+}
+
+static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
+{
+       if (IS_ENABLED(CONFIG_PREEMPT_RT) && wqh->rtlock_task) {
+               wake_up_state(wqh->rtlock_task, TASK_RTLOCK_WAIT);
+               put_task_struct(wqh->rtlock_task);
+               wqh->rtlock_task = NULL;
+       }
+
+       if (!wake_q_empty(&wqh->head))
+               wake_up_q(&wqh->head);
+
+       /* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
+       preempt_enable();
+}
+
 /*
  * Deadlock detection is conditional:
  *
@@ -348,12 +496,7 @@ rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
        return chwalk == RT_MUTEX_FULL_CHAINWALK;
 }
 
-/*
- * Max number of times we'll walk the boosting chain:
- */
-int max_lock_depth = 1024;
-
-static __always_inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
+static __always_inline struct rt_mutex_base *task_blocked_on_lock(struct task_struct *p)
 {
        return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 }
@@ -423,15 +566,15 @@ static __always_inline struct rt_mutex *task_blocked_on_lock(struct task_struct
  */
 static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
                                              enum rtmutex_chainwalk chwalk,
-                                             struct rt_mutex *orig_lock,
-                                             struct rt_mutex *next_lock,
+                                             struct rt_mutex_base *orig_lock,
+                                             struct rt_mutex_base *next_lock,
                                              struct rt_mutex_waiter *orig_waiter,
                                              struct task_struct *top_task)
 {
        struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
        struct rt_mutex_waiter *prerequeue_top_waiter;
        int ret = 0, depth = 0;
-       struct rt_mutex *lock;
+       struct rt_mutex_base *lock;
        bool detect_deadlock;
        bool requeue = true;
 
@@ -513,6 +656,31 @@ static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
        if (next_lock != waiter->lock)
                goto out_unlock_pi;
 
+       /*
+        * There could be 'spurious' loops in the lock graph due to ww_mutex,
+        * consider:
+        *
+        *   P1: A, ww_A, ww_B
+        *   P2: ww_B, ww_A
+        *   P3: A
+        *
+        * P3 should not return -EDEADLK because it gets trapped in the cycle
+        * created by P1 and P2 (which will resolve -- and runs into
+        * max_lock_depth above). Therefore disable detect_deadlock such that
+        * the below termination condition can trigger once all relevant tasks
+        * are boosted.
+        *
+        * Even when we start with ww_mutex we can disable deadlock detection,
+        * since we would supress a ww_mutex induced deadlock at [6] anyway.
+        * Supressing it here however is not sufficient since we might still
+        * hit [6] due to adjustment driven iteration.
+        *
+        * NOTE: if someone were to create a deadlock between 2 ww_classes we'd
+        * utterly fail to report it; lockdep should.
+        */
+       if (IS_ENABLED(CONFIG_PREEMPT_RT) && waiter->ww_ctx && detect_deadlock)
+               detect_deadlock = false;
+
        /*
         * Drop out, when the task has no waiters. Note,
         * top_waiter can be NULL, when we are in the deboosting
@@ -574,8 +742,21 @@ static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
         * walk, we detected a deadlock.
         */
        if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
-               raw_spin_unlock(&lock->wait_lock);
                ret = -EDEADLK;
+
+               /*
+                * When the deadlock is due to ww_mutex; also see above. Don't
+                * report the deadlock and instead let the ww_mutex wound/die
+                * logic pick which of the contending threads gets -EDEADLK.
+                *
+                * NOTE: assumes the cycle only contains a single ww_class; any
+                * other configuration and we fail to report; also, see
+                * lockdep.
+                */
+               if (IS_ENABLED(CONFIG_PREEMPT_RT) && orig_waiter && orig_waiter->ww_ctx)
+                       ret = 0;
+
+               raw_spin_unlock(&lock->wait_lock);
                goto out_unlock_pi;
        }
 
@@ -653,8 +834,7 @@ static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
         * serializes all pi_waiters access and rb_erase() does not care about
         * the values of the node being removed.
         */
-       waiter->prio = task->prio;
-       waiter->deadline = task->dl.deadline;
+       waiter_update_prio(waiter, task);
 
        rt_mutex_enqueue(lock, waiter);
 
@@ -676,7 +856,7 @@ static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
                 * to get the lock.
                 */
                if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
-                       wake_up_process(rt_mutex_top_waiter(lock)->task);
+                       wake_up_state(waiter->task, waiter->wake_state);
                raw_spin_unlock_irq(&lock->wait_lock);
                return 0;
        }
@@ -779,7 +959,7 @@ static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
  *         callsite called task_blocked_on_lock(), otherwise NULL
  */
 static int __sched
-try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
+try_to_take_rt_mutex(struct rt_mutex_base *lock, struct task_struct *task,
                     struct rt_mutex_waiter *waiter)
 {
        lockdep_assert_held(&lock->wait_lock);
@@ -815,19 +995,21 @@ try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
         * trylock attempt.
         */
        if (waiter) {
-               /*
-                * If waiter is not the highest priority waiter of
-                * @lock, give up.
-                */
-               if (waiter != rt_mutex_top_waiter(lock))
-                       return 0;
+               struct rt_mutex_waiter *top_waiter = rt_mutex_top_waiter(lock);
 
                /*
-                * We can acquire the lock. Remove the waiter from the
-                * lock waiters tree.
+                * If waiter is the highest priority waiter of @lock,
+                * or allowed to steal it, take it over.
                 */
-               rt_mutex_dequeue(lock, waiter);
-
+               if (waiter == top_waiter || rt_mutex_steal(waiter, top_waiter)) {
+                       /*
+                        * We can acquire the lock. Remove the waiter from the
+                        * lock waiters tree.
+                        */
+                       rt_mutex_dequeue(lock, waiter);
+               } else {
+                       return 0;
+               }
        } else {
                /*
                 * If the lock has waiters already we check whether @task is
@@ -838,13 +1020,9 @@ try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
                 * not need to be dequeued.
                 */
                if (rt_mutex_has_waiters(lock)) {
-                       /*
-                        * If @task->prio is greater than or equal to
-                        * the top waiter priority (kernel view),
-                        * @task lost.
-                        */
-                       if (!rt_mutex_waiter_less(task_to_waiter(task),
-                                                 rt_mutex_top_waiter(lock)))
+                       /* Check whether the trylock can steal it. */
+                       if (!rt_mutex_steal(task_to_waiter(task),
+                                           rt_mutex_top_waiter(lock)))
                                return 0;
 
                        /*
@@ -897,14 +1075,15 @@ takeit:
  *
  * This must be called with lock->wait_lock held and interrupts disabled
  */
-static int __sched task_blocks_on_rt_mutex(struct rt_mutex *lock,
+static int __sched task_blocks_on_rt_mutex(struct rt_mutex_base *lock,
                                           struct rt_mutex_waiter *waiter,
                                           struct task_struct *task,
+                                          struct ww_acquire_ctx *ww_ctx,
                                           enum rtmutex_chainwalk chwalk)
 {
        struct task_struct *owner = rt_mutex_owner(lock);
        struct rt_mutex_waiter *top_waiter = waiter;
-       struct rt_mutex *next_lock;
+       struct rt_mutex_base *next_lock;
        int chain_walk = 0, res;
 
        lockdep_assert_held(&lock->wait_lock);
@@ -924,8 +1103,7 @@ static int __sched task_blocks_on_rt_mutex(struct rt_mutex *lock,
        raw_spin_lock(&task->pi_lock);
        waiter->task = task;
        waiter->lock = lock;
-       waiter->prio = task->prio;
-       waiter->deadline = task->dl.deadline;
+       waiter_update_prio(waiter, task);
 
        /* Get the top priority waiter on the lock */
        if (rt_mutex_has_waiters(lock))
@@ -936,6 +1114,21 @@ static int __sched task_blocks_on_rt_mutex(struct rt_mutex *lock,
 
        raw_spin_unlock(&task->pi_lock);
 
+       if (build_ww_mutex() && ww_ctx) {
+               struct rt_mutex *rtm;
+
+               /* Check whether the waiter should back out immediately */
+               rtm = container_of(lock, struct rt_mutex, rtmutex);
+               res = __ww_mutex_add_waiter(waiter, rtm, ww_ctx);
+               if (res) {
+                       raw_spin_lock(&task->pi_lock);
+                       rt_mutex_dequeue(lock, waiter);
+                       task->pi_blocked_on = NULL;
+                       raw_spin_unlock(&task->pi_lock);
+                       return res;
+               }
+       }
+
        if (!owner)
                return 0;
 
@@ -986,8 +1179,8 @@ static int __sched task_blocks_on_rt_mutex(struct rt_mutex *lock,
  *
  * Called with lock->wait_lock held and interrupts disabled.
  */
-static void __sched mark_wakeup_next_waiter(struct wake_q_head *wake_q,
-                                           struct rt_mutex *lock)
+static void __sched mark_wakeup_next_waiter(struct rt_wake_q_head *wqh,
+                                           struct rt_mutex_base *lock)
 {
        struct rt_mutex_waiter *waiter;
 
@@ -1023,25 +1216,201 @@ static void __sched mark_wakeup_next_waiter(struct wake_q_head *wake_q,
         * deboost but before waking our donor task, hence the preempt_disable()
         * before unlock.
         *
-        * Pairs with preempt_enable() in rt_mutex_postunlock();
+        * Pairs with preempt_enable() in rt_mutex_wake_up_q();
         */
        preempt_disable();
-       wake_q_add(wake_q, waiter->task);
+       rt_mutex_wake_q_add(wqh, waiter);
        raw_spin_unlock(&current->pi_lock);
 }
 
+static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
+{
+       int ret = try_to_take_rt_mutex(lock, current, NULL);
+
+       /*
+        * try_to_take_rt_mutex() sets the lock waiters bit
+        * unconditionally. Clean this up.
+        */
+       fixup_rt_mutex_waiters(lock);
+
+       return ret;
+}
+
+/*
+ * Slow path try-lock function:
+ */
+static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock)
+{
+       unsigned long flags;
+       int ret;
+
+       /*
+        * If the lock already has an owner we fail to get the lock.
+        * This can be done without taking the @lock->wait_lock as
+        * it is only being read, and this is a trylock anyway.
+        */
+       if (rt_mutex_owner(lock))
+               return 0;
+
+       /*
+        * The mutex has currently no owner. Lock the wait lock and try to
+        * acquire the lock. We use irqsave here to support early boot calls.
+        */
+       raw_spin_lock_irqsave(&lock->wait_lock, flags);
+
+       ret = __rt_mutex_slowtrylock(lock);
+
+       raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
+
+       return ret;
+}
+
+static __always_inline int __rt_mutex_trylock(struct rt_mutex_base *lock)
+{
+       if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
+               return 1;
+
+       return rt_mutex_slowtrylock(lock);
+}
+
+/*
+ * Slow path to release a rt-mutex.
+ */
+static void __sched rt_mutex_slowunlock(struct rt_mutex_base *lock)
+{
+       DEFINE_RT_WAKE_Q(wqh);
+       unsigned long flags;
+
+       /* irqsave required to support early boot calls */
+       raw_spin_lock_irqsave(&lock->wait_lock, flags);
+
+       debug_rt_mutex_unlock(lock);
+
+       /*
+        * We must be careful here if the fast path is enabled. If we
+        * have no waiters queued we cannot set owner to NULL here
+        * because of:
+        *
+        * foo->lock->owner = NULL;
+        *                      rtmutex_lock(foo->lock);   <- fast path
+        *                      free = atomic_dec_and_test(foo->refcnt);
+        *                      rtmutex_unlock(foo->lock); <- fast path
+        *                      if (free)
+        *                              kfree(foo);
+        * raw_spin_unlock(foo->lock->wait_lock);
+        *
+        * So for the fastpath enabled kernel:
+        *
+        * Nothing can set the waiters bit as long as we hold
+        * lock->wait_lock. So we do the following sequence:
+        *
+        *      owner = rt_mutex_owner(lock);
+        *      clear_rt_mutex_waiters(lock);
+        *      raw_spin_unlock(&lock->wait_lock);
+        *      if (cmpxchg(&lock->owner, owner, 0) == owner)
+        *              return;
+        *      goto retry;
+        *
+        * The fastpath disabled variant is simple as all access to
+        * lock->owner is serialized by lock->wait_lock:
+        *
+        *      lock->owner = NULL;
+        *      raw_spin_unlock(&lock->wait_lock);
+        */
+       while (!rt_mutex_has_waiters(lock)) {
+               /* Drops lock->wait_lock ! */
+               if (unlock_rt_mutex_safe(lock, flags) == true)
+                       return;
+               /* Relock the rtmutex and try again */
+               raw_spin_lock_irqsave(&lock->wait_lock, flags);
+       }
+
+       /*
+        * The wakeup next waiter path does not suffer from the above
+        * race. See the comments there.
+        *
+        * Queue the next waiter for wakeup once we release the wait_lock.
+        */
+       mark_wakeup_next_waiter(&wqh, lock);
+       raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
+
+       rt_mutex_wake_up_q(&wqh);
+}
+
+static __always_inline void __rt_mutex_unlock(struct rt_mutex_base *lock)
+{
+       if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
+               return;
+
+       rt_mutex_slowunlock(lock);
+}
+
+#ifdef CONFIG_SMP
+static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
+                                 struct rt_mutex_waiter *waiter,
+                                 struct task_struct *owner)
+{
+       bool res = true;
+
+       rcu_read_lock();
+       for (;;) {
+               /* If owner changed, trylock again. */
+               if (owner != rt_mutex_owner(lock))
+                       break;
+               /*
+                * Ensure that @owner is dereferenced after checking that
+                * the lock owner still matches @owner. If that fails,
+                * @owner might point to freed memory. If it still matches,
+                * the rcu_read_lock() ensures the memory stays valid.
+                */
+               barrier();
+               /*
+                * Stop spinning when:
+                *  - the lock owner has been scheduled out
+                *  - current is not longer the top waiter
+                *  - current is requested to reschedule (redundant
+                *    for CONFIG_PREEMPT_RCU=y)
+                *  - the VCPU on which owner runs is preempted
+                */
+               if (!owner->on_cpu || need_resched() ||
+                   rt_mutex_waiter_is_top_waiter(lock, waiter) ||
+                   vcpu_is_preempted(task_cpu(owner))) {
+                       res = false;
+                       break;
+               }
+               cpu_relax();
+       }
+       rcu_read_unlock();
+       return res;
+}
+#else
+static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
+                                 struct rt_mutex_waiter *waiter,
+                                 struct task_struct *owner)
+{
+       return false;
+}
+#endif
+
+#ifdef RT_MUTEX_BUILD_MUTEX
+/*
+ * Functions required for:
+ *     - rtmutex, futex on all kernels
+ *     - mutex and rwsem substitutions on RT kernels
+ */
+
 /*
  * Remove a waiter from a lock and give up
  *
- * Must be called with lock->wait_lock held and interrupts disabled. I must
+ * Must be called with lock->wait_lock held and interrupts disabled. It must
  * have just failed to try_to_take_rt_mutex().
  */
-static void __sched remove_waiter(struct rt_mutex *lock,
+static void __sched remove_waiter(struct rt_mutex_base *lock,
                                  struct rt_mutex_waiter *waiter)
 {
        bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
        struct task_struct *owner = rt_mutex_owner(lock);
-       struct rt_mutex *next_lock;
+       struct rt_mutex_base *next_lock;
 
        lockdep_assert_held(&lock->wait_lock);
 
@@ -1089,56 +1458,25 @@ static void __sched remove_waiter(struct rt_mutex *lock,
        raw_spin_lock_irq(&lock->wait_lock);
 }
 
-/*
- * Recheck the pi chain, in case we got a priority setting
+/**
+ * rt_mutex_slowlock_block() - Perform the wait-wake-try-to-take loop
+ * @lock:               the rt_mutex to take
+ * @ww_ctx:             WW mutex context pointer
+ * @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
  *
- * Called from sched_setscheduler
+ * Must be called with lock->wait_lock held and interrupts disabled
  */
-void __sched rt_mutex_adjust_pi(struct task_struct *task)
-{
-       struct rt_mutex_waiter *waiter;
-       struct rt_mutex *next_lock;
-       unsigned long flags;
-
-       raw_spin_lock_irqsave(&task->pi_lock, flags);
-
-       waiter = task->pi_blocked_on;
-       if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
-               raw_spin_unlock_irqrestore(&task->pi_lock, flags);
-               return;
-       }
-       next_lock = waiter->lock;
-       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, RT_MUTEX_MIN_CHAINWALK, NULL,
-                                  next_lock, NULL, task);
-}
-
-void __sched rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
-{
-       debug_rt_mutex_init_waiter(waiter);
-       RB_CLEAR_NODE(&waiter->pi_tree_entry);
-       RB_CLEAR_NODE(&waiter->tree_entry);
-       waiter->task = NULL;
-}
-
-/**
- * __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
- *
- * Must be called with lock->wait_lock held and interrupts disabled
- */
-static int __sched __rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state,
-                                      struct hrtimer_sleeper *timeout,
-                                      struct rt_mutex_waiter *waiter)
+static int __sched rt_mutex_slowlock_block(struct rt_mutex_base *lock,
+                                          struct ww_acquire_ctx *ww_ctx,
+                                          unsigned int state,
+                                          struct hrtimer_sleeper *timeout,
+                                          struct rt_mutex_waiter *waiter)
 {
+       struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
+       struct task_struct *owner;
        int ret = 0;
 
        for (;;) {
@@ -1155,9 +1493,20 @@ static int __sched __rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state
                        break;
                }
 
+               if (build_ww_mutex() && ww_ctx) {
+                       ret = __ww_mutex_check_kill(rtm, waiter, ww_ctx);
+                       if (ret)
+                               break;
+               }
+
+               if (waiter == rt_mutex_top_waiter(lock))
+                       owner = rt_mutex_owner(lock);
+               else
+                       owner = NULL;
                raw_spin_unlock_irq(&lock->wait_lock);
 
-               schedule();
+               if (!owner || !rtmutex_spin_on_owner(lock, waiter, owner))
+                       schedule();
 
                raw_spin_lock_irq(&lock->wait_lock);
                set_current_state(state);
@@ -1177,6 +1526,9 @@ static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
        if (res != -EDEADLOCK || detect_deadlock)
                return;
 
+       if (build_ww_mutex() && w->ww_ctx)
+               return;
+
        /*
         * Yell loudly and stop the task right here.
         */
@@ -1187,51 +1539,52 @@ static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
        }
 }
 
-/*
- * Slow path lock function:
+/**
+ * __rt_mutex_slowlock - Locking slowpath invoked with lock::wait_lock held
+ * @lock:      The rtmutex to block lock
+ * @ww_ctx:    WW mutex context pointer
+ * @state:     The task state for sleeping
+ * @chwalk:    Indicator whether full or partial chainwalk is requested
+ * @waiter:    Initializer waiter for blocking
  */
-static int __sched rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state,
-                                    struct hrtimer_sleeper *timeout,
-                                    enum rtmutex_chainwalk chwalk)
+static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
+                                      struct ww_acquire_ctx *ww_ctx,
+                                      unsigned int state,
+                                      enum rtmutex_chainwalk chwalk,
+                                      struct rt_mutex_waiter *waiter)
 {
-       struct rt_mutex_waiter waiter;
-       unsigned long flags;
-       int ret = 0;
-
-       rt_mutex_init_waiter(&waiter);
+       struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
+       struct ww_mutex *ww = ww_container_of(rtm);
+       int ret;
 
-       /*
-        * Technically we could use raw_spin_[un]lock_irq() here, but this can
-        * be called in early boot if the cmpxchg() fast path is disabled
-        * (debug, no architecture support). In this case we will acquire the
-        * rtmutex with lock->wait_lock held. But we cannot unconditionally
-        * enable interrupts in that early boot case. So we need to use the
-        * irqsave/restore variants.
-        */
-       raw_spin_lock_irqsave(&lock->wait_lock, flags);
+       lockdep_assert_held(&lock->wait_lock);
 
        /* Try to acquire the lock again: */
        if (try_to_take_rt_mutex(lock, current, NULL)) {
-               raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
+               if (build_ww_mutex() && ww_ctx) {
+                       __ww_mutex_check_waiters(rtm, ww_ctx);
+                       ww_mutex_lock_acquired(ww, ww_ctx);
+               }
                return 0;
        }
 
        set_current_state(state);
 
-       /* Setup the timer, when timeout != NULL */
-       if (unlikely(timeout))
-               hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
-
-       ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
-
+       ret = task_blocks_on_rt_mutex(lock, waiter, current, ww_ctx, chwalk);
        if (likely(!ret))
-               /* sleep on the mutex */
-               ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
-
-       if (unlikely(ret)) {
+               ret = rt_mutex_slowlock_block(lock, ww_ctx, state, NULL, waiter);
+
+       if (likely(!ret)) {
+               /* acquired the lock */
+               if (build_ww_mutex() && ww_ctx) {
+                       if (!ww_ctx->is_wait_die)
+                               __ww_mutex_check_waiters(rtm, ww_ctx);
+                       ww_mutex_lock_acquired(ww, ww_ctx);
+               }
+       } else {
                __set_current_state(TASK_RUNNING);
-               remove_waiter(lock, &waiter);
-               rt_mutex_handle_deadlock(ret, chwalk, &waiter);
+               remove_waiter(lock, waiter);
+               rt_mutex_handle_deadlock(ret, chwalk, waiter);
        }
 
        /*
@@ -1239,547 +1592,126 @@ static int __sched rt_mutex_slowlock(struct rt_mutex *lock, unsigned int state,
         * unconditionally. We might have to fix that up.
         */
        fixup_rt_mutex_waiters(lock);
-
-       raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
-
-       /* Remove pending timer: */
-       if (unlikely(timeout))
-               hrtimer_cancel(&timeout->timer);
-
-       debug_rt_mutex_free_waiter(&waiter);
-
        return ret;
 }
 
-static int __sched __rt_mutex_slowtrylock(struct rt_mutex *lock)
+static inline int __rt_mutex_slowlock_locked(struct rt_mutex_base *lock,
+                                            struct ww_acquire_ctx *ww_ctx,
+                                            unsigned int state)
 {
-       int ret = try_to_take_rt_mutex(lock, current, NULL);
+       struct rt_mutex_waiter waiter;
+       int ret;
 
-       /*
-        * try_to_take_rt_mutex() sets the lock waiters bit
-        * unconditionally. Clean this up.
-        */
-       fixup_rt_mutex_waiters(lock);
+       rt_mutex_init_waiter(&waiter);
+       waiter.ww_ctx = ww_ctx;
+
+       ret = __rt_mutex_slowlock(lock, ww_ctx, state, RT_MUTEX_MIN_CHAINWALK,
+                                 &waiter);
 
+       debug_rt_mutex_free_waiter(&waiter);
        return ret;
 }
 
 /*
- * Slow path try-lock function:
+ * rt_mutex_slowlock - Locking slowpath invoked when fast path fails
+ * @lock:      The rtmutex to block lock
+ * @ww_ctx:    WW mutex context pointer
+ * @state:     The task state for sleeping
  */
-static int __sched rt_mutex_slowtrylock(struct rt_mutex *lock)
+static int __sched rt_mutex_slowlock(struct rt_mutex_base *lock,
+                                    struct ww_acquire_ctx *ww_ctx,
+                                    unsigned int state)
 {
        unsigned long flags;
        int ret;
 
        /*
-        * If the lock already has an owner we fail to get the lock.
-        * This can be done without taking the @lock->wait_lock as
-        * it is only being read, and this is a trylock anyway.
-        */
-       if (rt_mutex_owner(lock))
-               return 0;
-
-       /*
-        * The mutex has currently no owner. Lock the wait lock and try to
-        * acquire the lock. We use irqsave here to support early boot calls.
+        * Technically we could use raw_spin_[un]lock_irq() here, but this can
+        * be called in early boot if the cmpxchg() fast path is disabled
+        * (debug, no architecture support). In this case we will acquire the
+        * rtmutex with lock->wait_lock held. But we cannot unconditionally
+        * enable interrupts in that early boot case. So we need to use the
+        * irqsave/restore variants.
         */
        raw_spin_lock_irqsave(&lock->wait_lock, flags);
-
-       ret = __rt_mutex_slowtrylock(lock);
-
+       ret = __rt_mutex_slowlock_locked(lock, ww_ctx, state);
        raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 
        return ret;
 }
 
-/*
- * Performs the wakeup of the top-waiter and re-enables preemption.
- */
-void __sched rt_mutex_postunlock(struct wake_q_head *wake_q)
-{
-       wake_up_q(wake_q);
-
-       /* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
-       preempt_enable();
-}
-
-/*
- * Slow path to release a rt-mutex.
- *
- * Return whether the current task needs to call rt_mutex_postunlock().
- */
-static void __sched rt_mutex_slowunlock(struct rt_mutex *lock)
-{
-       DEFINE_WAKE_Q(wake_q);
-       unsigned long flags;
-
-       /* irqsave required to support early boot calls */
-       raw_spin_lock_irqsave(&lock->wait_lock, flags);
-
-       debug_rt_mutex_unlock(lock);
-
-       /*
-        * We must be careful here if the fast path is enabled. If we
-        * have no waiters queued we cannot set owner to NULL here
-        * because of:
-        *
-        * foo->lock->owner = NULL;
-        *                      rtmutex_lock(foo->lock);   <- fast path
-        *                      free = atomic_dec_and_test(foo->refcnt);
-        *                      rtmutex_unlock(foo->lock); <- fast path
-        *                      if (free)
-        *                              kfree(foo);
-        * raw_spin_unlock(foo->lock->wait_lock);
-        *
-        * So for the fastpath enabled kernel:
-        *
-        * Nothing can set the waiters bit as long as we hold
-        * lock->wait_lock. So we do the following sequence:
-        *
-        *      owner = rt_mutex_owner(lock);
-        *      clear_rt_mutex_waiters(lock);
-        *      raw_spin_unlock(&lock->wait_lock);
-        *      if (cmpxchg(&lock->owner, owner, 0) == owner)
-        *              return;
-        *      goto retry;
-        *
-        * The fastpath disabled variant is simple as all access to
-        * lock->owner is serialized by lock->wait_lock:
-        *
-        *      lock->owner = NULL;
-        *      raw_spin_unlock(&lock->wait_lock);
-        */
-       while (!rt_mutex_has_waiters(lock)) {
-               /* Drops lock->wait_lock ! */
-               if (unlock_rt_mutex_safe(lock, flags) == true)
-                       return;
-               /* Relock the rtmutex and try again */
-               raw_spin_lock_irqsave(&lock->wait_lock, flags);
-       }
-
-       /*
-        * The wakeup next waiter path does not suffer from the above
-        * race. See the comments there.
-        *
-        * Queue the next waiter for wakeup once we release the wait_lock.
-        */
-       mark_wakeup_next_waiter(&wake_q, lock);
-       raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
-
-       rt_mutex_postunlock(&wake_q);
-}
-
-/*
- * debug aware fast / slowpath lock,trylock,unlock
- *
- * The atomic acquire/release ops are compiled away, when either the
- * architecture does not support cmpxchg or when debugging is enabled.
- */
-static __always_inline int __rt_mutex_lock(struct rt_mutex *lock, long state,
-                                          unsigned int subclass)
+static __always_inline int __rt_mutex_lock(struct rt_mutex_base *lock,
+                                          unsigned int state)
 {
-       int ret;
-
-       might_sleep();
-       mutex_acquire(&lock->dep_map, subclass, 0, _RET_IP_);
-
        if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
                return 0;
 
-       ret = rt_mutex_slowlock(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
-       if (ret)
-               mutex_release(&lock->dep_map, _RET_IP_);
-       return ret;
+       return rt_mutex_slowlock(lock, NULL, state);
 }
+#endif /* RT_MUTEX_BUILD_MUTEX */
 
-#ifdef CONFIG_DEBUG_LOCK_ALLOC
-/**
- * rt_mutex_lock_nested - lock a rt_mutex
- *
- * @lock: the rt_mutex to be locked
- * @subclass: the lockdep subclass
- */
-void __sched rt_mutex_lock_nested(struct rt_mutex *lock, unsigned int subclass)
-{
-       __rt_mutex_lock(lock, TASK_UNINTERRUPTIBLE, subclass);
-}
-EXPORT_SYMBOL_GPL(rt_mutex_lock_nested);
-
-#else /* !CONFIG_DEBUG_LOCK_ALLOC */
-
-/**
- * rt_mutex_lock - lock a rt_mutex
- *
- * @lock: the rt_mutex to be locked
- */
-void __sched rt_mutex_lock(struct rt_mutex *lock)
-{
-       __rt_mutex_lock(lock, TASK_UNINTERRUPTIBLE, 0);
-}
-EXPORT_SYMBOL_GPL(rt_mutex_lock);
-#endif
-
-/**
- * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
- *
- * @lock:              the rt_mutex to be locked
- *
- * Returns:
- *  0          on success
- * -EINTR      when interrupted by a signal
- */
-int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
-{
-       return __rt_mutex_lock(lock, TASK_INTERRUPTIBLE, 0);
-}
-EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
-
-/**
- * rt_mutex_trylock - try to lock a rt_mutex
- *
- * @lock:      the rt_mutex to be locked
- *
- * This function can only be called in thread context. It's safe to call it
- * from atomic regions, but not from hard or soft interrupt context.
- *
- * Returns:
- *  1 on success
- *  0 on contention
- */
-int __sched rt_mutex_trylock(struct rt_mutex *lock)
-{
-       int ret;
-
-       if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES) && WARN_ON_ONCE(!in_task()))
-               return 0;
-
-       /*
-        * No lockdep annotation required because lockdep disables the fast
-        * path.
-        */
-       if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
-               return 1;
-
-       ret = rt_mutex_slowtrylock(lock);
-       if (ret)
-               mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
-
-       return ret;
-}
-EXPORT_SYMBOL_GPL(rt_mutex_trylock);
-
-/**
- * rt_mutex_unlock - unlock a rt_mutex
- *
- * @lock: the rt_mutex to be unlocked
- */
-void __sched rt_mutex_unlock(struct rt_mutex *lock)
-{
-       mutex_release(&lock->dep_map, _RET_IP_);
-       if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
-               return;
-
-       rt_mutex_slowunlock(lock);
-}
-EXPORT_SYMBOL_GPL(rt_mutex_unlock);
-
+#ifdef RT_MUTEX_BUILD_SPINLOCKS
 /*
- * Futex variants, must not use fastpath.
+ * Functions required for spin/rw_lock substitution on RT kernels
  */
-int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
-{
-       return rt_mutex_slowtrylock(lock);
-}
-
-int __sched __rt_mutex_futex_trylock(struct rt_mutex *lock)
-{
-       return __rt_mutex_slowtrylock(lock);
-}
 
 /**
- * __rt_mutex_futex_unlock - Futex variant, that since futex variants
- * do not use the fast-path, can be simple and will not need to retry.
- *
- * @lock:      The rt_mutex to be unlocked
- * @wake_q:    The wake queue head from which to get the next lock waiter
+ * rtlock_slowlock_locked - Slow path lock acquisition for RT locks
+ * @lock:      The underlying RT mutex
  */
-bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
-                                    struct wake_q_head *wake_q)
+static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock)
 {
-       lockdep_assert_held(&lock->wait_lock);
-
-       debug_rt_mutex_unlock(lock);
-
-       if (!rt_mutex_has_waiters(lock)) {
-               lock->owner = NULL;
-               return false; /* done */
-       }
-
-       /*
-        * We've already deboosted, mark_wakeup_next_waiter() will
-        * retain preempt_disabled when we drop the wait_lock, to
-        * avoid inversion prior to the wakeup.  preempt_disable()
-        * therein pairs with rt_mutex_postunlock().
-        */
-       mark_wakeup_next_waiter(wake_q, lock);
-
-       return true; /* call postunlock() */
-}
-
-void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
-{
-       DEFINE_WAKE_Q(wake_q);
-       unsigned long flags;
-       bool postunlock;
-
-       raw_spin_lock_irqsave(&lock->wait_lock, flags);
-       postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
-       raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
-
-       if (postunlock)
-               rt_mutex_postunlock(&wake_q);
-}
+       struct rt_mutex_waiter waiter;
+       struct task_struct *owner;
 
-/**
- * __rt_mutex_init - initialize the rt_mutex
- *
- * @lock:      The rt_mutex to be initialized
- * @name:      The lock name used for debugging
- * @key:       The lock class key used for debugging
- *
- * Initialize the rt_mutex to unlocked state.
- *
- * Initializing of a locked rt_mutex is not allowed
- */
-void __sched __rt_mutex_init(struct rt_mutex *lock, const char *name,
-                    struct lock_class_key *key)
-{
-       debug_check_no_locks_freed((void *)lock, sizeof(*lock));
-       lockdep_init_map(&lock->dep_map, name, key, 0);
+       lockdep_assert_held(&lock->wait_lock);
 
-       __rt_mutex_basic_init(lock);
-}
-EXPORT_SYMBOL_GPL(__rt_mutex_init);
+       if (try_to_take_rt_mutex(lock, current, NULL))
+               return;
 
-/**
- * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
- *                             proxy owner
- *
- * @lock:      the rt_mutex to be locked
- * @proxy_owner:the task to set as owner
- *
- * No locking. Caller has to do serializing itself
- *
- * Special API call for PI-futex support. This initializes the rtmutex and
- * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
- * possible at this point because the pi_state which contains the rtmutex
- * is not yet visible to other tasks.
- */
-void __sched rt_mutex_init_proxy_locked(struct rt_mutex *lock,
-                                       struct task_struct *proxy_owner)
-{
-       __rt_mutex_basic_init(lock);
-       rt_mutex_set_owner(lock, proxy_owner);
-}
+       rt_mutex_init_rtlock_waiter(&waiter);
 
-/**
- * rt_mutex_proxy_unlock - release a lock on behalf of owner
- *
- * @lock:      the rt_mutex to be locked
- *
- * No locking. Caller has to do serializing itself
- *
- * Special API call for PI-futex support. This merrily cleans up the rtmutex
- * (debugging) state. Concurrent operations on this rt_mutex are not
- * possible because it belongs to the pi_state which is about to be freed
- * and it is not longer visible to other tasks.
- */
-void __sched rt_mutex_proxy_unlock(struct rt_mutex *lock)
-{
-       debug_rt_mutex_proxy_unlock(lock);
-       rt_mutex_set_owner(lock, NULL);
-}
+       /* Save current state and set state to TASK_RTLOCK_WAIT */
+       current_save_and_set_rtlock_wait_state();
 
-/**
- * __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
- *
- * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
- * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
- *
- * NOTE: does _NOT_ remove the @waiter on failure; must either call
- * rt_mutex_wait_proxy_lock() or rt_mutex_cleanup_proxy_lock() after this.
- *
- * Returns:
- *  0 - task blocked on lock
- *  1 - acquired the lock for task, caller should wake it up
- * <0 - error
- *
- * Special API call for PI-futex support.
- */
-int __sched __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
-                                       struct rt_mutex_waiter *waiter,
-                                       struct task_struct *task)
-{
-       int ret;
+       task_blocks_on_rt_mutex(lock, &waiter, current, NULL, RT_MUTEX_MIN_CHAINWALK);
 
-       lockdep_assert_held(&lock->wait_lock);
+       for (;;) {
+               /* Try to acquire the lock again */
+               if (try_to_take_rt_mutex(lock, current, &waiter))
+                       break;
 
-       if (try_to_take_rt_mutex(lock, task, NULL))
-               return 1;
+               if (&waiter == rt_mutex_top_waiter(lock))
+                       owner = rt_mutex_owner(lock);
+               else
+                       owner = NULL;
+               raw_spin_unlock_irq(&lock->wait_lock);
 
-       /* We enforce deadlock detection for futexes */
-       ret = task_blocks_on_rt_mutex(lock, waiter, task,
-                                     RT_MUTEX_FULL_CHAINWALK);
+               if (!owner || !rtmutex_spin_on_owner(lock, &waiter, owner))
+                       schedule_rtlock();
 
-       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;
+               raw_spin_lock_irq(&lock->wait_lock);
+               set_current_state(TASK_RTLOCK_WAIT);
        }
 
-       return ret;
-}
-
-/**
- * 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
- *
- * Starts the rt_mutex acquire; it enqueues the @waiter and does deadlock
- * detection. It does not wait, see rt_mutex_wait_proxy_lock() for that.
- *
- * NOTE: unlike __rt_mutex_start_proxy_lock this _DOES_ remove the @waiter
- * on failure.
- *
- * Returns:
- *  0 - task blocked on lock
- *  1 - acquired the lock for task, caller should wake it up
- * <0 - error
- *
- * Special API call for PI-futex support.
- */
-int __sched rt_mutex_start_proxy_lock(struct rt_mutex *lock,
-                                     struct rt_mutex_waiter *waiter,
-                                     struct task_struct *task)
-{
-       int ret;
-
-       raw_spin_lock_irq(&lock->wait_lock);
-       ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
-       if (unlikely(ret))
-               remove_waiter(lock, waiter);
-       raw_spin_unlock_irq(&lock->wait_lock);
-
-       return ret;
-}
-
-/**
- * rt_mutex_wait_proxy_lock() - Wait for 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
- *
- * Wait for the lock acquisition started on our behalf by
- * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
- * rt_mutex_cleanup_proxy_lock().
- *
- * Returns:
- *  0 - success
- * <0 - error, one of -EINTR, -ETIMEDOUT
- *
- * Special API call for PI-futex support
- */
-int __sched rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
-                                    struct hrtimer_sleeper *to,
-                                    struct rt_mutex_waiter *waiter)
-{
-       int ret;
+       /* Restore the task state */
+       current_restore_rtlock_saved_state();
 
-       raw_spin_lock_irq(&lock->wait_lock);
-       /* sleep on the mutex */
-       set_current_state(TASK_INTERRUPTIBLE);
-       ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
        /*
-        * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
-        * have to fix that up.
+        * 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_irq(&lock->wait_lock);
-
-       return ret;
+       debug_rt_mutex_free_waiter(&waiter);
 }
 
-/**
- * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
- * @lock:              the rt_mutex we were woken on
- * @waiter:            the pre-initialized rt_mutex_waiter
- *
- * Attempt to clean up after a failed __rt_mutex_start_proxy_lock() or
- * rt_mutex_wait_proxy_lock().
- *
- * Unless we acquired the lock; we're still enqueued on the wait-list and can
- * in fact still be granted ownership until we're removed. Therefore we can
- * find we are in fact the owner and must disregard the
- * rt_mutex_wait_proxy_lock() failure.
- *
- * Returns:
- *  true  - did the cleanup, we done.
- *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
- *          caller should disregards its return value.
- *
- * Special API call for PI-futex support
- */
-bool __sched rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
-                                        struct rt_mutex_waiter *waiter)
+static __always_inline void __sched rtlock_slowlock(struct rt_mutex_base *lock)
 {
-       bool cleanup = false;
-
-       raw_spin_lock_irq(&lock->wait_lock);
-       /*
-        * Do an unconditional try-lock, this deals with the lock stealing
-        * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
-        * sets a NULL owner.
-        *
-        * We're not interested in the return value, because the subsequent
-        * test on rt_mutex_owner() will infer that. If the trylock succeeded,
-        * we will own the lock and it will have removed the waiter. If we
-        * failed the trylock, we're still not owner and we need to remove
-        * ourselves.
-        */
-       try_to_take_rt_mutex(lock, current, waiter);
-       /*
-        * Unless we're the owner; we're still enqueued on the wait_list.
-        * So check if we became owner, if not, take us off the wait_list.
-        */
-       if (rt_mutex_owner(lock) != current) {
-               remove_waiter(lock, waiter);
-               cleanup = true;
-       }
-       /*
-        * 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_irq(&lock->wait_lock);
+       unsigned long flags;
 
-       return cleanup;
+       raw_spin_lock_irqsave(&lock->wait_lock, flags);
+       rtlock_slowlock_locked(lock);
+       raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 }
 
-#ifdef CONFIG_DEBUG_RT_MUTEXES
-void rt_mutex_debug_task_free(struct task_struct *task)
-{
-       DEBUG_LOCKS_WARN_ON(!RB_EMPTY_ROOT(&task->pi_waiters.rb_root));
-       DEBUG_LOCKS_WARN_ON(task->pi_blocked_on);
-}
-#endif
+#endif /* RT_MUTEX_BUILD_SPINLOCKS */