kernel/locking/rtmutex.c

   1 /*
   2  * RT-Mutexes: simple blocking mutual exclusion locks with PI support
   3  *
   4  * started by Ingo Molnar and Thomas Gleixner.
   5  *
   6  *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   7  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
   8  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
   9  *  Copyright (C) 2006 Esben Nielsen
  10  *
  11  *  See Documentation/rt-mutex-design.txt for details.
  12  */
  13 #include <linux/spinlock.h>
  14 #include <linux/export.h>
  15 #include <linux/sched.h>
  16 #include <linux/sched/rt.h>
  17 #include <linux/sched/deadline.h>
  18 #include <linux/timer.h>
  19
  20 #include "rtmutex_common.h"
  21
  22 /*
  23  * lock->owner state tracking:
  24  *
  25  * lock->owner holds the task_struct pointer of the owner. Bit 0
  26  * is used to keep track of the "lock has waiters" state.
  27  *
  28  * owner        bit0
  29  * NULL         0       lock is free (fast acquire possible)
  30  * NULL         1       lock is free and has waiters and the top waiter
  31  *                              is going to take the lock*
  32  * taskpointer  0       lock is held (fast release possible)
  33  * taskpointer  1       lock is held and has waiters**
  34  *
  35  * The fast atomic compare exchange based acquire and release is only
  36  * possible when bit 0 of lock->owner is 0.
  37  *
  38  * (*) It also can be a transitional state when grabbing the lock
  39  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  40  * we need to set the bit0 before looking at the lock, and the owner may be
  41  * NULL in this small time, hence this can be a transitional state.
  42  *
  43  * (**) There is a small time when bit 0 is set but there are no
  44  * waiters. This can happen when grabbing the lock in the slow path.
  45  * To prevent a cmpxchg of the owner releasing the lock, we need to
  46  * set this bit before looking at the lock.
  47  */
  48
  49 static void
  50 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
  51 {
  52         unsigned long val = (unsigned long)owner;
  53
  54         if (rt_mutex_has_waiters(lock))
  55                 val |= RT_MUTEX_HAS_WAITERS;
  56
  57         lock->owner = (struct task_struct *)val;
  58 }
  59
  60 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  61 {
  62         lock->owner = (struct task_struct *)
  63                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  64 }
  65
  66 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  67 {
  68         unsigned long owner, *p = (unsigned long *) &lock->owner;
  69
  70         if (rt_mutex_has_waiters(lock))
  71                 return;
  72
  73         /*
  74          * The rbtree has no waiters enqueued, now make sure that the
  75          * lock->owner still has the waiters bit set, otherwise the
  76          * following can happen:
  77          *
  78          * CPU 0        CPU 1           CPU2
  79          * l->owner=T1
  80          *              rt_mutex_lock(l)
  81          *              lock(l->lock)
  82          *              l->owner = T1 | HAS_WAITERS;
  83          *              enqueue(T2)
  84          *              boost()
  85          *                unlock(l->lock)
  86          *              block()
  87          *
  88          *                              rt_mutex_lock(l)
  89          *                              lock(l->lock)
  90          *                              l->owner = T1 | HAS_WAITERS;
  91          *                              enqueue(T3)
  92          *                              boost()
  93          *                                unlock(l->lock)
  94          *                              block()
  95          *              signal(->T2)    signal(->T3)
  96          *              lock(l->lock)
  97          *              dequeue(T2)
  98          *              deboost()
  99          *                unlock(l->lock)
 100          *                              lock(l->lock)
 101          *                              dequeue(T3)
 102          *                               ==> wait list is empty
 103          *                              deboost()
 104          *                               unlock(l->lock)
 105          *              lock(l->lock)
 106          *              fixup_rt_mutex_waiters()
 107          *                if (wait_list_empty(l) {
 108          *                  l->owner = owner
 109          *                  owner = l->owner & ~HAS_WAITERS;
 110          *                    ==> l->owner = T1
 111          *                }
 112          *                              lock(l->lock)
 113          * rt_mutex_unlock(l)           fixup_rt_mutex_waiters()
 114          *                                if (wait_list_empty(l) {
 115          *                                  owner = l->owner & ~HAS_WAITERS;
 116          * cmpxchg(l->owner, T1, NULL)
 117          *  ===> Success (l->owner = NULL)
 118          *
 119          *                                  l->owner = owner
 120          *                                    ==> l->owner = T1
 121          *                                }
 122          *
 123          * With the check for the waiter bit in place T3 on CPU2 will not
 124          * overwrite. All tasks fiddling with the waiters bit are
 125          * serialized by l->lock, so nothing else can modify the waiters
 126          * bit. If the bit is set then nothing can change l->owner either
 127          * so the simple RMW is safe. The cmpxchg() will simply fail if it
 128          * happens in the middle of the RMW because the waiters bit is
 129          * still set.
 130          */
 131         owner = ACCESS_ONCE(*p);
 132         if (owner & RT_MUTEX_HAS_WAITERS)
 133                 ACCESS_ONCE(*p) = owner & ~RT_MUTEX_HAS_WAITERS;
 134 }
 135
 136 /*
 137  * We can speed up the acquire/release, if the architecture
 138  * supports cmpxchg and if there's no debugging state to be set up
 139  */
 140 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
 141 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
 142 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 143 {
 144         unsigned long owner, *p = (unsigned long *) &lock->owner;
 145
 146         do {
 147                 owner = *p;
 148         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
 149 }
 150
 151 /*
 152  * Safe fastpath aware unlock:
 153  * 1) Clear the waiters bit
 154  * 2) Drop lock->wait_lock
 155  * 3) Try to unlock the lock with cmpxchg
 156  */
 157 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
 158         __releases(lock->wait_lock)
 159 {
 160         struct task_struct *owner = rt_mutex_owner(lock);
 161
 162         clear_rt_mutex_waiters(lock);
 163         raw_spin_unlock(&lock->wait_lock);
 164         /*
 165          * If a new waiter comes in between the unlock and the cmpxchg
 166          * we have two situations:
 167          *
 168          * unlock(wait_lock);
 169          *                                      lock(wait_lock);
 170          * cmpxchg(p, owner, 0) == owner
 171          *                                      mark_rt_mutex_waiters(lock);
 172          *                                      acquire(lock);
 173          * or:
 174          *
 175          * unlock(wait_lock);
 176          *                                      lock(wait_lock);
 177          *                                      mark_rt_mutex_waiters(lock);
 178          *
 179          * cmpxchg(p, owner, 0) != owner
 180          *                                      enqueue_waiter();
 181          *                                      unlock(wait_lock);
 182          * lock(wait_lock);
 183          * wake waiter();
 184          * unlock(wait_lock);
 185          *                                      lock(wait_lock);
 186          *                                      acquire(lock);
 187          */
 188         return rt_mutex_cmpxchg(lock, owner, NULL);
 189 }
 190
 191 #else
 192 # define rt_mutex_cmpxchg(l,c,n)        (0)
 193 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 194 {
 195         lock->owner = (struct task_struct *)
 196                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 197 }
 198
 199 /*
 200  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 201  */
 202 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
 203         __releases(lock->wait_lock)
 204 {
 205         lock->owner = NULL;
 206         raw_spin_unlock(&lock->wait_lock);
 207         return true;
 208 }
 209 #endif
 210
 211 static inline int
 212 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
 213                      struct rt_mutex_waiter *right)
 214 {
 215         if (left->prio < right->prio)
 216                 return 1;
 217
 218         /*
 219          * If both waiters have dl_prio(), we check the deadlines of the
 220          * associated tasks.
 221          * If left waiter has a dl_prio(), and we didn't return 1 above,
 222          * then right waiter has a dl_prio() too.
 223          */
 224         if (dl_prio(left->prio))
 225                 return (left->task->dl.deadline < right->task->dl.deadline);
 226
 227         return 0;
 228 }
 229
 230 static void
 231 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 232 {
 233         struct rb_node **link = &lock->waiters.rb_node;
 234         struct rb_node *parent = NULL;
 235         struct rt_mutex_waiter *entry;
 236         int leftmost = 1;
 237
 238         while (*link) {
 239                 parent = *link;
 240                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
 241                 if (rt_mutex_waiter_less(waiter, entry)) {
 242                         link = &parent->rb_left;
 243                 } else {
 244                         link = &parent->rb_right;
 245                         leftmost = 0;
 246                 }
 247         }
 248
 249         if (leftmost)
 250                 lock->waiters_leftmost = &waiter->tree_entry;
 251
 252         rb_link_node(&waiter->tree_entry, parent, link);
 253         rb_insert_color(&waiter->tree_entry, &lock->waiters);
 254 }
 255
 256 static void
 257 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 258 {
 259         if (RB_EMPTY_NODE(&waiter->tree_entry))
 260                 return;
 261
 262         if (lock->waiters_leftmost == &waiter->tree_entry)
 263                 lock->waiters_leftmost = rb_next(&waiter->tree_entry);
 264
 265         rb_erase(&waiter->tree_entry, &lock->waiters);
 266         RB_CLEAR_NODE(&waiter->tree_entry);
 267 }
 268
 269 static void
 270 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 271 {
 272         struct rb_node **link = &task->pi_waiters.rb_node;
 273         struct rb_node *parent = NULL;
 274         struct rt_mutex_waiter *entry;
 275         int leftmost = 1;
 276
 277         while (*link) {
 278                 parent = *link;
 279                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
 280                 if (rt_mutex_waiter_less(waiter, entry)) {
 281                         link = &parent->rb_left;
 282                 } else {
 283                         link = &parent->rb_right;
 284                         leftmost = 0;
 285                 }
 286         }
 287
 288         if (leftmost)
 289                 task->pi_waiters_leftmost = &waiter->pi_tree_entry;
 290
 291         rb_link_node(&waiter->pi_tree_entry, parent, link);
 292         rb_insert_color(&waiter->pi_tree_entry, &task->pi_waiters);
 293 }
 294
 295 static void
 296 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 297 {
 298         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 299                 return;
 300
 301         if (task->pi_waiters_leftmost == &waiter->pi_tree_entry)
 302                 task->pi_waiters_leftmost = rb_next(&waiter->pi_tree_entry);
 303
 304         rb_erase(&waiter->pi_tree_entry, &task->pi_waiters);
 305         RB_CLEAR_NODE(&waiter->pi_tree_entry);
 306 }
 307
 308 /*
 309  * Calculate task priority from the waiter tree priority
 310  *
 311  * Return task->normal_prio when the waiter tree is empty or when
 312  * the waiter is not allowed to do priority boosting
 313  */
 314 int rt_mutex_getprio(struct task_struct *task)
 315 {
 316         if (likely(!task_has_pi_waiters(task)))
 317                 return task->normal_prio;
 318
 319         return min(task_top_pi_waiter(task)->prio,
 320                    task->normal_prio);
 321 }
 322
 323 struct task_struct *rt_mutex_get_top_task(struct task_struct *task)
 324 {
 325         if (likely(!task_has_pi_waiters(task)))
 326                 return NULL;
 327
 328         return task_top_pi_waiter(task)->task;
 329 }
 330
 331 /*
 332  * Called by sched_setscheduler() to get the priority which will be
 333  * effective after the change.
 334  */
 335 int rt_mutex_get_effective_prio(struct task_struct *task, int newprio)
 336 {
 337         if (!task_has_pi_waiters(task))
 338                 return newprio;
 339
 340         if (task_top_pi_waiter(task)->task->prio <= newprio)
 341                 return task_top_pi_waiter(task)->task->prio;
 342         return newprio;
 343 }
 344
 345 /*
 346  * Adjust the priority of a task, after its pi_waiters got modified.
 347  *
 348  * This can be both boosting and unboosting. task->pi_lock must be held.
 349  */
 350 static void __rt_mutex_adjust_prio(struct task_struct *task)
 351 {
 352         int prio = rt_mutex_getprio(task);
 353
 354         if (task->prio != prio || dl_prio(prio))
 355                 rt_mutex_setprio(task, prio);
 356 }
 357
 358 /*
 359  * Adjust task priority (undo boosting). Called from the exit path of
 360  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 361  *
 362  * (Note: We do this outside of the protection of lock->wait_lock to
 363  * allow the lock to be taken while or before we readjust the priority
 364  * of task. We do not use the spin_xx_mutex() variants here as we are
 365  * outside of the debug path.)
 366  */
 367 static void rt_mutex_adjust_prio(struct task_struct *task)
 368 {
 369         unsigned long flags;
 370
 371         raw_spin_lock_irqsave(&task->pi_lock, flags);
 372         __rt_mutex_adjust_prio(task);
 373         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 374 }
 375
 376 /*
 377  * Max number of times we'll walk the boosting chain:
 378  */
 379 int max_lock_depth = 1024;
 380
 381 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 382 {
 383         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 384 }
 385
 386 /*
 387  * Adjust the priority chain. Also used for deadlock detection.
 388  * Decreases task's usage by one - may thus free the task.
 389  *
 390  * @task:       the task owning the mutex (owner) for which a chain walk is
 391  *              probably needed
 392  * @deadlock_detect: do we have to carry out deadlock detection?
 393  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 394  *              things for a task that has just got its priority adjusted, and
 395  *              is waiting on a mutex)
 396  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 397  *              we dropped its pi_lock. Is never dereferenced, only used for
 398  *              comparison to detect lock chain changes.
 399  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 400  *              its priority to the mutex owner (can be NULL in the case
 401  *              depicted above or if the top waiter is gone away and we are
 402  *              actually deboosting the owner)
 403  * @top_task:   the current top waiter
 404  *
 405  * Returns 0 or -EDEADLK.
 406  */
 407 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 408                                       int deadlock_detect,
 409                                       struct rt_mutex *orig_lock,
 410                                       struct rt_mutex *next_lock,
 411                                       struct rt_mutex_waiter *orig_waiter,
 412                                       struct task_struct *top_task)
 413 {
 414         struct rt_mutex *lock;
 415         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 416         int detect_deadlock, ret = 0, depth = 0;
 417         unsigned long flags;
 418
 419         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 420                                                          deadlock_detect);
 421
 422         /*
 423          * The (de)boosting is a step by step approach with a lot of
 424          * pitfalls. We want this to be preemptible and we want hold a
 425          * maximum of two locks per step. So we have to check
 426          * carefully whether things change under us.
 427          */
 428  again:
 429         if (++depth > max_lock_depth) {
 430                 static int prev_max;
 431
 432                 /*
 433                  * Print this only once. If the admin changes the limit,
 434                  * print a new message when reaching the limit again.
 435                  */
 436                 if (prev_max != max_lock_depth) {
 437                         prev_max = max_lock_depth;
 438                         printk(KERN_WARNING "Maximum lock depth %d reached "
 439                                "task: %s (%d)\n", max_lock_depth,
 440                                top_task->comm, task_pid_nr(top_task));
 441                 }
 442                 put_task_struct(task);
 443
 444                 return -EDEADLK;
 445         }
 446  retry:
 447         /*
 448          * Task can not go away as we did a get_task() before !
 449          */
 450         raw_spin_lock_irqsave(&task->pi_lock, flags);
 451
 452         waiter = task->pi_blocked_on;
 453         /*
 454          * Check whether the end of the boosting chain has been
 455          * reached or the state of the chain has changed while we
 456          * dropped the locks.
 457          */
 458         if (!waiter)
 459                 goto out_unlock_pi;
 460
 461         /*
 462          * Check the orig_waiter state. After we dropped the locks,
 463          * the previous owner of the lock might have released the lock.
 464          */
 465         if (orig_waiter && !rt_mutex_owner(orig_lock))
 466                 goto out_unlock_pi;
 467
 468         /*
 469          * We dropped all locks after taking a refcount on @task, so
 470          * the task might have moved on in the lock chain or even left
 471          * the chain completely and blocks now on an unrelated lock or
 472          * on @orig_lock.
 473          *
 474          * We stored the lock on which @task was blocked in @next_lock,
 475          * so we can detect the chain change.
 476          */
 477         if (next_lock != waiter->lock)
 478                 goto out_unlock_pi;
 479
 480         /*
 481          * Drop out, when the task has no waiters. Note,
 482          * top_waiter can be NULL, when we are in the deboosting
 483          * mode!
 484          */
 485         if (top_waiter) {
 486                 if (!task_has_pi_waiters(task))
 487                         goto out_unlock_pi;
 488                 /*
 489                  * If deadlock detection is off, we stop here if we
 490                  * are not the top pi waiter of the task.
 491                  */
 492                 if (!detect_deadlock && top_waiter != task_top_pi_waiter(task))
 493                         goto out_unlock_pi;
 494         }
 495
 496         /*
 497          * When deadlock detection is off then we check, if further
 498          * priority adjustment is necessary.
 499          */
 500         if (!detect_deadlock && waiter->prio == task->prio)
 501                 goto out_unlock_pi;
 502
 503         lock = waiter->lock;
 504         if (!raw_spin_trylock(&lock->wait_lock)) {
 505                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 506                 cpu_relax();
 507                 goto retry;
 508         }
 509
 510         /*
 511          * Deadlock detection. If the lock is the same as the original
 512          * lock which caused us to walk the lock chain or if the
 513          * current lock is owned by the task which initiated the chain
 514          * walk, we detected a deadlock.
 515          */
 516         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 517                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 518                 raw_spin_unlock(&lock->wait_lock);
 519                 ret = -EDEADLK;
 520                 goto out_unlock_pi;
 521         }
 522
 523         top_waiter = rt_mutex_top_waiter(lock);
 524
 525         /* Requeue the waiter */
 526         rt_mutex_dequeue(lock, waiter);
 527         waiter->prio = task->prio;
 528         rt_mutex_enqueue(lock, waiter);
 529
 530         /* Release the task */
 531         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 532         if (!rt_mutex_owner(lock)) {
 533                 /*
 534                  * If the requeue above changed the top waiter, then we need
 535                  * to wake the new top waiter up to try to get the lock.
 536                  */
 537
 538                 if (top_waiter != rt_mutex_top_waiter(lock))
 539                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 540                 raw_spin_unlock(&lock->wait_lock);
 541                 goto out_put_task;
 542         }
 543         put_task_struct(task);
 544
 545         /* Grab the next task */
 546         task = rt_mutex_owner(lock);
 547         get_task_struct(task);
 548         raw_spin_lock_irqsave(&task->pi_lock, flags);
 549
 550         if (waiter == rt_mutex_top_waiter(lock)) {
 551                 /* Boost the owner */
 552                 rt_mutex_dequeue_pi(task, top_waiter);
 553                 rt_mutex_enqueue_pi(task, waiter);
 554                 __rt_mutex_adjust_prio(task);
 555
 556         } else if (top_waiter == waiter) {
 557                 /* Deboost the owner */
 558                 rt_mutex_dequeue_pi(task, waiter);
 559                 waiter = rt_mutex_top_waiter(lock);
 560                 rt_mutex_enqueue_pi(task, waiter);
 561                 __rt_mutex_adjust_prio(task);
 562         }
 563
 564         /*
 565          * Check whether the task which owns the current lock is pi
 566          * blocked itself. If yes we store a pointer to the lock for
 567          * the lock chain change detection above. After we dropped
 568          * task->pi_lock next_lock cannot be dereferenced anymore.
 569          */
 570         next_lock = task_blocked_on_lock(task);
 571
 572         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 573
 574         top_waiter = rt_mutex_top_waiter(lock);
 575         raw_spin_unlock(&lock->wait_lock);
 576
 577         /*
 578          * We reached the end of the lock chain. Stop right here. No
 579          * point to go back just to figure that out.
 580          */
 581         if (!next_lock)
 582                 goto out_put_task;
 583
 584         if (!detect_deadlock && waiter != top_waiter)
 585                 goto out_put_task;
 586
 587         goto again;
 588
 589  out_unlock_pi:
 590         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 591  out_put_task:
 592         put_task_struct(task);
 593
 594         return ret;
 595 }
 596
 597 /*
 598  * Try to take an rt-mutex
 599  *
 600  * Must be called with lock->wait_lock held.
 601  *
 602  * @lock:   the lock to be acquired.
 603  * @task:   the task which wants to acquire the lock
 604  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 605  */
 606 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 607                 struct rt_mutex_waiter *waiter)
 608 {
 609         /*
 610          * We have to be careful here if the atomic speedups are
 611          * enabled, such that, when
 612          *  - no other waiter is on the lock
 613          *  - the lock has been released since we did the cmpxchg
 614          * the lock can be released or taken while we are doing the
 615          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 616          *
 617          * The atomic acquire/release aware variant of
 618          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 619          * the WAITERS bit, the atomic release / acquire can not
 620          * happen anymore and lock->wait_lock protects us from the
 621          * non-atomic case.
 622          *
 623          * Note, that this might set lock->owner =
 624          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 625          * any more. This is fixed up when we take the ownership.
 626          * This is the transitional state explained at the top of this file.
 627          */
 628         mark_rt_mutex_waiters(lock);
 629
 630         if (rt_mutex_owner(lock))
 631                 return 0;
 632
 633         /*
 634          * It will get the lock because of one of these conditions:
 635          * 1) there is no waiter
 636          * 2) higher priority than waiters
 637          * 3) it is top waiter
 638          */
 639         if (rt_mutex_has_waiters(lock)) {
 640                 if (task->prio >= rt_mutex_top_waiter(lock)->prio) {
 641                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 642                                 return 0;
 643                 }
 644         }
 645
 646         if (waiter || rt_mutex_has_waiters(lock)) {
 647                 unsigned long flags;
 648                 struct rt_mutex_waiter *top;
 649
 650                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 651
 652                 /* remove the queued waiter. */
 653                 if (waiter) {
 654                         rt_mutex_dequeue(lock, waiter);
 655                         task->pi_blocked_on = NULL;
 656                 }
 657
 658                 /*
 659                  * We have to enqueue the top waiter(if it exists) into
 660                  * task->pi_waiters list.
 661                  */
 662                 if (rt_mutex_has_waiters(lock)) {
 663                         top = rt_mutex_top_waiter(lock);
 664                         rt_mutex_enqueue_pi(task, top);
 665                 }
 666                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 667         }
 668
 669         /* We got the lock. */
 670         debug_rt_mutex_lock(lock);
 671
 672         rt_mutex_set_owner(lock, task);
 673
 674         rt_mutex_deadlock_account_lock(lock, task);
 675
 676         return 1;
 677 }
 678
 679 /*
 680  * Task blocks on lock.
 681  *
 682  * Prepare waiter and propagate pi chain
 683  *
 684  * This must be called with lock->wait_lock held.
 685  */
 686 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 687                                    struct rt_mutex_waiter *waiter,
 688                                    struct task_struct *task,
 689                                    int detect_deadlock)
 690 {
 691         struct task_struct *owner = rt_mutex_owner(lock);
 692         struct rt_mutex_waiter *top_waiter = waiter;
 693         struct rt_mutex *next_lock;
 694         int chain_walk = 0, res;
 695         unsigned long flags;
 696
 697         /*
 698          * Early deadlock detection. We really don't want the task to
 699          * enqueue on itself just to untangle the mess later. It's not
 700          * only an optimization. We drop the locks, so another waiter
 701          * can come in before the chain walk detects the deadlock. So
 702          * the other will detect the deadlock and return -EDEADLOCK,
 703          * which is wrong, as the other waiter is not in a deadlock
 704          * situation.
 705          */
 706         if (owner == task)
 707                 return -EDEADLK;
 708
 709         raw_spin_lock_irqsave(&task->pi_lock, flags);
 710         __rt_mutex_adjust_prio(task);
 711         waiter->task = task;
 712         waiter->lock = lock;
 713         waiter->prio = task->prio;
 714
 715         /* Get the top priority waiter on the lock */
 716         if (rt_mutex_has_waiters(lock))
 717                 top_waiter = rt_mutex_top_waiter(lock);
 718         rt_mutex_enqueue(lock, waiter);
 719
 720         task->pi_blocked_on = waiter;
 721
 722         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 723
 724         if (!owner)
 725                 return 0;
 726
 727         raw_spin_lock_irqsave(&owner->pi_lock, flags);
 728         if (waiter == rt_mutex_top_waiter(lock)) {
 729                 rt_mutex_dequeue_pi(owner, top_waiter);
 730                 rt_mutex_enqueue_pi(owner, waiter);
 731
 732                 __rt_mutex_adjust_prio(owner);
 733                 if (owner->pi_blocked_on)
 734                         chain_walk = 1;
 735         } else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock)) {
 736                 chain_walk = 1;
 737         }
 738
 739         /* Store the lock on which owner is blocked or NULL */
 740         next_lock = task_blocked_on_lock(owner);
 741
 742         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 743         /*
 744          * Even if full deadlock detection is on, if the owner is not
 745          * blocked itself, we can avoid finding this out in the chain
 746          * walk.
 747          */
 748         if (!chain_walk || !next_lock)
 749                 return 0;
 750
 751         /*
 752          * The owner can't disappear while holding a lock,
 753          * so the owner struct is protected by wait_lock.
 754          * Gets dropped in rt_mutex_adjust_prio_chain()!
 755          */
 756         get_task_struct(owner);
 757
 758         raw_spin_unlock(&lock->wait_lock);
 759
 760         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock,
 761                                          next_lock, waiter, task);
 762
 763         raw_spin_lock(&lock->wait_lock);
 764
 765         return res;
 766 }
 767
 768 /*
 769  * Wake up the next waiter on the lock.
 770  *
 771  * Remove the top waiter from the current tasks pi waiter list and
 772  * wake it up.
 773  *
 774  * Called with lock->wait_lock held.
 775  */
 776 static void wakeup_next_waiter(struct rt_mutex *lock)
 777 {
 778         struct rt_mutex_waiter *waiter;
 779         unsigned long flags;
 780
 781         raw_spin_lock_irqsave(&current->pi_lock, flags);
 782
 783         waiter = rt_mutex_top_waiter(lock);
 784
 785         /*
 786          * Remove it from current->pi_waiters. We do not adjust a
 787          * possible priority boost right now. We execute wakeup in the
 788          * boosted mode and go back to normal after releasing
 789          * lock->wait_lock.
 790          */
 791         rt_mutex_dequeue_pi(current, waiter);
 792
 793         /*
 794          * As we are waking up the top waiter, and the waiter stays
 795          * queued on the lock until it gets the lock, this lock
 796          * obviously has waiters. Just set the bit here and this has
 797          * the added benefit of forcing all new tasks into the
 798          * slow path making sure no task of lower priority than
 799          * the top waiter can steal this lock.
 800          */
 801         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
 802
 803         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 804
 805         /*
 806          * It's safe to dereference waiter as it cannot go away as
 807          * long as we hold lock->wait_lock. The waiter task needs to
 808          * acquire it in order to dequeue the waiter.
 809          */
 810         wake_up_process(waiter->task);
 811 }
 812
 813 /*
 814  * Remove a waiter from a lock and give up
 815  *
 816  * Must be called with lock->wait_lock held and
 817  * have just failed to try_to_take_rt_mutex().
 818  */
 819 static void remove_waiter(struct rt_mutex *lock,
 820                           struct rt_mutex_waiter *waiter)
 821 {
 822         int first = (waiter == rt_mutex_top_waiter(lock));
 823         struct task_struct *owner = rt_mutex_owner(lock);
 824         struct rt_mutex *next_lock = NULL;
 825         unsigned long flags;
 826
 827         raw_spin_lock_irqsave(&current->pi_lock, flags);
 828         rt_mutex_dequeue(lock, waiter);
 829         current->pi_blocked_on = NULL;
 830         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 831
 832         if (!owner)
 833                 return;
 834
 835         if (first) {
 836
 837                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 838
 839                 rt_mutex_dequeue_pi(owner, waiter);
 840
 841                 if (rt_mutex_has_waiters(lock)) {
 842                         struct rt_mutex_waiter *next;
 843
 844                         next = rt_mutex_top_waiter(lock);
 845                         rt_mutex_enqueue_pi(owner, next);
 846                 }
 847                 __rt_mutex_adjust_prio(owner);
 848
 849                 /* Store the lock on which owner is blocked or NULL */
 850                 next_lock = task_blocked_on_lock(owner);
 851
 852                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 853         }
 854
 855         if (!next_lock)
 856                 return;
 857
 858         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 859         get_task_struct(owner);
 860
 861         raw_spin_unlock(&lock->wait_lock);
 862
 863         rt_mutex_adjust_prio_chain(owner, 0, lock, next_lock, NULL, current);
 864
 865         raw_spin_lock(&lock->wait_lock);
 866 }
 867
 868 /*
 869  * Recheck the pi chain, in case we got a priority setting
 870  *
 871  * Called from sched_setscheduler
 872  */
 873 void rt_mutex_adjust_pi(struct task_struct *task)
 874 {
 875         struct rt_mutex_waiter *waiter;
 876         struct rt_mutex *next_lock;
 877         unsigned long flags;
 878
 879         raw_spin_lock_irqsave(&task->pi_lock, flags);
 880
 881         waiter = task->pi_blocked_on;
 882         if (!waiter || (waiter->prio == task->prio &&
 883                         !dl_prio(task->prio))) {
 884                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 885                 return;
 886         }
 887         next_lock = waiter->lock;
 888         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 889
 890         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 891         get_task_struct(task);
 892
 893         rt_mutex_adjust_prio_chain(task, 0, NULL, next_lock, NULL, task);
 894 }
 895
 896 /**
 897  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 898  * @lock:                the rt_mutex to take
 899  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 900  *                       or TASK_UNINTERRUPTIBLE)
 901  * @timeout:             the pre-initialized and started timer, or NULL for none
 902  * @waiter:              the pre-initialized rt_mutex_waiter
 903  *
 904  * lock->wait_lock must be held by the caller.
 905  */
 906 static int __sched
 907 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 908                     struct hrtimer_sleeper *timeout,
 909                     struct rt_mutex_waiter *waiter)
 910 {
 911         int ret = 0;
 912
 913         for (;;) {
 914                 /* Try to acquire the lock: */
 915                 if (try_to_take_rt_mutex(lock, current, waiter))
 916                         break;
 917
 918                 /*
 919                  * TASK_INTERRUPTIBLE checks for signals and
 920                  * timeout. Ignored otherwise.
 921                  */
 922                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 923                         /* Signal pending? */
 924                         if (signal_pending(current))
 925                                 ret = -EINTR;
 926                         if (timeout && !timeout->task)
 927                                 ret = -ETIMEDOUT;
 928                         if (ret)
 929                                 break;
 930                 }
 931
 932                 raw_spin_unlock(&lock->wait_lock);
 933
 934                 debug_rt_mutex_print_deadlock(waiter);
 935
 936                 schedule_rt_mutex(lock);
 937
 938                 raw_spin_lock(&lock->wait_lock);
 939                 set_current_state(state);
 940         }
 941
 942         return ret;
 943 }
 944
 945 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
 946                                      struct rt_mutex_waiter *w)
 947 {
 948         /*
 949          * If the result is not -EDEADLOCK or the caller requested
 950          * deadlock detection, nothing to do here.
 951          */
 952         if (res != -EDEADLOCK || detect_deadlock)
 953                 return;
 954
 955         /*
 956          * Yell lowdly and stop the task right here.
 957          */
 958         rt_mutex_print_deadlock(w);
 959         while (1) {
 960                 set_current_state(TASK_INTERRUPTIBLE);
 961                 schedule();
 962         }
 963 }
 964
 965 /*
 966  * Slow path lock function:
 967  */
 968 static int __sched
 969 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 970                   struct hrtimer_sleeper *timeout,
 971                   int detect_deadlock)
 972 {
 973         struct rt_mutex_waiter waiter;
 974         int ret = 0;
 975
 976         debug_rt_mutex_init_waiter(&waiter);
 977         RB_CLEAR_NODE(&waiter.pi_tree_entry);
 978         RB_CLEAR_NODE(&waiter.tree_entry);
 979
 980         raw_spin_lock(&lock->wait_lock);
 981
 982         /* Try to acquire the lock again: */
 983         if (try_to_take_rt_mutex(lock, current, NULL)) {
 984                 raw_spin_unlock(&lock->wait_lock);
 985                 return 0;
 986         }
 987
 988         set_current_state(state);
 989
 990         /* Setup the timer, when timeout != NULL */
 991         if (unlikely(timeout)) {
 992                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 993                 if (!hrtimer_active(&timeout->timer))
 994                         timeout->task = NULL;
 995         }
 996
 997         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 998
 999         if (likely(!ret))
1000                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1001
1002         set_current_state(TASK_RUNNING);
1003
1004         if (unlikely(ret)) {
1005                 if (rt_mutex_has_waiters(lock))
1006                         remove_waiter(lock, &waiter);
1007                 rt_mutex_handle_deadlock(ret, detect_deadlock, &waiter);
1008         }
1009
1010         /*
1011          * try_to_take_rt_mutex() sets the waiter bit
1012          * unconditionally. We might have to fix that up.
1013          */
1014         fixup_rt_mutex_waiters(lock);
1015
1016         raw_spin_unlock(&lock->wait_lock);
1017
1018         /* Remove pending timer: */
1019         if (unlikely(timeout))
1020                 hrtimer_cancel(&timeout->timer);
1021
1022         debug_rt_mutex_free_waiter(&waiter);
1023
1024         return ret;
1025 }
1026
1027 /*
1028  * Slow path try-lock function:
1029  */
1030 static inline int
1031 rt_mutex_slowtrylock(struct rt_mutex *lock)
1032 {
1033         int ret = 0;
1034
1035         raw_spin_lock(&lock->wait_lock);
1036
1037         if (likely(rt_mutex_owner(lock) != current)) {
1038
1039                 ret = try_to_take_rt_mutex(lock, current, NULL);
1040                 /*
1041                  * try_to_take_rt_mutex() sets the lock waiters
1042                  * bit unconditionally. Clean this up.
1043                  */
1044                 fixup_rt_mutex_waiters(lock);
1045         }
1046
1047         raw_spin_unlock(&lock->wait_lock);
1048
1049         return ret;
1050 }
1051
1052 /*
1053  * Slow path to release a rt-mutex:
1054  */
1055 static void __sched
1056 rt_mutex_slowunlock(struct rt_mutex *lock)
1057 {
1058         raw_spin_lock(&lock->wait_lock);
1059
1060         debug_rt_mutex_unlock(lock);
1061
1062         rt_mutex_deadlock_account_unlock(current);
1063
1064         /*
1065          * We must be careful here if the fast path is enabled. If we
1066          * have no waiters queued we cannot set owner to NULL here
1067          * because of:
1068          *
1069          * foo->lock->owner = NULL;
1070          *                      rtmutex_lock(foo->lock);   <- fast path
1071          *                      free = atomic_dec_and_test(foo->refcnt);
1072          *                      rtmutex_unlock(foo->lock); <- fast path
1073          *                      if (free)
1074          *                              kfree(foo);
1075          * raw_spin_unlock(foo->lock->wait_lock);
1076          *
1077          * So for the fastpath enabled kernel:
1078          *
1079          * Nothing can set the waiters bit as long as we hold
1080          * lock->wait_lock. So we do the following sequence:
1081          *
1082          *      owner = rt_mutex_owner(lock);
1083          *      clear_rt_mutex_waiters(lock);
1084          *      raw_spin_unlock(&lock->wait_lock);
1085          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1086          *              return;
1087          *      goto retry;
1088          *
1089          * The fastpath disabled variant is simple as all access to
1090          * lock->owner is serialized by lock->wait_lock:
1091          *
1092          *      lock->owner = NULL;
1093          *      raw_spin_unlock(&lock->wait_lock);
1094          */
1095         while (!rt_mutex_has_waiters(lock)) {
1096                 /* Drops lock->wait_lock ! */
1097                 if (unlock_rt_mutex_safe(lock) == true)
1098                         return;
1099                 /* Relock the rtmutex and try again */
1100                 raw_spin_lock(&lock->wait_lock);
1101         }
1102
1103         /*
1104          * The wakeup next waiter path does not suffer from the above
1105          * race. See the comments there.
1106          */
1107         wakeup_next_waiter(lock);
1108
1109         raw_spin_unlock(&lock->wait_lock);
1110
1111         /* Undo pi boosting if necessary: */
1112         rt_mutex_adjust_prio(current);
1113 }
1114
1115 /*
1116  * debug aware fast / slowpath lock,trylock,unlock
1117  *
1118  * The atomic acquire/release ops are compiled away, when either the
1119  * architecture does not support cmpxchg or when debugging is enabled.
1120  */
1121 static inline int
1122 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1123                   int detect_deadlock,
1124                   int (*slowfn)(struct rt_mutex *lock, int state,
1125                                 struct hrtimer_sleeper *timeout,
1126                                 int detect_deadlock))
1127 {
1128         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1129                 rt_mutex_deadlock_account_lock(lock, current);
1130                 return 0;
1131         } else
1132                 return slowfn(lock, state, NULL, detect_deadlock);
1133 }
1134
1135 static inline int
1136 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1137                         struct hrtimer_sleeper *timeout, int detect_deadlock,
1138                         int (*slowfn)(struct rt_mutex *lock, int state,
1139                                       struct hrtimer_sleeper *timeout,
1140                                       int detect_deadlock))
1141 {
1142         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1143                 rt_mutex_deadlock_account_lock(lock, current);
1144                 return 0;
1145         } else
1146                 return slowfn(lock, state, timeout, detect_deadlock);
1147 }
1148
1149 static inline int
1150 rt_mutex_fasttrylock(struct rt_mutex *lock,
1151                      int (*slowfn)(struct rt_mutex *lock))
1152 {
1153         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
1154                 rt_mutex_deadlock_account_lock(lock, current);
1155                 return 1;
1156         }
1157         return slowfn(lock);
1158 }
1159
1160 static inline void
1161 rt_mutex_fastunlock(struct rt_mutex *lock,
1162                     void (*slowfn)(struct rt_mutex *lock))
1163 {
1164         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
1165                 rt_mutex_deadlock_account_unlock(current);
1166         else
1167                 slowfn(lock);
1168 }
1169
1170 /**
1171  * rt_mutex_lock - lock a rt_mutex
1172  *
1173  * @lock: the rt_mutex to be locked
1174  */
1175 void __sched rt_mutex_lock(struct rt_mutex *lock)
1176 {
1177         might_sleep();
1178
1179         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
1180 }
1181 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1182
1183 /**
1184  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1185  *
1186  * @lock:               the rt_mutex to be locked
1187  * @detect_deadlock:    deadlock detection on/off
1188  *
1189  * Returns:
1190  *  0           on success
1191  * -EINTR       when interrupted by a signal
1192  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
1193  */
1194 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
1195                                                  int detect_deadlock)
1196 {
1197         might_sleep();
1198
1199         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
1200                                  detect_deadlock, rt_mutex_slowlock);
1201 }
1202 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1203
1204 /**
1205  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1206  *                      the timeout structure is provided
1207  *                      by the caller
1208  *
1209  * @lock:               the rt_mutex to be locked
1210  * @timeout:            timeout structure or NULL (no timeout)
1211  * @detect_deadlock:    deadlock detection on/off
1212  *
1213  * Returns:
1214  *  0           on success
1215  * -EINTR       when interrupted by a signal
1216  * -ETIMEDOUT   when the timeout expired
1217  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
1218  */
1219 int
1220 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
1221                     int detect_deadlock)
1222 {
1223         might_sleep();
1224
1225         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1226                                        detect_deadlock, rt_mutex_slowlock);
1227 }
1228 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1229
1230 /**
1231  * rt_mutex_trylock - try to lock a rt_mutex
1232  *
1233  * @lock:       the rt_mutex to be locked
1234  *
1235  * Returns 1 on success and 0 on contention
1236  */
1237 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1238 {
1239         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1240 }
1241 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1242
1243 /**
1244  * rt_mutex_unlock - unlock a rt_mutex
1245  *
1246  * @lock: the rt_mutex to be unlocked
1247  */
1248 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1249 {
1250         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1251 }
1252 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1253
1254 /**
1255  * rt_mutex_destroy - mark a mutex unusable
1256  * @lock: the mutex to be destroyed
1257  *
1258  * This function marks the mutex uninitialized, and any subsequent
1259  * use of the mutex is forbidden. The mutex must not be locked when
1260  * this function is called.
1261  */
1262 void rt_mutex_destroy(struct rt_mutex *lock)
1263 {
1264         WARN_ON(rt_mutex_is_locked(lock));
1265 #ifdef CONFIG_DEBUG_RT_MUTEXES
1266         lock->magic = NULL;
1267 #endif
1268 }
1269
1270 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1271
1272 /**
1273  * __rt_mutex_init - initialize the rt lock
1274  *
1275  * @lock: the rt lock to be initialized
1276  *
1277  * Initialize the rt lock to unlocked state.
1278  *
1279  * Initializing of a locked rt lock is not allowed
1280  */
1281 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1282 {
1283         lock->owner = NULL;
1284         raw_spin_lock_init(&lock->wait_lock);
1285         lock->waiters = RB_ROOT;
1286         lock->waiters_leftmost = NULL;
1287
1288         debug_rt_mutex_init(lock, name);
1289 }
1290 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1291
1292 /**
1293  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1294  *                              proxy owner
1295  *
1296  * @lock:       the rt_mutex to be locked
1297  * @proxy_owner:the task to set as owner
1298  *
1299  * No locking. Caller has to do serializing itself
1300  * Special API call for PI-futex support
1301  */
1302 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1303                                 struct task_struct *proxy_owner)
1304 {
1305         __rt_mutex_init(lock, NULL);
1306         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1307         rt_mutex_set_owner(lock, proxy_owner);
1308         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1309 }
1310
1311 /**
1312  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1313  *
1314  * @lock:       the rt_mutex to be locked
1315  *
1316  * No locking. Caller has to do serializing itself
1317  * Special API call for PI-futex support
1318  */
1319 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1320                            struct task_struct *proxy_owner)
1321 {
1322         debug_rt_mutex_proxy_unlock(lock);
1323         rt_mutex_set_owner(lock, NULL);
1324         rt_mutex_deadlock_account_unlock(proxy_owner);
1325 }
1326
1327 /**
1328  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1329  * @lock:               the rt_mutex to take
1330  * @waiter:             the pre-initialized rt_mutex_waiter
1331  * @task:               the task to prepare
1332  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1333  *
1334  * Returns:
1335  *  0 - task blocked on lock
1336  *  1 - acquired the lock for task, caller should wake it up
1337  * <0 - error
1338  *
1339  * Special API call for FUTEX_REQUEUE_PI support.
1340  */
1341 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1342                               struct rt_mutex_waiter *waiter,
1343                               struct task_struct *task, int detect_deadlock)
1344 {
1345         int ret;
1346
1347         raw_spin_lock(&lock->wait_lock);
1348
1349         if (try_to_take_rt_mutex(lock, task, NULL)) {
1350                 raw_spin_unlock(&lock->wait_lock);
1351                 return 1;
1352         }
1353
1354         /* We enforce deadlock detection for futexes */
1355         ret = task_blocks_on_rt_mutex(lock, waiter, task, 1);
1356
1357         if (ret && !rt_mutex_owner(lock)) {
1358                 /*
1359                  * Reset the return value. We might have
1360                  * returned with -EDEADLK and the owner
1361                  * released the lock while we were walking the
1362                  * pi chain.  Let the waiter sort it out.
1363                  */
1364                 ret = 0;
1365         }
1366
1367         if (unlikely(ret))
1368                 remove_waiter(lock, waiter);
1369
1370         raw_spin_unlock(&lock->wait_lock);
1371
1372         debug_rt_mutex_print_deadlock(waiter);
1373
1374         return ret;
1375 }
1376
1377 /**
1378  * rt_mutex_next_owner - return the next owner of the lock
1379  *
1380  * @lock: the rt lock query
1381  *
1382  * Returns the next owner of the lock or NULL
1383  *
1384  * Caller has to serialize against other accessors to the lock
1385  * itself.
1386  *
1387  * Special API call for PI-futex support
1388  */
1389 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1390 {
1391         if (!rt_mutex_has_waiters(lock))
1392                 return NULL;
1393
1394         return rt_mutex_top_waiter(lock)->task;
1395 }
1396
1397 /**
1398  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1399  * @lock:               the rt_mutex we were woken on
1400  * @to:                 the timeout, null if none. hrtimer should already have
1401  *                      been started.
1402  * @waiter:             the pre-initialized rt_mutex_waiter
1403  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1404  *
1405  * Complete the lock acquisition started our behalf by another thread.
1406  *
1407  * Returns:
1408  *  0 - success
1409  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1410  *
1411  * Special API call for PI-futex requeue support
1412  */
1413 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1414                                struct hrtimer_sleeper *to,
1415                                struct rt_mutex_waiter *waiter,
1416                                int detect_deadlock)
1417 {
1418         int ret;
1419
1420         raw_spin_lock(&lock->wait_lock);
1421
1422         set_current_state(TASK_INTERRUPTIBLE);
1423
1424         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1425
1426         set_current_state(TASK_RUNNING);
1427
1428         if (unlikely(ret))
1429                 remove_waiter(lock, waiter);
1430
1431         /*
1432          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1433          * have to fix that up.
1434          */
1435         fixup_rt_mutex_waiters(lock);
1436
1437         raw_spin_unlock(&lock->wait_lock);
1438
1439         return ret;
1440 }