kernel/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/module.h>
  15 #include <linux/sched.h>
  16 #include <linux/timer.h>
  17
  18 #include "rtmutex_common.h"
  19
  20 #ifdef CONFIG_DEBUG_RT_MUTEXES
  21 # include "rtmutex-debug.h"
  22 #else
  23 # include "rtmutex.h"
  24 #endif
  25
  26 /*
  27  * lock->owner state tracking:
  28  *
  29  * lock->owner holds the task_struct pointer of the owner. Bit 0 and 1
  30  * are used to keep track of the "owner is pending" and "lock has
  31  * waiters" state.
  32  *
  33  * owner        bit1    bit0
  34  * NULL         0       0       lock is free (fast acquire possible)
  35  * NULL         0       1       invalid state
  36  * NULL         1       0       Transitional State*
  37  * NULL         1       1       invalid state
  38  * taskpointer  0       0       lock is held (fast release possible)
  39  * taskpointer  0       1       task is pending owner
  40  * taskpointer  1       0       lock is held and has waiters
  41  * taskpointer  1       1       task is pending owner and lock has more waiters
  42  *
  43  * Pending ownership is assigned to the top (highest priority)
  44  * waiter of the lock, when the lock is released. The thread is woken
  45  * up and can now take the lock. Until the lock is taken (bit 0
  46  * cleared) a competing higher priority thread can steal the lock
  47  * which puts the woken up thread back on the waiters list.
  48  *
  49  * The fast atomic compare exchange based acquire and release is only
  50  * possible when bit 0 and 1 of lock->owner are 0.
  51  *
  52  * (*) There's a small time where the owner can be NULL and the
  53  * "lock has waiters" bit is set.  This can happen when grabbing the lock.
  54  * To prevent a cmpxchg of the owner releasing the lock, we need to set this
  55  * bit before looking at the lock, hence the reason this is a transitional
  56  * state.
  57  */
  58
  59 void
  60 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner,
  61                    unsigned long mask)
  62 {
  63         unsigned long val = (unsigned long)owner | mask;
  64
  65         if (rt_mutex_has_waiters(lock))
  66                 val |= RT_MUTEX_HAS_WAITERS;
  67
  68         lock->owner = (struct task_struct *)val;
  69 }
  70
  71 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  72 {
  73         lock->owner = (struct task_struct *)
  74                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  75 }
  76
  77 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  78 {
  79         if (!rt_mutex_has_waiters(lock))
  80                 clear_rt_mutex_waiters(lock);
  81 }
  82
  83 /*
  84  * Calculate task priority from the waiter list priority
  85  *
  86  * Return task->normal_prio when the waiter list is empty or when
  87  * the waiter is not allowed to do priority boosting
  88  */
  89 int rt_mutex_getprio(struct task_struct *task)
  90 {
  91         if (likely(!task_has_pi_waiters(task)))
  92                 return task->normal_prio;
  93
  94         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
  95                    task->normal_prio);
  96 }
  97
  98 /*
  99  * Adjust the priority of a task, after its pi_waiters got modified.
 100  *
 101  * This can be both boosting and unboosting. task->pi_lock must be held.
 102  */
 103 void __rt_mutex_adjust_prio(struct task_struct *task)
 104 {
 105         int prio = rt_mutex_getprio(task);
 106
 107         if (task->prio != prio)
 108                 rt_mutex_setprio(task, prio);
 109 }
 110
 111 /*
 112  * Adjust task priority (undo boosting). Called from the exit path of
 113  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 114  *
 115  * (Note: We do this outside of the protection of lock->wait_lock to
 116  * allow the lock to be taken while or before we readjust the priority
 117  * of task. We do not use the spin_xx_mutex() variants here as we are
 118  * outside of the debug path.)
 119  */
 120 static void rt_mutex_adjust_prio(struct task_struct *task)
 121 {
 122         unsigned long flags;
 123
 124         spin_lock_irqsave(&task->pi_lock, flags);
 125         __rt_mutex_adjust_prio(task);
 126         spin_unlock_irqrestore(&task->pi_lock, flags);
 127 }
 128
 129 /*
 130  * Max number of times we'll walk the boosting chain:
 131  */
 132 int max_lock_depth = 1024;
 133
 134 /*
 135  * Adjust the priority chain. Also used for deadlock detection.
 136  * Decreases task's usage by one - may thus free the task.
 137  * Returns 0 or -EDEADLK.
 138  */
 139 int rt_mutex_adjust_prio_chain(struct task_struct *task,
 140                                int deadlock_detect,
 141                                struct rt_mutex *orig_lock,
 142                                struct rt_mutex_waiter *orig_waiter,
 143                                struct task_struct *top_task)
 144 {
 145         struct rt_mutex *lock;
 146         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 147         int detect_deadlock, ret = 0, depth = 0;
 148         unsigned long flags;
 149
 150         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 151                                                          deadlock_detect);
 152
 153         /*
 154          * The (de)boosting is a step by step approach with a lot of
 155          * pitfalls. We want this to be preemptible and we want hold a
 156          * maximum of two locks per step. So we have to check
 157          * carefully whether things change under us.
 158          */
 159  again:
 160         if (++depth > max_lock_depth) {
 161                 static int prev_max;
 162
 163                 /*
 164                  * Print this only once. If the admin changes the limit,
 165                  * print a new message when reaching the limit again.
 166                  */
 167                 if (prev_max != max_lock_depth) {
 168                         prev_max = max_lock_depth;
 169                         printk(KERN_WARNING "Maximum lock depth %d reached "
 170                                "task: %s (%d)\n", max_lock_depth,
 171                                top_task->comm, top_task->pid);
 172                 }
 173                 put_task_struct(task);
 174
 175                 return deadlock_detect ? -EDEADLK : 0;
 176         }
 177  retry:
 178         /*
 179          * Task can not go away as we did a get_task() before !
 180          */
 181         spin_lock_irqsave(&task->pi_lock, flags);
 182
 183         waiter = task->pi_blocked_on;
 184         /*
 185          * Check whether the end of the boosting chain has been
 186          * reached or the state of the chain has changed while we
 187          * dropped the locks.
 188          */
 189         if (!waiter || !waiter->task)
 190                 goto out_unlock_pi;
 191
 192         /*
 193          * Check the orig_waiter state. After we dropped the locks,
 194          * the previous owner of the lock might have released the lock
 195          * and made us the pending owner:
 196          */
 197         if (orig_waiter && !orig_waiter->task)
 198                 goto out_unlock_pi;
 199
 200         /*
 201          * Drop out, when the task has no waiters. Note,
 202          * top_waiter can be NULL, when we are in the deboosting
 203          * mode!
 204          */
 205         if (top_waiter && (!task_has_pi_waiters(task) ||
 206                            top_waiter != task_top_pi_waiter(task)))
 207                 goto out_unlock_pi;
 208
 209         /*
 210          * When deadlock detection is off then we check, if further
 211          * priority adjustment is necessary.
 212          */
 213         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 214                 goto out_unlock_pi;
 215
 216         lock = waiter->lock;
 217         if (!spin_trylock(&lock->wait_lock)) {
 218                 spin_unlock_irqrestore(&task->pi_lock, flags);
 219                 cpu_relax();
 220                 goto retry;
 221         }
 222
 223         /* Deadlock detection */
 224         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 225                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 226                 spin_unlock(&lock->wait_lock);
 227                 ret = deadlock_detect ? -EDEADLK : 0;
 228                 goto out_unlock_pi;
 229         }
 230
 231         top_waiter = rt_mutex_top_waiter(lock);
 232
 233         /* Requeue the waiter */
 234         plist_del(&waiter->list_entry, &lock->wait_list);
 235         waiter->list_entry.prio = task->prio;
 236         plist_add(&waiter->list_entry, &lock->wait_list);
 237
 238         /* Release the task */
 239         spin_unlock_irqrestore(&task->pi_lock, flags);
 240         put_task_struct(task);
 241
 242         /* Grab the next task */
 243         task = rt_mutex_owner(lock);
 244         get_task_struct(task);
 245         spin_lock_irqsave(&task->pi_lock, flags);
 246
 247         if (waiter == rt_mutex_top_waiter(lock)) {
 248                 /* Boost the owner */
 249                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 250                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 251                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 252                 __rt_mutex_adjust_prio(task);
 253
 254         } else if (top_waiter == waiter) {
 255                 /* Deboost the owner */
 256                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 257                 waiter = rt_mutex_top_waiter(lock);
 258                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 259                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 260                 __rt_mutex_adjust_prio(task);
 261         }
 262
 263         spin_unlock_irqrestore(&task->pi_lock, flags);
 264
 265         top_waiter = rt_mutex_top_waiter(lock);
 266         spin_unlock(&lock->wait_lock);
 267
 268         if (!detect_deadlock && waiter != top_waiter)
 269                 goto out_put_task;
 270
 271         goto again;
 272
 273  out_unlock_pi:
 274         spin_unlock_irqrestore(&task->pi_lock, flags);
 275  out_put_task:
 276         put_task_struct(task);
 277
 278         return ret;
 279 }
 280
 281 /*
 282  * Optimization: check if we can steal the lock from the
 283  * assigned pending owner [which might not have taken the
 284  * lock yet]:
 285  */
 286 static inline int try_to_steal_lock(struct rt_mutex *lock)
 287 {
 288         struct task_struct *pendowner = rt_mutex_owner(lock);
 289         struct rt_mutex_waiter *next;
 290         unsigned long flags;
 291
 292         if (!rt_mutex_owner_pending(lock))
 293                 return 0;
 294
 295         if (pendowner == current)
 296                 return 1;
 297
 298         spin_lock_irqsave(&pendowner->pi_lock, flags);
 299         if (current->prio >= pendowner->prio) {
 300                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 301                 return 0;
 302         }
 303
 304         /*
 305          * Check if a waiter is enqueued on the pending owners
 306          * pi_waiters list. Remove it and readjust pending owners
 307          * priority.
 308          */
 309         if (likely(!rt_mutex_has_waiters(lock))) {
 310                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 311                 return 1;
 312         }
 313
 314         /* No chain handling, pending owner is not blocked on anything: */
 315         next = rt_mutex_top_waiter(lock);
 316         plist_del(&next->pi_list_entry, &pendowner->pi_waiters);
 317         __rt_mutex_adjust_prio(pendowner);
 318         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 319
 320         /*
 321          * We are going to steal the lock and a waiter was
 322          * enqueued on the pending owners pi_waiters queue. So
 323          * we have to enqueue this waiter into
 324          * current->pi_waiters list. This covers the case,
 325          * where current is boosted because it holds another
 326          * lock and gets unboosted because the booster is
 327          * interrupted, so we would delay a waiter with higher
 328          * priority as current->normal_prio.
 329          *
 330          * Note: in the rare case of a SCHED_OTHER task changing
 331          * its priority and thus stealing the lock, next->task
 332          * might be current:
 333          */
 334         if (likely(next->task != current)) {
 335                 spin_lock_irqsave(&current->pi_lock, flags);
 336                 plist_add(&next->pi_list_entry, &current->pi_waiters);
 337                 __rt_mutex_adjust_prio(current);
 338                 spin_unlock_irqrestore(&current->pi_lock, flags);
 339         }
 340         return 1;
 341 }
 342
 343 /*
 344  * Try to take an rt-mutex
 345  *
 346  * This fails
 347  * - when the lock has a real owner
 348  * - when a different pending owner exists and has higher priority than current
 349  *
 350  * Must be called with lock->wait_lock held.
 351  */
 352 static int try_to_take_rt_mutex(struct rt_mutex *lock)
 353 {
 354         /*
 355          * We have to be careful here if the atomic speedups are
 356          * enabled, such that, when
 357          *  - no other waiter is on the lock
 358          *  - the lock has been released since we did the cmpxchg
 359          * the lock can be released or taken while we are doing the
 360          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 361          *
 362          * The atomic acquire/release aware variant of
 363          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 364          * the WAITERS bit, the atomic release / acquire can not
 365          * happen anymore and lock->wait_lock protects us from the
 366          * non-atomic case.
 367          *
 368          * Note, that this might set lock->owner =
 369          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 370          * any more. This is fixed up when we take the ownership.
 371          * This is the transitional state explained at the top of this file.
 372          */
 373         mark_rt_mutex_waiters(lock);
 374
 375         if (rt_mutex_owner(lock) && !try_to_steal_lock(lock))
 376                 return 0;
 377
 378         /* We got the lock. */
 379         debug_rt_mutex_lock(lock);
 380
 381         rt_mutex_set_owner(lock, current, 0);
 382
 383         rt_mutex_deadlock_account_lock(lock, current);
 384
 385         return 1;
 386 }
 387
 388 /*
 389  * Task blocks on lock.
 390  *
 391  * Prepare waiter and propagate pi chain
 392  *
 393  * This must be called with lock->wait_lock held.
 394  */
 395 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 396                                    struct rt_mutex_waiter *waiter,
 397                                    int detect_deadlock)
 398 {
 399         struct task_struct *owner = rt_mutex_owner(lock);
 400         struct rt_mutex_waiter *top_waiter = waiter;
 401         unsigned long flags;
 402         int chain_walk = 0, res;
 403
 404         spin_lock_irqsave(&current->pi_lock, flags);
 405         __rt_mutex_adjust_prio(current);
 406         waiter->task = current;
 407         waiter->lock = lock;
 408         plist_node_init(&waiter->list_entry, current->prio);
 409         plist_node_init(&waiter->pi_list_entry, current->prio);
 410
 411         /* Get the top priority waiter on the lock */
 412         if (rt_mutex_has_waiters(lock))
 413                 top_waiter = rt_mutex_top_waiter(lock);
 414         plist_add(&waiter->list_entry, &lock->wait_list);
 415
 416         current->pi_blocked_on = waiter;
 417
 418         spin_unlock_irqrestore(&current->pi_lock, flags);
 419
 420         if (waiter == rt_mutex_top_waiter(lock)) {
 421                 spin_lock_irqsave(&owner->pi_lock, flags);
 422                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 423                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 424
 425                 __rt_mutex_adjust_prio(owner);
 426                 if (owner->pi_blocked_on)
 427                         chain_walk = 1;
 428                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 429         }
 430         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 431                 chain_walk = 1;
 432
 433         if (!chain_walk)
 434                 return 0;
 435
 436         /*
 437          * The owner can't disappear while holding a lock,
 438          * so the owner struct is protected by wait_lock.
 439          * Gets dropped in rt_mutex_adjust_prio_chain()!
 440          */
 441         get_task_struct(owner);
 442
 443         spin_unlock(&lock->wait_lock);
 444
 445         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 446                                          current);
 447
 448         spin_lock(&lock->wait_lock);
 449
 450         return res;
 451 }
 452
 453 /*
 454  * Wake up the next waiter on the lock.
 455  *
 456  * Remove the top waiter from the current tasks waiter list and from
 457  * the lock waiter list. Set it as pending owner. Then wake it up.
 458  *
 459  * Called with lock->wait_lock held.
 460  */
 461 static void wakeup_next_waiter(struct rt_mutex *lock)
 462 {
 463         struct rt_mutex_waiter *waiter;
 464         struct task_struct *pendowner;
 465         unsigned long flags;
 466
 467         spin_lock_irqsave(&current->pi_lock, flags);
 468
 469         waiter = rt_mutex_top_waiter(lock);
 470         plist_del(&waiter->list_entry, &lock->wait_list);
 471
 472         /*
 473          * Remove it from current->pi_waiters. We do not adjust a
 474          * possible priority boost right now. We execute wakeup in the
 475          * boosted mode and go back to normal after releasing
 476          * lock->wait_lock.
 477          */
 478         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 479         pendowner = waiter->task;
 480         waiter->task = NULL;
 481
 482         rt_mutex_set_owner(lock, pendowner, RT_MUTEX_OWNER_PENDING);
 483
 484         spin_unlock_irqrestore(&current->pi_lock, flags);
 485
 486         /*
 487          * Clear the pi_blocked_on variable and enqueue a possible
 488          * waiter into the pi_waiters list of the pending owner. This
 489          * prevents that in case the pending owner gets unboosted a
 490          * waiter with higher priority than pending-owner->normal_prio
 491          * is blocked on the unboosted (pending) owner.
 492          */
 493         spin_lock_irqsave(&pendowner->pi_lock, flags);
 494
 495         WARN_ON(!pendowner->pi_blocked_on);
 496         WARN_ON(pendowner->pi_blocked_on != waiter);
 497         WARN_ON(pendowner->pi_blocked_on->lock != lock);
 498
 499         pendowner->pi_blocked_on = NULL;
 500
 501         if (rt_mutex_has_waiters(lock)) {
 502                 struct rt_mutex_waiter *next;
 503
 504                 next = rt_mutex_top_waiter(lock);
 505                 plist_add(&next->pi_list_entry, &pendowner->pi_waiters);
 506         }
 507         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 508
 509         wake_up_process(pendowner);
 510 }
 511
 512 /*
 513  * Remove a waiter from a lock
 514  *
 515  * Must be called with lock->wait_lock held
 516  */
 517 void remove_waiter(struct rt_mutex *lock,
 518                    struct rt_mutex_waiter *waiter)
 519 {
 520         int first = (waiter == rt_mutex_top_waiter(lock));
 521         struct task_struct *owner = rt_mutex_owner(lock);
 522         unsigned long flags;
 523         int chain_walk = 0;
 524
 525         spin_lock_irqsave(&current->pi_lock, flags);
 526         plist_del(&waiter->list_entry, &lock->wait_list);
 527         waiter->task = NULL;
 528         current->pi_blocked_on = NULL;
 529         spin_unlock_irqrestore(&current->pi_lock, flags);
 530
 531         if (first && owner != current) {
 532
 533                 spin_lock_irqsave(&owner->pi_lock, flags);
 534
 535                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 536
 537                 if (rt_mutex_has_waiters(lock)) {
 538                         struct rt_mutex_waiter *next;
 539
 540                         next = rt_mutex_top_waiter(lock);
 541                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 542                 }
 543                 __rt_mutex_adjust_prio(owner);
 544
 545                 if (owner->pi_blocked_on)
 546                         chain_walk = 1;
 547
 548                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 549         }
 550
 551         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 552
 553         if (!chain_walk)
 554                 return;
 555
 556         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 557         get_task_struct(owner);
 558
 559         spin_unlock(&lock->wait_lock);
 560
 561         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 562
 563         spin_lock(&lock->wait_lock);
 564 }
 565
 566 /*
 567  * Recheck the pi chain, in case we got a priority setting
 568  *
 569  * Called from sched_setscheduler
 570  */
 571 void rt_mutex_adjust_pi(struct task_struct *task)
 572 {
 573         struct rt_mutex_waiter *waiter;
 574         unsigned long flags;
 575
 576         spin_lock_irqsave(&task->pi_lock, flags);
 577
 578         waiter = task->pi_blocked_on;
 579         if (!waiter || waiter->list_entry.prio == task->prio) {
 580                 spin_unlock_irqrestore(&task->pi_lock, flags);
 581                 return;
 582         }
 583
 584         spin_unlock_irqrestore(&task->pi_lock, flags);
 585
 586         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 587         get_task_struct(task);
 588         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 589 }
 590
 591 /*
 592  * Slow path lock function:
 593  */
 594 static int __sched
 595 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 596                   struct hrtimer_sleeper *timeout,
 597                   int detect_deadlock)
 598 {
 599         struct rt_mutex_waiter waiter;
 600         int ret = 0;
 601
 602         debug_rt_mutex_init_waiter(&waiter);
 603         waiter.task = NULL;
 604
 605         spin_lock(&lock->wait_lock);
 606
 607         /* Try to acquire the lock again: */
 608         if (try_to_take_rt_mutex(lock)) {
 609                 spin_unlock(&lock->wait_lock);
 610                 return 0;
 611         }
 612
 613         set_current_state(state);
 614
 615         /* Setup the timer, when timeout != NULL */
 616         if (unlikely(timeout))
 617                 hrtimer_start(&timeout->timer, timeout->timer.expires,
 618                               HRTIMER_MODE_ABS);
 619
 620         for (;;) {
 621                 /* Try to acquire the lock: */
 622                 if (try_to_take_rt_mutex(lock))
 623                         break;
 624
 625                 /*
 626                  * TASK_INTERRUPTIBLE checks for signals and
 627                  * timeout. Ignored otherwise.
 628                  */
 629                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 630                         /* Signal pending? */
 631                         if (signal_pending(current))
 632                                 ret = -EINTR;
 633                         if (timeout && !timeout->task)
 634                                 ret = -ETIMEDOUT;
 635                         if (ret)
 636                                 break;
 637                 }
 638
 639                 /*
 640                  * waiter.task is NULL the first time we come here and
 641                  * when we have been woken up by the previous owner
 642                  * but the lock got stolen by a higher prio task.
 643                  */
 644                 if (!waiter.task) {
 645                         ret = task_blocks_on_rt_mutex(lock, &waiter,
 646                                                       detect_deadlock);
 647                         /*
 648                          * If we got woken up by the owner then start loop
 649                          * all over without going into schedule to try
 650                          * to get the lock now:
 651                          */
 652                         if (unlikely(!waiter.task)) {
 653                                 /*
 654                                  * Reset the return value. We might
 655                                  * have returned with -EDEADLK and the
 656                                  * owner released the lock while we
 657                                  * were walking the pi chain.
 658                                  */
 659                                 ret = 0;
 660                                 continue;
 661                         }
 662                         if (unlikely(ret))
 663                                 break;
 664                 }
 665
 666                 spin_unlock(&lock->wait_lock);
 667
 668                 debug_rt_mutex_print_deadlock(&waiter);
 669
 670                 if (waiter.task)
 671                         schedule_rt_mutex(lock);
 672
 673                 spin_lock(&lock->wait_lock);
 674                 set_current_state(state);
 675         }
 676
 677         set_current_state(TASK_RUNNING);
 678
 679         if (unlikely(waiter.task))
 680                 remove_waiter(lock, &waiter);
 681
 682         /*
 683          * try_to_take_rt_mutex() sets the waiter bit
 684          * unconditionally. We might have to fix that up.
 685          */
 686         fixup_rt_mutex_waiters(lock);
 687
 688         spin_unlock(&lock->wait_lock);
 689
 690         /* Remove pending timer: */
 691         if (unlikely(timeout))
 692                 hrtimer_cancel(&timeout->timer);
 693
 694         /*
 695          * Readjust priority, when we did not get the lock. We might
 696          * have been the pending owner and boosted. Since we did not
 697          * take the lock, the PI boost has to go.
 698          */
 699         if (unlikely(ret))
 700                 rt_mutex_adjust_prio(current);
 701
 702         debug_rt_mutex_free_waiter(&waiter);
 703
 704         return ret;
 705 }
 706
 707 /*
 708  * Slow path try-lock function:
 709  */
 710 static inline int
 711 rt_mutex_slowtrylock(struct rt_mutex *lock)
 712 {
 713         int ret = 0;
 714
 715         spin_lock(&lock->wait_lock);
 716
 717         if (likely(rt_mutex_owner(lock) != current)) {
 718
 719                 ret = try_to_take_rt_mutex(lock);
 720                 /*
 721                  * try_to_take_rt_mutex() sets the lock waiters
 722                  * bit unconditionally. Clean this up.
 723                  */
 724                 fixup_rt_mutex_waiters(lock);
 725         }
 726
 727         spin_unlock(&lock->wait_lock);
 728
 729         return ret;
 730 }
 731
 732 /*
 733  * Slow path to release a rt-mutex:
 734  */
 735 static void __sched
 736 rt_mutex_slowunlock(struct rt_mutex *lock)
 737 {
 738         spin_lock(&lock->wait_lock);
 739
 740         debug_rt_mutex_unlock(lock);
 741
 742         rt_mutex_deadlock_account_unlock(current);
 743
 744         if (!rt_mutex_has_waiters(lock)) {
 745                 lock->owner = NULL;
 746                 spin_unlock(&lock->wait_lock);
 747                 return;
 748         }
 749
 750         wakeup_next_waiter(lock);
 751
 752         spin_unlock(&lock->wait_lock);
 753
 754         /* Undo pi boosting if necessary: */
 755         rt_mutex_adjust_prio(current);
 756 }
 757
 758 /*
 759  * debug aware fast / slowpath lock,trylock,unlock
 760  *
 761  * The atomic acquire/release ops are compiled away, when either the
 762  * architecture does not support cmpxchg or when debugging is enabled.
 763  */
 764 static inline int
 765 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 766                   int detect_deadlock,
 767                   int (*slowfn)(struct rt_mutex *lock, int state,
 768                                 struct hrtimer_sleeper *timeout,
 769                                 int detect_deadlock))
 770 {
 771         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 772                 rt_mutex_deadlock_account_lock(lock, current);
 773                 return 0;
 774         } else
 775                 return slowfn(lock, state, NULL, detect_deadlock);
 776 }
 777
 778 static inline int
 779 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 780                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 781                         int (*slowfn)(struct rt_mutex *lock, int state,
 782                                       struct hrtimer_sleeper *timeout,
 783                                       int detect_deadlock))
 784 {
 785         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 786                 rt_mutex_deadlock_account_lock(lock, current);
 787                 return 0;
 788         } else
 789                 return slowfn(lock, state, timeout, detect_deadlock);
 790 }
 791
 792 static inline int
 793 rt_mutex_fasttrylock(struct rt_mutex *lock,
 794                      int (*slowfn)(struct rt_mutex *lock))
 795 {
 796         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 797                 rt_mutex_deadlock_account_lock(lock, current);
 798                 return 1;
 799         }
 800         return slowfn(lock);
 801 }
 802
 803 static inline void
 804 rt_mutex_fastunlock(struct rt_mutex *lock,
 805                     void (*slowfn)(struct rt_mutex *lock))
 806 {
 807         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 808                 rt_mutex_deadlock_account_unlock(current);
 809         else
 810                 slowfn(lock);
 811 }
 812
 813 /**
 814  * rt_mutex_lock - lock a rt_mutex
 815  *
 816  * @lock: the rt_mutex to be locked
 817  */
 818 void __sched rt_mutex_lock(struct rt_mutex *lock)
 819 {
 820         might_sleep();
 821
 822         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 823 }
 824 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 825
 826 /**
 827  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 828  *
 829  * @lock:               the rt_mutex to be locked
 830  * @detect_deadlock:    deadlock detection on/off
 831  *
 832  * Returns:
 833  *  0           on success
 834  * -EINTR       when interrupted by a signal
 835  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 836  */
 837 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 838                                                  int detect_deadlock)
 839 {
 840         might_sleep();
 841
 842         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 843                                  detect_deadlock, rt_mutex_slowlock);
 844 }
 845 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 846
 847 /**
 848  * rt_mutex_lock_interruptible_ktime - lock a rt_mutex interruptible
 849  *                                     the timeout structure is provided
 850  *                                     by the caller
 851  *
 852  * @lock:               the rt_mutex to be locked
 853  * @timeout:            timeout structure or NULL (no timeout)
 854  * @detect_deadlock:    deadlock detection on/off
 855  *
 856  * Returns:
 857  *  0           on success
 858  * -EINTR       when interrupted by a signal
 859  * -ETIMEOUT    when the timeout expired
 860  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 861  */
 862 int
 863 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 864                     int detect_deadlock)
 865 {
 866         might_sleep();
 867
 868         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 869                                        detect_deadlock, rt_mutex_slowlock);
 870 }
 871 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 872
 873 /**
 874  * rt_mutex_trylock - try to lock a rt_mutex
 875  *
 876  * @lock:       the rt_mutex to be locked
 877  *
 878  * Returns 1 on success and 0 on contention
 879  */
 880 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 881 {
 882         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 883 }
 884 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 885
 886 /**
 887  * rt_mutex_unlock - unlock a rt_mutex
 888  *
 889  * @lock: the rt_mutex to be unlocked
 890  */
 891 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 892 {
 893         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 894 }
 895 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 896
 897 /***
 898  * rt_mutex_destroy - mark a mutex unusable
 899  * @lock: the mutex to be destroyed
 900  *
 901  * This function marks the mutex uninitialized, and any subsequent
 902  * use of the mutex is forbidden. The mutex must not be locked when
 903  * this function is called.
 904  */
 905 void rt_mutex_destroy(struct rt_mutex *lock)
 906 {
 907         WARN_ON(rt_mutex_is_locked(lock));
 908 #ifdef CONFIG_DEBUG_RT_MUTEXES
 909         lock->magic = NULL;
 910 #endif
 911 }
 912
 913 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 914
 915 /**
 916  * __rt_mutex_init - initialize the rt lock
 917  *
 918  * @lock: the rt lock to be initialized
 919  *
 920  * Initialize the rt lock to unlocked state.
 921  *
 922  * Initializing of a locked rt lock is not allowed
 923  */
 924 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 925 {
 926         lock->owner = NULL;
 927         spin_lock_init(&lock->wait_lock);
 928         plist_head_init(&lock->wait_list, &lock->wait_lock);
 929
 930         debug_rt_mutex_init(lock, name);
 931 }
 932 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 933
 934 /**
 935  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 936  *                              proxy owner
 937  *
 938  * @lock:       the rt_mutex to be locked
 939  * @proxy_owner:the task to set as owner
 940  *
 941  * No locking. Caller has to do serializing itself
 942  * Special API call for PI-futex support
 943  */
 944 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 945                                 struct task_struct *proxy_owner)
 946 {
 947         __rt_mutex_init(lock, NULL);
 948         debug_rt_mutex_proxy_lock(lock, proxy_owner);
 949         rt_mutex_set_owner(lock, proxy_owner, 0);
 950         rt_mutex_deadlock_account_lock(lock, proxy_owner);
 951 }
 952
 953 /**
 954  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 955  *
 956  * @lock:       the rt_mutex to be locked
 957  *
 958  * No locking. Caller has to do serializing itself
 959  * Special API call for PI-futex support
 960  */
 961 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 962                            struct task_struct *proxy_owner)
 963 {
 964         debug_rt_mutex_proxy_unlock(lock);
 965         rt_mutex_set_owner(lock, NULL, 0);
 966         rt_mutex_deadlock_account_unlock(proxy_owner);
 967 }
 968
 969 /**
 970  * rt_mutex_next_owner - return the next owner of the lock
 971  *
 972  * @lock: the rt lock query
 973  *
 974  * Returns the next owner of the lock or NULL
 975  *
 976  * Caller has to serialize against other accessors to the lock
 977  * itself.
 978  *
 979  * Special API call for PI-futex support
 980  */
 981 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
 982 {
 983         if (!rt_mutex_has_waiters(lock))
 984                 return NULL;
 985
 986         return rt_mutex_top_waiter(lock)->task;
 987 }