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 static 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  * We can speed up the acquire/release, if the architecture
  85  * supports cmpxchg and if there's no debugging state to be set up
  86  */
  87 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
  88 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
  89 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  90 {
  91         unsigned long owner, *p = (unsigned long *) &lock->owner;
  92
  93         do {
  94                 owner = *p;
  95         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
  96 }
  97 #else
  98 # define rt_mutex_cmpxchg(l,c,n)        (0)
  99 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 100 {
 101         lock->owner = (struct task_struct *)
 102                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 103 }
 104 #endif
 105
 106 /*
 107  * Calculate task priority from the waiter list priority
 108  *
 109  * Return task->normal_prio when the waiter list is empty or when
 110  * the waiter is not allowed to do priority boosting
 111  */
 112 int rt_mutex_getprio(struct task_struct *task)
 113 {
 114         if (likely(!task_has_pi_waiters(task)))
 115                 return task->normal_prio;
 116
 117         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
 118                    task->normal_prio);
 119 }
 120
 121 /*
 122  * Adjust the priority of a task, after its pi_waiters got modified.
 123  *
 124  * This can be both boosting and unboosting. task->pi_lock must be held.
 125  */
 126 static void __rt_mutex_adjust_prio(struct task_struct *task)
 127 {
 128         int prio = rt_mutex_getprio(task);
 129
 130         if (task->prio != prio)
 131                 rt_mutex_setprio(task, prio);
 132 }
 133
 134 /*
 135  * Adjust task priority (undo boosting). Called from the exit path of
 136  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 137  *
 138  * (Note: We do this outside of the protection of lock->wait_lock to
 139  * allow the lock to be taken while or before we readjust the priority
 140  * of task. We do not use the spin_xx_mutex() variants here as we are
 141  * outside of the debug path.)
 142  */
 143 static void rt_mutex_adjust_prio(struct task_struct *task)
 144 {
 145         unsigned long flags;
 146
 147         spin_lock_irqsave(&task->pi_lock, flags);
 148         __rt_mutex_adjust_prio(task);
 149         spin_unlock_irqrestore(&task->pi_lock, flags);
 150 }
 151
 152 /*
 153  * Max number of times we'll walk the boosting chain:
 154  */
 155 int max_lock_depth = 1024;
 156
 157 /*
 158  * Adjust the priority chain. Also used for deadlock detection.
 159  * Decreases task's usage by one - may thus free the task.
 160  * Returns 0 or -EDEADLK.
 161  */
 162 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 163                                       int deadlock_detect,
 164                                       struct rt_mutex *orig_lock,
 165                                       struct rt_mutex_waiter *orig_waiter,
 166                                       struct task_struct *top_task)
 167 {
 168         struct rt_mutex *lock;
 169         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 170         int detect_deadlock, ret = 0, depth = 0;
 171         unsigned long flags;
 172
 173         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 174                                                          deadlock_detect);
 175
 176         /*
 177          * The (de)boosting is a step by step approach with a lot of
 178          * pitfalls. We want this to be preemptible and we want hold a
 179          * maximum of two locks per step. So we have to check
 180          * carefully whether things change under us.
 181          */
 182  again:
 183         if (++depth > max_lock_depth) {
 184                 static int prev_max;
 185
 186                 /*
 187                  * Print this only once. If the admin changes the limit,
 188                  * print a new message when reaching the limit again.
 189                  */
 190                 if (prev_max != max_lock_depth) {
 191                         prev_max = max_lock_depth;
 192                         printk(KERN_WARNING "Maximum lock depth %d reached "
 193                                "task: %s (%d)\n", max_lock_depth,
 194                                top_task->comm, top_task->pid);
 195                 }
 196                 put_task_struct(task);
 197
 198                 return deadlock_detect ? -EDEADLK : 0;
 199         }
 200  retry:
 201         /*
 202          * Task can not go away as we did a get_task() before !
 203          */
 204         spin_lock_irqsave(&task->pi_lock, flags);
 205
 206         waiter = task->pi_blocked_on;
 207         /*
 208          * Check whether the end of the boosting chain has been
 209          * reached or the state of the chain has changed while we
 210          * dropped the locks.
 211          */
 212         if (!waiter || !waiter->task)
 213                 goto out_unlock_pi;
 214
 215         if (top_waiter && (!task_has_pi_waiters(task) ||
 216                            top_waiter != task_top_pi_waiter(task)))
 217                 goto out_unlock_pi;
 218
 219         /*
 220          * When deadlock detection is off then we check, if further
 221          * priority adjustment is necessary.
 222          */
 223         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 224                 goto out_unlock_pi;
 225
 226         lock = waiter->lock;
 227         if (!spin_trylock(&lock->wait_lock)) {
 228                 spin_unlock_irqrestore(&task->pi_lock, flags);
 229                 cpu_relax();
 230                 goto retry;
 231         }
 232
 233         /* Deadlock detection */
 234         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 235                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 236                 spin_unlock(&lock->wait_lock);
 237                 ret = deadlock_detect ? -EDEADLK : 0;
 238                 goto out_unlock_pi;
 239         }
 240
 241         top_waiter = rt_mutex_top_waiter(lock);
 242
 243         /* Requeue the waiter */
 244         plist_del(&waiter->list_entry, &lock->wait_list);
 245         waiter->list_entry.prio = task->prio;
 246         plist_add(&waiter->list_entry, &lock->wait_list);
 247
 248         /* Release the task */
 249         spin_unlock_irqrestore(&task->pi_lock, flags);
 250         put_task_struct(task);
 251
 252         /* Grab the next task */
 253         task = rt_mutex_owner(lock);
 254         get_task_struct(task);
 255         spin_lock_irqsave(&task->pi_lock, flags);
 256
 257         if (waiter == rt_mutex_top_waiter(lock)) {
 258                 /* Boost the owner */
 259                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 260                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 261                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 262                 __rt_mutex_adjust_prio(task);
 263
 264         } else if (top_waiter == waiter) {
 265                 /* Deboost the owner */
 266                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 267                 waiter = rt_mutex_top_waiter(lock);
 268                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 269                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 270                 __rt_mutex_adjust_prio(task);
 271         }
 272
 273         spin_unlock_irqrestore(&task->pi_lock, flags);
 274
 275         top_waiter = rt_mutex_top_waiter(lock);
 276         spin_unlock(&lock->wait_lock);
 277
 278         if (!detect_deadlock && waiter != top_waiter)
 279                 goto out_put_task;
 280
 281         goto again;
 282
 283  out_unlock_pi:
 284         spin_unlock_irqrestore(&task->pi_lock, flags);
 285  out_put_task:
 286         put_task_struct(task);
 287
 288         return ret;
 289 }
 290
 291 /*
 292  * Optimization: check if we can steal the lock from the
 293  * assigned pending owner [which might not have taken the
 294  * lock yet]:
 295  */
 296 static inline int try_to_steal_lock(struct rt_mutex *lock)
 297 {
 298         struct task_struct *pendowner = rt_mutex_owner(lock);
 299         struct rt_mutex_waiter *next;
 300         unsigned long flags;
 301
 302         if (!rt_mutex_owner_pending(lock))
 303                 return 0;
 304
 305         if (pendowner == current)
 306                 return 1;
 307
 308         spin_lock_irqsave(&pendowner->pi_lock, flags);
 309         if (current->prio >= pendowner->prio) {
 310                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 311                 return 0;
 312         }
 313
 314         /*
 315          * Check if a waiter is enqueued on the pending owners
 316          * pi_waiters list. Remove it and readjust pending owners
 317          * priority.
 318          */
 319         if (likely(!rt_mutex_has_waiters(lock))) {
 320                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 321                 return 1;
 322         }
 323
 324         /* No chain handling, pending owner is not blocked on anything: */
 325         next = rt_mutex_top_waiter(lock);
 326         plist_del(&next->pi_list_entry, &pendowner->pi_waiters);
 327         __rt_mutex_adjust_prio(pendowner);
 328         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 329
 330         /*
 331          * We are going to steal the lock and a waiter was
 332          * enqueued on the pending owners pi_waiters queue. So
 333          * we have to enqueue this waiter into
 334          * current->pi_waiters list. This covers the case,
 335          * where current is boosted because it holds another
 336          * lock and gets unboosted because the booster is
 337          * interrupted, so we would delay a waiter with higher
 338          * priority as current->normal_prio.
 339          *
 340          * Note: in the rare case of a SCHED_OTHER task changing
 341          * its priority and thus stealing the lock, next->task
 342          * might be current:
 343          */
 344         if (likely(next->task != current)) {
 345                 spin_lock_irqsave(&current->pi_lock, flags);
 346                 plist_add(&next->pi_list_entry, &current->pi_waiters);
 347                 __rt_mutex_adjust_prio(current);
 348                 spin_unlock_irqrestore(&current->pi_lock, flags);
 349         }
 350         return 1;
 351 }
 352
 353 /*
 354  * Try to take an rt-mutex
 355  *
 356  * This fails
 357  * - when the lock has a real owner
 358  * - when a different pending owner exists and has higher priority than current
 359  *
 360  * Must be called with lock->wait_lock held.
 361  */
 362 static int try_to_take_rt_mutex(struct rt_mutex *lock)
 363 {
 364         /*
 365          * We have to be careful here if the atomic speedups are
 366          * enabled, such that, when
 367          *  - no other waiter is on the lock
 368          *  - the lock has been released since we did the cmpxchg
 369          * the lock can be released or taken while we are doing the
 370          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 371          *
 372          * The atomic acquire/release aware variant of
 373          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 374          * the WAITERS bit, the atomic release / acquire can not
 375          * happen anymore and lock->wait_lock protects us from the
 376          * non-atomic case.
 377          *
 378          * Note, that this might set lock->owner =
 379          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 380          * any more. This is fixed up when we take the ownership.
 381          * This is the transitional state explained at the top of this file.
 382          */
 383         mark_rt_mutex_waiters(lock);
 384
 385         if (rt_mutex_owner(lock) && !try_to_steal_lock(lock))
 386                 return 0;
 387
 388         /* We got the lock. */
 389         debug_rt_mutex_lock(lock);
 390
 391         rt_mutex_set_owner(lock, current, 0);
 392
 393         rt_mutex_deadlock_account_lock(lock, current);
 394
 395         return 1;
 396 }
 397
 398 /*
 399  * Task blocks on lock.
 400  *
 401  * Prepare waiter and propagate pi chain
 402  *
 403  * This must be called with lock->wait_lock held.
 404  */
 405 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 406                                    struct rt_mutex_waiter *waiter,
 407                                    int detect_deadlock)
 408 {
 409         struct task_struct *owner = rt_mutex_owner(lock);
 410         struct rt_mutex_waiter *top_waiter = waiter;
 411         unsigned long flags;
 412         int chain_walk = 0, res;
 413
 414         spin_lock_irqsave(&current->pi_lock, flags);
 415         __rt_mutex_adjust_prio(current);
 416         waiter->task = current;
 417         waiter->lock = lock;
 418         plist_node_init(&waiter->list_entry, current->prio);
 419         plist_node_init(&waiter->pi_list_entry, current->prio);
 420
 421         /* Get the top priority waiter on the lock */
 422         if (rt_mutex_has_waiters(lock))
 423                 top_waiter = rt_mutex_top_waiter(lock);
 424         plist_add(&waiter->list_entry, &lock->wait_list);
 425
 426         current->pi_blocked_on = waiter;
 427
 428         spin_unlock_irqrestore(&current->pi_lock, flags);
 429
 430         if (waiter == rt_mutex_top_waiter(lock)) {
 431                 spin_lock_irqsave(&owner->pi_lock, flags);
 432                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 433                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 434
 435                 __rt_mutex_adjust_prio(owner);
 436                 if (owner->pi_blocked_on)
 437                         chain_walk = 1;
 438                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 439         }
 440         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 441                 chain_walk = 1;
 442
 443         if (!chain_walk)
 444                 return 0;
 445
 446         /*
 447          * The owner can't disappear while holding a lock,
 448          * so the owner struct is protected by wait_lock.
 449          * Gets dropped in rt_mutex_adjust_prio_chain()!
 450          */
 451         get_task_struct(owner);
 452
 453         spin_unlock(&lock->wait_lock);
 454
 455         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 456                                          current);
 457
 458         spin_lock(&lock->wait_lock);
 459
 460         return res;
 461 }
 462
 463 /*
 464  * Wake up the next waiter on the lock.
 465  *
 466  * Remove the top waiter from the current tasks waiter list and from
 467  * the lock waiter list. Set it as pending owner. Then wake it up.
 468  *
 469  * Called with lock->wait_lock held.
 470  */
 471 static void wakeup_next_waiter(struct rt_mutex *lock)
 472 {
 473         struct rt_mutex_waiter *waiter;
 474         struct task_struct *pendowner;
 475         unsigned long flags;
 476
 477         spin_lock_irqsave(&current->pi_lock, flags);
 478
 479         waiter = rt_mutex_top_waiter(lock);
 480         plist_del(&waiter->list_entry, &lock->wait_list);
 481
 482         /*
 483          * Remove it from current->pi_waiters. We do not adjust a
 484          * possible priority boost right now. We execute wakeup in the
 485          * boosted mode and go back to normal after releasing
 486          * lock->wait_lock.
 487          */
 488         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 489         pendowner = waiter->task;
 490         waiter->task = NULL;
 491
 492         rt_mutex_set_owner(lock, pendowner, RT_MUTEX_OWNER_PENDING);
 493
 494         spin_unlock_irqrestore(&current->pi_lock, flags);
 495
 496         /*
 497          * Clear the pi_blocked_on variable and enqueue a possible
 498          * waiter into the pi_waiters list of the pending owner. This
 499          * prevents that in case the pending owner gets unboosted a
 500          * waiter with higher priority than pending-owner->normal_prio
 501          * is blocked on the unboosted (pending) owner.
 502          */
 503         spin_lock_irqsave(&pendowner->pi_lock, flags);
 504
 505         WARN_ON(!pendowner->pi_blocked_on);
 506         WARN_ON(pendowner->pi_blocked_on != waiter);
 507         WARN_ON(pendowner->pi_blocked_on->lock != lock);
 508
 509         pendowner->pi_blocked_on = NULL;
 510
 511         if (rt_mutex_has_waiters(lock)) {
 512                 struct rt_mutex_waiter *next;
 513
 514                 next = rt_mutex_top_waiter(lock);
 515                 plist_add(&next->pi_list_entry, &pendowner->pi_waiters);
 516         }
 517         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 518
 519         wake_up_process(pendowner);
 520 }
 521
 522 /*
 523  * Remove a waiter from a lock
 524  *
 525  * Must be called with lock->wait_lock held
 526  */
 527 static void remove_waiter(struct rt_mutex *lock,
 528                           struct rt_mutex_waiter *waiter)
 529 {
 530         int first = (waiter == rt_mutex_top_waiter(lock));
 531         struct task_struct *owner = rt_mutex_owner(lock);
 532         unsigned long flags;
 533         int chain_walk = 0;
 534
 535         spin_lock_irqsave(&current->pi_lock, flags);
 536         plist_del(&waiter->list_entry, &lock->wait_list);
 537         waiter->task = NULL;
 538         current->pi_blocked_on = NULL;
 539         spin_unlock_irqrestore(&current->pi_lock, flags);
 540
 541         if (first && owner != current) {
 542
 543                 spin_lock_irqsave(&owner->pi_lock, flags);
 544
 545                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 546
 547                 if (rt_mutex_has_waiters(lock)) {
 548                         struct rt_mutex_waiter *next;
 549
 550                         next = rt_mutex_top_waiter(lock);
 551                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 552                 }
 553                 __rt_mutex_adjust_prio(owner);
 554
 555                 if (owner->pi_blocked_on)
 556                         chain_walk = 1;
 557
 558                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 559         }
 560
 561         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 562
 563         if (!chain_walk)
 564                 return;
 565
 566         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 567         get_task_struct(owner);
 568
 569         spin_unlock(&lock->wait_lock);
 570
 571         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 572
 573         spin_lock(&lock->wait_lock);
 574 }
 575
 576 /*
 577  * Recheck the pi chain, in case we got a priority setting
 578  *
 579  * Called from sched_setscheduler
 580  */
 581 void rt_mutex_adjust_pi(struct task_struct *task)
 582 {
 583         struct rt_mutex_waiter *waiter;
 584         unsigned long flags;
 585
 586         spin_lock_irqsave(&task->pi_lock, flags);
 587
 588         waiter = task->pi_blocked_on;
 589         if (!waiter || waiter->list_entry.prio == task->prio) {
 590                 spin_unlock_irqrestore(&task->pi_lock, flags);
 591                 return;
 592         }
 593
 594         spin_unlock_irqrestore(&task->pi_lock, flags);
 595
 596         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 597         get_task_struct(task);
 598         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 599 }
 600
 601 /*
 602  * Slow path lock function:
 603  */
 604 static int __sched
 605 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 606                   struct hrtimer_sleeper *timeout,
 607                   int detect_deadlock)
 608 {
 609         struct rt_mutex_waiter waiter;
 610         int ret = 0;
 611
 612         debug_rt_mutex_init_waiter(&waiter);
 613         waiter.task = NULL;
 614
 615         spin_lock(&lock->wait_lock);
 616
 617         /* Try to acquire the lock again: */
 618         if (try_to_take_rt_mutex(lock)) {
 619                 spin_unlock(&lock->wait_lock);
 620                 return 0;
 621         }
 622
 623         set_current_state(state);
 624
 625         /* Setup the timer, when timeout != NULL */
 626         if (unlikely(timeout))
 627                 hrtimer_start(&timeout->timer, timeout->timer.expires,
 628                               HRTIMER_ABS);
 629
 630         for (;;) {
 631                 /* Try to acquire the lock: */
 632                 if (try_to_take_rt_mutex(lock))
 633                         break;
 634
 635                 /*
 636                  * TASK_INTERRUPTIBLE checks for signals and
 637                  * timeout. Ignored otherwise.
 638                  */
 639                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 640                         /* Signal pending? */
 641                         if (signal_pending(current))
 642                                 ret = -EINTR;
 643                         if (timeout && !timeout->task)
 644                                 ret = -ETIMEDOUT;
 645                         if (ret)
 646                                 break;
 647                 }
 648
 649                 /*
 650                  * waiter.task is NULL the first time we come here and
 651                  * when we have been woken up by the previous owner
 652                  * but the lock got stolen by a higher prio task.
 653                  */
 654                 if (!waiter.task) {
 655                         ret = task_blocks_on_rt_mutex(lock, &waiter,
 656                                                       detect_deadlock);
 657                         /*
 658                          * If we got woken up by the owner then start loop
 659                          * all over without going into schedule to try
 660                          * to get the lock now:
 661                          */
 662                         if (unlikely(!waiter.task))
 663                                 continue;
 664
 665                         if (unlikely(ret))
 666                                 break;
 667                 }
 668
 669                 spin_unlock(&lock->wait_lock);
 670
 671                 debug_rt_mutex_print_deadlock(&waiter);
 672
 673                 if (waiter.task)
 674                         schedule_rt_mutex(lock);
 675
 676                 spin_lock(&lock->wait_lock);
 677                 set_current_state(state);
 678         }
 679
 680         set_current_state(TASK_RUNNING);
 681
 682         if (unlikely(waiter.task))
 683                 remove_waiter(lock, &waiter);
 684
 685         /*
 686          * try_to_take_rt_mutex() sets the waiter bit
 687          * unconditionally. We might have to fix that up.
 688          */
 689         fixup_rt_mutex_waiters(lock);
 690
 691         spin_unlock(&lock->wait_lock);
 692
 693         /* Remove pending timer: */
 694         if (unlikely(timeout))
 695                 hrtimer_cancel(&timeout->timer);
 696
 697         /*
 698          * Readjust priority, when we did not get the lock. We might
 699          * have been the pending owner and boosted. Since we did not
 700          * take the lock, the PI boost has to go.
 701          */
 702         if (unlikely(ret))
 703                 rt_mutex_adjust_prio(current);
 704
 705         debug_rt_mutex_free_waiter(&waiter);
 706
 707         return ret;
 708 }
 709
 710 /*
 711  * Slow path try-lock function:
 712  */
 713 static inline int
 714 rt_mutex_slowtrylock(struct rt_mutex *lock)
 715 {
 716         int ret = 0;
 717
 718         spin_lock(&lock->wait_lock);
 719
 720         if (likely(rt_mutex_owner(lock) != current)) {
 721
 722                 ret = try_to_take_rt_mutex(lock);
 723                 /*
 724                  * try_to_take_rt_mutex() sets the lock waiters
 725                  * bit unconditionally. Clean this up.
 726                  */
 727                 fixup_rt_mutex_waiters(lock);
 728         }
 729
 730         spin_unlock(&lock->wait_lock);
 731
 732         return ret;
 733 }
 734
 735 /*
 736  * Slow path to release a rt-mutex:
 737  */
 738 static void __sched
 739 rt_mutex_slowunlock(struct rt_mutex *lock)
 740 {
 741         spin_lock(&lock->wait_lock);
 742
 743         debug_rt_mutex_unlock(lock);
 744
 745         rt_mutex_deadlock_account_unlock(current);
 746
 747         if (!rt_mutex_has_waiters(lock)) {
 748                 lock->owner = NULL;
 749                 spin_unlock(&lock->wait_lock);
 750                 return;
 751         }
 752
 753         wakeup_next_waiter(lock);
 754
 755         spin_unlock(&lock->wait_lock);
 756
 757         /* Undo pi boosting if necessary: */
 758         rt_mutex_adjust_prio(current);
 759 }
 760
 761 /*
 762  * debug aware fast / slowpath lock,trylock,unlock
 763  *
 764  * The atomic acquire/release ops are compiled away, when either the
 765  * architecture does not support cmpxchg or when debugging is enabled.
 766  */
 767 static inline int
 768 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 769                   int detect_deadlock,
 770                   int (*slowfn)(struct rt_mutex *lock, int state,
 771                                 struct hrtimer_sleeper *timeout,
 772                                 int detect_deadlock))
 773 {
 774         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 775                 rt_mutex_deadlock_account_lock(lock, current);
 776                 return 0;
 777         } else
 778                 return slowfn(lock, state, NULL, detect_deadlock);
 779 }
 780
 781 static inline int
 782 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 783                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 784                         int (*slowfn)(struct rt_mutex *lock, int state,
 785                                       struct hrtimer_sleeper *timeout,
 786                                       int detect_deadlock))
 787 {
 788         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 789                 rt_mutex_deadlock_account_lock(lock, current);
 790                 return 0;
 791         } else
 792                 return slowfn(lock, state, timeout, detect_deadlock);
 793 }
 794
 795 static inline int
 796 rt_mutex_fasttrylock(struct rt_mutex *lock,
 797                      int (*slowfn)(struct rt_mutex *lock))
 798 {
 799         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 800                 rt_mutex_deadlock_account_lock(lock, current);
 801                 return 1;
 802         }
 803         return slowfn(lock);
 804 }
 805
 806 static inline void
 807 rt_mutex_fastunlock(struct rt_mutex *lock,
 808                     void (*slowfn)(struct rt_mutex *lock))
 809 {
 810         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 811                 rt_mutex_deadlock_account_unlock(current);
 812         else
 813                 slowfn(lock);
 814 }
 815
 816 /**
 817  * rt_mutex_lock - lock a rt_mutex
 818  *
 819  * @lock: the rt_mutex to be locked
 820  */
 821 void __sched rt_mutex_lock(struct rt_mutex *lock)
 822 {
 823         might_sleep();
 824
 825         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 826 }
 827 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 828
 829 /**
 830  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 831  *
 832  * @lock:               the rt_mutex to be locked
 833  * @detect_deadlock:    deadlock detection on/off
 834  *
 835  * Returns:
 836  *  0           on success
 837  * -EINTR       when interrupted by a signal
 838  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 839  */
 840 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 841                                                  int detect_deadlock)
 842 {
 843         might_sleep();
 844
 845         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 846                                  detect_deadlock, rt_mutex_slowlock);
 847 }
 848 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 849
 850 /**
 851  * rt_mutex_lock_interruptible_ktime - lock a rt_mutex interruptible
 852  *                                     the timeout structure is provided
 853  *                                     by the caller
 854  *
 855  * @lock:               the rt_mutex to be locked
 856  * @timeout:            timeout structure or NULL (no timeout)
 857  * @detect_deadlock:    deadlock detection on/off
 858  *
 859  * Returns:
 860  *  0           on success
 861  * -EINTR       when interrupted by a signal
 862  * -ETIMEOUT    when the timeout expired
 863  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 864  */
 865 int
 866 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 867                     int detect_deadlock)
 868 {
 869         might_sleep();
 870
 871         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 872                                        detect_deadlock, rt_mutex_slowlock);
 873 }
 874 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 875
 876 /**
 877  * rt_mutex_trylock - try to lock a rt_mutex
 878  *
 879  * @lock:       the rt_mutex to be locked
 880  *
 881  * Returns 1 on success and 0 on contention
 882  */
 883 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 884 {
 885         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 886 }
 887 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 888
 889 /**
 890  * rt_mutex_unlock - unlock a rt_mutex
 891  *
 892  * @lock: the rt_mutex to be unlocked
 893  */
 894 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 895 {
 896         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 897 }
 898 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 899
 900 /***
 901  * rt_mutex_destroy - mark a mutex unusable
 902  * @lock: the mutex to be destroyed
 903  *
 904  * This function marks the mutex uninitialized, and any subsequent
 905  * use of the mutex is forbidden. The mutex must not be locked when
 906  * this function is called.
 907  */
 908 void rt_mutex_destroy(struct rt_mutex *lock)
 909 {
 910         WARN_ON(rt_mutex_is_locked(lock));
 911 #ifdef CONFIG_DEBUG_RT_MUTEXES
 912         lock->magic = NULL;
 913 #endif
 914 }
 915
 916 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 917
 918 /**
 919  * __rt_mutex_init - initialize the rt lock
 920  *
 921  * @lock: the rt lock to be initialized
 922  *
 923  * Initialize the rt lock to unlocked state.
 924  *
 925  * Initializing of a locked rt lock is not allowed
 926  */
 927 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 928 {
 929         lock->owner = NULL;
 930         spin_lock_init(&lock->wait_lock);
 931         plist_head_init(&lock->wait_list, &lock->wait_lock);
 932
 933         debug_rt_mutex_init(lock, name);
 934 }
 935 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 936
 937 /**
 938  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 939  *                              proxy owner
 940  *
 941  * @lock:       the rt_mutex to be locked
 942  * @proxy_owner:the task to set as owner
 943  *
 944  * No locking. Caller has to do serializing itself
 945  * Special API call for PI-futex support
 946  */
 947 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 948                                 struct task_struct *proxy_owner)
 949 {
 950         __rt_mutex_init(lock, NULL);
 951         debug_rt_mutex_proxy_lock(lock, proxy_owner);
 952         rt_mutex_set_owner(lock, proxy_owner, 0);
 953         rt_mutex_deadlock_account_lock(lock, proxy_owner);
 954 }
 955
 956 /**
 957  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 958  *
 959  * @lock:       the rt_mutex to be locked
 960  *
 961  * No locking. Caller has to do serializing itself
 962  * Special API call for PI-futex support
 963  */
 964 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 965                            struct task_struct *proxy_owner)
 966 {
 967         debug_rt_mutex_proxy_unlock(lock);
 968         rt_mutex_set_owner(lock, NULL, 0);
 969         rt_mutex_deadlock_account_unlock(proxy_owner);
 970 }
 971
 972 /**
 973  * rt_mutex_next_owner - return the next owner of the lock
 974  *
 975  * @lock: the rt lock query
 976  *
 977  * Returns the next owner of the lock or NULL
 978  *
 979  * Caller has to serialize against other accessors to the lock
 980  * itself.
 981  *
 982  * Special API call for PI-futex support
 983  */
 984 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
 985 {
 986         if (!rt_mutex_has_waiters(lock))
 987                 return NULL;
 988
 989         return rt_mutex_top_waiter(lock)->task;
 990 }