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/export.h>
  15 #include <linux/sched.h>
  16 #include <linux/timer.h>
  17
  18 #include "rtmutex_common.h"
  19
  20 /*
  21  * lock->owner state tracking:
  22  *
  23  * lock->owner holds the task_struct pointer of the owner. Bit 0
  24  * is used to keep track of the "lock has waiters" state.
  25  *
  26  * owner        bit0
  27  * NULL         0       lock is free (fast acquire possible)
  28  * NULL         1       lock is free and has waiters and the top waiter
  29  *                              is going to take the lock*
  30  * taskpointer  0       lock is held (fast release possible)
  31  * taskpointer  1       lock is held and has waiters**
  32  *
  33  * The fast atomic compare exchange based acquire and release is only
  34  * possible when bit 0 of lock->owner is 0.
  35  *
  36  * (*) It also can be a transitional state when grabbing the lock
  37  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  38  * we need to set the bit0 before looking at the lock, and the owner may be
  39  * NULL in this small time, hence this can be a transitional state.
  40  *
  41  * (**) There is a small time when bit 0 is set but there are no
  42  * waiters. This can happen when grabbing the lock in the slow path.
  43  * To prevent a cmpxchg of the owner releasing the lock, we need to
  44  * set this bit before looking at the lock.
  45  */
  46
  47 static void
  48 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
  49 {
  50         unsigned long val = (unsigned long)owner;
  51
  52         if (rt_mutex_has_waiters(lock))
  53                 val |= RT_MUTEX_HAS_WAITERS;
  54
  55         lock->owner = (struct task_struct *)val;
  56 }
  57
  58 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  59 {
  60         lock->owner = (struct task_struct *)
  61                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  62 }
  63
  64 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  65 {
  66         if (!rt_mutex_has_waiters(lock))
  67                 clear_rt_mutex_waiters(lock);
  68 }
  69
  70 /*
  71  * We can speed up the acquire/release, if the architecture
  72  * supports cmpxchg and if there's no debugging state to be set up
  73  */
  74 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
  75 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
  76 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  77 {
  78         unsigned long owner, *p = (unsigned long *) &lock->owner;
  79
  80         do {
  81                 owner = *p;
  82         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
  83 }
  84 #else
  85 # define rt_mutex_cmpxchg(l,c,n)        (0)
  86 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  87 {
  88         lock->owner = (struct task_struct *)
  89                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
  90 }
  91 #endif
  92
  93 /*
  94  * Calculate task priority from the waiter list priority
  95  *
  96  * Return task->normal_prio when the waiter list is empty or when
  97  * the waiter is not allowed to do priority boosting
  98  */
  99 int rt_mutex_getprio(struct task_struct *task)
 100 {
 101         if (likely(!task_has_pi_waiters(task)))
 102                 return task->normal_prio;
 103
 104         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
 105                    task->normal_prio);
 106 }
 107
 108 /*
 109  * Adjust the priority of a task, after its pi_waiters got modified.
 110  *
 111  * This can be both boosting and unboosting. task->pi_lock must be held.
 112  */
 113 static void __rt_mutex_adjust_prio(struct task_struct *task)
 114 {
 115         int prio = rt_mutex_getprio(task);
 116
 117         if (task->prio != prio)
 118                 rt_mutex_setprio(task, prio);
 119 }
 120
 121 /*
 122  * Adjust task priority (undo boosting). Called from the exit path of
 123  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 124  *
 125  * (Note: We do this outside of the protection of lock->wait_lock to
 126  * allow the lock to be taken while or before we readjust the priority
 127  * of task. We do not use the spin_xx_mutex() variants here as we are
 128  * outside of the debug path.)
 129  */
 130 static void rt_mutex_adjust_prio(struct task_struct *task)
 131 {
 132         unsigned long flags;
 133
 134         raw_spin_lock_irqsave(&task->pi_lock, flags);
 135         __rt_mutex_adjust_prio(task);
 136         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 137 }
 138
 139 /*
 140  * Max number of times we'll walk the boosting chain:
 141  */
 142 int max_lock_depth = 1024;
 143
 144 /*
 145  * Adjust the priority chain. Also used for deadlock detection.
 146  * Decreases task's usage by one - may thus free the task.
 147  * Returns 0 or -EDEADLK.
 148  */
 149 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 150                                       int deadlock_detect,
 151                                       struct rt_mutex *orig_lock,
 152                                       struct rt_mutex_waiter *orig_waiter,
 153                                       struct task_struct *top_task)
 154 {
 155         struct rt_mutex *lock;
 156         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 157         int detect_deadlock, ret = 0, depth = 0;
 158         unsigned long flags;
 159
 160         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 161                                                          deadlock_detect);
 162
 163         /*
 164          * The (de)boosting is a step by step approach with a lot of
 165          * pitfalls. We want this to be preemptible and we want hold a
 166          * maximum of two locks per step. So we have to check
 167          * carefully whether things change under us.
 168          */
 169  again:
 170         if (++depth > max_lock_depth) {
 171                 static int prev_max;
 172
 173                 /*
 174                  * Print this only once. If the admin changes the limit,
 175                  * print a new message when reaching the limit again.
 176                  */
 177                 if (prev_max != max_lock_depth) {
 178                         prev_max = max_lock_depth;
 179                         printk(KERN_WARNING "Maximum lock depth %d reached "
 180                                "task: %s (%d)\n", max_lock_depth,
 181                                top_task->comm, task_pid_nr(top_task));
 182                 }
 183                 put_task_struct(task);
 184
 185                 return deadlock_detect ? -EDEADLK : 0;
 186         }
 187  retry:
 188         /*
 189          * Task can not go away as we did a get_task() before !
 190          */
 191         raw_spin_lock_irqsave(&task->pi_lock, flags);
 192
 193         waiter = task->pi_blocked_on;
 194         /*
 195          * Check whether the end of the boosting chain has been
 196          * reached or the state of the chain has changed while we
 197          * dropped the locks.
 198          */
 199         if (!waiter)
 200                 goto out_unlock_pi;
 201
 202         /*
 203          * Check the orig_waiter state. After we dropped the locks,
 204          * the previous owner of the lock might have released the lock.
 205          */
 206         if (orig_waiter && !rt_mutex_owner(orig_lock))
 207                 goto out_unlock_pi;
 208
 209         /*
 210          * Drop out, when the task has no waiters. Note,
 211          * top_waiter can be NULL, when we are in the deboosting
 212          * mode!
 213          */
 214         if (top_waiter) {
 215                 if (!task_has_pi_waiters(task))
 216                         goto out_unlock_pi;
 217                 /*
 218                  * If deadlock detection is off, we stop here if we
 219                  * are not the top pi waiter of the task.
 220                  */
 221                 if (!detect_deadlock && top_waiter != task_top_pi_waiter(task))
 222                         goto out_unlock_pi;
 223         }
 224
 225         /*
 226          * When deadlock detection is off then we check, if further
 227          * priority adjustment is necessary.
 228          */
 229         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 230                 goto out_unlock_pi;
 231
 232         lock = waiter->lock;
 233         if (!raw_spin_trylock(&lock->wait_lock)) {
 234                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 235                 cpu_relax();
 236                 goto retry;
 237         }
 238
 239         /*
 240          * Deadlock detection. If the lock is the same as the original
 241          * lock which caused us to walk the lock chain or if the
 242          * current lock is owned by the task which initiated the chain
 243          * walk, we detected a deadlock.
 244          */
 245         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 246                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 247                 raw_spin_unlock(&lock->wait_lock);
 248                 ret = deadlock_detect ? -EDEADLK : 0;
 249                 goto out_unlock_pi;
 250         }
 251
 252         top_waiter = rt_mutex_top_waiter(lock);
 253
 254         /* Requeue the waiter */
 255         plist_del(&waiter->list_entry, &lock->wait_list);
 256         waiter->list_entry.prio = task->prio;
 257         plist_add(&waiter->list_entry, &lock->wait_list);
 258
 259         /* Release the task */
 260         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 261         if (!rt_mutex_owner(lock)) {
 262                 /*
 263                  * If the requeue above changed the top waiter, then we need
 264                  * to wake the new top waiter up to try to get the lock.
 265                  */
 266
 267                 if (top_waiter != rt_mutex_top_waiter(lock))
 268                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 269                 raw_spin_unlock(&lock->wait_lock);
 270                 goto out_put_task;
 271         }
 272         put_task_struct(task);
 273
 274         /* Grab the next task */
 275         task = rt_mutex_owner(lock);
 276         get_task_struct(task);
 277         raw_spin_lock_irqsave(&task->pi_lock, flags);
 278
 279         if (waiter == rt_mutex_top_waiter(lock)) {
 280                 /* Boost the owner */
 281                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 282                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 283                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 284                 __rt_mutex_adjust_prio(task);
 285
 286         } else if (top_waiter == waiter) {
 287                 /* Deboost the owner */
 288                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 289                 waiter = rt_mutex_top_waiter(lock);
 290                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 291                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 292                 __rt_mutex_adjust_prio(task);
 293         }
 294
 295         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 296
 297         top_waiter = rt_mutex_top_waiter(lock);
 298         raw_spin_unlock(&lock->wait_lock);
 299
 300         if (!detect_deadlock && waiter != top_waiter)
 301                 goto out_put_task;
 302
 303         goto again;
 304
 305  out_unlock_pi:
 306         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 307  out_put_task:
 308         put_task_struct(task);
 309
 310         return ret;
 311 }
 312
 313 /*
 314  * Try to take an rt-mutex
 315  *
 316  * Must be called with lock->wait_lock held.
 317  *
 318  * @lock:   the lock to be acquired.
 319  * @task:   the task which wants to acquire the lock
 320  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 321  */
 322 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 323                 struct rt_mutex_waiter *waiter)
 324 {
 325         /*
 326          * We have to be careful here if the atomic speedups are
 327          * enabled, such that, when
 328          *  - no other waiter is on the lock
 329          *  - the lock has been released since we did the cmpxchg
 330          * the lock can be released or taken while we are doing the
 331          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 332          *
 333          * The atomic acquire/release aware variant of
 334          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 335          * the WAITERS bit, the atomic release / acquire can not
 336          * happen anymore and lock->wait_lock protects us from the
 337          * non-atomic case.
 338          *
 339          * Note, that this might set lock->owner =
 340          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 341          * any more. This is fixed up when we take the ownership.
 342          * This is the transitional state explained at the top of this file.
 343          */
 344         mark_rt_mutex_waiters(lock);
 345
 346         if (rt_mutex_owner(lock))
 347                 return 0;
 348
 349         /*
 350          * It will get the lock because of one of these conditions:
 351          * 1) there is no waiter
 352          * 2) higher priority than waiters
 353          * 3) it is top waiter
 354          */
 355         if (rt_mutex_has_waiters(lock)) {
 356                 if (task->prio >= rt_mutex_top_waiter(lock)->list_entry.prio) {
 357                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 358                                 return 0;
 359                 }
 360         }
 361
 362         if (waiter || rt_mutex_has_waiters(lock)) {
 363                 unsigned long flags;
 364                 struct rt_mutex_waiter *top;
 365
 366                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 367
 368                 /* remove the queued waiter. */
 369                 if (waiter) {
 370                         plist_del(&waiter->list_entry, &lock->wait_list);
 371                         task->pi_blocked_on = NULL;
 372                 }
 373
 374                 /*
 375                  * We have to enqueue the top waiter(if it exists) into
 376                  * task->pi_waiters list.
 377                  */
 378                 if (rt_mutex_has_waiters(lock)) {
 379                         top = rt_mutex_top_waiter(lock);
 380                         top->pi_list_entry.prio = top->list_entry.prio;
 381                         plist_add(&top->pi_list_entry, &task->pi_waiters);
 382                 }
 383                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 384         }
 385
 386         /* We got the lock. */
 387         debug_rt_mutex_lock(lock);
 388
 389         rt_mutex_set_owner(lock, task);
 390
 391         rt_mutex_deadlock_account_lock(lock, task);
 392
 393         return 1;
 394 }
 395
 396 /*
 397  * Task blocks on lock.
 398  *
 399  * Prepare waiter and propagate pi chain
 400  *
 401  * This must be called with lock->wait_lock held.
 402  */
 403 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 404                                    struct rt_mutex_waiter *waiter,
 405                                    struct task_struct *task,
 406                                    int detect_deadlock)
 407 {
 408         struct task_struct *owner = rt_mutex_owner(lock);
 409         struct rt_mutex_waiter *top_waiter = waiter;
 410         unsigned long flags;
 411         int chain_walk = 0, res;
 412
 413         /*
 414          * Early deadlock detection. We really don't want the task to
 415          * enqueue on itself just to untangle the mess later. It's not
 416          * only an optimization. We drop the locks, so another waiter
 417          * can come in before the chain walk detects the deadlock. So
 418          * the other will detect the deadlock and return -EDEADLOCK,
 419          * which is wrong, as the other waiter is not in a deadlock
 420          * situation.
 421          */
 422         if (detect_deadlock && owner == task)
 423                 return -EDEADLK;
 424
 425         raw_spin_lock_irqsave(&task->pi_lock, flags);
 426         __rt_mutex_adjust_prio(task);
 427         waiter->task = task;
 428         waiter->lock = lock;
 429         plist_node_init(&waiter->list_entry, task->prio);
 430         plist_node_init(&waiter->pi_list_entry, task->prio);
 431
 432         /* Get the top priority waiter on the lock */
 433         if (rt_mutex_has_waiters(lock))
 434                 top_waiter = rt_mutex_top_waiter(lock);
 435         plist_add(&waiter->list_entry, &lock->wait_list);
 436
 437         task->pi_blocked_on = waiter;
 438
 439         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 440
 441         if (!owner)
 442                 return 0;
 443
 444         if (waiter == rt_mutex_top_waiter(lock)) {
 445                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 446                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 447                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 448
 449                 __rt_mutex_adjust_prio(owner);
 450                 if (owner->pi_blocked_on)
 451                         chain_walk = 1;
 452                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 453         }
 454         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 455                 chain_walk = 1;
 456
 457         if (!chain_walk)
 458                 return 0;
 459
 460         /*
 461          * The owner can't disappear while holding a lock,
 462          * so the owner struct is protected by wait_lock.
 463          * Gets dropped in rt_mutex_adjust_prio_chain()!
 464          */
 465         get_task_struct(owner);
 466
 467         raw_spin_unlock(&lock->wait_lock);
 468
 469         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 470                                          task);
 471
 472         raw_spin_lock(&lock->wait_lock);
 473
 474         return res;
 475 }
 476
 477 /*
 478  * Wake up the next waiter on the lock.
 479  *
 480  * Remove the top waiter from the current tasks waiter list and wake it up.
 481  *
 482  * Called with lock->wait_lock held.
 483  */
 484 static void wakeup_next_waiter(struct rt_mutex *lock)
 485 {
 486         struct rt_mutex_waiter *waiter;
 487         unsigned long flags;
 488
 489         raw_spin_lock_irqsave(&current->pi_lock, flags);
 490
 491         waiter = rt_mutex_top_waiter(lock);
 492
 493         /*
 494          * Remove it from current->pi_waiters. We do not adjust a
 495          * possible priority boost right now. We execute wakeup in the
 496          * boosted mode and go back to normal after releasing
 497          * lock->wait_lock.
 498          */
 499         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 500
 501         rt_mutex_set_owner(lock, NULL);
 502
 503         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 504
 505         wake_up_process(waiter->task);
 506 }
 507
 508 /*
 509  * Remove a waiter from a lock and give up
 510  *
 511  * Must be called with lock->wait_lock held and
 512  * have just failed to try_to_take_rt_mutex().
 513  */
 514 static void remove_waiter(struct rt_mutex *lock,
 515                           struct rt_mutex_waiter *waiter)
 516 {
 517         int first = (waiter == rt_mutex_top_waiter(lock));
 518         struct task_struct *owner = rt_mutex_owner(lock);
 519         unsigned long flags;
 520         int chain_walk = 0;
 521
 522         raw_spin_lock_irqsave(&current->pi_lock, flags);
 523         plist_del(&waiter->list_entry, &lock->wait_list);
 524         current->pi_blocked_on = NULL;
 525         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 526
 527         if (!owner)
 528                 return;
 529
 530         if (first) {
 531
 532                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 533
 534                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 535
 536                 if (rt_mutex_has_waiters(lock)) {
 537                         struct rt_mutex_waiter *next;
 538
 539                         next = rt_mutex_top_waiter(lock);
 540                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 541                 }
 542                 __rt_mutex_adjust_prio(owner);
 543
 544                 if (owner->pi_blocked_on)
 545                         chain_walk = 1;
 546
 547                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 548         }
 549
 550         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 551
 552         if (!chain_walk)
 553                 return;
 554
 555         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 556         get_task_struct(owner);
 557
 558         raw_spin_unlock(&lock->wait_lock);
 559
 560         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 561
 562         raw_spin_lock(&lock->wait_lock);
 563 }
 564
 565 /*
 566  * Recheck the pi chain, in case we got a priority setting
 567  *
 568  * Called from sched_setscheduler
 569  */
 570 void rt_mutex_adjust_pi(struct task_struct *task)
 571 {
 572         struct rt_mutex_waiter *waiter;
 573         unsigned long flags;
 574
 575         raw_spin_lock_irqsave(&task->pi_lock, flags);
 576
 577         waiter = task->pi_blocked_on;
 578         if (!waiter || waiter->list_entry.prio == task->prio) {
 579                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 580                 return;
 581         }
 582
 583         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 584
 585         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 586         get_task_struct(task);
 587         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 588 }
 589
 590 /**
 591  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 592  * @lock:                the rt_mutex to take
 593  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 594  *                       or TASK_UNINTERRUPTIBLE)
 595  * @timeout:             the pre-initialized and started timer, or NULL for none
 596  * @waiter:              the pre-initialized rt_mutex_waiter
 597  *
 598  * lock->wait_lock must be held by the caller.
 599  */
 600 static int __sched
 601 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 602                     struct hrtimer_sleeper *timeout,
 603                     struct rt_mutex_waiter *waiter)
 604 {
 605         int ret = 0;
 606         int was_disabled;
 607
 608         for (;;) {
 609                 /* Try to acquire the lock: */
 610                 if (try_to_take_rt_mutex(lock, current, waiter))
 611                         break;
 612
 613                 /*
 614                  * TASK_INTERRUPTIBLE checks for signals and
 615                  * timeout. Ignored otherwise.
 616                  */
 617                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 618                         /* Signal pending? */
 619                         if (signal_pending(current))
 620                                 ret = -EINTR;
 621                         if (timeout && !timeout->task)
 622                                 ret = -ETIMEDOUT;
 623                         if (ret)
 624                                 break;
 625                 }
 626
 627                 raw_spin_unlock(&lock->wait_lock);
 628
 629                 was_disabled = irqs_disabled();
 630                 if (was_disabled)
 631                         local_irq_enable();
 632
 633                 debug_rt_mutex_print_deadlock(waiter);
 634
 635                 schedule_rt_mutex(lock);
 636
 637                 if (was_disabled)
 638                         local_irq_disable();
 639
 640                 raw_spin_lock(&lock->wait_lock);
 641                 set_current_state(state);
 642         }
 643
 644         return ret;
 645 }
 646
 647 /*
 648  * Slow path lock function:
 649  */
 650 static int __sched
 651 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 652                   struct hrtimer_sleeper *timeout,
 653                   int detect_deadlock)
 654 {
 655         struct rt_mutex_waiter waiter;
 656         int ret = 0;
 657
 658         debug_rt_mutex_init_waiter(&waiter);
 659
 660         raw_spin_lock(&lock->wait_lock);
 661
 662         /* Try to acquire the lock again: */
 663         if (try_to_take_rt_mutex(lock, current, NULL)) {
 664                 raw_spin_unlock(&lock->wait_lock);
 665                 return 0;
 666         }
 667
 668         set_current_state(state);
 669
 670         /* Setup the timer, when timeout != NULL */
 671         if (unlikely(timeout)) {
 672                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 673                 if (!hrtimer_active(&timeout->timer))
 674                         timeout->task = NULL;
 675         }
 676
 677         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 678
 679         if (likely(!ret))
 680                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
 681
 682         set_current_state(TASK_RUNNING);
 683
 684         if (unlikely(ret))
 685                 remove_waiter(lock, &waiter);
 686
 687         /*
 688          * try_to_take_rt_mutex() sets the waiter bit
 689          * unconditionally. We might have to fix that up.
 690          */
 691         fixup_rt_mutex_waiters(lock);
 692
 693         raw_spin_unlock(&lock->wait_lock);
 694
 695         /* Remove pending timer: */
 696         if (unlikely(timeout))
 697                 hrtimer_cancel(&timeout->timer);
 698
 699         debug_rt_mutex_free_waiter(&waiter);
 700
 701         return ret;
 702 }
 703
 704 /*
 705  * Slow path try-lock function:
 706  */
 707 static inline int
 708 rt_mutex_slowtrylock(struct rt_mutex *lock)
 709 {
 710         int ret = 0;
 711
 712         raw_spin_lock(&lock->wait_lock);
 713
 714         if (likely(rt_mutex_owner(lock) != current)) {
 715
 716                 ret = try_to_take_rt_mutex(lock, current, NULL);
 717                 /*
 718                  * try_to_take_rt_mutex() sets the lock waiters
 719                  * bit unconditionally. Clean this up.
 720                  */
 721                 fixup_rt_mutex_waiters(lock);
 722         }
 723
 724         raw_spin_unlock(&lock->wait_lock);
 725
 726         return ret;
 727 }
 728
 729 /*
 730  * Slow path to release a rt-mutex:
 731  */
 732 static void __sched
 733 rt_mutex_slowunlock(struct rt_mutex *lock)
 734 {
 735         raw_spin_lock(&lock->wait_lock);
 736
 737         debug_rt_mutex_unlock(lock);
 738
 739         rt_mutex_deadlock_account_unlock(current);
 740
 741         if (!rt_mutex_has_waiters(lock)) {
 742                 lock->owner = NULL;
 743                 raw_spin_unlock(&lock->wait_lock);
 744                 return;
 745         }
 746
 747         wakeup_next_waiter(lock);
 748
 749         raw_spin_unlock(&lock->wait_lock);
 750
 751         /* Undo pi boosting if necessary: */
 752         rt_mutex_adjust_prio(current);
 753 }
 754
 755 /*
 756  * debug aware fast / slowpath lock,trylock,unlock
 757  *
 758  * The atomic acquire/release ops are compiled away, when either the
 759  * architecture does not support cmpxchg or when debugging is enabled.
 760  */
 761 static inline int
 762 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 763                   int detect_deadlock,
 764                   int (*slowfn)(struct rt_mutex *lock, int state,
 765                                 struct hrtimer_sleeper *timeout,
 766                                 int detect_deadlock))
 767 {
 768         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 769                 rt_mutex_deadlock_account_lock(lock, current);
 770                 return 0;
 771         } else
 772                 return slowfn(lock, state, NULL, detect_deadlock);
 773 }
 774
 775 static inline int
 776 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 777                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 778                         int (*slowfn)(struct rt_mutex *lock, int state,
 779                                       struct hrtimer_sleeper *timeout,
 780                                       int detect_deadlock))
 781 {
 782         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 783                 rt_mutex_deadlock_account_lock(lock, current);
 784                 return 0;
 785         } else
 786                 return slowfn(lock, state, timeout, detect_deadlock);
 787 }
 788
 789 static inline int
 790 rt_mutex_fasttrylock(struct rt_mutex *lock,
 791                      int (*slowfn)(struct rt_mutex *lock))
 792 {
 793         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 794                 rt_mutex_deadlock_account_lock(lock, current);
 795                 return 1;
 796         }
 797         return slowfn(lock);
 798 }
 799
 800 static inline void
 801 rt_mutex_fastunlock(struct rt_mutex *lock,
 802                     void (*slowfn)(struct rt_mutex *lock))
 803 {
 804         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 805                 rt_mutex_deadlock_account_unlock(current);
 806         else
 807                 slowfn(lock);
 808 }
 809
 810 /**
 811  * rt_mutex_lock - lock a rt_mutex
 812  *
 813  * @lock: the rt_mutex to be locked
 814  */
 815 void __sched rt_mutex_lock(struct rt_mutex *lock)
 816 {
 817         might_sleep();
 818
 819         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 820 }
 821 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 822
 823 /**
 824  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 825  *
 826  * @lock:               the rt_mutex to be locked
 827  * @detect_deadlock:    deadlock detection on/off
 828  *
 829  * Returns:
 830  *  0           on success
 831  * -EINTR       when interrupted by a signal
 832  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 833  */
 834 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 835                                                  int detect_deadlock)
 836 {
 837         might_sleep();
 838
 839         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 840                                  detect_deadlock, rt_mutex_slowlock);
 841 }
 842 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 843
 844 /**
 845  * rt_mutex_timed_lock - lock a rt_mutex interruptible
 846  *                      the timeout structure is provided
 847  *                      by the caller
 848  *
 849  * @lock:               the rt_mutex to be locked
 850  * @timeout:            timeout structure or NULL (no timeout)
 851  * @detect_deadlock:    deadlock detection on/off
 852  *
 853  * Returns:
 854  *  0           on success
 855  * -EINTR       when interrupted by a signal
 856  * -ETIMEDOUT   when the timeout expired
 857  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 858  */
 859 int
 860 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 861                     int detect_deadlock)
 862 {
 863         might_sleep();
 864
 865         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 866                                        detect_deadlock, rt_mutex_slowlock);
 867 }
 868 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 869
 870 /**
 871  * rt_mutex_trylock - try to lock a rt_mutex
 872  *
 873  * @lock:       the rt_mutex to be locked
 874  *
 875  * Returns 1 on success and 0 on contention
 876  */
 877 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 878 {
 879         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 880 }
 881 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 882
 883 /**
 884  * rt_mutex_unlock - unlock a rt_mutex
 885  *
 886  * @lock: the rt_mutex to be unlocked
 887  */
 888 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 889 {
 890         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 891 }
 892 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 893
 894 /**
 895  * rt_mutex_destroy - mark a mutex unusable
 896  * @lock: the mutex to be destroyed
 897  *
 898  * This function marks the mutex uninitialized, and any subsequent
 899  * use of the mutex is forbidden. The mutex must not be locked when
 900  * this function is called.
 901  */
 902 void rt_mutex_destroy(struct rt_mutex *lock)
 903 {
 904         WARN_ON(rt_mutex_is_locked(lock));
 905 #ifdef CONFIG_DEBUG_RT_MUTEXES
 906         lock->magic = NULL;
 907 #endif
 908 }
 909
 910 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 911
 912 /**
 913  * __rt_mutex_init - initialize the rt lock
 914  *
 915  * @lock: the rt lock to be initialized
 916  *
 917  * Initialize the rt lock to unlocked state.
 918  *
 919  * Initializing of a locked rt lock is not allowed
 920  */
 921 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 922 {
 923         lock->owner = NULL;
 924         raw_spin_lock_init(&lock->wait_lock);
 925         plist_head_init(&lock->wait_list);
 926
 927         debug_rt_mutex_init(lock, name);
 928 }
 929 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 930
 931 /**
 932  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 933  *                              proxy owner
 934  *
 935  * @lock:       the rt_mutex to be locked
 936  * @proxy_owner:the task to set as owner
 937  *
 938  * No locking. Caller has to do serializing itself
 939  * Special API call for PI-futex support
 940  */
 941 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 942                                 struct task_struct *proxy_owner)
 943 {
 944         __rt_mutex_init(lock, NULL);
 945         debug_rt_mutex_proxy_lock(lock, proxy_owner);
 946         rt_mutex_set_owner(lock, proxy_owner);
 947         rt_mutex_deadlock_account_lock(lock, proxy_owner);
 948 }
 949
 950 /**
 951  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 952  *
 953  * @lock:       the rt_mutex to be locked
 954  *
 955  * No locking. Caller has to do serializing itself
 956  * Special API call for PI-futex support
 957  */
 958 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 959                            struct task_struct *proxy_owner)
 960 {
 961         debug_rt_mutex_proxy_unlock(lock);
 962         rt_mutex_set_owner(lock, NULL);
 963         rt_mutex_deadlock_account_unlock(proxy_owner);
 964 }
 965
 966 /**
 967  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
 968  * @lock:               the rt_mutex to take
 969  * @waiter:             the pre-initialized rt_mutex_waiter
 970  * @task:               the task to prepare
 971  * @detect_deadlock:    perform deadlock detection (1) or not (0)
 972  *
 973  * Returns:
 974  *  0 - task blocked on lock
 975  *  1 - acquired the lock for task, caller should wake it up
 976  * <0 - error
 977  *
 978  * Special API call for FUTEX_REQUEUE_PI support.
 979  */
 980 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
 981                               struct rt_mutex_waiter *waiter,
 982                               struct task_struct *task, int detect_deadlock)
 983 {
 984         int ret;
 985
 986         raw_spin_lock(&lock->wait_lock);
 987
 988         if (try_to_take_rt_mutex(lock, task, NULL)) {
 989                 raw_spin_unlock(&lock->wait_lock);
 990                 return 1;
 991         }
 992
 993         ret = task_blocks_on_rt_mutex(lock, waiter, task, detect_deadlock);
 994
 995         if (ret && !rt_mutex_owner(lock)) {
 996                 /*
 997                  * Reset the return value. We might have
 998                  * returned with -EDEADLK and the owner
 999                  * released the lock while we were walking the
1000                  * pi chain.  Let the waiter sort it out.
1001                  */
1002                 ret = 0;
1003         }
1004
1005         if (unlikely(ret))
1006                 remove_waiter(lock, waiter);
1007
1008         raw_spin_unlock(&lock->wait_lock);
1009
1010         debug_rt_mutex_print_deadlock(waiter);
1011
1012         return ret;
1013 }
1014
1015 /**
1016  * rt_mutex_next_owner - return the next owner of the lock
1017  *
1018  * @lock: the rt lock query
1019  *
1020  * Returns the next owner of the lock or NULL
1021  *
1022  * Caller has to serialize against other accessors to the lock
1023  * itself.
1024  *
1025  * Special API call for PI-futex support
1026  */
1027 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1028 {
1029         if (!rt_mutex_has_waiters(lock))
1030                 return NULL;
1031
1032         return rt_mutex_top_waiter(lock)->task;
1033 }
1034
1035 /**
1036  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1037  * @lock:               the rt_mutex we were woken on
1038  * @to:                 the timeout, null if none. hrtimer should already have
1039  *                      been started.
1040  * @waiter:             the pre-initialized rt_mutex_waiter
1041  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1042  *
1043  * Complete the lock acquisition started our behalf by another thread.
1044  *
1045  * Returns:
1046  *  0 - success
1047  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1048  *
1049  * Special API call for PI-futex requeue support
1050  */
1051 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1052                                struct hrtimer_sleeper *to,
1053                                struct rt_mutex_waiter *waiter,
1054                                int detect_deadlock)
1055 {
1056         int ret;
1057
1058         raw_spin_lock(&lock->wait_lock);
1059
1060         set_current_state(TASK_INTERRUPTIBLE);
1061
1062         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1063
1064         set_current_state(TASK_RUNNING);
1065
1066         if (unlikely(ret))
1067                 remove_waiter(lock, waiter);
1068
1069         /*
1070          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1071          * have to fix that up.
1072          */
1073         fixup_rt_mutex_waiters(lock);
1074
1075         raw_spin_unlock(&lock->wait_lock);
1076
1077         return ret;
1078 }