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 /*
  21  * lock->owner state tracking:
  22  *
  23  * lock->owner holds the task_struct pointer of the owner. Bit 0 and 1
  24  * are used to keep track of the "owner is pending" and "lock has
  25  * waiters" state.
  26  *
  27  * owner        bit1    bit0
  28  * NULL         0       0       lock is free (fast acquire possible)
  29  * NULL         0       1       invalid state
  30  * NULL         1       0       Transitional State*
  31  * NULL         1       1       invalid state
  32  * taskpointer  0       0       lock is held (fast release possible)
  33  * taskpointer  0       1       task is pending owner
  34  * taskpointer  1       0       lock is held and has waiters
  35  * taskpointer  1       1       task is pending owner and lock has more waiters
  36  *
  37  * Pending ownership is assigned to the top (highest priority)
  38  * waiter of the lock, when the lock is released. The thread is woken
  39  * up and can now take the lock. Until the lock is taken (bit 0
  40  * cleared) a competing higher priority thread can steal the lock
  41  * which puts the woken up thread back on the waiters list.
  42  *
  43  * The fast atomic compare exchange based acquire and release is only
  44  * possible when bit 0 and 1 of lock->owner are 0.
  45  *
  46  * (*) There's a small time where the owner can be NULL and the
  47  * "lock has waiters" bit is set.  This can happen when grabbing the lock.
  48  * To prevent a cmpxchg of the owner releasing the lock, we need to set this
  49  * bit before looking at the lock, hence the reason this is a transitional
  50  * state.
  51  */
  52
  53 static void
  54 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner,
  55                    unsigned long mask)
  56 {
  57         unsigned long val = (unsigned long)owner | mask;
  58
  59         if (rt_mutex_has_waiters(lock))
  60                 val |= RT_MUTEX_HAS_WAITERS;
  61
  62         lock->owner = (struct task_struct *)val;
  63 }
  64
  65 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  66 {
  67         lock->owner = (struct task_struct *)
  68                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  69 }
  70
  71 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  72 {
  73         if (!rt_mutex_has_waiters(lock))
  74                 clear_rt_mutex_waiters(lock);
  75 }
  76
  77 /*
  78  * We can speed up the acquire/release, if the architecture
  79  * supports cmpxchg and if there's no debugging state to be set up
  80  */
  81 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
  82 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
  83 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  84 {
  85         unsigned long owner, *p = (unsigned long *) &lock->owner;
  86
  87         do {
  88                 owner = *p;
  89         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
  90 }
  91 #else
  92 # define rt_mutex_cmpxchg(l,c,n)        (0)
  93 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  94 {
  95         lock->owner = (struct task_struct *)
  96                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
  97 }
  98 #endif
  99
 100 /*
 101  * Calculate task priority from the waiter list priority
 102  *
 103  * Return task->normal_prio when the waiter list is empty or when
 104  * the waiter is not allowed to do priority boosting
 105  */
 106 int rt_mutex_getprio(struct task_struct *task)
 107 {
 108         if (likely(!task_has_pi_waiters(task)))
 109                 return task->normal_prio;
 110
 111         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
 112                    task->normal_prio);
 113 }
 114
 115 /*
 116  * Adjust the priority of a task, after its pi_waiters got modified.
 117  *
 118  * This can be both boosting and unboosting. task->pi_lock must be held.
 119  */
 120 static void __rt_mutex_adjust_prio(struct task_struct *task)
 121 {
 122         int prio = rt_mutex_getprio(task);
 123
 124         if (task->prio != prio)
 125                 rt_mutex_setprio(task, prio);
 126 }
 127
 128 /*
 129  * Adjust task priority (undo boosting). Called from the exit path of
 130  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 131  *
 132  * (Note: We do this outside of the protection of lock->wait_lock to
 133  * allow the lock to be taken while or before we readjust the priority
 134  * of task. We do not use the spin_xx_mutex() variants here as we are
 135  * outside of the debug path.)
 136  */
 137 static void rt_mutex_adjust_prio(struct task_struct *task)
 138 {
 139         unsigned long flags;
 140
 141         spin_lock_irqsave(&task->pi_lock, flags);
 142         __rt_mutex_adjust_prio(task);
 143         spin_unlock_irqrestore(&task->pi_lock, flags);
 144 }
 145
 146 /*
 147  * Max number of times we'll walk the boosting chain:
 148  */
 149 int max_lock_depth = 1024;
 150
 151 /*
 152  * Adjust the priority chain. Also used for deadlock detection.
 153  * Decreases task's usage by one - may thus free the task.
 154  * Returns 0 or -EDEADLK.
 155  */
 156 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 157                                       int deadlock_detect,
 158                                       struct rt_mutex *orig_lock,
 159                                       struct rt_mutex_waiter *orig_waiter,
 160                                       struct task_struct *top_task)
 161 {
 162         struct rt_mutex *lock;
 163         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 164         int detect_deadlock, ret = 0, depth = 0;
 165         unsigned long flags;
 166
 167         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 168                                                          deadlock_detect);
 169
 170         /*
 171          * The (de)boosting is a step by step approach with a lot of
 172          * pitfalls. We want this to be preemptible and we want hold a
 173          * maximum of two locks per step. So we have to check
 174          * carefully whether things change under us.
 175          */
 176  again:
 177         if (++depth > max_lock_depth) {
 178                 static int prev_max;
 179
 180                 /*
 181                  * Print this only once. If the admin changes the limit,
 182                  * print a new message when reaching the limit again.
 183                  */
 184                 if (prev_max != max_lock_depth) {
 185                         prev_max = max_lock_depth;
 186                         printk(KERN_WARNING "Maximum lock depth %d reached "
 187                                "task: %s (%d)\n", max_lock_depth,
 188                                top_task->comm, task_pid_nr(top_task));
 189                 }
 190                 put_task_struct(task);
 191
 192                 return deadlock_detect ? -EDEADLK : 0;
 193         }
 194  retry:
 195         /*
 196          * Task can not go away as we did a get_task() before !
 197          */
 198         spin_lock_irqsave(&task->pi_lock, flags);
 199
 200         waiter = task->pi_blocked_on;
 201         /*
 202          * Check whether the end of the boosting chain has been
 203          * reached or the state of the chain has changed while we
 204          * dropped the locks.
 205          */
 206         if (!waiter || !waiter->task)
 207                 goto out_unlock_pi;
 208
 209         /*
 210          * Check the orig_waiter state. After we dropped the locks,
 211          * the previous owner of the lock might have released the lock
 212          * and made us the pending owner:
 213          */
 214         if (orig_waiter && !orig_waiter->task)
 215                 goto out_unlock_pi;
 216
 217         /*
 218          * Drop out, when the task has no waiters. Note,
 219          * top_waiter can be NULL, when we are in the deboosting
 220          * mode!
 221          */
 222         if (top_waiter && (!task_has_pi_waiters(task) ||
 223                            top_waiter != task_top_pi_waiter(task)))
 224                 goto out_unlock_pi;
 225
 226         /*
 227          * When deadlock detection is off then we check, if further
 228          * priority adjustment is necessary.
 229          */
 230         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 231                 goto out_unlock_pi;
 232
 233         lock = waiter->lock;
 234         if (!spin_trylock(&lock->wait_lock)) {
 235                 spin_unlock_irqrestore(&task->pi_lock, flags);
 236                 cpu_relax();
 237                 goto retry;
 238         }
 239
 240         /* Deadlock detection */
 241         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 242                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 243                 spin_unlock(&lock->wait_lock);
 244                 ret = deadlock_detect ? -EDEADLK : 0;
 245                 goto out_unlock_pi;
 246         }
 247
 248         top_waiter = rt_mutex_top_waiter(lock);
 249
 250         /* Requeue the waiter */
 251         plist_del(&waiter->list_entry, &lock->wait_list);
 252         waiter->list_entry.prio = task->prio;
 253         plist_add(&waiter->list_entry, &lock->wait_list);
 254
 255         /* Release the task */
 256         spin_unlock_irqrestore(&task->pi_lock, flags);
 257         put_task_struct(task);
 258
 259         /* Grab the next task */
 260         task = rt_mutex_owner(lock);
 261         get_task_struct(task);
 262         spin_lock_irqsave(&task->pi_lock, flags);
 263
 264         if (waiter == rt_mutex_top_waiter(lock)) {
 265                 /* Boost the owner */
 266                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 267                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 268                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 269                 __rt_mutex_adjust_prio(task);
 270
 271         } else if (top_waiter == waiter) {
 272                 /* Deboost the owner */
 273                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 274                 waiter = rt_mutex_top_waiter(lock);
 275                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 276                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 277                 __rt_mutex_adjust_prio(task);
 278         }
 279
 280         spin_unlock_irqrestore(&task->pi_lock, flags);
 281
 282         top_waiter = rt_mutex_top_waiter(lock);
 283         spin_unlock(&lock->wait_lock);
 284
 285         if (!detect_deadlock && waiter != top_waiter)
 286                 goto out_put_task;
 287
 288         goto again;
 289
 290  out_unlock_pi:
 291         spin_unlock_irqrestore(&task->pi_lock, flags);
 292  out_put_task:
 293         put_task_struct(task);
 294
 295         return ret;
 296 }
 297
 298 /*
 299  * Optimization: check if we can steal the lock from the
 300  * assigned pending owner [which might not have taken the
 301  * lock yet]:
 302  */
 303 static inline int try_to_steal_lock(struct rt_mutex *lock)
 304 {
 305         struct task_struct *pendowner = rt_mutex_owner(lock);
 306         struct rt_mutex_waiter *next;
 307         unsigned long flags;
 308
 309         if (!rt_mutex_owner_pending(lock))
 310                 return 0;
 311
 312         if (pendowner == current)
 313                 return 1;
 314
 315         spin_lock_irqsave(&pendowner->pi_lock, flags);
 316         if (current->prio >= pendowner->prio) {
 317                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 318                 return 0;
 319         }
 320
 321         /*
 322          * Check if a waiter is enqueued on the pending owners
 323          * pi_waiters list. Remove it and readjust pending owners
 324          * priority.
 325          */
 326         if (likely(!rt_mutex_has_waiters(lock))) {
 327                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 328                 return 1;
 329         }
 330
 331         /* No chain handling, pending owner is not blocked on anything: */
 332         next = rt_mutex_top_waiter(lock);
 333         plist_del(&next->pi_list_entry, &pendowner->pi_waiters);
 334         __rt_mutex_adjust_prio(pendowner);
 335         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 336
 337         /*
 338          * We are going to steal the lock and a waiter was
 339          * enqueued on the pending owners pi_waiters queue. So
 340          * we have to enqueue this waiter into
 341          * current->pi_waiters list. This covers the case,
 342          * where current is boosted because it holds another
 343          * lock and gets unboosted because the booster is
 344          * interrupted, so we would delay a waiter with higher
 345          * priority as current->normal_prio.
 346          *
 347          * Note: in the rare case of a SCHED_OTHER task changing
 348          * its priority and thus stealing the lock, next->task
 349          * might be current:
 350          */
 351         if (likely(next->task != current)) {
 352                 spin_lock_irqsave(&current->pi_lock, flags);
 353                 plist_add(&next->pi_list_entry, &current->pi_waiters);
 354                 __rt_mutex_adjust_prio(current);
 355                 spin_unlock_irqrestore(&current->pi_lock, flags);
 356         }
 357         return 1;
 358 }
 359
 360 /*
 361  * Try to take an rt-mutex
 362  *
 363  * This fails
 364  * - when the lock has a real owner
 365  * - when a different pending owner exists and has higher priority than current
 366  *
 367  * Must be called with lock->wait_lock held.
 368  */
 369 static int try_to_take_rt_mutex(struct rt_mutex *lock)
 370 {
 371         /*
 372          * We have to be careful here if the atomic speedups are
 373          * enabled, such that, when
 374          *  - no other waiter is on the lock
 375          *  - the lock has been released since we did the cmpxchg
 376          * the lock can be released or taken while we are doing the
 377          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 378          *
 379          * The atomic acquire/release aware variant of
 380          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 381          * the WAITERS bit, the atomic release / acquire can not
 382          * happen anymore and lock->wait_lock protects us from the
 383          * non-atomic case.
 384          *
 385          * Note, that this might set lock->owner =
 386          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 387          * any more. This is fixed up when we take the ownership.
 388          * This is the transitional state explained at the top of this file.
 389          */
 390         mark_rt_mutex_waiters(lock);
 391
 392         if (rt_mutex_owner(lock) && !try_to_steal_lock(lock))
 393                 return 0;
 394
 395         /* We got the lock. */
 396         debug_rt_mutex_lock(lock);
 397
 398         rt_mutex_set_owner(lock, current, 0);
 399
 400         rt_mutex_deadlock_account_lock(lock, current);
 401
 402         return 1;
 403 }
 404
 405 /*
 406  * Task blocks on lock.
 407  *
 408  * Prepare waiter and propagate pi chain
 409  *
 410  * This must be called with lock->wait_lock held.
 411  */
 412 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 413                                    struct rt_mutex_waiter *waiter,
 414                                    int detect_deadlock)
 415 {
 416         struct task_struct *owner = rt_mutex_owner(lock);
 417         struct rt_mutex_waiter *top_waiter = waiter;
 418         unsigned long flags;
 419         int chain_walk = 0, res;
 420
 421         spin_lock_irqsave(&current->pi_lock, flags);
 422         __rt_mutex_adjust_prio(current);
 423         waiter->task = current;
 424         waiter->lock = lock;
 425         plist_node_init(&waiter->list_entry, current->prio);
 426         plist_node_init(&waiter->pi_list_entry, current->prio);
 427
 428         /* Get the top priority waiter on the lock */
 429         if (rt_mutex_has_waiters(lock))
 430                 top_waiter = rt_mutex_top_waiter(lock);
 431         plist_add(&waiter->list_entry, &lock->wait_list);
 432
 433         current->pi_blocked_on = waiter;
 434
 435         spin_unlock_irqrestore(&current->pi_lock, flags);
 436
 437         if (waiter == rt_mutex_top_waiter(lock)) {
 438                 spin_lock_irqsave(&owner->pi_lock, flags);
 439                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 440                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 441
 442                 __rt_mutex_adjust_prio(owner);
 443                 if (owner->pi_blocked_on)
 444                         chain_walk = 1;
 445                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 446         }
 447         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 448                 chain_walk = 1;
 449
 450         if (!chain_walk)
 451                 return 0;
 452
 453         /*
 454          * The owner can't disappear while holding a lock,
 455          * so the owner struct is protected by wait_lock.
 456          * Gets dropped in rt_mutex_adjust_prio_chain()!
 457          */
 458         get_task_struct(owner);
 459
 460         spin_unlock(&lock->wait_lock);
 461
 462         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 463                                          current);
 464
 465         spin_lock(&lock->wait_lock);
 466
 467         return res;
 468 }
 469
 470 /*
 471  * Wake up the next waiter on the lock.
 472  *
 473  * Remove the top waiter from the current tasks waiter list and from
 474  * the lock waiter list. Set it as pending owner. Then wake it up.
 475  *
 476  * Called with lock->wait_lock held.
 477  */
 478 static void wakeup_next_waiter(struct rt_mutex *lock)
 479 {
 480         struct rt_mutex_waiter *waiter;
 481         struct task_struct *pendowner;
 482         unsigned long flags;
 483
 484         spin_lock_irqsave(&current->pi_lock, flags);
 485
 486         waiter = rt_mutex_top_waiter(lock);
 487         plist_del(&waiter->list_entry, &lock->wait_list);
 488
 489         /*
 490          * Remove it from current->pi_waiters. We do not adjust a
 491          * possible priority boost right now. We execute wakeup in the
 492          * boosted mode and go back to normal after releasing
 493          * lock->wait_lock.
 494          */
 495         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 496         pendowner = waiter->task;
 497         waiter->task = NULL;
 498
 499         rt_mutex_set_owner(lock, pendowner, RT_MUTEX_OWNER_PENDING);
 500
 501         spin_unlock_irqrestore(&current->pi_lock, flags);
 502
 503         /*
 504          * Clear the pi_blocked_on variable and enqueue a possible
 505          * waiter into the pi_waiters list of the pending owner. This
 506          * prevents that in case the pending owner gets unboosted a
 507          * waiter with higher priority than pending-owner->normal_prio
 508          * is blocked on the unboosted (pending) owner.
 509          */
 510         spin_lock_irqsave(&pendowner->pi_lock, flags);
 511
 512         WARN_ON(!pendowner->pi_blocked_on);
 513         WARN_ON(pendowner->pi_blocked_on != waiter);
 514         WARN_ON(pendowner->pi_blocked_on->lock != lock);
 515
 516         pendowner->pi_blocked_on = NULL;
 517
 518         if (rt_mutex_has_waiters(lock)) {
 519                 struct rt_mutex_waiter *next;
 520
 521                 next = rt_mutex_top_waiter(lock);
 522                 plist_add(&next->pi_list_entry, &pendowner->pi_waiters);
 523         }
 524         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 525
 526         wake_up_process(pendowner);
 527 }
 528
 529 /*
 530  * Remove a waiter from a lock
 531  *
 532  * Must be called with lock->wait_lock held
 533  */
 534 static void remove_waiter(struct rt_mutex *lock,
 535                           struct rt_mutex_waiter *waiter)
 536 {
 537         int first = (waiter == rt_mutex_top_waiter(lock));
 538         struct task_struct *owner = rt_mutex_owner(lock);
 539         unsigned long flags;
 540         int chain_walk = 0;
 541
 542         spin_lock_irqsave(&current->pi_lock, flags);
 543         plist_del(&waiter->list_entry, &lock->wait_list);
 544         waiter->task = NULL;
 545         current->pi_blocked_on = NULL;
 546         spin_unlock_irqrestore(&current->pi_lock, flags);
 547
 548         if (first && owner != current) {
 549
 550                 spin_lock_irqsave(&owner->pi_lock, flags);
 551
 552                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 553
 554                 if (rt_mutex_has_waiters(lock)) {
 555                         struct rt_mutex_waiter *next;
 556
 557                         next = rt_mutex_top_waiter(lock);
 558                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 559                 }
 560                 __rt_mutex_adjust_prio(owner);
 561
 562                 if (owner->pi_blocked_on)
 563                         chain_walk = 1;
 564
 565                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 566         }
 567
 568         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 569
 570         if (!chain_walk)
 571                 return;
 572
 573         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 574         get_task_struct(owner);
 575
 576         spin_unlock(&lock->wait_lock);
 577
 578         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 579
 580         spin_lock(&lock->wait_lock);
 581 }
 582
 583 /*
 584  * Recheck the pi chain, in case we got a priority setting
 585  *
 586  * Called from sched_setscheduler
 587  */
 588 void rt_mutex_adjust_pi(struct task_struct *task)
 589 {
 590         struct rt_mutex_waiter *waiter;
 591         unsigned long flags;
 592
 593         spin_lock_irqsave(&task->pi_lock, flags);
 594
 595         waiter = task->pi_blocked_on;
 596         if (!waiter || waiter->list_entry.prio == task->prio) {
 597                 spin_unlock_irqrestore(&task->pi_lock, flags);
 598                 return;
 599         }
 600
 601         spin_unlock_irqrestore(&task->pi_lock, flags);
 602
 603         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 604         get_task_struct(task);
 605         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 606 }
 607
 608 /*
 609  * Slow path lock function:
 610  */
 611 static int __sched
 612 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 613                   struct hrtimer_sleeper *timeout,
 614                   int detect_deadlock)
 615 {
 616         struct rt_mutex_waiter waiter;
 617         int ret = 0;
 618
 619         debug_rt_mutex_init_waiter(&waiter);
 620         waiter.task = NULL;
 621
 622         spin_lock(&lock->wait_lock);
 623
 624         /* Try to acquire the lock again: */
 625         if (try_to_take_rt_mutex(lock)) {
 626                 spin_unlock(&lock->wait_lock);
 627                 return 0;
 628         }
 629
 630         set_current_state(state);
 631
 632         /* Setup the timer, when timeout != NULL */
 633         if (unlikely(timeout)) {
 634                 hrtimer_start(&timeout->timer, timeout->timer.expires,
 635                               HRTIMER_MODE_ABS);
 636                 if (!hrtimer_active(&timeout->timer))
 637                         timeout->task = NULL;
 638         }
 639
 640         for (;;) {
 641                 /* Try to acquire the lock: */
 642                 if (try_to_take_rt_mutex(lock))
 643                         break;
 644
 645                 /*
 646                  * TASK_INTERRUPTIBLE checks for signals and
 647                  * timeout. Ignored otherwise.
 648                  */
 649                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 650                         /* Signal pending? */
 651                         if (signal_pending(current))
 652                                 ret = -EINTR;
 653                         if (timeout && !timeout->task)
 654                                 ret = -ETIMEDOUT;
 655                         if (ret)
 656                                 break;
 657                 }
 658
 659                 /*
 660                  * waiter.task is NULL the first time we come here and
 661                  * when we have been woken up by the previous owner
 662                  * but the lock got stolen by a higher prio task.
 663                  */
 664                 if (!waiter.task) {
 665                         ret = task_blocks_on_rt_mutex(lock, &waiter,
 666                                                       detect_deadlock);
 667                         /*
 668                          * If we got woken up by the owner then start loop
 669                          * all over without going into schedule to try
 670                          * to get the lock now:
 671                          */
 672                         if (unlikely(!waiter.task)) {
 673                                 /*
 674                                  * Reset the return value. We might
 675                                  * have returned with -EDEADLK and the
 676                                  * owner released the lock while we
 677                                  * were walking the pi chain.
 678                                  */
 679                                 ret = 0;
 680                                 continue;
 681                         }
 682                         if (unlikely(ret))
 683                                 break;
 684                 }
 685
 686                 spin_unlock(&lock->wait_lock);
 687
 688                 debug_rt_mutex_print_deadlock(&waiter);
 689
 690                 if (waiter.task)
 691                         schedule_rt_mutex(lock);
 692
 693                 spin_lock(&lock->wait_lock);
 694                 set_current_state(state);
 695         }
 696
 697         set_current_state(TASK_RUNNING);
 698
 699         if (unlikely(waiter.task))
 700                 remove_waiter(lock, &waiter);
 701
 702         /*
 703          * try_to_take_rt_mutex() sets the waiter bit
 704          * unconditionally. We might have to fix that up.
 705          */
 706         fixup_rt_mutex_waiters(lock);
 707
 708         spin_unlock(&lock->wait_lock);
 709
 710         /* Remove pending timer: */
 711         if (unlikely(timeout))
 712                 hrtimer_cancel(&timeout->timer);
 713
 714         /*
 715          * Readjust priority, when we did not get the lock. We might
 716          * have been the pending owner and boosted. Since we did not
 717          * take the lock, the PI boost has to go.
 718          */
 719         if (unlikely(ret))
 720                 rt_mutex_adjust_prio(current);
 721
 722         debug_rt_mutex_free_waiter(&waiter);
 723
 724         return ret;
 725 }
 726
 727 /*
 728  * Slow path try-lock function:
 729  */
 730 static inline int
 731 rt_mutex_slowtrylock(struct rt_mutex *lock)
 732 {
 733         int ret = 0;
 734
 735         spin_lock(&lock->wait_lock);
 736
 737         if (likely(rt_mutex_owner(lock) != current)) {
 738
 739                 ret = try_to_take_rt_mutex(lock);
 740                 /*
 741                  * try_to_take_rt_mutex() sets the lock waiters
 742                  * bit unconditionally. Clean this up.
 743                  */
 744                 fixup_rt_mutex_waiters(lock);
 745         }
 746
 747         spin_unlock(&lock->wait_lock);
 748
 749         return ret;
 750 }
 751
 752 /*
 753  * Slow path to release a rt-mutex:
 754  */
 755 static void __sched
 756 rt_mutex_slowunlock(struct rt_mutex *lock)
 757 {
 758         spin_lock(&lock->wait_lock);
 759
 760         debug_rt_mutex_unlock(lock);
 761
 762         rt_mutex_deadlock_account_unlock(current);
 763
 764         if (!rt_mutex_has_waiters(lock)) {
 765                 lock->owner = NULL;
 766                 spin_unlock(&lock->wait_lock);
 767                 return;
 768         }
 769
 770         wakeup_next_waiter(lock);
 771
 772         spin_unlock(&lock->wait_lock);
 773
 774         /* Undo pi boosting if necessary: */
 775         rt_mutex_adjust_prio(current);
 776 }
 777
 778 /*
 779  * debug aware fast / slowpath lock,trylock,unlock
 780  *
 781  * The atomic acquire/release ops are compiled away, when either the
 782  * architecture does not support cmpxchg or when debugging is enabled.
 783  */
 784 static inline int
 785 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 786                   int detect_deadlock,
 787                   int (*slowfn)(struct rt_mutex *lock, int state,
 788                                 struct hrtimer_sleeper *timeout,
 789                                 int detect_deadlock))
 790 {
 791         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 792                 rt_mutex_deadlock_account_lock(lock, current);
 793                 return 0;
 794         } else
 795                 return slowfn(lock, state, NULL, detect_deadlock);
 796 }
 797
 798 static inline int
 799 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 800                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 801                         int (*slowfn)(struct rt_mutex *lock, int state,
 802                                       struct hrtimer_sleeper *timeout,
 803                                       int detect_deadlock))
 804 {
 805         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 806                 rt_mutex_deadlock_account_lock(lock, current);
 807                 return 0;
 808         } else
 809                 return slowfn(lock, state, timeout, detect_deadlock);
 810 }
 811
 812 static inline int
 813 rt_mutex_fasttrylock(struct rt_mutex *lock,
 814                      int (*slowfn)(struct rt_mutex *lock))
 815 {
 816         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 817                 rt_mutex_deadlock_account_lock(lock, current);
 818                 return 1;
 819         }
 820         return slowfn(lock);
 821 }
 822
 823 static inline void
 824 rt_mutex_fastunlock(struct rt_mutex *lock,
 825                     void (*slowfn)(struct rt_mutex *lock))
 826 {
 827         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 828                 rt_mutex_deadlock_account_unlock(current);
 829         else
 830                 slowfn(lock);
 831 }
 832
 833 /**
 834  * rt_mutex_lock - lock a rt_mutex
 835  *
 836  * @lock: the rt_mutex to be locked
 837  */
 838 void __sched rt_mutex_lock(struct rt_mutex *lock)
 839 {
 840         might_sleep();
 841
 842         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 843 }
 844 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 845
 846 /**
 847  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 848  *
 849  * @lock:               the rt_mutex to be locked
 850  * @detect_deadlock:    deadlock detection on/off
 851  *
 852  * Returns:
 853  *  0           on success
 854  * -EINTR       when interrupted by a signal
 855  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 856  */
 857 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 858                                                  int detect_deadlock)
 859 {
 860         might_sleep();
 861
 862         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 863                                  detect_deadlock, rt_mutex_slowlock);
 864 }
 865 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 866
 867 /**
 868  * rt_mutex_lock_interruptible_ktime - lock a rt_mutex interruptible
 869  *                                     the timeout structure is provided
 870  *                                     by the caller
 871  *
 872  * @lock:               the rt_mutex to be locked
 873  * @timeout:            timeout structure or NULL (no timeout)
 874  * @detect_deadlock:    deadlock detection on/off
 875  *
 876  * Returns:
 877  *  0           on success
 878  * -EINTR       when interrupted by a signal
 879  * -ETIMEOUT    when the timeout expired
 880  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 881  */
 882 int
 883 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 884                     int detect_deadlock)
 885 {
 886         might_sleep();
 887
 888         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 889                                        detect_deadlock, rt_mutex_slowlock);
 890 }
 891 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 892
 893 /**
 894  * rt_mutex_trylock - try to lock a rt_mutex
 895  *
 896  * @lock:       the rt_mutex to be locked
 897  *
 898  * Returns 1 on success and 0 on contention
 899  */
 900 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 901 {
 902         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 903 }
 904 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 905
 906 /**
 907  * rt_mutex_unlock - unlock a rt_mutex
 908  *
 909  * @lock: the rt_mutex to be unlocked
 910  */
 911 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 912 {
 913         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 914 }
 915 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 916
 917 /***
 918  * rt_mutex_destroy - mark a mutex unusable
 919  * @lock: the mutex to be destroyed
 920  *
 921  * This function marks the mutex uninitialized, and any subsequent
 922  * use of the mutex is forbidden. The mutex must not be locked when
 923  * this function is called.
 924  */
 925 void rt_mutex_destroy(struct rt_mutex *lock)
 926 {
 927         WARN_ON(rt_mutex_is_locked(lock));
 928 #ifdef CONFIG_DEBUG_RT_MUTEXES
 929         lock->magic = NULL;
 930 #endif
 931 }
 932
 933 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 934
 935 /**
 936  * __rt_mutex_init - initialize the rt lock
 937  *
 938  * @lock: the rt lock to be initialized
 939  *
 940  * Initialize the rt lock to unlocked state.
 941  *
 942  * Initializing of a locked rt lock is not allowed
 943  */
 944 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 945 {
 946         lock->owner = NULL;
 947         spin_lock_init(&lock->wait_lock);
 948         plist_head_init(&lock->wait_list, &lock->wait_lock);
 949
 950         debug_rt_mutex_init(lock, name);
 951 }
 952 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 953
 954 /**
 955  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 956  *                              proxy owner
 957  *
 958  * @lock:       the rt_mutex to be locked
 959  * @proxy_owner:the task to set as owner
 960  *
 961  * No locking. Caller has to do serializing itself
 962  * Special API call for PI-futex support
 963  */
 964 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 965                                 struct task_struct *proxy_owner)
 966 {
 967         __rt_mutex_init(lock, NULL);
 968         debug_rt_mutex_proxy_lock(lock, proxy_owner);
 969         rt_mutex_set_owner(lock, proxy_owner, 0);
 970         rt_mutex_deadlock_account_lock(lock, proxy_owner);
 971 }
 972
 973 /**
 974  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 975  *
 976  * @lock:       the rt_mutex to be locked
 977  *
 978  * No locking. Caller has to do serializing itself
 979  * Special API call for PI-futex support
 980  */
 981 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 982                            struct task_struct *proxy_owner)
 983 {
 984         debug_rt_mutex_proxy_unlock(lock);
 985         rt_mutex_set_owner(lock, NULL, 0);
 986         rt_mutex_deadlock_account_unlock(proxy_owner);
 987 }
 988
 989 /**
 990  * rt_mutex_next_owner - return the next owner of the lock
 991  *
 992  * @lock: the rt lock query
 993  *
 994  * Returns the next owner of the lock or NULL
 995  *
 996  * Caller has to serialize against other accessors to the lock
 997  * itself.
 998  *
 999  * Special API call for PI-futex support
1000  */
1001 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1002 {
1003         if (!rt_mutex_has_waiters(lock))
1004                 return NULL;
1005
1006         return rt_mutex_top_waiter(lock)->task;
1007 }