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 #include <linux/spinlock.h>
  12 #include <linux/module.h>
  13 #include <linux/sched.h>
  14 #include <linux/timer.h>
  15
  16 #include "rtmutex_common.h"
  17
  18 #ifdef CONFIG_DEBUG_RT_MUTEXES
  19 # include "rtmutex-debug.h"
  20 #else
  21 # include "rtmutex.h"
  22 #endif
  23
  24 /*
  25  * lock->owner state tracking:
  26  *
  27  * lock->owner holds the task_struct pointer of the owner. Bit 0 and 1
  28  * are used to keep track of the "owner is pending" and "lock has
  29  * waiters" state.
  30  *
  31  * owner        bit1    bit0
  32  * NULL         0       0       lock is free (fast acquire possible)
  33  * NULL         0       1       invalid state
  34  * NULL         1       0       Transitional State*
  35  * NULL         1       1       invalid state
  36  * taskpointer  0       0       lock is held (fast release possible)
  37  * taskpointer  0       1       task is pending owner
  38  * taskpointer  1       0       lock is held and has waiters
  39  * taskpointer  1       1       task is pending owner and lock has more waiters
  40  *
  41  * Pending ownership is assigned to the top (highest priority)
  42  * waiter of the lock, when the lock is released. The thread is woken
  43  * up and can now take the lock. Until the lock is taken (bit 0
  44  * cleared) a competing higher priority thread can steal the lock
  45  * which puts the woken up thread back on the waiters list.
  46  *
  47  * The fast atomic compare exchange based acquire and release is only
  48  * possible when bit 0 and 1 of lock->owner are 0.
  49  *
  50  * (*) There's a small time where the owner can be NULL and the
  51  * "lock has waiters" bit is set.  This can happen when grabbing the lock.
  52  * To prevent a cmpxchg of the owner releasing the lock, we need to set this
  53  * bit before looking at the lock, hence the reason this is a transitional
  54  * state.
  55  */
  56
  57 static void
  58 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner,
  59                    unsigned long mask)
  60 {
  61         unsigned long val = (unsigned long)owner | mask;
  62
  63         if (rt_mutex_has_waiters(lock))
  64                 val |= RT_MUTEX_HAS_WAITERS;
  65
  66         lock->owner = (struct task_struct *)val;
  67 }
  68
  69 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  70 {
  71         lock->owner = (struct task_struct *)
  72                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  73 }
  74
  75 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  76 {
  77         if (!rt_mutex_has_waiters(lock))
  78                 clear_rt_mutex_waiters(lock);
  79 }
  80
  81 /*
  82  * We can speed up the acquire/release, if the architecture
  83  * supports cmpxchg and if there's no debugging state to be set up
  84  */
  85 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
  86 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
  87 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  88 {
  89         unsigned long owner, *p = (unsigned long *) &lock->owner;
  90
  91         do {
  92                 owner = *p;
  93         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
  94 }
  95 #else
  96 # define rt_mutex_cmpxchg(l,c,n)        (0)
  97 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  98 {
  99         lock->owner = (struct task_struct *)
 100                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 101 }
 102 #endif
 103
 104 /*
 105  * Calculate task priority from the waiter list priority
 106  *
 107  * Return task->normal_prio when the waiter list is empty or when
 108  * the waiter is not allowed to do priority boosting
 109  */
 110 int rt_mutex_getprio(struct task_struct *task)
 111 {
 112         if (likely(!task_has_pi_waiters(task)))
 113                 return task->normal_prio;
 114
 115         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
 116                    task->normal_prio);
 117 }
 118
 119 /*
 120  * Adjust the priority of a task, after its pi_waiters got modified.
 121  *
 122  * This can be both boosting and unboosting. task->pi_lock must be held.
 123  */
 124 static void __rt_mutex_adjust_prio(struct task_struct *task)
 125 {
 126         int prio = rt_mutex_getprio(task);
 127
 128         if (task->prio != prio)
 129                 rt_mutex_setprio(task, prio);
 130 }
 131
 132 /*
 133  * Adjust task priority (undo boosting). Called from the exit path of
 134  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 135  *
 136  * (Note: We do this outside of the protection of lock->wait_lock to
 137  * allow the lock to be taken while or before we readjust the priority
 138  * of task. We do not use the spin_xx_mutex() variants here as we are
 139  * outside of the debug path.)
 140  */
 141 static void rt_mutex_adjust_prio(struct task_struct *task)
 142 {
 143         unsigned long flags;
 144
 145         spin_lock_irqsave(&task->pi_lock, flags);
 146         __rt_mutex_adjust_prio(task);
 147         spin_unlock_irqrestore(&task->pi_lock, flags);
 148 }
 149
 150 /*
 151  * Max number of times we'll walk the boosting chain:
 152  */
 153 int max_lock_depth = 1024;
 154
 155 /*
 156  * Adjust the priority chain. Also used for deadlock detection.
 157  * Decreases task's usage by one - may thus free the task.
 158  * Returns 0 or -EDEADLK.
 159  */
 160 static int rt_mutex_adjust_prio_chain(task_t *task,
 161                                       int deadlock_detect,
 162                                       struct rt_mutex *orig_lock,
 163                                       struct rt_mutex_waiter *orig_waiter,
 164                                       struct task_struct *top_task)
 165 {
 166         struct rt_mutex *lock;
 167         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 168         int detect_deadlock, ret = 0, depth = 0;
 169         unsigned long flags;
 170
 171         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 172                                                          deadlock_detect);
 173
 174         /*
 175          * The (de)boosting is a step by step approach with a lot of
 176          * pitfalls. We want this to be preemptible and we want hold a
 177          * maximum of two locks per step. So we have to check
 178          * carefully whether things change under us.
 179          */
 180  again:
 181         if (++depth > max_lock_depth) {
 182                 static int prev_max;
 183
 184                 /*
 185                  * Print this only once. If the admin changes the limit,
 186                  * print a new message when reaching the limit again.
 187                  */
 188                 if (prev_max != max_lock_depth) {
 189                         prev_max = max_lock_depth;
 190                         printk(KERN_WARNING "Maximum lock depth %d reached "
 191                                "task: %s (%d)\n", max_lock_depth,
 192                                top_task->comm, top_task->pid);
 193                 }
 194                 put_task_struct(task);
 195
 196                 return deadlock_detect ? -EDEADLK : 0;
 197         }
 198  retry:
 199         /*
 200          * Task can not go away as we did a get_task() before !
 201          */
 202         spin_lock_irqsave(&task->pi_lock, flags);
 203
 204         waiter = task->pi_blocked_on;
 205         /*
 206          * Check whether the end of the boosting chain has been
 207          * reached or the state of the chain has changed while we
 208          * dropped the locks.
 209          */
 210         if (!waiter || !waiter->task)
 211                 goto out_unlock_pi;
 212
 213         if (top_waiter && (!task_has_pi_waiters(task) ||
 214                            top_waiter != task_top_pi_waiter(task)))
 215                 goto out_unlock_pi;
 216
 217         /*
 218          * When deadlock detection is off then we check, if further
 219          * priority adjustment is necessary.
 220          */
 221         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 222                 goto out_unlock_pi;
 223
 224         lock = waiter->lock;
 225         if (!spin_trylock(&lock->wait_lock)) {
 226                 spin_unlock_irqrestore(&task->pi_lock, flags);
 227                 cpu_relax();
 228                 goto retry;
 229         }
 230
 231         /* Deadlock detection */
 232         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 233                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 234                 spin_unlock(&lock->wait_lock);
 235                 ret = deadlock_detect ? -EDEADLK : 0;
 236                 goto out_unlock_pi;
 237         }
 238
 239         top_waiter = rt_mutex_top_waiter(lock);
 240
 241         /* Requeue the waiter */
 242         plist_del(&waiter->list_entry, &lock->wait_list);
 243         waiter->list_entry.prio = task->prio;
 244         plist_add(&waiter->list_entry, &lock->wait_list);
 245
 246         /* Release the task */
 247         spin_unlock_irqrestore(&task->pi_lock, flags);
 248         put_task_struct(task);
 249
 250         /* Grab the next task */
 251         task = rt_mutex_owner(lock);
 252         spin_lock_irqsave(&task->pi_lock, flags);
 253
 254         if (waiter == rt_mutex_top_waiter(lock)) {
 255                 /* Boost the owner */
 256                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 257                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 258                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 259                 __rt_mutex_adjust_prio(task);
 260
 261         } else if (top_waiter == waiter) {
 262                 /* Deboost the owner */
 263                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 264                 waiter = rt_mutex_top_waiter(lock);
 265                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 266                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 267                 __rt_mutex_adjust_prio(task);
 268         }
 269
 270         get_task_struct(task);
 271         spin_unlock_irqrestore(&task->pi_lock, flags);
 272
 273         top_waiter = rt_mutex_top_waiter(lock);
 274         spin_unlock(&lock->wait_lock);
 275
 276         if (!detect_deadlock && waiter != top_waiter)
 277                 goto out_put_task;
 278
 279         goto again;
 280
 281  out_unlock_pi:
 282         spin_unlock_irqrestore(&task->pi_lock, flags);
 283  out_put_task:
 284         put_task_struct(task);
 285         return ret;
 286 }
 287
 288 /*
 289  * Optimization: check if we can steal the lock from the
 290  * assigned pending owner [which might not have taken the
 291  * lock yet]:
 292  */
 293 static inline int try_to_steal_lock(struct rt_mutex *lock)
 294 {
 295         struct task_struct *pendowner = rt_mutex_owner(lock);
 296         struct rt_mutex_waiter *next;
 297         unsigned long flags;
 298
 299         if (!rt_mutex_owner_pending(lock))
 300                 return 0;
 301
 302         if (pendowner == current)
 303                 return 1;
 304
 305         spin_lock_irqsave(&pendowner->pi_lock, flags);
 306         if (current->prio >= pendowner->prio) {
 307                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 308                 return 0;
 309         }
 310
 311         /*
 312          * Check if a waiter is enqueued on the pending owners
 313          * pi_waiters list. Remove it and readjust pending owners
 314          * priority.
 315          */
 316         if (likely(!rt_mutex_has_waiters(lock))) {
 317                 spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 318                 return 1;
 319         }
 320
 321         /* No chain handling, pending owner is not blocked on anything: */
 322         next = rt_mutex_top_waiter(lock);
 323         plist_del(&next->pi_list_entry, &pendowner->pi_waiters);
 324         __rt_mutex_adjust_prio(pendowner);
 325         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 326
 327         /*
 328          * We are going to steal the lock and a waiter was
 329          * enqueued on the pending owners pi_waiters queue. So
 330          * we have to enqueue this waiter into
 331          * current->pi_waiters list. This covers the case,
 332          * where current is boosted because it holds another
 333          * lock and gets unboosted because the booster is
 334          * interrupted, so we would delay a waiter with higher
 335          * priority as current->normal_prio.
 336          *
 337          * Note: in the rare case of a SCHED_OTHER task changing
 338          * its priority and thus stealing the lock, next->task
 339          * might be current:
 340          */
 341         if (likely(next->task != current)) {
 342                 spin_lock_irqsave(&current->pi_lock, flags);
 343                 plist_add(&next->pi_list_entry, &current->pi_waiters);
 344                 __rt_mutex_adjust_prio(current);
 345                 spin_unlock_irqrestore(&current->pi_lock, flags);
 346         }
 347         return 1;
 348 }
 349
 350 /*
 351  * Try to take an rt-mutex
 352  *
 353  * This fails
 354  * - when the lock has a real owner
 355  * - when a different pending owner exists and has higher priority than current
 356  *
 357  * Must be called with lock->wait_lock held.
 358  */
 359 static int try_to_take_rt_mutex(struct rt_mutex *lock)
 360 {
 361         /*
 362          * We have to be careful here if the atomic speedups are
 363          * enabled, such that, when
 364          *  - no other waiter is on the lock
 365          *  - the lock has been released since we did the cmpxchg
 366          * the lock can be released or taken while we are doing the
 367          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 368          *
 369          * The atomic acquire/release aware variant of
 370          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 371          * the WAITERS bit, the atomic release / acquire can not
 372          * happen anymore and lock->wait_lock protects us from the
 373          * non-atomic case.
 374          *
 375          * Note, that this might set lock->owner =
 376          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 377          * any more. This is fixed up when we take the ownership.
 378          * This is the transitional state explained at the top of this file.
 379          */
 380         mark_rt_mutex_waiters(lock);
 381
 382         if (rt_mutex_owner(lock) && !try_to_steal_lock(lock))
 383                 return 0;
 384
 385         /* We got the lock. */
 386         debug_rt_mutex_lock(lock);
 387
 388         rt_mutex_set_owner(lock, current, 0);
 389
 390         rt_mutex_deadlock_account_lock(lock, current);
 391
 392         return 1;
 393 }
 394
 395 /*
 396  * Task blocks on lock.
 397  *
 398  * Prepare waiter and propagate pi chain
 399  *
 400  * This must be called with lock->wait_lock held.
 401  */
 402 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 403                                    struct rt_mutex_waiter *waiter,
 404                                    int detect_deadlock)
 405 {
 406         struct rt_mutex_waiter *top_waiter = waiter;
 407         task_t *owner = rt_mutex_owner(lock);
 408         int boost = 0, res;
 409         unsigned long flags;
 410
 411         spin_lock_irqsave(&current->pi_lock, flags);
 412         __rt_mutex_adjust_prio(current);
 413         waiter->task = current;
 414         waiter->lock = lock;
 415         plist_node_init(&waiter->list_entry, current->prio);
 416         plist_node_init(&waiter->pi_list_entry, current->prio);
 417
 418         /* Get the top priority waiter on the lock */
 419         if (rt_mutex_has_waiters(lock))
 420                 top_waiter = rt_mutex_top_waiter(lock);
 421         plist_add(&waiter->list_entry, &lock->wait_list);
 422
 423         current->pi_blocked_on = waiter;
 424
 425         spin_unlock_irqrestore(&current->pi_lock, flags);
 426
 427         if (waiter == rt_mutex_top_waiter(lock)) {
 428                 spin_lock_irqsave(&owner->pi_lock, flags);
 429                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 430                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 431
 432                 __rt_mutex_adjust_prio(owner);
 433                 if (owner->pi_blocked_on) {
 434                         boost = 1;
 435                         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 436                         get_task_struct(owner);
 437                 }
 438                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 439         }
 440         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock)) {
 441                 spin_lock_irqsave(&owner->pi_lock, flags);
 442                 if (owner->pi_blocked_on) {
 443                         boost = 1;
 444                         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 445                         get_task_struct(owner);
 446                 }
 447                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 448         }
 449         if (!boost)
 450                 return 0;
 451
 452         spin_unlock(&lock->wait_lock);
 453
 454         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 455                                          current);
 456
 457         spin_lock(&lock->wait_lock);
 458
 459         return res;
 460 }
 461
 462 /*
 463  * Wake up the next waiter on the lock.
 464  *
 465  * Remove the top waiter from the current tasks waiter list and from
 466  * the lock waiter list. Set it as pending owner. Then wake it up.
 467  *
 468  * Called with lock->wait_lock held.
 469  */
 470 static void wakeup_next_waiter(struct rt_mutex *lock)
 471 {
 472         struct rt_mutex_waiter *waiter;
 473         struct task_struct *pendowner;
 474         unsigned long flags;
 475
 476         spin_lock_irqsave(&current->pi_lock, flags);
 477
 478         waiter = rt_mutex_top_waiter(lock);
 479         plist_del(&waiter->list_entry, &lock->wait_list);
 480
 481         /*
 482          * Remove it from current->pi_waiters. We do not adjust a
 483          * possible priority boost right now. We execute wakeup in the
 484          * boosted mode and go back to normal after releasing
 485          * lock->wait_lock.
 486          */
 487         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 488         pendowner = waiter->task;
 489         waiter->task = NULL;
 490
 491         rt_mutex_set_owner(lock, pendowner, RT_MUTEX_OWNER_PENDING);
 492
 493         spin_unlock_irqrestore(&current->pi_lock, flags);
 494
 495         /*
 496          * Clear the pi_blocked_on variable and enqueue a possible
 497          * waiter into the pi_waiters list of the pending owner. This
 498          * prevents that in case the pending owner gets unboosted a
 499          * waiter with higher priority than pending-owner->normal_prio
 500          * is blocked on the unboosted (pending) owner.
 501          */
 502         spin_lock_irqsave(&pendowner->pi_lock, flags);
 503
 504         WARN_ON(!pendowner->pi_blocked_on);
 505         WARN_ON(pendowner->pi_blocked_on != waiter);
 506         WARN_ON(pendowner->pi_blocked_on->lock != lock);
 507
 508         pendowner->pi_blocked_on = NULL;
 509
 510         if (rt_mutex_has_waiters(lock)) {
 511                 struct rt_mutex_waiter *next;
 512
 513                 next = rt_mutex_top_waiter(lock);
 514                 plist_add(&next->pi_list_entry, &pendowner->pi_waiters);
 515         }
 516         spin_unlock_irqrestore(&pendowner->pi_lock, flags);
 517
 518         wake_up_process(pendowner);
 519 }
 520
 521 /*
 522  * Remove a waiter from a lock
 523  *
 524  * Must be called with lock->wait_lock held
 525  */
 526 static void remove_waiter(struct rt_mutex *lock,
 527                           struct rt_mutex_waiter *waiter)
 528 {
 529         int first = (waiter == rt_mutex_top_waiter(lock));
 530         int boost = 0;
 531         task_t *owner = rt_mutex_owner(lock);
 532         unsigned long flags;
 533
 534         spin_lock_irqsave(&current->pi_lock, flags);
 535         plist_del(&waiter->list_entry, &lock->wait_list);
 536         waiter->task = NULL;
 537         current->pi_blocked_on = NULL;
 538         spin_unlock_irqrestore(&current->pi_lock, flags);
 539
 540         if (first && owner != current) {
 541
 542                 spin_lock_irqsave(&owner->pi_lock, flags);
 543
 544                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 545
 546                 if (rt_mutex_has_waiters(lock)) {
 547                         struct rt_mutex_waiter *next;
 548
 549                         next = rt_mutex_top_waiter(lock);
 550                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 551                 }
 552                 __rt_mutex_adjust_prio(owner);
 553
 554                 if (owner->pi_blocked_on) {
 555                         boost = 1;
 556                         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 557                         get_task_struct(owner);
 558                 }
 559                 spin_unlock_irqrestore(&owner->pi_lock, flags);
 560         }
 561
 562         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 563
 564         if (!boost)
 565                 return;
 566
 567         spin_unlock(&lock->wait_lock);
 568
 569         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 570
 571         spin_lock(&lock->wait_lock);
 572 }
 573
 574 /*
 575  * Recheck the pi chain, in case we got a priority setting
 576  *
 577  * Called from sched_setscheduler
 578  */
 579 void rt_mutex_adjust_pi(struct task_struct *task)
 580 {
 581         struct rt_mutex_waiter *waiter;
 582         unsigned long flags;
 583
 584         spin_lock_irqsave(&task->pi_lock, flags);
 585
 586         waiter = task->pi_blocked_on;
 587         if (!waiter || waiter->list_entry.prio == task->prio) {
 588                 spin_unlock_irqrestore(&task->pi_lock, flags);
 589                 return;
 590         }
 591
 592         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 593         get_task_struct(task);
 594         spin_unlock_irqrestore(&task->pi_lock, flags);
 595
 596         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 597 }
 598
 599 /*
 600  * Slow path lock function:
 601  */
 602 static int __sched
 603 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 604                   struct hrtimer_sleeper *timeout,
 605                   int detect_deadlock)
 606 {
 607         struct rt_mutex_waiter waiter;
 608         int ret = 0;
 609
 610         debug_rt_mutex_init_waiter(&waiter);
 611         waiter.task = NULL;
 612
 613         spin_lock(&lock->wait_lock);
 614
 615         /* Try to acquire the lock again: */
 616         if (try_to_take_rt_mutex(lock)) {
 617                 spin_unlock(&lock->wait_lock);
 618                 return 0;
 619         }
 620
 621         set_current_state(state);
 622
 623         /* Setup the timer, when timeout != NULL */
 624         if (unlikely(timeout))
 625                 hrtimer_start(&timeout->timer, timeout->timer.expires,
 626                               HRTIMER_ABS);
 627
 628         for (;;) {
 629                 /* Try to acquire the lock: */
 630                 if (try_to_take_rt_mutex(lock))
 631                         break;
 632
 633                 /*
 634                  * TASK_INTERRUPTIBLE checks for signals and
 635                  * timeout. Ignored otherwise.
 636                  */
 637                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 638                         /* Signal pending? */
 639                         if (signal_pending(current))
 640                                 ret = -EINTR;
 641                         if (timeout && !timeout->task)
 642                                 ret = -ETIMEDOUT;
 643                         if (ret)
 644                                 break;
 645                 }
 646
 647                 /*
 648                  * waiter.task is NULL the first time we come here and
 649                  * when we have been woken up by the previous owner
 650                  * but the lock got stolen by a higher prio task.
 651                  */
 652                 if (!waiter.task) {
 653                         ret = task_blocks_on_rt_mutex(lock, &waiter,
 654                                                       detect_deadlock);
 655                         /*
 656                          * If we got woken up by the owner then start loop
 657                          * all over without going into schedule to try
 658                          * to get the lock now:
 659                          */
 660                         if (unlikely(!waiter.task))
 661                                 continue;
 662
 663                         if (unlikely(ret))
 664                                 break;
 665                 }
 666
 667                 spin_unlock(&lock->wait_lock);
 668
 669                 debug_rt_mutex_print_deadlock(&waiter);
 670
 671                 if (waiter.task)
 672                         schedule_rt_mutex(lock);
 673
 674                 spin_lock(&lock->wait_lock);
 675                 set_current_state(state);
 676         }
 677
 678         set_current_state(TASK_RUNNING);
 679
 680         if (unlikely(waiter.task))
 681                 remove_waiter(lock, &waiter);
 682
 683         /*
 684          * try_to_take_rt_mutex() sets the waiter bit
 685          * unconditionally. We might have to fix that up.
 686          */
 687         fixup_rt_mutex_waiters(lock);
 688
 689         spin_unlock(&lock->wait_lock);
 690
 691         /* Remove pending timer: */
 692         if (unlikely(timeout))
 693                 hrtimer_cancel(&timeout->timer);
 694
 695         /*
 696          * Readjust priority, when we did not get the lock. We might
 697          * have been the pending owner and boosted. Since we did not
 698          * take the lock, the PI boost has to go.
 699          */
 700         if (unlikely(ret))
 701                 rt_mutex_adjust_prio(current);
 702
 703         debug_rt_mutex_free_waiter(&waiter);
 704
 705         return ret;
 706 }
 707
 708 /*
 709  * Slow path try-lock function:
 710  */
 711 static inline int
 712 rt_mutex_slowtrylock(struct rt_mutex *lock)
 713 {
 714         int ret = 0;
 715
 716         spin_lock(&lock->wait_lock);
 717
 718         if (likely(rt_mutex_owner(lock) != current)) {
 719
 720                 ret = try_to_take_rt_mutex(lock);
 721                 /*
 722                  * try_to_take_rt_mutex() sets the lock waiters
 723                  * bit unconditionally. Clean this up.
 724                  */
 725                 fixup_rt_mutex_waiters(lock);
 726         }
 727
 728         spin_unlock(&lock->wait_lock);
 729
 730         return ret;
 731 }
 732
 733 /*
 734  * Slow path to release a rt-mutex:
 735  */
 736 static void __sched
 737 rt_mutex_slowunlock(struct rt_mutex *lock)
 738 {
 739         spin_lock(&lock->wait_lock);
 740
 741         debug_rt_mutex_unlock(lock);
 742
 743         rt_mutex_deadlock_account_unlock(current);
 744
 745         if (!rt_mutex_has_waiters(lock)) {
 746                 lock->owner = NULL;
 747                 spin_unlock(&lock->wait_lock);
 748                 return;
 749         }
 750
 751         wakeup_next_waiter(lock);
 752
 753         spin_unlock(&lock->wait_lock);
 754
 755         /* Undo pi boosting if necessary: */
 756         rt_mutex_adjust_prio(current);
 757 }
 758
 759 /*
 760  * debug aware fast / slowpath lock,trylock,unlock
 761  *
 762  * The atomic acquire/release ops are compiled away, when either the
 763  * architecture does not support cmpxchg or when debugging is enabled.
 764  */
 765 static inline int
 766 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 767                   int detect_deadlock,
 768                   int (*slowfn)(struct rt_mutex *lock, int state,
 769                                 struct hrtimer_sleeper *timeout,
 770                                 int detect_deadlock))
 771 {
 772         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 773                 rt_mutex_deadlock_account_lock(lock, current);
 774                 return 0;
 775         } else
 776                 return slowfn(lock, state, NULL, detect_deadlock);
 777 }
 778
 779 static inline int
 780 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 781                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 782                         int (*slowfn)(struct rt_mutex *lock, int state,
 783                                       struct hrtimer_sleeper *timeout,
 784                                       int detect_deadlock))
 785 {
 786         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 787                 rt_mutex_deadlock_account_lock(lock, current);
 788                 return 0;
 789         } else
 790                 return slowfn(lock, state, timeout, detect_deadlock);
 791 }
 792
 793 static inline int
 794 rt_mutex_fasttrylock(struct rt_mutex *lock,
 795                      int (*slowfn)(struct rt_mutex *lock))
 796 {
 797         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 798                 rt_mutex_deadlock_account_lock(lock, current);
 799                 return 1;
 800         }
 801         return slowfn(lock);
 802 }
 803
 804 static inline void
 805 rt_mutex_fastunlock(struct rt_mutex *lock,
 806                     void (*slowfn)(struct rt_mutex *lock))
 807 {
 808         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 809                 rt_mutex_deadlock_account_unlock(current);
 810         else
 811                 slowfn(lock);
 812 }
 813
 814 /**
 815  * rt_mutex_lock - lock a rt_mutex
 816  *
 817  * @lock: the rt_mutex to be locked
 818  */
 819 void __sched rt_mutex_lock(struct rt_mutex *lock)
 820 {
 821         might_sleep();
 822
 823         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 824 }
 825 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 826
 827 /**
 828  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 829  *
 830  * @lock:               the rt_mutex to be locked
 831  * @detect_deadlock:    deadlock detection on/off
 832  *
 833  * Returns:
 834  *  0           on success
 835  * -EINTR       when interrupted by a signal
 836  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 837  */
 838 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 839                                                  int detect_deadlock)
 840 {
 841         might_sleep();
 842
 843         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 844                                  detect_deadlock, rt_mutex_slowlock);
 845 }
 846 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 847
 848 /**
 849  * rt_mutex_lock_interruptible_ktime - lock a rt_mutex interruptible
 850  *                                     the timeout structure is provided
 851  *                                     by the caller
 852  *
 853  * @lock:               the rt_mutex to be locked
 854  * @timeout:            timeout structure or NULL (no timeout)
 855  * @detect_deadlock:    deadlock detection on/off
 856  *
 857  * Returns:
 858  *  0           on success
 859  * -EINTR       when interrupted by a signal
 860  * -ETIMEOUT    when the timeout expired
 861  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 862  */
 863 int
 864 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 865                     int detect_deadlock)
 866 {
 867         might_sleep();
 868
 869         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 870                                        detect_deadlock, rt_mutex_slowlock);
 871 }
 872 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 873
 874 /**
 875  * rt_mutex_trylock - try to lock a rt_mutex
 876  *
 877  * @lock:       the rt_mutex to be locked
 878  *
 879  * Returns 1 on success and 0 on contention
 880  */
 881 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 882 {
 883         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 884 }
 885 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 886
 887 /**
 888  * rt_mutex_unlock - unlock a rt_mutex
 889  *
 890  * @lock: the rt_mutex to be unlocked
 891  */
 892 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 893 {
 894         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 895 }
 896 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 897
 898 /***
 899  * rt_mutex_destroy - mark a mutex unusable
 900  * @lock: the mutex to be destroyed
 901  *
 902  * This function marks the mutex uninitialized, and any subsequent
 903  * use of the mutex is forbidden. The mutex must not be locked when
 904  * this function is called.
 905  */
 906 void rt_mutex_destroy(struct rt_mutex *lock)
 907 {
 908         WARN_ON(rt_mutex_is_locked(lock));
 909 #ifdef CONFIG_DEBUG_RT_MUTEXES
 910         lock->magic = NULL;
 911 #endif
 912 }
 913
 914 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 915
 916 /**
 917  * __rt_mutex_init - initialize the rt lock
 918  *
 919  * @lock: the rt lock to be initialized
 920  *
 921  * Initialize the rt lock to unlocked state.
 922  *
 923  * Initializing of a locked rt lock is not allowed
 924  */
 925 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 926 {
 927         lock->owner = NULL;
 928         spin_lock_init(&lock->wait_lock);
 929         plist_head_init(&lock->wait_list, &lock->wait_lock);
 930
 931         debug_rt_mutex_init(lock, name);
 932 }
 933 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 934
 935 /**
 936  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 937  *                              proxy owner
 938  *
 939  * @lock:       the rt_mutex to be locked
 940  * @proxy_owner:the task to set as owner
 941  *
 942  * No locking. Caller has to do serializing itself
 943  * Special API call for PI-futex support
 944  */
 945 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 946                                 struct task_struct *proxy_owner)
 947 {
 948         __rt_mutex_init(lock, NULL);
 949         debug_rt_mutex_proxy_lock(lock, proxy_owner);
 950         rt_mutex_set_owner(lock, proxy_owner, 0);
 951         rt_mutex_deadlock_account_lock(lock, proxy_owner);
 952 }
 953
 954 /**
 955  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 956  *
 957  * @lock:       the rt_mutex to be locked
 958  *
 959  * No locking. Caller has to do serializing itself
 960  * Special API call for PI-futex support
 961  */
 962 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 963                            struct task_struct *proxy_owner)
 964 {
 965         debug_rt_mutex_proxy_unlock(lock);
 966         rt_mutex_set_owner(lock, NULL, 0);
 967         rt_mutex_deadlock_account_unlock(proxy_owner);
 968 }
 969
 970 /**
 971  * rt_mutex_next_owner - return the next owner of the lock
 972  *
 973  * @lock: the rt lock query
 974  *
 975  * Returns the next owner of the lock or NULL
 976  *
 977  * Caller has to serialize against other accessors to the lock
 978  * itself.
 979  *
 980  * Special API call for PI-futex support
 981  */
 982 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
 983 {
 984         if (!rt_mutex_has_waiters(lock))
 985                 return NULL;
 986
 987         return rt_mutex_top_waiter(lock)->task;
 988 }