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 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 145 {
 146         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 147 }
 148
 149 /*
 150  * Adjust the priority chain. Also used for deadlock detection.
 151  * Decreases task's usage by one - may thus free the task.
 152  *
 153  * @task:       the task owning the mutex (owner) for which a chain walk is
 154  *              probably needed
 155  * @deadlock_detect: do we have to carry out deadlock detection?
 156  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 157  *              things for a task that has just got its priority adjusted, and
 158  *              is waiting on a mutex)
 159  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 160  *              we dropped its pi_lock. Is never dereferenced, only used for
 161  *              comparison to detect lock chain changes.
 162  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 163  *              its priority to the mutex owner (can be NULL in the case
 164  *              depicted above or if the top waiter is gone away and we are
 165  *              actually deboosting the owner)
 166  * @top_task:   the current top waiter
 167  *
 168  * Returns 0 or -EDEADLK.
 169  */
 170 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 171                                       int deadlock_detect,
 172                                       struct rt_mutex *orig_lock,
 173                                       struct rt_mutex *next_lock,
 174                                       struct rt_mutex_waiter *orig_waiter,
 175                                       struct task_struct *top_task)
 176 {
 177         struct rt_mutex *lock;
 178         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 179         int detect_deadlock, ret = 0, depth = 0;
 180         unsigned long flags;
 181
 182         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 183                                                          deadlock_detect);
 184
 185         /*
 186          * The (de)boosting is a step by step approach with a lot of
 187          * pitfalls. We want this to be preemptible and we want hold a
 188          * maximum of two locks per step. So we have to check
 189          * carefully whether things change under us.
 190          */
 191  again:
 192         if (++depth > max_lock_depth) {
 193                 static int prev_max;
 194
 195                 /*
 196                  * Print this only once. If the admin changes the limit,
 197                  * print a new message when reaching the limit again.
 198                  */
 199                 if (prev_max != max_lock_depth) {
 200                         prev_max = max_lock_depth;
 201                         printk(KERN_WARNING "Maximum lock depth %d reached "
 202                                "task: %s (%d)\n", max_lock_depth,
 203                                top_task->comm, task_pid_nr(top_task));
 204                 }
 205                 put_task_struct(task);
 206
 207                 return -EDEADLK;
 208         }
 209  retry:
 210         /*
 211          * Task can not go away as we did a get_task() before !
 212          */
 213         raw_spin_lock_irqsave(&task->pi_lock, flags);
 214
 215         waiter = task->pi_blocked_on;
 216         /*
 217          * Check whether the end of the boosting chain has been
 218          * reached or the state of the chain has changed while we
 219          * dropped the locks.
 220          */
 221         if (!waiter)
 222                 goto out_unlock_pi;
 223
 224         /*
 225          * Check the orig_waiter state. After we dropped the locks,
 226          * the previous owner of the lock might have released the lock.
 227          */
 228         if (orig_waiter && !rt_mutex_owner(orig_lock))
 229                 goto out_unlock_pi;
 230
 231         /*
 232          * We dropped all locks after taking a refcount on @task, so
 233          * the task might have moved on in the lock chain or even left
 234          * the chain completely and blocks now on an unrelated lock or
 235          * on @orig_lock.
 236          *
 237          * We stored the lock on which @task was blocked in @next_lock,
 238          * so we can detect the chain change.
 239          */
 240         if (next_lock != waiter->lock)
 241                 goto out_unlock_pi;
 242
 243         /*
 244          * Drop out, when the task has no waiters. Note,
 245          * top_waiter can be NULL, when we are in the deboosting
 246          * mode!
 247          */
 248         if (top_waiter) {
 249                 if (!task_has_pi_waiters(task))
 250                         goto out_unlock_pi;
 251                 /*
 252                  * If deadlock detection is off, we stop here if we
 253                  * are not the top pi waiter of the task.
 254                  */
 255                 if (!detect_deadlock && top_waiter != task_top_pi_waiter(task))
 256                         goto out_unlock_pi;
 257         }
 258
 259         /*
 260          * When deadlock detection is off then we check, if further
 261          * priority adjustment is necessary.
 262          */
 263         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 264                 goto out_unlock_pi;
 265
 266         lock = waiter->lock;
 267         if (!raw_spin_trylock(&lock->wait_lock)) {
 268                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 269                 cpu_relax();
 270                 goto retry;
 271         }
 272
 273         /*
 274          * Deadlock detection. If the lock is the same as the original
 275          * lock which caused us to walk the lock chain or if the
 276          * current lock is owned by the task which initiated the chain
 277          * walk, we detected a deadlock.
 278          */
 279         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 280                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 281                 raw_spin_unlock(&lock->wait_lock);
 282                 ret = -EDEADLK;
 283                 goto out_unlock_pi;
 284         }
 285
 286         top_waiter = rt_mutex_top_waiter(lock);
 287
 288         /* Requeue the waiter */
 289         plist_del(&waiter->list_entry, &lock->wait_list);
 290         waiter->list_entry.prio = task->prio;
 291         plist_add(&waiter->list_entry, &lock->wait_list);
 292
 293         /* Release the task */
 294         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 295         if (!rt_mutex_owner(lock)) {
 296                 /*
 297                  * If the requeue above changed the top waiter, then we need
 298                  * to wake the new top waiter up to try to get the lock.
 299                  */
 300
 301                 if (top_waiter != rt_mutex_top_waiter(lock))
 302                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 303                 raw_spin_unlock(&lock->wait_lock);
 304                 goto out_put_task;
 305         }
 306         put_task_struct(task);
 307
 308         /* Grab the next task */
 309         task = rt_mutex_owner(lock);
 310         get_task_struct(task);
 311         raw_spin_lock_irqsave(&task->pi_lock, flags);
 312
 313         if (waiter == rt_mutex_top_waiter(lock)) {
 314                 /* Boost the owner */
 315                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 316                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 317                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 318                 __rt_mutex_adjust_prio(task);
 319
 320         } else if (top_waiter == waiter) {
 321                 /* Deboost the owner */
 322                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 323                 waiter = rt_mutex_top_waiter(lock);
 324                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 325                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 326                 __rt_mutex_adjust_prio(task);
 327         }
 328
 329         /*
 330          * Check whether the task which owns the current lock is pi
 331          * blocked itself. If yes we store a pointer to the lock for
 332          * the lock chain change detection above. After we dropped
 333          * task->pi_lock next_lock cannot be dereferenced anymore.
 334          */
 335         next_lock = task_blocked_on_lock(task);
 336
 337         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 338
 339         top_waiter = rt_mutex_top_waiter(lock);
 340         raw_spin_unlock(&lock->wait_lock);
 341
 342         /*
 343          * We reached the end of the lock chain. Stop right here. No
 344          * point to go back just to figure that out.
 345          */
 346         if (!next_lock)
 347                 goto out_put_task;
 348
 349         if (!detect_deadlock && waiter != top_waiter)
 350                 goto out_put_task;
 351
 352         goto again;
 353
 354  out_unlock_pi:
 355         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 356  out_put_task:
 357         put_task_struct(task);
 358
 359         return ret;
 360 }
 361
 362 /*
 363  * Try to take an rt-mutex
 364  *
 365  * Must be called with lock->wait_lock held.
 366  *
 367  * @lock:   the lock to be acquired.
 368  * @task:   the task which wants to acquire the lock
 369  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 370  */
 371 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 372                 struct rt_mutex_waiter *waiter)
 373 {
 374         /*
 375          * We have to be careful here if the atomic speedups are
 376          * enabled, such that, when
 377          *  - no other waiter is on the lock
 378          *  - the lock has been released since we did the cmpxchg
 379          * the lock can be released or taken while we are doing the
 380          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 381          *
 382          * The atomic acquire/release aware variant of
 383          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 384          * the WAITERS bit, the atomic release / acquire can not
 385          * happen anymore and lock->wait_lock protects us from the
 386          * non-atomic case.
 387          *
 388          * Note, that this might set lock->owner =
 389          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 390          * any more. This is fixed up when we take the ownership.
 391          * This is the transitional state explained at the top of this file.
 392          */
 393         mark_rt_mutex_waiters(lock);
 394
 395         if (rt_mutex_owner(lock))
 396                 return 0;
 397
 398         /*
 399          * It will get the lock because of one of these conditions:
 400          * 1) there is no waiter
 401          * 2) higher priority than waiters
 402          * 3) it is top waiter
 403          */
 404         if (rt_mutex_has_waiters(lock)) {
 405                 if (task->prio >= rt_mutex_top_waiter(lock)->list_entry.prio) {
 406                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 407                                 return 0;
 408                 }
 409         }
 410
 411         if (waiter || rt_mutex_has_waiters(lock)) {
 412                 unsigned long flags;
 413                 struct rt_mutex_waiter *top;
 414
 415                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 416
 417                 /* remove the queued waiter. */
 418                 if (waiter) {
 419                         plist_del(&waiter->list_entry, &lock->wait_list);
 420                         task->pi_blocked_on = NULL;
 421                 }
 422
 423                 /*
 424                  * We have to enqueue the top waiter(if it exists) into
 425                  * task->pi_waiters list.
 426                  */
 427                 if (rt_mutex_has_waiters(lock)) {
 428                         top = rt_mutex_top_waiter(lock);
 429                         top->pi_list_entry.prio = top->list_entry.prio;
 430                         plist_add(&top->pi_list_entry, &task->pi_waiters);
 431                 }
 432                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 433         }
 434
 435         /* We got the lock. */
 436         debug_rt_mutex_lock(lock);
 437
 438         rt_mutex_set_owner(lock, task);
 439
 440         rt_mutex_deadlock_account_lock(lock, task);
 441
 442         return 1;
 443 }
 444
 445 /*
 446  * Task blocks on lock.
 447  *
 448  * Prepare waiter and propagate pi chain
 449  *
 450  * This must be called with lock->wait_lock held.
 451  */
 452 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 453                                    struct rt_mutex_waiter *waiter,
 454                                    struct task_struct *task,
 455                                    int detect_deadlock)
 456 {
 457         struct task_struct *owner = rt_mutex_owner(lock);
 458         struct rt_mutex_waiter *top_waiter = waiter;
 459         struct rt_mutex *next_lock;
 460         int chain_walk = 0, res;
 461         unsigned long flags;
 462
 463         /*
 464          * Early deadlock detection. We really don't want the task to
 465          * enqueue on itself just to untangle the mess later. It's not
 466          * only an optimization. We drop the locks, so another waiter
 467          * can come in before the chain walk detects the deadlock. So
 468          * the other will detect the deadlock and return -EDEADLOCK,
 469          * which is wrong, as the other waiter is not in a deadlock
 470          * situation.
 471          */
 472         if (owner == task)
 473                 return -EDEADLK;
 474
 475         raw_spin_lock_irqsave(&task->pi_lock, flags);
 476         __rt_mutex_adjust_prio(task);
 477         waiter->task = task;
 478         waiter->lock = lock;
 479         plist_node_init(&waiter->list_entry, task->prio);
 480         plist_node_init(&waiter->pi_list_entry, task->prio);
 481
 482         /* Get the top priority waiter on the lock */
 483         if (rt_mutex_has_waiters(lock))
 484                 top_waiter = rt_mutex_top_waiter(lock);
 485         plist_add(&waiter->list_entry, &lock->wait_list);
 486
 487         task->pi_blocked_on = waiter;
 488
 489         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 490
 491         if (!owner)
 492                 return 0;
 493
 494         raw_spin_lock_irqsave(&owner->pi_lock, flags);
 495         if (waiter == rt_mutex_top_waiter(lock)) {
 496                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 497                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 498
 499                 __rt_mutex_adjust_prio(owner);
 500                 if (owner->pi_blocked_on)
 501                         chain_walk = 1;
 502         } else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock)) {
 503                 chain_walk = 1;
 504         }
 505
 506         /* Store the lock on which owner is blocked or NULL */
 507         next_lock = task_blocked_on_lock(owner);
 508
 509         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 510         /*
 511          * Even if full deadlock detection is on, if the owner is not
 512          * blocked itself, we can avoid finding this out in the chain
 513          * walk.
 514          */
 515         if (!chain_walk || !next_lock)
 516                 return 0;
 517
 518         /*
 519          * The owner can't disappear while holding a lock,
 520          * so the owner struct is protected by wait_lock.
 521          * Gets dropped in rt_mutex_adjust_prio_chain()!
 522          */
 523         get_task_struct(owner);
 524
 525         raw_spin_unlock(&lock->wait_lock);
 526
 527         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock,
 528                                          next_lock, waiter, task);
 529
 530         raw_spin_lock(&lock->wait_lock);
 531
 532         return res;
 533 }
 534
 535 /*
 536  * Wake up the next waiter on the lock.
 537  *
 538  * Remove the top waiter from the current tasks waiter list and wake it up.
 539  *
 540  * Called with lock->wait_lock held.
 541  */
 542 static void wakeup_next_waiter(struct rt_mutex *lock)
 543 {
 544         struct rt_mutex_waiter *waiter;
 545         unsigned long flags;
 546
 547         raw_spin_lock_irqsave(&current->pi_lock, flags);
 548
 549         waiter = rt_mutex_top_waiter(lock);
 550
 551         /*
 552          * Remove it from current->pi_waiters. We do not adjust a
 553          * possible priority boost right now. We execute wakeup in the
 554          * boosted mode and go back to normal after releasing
 555          * lock->wait_lock.
 556          */
 557         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 558
 559         rt_mutex_set_owner(lock, NULL);
 560
 561         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 562
 563         wake_up_process(waiter->task);
 564 }
 565
 566 /*
 567  * Remove a waiter from a lock and give up
 568  *
 569  * Must be called with lock->wait_lock held and
 570  * have just failed to try_to_take_rt_mutex().
 571  */
 572 static void remove_waiter(struct rt_mutex *lock,
 573                           struct rt_mutex_waiter *waiter)
 574 {
 575         int first = (waiter == rt_mutex_top_waiter(lock));
 576         struct task_struct *owner = rt_mutex_owner(lock);
 577         struct rt_mutex *next_lock = NULL;
 578         unsigned long flags;
 579
 580         raw_spin_lock_irqsave(&current->pi_lock, flags);
 581         plist_del(&waiter->list_entry, &lock->wait_list);
 582         current->pi_blocked_on = NULL;
 583         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 584
 585         if (!owner)
 586                 return;
 587
 588         if (first) {
 589
 590                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 591
 592                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 593
 594                 if (rt_mutex_has_waiters(lock)) {
 595                         struct rt_mutex_waiter *next;
 596
 597                         next = rt_mutex_top_waiter(lock);
 598                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 599                 }
 600                 __rt_mutex_adjust_prio(owner);
 601
 602                 /* Store the lock on which owner is blocked or NULL */
 603                 next_lock = task_blocked_on_lock(owner);
 604
 605                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 606         }
 607
 608         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 609
 610         if (!next_lock)
 611                 return;
 612
 613         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 614         get_task_struct(owner);
 615
 616         raw_spin_unlock(&lock->wait_lock);
 617
 618         rt_mutex_adjust_prio_chain(owner, 0, lock, next_lock, NULL, current);
 619
 620         raw_spin_lock(&lock->wait_lock);
 621 }
 622
 623 /*
 624  * Recheck the pi chain, in case we got a priority setting
 625  *
 626  * Called from sched_setscheduler
 627  */
 628 void rt_mutex_adjust_pi(struct task_struct *task)
 629 {
 630         struct rt_mutex_waiter *waiter;
 631         struct rt_mutex *next_lock;
 632         unsigned long flags;
 633
 634         raw_spin_lock_irqsave(&task->pi_lock, flags);
 635
 636         waiter = task->pi_blocked_on;
 637         if (!waiter || waiter->list_entry.prio == task->prio) {
 638                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 639                 return;
 640         }
 641         next_lock = waiter->lock;
 642         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 643
 644         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 645         get_task_struct(task);
 646
 647         rt_mutex_adjust_prio_chain(task, 0, NULL, next_lock, NULL, task);
 648 }
 649
 650 /**
 651  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 652  * @lock:                the rt_mutex to take
 653  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 654  *                       or TASK_UNINTERRUPTIBLE)
 655  * @timeout:             the pre-initialized and started timer, or NULL for none
 656  * @waiter:              the pre-initialized rt_mutex_waiter
 657  *
 658  * lock->wait_lock must be held by the caller.
 659  */
 660 static int __sched
 661 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 662                     struct hrtimer_sleeper *timeout,
 663                     struct rt_mutex_waiter *waiter)
 664 {
 665         int ret = 0;
 666         int was_disabled;
 667
 668         for (;;) {
 669                 /* Try to acquire the lock: */
 670                 if (try_to_take_rt_mutex(lock, current, waiter))
 671                         break;
 672
 673                 /*
 674                  * TASK_INTERRUPTIBLE checks for signals and
 675                  * timeout. Ignored otherwise.
 676                  */
 677                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 678                         /* Signal pending? */
 679                         if (signal_pending(current))
 680                                 ret = -EINTR;
 681                         if (timeout && !timeout->task)
 682                                 ret = -ETIMEDOUT;
 683                         if (ret)
 684                                 break;
 685                 }
 686
 687                 raw_spin_unlock(&lock->wait_lock);
 688
 689                 was_disabled = irqs_disabled();
 690                 if (was_disabled)
 691                         local_irq_enable();
 692
 693                 debug_rt_mutex_print_deadlock(waiter);
 694
 695                 schedule_rt_mutex(lock);
 696
 697                 if (was_disabled)
 698                         local_irq_disable();
 699
 700                 raw_spin_lock(&lock->wait_lock);
 701                 set_current_state(state);
 702         }
 703
 704         return ret;
 705 }
 706
 707 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
 708                                      struct rt_mutex_waiter *w)
 709 {
 710         /*
 711          * If the result is not -EDEADLOCK or the caller requested
 712          * deadlock detection, nothing to do here.
 713          */
 714         if (res != -EDEADLOCK || detect_deadlock)
 715                 return;
 716
 717         /*
 718          * Yell lowdly and stop the task right here.
 719          */
 720         rt_mutex_print_deadlock(w);
 721         while (1) {
 722                 set_current_state(TASK_INTERRUPTIBLE);
 723                 schedule();
 724         }
 725 }
 726
 727 /*
 728  * Slow path lock function:
 729  */
 730 static int __sched
 731 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 732                   struct hrtimer_sleeper *timeout,
 733                   int detect_deadlock)
 734 {
 735         struct rt_mutex_waiter waiter;
 736         int ret = 0;
 737
 738         debug_rt_mutex_init_waiter(&waiter);
 739
 740         raw_spin_lock(&lock->wait_lock);
 741
 742         /* Try to acquire the lock again: */
 743         if (try_to_take_rt_mutex(lock, current, NULL)) {
 744                 raw_spin_unlock(&lock->wait_lock);
 745                 return 0;
 746         }
 747
 748         set_current_state(state);
 749
 750         /* Setup the timer, when timeout != NULL */
 751         if (unlikely(timeout)) {
 752                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 753                 if (!hrtimer_active(&timeout->timer))
 754                         timeout->task = NULL;
 755         }
 756
 757         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 758
 759         if (likely(!ret))
 760                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
 761
 762         set_current_state(TASK_RUNNING);
 763
 764         if (unlikely(ret)) {
 765                 remove_waiter(lock, &waiter);
 766                 rt_mutex_handle_deadlock(ret, detect_deadlock, &waiter);
 767         }
 768
 769         /*
 770          * try_to_take_rt_mutex() sets the waiter bit
 771          * unconditionally. We might have to fix that up.
 772          */
 773         fixup_rt_mutex_waiters(lock);
 774
 775         raw_spin_unlock(&lock->wait_lock);
 776
 777         /* Remove pending timer: */
 778         if (unlikely(timeout))
 779                 hrtimer_cancel(&timeout->timer);
 780
 781         debug_rt_mutex_free_waiter(&waiter);
 782
 783         return ret;
 784 }
 785
 786 /*
 787  * Slow path try-lock function:
 788  */
 789 static inline int
 790 rt_mutex_slowtrylock(struct rt_mutex *lock)
 791 {
 792         int ret = 0;
 793
 794         raw_spin_lock(&lock->wait_lock);
 795
 796         if (likely(rt_mutex_owner(lock) != current)) {
 797
 798                 ret = try_to_take_rt_mutex(lock, current, NULL);
 799                 /*
 800                  * try_to_take_rt_mutex() sets the lock waiters
 801                  * bit unconditionally. Clean this up.
 802                  */
 803                 fixup_rt_mutex_waiters(lock);
 804         }
 805
 806         raw_spin_unlock(&lock->wait_lock);
 807
 808         return ret;
 809 }
 810
 811 /*
 812  * Slow path to release a rt-mutex:
 813  */
 814 static void __sched
 815 rt_mutex_slowunlock(struct rt_mutex *lock)
 816 {
 817         raw_spin_lock(&lock->wait_lock);
 818
 819         debug_rt_mutex_unlock(lock);
 820
 821         rt_mutex_deadlock_account_unlock(current);
 822
 823         if (!rt_mutex_has_waiters(lock)) {
 824                 lock->owner = NULL;
 825                 raw_spin_unlock(&lock->wait_lock);
 826                 return;
 827         }
 828
 829         wakeup_next_waiter(lock);
 830
 831         raw_spin_unlock(&lock->wait_lock);
 832
 833         /* Undo pi boosting if necessary: */
 834         rt_mutex_adjust_prio(current);
 835 }
 836
 837 /*
 838  * debug aware fast / slowpath lock,trylock,unlock
 839  *
 840  * The atomic acquire/release ops are compiled away, when either the
 841  * architecture does not support cmpxchg or when debugging is enabled.
 842  */
 843 static inline int
 844 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 845                   int detect_deadlock,
 846                   int (*slowfn)(struct rt_mutex *lock, int state,
 847                                 struct hrtimer_sleeper *timeout,
 848                                 int detect_deadlock))
 849 {
 850         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 851                 rt_mutex_deadlock_account_lock(lock, current);
 852                 return 0;
 853         } else
 854                 return slowfn(lock, state, NULL, detect_deadlock);
 855 }
 856
 857 static inline int
 858 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 859                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 860                         int (*slowfn)(struct rt_mutex *lock, int state,
 861                                       struct hrtimer_sleeper *timeout,
 862                                       int detect_deadlock))
 863 {
 864         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 865                 rt_mutex_deadlock_account_lock(lock, current);
 866                 return 0;
 867         } else
 868                 return slowfn(lock, state, timeout, detect_deadlock);
 869 }
 870
 871 static inline int
 872 rt_mutex_fasttrylock(struct rt_mutex *lock,
 873                      int (*slowfn)(struct rt_mutex *lock))
 874 {
 875         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 876                 rt_mutex_deadlock_account_lock(lock, current);
 877                 return 1;
 878         }
 879         return slowfn(lock);
 880 }
 881
 882 static inline void
 883 rt_mutex_fastunlock(struct rt_mutex *lock,
 884                     void (*slowfn)(struct rt_mutex *lock))
 885 {
 886         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 887                 rt_mutex_deadlock_account_unlock(current);
 888         else
 889                 slowfn(lock);
 890 }
 891
 892 /**
 893  * rt_mutex_lock - lock a rt_mutex
 894  *
 895  * @lock: the rt_mutex to be locked
 896  */
 897 void __sched rt_mutex_lock(struct rt_mutex *lock)
 898 {
 899         might_sleep();
 900
 901         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 902 }
 903 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 904
 905 /**
 906  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 907  *
 908  * @lock:               the rt_mutex to be locked
 909  * @detect_deadlock:    deadlock detection on/off
 910  *
 911  * Returns:
 912  *  0           on success
 913  * -EINTR       when interrupted by a signal
 914  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 915  */
 916 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 917                                                  int detect_deadlock)
 918 {
 919         might_sleep();
 920
 921         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 922                                  detect_deadlock, rt_mutex_slowlock);
 923 }
 924 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 925
 926 /**
 927  * rt_mutex_timed_lock - lock a rt_mutex interruptible
 928  *                      the timeout structure is provided
 929  *                      by the caller
 930  *
 931  * @lock:               the rt_mutex to be locked
 932  * @timeout:            timeout structure or NULL (no timeout)
 933  * @detect_deadlock:    deadlock detection on/off
 934  *
 935  * Returns:
 936  *  0           on success
 937  * -EINTR       when interrupted by a signal
 938  * -ETIMEDOUT   when the timeout expired
 939  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 940  */
 941 int
 942 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 943                     int detect_deadlock)
 944 {
 945         might_sleep();
 946
 947         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 948                                        detect_deadlock, rt_mutex_slowlock);
 949 }
 950 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 951
 952 /**
 953  * rt_mutex_trylock - try to lock a rt_mutex
 954  *
 955  * @lock:       the rt_mutex to be locked
 956  *
 957  * Returns 1 on success and 0 on contention
 958  */
 959 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 960 {
 961         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 962 }
 963 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 964
 965 /**
 966  * rt_mutex_unlock - unlock a rt_mutex
 967  *
 968  * @lock: the rt_mutex to be unlocked
 969  */
 970 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 971 {
 972         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 973 }
 974 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 975
 976 /**
 977  * rt_mutex_destroy - mark a mutex unusable
 978  * @lock: the mutex to be destroyed
 979  *
 980  * This function marks the mutex uninitialized, and any subsequent
 981  * use of the mutex is forbidden. The mutex must not be locked when
 982  * this function is called.
 983  */
 984 void rt_mutex_destroy(struct rt_mutex *lock)
 985 {
 986         WARN_ON(rt_mutex_is_locked(lock));
 987 #ifdef CONFIG_DEBUG_RT_MUTEXES
 988         lock->magic = NULL;
 989 #endif
 990 }
 991
 992 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 993
 994 /**
 995  * __rt_mutex_init - initialize the rt lock
 996  *
 997  * @lock: the rt lock to be initialized
 998  *
 999  * Initialize the rt lock to unlocked state.
1000  *
1001  * Initializing of a locked rt lock is not allowed
1002  */
1003 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1004 {
1005         lock->owner = NULL;
1006         raw_spin_lock_init(&lock->wait_lock);
1007         plist_head_init(&lock->wait_list);
1008
1009         debug_rt_mutex_init(lock, name);
1010 }
1011 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1012
1013 /**
1014  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1015  *                              proxy owner
1016  *
1017  * @lock:       the rt_mutex to be locked
1018  * @proxy_owner:the task to set as owner
1019  *
1020  * No locking. Caller has to do serializing itself
1021  * Special API call for PI-futex support
1022  */
1023 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1024                                 struct task_struct *proxy_owner)
1025 {
1026         __rt_mutex_init(lock, NULL);
1027         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1028         rt_mutex_set_owner(lock, proxy_owner);
1029         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1030 }
1031
1032 /**
1033  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1034  *
1035  * @lock:       the rt_mutex to be locked
1036  *
1037  * No locking. Caller has to do serializing itself
1038  * Special API call for PI-futex support
1039  */
1040 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1041                            struct task_struct *proxy_owner)
1042 {
1043         debug_rt_mutex_proxy_unlock(lock);
1044         rt_mutex_set_owner(lock, NULL);
1045         rt_mutex_deadlock_account_unlock(proxy_owner);
1046 }
1047
1048 /**
1049  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1050  * @lock:               the rt_mutex to take
1051  * @waiter:             the pre-initialized rt_mutex_waiter
1052  * @task:               the task to prepare
1053  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1054  *
1055  * Returns:
1056  *  0 - task blocked on lock
1057  *  1 - acquired the lock for task, caller should wake it up
1058  * <0 - error
1059  *
1060  * Special API call for FUTEX_REQUEUE_PI support.
1061  */
1062 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1063                               struct rt_mutex_waiter *waiter,
1064                               struct task_struct *task, int detect_deadlock)
1065 {
1066         int ret;
1067
1068         raw_spin_lock(&lock->wait_lock);
1069
1070         if (try_to_take_rt_mutex(lock, task, NULL)) {
1071                 raw_spin_unlock(&lock->wait_lock);
1072                 return 1;
1073         }
1074
1075         /* We enforce deadlock detection for futexes */
1076         ret = task_blocks_on_rt_mutex(lock, waiter, task, 1);
1077
1078         if (ret && !rt_mutex_owner(lock)) {
1079                 /*
1080                  * Reset the return value. We might have
1081                  * returned with -EDEADLK and the owner
1082                  * released the lock while we were walking the
1083                  * pi chain.  Let the waiter sort it out.
1084                  */
1085                 ret = 0;
1086         }
1087
1088         if (unlikely(ret))
1089                 remove_waiter(lock, waiter);
1090
1091         raw_spin_unlock(&lock->wait_lock);
1092
1093         debug_rt_mutex_print_deadlock(waiter);
1094
1095         return ret;
1096 }
1097
1098 /**
1099  * rt_mutex_next_owner - return the next owner of the lock
1100  *
1101  * @lock: the rt lock query
1102  *
1103  * Returns the next owner of the lock or NULL
1104  *
1105  * Caller has to serialize against other accessors to the lock
1106  * itself.
1107  *
1108  * Special API call for PI-futex support
1109  */
1110 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1111 {
1112         if (!rt_mutex_has_waiters(lock))
1113                 return NULL;
1114
1115         return rt_mutex_top_waiter(lock)->task;
1116 }
1117
1118 /**
1119  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1120  * @lock:               the rt_mutex we were woken on
1121  * @to:                 the timeout, null if none. hrtimer should already have
1122  *                      been started.
1123  * @waiter:             the pre-initialized rt_mutex_waiter
1124  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1125  *
1126  * Complete the lock acquisition started our behalf by another thread.
1127  *
1128  * Returns:
1129  *  0 - success
1130  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1131  *
1132  * Special API call for PI-futex requeue support
1133  */
1134 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1135                                struct hrtimer_sleeper *to,
1136                                struct rt_mutex_waiter *waiter,
1137                                int detect_deadlock)
1138 {
1139         int ret;
1140
1141         raw_spin_lock(&lock->wait_lock);
1142
1143         set_current_state(TASK_INTERRUPTIBLE);
1144
1145         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1146
1147         set_current_state(TASK_RUNNING);
1148
1149         if (unlikely(ret))
1150                 remove_waiter(lock, waiter);
1151
1152         /*
1153          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1154          * have to fix that up.
1155          */
1156         fixup_rt_mutex_waiters(lock);
1157
1158         raw_spin_unlock(&lock->wait_lock);
1159
1160         return ret;
1161 }