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
  85 /*
  86  * Safe fastpath aware unlock:
  87  * 1) Clear the waiters bit
  88  * 2) Drop lock->wait_lock
  89  * 3) Try to unlock the lock with cmpxchg
  90  */
  91 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
  92         __releases(lock->wait_lock)
  93 {
  94         struct task_struct *owner = rt_mutex_owner(lock);
  95
  96         clear_rt_mutex_waiters(lock);
  97         raw_spin_unlock(&lock->wait_lock);
  98         /*
  99          * If a new waiter comes in between the unlock and the cmpxchg
 100          * we have two situations:
 101          *
 102          * unlock(wait_lock);
 103          *                                      lock(wait_lock);
 104          * cmpxchg(p, owner, 0) == owner
 105          *                                      mark_rt_mutex_waiters(lock);
 106          *                                      acquire(lock);
 107          * or:
 108          *
 109          * unlock(wait_lock);
 110          *                                      lock(wait_lock);
 111          *                                      mark_rt_mutex_waiters(lock);
 112          *
 113          * cmpxchg(p, owner, 0) != owner
 114          *                                      enqueue_waiter();
 115          *                                      unlock(wait_lock);
 116          * lock(wait_lock);
 117          * wake waiter();
 118          * unlock(wait_lock);
 119          *                                      lock(wait_lock);
 120          *                                      acquire(lock);
 121          */
 122         return rt_mutex_cmpxchg(lock, owner, NULL);
 123 }
 124
 125 #else
 126 # define rt_mutex_cmpxchg(l,c,n)        (0)
 127 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 128 {
 129         lock->owner = (struct task_struct *)
 130                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 131 }
 132
 133 /*
 134  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 135  */
 136 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
 137         __releases(lock->wait_lock)
 138 {
 139         lock->owner = NULL;
 140         raw_spin_unlock(&lock->wait_lock);
 141         return true;
 142 }
 143 #endif
 144
 145 /*
 146  * Calculate task priority from the waiter list priority
 147  *
 148  * Return task->normal_prio when the waiter list is empty or when
 149  * the waiter is not allowed to do priority boosting
 150  */
 151 int rt_mutex_getprio(struct task_struct *task)
 152 {
 153         if (likely(!task_has_pi_waiters(task)))
 154                 return task->normal_prio;
 155
 156         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
 157                    task->normal_prio);
 158 }
 159
 160 /*
 161  * Adjust the priority of a task, after its pi_waiters got modified.
 162  *
 163  * This can be both boosting and unboosting. task->pi_lock must be held.
 164  */
 165 static void __rt_mutex_adjust_prio(struct task_struct *task)
 166 {
 167         int prio = rt_mutex_getprio(task);
 168
 169         if (task->prio != prio)
 170                 rt_mutex_setprio(task, prio);
 171 }
 172
 173 /*
 174  * Adjust task priority (undo boosting). Called from the exit path of
 175  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 176  *
 177  * (Note: We do this outside of the protection of lock->wait_lock to
 178  * allow the lock to be taken while or before we readjust the priority
 179  * of task. We do not use the spin_xx_mutex() variants here as we are
 180  * outside of the debug path.)
 181  */
 182 static void rt_mutex_adjust_prio(struct task_struct *task)
 183 {
 184         unsigned long flags;
 185
 186         raw_spin_lock_irqsave(&task->pi_lock, flags);
 187         __rt_mutex_adjust_prio(task);
 188         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 189 }
 190
 191 /*
 192  * Max number of times we'll walk the boosting chain:
 193  */
 194 int max_lock_depth = 1024;
 195
 196 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 197 {
 198         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 199 }
 200
 201 /*
 202  * Adjust the priority chain. Also used for deadlock detection.
 203  * Decreases task's usage by one - may thus free the task.
 204  *
 205  * @task:       the task owning the mutex (owner) for which a chain walk is
 206  *              probably needed
 207  * @deadlock_detect: do we have to carry out deadlock detection?
 208  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 209  *              things for a task that has just got its priority adjusted, and
 210  *              is waiting on a mutex)
 211  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 212  *              we dropped its pi_lock. Is never dereferenced, only used for
 213  *              comparison to detect lock chain changes.
 214  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 215  *              its priority to the mutex owner (can be NULL in the case
 216  *              depicted above or if the top waiter is gone away and we are
 217  *              actually deboosting the owner)
 218  * @top_task:   the current top waiter
 219  *
 220  * Returns 0 or -EDEADLK.
 221  */
 222 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 223                                       int deadlock_detect,
 224                                       struct rt_mutex *orig_lock,
 225                                       struct rt_mutex *next_lock,
 226                                       struct rt_mutex_waiter *orig_waiter,
 227                                       struct task_struct *top_task)
 228 {
 229         struct rt_mutex *lock;
 230         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 231         int detect_deadlock, ret = 0, depth = 0;
 232         unsigned long flags;
 233
 234         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 235                                                          deadlock_detect);
 236
 237         /*
 238          * The (de)boosting is a step by step approach with a lot of
 239          * pitfalls. We want this to be preemptible and we want hold a
 240          * maximum of two locks per step. So we have to check
 241          * carefully whether things change under us.
 242          */
 243  again:
 244         if (++depth > max_lock_depth) {
 245                 static int prev_max;
 246
 247                 /*
 248                  * Print this only once. If the admin changes the limit,
 249                  * print a new message when reaching the limit again.
 250                  */
 251                 if (prev_max != max_lock_depth) {
 252                         prev_max = max_lock_depth;
 253                         printk(KERN_WARNING "Maximum lock depth %d reached "
 254                                "task: %s (%d)\n", max_lock_depth,
 255                                top_task->comm, task_pid_nr(top_task));
 256                 }
 257                 put_task_struct(task);
 258
 259                 return -EDEADLK;
 260         }
 261  retry:
 262         /*
 263          * Task can not go away as we did a get_task() before !
 264          */
 265         raw_spin_lock_irqsave(&task->pi_lock, flags);
 266
 267         waiter = task->pi_blocked_on;
 268         /*
 269          * Check whether the end of the boosting chain has been
 270          * reached or the state of the chain has changed while we
 271          * dropped the locks.
 272          */
 273         if (!waiter)
 274                 goto out_unlock_pi;
 275
 276         /*
 277          * Check the orig_waiter state. After we dropped the locks,
 278          * the previous owner of the lock might have released the lock.
 279          */
 280         if (orig_waiter && !rt_mutex_owner(orig_lock))
 281                 goto out_unlock_pi;
 282
 283         /*
 284          * We dropped all locks after taking a refcount on @task, so
 285          * the task might have moved on in the lock chain or even left
 286          * the chain completely and blocks now on an unrelated lock or
 287          * on @orig_lock.
 288          *
 289          * We stored the lock on which @task was blocked in @next_lock,
 290          * so we can detect the chain change.
 291          */
 292         if (next_lock != waiter->lock)
 293                 goto out_unlock_pi;
 294
 295         /*
 296          * Drop out, when the task has no waiters. Note,
 297          * top_waiter can be NULL, when we are in the deboosting
 298          * mode!
 299          */
 300         if (top_waiter) {
 301                 if (!task_has_pi_waiters(task))
 302                         goto out_unlock_pi;
 303                 /*
 304                  * If deadlock detection is off, we stop here if we
 305                  * are not the top pi waiter of the task.
 306                  */
 307                 if (!detect_deadlock && top_waiter != task_top_pi_waiter(task))
 308                         goto out_unlock_pi;
 309         }
 310
 311         /*
 312          * When deadlock detection is off then we check, if further
 313          * priority adjustment is necessary.
 314          */
 315         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 316                 goto out_unlock_pi;
 317
 318         lock = waiter->lock;
 319         if (!raw_spin_trylock(&lock->wait_lock)) {
 320                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 321                 cpu_relax();
 322                 goto retry;
 323         }
 324
 325         /*
 326          * Deadlock detection. If the lock is the same as the original
 327          * lock which caused us to walk the lock chain or if the
 328          * current lock is owned by the task which initiated the chain
 329          * walk, we detected a deadlock.
 330          */
 331         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 332                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 333                 raw_spin_unlock(&lock->wait_lock);
 334                 ret = -EDEADLK;
 335                 goto out_unlock_pi;
 336         }
 337
 338         top_waiter = rt_mutex_top_waiter(lock);
 339
 340         /* Requeue the waiter */
 341         plist_del(&waiter->list_entry, &lock->wait_list);
 342         waiter->list_entry.prio = task->prio;
 343         plist_add(&waiter->list_entry, &lock->wait_list);
 344
 345         /* Release the task */
 346         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 347         if (!rt_mutex_owner(lock)) {
 348                 /*
 349                  * If the requeue above changed the top waiter, then we need
 350                  * to wake the new top waiter up to try to get the lock.
 351                  */
 352
 353                 if (top_waiter != rt_mutex_top_waiter(lock))
 354                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 355                 raw_spin_unlock(&lock->wait_lock);
 356                 goto out_put_task;
 357         }
 358         put_task_struct(task);
 359
 360         /* Grab the next task */
 361         task = rt_mutex_owner(lock);
 362         get_task_struct(task);
 363         raw_spin_lock_irqsave(&task->pi_lock, flags);
 364
 365         if (waiter == rt_mutex_top_waiter(lock)) {
 366                 /* Boost the owner */
 367                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 368                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 369                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 370                 __rt_mutex_adjust_prio(task);
 371
 372         } else if (top_waiter == waiter) {
 373                 /* Deboost the owner */
 374                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 375                 waiter = rt_mutex_top_waiter(lock);
 376                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 377                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 378                 __rt_mutex_adjust_prio(task);
 379         }
 380
 381         /*
 382          * Check whether the task which owns the current lock is pi
 383          * blocked itself. If yes we store a pointer to the lock for
 384          * the lock chain change detection above. After we dropped
 385          * task->pi_lock next_lock cannot be dereferenced anymore.
 386          */
 387         next_lock = task_blocked_on_lock(task);
 388
 389         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 390
 391         top_waiter = rt_mutex_top_waiter(lock);
 392         raw_spin_unlock(&lock->wait_lock);
 393
 394         /*
 395          * We reached the end of the lock chain. Stop right here. No
 396          * point to go back just to figure that out.
 397          */
 398         if (!next_lock)
 399                 goto out_put_task;
 400
 401         if (!detect_deadlock && waiter != top_waiter)
 402                 goto out_put_task;
 403
 404         goto again;
 405
 406  out_unlock_pi:
 407         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 408  out_put_task:
 409         put_task_struct(task);
 410
 411         return ret;
 412 }
 413
 414 /*
 415  * Try to take an rt-mutex
 416  *
 417  * Must be called with lock->wait_lock held.
 418  *
 419  * @lock:   the lock to be acquired.
 420  * @task:   the task which wants to acquire the lock
 421  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 422  */
 423 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 424                 struct rt_mutex_waiter *waiter)
 425 {
 426         /*
 427          * We have to be careful here if the atomic speedups are
 428          * enabled, such that, when
 429          *  - no other waiter is on the lock
 430          *  - the lock has been released since we did the cmpxchg
 431          * the lock can be released or taken while we are doing the
 432          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 433          *
 434          * The atomic acquire/release aware variant of
 435          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 436          * the WAITERS bit, the atomic release / acquire can not
 437          * happen anymore and lock->wait_lock protects us from the
 438          * non-atomic case.
 439          *
 440          * Note, that this might set lock->owner =
 441          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 442          * any more. This is fixed up when we take the ownership.
 443          * This is the transitional state explained at the top of this file.
 444          */
 445         mark_rt_mutex_waiters(lock);
 446
 447         if (rt_mutex_owner(lock))
 448                 return 0;
 449
 450         /*
 451          * It will get the lock because of one of these conditions:
 452          * 1) there is no waiter
 453          * 2) higher priority than waiters
 454          * 3) it is top waiter
 455          */
 456         if (rt_mutex_has_waiters(lock)) {
 457                 if (task->prio >= rt_mutex_top_waiter(lock)->list_entry.prio) {
 458                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 459                                 return 0;
 460                 }
 461         }
 462
 463         if (waiter || rt_mutex_has_waiters(lock)) {
 464                 unsigned long flags;
 465                 struct rt_mutex_waiter *top;
 466
 467                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 468
 469                 /* remove the queued waiter. */
 470                 if (waiter) {
 471                         plist_del(&waiter->list_entry, &lock->wait_list);
 472                         task->pi_blocked_on = NULL;
 473                 }
 474
 475                 /*
 476                  * We have to enqueue the top waiter(if it exists) into
 477                  * task->pi_waiters list.
 478                  */
 479                 if (rt_mutex_has_waiters(lock)) {
 480                         top = rt_mutex_top_waiter(lock);
 481                         top->pi_list_entry.prio = top->list_entry.prio;
 482                         plist_add(&top->pi_list_entry, &task->pi_waiters);
 483                 }
 484                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 485         }
 486
 487         /* We got the lock. */
 488         debug_rt_mutex_lock(lock);
 489
 490         rt_mutex_set_owner(lock, task);
 491
 492         rt_mutex_deadlock_account_lock(lock, task);
 493
 494         return 1;
 495 }
 496
 497 /*
 498  * Task blocks on lock.
 499  *
 500  * Prepare waiter and propagate pi chain
 501  *
 502  * This must be called with lock->wait_lock held.
 503  */
 504 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 505                                    struct rt_mutex_waiter *waiter,
 506                                    struct task_struct *task,
 507                                    int detect_deadlock)
 508 {
 509         struct task_struct *owner = rt_mutex_owner(lock);
 510         struct rt_mutex_waiter *top_waiter = waiter;
 511         struct rt_mutex *next_lock;
 512         int chain_walk = 0, res;
 513         unsigned long flags;
 514
 515         /*
 516          * Early deadlock detection. We really don't want the task to
 517          * enqueue on itself just to untangle the mess later. It's not
 518          * only an optimization. We drop the locks, so another waiter
 519          * can come in before the chain walk detects the deadlock. So
 520          * the other will detect the deadlock and return -EDEADLOCK,
 521          * which is wrong, as the other waiter is not in a deadlock
 522          * situation.
 523          */
 524         if (owner == task)
 525                 return -EDEADLK;
 526
 527         raw_spin_lock_irqsave(&task->pi_lock, flags);
 528         __rt_mutex_adjust_prio(task);
 529         waiter->task = task;
 530         waiter->lock = lock;
 531         plist_node_init(&waiter->list_entry, task->prio);
 532         plist_node_init(&waiter->pi_list_entry, task->prio);
 533
 534         /* Get the top priority waiter on the lock */
 535         if (rt_mutex_has_waiters(lock))
 536                 top_waiter = rt_mutex_top_waiter(lock);
 537         plist_add(&waiter->list_entry, &lock->wait_list);
 538
 539         task->pi_blocked_on = waiter;
 540
 541         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 542
 543         if (!owner)
 544                 return 0;
 545
 546         raw_spin_lock_irqsave(&owner->pi_lock, flags);
 547         if (waiter == rt_mutex_top_waiter(lock)) {
 548                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 549                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 550
 551                 __rt_mutex_adjust_prio(owner);
 552                 if (owner->pi_blocked_on)
 553                         chain_walk = 1;
 554         } else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock)) {
 555                 chain_walk = 1;
 556         }
 557
 558         /* Store the lock on which owner is blocked or NULL */
 559         next_lock = task_blocked_on_lock(owner);
 560
 561         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 562         /*
 563          * Even if full deadlock detection is on, if the owner is not
 564          * blocked itself, we can avoid finding this out in the chain
 565          * walk.
 566          */
 567         if (!chain_walk || !next_lock)
 568                 return 0;
 569
 570         /*
 571          * The owner can't disappear while holding a lock,
 572          * so the owner struct is protected by wait_lock.
 573          * Gets dropped in rt_mutex_adjust_prio_chain()!
 574          */
 575         get_task_struct(owner);
 576
 577         raw_spin_unlock(&lock->wait_lock);
 578
 579         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock,
 580                                          next_lock, waiter, task);
 581
 582         raw_spin_lock(&lock->wait_lock);
 583
 584         return res;
 585 }
 586
 587 /*
 588  * Wake up the next waiter on the lock.
 589  *
 590  * Remove the top waiter from the current tasks pi waiter list and
 591  * wake it up.
 592  *
 593  * Called with lock->wait_lock held.
 594  */
 595 static void wakeup_next_waiter(struct rt_mutex *lock)
 596 {
 597         struct rt_mutex_waiter *waiter;
 598         unsigned long flags;
 599
 600         raw_spin_lock_irqsave(&current->pi_lock, flags);
 601
 602         waiter = rt_mutex_top_waiter(lock);
 603
 604         /*
 605          * Remove it from current->pi_waiters. We do not adjust a
 606          * possible priority boost right now. We execute wakeup in the
 607          * boosted mode and go back to normal after releasing
 608          * lock->wait_lock.
 609          */
 610         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 611
 612         /*
 613          * As we are waking up the top waiter, and the waiter stays
 614          * queued on the lock until it gets the lock, this lock
 615          * obviously has waiters. Just set the bit here and this has
 616          * the added benefit of forcing all new tasks into the
 617          * slow path making sure no task of lower priority than
 618          * the top waiter can steal this lock.
 619          */
 620         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
 621
 622         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 623
 624         /*
 625          * It's safe to dereference waiter as it cannot go away as
 626          * long as we hold lock->wait_lock. The waiter task needs to
 627          * acquire it in order to dequeue the waiter.
 628          */
 629         wake_up_process(waiter->task);
 630 }
 631
 632 /*
 633  * Remove a waiter from a lock and give up
 634  *
 635  * Must be called with lock->wait_lock held and
 636  * have just failed to try_to_take_rt_mutex().
 637  */
 638 static void remove_waiter(struct rt_mutex *lock,
 639                           struct rt_mutex_waiter *waiter)
 640 {
 641         int first = (waiter == rt_mutex_top_waiter(lock));
 642         struct task_struct *owner = rt_mutex_owner(lock);
 643         struct rt_mutex *next_lock = NULL;
 644         unsigned long flags;
 645
 646         raw_spin_lock_irqsave(&current->pi_lock, flags);
 647         plist_del(&waiter->list_entry, &lock->wait_list);
 648         current->pi_blocked_on = NULL;
 649         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 650
 651         if (!owner)
 652                 return;
 653
 654         if (first) {
 655
 656                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 657
 658                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 659
 660                 if (rt_mutex_has_waiters(lock)) {
 661                         struct rt_mutex_waiter *next;
 662
 663                         next = rt_mutex_top_waiter(lock);
 664                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 665                 }
 666                 __rt_mutex_adjust_prio(owner);
 667
 668                 /* Store the lock on which owner is blocked or NULL */
 669                 next_lock = task_blocked_on_lock(owner);
 670
 671                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 672         }
 673
 674         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 675
 676         if (!next_lock)
 677                 return;
 678
 679         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 680         get_task_struct(owner);
 681
 682         raw_spin_unlock(&lock->wait_lock);
 683
 684         rt_mutex_adjust_prio_chain(owner, 0, lock, next_lock, NULL, current);
 685
 686         raw_spin_lock(&lock->wait_lock);
 687 }
 688
 689 /*
 690  * Recheck the pi chain, in case we got a priority setting
 691  *
 692  * Called from sched_setscheduler
 693  */
 694 void rt_mutex_adjust_pi(struct task_struct *task)
 695 {
 696         struct rt_mutex_waiter *waiter;
 697         struct rt_mutex *next_lock;
 698         unsigned long flags;
 699
 700         raw_spin_lock_irqsave(&task->pi_lock, flags);
 701
 702         waiter = task->pi_blocked_on;
 703         if (!waiter || waiter->list_entry.prio == task->prio) {
 704                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 705                 return;
 706         }
 707         next_lock = waiter->lock;
 708         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 709
 710         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 711         get_task_struct(task);
 712
 713         rt_mutex_adjust_prio_chain(task, 0, NULL, next_lock, NULL, task);
 714 }
 715
 716 /**
 717  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 718  * @lock:                the rt_mutex to take
 719  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 720  *                       or TASK_UNINTERRUPTIBLE)
 721  * @timeout:             the pre-initialized and started timer, or NULL for none
 722  * @waiter:              the pre-initialized rt_mutex_waiter
 723  *
 724  * lock->wait_lock must be held by the caller.
 725  */
 726 static int __sched
 727 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 728                     struct hrtimer_sleeper *timeout,
 729                     struct rt_mutex_waiter *waiter)
 730 {
 731         int ret = 0;
 732         int was_disabled;
 733
 734         for (;;) {
 735                 /* Try to acquire the lock: */
 736                 if (try_to_take_rt_mutex(lock, current, waiter))
 737                         break;
 738
 739                 /*
 740                  * TASK_INTERRUPTIBLE checks for signals and
 741                  * timeout. Ignored otherwise.
 742                  */
 743                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 744                         /* Signal pending? */
 745                         if (signal_pending(current))
 746                                 ret = -EINTR;
 747                         if (timeout && !timeout->task)
 748                                 ret = -ETIMEDOUT;
 749                         if (ret)
 750                                 break;
 751                 }
 752
 753                 raw_spin_unlock(&lock->wait_lock);
 754
 755                 was_disabled = irqs_disabled();
 756                 if (was_disabled)
 757                         local_irq_enable();
 758
 759                 debug_rt_mutex_print_deadlock(waiter);
 760
 761                 schedule_rt_mutex(lock);
 762
 763                 if (was_disabled)
 764                         local_irq_disable();
 765
 766                 raw_spin_lock(&lock->wait_lock);
 767                 set_current_state(state);
 768         }
 769
 770         return ret;
 771 }
 772
 773 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
 774                                      struct rt_mutex_waiter *w)
 775 {
 776         /*
 777          * If the result is not -EDEADLOCK or the caller requested
 778          * deadlock detection, nothing to do here.
 779          */
 780         if (res != -EDEADLOCK || detect_deadlock)
 781                 return;
 782
 783         /*
 784          * Yell lowdly and stop the task right here.
 785          */
 786         rt_mutex_print_deadlock(w);
 787         while (1) {
 788                 set_current_state(TASK_INTERRUPTIBLE);
 789                 schedule();
 790         }
 791 }
 792
 793 /*
 794  * Slow path lock function:
 795  */
 796 static int __sched
 797 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 798                   struct hrtimer_sleeper *timeout,
 799                   int detect_deadlock)
 800 {
 801         struct rt_mutex_waiter waiter;
 802         int ret = 0;
 803
 804         debug_rt_mutex_init_waiter(&waiter);
 805
 806         raw_spin_lock(&lock->wait_lock);
 807
 808         /* Try to acquire the lock again: */
 809         if (try_to_take_rt_mutex(lock, current, NULL)) {
 810                 raw_spin_unlock(&lock->wait_lock);
 811                 return 0;
 812         }
 813
 814         set_current_state(state);
 815
 816         /* Setup the timer, when timeout != NULL */
 817         if (unlikely(timeout)) {
 818                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 819                 if (!hrtimer_active(&timeout->timer))
 820                         timeout->task = NULL;
 821         }
 822
 823         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 824
 825         if (likely(!ret))
 826                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
 827
 828         set_current_state(TASK_RUNNING);
 829
 830         if (unlikely(ret)) {
 831                 remove_waiter(lock, &waiter);
 832                 rt_mutex_handle_deadlock(ret, detect_deadlock, &waiter);
 833         }
 834
 835         /*
 836          * try_to_take_rt_mutex() sets the waiter bit
 837          * unconditionally. We might have to fix that up.
 838          */
 839         fixup_rt_mutex_waiters(lock);
 840
 841         raw_spin_unlock(&lock->wait_lock);
 842
 843         /* Remove pending timer: */
 844         if (unlikely(timeout))
 845                 hrtimer_cancel(&timeout->timer);
 846
 847         debug_rt_mutex_free_waiter(&waiter);
 848
 849         return ret;
 850 }
 851
 852 /*
 853  * Slow path try-lock function:
 854  */
 855 static inline int
 856 rt_mutex_slowtrylock(struct rt_mutex *lock)
 857 {
 858         int ret = 0;
 859
 860         raw_spin_lock(&lock->wait_lock);
 861
 862         if (likely(rt_mutex_owner(lock) != current)) {
 863
 864                 ret = try_to_take_rt_mutex(lock, current, NULL);
 865                 /*
 866                  * try_to_take_rt_mutex() sets the lock waiters
 867                  * bit unconditionally. Clean this up.
 868                  */
 869                 fixup_rt_mutex_waiters(lock);
 870         }
 871
 872         raw_spin_unlock(&lock->wait_lock);
 873
 874         return ret;
 875 }
 876
 877 /*
 878  * Slow path to release a rt-mutex:
 879  */
 880 static void __sched
 881 rt_mutex_slowunlock(struct rt_mutex *lock)
 882 {
 883         raw_spin_lock(&lock->wait_lock);
 884
 885         debug_rt_mutex_unlock(lock);
 886
 887         rt_mutex_deadlock_account_unlock(current);
 888
 889         /*
 890          * We must be careful here if the fast path is enabled. If we
 891          * have no waiters queued we cannot set owner to NULL here
 892          * because of:
 893          *
 894          * foo->lock->owner = NULL;
 895          *                      rtmutex_lock(foo->lock);   <- fast path
 896          *                      free = atomic_dec_and_test(foo->refcnt);
 897          *                      rtmutex_unlock(foo->lock); <- fast path
 898          *                      if (free)
 899          *                              kfree(foo);
 900          * raw_spin_unlock(foo->lock->wait_lock);
 901          *
 902          * So for the fastpath enabled kernel:
 903          *
 904          * Nothing can set the waiters bit as long as we hold
 905          * lock->wait_lock. So we do the following sequence:
 906          *
 907          *      owner = rt_mutex_owner(lock);
 908          *      clear_rt_mutex_waiters(lock);
 909          *      raw_spin_unlock(&lock->wait_lock);
 910          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
 911          *              return;
 912          *      goto retry;
 913          *
 914          * The fastpath disabled variant is simple as all access to
 915          * lock->owner is serialized by lock->wait_lock:
 916          *
 917          *      lock->owner = NULL;
 918          *      raw_spin_unlock(&lock->wait_lock);
 919          */
 920         while (!rt_mutex_has_waiters(lock)) {
 921                 /* Drops lock->wait_lock ! */
 922                 if (unlock_rt_mutex_safe(lock) == true)
 923                         return;
 924                 /* Relock the rtmutex and try again */
 925                 raw_spin_lock(&lock->wait_lock);
 926         }
 927
 928         /*
 929          * The wakeup next waiter path does not suffer from the above
 930          * race. See the comments there.
 931          */
 932         wakeup_next_waiter(lock);
 933
 934         raw_spin_unlock(&lock->wait_lock);
 935
 936         /* Undo pi boosting if necessary: */
 937         rt_mutex_adjust_prio(current);
 938 }
 939
 940 /*
 941  * debug aware fast / slowpath lock,trylock,unlock
 942  *
 943  * The atomic acquire/release ops are compiled away, when either the
 944  * architecture does not support cmpxchg or when debugging is enabled.
 945  */
 946 static inline int
 947 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 948                   int detect_deadlock,
 949                   int (*slowfn)(struct rt_mutex *lock, int state,
 950                                 struct hrtimer_sleeper *timeout,
 951                                 int detect_deadlock))
 952 {
 953         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 954                 rt_mutex_deadlock_account_lock(lock, current);
 955                 return 0;
 956         } else
 957                 return slowfn(lock, state, NULL, detect_deadlock);
 958 }
 959
 960 static inline int
 961 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 962                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 963                         int (*slowfn)(struct rt_mutex *lock, int state,
 964                                       struct hrtimer_sleeper *timeout,
 965                                       int detect_deadlock))
 966 {
 967         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 968                 rt_mutex_deadlock_account_lock(lock, current);
 969                 return 0;
 970         } else
 971                 return slowfn(lock, state, timeout, detect_deadlock);
 972 }
 973
 974 static inline int
 975 rt_mutex_fasttrylock(struct rt_mutex *lock,
 976                      int (*slowfn)(struct rt_mutex *lock))
 977 {
 978         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 979                 rt_mutex_deadlock_account_lock(lock, current);
 980                 return 1;
 981         }
 982         return slowfn(lock);
 983 }
 984
 985 static inline void
 986 rt_mutex_fastunlock(struct rt_mutex *lock,
 987                     void (*slowfn)(struct rt_mutex *lock))
 988 {
 989         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 990                 rt_mutex_deadlock_account_unlock(current);
 991         else
 992                 slowfn(lock);
 993 }
 994
 995 /**
 996  * rt_mutex_lock - lock a rt_mutex
 997  *
 998  * @lock: the rt_mutex to be locked
 999  */
1000 void __sched rt_mutex_lock(struct rt_mutex *lock)
1001 {
1002         might_sleep();
1003
1004         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
1005 }
1006 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1007
1008 /**
1009  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1010  *
1011  * @lock:               the rt_mutex to be locked
1012  * @detect_deadlock:    deadlock detection on/off
1013  *
1014  * Returns:
1015  *  0           on success
1016  * -EINTR       when interrupted by a signal
1017  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
1018  */
1019 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
1020                                                  int detect_deadlock)
1021 {
1022         might_sleep();
1023
1024         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
1025                                  detect_deadlock, rt_mutex_slowlock);
1026 }
1027 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1028
1029 /**
1030  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1031  *                      the timeout structure is provided
1032  *                      by the caller
1033  *
1034  * @lock:               the rt_mutex to be locked
1035  * @timeout:            timeout structure or NULL (no timeout)
1036  * @detect_deadlock:    deadlock detection on/off
1037  *
1038  * Returns:
1039  *  0           on success
1040  * -EINTR       when interrupted by a signal
1041  * -ETIMEDOUT   when the timeout expired
1042  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
1043  */
1044 int
1045 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
1046                     int detect_deadlock)
1047 {
1048         might_sleep();
1049
1050         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1051                                        detect_deadlock, rt_mutex_slowlock);
1052 }
1053 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1054
1055 /**
1056  * rt_mutex_trylock - try to lock a rt_mutex
1057  *
1058  * @lock:       the rt_mutex to be locked
1059  *
1060  * Returns 1 on success and 0 on contention
1061  */
1062 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1063 {
1064         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1065 }
1066 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1067
1068 /**
1069  * rt_mutex_unlock - unlock a rt_mutex
1070  *
1071  * @lock: the rt_mutex to be unlocked
1072  */
1073 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1074 {
1075         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1076 }
1077 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1078
1079 /**
1080  * rt_mutex_destroy - mark a mutex unusable
1081  * @lock: the mutex to be destroyed
1082  *
1083  * This function marks the mutex uninitialized, and any subsequent
1084  * use of the mutex is forbidden. The mutex must not be locked when
1085  * this function is called.
1086  */
1087 void rt_mutex_destroy(struct rt_mutex *lock)
1088 {
1089         WARN_ON(rt_mutex_is_locked(lock));
1090 #ifdef CONFIG_DEBUG_RT_MUTEXES
1091         lock->magic = NULL;
1092 #endif
1093 }
1094
1095 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1096
1097 /**
1098  * __rt_mutex_init - initialize the rt lock
1099  *
1100  * @lock: the rt lock to be initialized
1101  *
1102  * Initialize the rt lock to unlocked state.
1103  *
1104  * Initializing of a locked rt lock is not allowed
1105  */
1106 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1107 {
1108         lock->owner = NULL;
1109         raw_spin_lock_init(&lock->wait_lock);
1110         plist_head_init(&lock->wait_list);
1111
1112         debug_rt_mutex_init(lock, name);
1113 }
1114 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1115
1116 /**
1117  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1118  *                              proxy owner
1119  *
1120  * @lock:       the rt_mutex to be locked
1121  * @proxy_owner:the task to set as owner
1122  *
1123  * No locking. Caller has to do serializing itself
1124  * Special API call for PI-futex support
1125  */
1126 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1127                                 struct task_struct *proxy_owner)
1128 {
1129         __rt_mutex_init(lock, NULL);
1130         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1131         rt_mutex_set_owner(lock, proxy_owner);
1132         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1133 }
1134
1135 /**
1136  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1137  *
1138  * @lock:       the rt_mutex to be locked
1139  *
1140  * No locking. Caller has to do serializing itself
1141  * Special API call for PI-futex support
1142  */
1143 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1144                            struct task_struct *proxy_owner)
1145 {
1146         debug_rt_mutex_proxy_unlock(lock);
1147         rt_mutex_set_owner(lock, NULL);
1148         rt_mutex_deadlock_account_unlock(proxy_owner);
1149 }
1150
1151 /**
1152  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1153  * @lock:               the rt_mutex to take
1154  * @waiter:             the pre-initialized rt_mutex_waiter
1155  * @task:               the task to prepare
1156  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1157  *
1158  * Returns:
1159  *  0 - task blocked on lock
1160  *  1 - acquired the lock for task, caller should wake it up
1161  * <0 - error
1162  *
1163  * Special API call for FUTEX_REQUEUE_PI support.
1164  */
1165 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1166                               struct rt_mutex_waiter *waiter,
1167                               struct task_struct *task, int detect_deadlock)
1168 {
1169         int ret;
1170
1171         raw_spin_lock(&lock->wait_lock);
1172
1173         if (try_to_take_rt_mutex(lock, task, NULL)) {
1174                 raw_spin_unlock(&lock->wait_lock);
1175                 return 1;
1176         }
1177
1178         /* We enforce deadlock detection for futexes */
1179         ret = task_blocks_on_rt_mutex(lock, waiter, task, 1);
1180
1181         if (ret && !rt_mutex_owner(lock)) {
1182                 /*
1183                  * Reset the return value. We might have
1184                  * returned with -EDEADLK and the owner
1185                  * released the lock while we were walking the
1186                  * pi chain.  Let the waiter sort it out.
1187                  */
1188                 ret = 0;
1189         }
1190
1191         if (unlikely(ret))
1192                 remove_waiter(lock, waiter);
1193
1194         raw_spin_unlock(&lock->wait_lock);
1195
1196         debug_rt_mutex_print_deadlock(waiter);
1197
1198         return ret;
1199 }
1200
1201 /**
1202  * rt_mutex_next_owner - return the next owner of the lock
1203  *
1204  * @lock: the rt lock query
1205  *
1206  * Returns the next owner of the lock or NULL
1207  *
1208  * Caller has to serialize against other accessors to the lock
1209  * itself.
1210  *
1211  * Special API call for PI-futex support
1212  */
1213 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1214 {
1215         if (!rt_mutex_has_waiters(lock))
1216                 return NULL;
1217
1218         return rt_mutex_top_waiter(lock)->task;
1219 }
1220
1221 /**
1222  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1223  * @lock:               the rt_mutex we were woken on
1224  * @to:                 the timeout, null if none. hrtimer should already have
1225  *                      been started.
1226  * @waiter:             the pre-initialized rt_mutex_waiter
1227  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1228  *
1229  * Complete the lock acquisition started our behalf by another thread.
1230  *
1231  * Returns:
1232  *  0 - success
1233  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1234  *
1235  * Special API call for PI-futex requeue support
1236  */
1237 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1238                                struct hrtimer_sleeper *to,
1239                                struct rt_mutex_waiter *waiter,
1240                                int detect_deadlock)
1241 {
1242         int ret;
1243
1244         raw_spin_lock(&lock->wait_lock);
1245
1246         set_current_state(TASK_INTERRUPTIBLE);
1247
1248         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1249
1250         set_current_state(TASK_RUNNING);
1251
1252         if (unlikely(ret))
1253                 remove_waiter(lock, waiter);
1254
1255         /*
1256          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1257          * have to fix that up.
1258          */
1259         fixup_rt_mutex_waiters(lock);
1260
1261         raw_spin_unlock(&lock->wait_lock);
1262
1263         return ret;
1264 }