kernel/locking/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/sched/rt.h>
  17 #include <linux/sched/deadline.h>
  18 #include <linux/timer.h>
  19
  20 #include "rtmutex_common.h"
  21
  22 /*
  23  * lock->owner state tracking:
  24  *
  25  * lock->owner holds the task_struct pointer of the owner. Bit 0
  26  * is used to keep track of the "lock has waiters" state.
  27  *
  28  * owner        bit0
  29  * NULL         0       lock is free (fast acquire possible)
  30  * NULL         1       lock is free and has waiters and the top waiter
  31  *                              is going to take the lock*
  32  * taskpointer  0       lock is held (fast release possible)
  33  * taskpointer  1       lock is held and has waiters**
  34  *
  35  * The fast atomic compare exchange based acquire and release is only
  36  * possible when bit 0 of lock->owner is 0.
  37  *
  38  * (*) It also can be a transitional state when grabbing the lock
  39  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  40  * we need to set the bit0 before looking at the lock, and the owner may be
  41  * NULL in this small time, hence this can be a transitional state.
  42  *
  43  * (**) There is a small time when bit 0 is set but there are no
  44  * waiters. This can happen when grabbing the lock in the slow path.
  45  * To prevent a cmpxchg of the owner releasing the lock, we need to
  46  * set this bit before looking at the lock.
  47  */
  48
  49 static void
  50 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
  51 {
  52         unsigned long val = (unsigned long)owner;
  53
  54         if (rt_mutex_has_waiters(lock))
  55                 val |= RT_MUTEX_HAS_WAITERS;
  56
  57         lock->owner = (struct task_struct *)val;
  58 }
  59
  60 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  61 {
  62         lock->owner = (struct task_struct *)
  63                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  64 }
  65
  66 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  67 {
  68         if (!rt_mutex_has_waiters(lock))
  69                 clear_rt_mutex_waiters(lock);
  70 }
  71
  72 /*
  73  * We can speed up the acquire/release, if the architecture
  74  * supports cmpxchg and if there's no debugging state to be set up
  75  */
  76 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
  77 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
  78 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  79 {
  80         unsigned long owner, *p = (unsigned long *) &lock->owner;
  81
  82         do {
  83                 owner = *p;
  84         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
  85 }
  86 #else
  87 # define rt_mutex_cmpxchg(l,c,n)        (0)
  88 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  89 {
  90         lock->owner = (struct task_struct *)
  91                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
  92 }
  93 #endif
  94
  95 static inline int
  96 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
  97                      struct rt_mutex_waiter *right)
  98 {
  99         if (left->prio < right->prio)
 100                 return 1;
 101
 102         /*
 103          * If both waiters have dl_prio(), we check the deadlines of the
 104          * associated tasks.
 105          * If left waiter has a dl_prio(), and we didn't return 1 above,
 106          * then right waiter has a dl_prio() too.
 107          */
 108         if (dl_prio(left->prio))
 109                 return (left->task->dl.deadline < right->task->dl.deadline);
 110
 111         return 0;
 112 }
 113
 114 static void
 115 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 116 {
 117         struct rb_node **link = &lock->waiters.rb_node;
 118         struct rb_node *parent = NULL;
 119         struct rt_mutex_waiter *entry;
 120         int leftmost = 1;
 121
 122         while (*link) {
 123                 parent = *link;
 124                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
 125                 if (rt_mutex_waiter_less(waiter, entry)) {
 126                         link = &parent->rb_left;
 127                 } else {
 128                         link = &parent->rb_right;
 129                         leftmost = 0;
 130                 }
 131         }
 132
 133         if (leftmost)
 134                 lock->waiters_leftmost = &waiter->tree_entry;
 135
 136         rb_link_node(&waiter->tree_entry, parent, link);
 137         rb_insert_color(&waiter->tree_entry, &lock->waiters);
 138 }
 139
 140 static void
 141 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 142 {
 143         if (RB_EMPTY_NODE(&waiter->tree_entry))
 144                 return;
 145
 146         if (lock->waiters_leftmost == &waiter->tree_entry)
 147                 lock->waiters_leftmost = rb_next(&waiter->tree_entry);
 148
 149         rb_erase(&waiter->tree_entry, &lock->waiters);
 150         RB_CLEAR_NODE(&waiter->tree_entry);
 151 }
 152
 153 static void
 154 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 155 {
 156         struct rb_node **link = &task->pi_waiters.rb_node;
 157         struct rb_node *parent = NULL;
 158         struct rt_mutex_waiter *entry;
 159         int leftmost = 1;
 160
 161         while (*link) {
 162                 parent = *link;
 163                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
 164                 if (rt_mutex_waiter_less(waiter, entry)) {
 165                         link = &parent->rb_left;
 166                 } else {
 167                         link = &parent->rb_right;
 168                         leftmost = 0;
 169                 }
 170         }
 171
 172         if (leftmost)
 173                 task->pi_waiters_leftmost = &waiter->pi_tree_entry;
 174
 175         rb_link_node(&waiter->pi_tree_entry, parent, link);
 176         rb_insert_color(&waiter->pi_tree_entry, &task->pi_waiters);
 177 }
 178
 179 static void
 180 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 181 {
 182         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 183                 return;
 184
 185         if (task->pi_waiters_leftmost == &waiter->pi_tree_entry)
 186                 task->pi_waiters_leftmost = rb_next(&waiter->pi_tree_entry);
 187
 188         rb_erase(&waiter->pi_tree_entry, &task->pi_waiters);
 189         RB_CLEAR_NODE(&waiter->pi_tree_entry);
 190 }
 191
 192 /*
 193  * Calculate task priority from the waiter tree priority
 194  *
 195  * Return task->normal_prio when the waiter tree is empty or when
 196  * the waiter is not allowed to do priority boosting
 197  */
 198 int rt_mutex_getprio(struct task_struct *task)
 199 {
 200         if (likely(!task_has_pi_waiters(task)))
 201                 return task->normal_prio;
 202
 203         return min(task_top_pi_waiter(task)->prio,
 204                    task->normal_prio);
 205 }
 206
 207 struct task_struct *rt_mutex_get_top_task(struct task_struct *task)
 208 {
 209         if (likely(!task_has_pi_waiters(task)))
 210                 return NULL;
 211
 212         return task_top_pi_waiter(task)->task;
 213 }
 214
 215 /*
 216  * Called by sched_setscheduler() to check whether the priority change
 217  * is overruled by a possible priority boosting.
 218  */
 219 int rt_mutex_check_prio(struct task_struct *task, int newprio)
 220 {
 221         if (!task_has_pi_waiters(task))
 222                 return 0;
 223
 224         return task_top_pi_waiter(task)->task->prio <= newprio;
 225 }
 226
 227 /*
 228  * Adjust the priority of a task, after its pi_waiters got modified.
 229  *
 230  * This can be both boosting and unboosting. task->pi_lock must be held.
 231  */
 232 static void __rt_mutex_adjust_prio(struct task_struct *task)
 233 {
 234         int prio = rt_mutex_getprio(task);
 235
 236         if (task->prio != prio || dl_prio(prio))
 237                 rt_mutex_setprio(task, prio);
 238 }
 239
 240 /*
 241  * Adjust task priority (undo boosting). Called from the exit path of
 242  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 243  *
 244  * (Note: We do this outside of the protection of lock->wait_lock to
 245  * allow the lock to be taken while or before we readjust the priority
 246  * of task. We do not use the spin_xx_mutex() variants here as we are
 247  * outside of the debug path.)
 248  */
 249 static void rt_mutex_adjust_prio(struct task_struct *task)
 250 {
 251         unsigned long flags;
 252
 253         raw_spin_lock_irqsave(&task->pi_lock, flags);
 254         __rt_mutex_adjust_prio(task);
 255         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 256 }
 257
 258 /*
 259  * Max number of times we'll walk the boosting chain:
 260  */
 261 int max_lock_depth = 1024;
 262
 263 /*
 264  * Adjust the priority chain. Also used for deadlock detection.
 265  * Decreases task's usage by one - may thus free the task.
 266  *
 267  * @task: the task owning the mutex (owner) for which a chain walk is probably
 268  *        needed
 269  * @deadlock_detect: do we have to carry out deadlock detection?
 270  * @orig_lock: the mutex (can be NULL if we are walking the chain to recheck
 271  *             things for a task that has just got its priority adjusted, and
 272  *             is waiting on a mutex)
 273  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 274  *               its priority to the mutex owner (can be NULL in the case
 275  *               depicted above or if the top waiter is gone away and we are
 276  *               actually deboosting the owner)
 277  * @top_task: the current top waiter
 278  *
 279  * Returns 0 or -EDEADLK.
 280  */
 281 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 282                                       int deadlock_detect,
 283                                       struct rt_mutex *orig_lock,
 284                                       struct rt_mutex_waiter *orig_waiter,
 285                                       struct task_struct *top_task)
 286 {
 287         struct rt_mutex *lock;
 288         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 289         int detect_deadlock, ret = 0, depth = 0;
 290         unsigned long flags;
 291
 292         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 293                                                          deadlock_detect);
 294
 295         /*
 296          * The (de)boosting is a step by step approach with a lot of
 297          * pitfalls. We want this to be preemptible and we want hold a
 298          * maximum of two locks per step. So we have to check
 299          * carefully whether things change under us.
 300          */
 301  again:
 302         if (++depth > max_lock_depth) {
 303                 static int prev_max;
 304
 305                 /*
 306                  * Print this only once. If the admin changes the limit,
 307                  * print a new message when reaching the limit again.
 308                  */
 309                 if (prev_max != max_lock_depth) {
 310                         prev_max = max_lock_depth;
 311                         printk(KERN_WARNING "Maximum lock depth %d reached "
 312                                "task: %s (%d)\n", max_lock_depth,
 313                                top_task->comm, task_pid_nr(top_task));
 314                 }
 315                 put_task_struct(task);
 316
 317                 return deadlock_detect ? -EDEADLK : 0;
 318         }
 319  retry:
 320         /*
 321          * Task can not go away as we did a get_task() before !
 322          */
 323         raw_spin_lock_irqsave(&task->pi_lock, flags);
 324
 325         waiter = task->pi_blocked_on;
 326         /*
 327          * Check whether the end of the boosting chain has been
 328          * reached or the state of the chain has changed while we
 329          * dropped the locks.
 330          */
 331         if (!waiter)
 332                 goto out_unlock_pi;
 333
 334         /*
 335          * Check the orig_waiter state. After we dropped the locks,
 336          * the previous owner of the lock might have released the lock.
 337          */
 338         if (orig_waiter && !rt_mutex_owner(orig_lock))
 339                 goto out_unlock_pi;
 340
 341         /*
 342          * Drop out, when the task has no waiters. Note,
 343          * top_waiter can be NULL, when we are in the deboosting
 344          * mode!
 345          */
 346         if (top_waiter) {
 347                 if (!task_has_pi_waiters(task))
 348                         goto out_unlock_pi;
 349                 /*
 350                  * If deadlock detection is off, we stop here if we
 351                  * are not the top pi waiter of the task.
 352                  */
 353                 if (!detect_deadlock && top_waiter != task_top_pi_waiter(task))
 354                         goto out_unlock_pi;
 355         }
 356
 357         /*
 358          * When deadlock detection is off then we check, if further
 359          * priority adjustment is necessary.
 360          */
 361         if (!detect_deadlock && waiter->prio == task->prio)
 362                 goto out_unlock_pi;
 363
 364         lock = waiter->lock;
 365         if (!raw_spin_trylock(&lock->wait_lock)) {
 366                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 367                 cpu_relax();
 368                 goto retry;
 369         }
 370
 371         /*
 372          * Deadlock detection. If the lock is the same as the original
 373          * lock which caused us to walk the lock chain or if the
 374          * current lock is owned by the task which initiated the chain
 375          * walk, we detected a deadlock.
 376          */
 377         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 378                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 379                 raw_spin_unlock(&lock->wait_lock);
 380                 ret = deadlock_detect ? -EDEADLK : 0;
 381                 goto out_unlock_pi;
 382         }
 383
 384         top_waiter = rt_mutex_top_waiter(lock);
 385
 386         /* Requeue the waiter */
 387         rt_mutex_dequeue(lock, waiter);
 388         waiter->prio = task->prio;
 389         rt_mutex_enqueue(lock, waiter);
 390
 391         /* Release the task */
 392         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 393         if (!rt_mutex_owner(lock)) {
 394                 /*
 395                  * If the requeue above changed the top waiter, then we need
 396                  * to wake the new top waiter up to try to get the lock.
 397                  */
 398
 399                 if (top_waiter != rt_mutex_top_waiter(lock))
 400                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 401                 raw_spin_unlock(&lock->wait_lock);
 402                 goto out_put_task;
 403         }
 404         put_task_struct(task);
 405
 406         /* Grab the next task */
 407         task = rt_mutex_owner(lock);
 408         get_task_struct(task);
 409         raw_spin_lock_irqsave(&task->pi_lock, flags);
 410
 411         if (waiter == rt_mutex_top_waiter(lock)) {
 412                 /* Boost the owner */
 413                 rt_mutex_dequeue_pi(task, top_waiter);
 414                 rt_mutex_enqueue_pi(task, waiter);
 415                 __rt_mutex_adjust_prio(task);
 416
 417         } else if (top_waiter == waiter) {
 418                 /* Deboost the owner */
 419                 rt_mutex_dequeue_pi(task, waiter);
 420                 waiter = rt_mutex_top_waiter(lock);
 421                 rt_mutex_enqueue_pi(task, waiter);
 422                 __rt_mutex_adjust_prio(task);
 423         }
 424
 425         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 426
 427         top_waiter = rt_mutex_top_waiter(lock);
 428         raw_spin_unlock(&lock->wait_lock);
 429
 430         if (!detect_deadlock && waiter != top_waiter)
 431                 goto out_put_task;
 432
 433         goto again;
 434
 435  out_unlock_pi:
 436         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 437  out_put_task:
 438         put_task_struct(task);
 439
 440         return ret;
 441 }
 442
 443 /*
 444  * Try to take an rt-mutex
 445  *
 446  * Must be called with lock->wait_lock held.
 447  *
 448  * @lock:   the lock to be acquired.
 449  * @task:   the task which wants to acquire the lock
 450  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 451  */
 452 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 453                 struct rt_mutex_waiter *waiter)
 454 {
 455         /*
 456          * We have to be careful here if the atomic speedups are
 457          * enabled, such that, when
 458          *  - no other waiter is on the lock
 459          *  - the lock has been released since we did the cmpxchg
 460          * the lock can be released or taken while we are doing the
 461          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 462          *
 463          * The atomic acquire/release aware variant of
 464          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 465          * the WAITERS bit, the atomic release / acquire can not
 466          * happen anymore and lock->wait_lock protects us from the
 467          * non-atomic case.
 468          *
 469          * Note, that this might set lock->owner =
 470          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 471          * any more. This is fixed up when we take the ownership.
 472          * This is the transitional state explained at the top of this file.
 473          */
 474         mark_rt_mutex_waiters(lock);
 475
 476         if (rt_mutex_owner(lock))
 477                 return 0;
 478
 479         /*
 480          * It will get the lock because of one of these conditions:
 481          * 1) there is no waiter
 482          * 2) higher priority than waiters
 483          * 3) it is top waiter
 484          */
 485         if (rt_mutex_has_waiters(lock)) {
 486                 if (task->prio >= rt_mutex_top_waiter(lock)->prio) {
 487                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 488                                 return 0;
 489                 }
 490         }
 491
 492         if (waiter || rt_mutex_has_waiters(lock)) {
 493                 unsigned long flags;
 494                 struct rt_mutex_waiter *top;
 495
 496                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 497
 498                 /* remove the queued waiter. */
 499                 if (waiter) {
 500                         rt_mutex_dequeue(lock, waiter);
 501                         task->pi_blocked_on = NULL;
 502                 }
 503
 504                 /*
 505                  * We have to enqueue the top waiter(if it exists) into
 506                  * task->pi_waiters list.
 507                  */
 508                 if (rt_mutex_has_waiters(lock)) {
 509                         top = rt_mutex_top_waiter(lock);
 510                         rt_mutex_enqueue_pi(task, top);
 511                 }
 512                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 513         }
 514
 515         /* We got the lock. */
 516         debug_rt_mutex_lock(lock);
 517
 518         rt_mutex_set_owner(lock, task);
 519
 520         rt_mutex_deadlock_account_lock(lock, task);
 521
 522         return 1;
 523 }
 524
 525 /*
 526  * Task blocks on lock.
 527  *
 528  * Prepare waiter and propagate pi chain
 529  *
 530  * This must be called with lock->wait_lock held.
 531  */
 532 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 533                                    struct rt_mutex_waiter *waiter,
 534                                    struct task_struct *task,
 535                                    int detect_deadlock)
 536 {
 537         struct task_struct *owner = rt_mutex_owner(lock);
 538         struct rt_mutex_waiter *top_waiter = waiter;
 539         unsigned long flags;
 540         int chain_walk = 0, res;
 541
 542         /*
 543          * Early deadlock detection. We really don't want the task to
 544          * enqueue on itself just to untangle the mess later. It's not
 545          * only an optimization. We drop the locks, so another waiter
 546          * can come in before the chain walk detects the deadlock. So
 547          * the other will detect the deadlock and return -EDEADLOCK,
 548          * which is wrong, as the other waiter is not in a deadlock
 549          * situation.
 550          */
 551         if (detect_deadlock && owner == task)
 552                 return -EDEADLK;
 553
 554         raw_spin_lock_irqsave(&task->pi_lock, flags);
 555         __rt_mutex_adjust_prio(task);
 556         waiter->task = task;
 557         waiter->lock = lock;
 558         waiter->prio = task->prio;
 559
 560         /* Get the top priority waiter on the lock */
 561         if (rt_mutex_has_waiters(lock))
 562                 top_waiter = rt_mutex_top_waiter(lock);
 563         rt_mutex_enqueue(lock, waiter);
 564
 565         task->pi_blocked_on = waiter;
 566
 567         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 568
 569         if (!owner)
 570                 return 0;
 571
 572         if (waiter == rt_mutex_top_waiter(lock)) {
 573                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 574                 rt_mutex_dequeue_pi(owner, top_waiter);
 575                 rt_mutex_enqueue_pi(owner, waiter);
 576
 577                 __rt_mutex_adjust_prio(owner);
 578                 if (owner->pi_blocked_on)
 579                         chain_walk = 1;
 580                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 581         }
 582         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 583                 chain_walk = 1;
 584
 585         if (!chain_walk)
 586                 return 0;
 587
 588         /*
 589          * The owner can't disappear while holding a lock,
 590          * so the owner struct is protected by wait_lock.
 591          * Gets dropped in rt_mutex_adjust_prio_chain()!
 592          */
 593         get_task_struct(owner);
 594
 595         raw_spin_unlock(&lock->wait_lock);
 596
 597         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 598                                          task);
 599
 600         raw_spin_lock(&lock->wait_lock);
 601
 602         return res;
 603 }
 604
 605 /*
 606  * Wake up the next waiter on the lock.
 607  *
 608  * Remove the top waiter from the current tasks waiter list and wake it up.
 609  *
 610  * Called with lock->wait_lock held.
 611  */
 612 static void wakeup_next_waiter(struct rt_mutex *lock)
 613 {
 614         struct rt_mutex_waiter *waiter;
 615         unsigned long flags;
 616
 617         raw_spin_lock_irqsave(&current->pi_lock, flags);
 618
 619         waiter = rt_mutex_top_waiter(lock);
 620
 621         /*
 622          * Remove it from current->pi_waiters. We do not adjust a
 623          * possible priority boost right now. We execute wakeup in the
 624          * boosted mode and go back to normal after releasing
 625          * lock->wait_lock.
 626          */
 627         rt_mutex_dequeue_pi(current, waiter);
 628
 629         rt_mutex_set_owner(lock, NULL);
 630
 631         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 632
 633         wake_up_process(waiter->task);
 634 }
 635
 636 /*
 637  * Remove a waiter from a lock and give up
 638  *
 639  * Must be called with lock->wait_lock held and
 640  * have just failed to try_to_take_rt_mutex().
 641  */
 642 static void remove_waiter(struct rt_mutex *lock,
 643                           struct rt_mutex_waiter *waiter)
 644 {
 645         int first = (waiter == rt_mutex_top_waiter(lock));
 646         struct task_struct *owner = rt_mutex_owner(lock);
 647         unsigned long flags;
 648         int chain_walk = 0;
 649
 650         raw_spin_lock_irqsave(&current->pi_lock, flags);
 651         rt_mutex_dequeue(lock, waiter);
 652         current->pi_blocked_on = NULL;
 653         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 654
 655         if (!owner)
 656                 return;
 657
 658         if (first) {
 659
 660                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 661
 662                 rt_mutex_dequeue_pi(owner, waiter);
 663
 664                 if (rt_mutex_has_waiters(lock)) {
 665                         struct rt_mutex_waiter *next;
 666
 667                         next = rt_mutex_top_waiter(lock);
 668                         rt_mutex_enqueue_pi(owner, next);
 669                 }
 670                 __rt_mutex_adjust_prio(owner);
 671
 672                 if (owner->pi_blocked_on)
 673                         chain_walk = 1;
 674
 675                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 676         }
 677
 678         if (!chain_walk)
 679                 return;
 680
 681         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 682         get_task_struct(owner);
 683
 684         raw_spin_unlock(&lock->wait_lock);
 685
 686         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 687
 688         raw_spin_lock(&lock->wait_lock);
 689 }
 690
 691 /*
 692  * Recheck the pi chain, in case we got a priority setting
 693  *
 694  * Called from sched_setscheduler
 695  */
 696 void rt_mutex_adjust_pi(struct task_struct *task)
 697 {
 698         struct rt_mutex_waiter *waiter;
 699         unsigned long flags;
 700
 701         raw_spin_lock_irqsave(&task->pi_lock, flags);
 702
 703         waiter = task->pi_blocked_on;
 704         if (!waiter || (waiter->prio == task->prio &&
 705                         !dl_prio(task->prio))) {
 706                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 707                 return;
 708         }
 709
 710         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 711
 712         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 713         get_task_struct(task);
 714         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 715 }
 716
 717 /**
 718  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 719  * @lock:                the rt_mutex to take
 720  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 721  *                       or TASK_UNINTERRUPTIBLE)
 722  * @timeout:             the pre-initialized and started timer, or NULL for none
 723  * @waiter:              the pre-initialized rt_mutex_waiter
 724  *
 725  * lock->wait_lock must be held by the caller.
 726  */
 727 static int __sched
 728 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 729                     struct hrtimer_sleeper *timeout,
 730                     struct rt_mutex_waiter *waiter)
 731 {
 732         int ret = 0;
 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                 debug_rt_mutex_print_deadlock(waiter);
 756
 757                 schedule_rt_mutex(lock);
 758
 759                 raw_spin_lock(&lock->wait_lock);
 760                 set_current_state(state);
 761         }
 762
 763         return ret;
 764 }
 765
 766 /*
 767  * Slow path lock function:
 768  */
 769 static int __sched
 770 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 771                   struct hrtimer_sleeper *timeout,
 772                   int detect_deadlock)
 773 {
 774         struct rt_mutex_waiter waiter;
 775         int ret = 0;
 776
 777         debug_rt_mutex_init_waiter(&waiter);
 778         RB_CLEAR_NODE(&waiter.pi_tree_entry);
 779         RB_CLEAR_NODE(&waiter.tree_entry);
 780
 781         raw_spin_lock(&lock->wait_lock);
 782
 783         /* Try to acquire the lock again: */
 784         if (try_to_take_rt_mutex(lock, current, NULL)) {
 785                 raw_spin_unlock(&lock->wait_lock);
 786                 return 0;
 787         }
 788
 789         set_current_state(state);
 790
 791         /* Setup the timer, when timeout != NULL */
 792         if (unlikely(timeout)) {
 793                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 794                 if (!hrtimer_active(&timeout->timer))
 795                         timeout->task = NULL;
 796         }
 797
 798         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 799
 800         if (likely(!ret))
 801                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
 802
 803         set_current_state(TASK_RUNNING);
 804
 805         if (unlikely(ret))
 806                 remove_waiter(lock, &waiter);
 807
 808         /*
 809          * try_to_take_rt_mutex() sets the waiter bit
 810          * unconditionally. We might have to fix that up.
 811          */
 812         fixup_rt_mutex_waiters(lock);
 813
 814         raw_spin_unlock(&lock->wait_lock);
 815
 816         /* Remove pending timer: */
 817         if (unlikely(timeout))
 818                 hrtimer_cancel(&timeout->timer);
 819
 820         debug_rt_mutex_free_waiter(&waiter);
 821
 822         return ret;
 823 }
 824
 825 /*
 826  * Slow path try-lock function:
 827  */
 828 static inline int
 829 rt_mutex_slowtrylock(struct rt_mutex *lock)
 830 {
 831         int ret = 0;
 832
 833         raw_spin_lock(&lock->wait_lock);
 834
 835         if (likely(rt_mutex_owner(lock) != current)) {
 836
 837                 ret = try_to_take_rt_mutex(lock, current, NULL);
 838                 /*
 839                  * try_to_take_rt_mutex() sets the lock waiters
 840                  * bit unconditionally. Clean this up.
 841                  */
 842                 fixup_rt_mutex_waiters(lock);
 843         }
 844
 845         raw_spin_unlock(&lock->wait_lock);
 846
 847         return ret;
 848 }
 849
 850 /*
 851  * Slow path to release a rt-mutex:
 852  */
 853 static void __sched
 854 rt_mutex_slowunlock(struct rt_mutex *lock)
 855 {
 856         raw_spin_lock(&lock->wait_lock);
 857
 858         debug_rt_mutex_unlock(lock);
 859
 860         rt_mutex_deadlock_account_unlock(current);
 861
 862         if (!rt_mutex_has_waiters(lock)) {
 863                 lock->owner = NULL;
 864                 raw_spin_unlock(&lock->wait_lock);
 865                 return;
 866         }
 867
 868         wakeup_next_waiter(lock);
 869
 870         raw_spin_unlock(&lock->wait_lock);
 871
 872         /* Undo pi boosting if necessary: */
 873         rt_mutex_adjust_prio(current);
 874 }
 875
 876 /*
 877  * debug aware fast / slowpath lock,trylock,unlock
 878  *
 879  * The atomic acquire/release ops are compiled away, when either the
 880  * architecture does not support cmpxchg or when debugging is enabled.
 881  */
 882 static inline int
 883 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 884                   int detect_deadlock,
 885                   int (*slowfn)(struct rt_mutex *lock, int state,
 886                                 struct hrtimer_sleeper *timeout,
 887                                 int detect_deadlock))
 888 {
 889         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 890                 rt_mutex_deadlock_account_lock(lock, current);
 891                 return 0;
 892         } else
 893                 return slowfn(lock, state, NULL, detect_deadlock);
 894 }
 895
 896 static inline int
 897 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 898                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 899                         int (*slowfn)(struct rt_mutex *lock, int state,
 900                                       struct hrtimer_sleeper *timeout,
 901                                       int detect_deadlock))
 902 {
 903         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 904                 rt_mutex_deadlock_account_lock(lock, current);
 905                 return 0;
 906         } else
 907                 return slowfn(lock, state, timeout, detect_deadlock);
 908 }
 909
 910 static inline int
 911 rt_mutex_fasttrylock(struct rt_mutex *lock,
 912                      int (*slowfn)(struct rt_mutex *lock))
 913 {
 914         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 915                 rt_mutex_deadlock_account_lock(lock, current);
 916                 return 1;
 917         }
 918         return slowfn(lock);
 919 }
 920
 921 static inline void
 922 rt_mutex_fastunlock(struct rt_mutex *lock,
 923                     void (*slowfn)(struct rt_mutex *lock))
 924 {
 925         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 926                 rt_mutex_deadlock_account_unlock(current);
 927         else
 928                 slowfn(lock);
 929 }
 930
 931 /**
 932  * rt_mutex_lock - lock a rt_mutex
 933  *
 934  * @lock: the rt_mutex to be locked
 935  */
 936 void __sched rt_mutex_lock(struct rt_mutex *lock)
 937 {
 938         might_sleep();
 939
 940         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 941 }
 942 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 943
 944 /**
 945  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 946  *
 947  * @lock:               the rt_mutex to be locked
 948  * @detect_deadlock:    deadlock detection on/off
 949  *
 950  * Returns:
 951  *  0           on success
 952  * -EINTR       when interrupted by a signal
 953  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 954  */
 955 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 956                                                  int detect_deadlock)
 957 {
 958         might_sleep();
 959
 960         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 961                                  detect_deadlock, rt_mutex_slowlock);
 962 }
 963 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 964
 965 /**
 966  * rt_mutex_timed_lock - lock a rt_mutex interruptible
 967  *                      the timeout structure is provided
 968  *                      by the caller
 969  *
 970  * @lock:               the rt_mutex to be locked
 971  * @timeout:            timeout structure or NULL (no timeout)
 972  * @detect_deadlock:    deadlock detection on/off
 973  *
 974  * Returns:
 975  *  0           on success
 976  * -EINTR       when interrupted by a signal
 977  * -ETIMEDOUT   when the timeout expired
 978  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 979  */
 980 int
 981 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 982                     int detect_deadlock)
 983 {
 984         might_sleep();
 985
 986         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 987                                        detect_deadlock, rt_mutex_slowlock);
 988 }
 989 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 990
 991 /**
 992  * rt_mutex_trylock - try to lock a rt_mutex
 993  *
 994  * @lock:       the rt_mutex to be locked
 995  *
 996  * Returns 1 on success and 0 on contention
 997  */
 998 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 999 {
1000         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1001 }
1002 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1003
1004 /**
1005  * rt_mutex_unlock - unlock a rt_mutex
1006  *
1007  * @lock: the rt_mutex to be unlocked
1008  */
1009 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1010 {
1011         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1012 }
1013 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1014
1015 /**
1016  * rt_mutex_destroy - mark a mutex unusable
1017  * @lock: the mutex to be destroyed
1018  *
1019  * This function marks the mutex uninitialized, and any subsequent
1020  * use of the mutex is forbidden. The mutex must not be locked when
1021  * this function is called.
1022  */
1023 void rt_mutex_destroy(struct rt_mutex *lock)
1024 {
1025         WARN_ON(rt_mutex_is_locked(lock));
1026 #ifdef CONFIG_DEBUG_RT_MUTEXES
1027         lock->magic = NULL;
1028 #endif
1029 }
1030
1031 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1032
1033 /**
1034  * __rt_mutex_init - initialize the rt lock
1035  *
1036  * @lock: the rt lock to be initialized
1037  *
1038  * Initialize the rt lock to unlocked state.
1039  *
1040  * Initializing of a locked rt lock is not allowed
1041  */
1042 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1043 {
1044         lock->owner = NULL;
1045         raw_spin_lock_init(&lock->wait_lock);
1046         lock->waiters = RB_ROOT;
1047         lock->waiters_leftmost = NULL;
1048
1049         debug_rt_mutex_init(lock, name);
1050 }
1051 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1052
1053 /**
1054  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1055  *                              proxy owner
1056  *
1057  * @lock:       the rt_mutex to be locked
1058  * @proxy_owner:the task to set as owner
1059  *
1060  * No locking. Caller has to do serializing itself
1061  * Special API call for PI-futex support
1062  */
1063 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1064                                 struct task_struct *proxy_owner)
1065 {
1066         __rt_mutex_init(lock, NULL);
1067         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1068         rt_mutex_set_owner(lock, proxy_owner);
1069         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1070 }
1071
1072 /**
1073  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1074  *
1075  * @lock:       the rt_mutex to be locked
1076  *
1077  * No locking. Caller has to do serializing itself
1078  * Special API call for PI-futex support
1079  */
1080 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1081                            struct task_struct *proxy_owner)
1082 {
1083         debug_rt_mutex_proxy_unlock(lock);
1084         rt_mutex_set_owner(lock, NULL);
1085         rt_mutex_deadlock_account_unlock(proxy_owner);
1086 }
1087
1088 /**
1089  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1090  * @lock:               the rt_mutex to take
1091  * @waiter:             the pre-initialized rt_mutex_waiter
1092  * @task:               the task to prepare
1093  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1094  *
1095  * Returns:
1096  *  0 - task blocked on lock
1097  *  1 - acquired the lock for task, caller should wake it up
1098  * <0 - error
1099  *
1100  * Special API call for FUTEX_REQUEUE_PI support.
1101  */
1102 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1103                               struct rt_mutex_waiter *waiter,
1104                               struct task_struct *task, int detect_deadlock)
1105 {
1106         int ret;
1107
1108         raw_spin_lock(&lock->wait_lock);
1109
1110         if (try_to_take_rt_mutex(lock, task, NULL)) {
1111                 raw_spin_unlock(&lock->wait_lock);
1112                 return 1;
1113         }
1114
1115         ret = task_blocks_on_rt_mutex(lock, waiter, task, detect_deadlock);
1116
1117         if (ret && !rt_mutex_owner(lock)) {
1118                 /*
1119                  * Reset the return value. We might have
1120                  * returned with -EDEADLK and the owner
1121                  * released the lock while we were walking the
1122                  * pi chain.  Let the waiter sort it out.
1123                  */
1124                 ret = 0;
1125         }
1126
1127         if (unlikely(ret))
1128                 remove_waiter(lock, waiter);
1129
1130         raw_spin_unlock(&lock->wait_lock);
1131
1132         debug_rt_mutex_print_deadlock(waiter);
1133
1134         return ret;
1135 }
1136
1137 /**
1138  * rt_mutex_next_owner - return the next owner of the lock
1139  *
1140  * @lock: the rt lock query
1141  *
1142  * Returns the next owner of the lock or NULL
1143  *
1144  * Caller has to serialize against other accessors to the lock
1145  * itself.
1146  *
1147  * Special API call for PI-futex support
1148  */
1149 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1150 {
1151         if (!rt_mutex_has_waiters(lock))
1152                 return NULL;
1153
1154         return rt_mutex_top_waiter(lock)->task;
1155 }
1156
1157 /**
1158  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1159  * @lock:               the rt_mutex we were woken on
1160  * @to:                 the timeout, null if none. hrtimer should already have
1161  *                      been started.
1162  * @waiter:             the pre-initialized rt_mutex_waiter
1163  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1164  *
1165  * Complete the lock acquisition started our behalf by another thread.
1166  *
1167  * Returns:
1168  *  0 - success
1169  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1170  *
1171  * Special API call for PI-futex requeue support
1172  */
1173 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1174                                struct hrtimer_sleeper *to,
1175                                struct rt_mutex_waiter *waiter,
1176                                int detect_deadlock)
1177 {
1178         int ret;
1179
1180         raw_spin_lock(&lock->wait_lock);
1181
1182         set_current_state(TASK_INTERRUPTIBLE);
1183
1184         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1185
1186         set_current_state(TASK_RUNNING);
1187
1188         if (unlikely(ret))
1189                 remove_waiter(lock, waiter);
1190
1191         /*
1192          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1193          * have to fix that up.
1194          */
1195         fixup_rt_mutex_waiters(lock);
1196
1197         raw_spin_unlock(&lock->wait_lock);
1198
1199         return ret;
1200 }