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