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 && (!task_has_pi_waiters(task) ||
 347                            top_waiter != task_top_pi_waiter(task)))
 348                 goto out_unlock_pi;
 349
 350         /*
 351          * When deadlock detection is off then we check, if further
 352          * priority adjustment is necessary.
 353          */
 354         if (!detect_deadlock && waiter->prio == task->prio)
 355                 goto out_unlock_pi;
 356
 357         lock = waiter->lock;
 358         if (!raw_spin_trylock(&lock->wait_lock)) {
 359                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 360                 cpu_relax();
 361                 goto retry;
 362         }
 363
 364         /* Deadlock detection */
 365         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 366                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 367                 raw_spin_unlock(&lock->wait_lock);
 368                 ret = deadlock_detect ? -EDEADLK : 0;
 369                 goto out_unlock_pi;
 370         }
 371
 372         top_waiter = rt_mutex_top_waiter(lock);
 373
 374         /* Requeue the waiter */
 375         rt_mutex_dequeue(lock, waiter);
 376         waiter->prio = task->prio;
 377         rt_mutex_enqueue(lock, waiter);
 378
 379         /* Release the task */
 380         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 381         if (!rt_mutex_owner(lock)) {
 382                 /*
 383                  * If the requeue above changed the top waiter, then we need
 384                  * to wake the new top waiter up to try to get the lock.
 385                  */
 386
 387                 if (top_waiter != rt_mutex_top_waiter(lock))
 388                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 389                 raw_spin_unlock(&lock->wait_lock);
 390                 goto out_put_task;
 391         }
 392         put_task_struct(task);
 393
 394         /* Grab the next task */
 395         task = rt_mutex_owner(lock);
 396         get_task_struct(task);
 397         raw_spin_lock_irqsave(&task->pi_lock, flags);
 398
 399         if (waiter == rt_mutex_top_waiter(lock)) {
 400                 /* Boost the owner */
 401                 rt_mutex_dequeue_pi(task, top_waiter);
 402                 rt_mutex_enqueue_pi(task, waiter);
 403                 __rt_mutex_adjust_prio(task);
 404
 405         } else if (top_waiter == waiter) {
 406                 /* Deboost the owner */
 407                 rt_mutex_dequeue_pi(task, waiter);
 408                 waiter = rt_mutex_top_waiter(lock);
 409                 rt_mutex_enqueue_pi(task, waiter);
 410                 __rt_mutex_adjust_prio(task);
 411         }
 412
 413         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 414
 415         top_waiter = rt_mutex_top_waiter(lock);
 416         raw_spin_unlock(&lock->wait_lock);
 417
 418         if (!detect_deadlock && waiter != top_waiter)
 419                 goto out_put_task;
 420
 421         goto again;
 422
 423  out_unlock_pi:
 424         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 425  out_put_task:
 426         put_task_struct(task);
 427
 428         return ret;
 429 }
 430
 431 /*
 432  * Try to take an rt-mutex
 433  *
 434  * Must be called with lock->wait_lock held.
 435  *
 436  * @lock:   the lock to be acquired.
 437  * @task:   the task which wants to acquire the lock
 438  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 439  */
 440 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 441                 struct rt_mutex_waiter *waiter)
 442 {
 443         /*
 444          * We have to be careful here if the atomic speedups are
 445          * enabled, such that, when
 446          *  - no other waiter is on the lock
 447          *  - the lock has been released since we did the cmpxchg
 448          * the lock can be released or taken while we are doing the
 449          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 450          *
 451          * The atomic acquire/release aware variant of
 452          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 453          * the WAITERS bit, the atomic release / acquire can not
 454          * happen anymore and lock->wait_lock protects us from the
 455          * non-atomic case.
 456          *
 457          * Note, that this might set lock->owner =
 458          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 459          * any more. This is fixed up when we take the ownership.
 460          * This is the transitional state explained at the top of this file.
 461          */
 462         mark_rt_mutex_waiters(lock);
 463
 464         if (rt_mutex_owner(lock))
 465                 return 0;
 466
 467         /*
 468          * It will get the lock because of one of these conditions:
 469          * 1) there is no waiter
 470          * 2) higher priority than waiters
 471          * 3) it is top waiter
 472          */
 473         if (rt_mutex_has_waiters(lock)) {
 474                 if (task->prio >= rt_mutex_top_waiter(lock)->prio) {
 475                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 476                                 return 0;
 477                 }
 478         }
 479
 480         if (waiter || rt_mutex_has_waiters(lock)) {
 481                 unsigned long flags;
 482                 struct rt_mutex_waiter *top;
 483
 484                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 485
 486                 /* remove the queued waiter. */
 487                 if (waiter) {
 488                         rt_mutex_dequeue(lock, waiter);
 489                         task->pi_blocked_on = NULL;
 490                 }
 491
 492                 /*
 493                  * We have to enqueue the top waiter(if it exists) into
 494                  * task->pi_waiters list.
 495                  */
 496                 if (rt_mutex_has_waiters(lock)) {
 497                         top = rt_mutex_top_waiter(lock);
 498                         rt_mutex_enqueue_pi(task, top);
 499                 }
 500                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 501         }
 502
 503         /* We got the lock. */
 504         debug_rt_mutex_lock(lock);
 505
 506         rt_mutex_set_owner(lock, task);
 507
 508         rt_mutex_deadlock_account_lock(lock, task);
 509
 510         return 1;
 511 }
 512
 513 /*
 514  * Task blocks on lock.
 515  *
 516  * Prepare waiter and propagate pi chain
 517  *
 518  * This must be called with lock->wait_lock held.
 519  */
 520 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 521                                    struct rt_mutex_waiter *waiter,
 522                                    struct task_struct *task,
 523                                    int detect_deadlock)
 524 {
 525         struct task_struct *owner = rt_mutex_owner(lock);
 526         struct rt_mutex_waiter *top_waiter = waiter;
 527         unsigned long flags;
 528         int chain_walk = 0, res;
 529
 530         raw_spin_lock_irqsave(&task->pi_lock, flags);
 531         __rt_mutex_adjust_prio(task);
 532         waiter->task = task;
 533         waiter->lock = lock;
 534         waiter->prio = task->prio;
 535
 536         /* Get the top priority waiter on the lock */
 537         if (rt_mutex_has_waiters(lock))
 538                 top_waiter = rt_mutex_top_waiter(lock);
 539         rt_mutex_enqueue(lock, waiter);
 540
 541         task->pi_blocked_on = waiter;
 542
 543         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 544
 545         if (!owner)
 546                 return 0;
 547
 548         if (waiter == rt_mutex_top_waiter(lock)) {
 549                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 550                 rt_mutex_dequeue_pi(owner, top_waiter);
 551                 rt_mutex_enqueue_pi(owner, waiter);
 552
 553                 __rt_mutex_adjust_prio(owner);
 554                 if (owner->pi_blocked_on)
 555                         chain_walk = 1;
 556                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 557         }
 558         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 559                 chain_walk = 1;
 560
 561         if (!chain_walk)
 562                 return 0;
 563
 564         /*
 565          * The owner can't disappear while holding a lock,
 566          * so the owner struct is protected by wait_lock.
 567          * Gets dropped in rt_mutex_adjust_prio_chain()!
 568          */
 569         get_task_struct(owner);
 570
 571         raw_spin_unlock(&lock->wait_lock);
 572
 573         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 574                                          task);
 575
 576         raw_spin_lock(&lock->wait_lock);
 577
 578         return res;
 579 }
 580
 581 /*
 582  * Wake up the next waiter on the lock.
 583  *
 584  * Remove the top waiter from the current tasks waiter list and wake it up.
 585  *
 586  * Called with lock->wait_lock held.
 587  */
 588 static void wakeup_next_waiter(struct rt_mutex *lock)
 589 {
 590         struct rt_mutex_waiter *waiter;
 591         unsigned long flags;
 592
 593         raw_spin_lock_irqsave(&current->pi_lock, flags);
 594
 595         waiter = rt_mutex_top_waiter(lock);
 596
 597         /*
 598          * Remove it from current->pi_waiters. We do not adjust a
 599          * possible priority boost right now. We execute wakeup in the
 600          * boosted mode and go back to normal after releasing
 601          * lock->wait_lock.
 602          */
 603         rt_mutex_dequeue_pi(current, waiter);
 604
 605         rt_mutex_set_owner(lock, NULL);
 606
 607         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 608
 609         wake_up_process(waiter->task);
 610 }
 611
 612 /*
 613  * Remove a waiter from a lock and give up
 614  *
 615  * Must be called with lock->wait_lock held and
 616  * have just failed to try_to_take_rt_mutex().
 617  */
 618 static void remove_waiter(struct rt_mutex *lock,
 619                           struct rt_mutex_waiter *waiter)
 620 {
 621         int first = (waiter == rt_mutex_top_waiter(lock));
 622         struct task_struct *owner = rt_mutex_owner(lock);
 623         unsigned long flags;
 624         int chain_walk = 0;
 625
 626         raw_spin_lock_irqsave(&current->pi_lock, flags);
 627         rt_mutex_dequeue(lock, waiter);
 628         current->pi_blocked_on = NULL;
 629         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 630
 631         if (!owner)
 632                 return;
 633
 634         if (first) {
 635
 636                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 637
 638                 rt_mutex_dequeue_pi(owner, waiter);
 639
 640                 if (rt_mutex_has_waiters(lock)) {
 641                         struct rt_mutex_waiter *next;
 642
 643                         next = rt_mutex_top_waiter(lock);
 644                         rt_mutex_enqueue_pi(owner, next);
 645                 }
 646                 __rt_mutex_adjust_prio(owner);
 647
 648                 if (owner->pi_blocked_on)
 649                         chain_walk = 1;
 650
 651                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 652         }
 653
 654         if (!chain_walk)
 655                 return;
 656
 657         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 658         get_task_struct(owner);
 659
 660         raw_spin_unlock(&lock->wait_lock);
 661
 662         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 663
 664         raw_spin_lock(&lock->wait_lock);
 665 }
 666
 667 /*
 668  * Recheck the pi chain, in case we got a priority setting
 669  *
 670  * Called from sched_setscheduler
 671  */
 672 void rt_mutex_adjust_pi(struct task_struct *task)
 673 {
 674         struct rt_mutex_waiter *waiter;
 675         unsigned long flags;
 676
 677         raw_spin_lock_irqsave(&task->pi_lock, flags);
 678
 679         waiter = task->pi_blocked_on;
 680         if (!waiter || (waiter->prio == task->prio &&
 681                         !dl_prio(task->prio))) {
 682                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 683                 return;
 684         }
 685
 686         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 687
 688         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 689         get_task_struct(task);
 690         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 691 }
 692
 693 /**
 694  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 695  * @lock:                the rt_mutex to take
 696  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 697  *                       or TASK_UNINTERRUPTIBLE)
 698  * @timeout:             the pre-initialized and started timer, or NULL for none
 699  * @waiter:              the pre-initialized rt_mutex_waiter
 700  *
 701  * lock->wait_lock must be held by the caller.
 702  */
 703 static int __sched
 704 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 705                     struct hrtimer_sleeper *timeout,
 706                     struct rt_mutex_waiter *waiter)
 707 {
 708         int ret = 0;
 709
 710         for (;;) {
 711                 /* Try to acquire the lock: */
 712                 if (try_to_take_rt_mutex(lock, current, waiter))
 713                         break;
 714
 715                 /*
 716                  * TASK_INTERRUPTIBLE checks for signals and
 717                  * timeout. Ignored otherwise.
 718                  */
 719                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 720                         /* Signal pending? */
 721                         if (signal_pending(current))
 722                                 ret = -EINTR;
 723                         if (timeout && !timeout->task)
 724                                 ret = -ETIMEDOUT;
 725                         if (ret)
 726                                 break;
 727                 }
 728
 729                 raw_spin_unlock(&lock->wait_lock);
 730
 731                 debug_rt_mutex_print_deadlock(waiter);
 732
 733                 schedule_rt_mutex(lock);
 734
 735                 raw_spin_lock(&lock->wait_lock);
 736                 set_current_state(state);
 737         }
 738
 739         return ret;
 740 }
 741
 742 /*
 743  * Slow path lock function:
 744  */
 745 static int __sched
 746 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 747                   struct hrtimer_sleeper *timeout,
 748                   int detect_deadlock)
 749 {
 750         struct rt_mutex_waiter waiter;
 751         int ret = 0;
 752
 753         debug_rt_mutex_init_waiter(&waiter);
 754         RB_CLEAR_NODE(&waiter.pi_tree_entry);
 755         RB_CLEAR_NODE(&waiter.tree_entry);
 756
 757         raw_spin_lock(&lock->wait_lock);
 758
 759         /* Try to acquire the lock again: */
 760         if (try_to_take_rt_mutex(lock, current, NULL)) {
 761                 raw_spin_unlock(&lock->wait_lock);
 762                 return 0;
 763         }
 764
 765         set_current_state(state);
 766
 767         /* Setup the timer, when timeout != NULL */
 768         if (unlikely(timeout)) {
 769                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 770                 if (!hrtimer_active(&timeout->timer))
 771                         timeout->task = NULL;
 772         }
 773
 774         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 775
 776         if (likely(!ret))
 777                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
 778
 779         set_current_state(TASK_RUNNING);
 780
 781         if (unlikely(ret))
 782                 remove_waiter(lock, &waiter);
 783
 784         /*
 785          * try_to_take_rt_mutex() sets the waiter bit
 786          * unconditionally. We might have to fix that up.
 787          */
 788         fixup_rt_mutex_waiters(lock);
 789
 790         raw_spin_unlock(&lock->wait_lock);
 791
 792         /* Remove pending timer: */
 793         if (unlikely(timeout))
 794                 hrtimer_cancel(&timeout->timer);
 795
 796         debug_rt_mutex_free_waiter(&waiter);
 797
 798         return ret;
 799 }
 800
 801 /*
 802  * Slow path try-lock function:
 803  */
 804 static inline int
 805 rt_mutex_slowtrylock(struct rt_mutex *lock)
 806 {
 807         int ret = 0;
 808
 809         raw_spin_lock(&lock->wait_lock);
 810
 811         if (likely(rt_mutex_owner(lock) != current)) {
 812
 813                 ret = try_to_take_rt_mutex(lock, current, NULL);
 814                 /*
 815                  * try_to_take_rt_mutex() sets the lock waiters
 816                  * bit unconditionally. Clean this up.
 817                  */
 818                 fixup_rt_mutex_waiters(lock);
 819         }
 820
 821         raw_spin_unlock(&lock->wait_lock);
 822
 823         return ret;
 824 }
 825
 826 /*
 827  * Slow path to release a rt-mutex:
 828  */
 829 static void __sched
 830 rt_mutex_slowunlock(struct rt_mutex *lock)
 831 {
 832         raw_spin_lock(&lock->wait_lock);
 833
 834         debug_rt_mutex_unlock(lock);
 835
 836         rt_mutex_deadlock_account_unlock(current);
 837
 838         if (!rt_mutex_has_waiters(lock)) {
 839                 lock->owner = NULL;
 840                 raw_spin_unlock(&lock->wait_lock);
 841                 return;
 842         }
 843
 844         wakeup_next_waiter(lock);
 845
 846         raw_spin_unlock(&lock->wait_lock);
 847
 848         /* Undo pi boosting if necessary: */
 849         rt_mutex_adjust_prio(current);
 850 }
 851
 852 /*
 853  * debug aware fast / slowpath lock,trylock,unlock
 854  *
 855  * The atomic acquire/release ops are compiled away, when either the
 856  * architecture does not support cmpxchg or when debugging is enabled.
 857  */
 858 static inline int
 859 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 860                   int detect_deadlock,
 861                   int (*slowfn)(struct rt_mutex *lock, int state,
 862                                 struct hrtimer_sleeper *timeout,
 863                                 int detect_deadlock))
 864 {
 865         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 866                 rt_mutex_deadlock_account_lock(lock, current);
 867                 return 0;
 868         } else
 869                 return slowfn(lock, state, NULL, detect_deadlock);
 870 }
 871
 872 static inline int
 873 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 874                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 875                         int (*slowfn)(struct rt_mutex *lock, int state,
 876                                       struct hrtimer_sleeper *timeout,
 877                                       int detect_deadlock))
 878 {
 879         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 880                 rt_mutex_deadlock_account_lock(lock, current);
 881                 return 0;
 882         } else
 883                 return slowfn(lock, state, timeout, detect_deadlock);
 884 }
 885
 886 static inline int
 887 rt_mutex_fasttrylock(struct rt_mutex *lock,
 888                      int (*slowfn)(struct rt_mutex *lock))
 889 {
 890         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 891                 rt_mutex_deadlock_account_lock(lock, current);
 892                 return 1;
 893         }
 894         return slowfn(lock);
 895 }
 896
 897 static inline void
 898 rt_mutex_fastunlock(struct rt_mutex *lock,
 899                     void (*slowfn)(struct rt_mutex *lock))
 900 {
 901         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 902                 rt_mutex_deadlock_account_unlock(current);
 903         else
 904                 slowfn(lock);
 905 }
 906
 907 /**
 908  * rt_mutex_lock - lock a rt_mutex
 909  *
 910  * @lock: the rt_mutex to be locked
 911  */
 912 void __sched rt_mutex_lock(struct rt_mutex *lock)
 913 {
 914         might_sleep();
 915
 916         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 917 }
 918 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 919
 920 /**
 921  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 922  *
 923  * @lock:               the rt_mutex to be locked
 924  * @detect_deadlock:    deadlock detection on/off
 925  *
 926  * Returns:
 927  *  0           on success
 928  * -EINTR       when interrupted by a signal
 929  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 930  */
 931 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 932                                                  int detect_deadlock)
 933 {
 934         might_sleep();
 935
 936         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 937                                  detect_deadlock, rt_mutex_slowlock);
 938 }
 939 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 940
 941 /**
 942  * rt_mutex_timed_lock - lock a rt_mutex interruptible
 943  *                      the timeout structure is provided
 944  *                      by the caller
 945  *
 946  * @lock:               the rt_mutex to be locked
 947  * @timeout:            timeout structure or NULL (no timeout)
 948  * @detect_deadlock:    deadlock detection on/off
 949  *
 950  * Returns:
 951  *  0           on success
 952  * -EINTR       when interrupted by a signal
 953  * -ETIMEDOUT   when the timeout expired
 954  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 955  */
 956 int
 957 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 958                     int detect_deadlock)
 959 {
 960         might_sleep();
 961
 962         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 963                                        detect_deadlock, rt_mutex_slowlock);
 964 }
 965 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 966
 967 /**
 968  * rt_mutex_trylock - try to lock a rt_mutex
 969  *
 970  * @lock:       the rt_mutex to be locked
 971  *
 972  * Returns 1 on success and 0 on contention
 973  */
 974 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 975 {
 976         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 977 }
 978 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 979
 980 /**
 981  * rt_mutex_unlock - unlock a rt_mutex
 982  *
 983  * @lock: the rt_mutex to be unlocked
 984  */
 985 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 986 {
 987         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 988 }
 989 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 990
 991 /**
 992  * rt_mutex_destroy - mark a mutex unusable
 993  * @lock: the mutex to be destroyed
 994  *
 995  * This function marks the mutex uninitialized, and any subsequent
 996  * use of the mutex is forbidden. The mutex must not be locked when
 997  * this function is called.
 998  */
 999 void rt_mutex_destroy(struct rt_mutex *lock)
1000 {
1001         WARN_ON(rt_mutex_is_locked(lock));
1002 #ifdef CONFIG_DEBUG_RT_MUTEXES
1003         lock->magic = NULL;
1004 #endif
1005 }
1006
1007 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1008
1009 /**
1010  * __rt_mutex_init - initialize the rt lock
1011  *
1012  * @lock: the rt lock to be initialized
1013  *
1014  * Initialize the rt lock to unlocked state.
1015  *
1016  * Initializing of a locked rt lock is not allowed
1017  */
1018 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1019 {
1020         lock->owner = NULL;
1021         raw_spin_lock_init(&lock->wait_lock);
1022         lock->waiters = RB_ROOT;
1023         lock->waiters_leftmost = NULL;
1024
1025         debug_rt_mutex_init(lock, name);
1026 }
1027 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1028
1029 /**
1030  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1031  *                              proxy owner
1032  *
1033  * @lock:       the rt_mutex to be locked
1034  * @proxy_owner:the task to set as owner
1035  *
1036  * No locking. Caller has to do serializing itself
1037  * Special API call for PI-futex support
1038  */
1039 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1040                                 struct task_struct *proxy_owner)
1041 {
1042         __rt_mutex_init(lock, NULL);
1043         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1044         rt_mutex_set_owner(lock, proxy_owner);
1045         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1046 }
1047
1048 /**
1049  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1050  *
1051  * @lock:       the rt_mutex to be locked
1052  *
1053  * No locking. Caller has to do serializing itself
1054  * Special API call for PI-futex support
1055  */
1056 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1057                            struct task_struct *proxy_owner)
1058 {
1059         debug_rt_mutex_proxy_unlock(lock);
1060         rt_mutex_set_owner(lock, NULL);
1061         rt_mutex_deadlock_account_unlock(proxy_owner);
1062 }
1063
1064 /**
1065  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1066  * @lock:               the rt_mutex to take
1067  * @waiter:             the pre-initialized rt_mutex_waiter
1068  * @task:               the task to prepare
1069  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1070  *
1071  * Returns:
1072  *  0 - task blocked on lock
1073  *  1 - acquired the lock for task, caller should wake it up
1074  * <0 - error
1075  *
1076  * Special API call for FUTEX_REQUEUE_PI support.
1077  */
1078 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1079                               struct rt_mutex_waiter *waiter,
1080                               struct task_struct *task, int detect_deadlock)
1081 {
1082         int ret;
1083
1084         raw_spin_lock(&lock->wait_lock);
1085
1086         if (try_to_take_rt_mutex(lock, task, NULL)) {
1087                 raw_spin_unlock(&lock->wait_lock);
1088                 return 1;
1089         }
1090
1091         ret = task_blocks_on_rt_mutex(lock, waiter, task, detect_deadlock);
1092
1093         if (ret && !rt_mutex_owner(lock)) {
1094                 /*
1095                  * Reset the return value. We might have
1096                  * returned with -EDEADLK and the owner
1097                  * released the lock while we were walking the
1098                  * pi chain.  Let the waiter sort it out.
1099                  */
1100                 ret = 0;
1101         }
1102
1103         if (unlikely(ret))
1104                 remove_waiter(lock, waiter);
1105
1106         raw_spin_unlock(&lock->wait_lock);
1107
1108         debug_rt_mutex_print_deadlock(waiter);
1109
1110         return ret;
1111 }
1112
1113 /**
1114  * rt_mutex_next_owner - return the next owner of the lock
1115  *
1116  * @lock: the rt lock query
1117  *
1118  * Returns the next owner of the lock or NULL
1119  *
1120  * Caller has to serialize against other accessors to the lock
1121  * itself.
1122  *
1123  * Special API call for PI-futex support
1124  */
1125 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1126 {
1127         if (!rt_mutex_has_waiters(lock))
1128                 return NULL;
1129
1130         return rt_mutex_top_waiter(lock)->task;
1131 }
1132
1133 /**
1134  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1135  * @lock:               the rt_mutex we were woken on
1136  * @to:                 the timeout, null if none. hrtimer should already have
1137  *                      been started.
1138  * @waiter:             the pre-initialized rt_mutex_waiter
1139  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1140  *
1141  * Complete the lock acquisition started our behalf by another thread.
1142  *
1143  * Returns:
1144  *  0 - success
1145  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1146  *
1147  * Special API call for PI-futex requeue support
1148  */
1149 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1150                                struct hrtimer_sleeper *to,
1151                                struct rt_mutex_waiter *waiter,
1152                                int detect_deadlock)
1153 {
1154         int ret;
1155
1156         raw_spin_lock(&lock->wait_lock);
1157
1158         set_current_state(TASK_INTERRUPTIBLE);
1159
1160         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1161
1162         set_current_state(TASK_RUNNING);
1163
1164         if (unlikely(ret))
1165                 remove_waiter(lock, waiter);
1166
1167         /*
1168          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1169          * have to fix that up.
1170          */
1171         fixup_rt_mutex_waiters(lock);
1172
1173         raw_spin_unlock(&lock->wait_lock);
1174
1175         return ret;
1176 }