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/locking/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 there's no debugging state to be
  74  * set up.
  75  */
  76 #ifndef CONFIG_DEBUG_RT_MUTEXES
  77 # define rt_mutex_cmpxchg_relaxed(l,c,n) (cmpxchg_relaxed(&l->owner, c, n) == c)
  78 # define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
  79 # define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
  80
  81 /*
  82  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
  83  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
  84  * relaxed semantics suffice.
  85  */
  86 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  87 {
  88         unsigned long owner, *p = (unsigned long *) &lock->owner;
  89
  90         do {
  91                 owner = *p;
  92         } while (cmpxchg_relaxed(p, owner,
  93                                  owner | RT_MUTEX_HAS_WAITERS) != owner);
  94 }
  95
  96 /*
  97  * Safe fastpath aware unlock:
  98  * 1) Clear the waiters bit
  99  * 2) Drop lock->wait_lock
 100  * 3) Try to unlock the lock with cmpxchg
 101  */
 102 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
 103         __releases(lock->wait_lock)
 104 {
 105         struct task_struct *owner = rt_mutex_owner(lock);
 106
 107         clear_rt_mutex_waiters(lock);
 108         raw_spin_unlock(&lock->wait_lock);
 109         /*
 110          * If a new waiter comes in between the unlock and the cmpxchg
 111          * we have two situations:
 112          *
 113          * unlock(wait_lock);
 114          *                                      lock(wait_lock);
 115          * cmpxchg(p, owner, 0) == owner
 116          *                                      mark_rt_mutex_waiters(lock);
 117          *                                      acquire(lock);
 118          * or:
 119          *
 120          * unlock(wait_lock);
 121          *                                      lock(wait_lock);
 122          *                                      mark_rt_mutex_waiters(lock);
 123          *
 124          * cmpxchg(p, owner, 0) != owner
 125          *                                      enqueue_waiter();
 126          *                                      unlock(wait_lock);
 127          * lock(wait_lock);
 128          * wake waiter();
 129          * unlock(wait_lock);
 130          *                                      lock(wait_lock);
 131          *                                      acquire(lock);
 132          */
 133         return rt_mutex_cmpxchg_release(lock, owner, NULL);
 134 }
 135
 136 #else
 137 # define rt_mutex_cmpxchg_relaxed(l,c,n)        (0)
 138 # define rt_mutex_cmpxchg_acquire(l,c,n)        (0)
 139 # define rt_mutex_cmpxchg_release(l,c,n)        (0)
 140
 141 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 142 {
 143         lock->owner = (struct task_struct *)
 144                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 145 }
 146
 147 /*
 148  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 149  */
 150 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock)
 151         __releases(lock->wait_lock)
 152 {
 153         lock->owner = NULL;
 154         raw_spin_unlock(&lock->wait_lock);
 155         return true;
 156 }
 157 #endif
 158
 159 static inline int
 160 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
 161                      struct rt_mutex_waiter *right)
 162 {
 163         if (left->prio < right->prio)
 164                 return 1;
 165
 166         /*
 167          * If both waiters have dl_prio(), we check the deadlines of the
 168          * associated tasks.
 169          * If left waiter has a dl_prio(), and we didn't return 1 above,
 170          * then right waiter has a dl_prio() too.
 171          */
 172         if (dl_prio(left->prio))
 173                 return dl_time_before(left->task->dl.deadline,
 174                                       right->task->dl.deadline);
 175
 176         return 0;
 177 }
 178
 179 static void
 180 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 181 {
 182         struct rb_node **link = &lock->waiters.rb_node;
 183         struct rb_node *parent = NULL;
 184         struct rt_mutex_waiter *entry;
 185         int leftmost = 1;
 186
 187         while (*link) {
 188                 parent = *link;
 189                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
 190                 if (rt_mutex_waiter_less(waiter, entry)) {
 191                         link = &parent->rb_left;
 192                 } else {
 193                         link = &parent->rb_right;
 194                         leftmost = 0;
 195                 }
 196         }
 197
 198         if (leftmost)
 199                 lock->waiters_leftmost = &waiter->tree_entry;
 200
 201         rb_link_node(&waiter->tree_entry, parent, link);
 202         rb_insert_color(&waiter->tree_entry, &lock->waiters);
 203 }
 204
 205 static void
 206 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 207 {
 208         if (RB_EMPTY_NODE(&waiter->tree_entry))
 209                 return;
 210
 211         if (lock->waiters_leftmost == &waiter->tree_entry)
 212                 lock->waiters_leftmost = rb_next(&waiter->tree_entry);
 213
 214         rb_erase(&waiter->tree_entry, &lock->waiters);
 215         RB_CLEAR_NODE(&waiter->tree_entry);
 216 }
 217
 218 static void
 219 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 220 {
 221         struct rb_node **link = &task->pi_waiters.rb_node;
 222         struct rb_node *parent = NULL;
 223         struct rt_mutex_waiter *entry;
 224         int leftmost = 1;
 225
 226         while (*link) {
 227                 parent = *link;
 228                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
 229                 if (rt_mutex_waiter_less(waiter, entry)) {
 230                         link = &parent->rb_left;
 231                 } else {
 232                         link = &parent->rb_right;
 233                         leftmost = 0;
 234                 }
 235         }
 236
 237         if (leftmost)
 238                 task->pi_waiters_leftmost = &waiter->pi_tree_entry;
 239
 240         rb_link_node(&waiter->pi_tree_entry, parent, link);
 241         rb_insert_color(&waiter->pi_tree_entry, &task->pi_waiters);
 242 }
 243
 244 static void
 245 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 246 {
 247         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 248                 return;
 249
 250         if (task->pi_waiters_leftmost == &waiter->pi_tree_entry)
 251                 task->pi_waiters_leftmost = rb_next(&waiter->pi_tree_entry);
 252
 253         rb_erase(&waiter->pi_tree_entry, &task->pi_waiters);
 254         RB_CLEAR_NODE(&waiter->pi_tree_entry);
 255 }
 256
 257 /*
 258  * Calculate task priority from the waiter tree priority
 259  *
 260  * Return task->normal_prio when the waiter tree is empty or when
 261  * the waiter is not allowed to do priority boosting
 262  */
 263 int rt_mutex_getprio(struct task_struct *task)
 264 {
 265         if (likely(!task_has_pi_waiters(task)))
 266                 return task->normal_prio;
 267
 268         return min(task_top_pi_waiter(task)->prio,
 269                    task->normal_prio);
 270 }
 271
 272 struct task_struct *rt_mutex_get_top_task(struct task_struct *task)
 273 {
 274         if (likely(!task_has_pi_waiters(task)))
 275                 return NULL;
 276
 277         return task_top_pi_waiter(task)->task;
 278 }
 279
 280 /*
 281  * Called by sched_setscheduler() to get the priority which will be
 282  * effective after the change.
 283  */
 284 int rt_mutex_get_effective_prio(struct task_struct *task, int newprio)
 285 {
 286         if (!task_has_pi_waiters(task))
 287                 return newprio;
 288
 289         if (task_top_pi_waiter(task)->task->prio <= newprio)
 290                 return task_top_pi_waiter(task)->task->prio;
 291         return newprio;
 292 }
 293
 294 /*
 295  * Adjust the priority of a task, after its pi_waiters got modified.
 296  *
 297  * This can be both boosting and unboosting. task->pi_lock must be held.
 298  */
 299 static void __rt_mutex_adjust_prio(struct task_struct *task)
 300 {
 301         int prio = rt_mutex_getprio(task);
 302
 303         if (task->prio != prio || dl_prio(prio))
 304                 rt_mutex_setprio(task, prio);
 305 }
 306
 307 /*
 308  * Adjust task priority (undo boosting). Called from the exit path of
 309  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 310  *
 311  * (Note: We do this outside of the protection of lock->wait_lock to
 312  * allow the lock to be taken while or before we readjust the priority
 313  * of task. We do not use the spin_xx_mutex() variants here as we are
 314  * outside of the debug path.)
 315  */
 316 void rt_mutex_adjust_prio(struct task_struct *task)
 317 {
 318         unsigned long flags;
 319
 320         raw_spin_lock_irqsave(&task->pi_lock, flags);
 321         __rt_mutex_adjust_prio(task);
 322         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 323 }
 324
 325 /*
 326  * Deadlock detection is conditional:
 327  *
 328  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
 329  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
 330  *
 331  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
 332  * conducted independent of the detect argument.
 333  *
 334  * If the waiter argument is NULL this indicates the deboost path and
 335  * deadlock detection is disabled independent of the detect argument
 336  * and the config settings.
 337  */
 338 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
 339                                           enum rtmutex_chainwalk chwalk)
 340 {
 341         /*
 342          * This is just a wrapper function for the following call,
 343          * because debug_rt_mutex_detect_deadlock() smells like a magic
 344          * debug feature and I wanted to keep the cond function in the
 345          * main source file along with the comments instead of having
 346          * two of the same in the headers.
 347          */
 348         return debug_rt_mutex_detect_deadlock(waiter, chwalk);
 349 }
 350
 351 /*
 352  * Max number of times we'll walk the boosting chain:
 353  */
 354 int max_lock_depth = 1024;
 355
 356 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 357 {
 358         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 359 }
 360
 361 /*
 362  * Adjust the priority chain. Also used for deadlock detection.
 363  * Decreases task's usage by one - may thus free the task.
 364  *
 365  * @task:       the task owning the mutex (owner) for which a chain walk is
 366  *              probably needed
 367  * @chwalk:     do we have to carry out deadlock detection?
 368  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 369  *              things for a task that has just got its priority adjusted, and
 370  *              is waiting on a mutex)
 371  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 372  *              we dropped its pi_lock. Is never dereferenced, only used for
 373  *              comparison to detect lock chain changes.
 374  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 375  *              its priority to the mutex owner (can be NULL in the case
 376  *              depicted above or if the top waiter is gone away and we are
 377  *              actually deboosting the owner)
 378  * @top_task:   the current top waiter
 379  *
 380  * Returns 0 or -EDEADLK.
 381  *
 382  * Chain walk basics and protection scope
 383  *
 384  * [R] refcount on task
 385  * [P] task->pi_lock held
 386  * [L] rtmutex->wait_lock held
 387  *
 388  * Step Description                             Protected by
 389  *      function arguments:
 390  *      @task                                   [R]
 391  *      @orig_lock if != NULL                   @top_task is blocked on it
 392  *      @next_lock                              Unprotected. Cannot be
 393  *                                              dereferenced. Only used for
 394  *                                              comparison.
 395  *      @orig_waiter if != NULL                 @top_task is blocked on it
 396  *      @top_task                               current, or in case of proxy
 397  *                                              locking protected by calling
 398  *                                              code
 399  *      again:
 400  *        loop_sanity_check();
 401  *      retry:
 402  * [1]    lock(task->pi_lock);                  [R] acquire [P]
 403  * [2]    waiter = task->pi_blocked_on;         [P]
 404  * [3]    check_exit_conditions_1();            [P]
 405  * [4]    lock = waiter->lock;                  [P]
 406  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
 407  *          unlock(task->pi_lock);              release [P]
 408  *          goto retry;
 409  *        }
 410  * [6]    check_exit_conditions_2();            [P] + [L]
 411  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
 412  * [8]    unlock(task->pi_lock);                release [P]
 413  *        put_task_struct(task);                release [R]
 414  * [9]    check_exit_conditions_3();            [L]
 415  * [10]   task = owner(lock);                   [L]
 416  *        get_task_struct(task);                [L] acquire [R]
 417  *        lock(task->pi_lock);                  [L] acquire [P]
 418  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
 419  * [12]   check_exit_conditions_4();            [P] + [L]
 420  * [13]   unlock(task->pi_lock);                release [P]
 421  *        unlock(lock->wait_lock);              release [L]
 422  *        goto again;
 423  */
 424 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 425                                       enum rtmutex_chainwalk chwalk,
 426                                       struct rt_mutex *orig_lock,
 427                                       struct rt_mutex *next_lock,
 428                                       struct rt_mutex_waiter *orig_waiter,
 429                                       struct task_struct *top_task)
 430 {
 431         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 432         struct rt_mutex_waiter *prerequeue_top_waiter;
 433         int ret = 0, depth = 0;
 434         struct rt_mutex *lock;
 435         bool detect_deadlock;
 436         unsigned long flags;
 437         bool requeue = true;
 438
 439         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
 440
 441         /*
 442          * The (de)boosting is a step by step approach with a lot of
 443          * pitfalls. We want this to be preemptible and we want hold a
 444          * maximum of two locks per step. So we have to check
 445          * carefully whether things change under us.
 446          */
 447  again:
 448         /*
 449          * We limit the lock chain length for each invocation.
 450          */
 451         if (++depth > max_lock_depth) {
 452                 static int prev_max;
 453
 454                 /*
 455                  * Print this only once. If the admin changes the limit,
 456                  * print a new message when reaching the limit again.
 457                  */
 458                 if (prev_max != max_lock_depth) {
 459                         prev_max = max_lock_depth;
 460                         printk(KERN_WARNING "Maximum lock depth %d reached "
 461                                "task: %s (%d)\n", max_lock_depth,
 462                                top_task->comm, task_pid_nr(top_task));
 463                 }
 464                 put_task_struct(task);
 465
 466                 return -EDEADLK;
 467         }
 468
 469         /*
 470          * We are fully preemptible here and only hold the refcount on
 471          * @task. So everything can have changed under us since the
 472          * caller or our own code below (goto retry/again) dropped all
 473          * locks.
 474          */
 475  retry:
 476         /*
 477          * [1] Task cannot go away as we did a get_task() before !
 478          */
 479         raw_spin_lock_irqsave(&task->pi_lock, flags);
 480
 481         /*
 482          * [2] Get the waiter on which @task is blocked on.
 483          */
 484         waiter = task->pi_blocked_on;
 485
 486         /*
 487          * [3] check_exit_conditions_1() protected by task->pi_lock.
 488          */
 489
 490         /*
 491          * Check whether the end of the boosting chain has been
 492          * reached or the state of the chain has changed while we
 493          * dropped the locks.
 494          */
 495         if (!waiter)
 496                 goto out_unlock_pi;
 497
 498         /*
 499          * Check the orig_waiter state. After we dropped the locks,
 500          * the previous owner of the lock might have released the lock.
 501          */
 502         if (orig_waiter && !rt_mutex_owner(orig_lock))
 503                 goto out_unlock_pi;
 504
 505         /*
 506          * We dropped all locks after taking a refcount on @task, so
 507          * the task might have moved on in the lock chain or even left
 508          * the chain completely and blocks now on an unrelated lock or
 509          * on @orig_lock.
 510          *
 511          * We stored the lock on which @task was blocked in @next_lock,
 512          * so we can detect the chain change.
 513          */
 514         if (next_lock != waiter->lock)
 515                 goto out_unlock_pi;
 516
 517         /*
 518          * Drop out, when the task has no waiters. Note,
 519          * top_waiter can be NULL, when we are in the deboosting
 520          * mode!
 521          */
 522         if (top_waiter) {
 523                 if (!task_has_pi_waiters(task))
 524                         goto out_unlock_pi;
 525                 /*
 526                  * If deadlock detection is off, we stop here if we
 527                  * are not the top pi waiter of the task. If deadlock
 528                  * detection is enabled we continue, but stop the
 529                  * requeueing in the chain walk.
 530                  */
 531                 if (top_waiter != task_top_pi_waiter(task)) {
 532                         if (!detect_deadlock)
 533                                 goto out_unlock_pi;
 534                         else
 535                                 requeue = false;
 536                 }
 537         }
 538
 539         /*
 540          * If the waiter priority is the same as the task priority
 541          * then there is no further priority adjustment necessary.  If
 542          * deadlock detection is off, we stop the chain walk. If its
 543          * enabled we continue, but stop the requeueing in the chain
 544          * walk.
 545          */
 546         if (waiter->prio == task->prio) {
 547                 if (!detect_deadlock)
 548                         goto out_unlock_pi;
 549                 else
 550                         requeue = false;
 551         }
 552
 553         /*
 554          * [4] Get the next lock
 555          */
 556         lock = waiter->lock;
 557         /*
 558          * [5] We need to trylock here as we are holding task->pi_lock,
 559          * which is the reverse lock order versus the other rtmutex
 560          * operations.
 561          */
 562         if (!raw_spin_trylock(&lock->wait_lock)) {
 563                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 564                 cpu_relax();
 565                 goto retry;
 566         }
 567
 568         /*
 569          * [6] check_exit_conditions_2() protected by task->pi_lock and
 570          * lock->wait_lock.
 571          *
 572          * Deadlock detection. If the lock is the same as the original
 573          * lock which caused us to walk the lock chain or if the
 574          * current lock is owned by the task which initiated the chain
 575          * walk, we detected a deadlock.
 576          */
 577         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 578                 debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
 579                 raw_spin_unlock(&lock->wait_lock);
 580                 ret = -EDEADLK;
 581                 goto out_unlock_pi;
 582         }
 583
 584         /*
 585          * If we just follow the lock chain for deadlock detection, no
 586          * need to do all the requeue operations. To avoid a truckload
 587          * of conditionals around the various places below, just do the
 588          * minimum chain walk checks.
 589          */
 590         if (!requeue) {
 591                 /*
 592                  * No requeue[7] here. Just release @task [8]
 593                  */
 594                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 595                 put_task_struct(task);
 596
 597                 /*
 598                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
 599                  * If there is no owner of the lock, end of chain.
 600                  */
 601                 if (!rt_mutex_owner(lock)) {
 602                         raw_spin_unlock(&lock->wait_lock);
 603                         return 0;
 604                 }
 605
 606                 /* [10] Grab the next task, i.e. owner of @lock */
 607                 task = rt_mutex_owner(lock);
 608                 get_task_struct(task);
 609                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 610
 611                 /*
 612                  * No requeue [11] here. We just do deadlock detection.
 613                  *
 614                  * [12] Store whether owner is blocked
 615                  * itself. Decision is made after dropping the locks
 616                  */
 617                 next_lock = task_blocked_on_lock(task);
 618                 /*
 619                  * Get the top waiter for the next iteration
 620                  */
 621                 top_waiter = rt_mutex_top_waiter(lock);
 622
 623                 /* [13] Drop locks */
 624                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 625                 raw_spin_unlock(&lock->wait_lock);
 626
 627                 /* If owner is not blocked, end of chain. */
 628                 if (!next_lock)
 629                         goto out_put_task;
 630                 goto again;
 631         }
 632
 633         /*
 634          * Store the current top waiter before doing the requeue
 635          * operation on @lock. We need it for the boost/deboost
 636          * decision below.
 637          */
 638         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
 639
 640         /* [7] Requeue the waiter in the lock waiter tree. */
 641         rt_mutex_dequeue(lock, waiter);
 642         waiter->prio = task->prio;
 643         rt_mutex_enqueue(lock, waiter);
 644
 645         /* [8] Release the task */
 646         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 647         put_task_struct(task);
 648
 649         /*
 650          * [9] check_exit_conditions_3 protected by lock->wait_lock.
 651          *
 652          * We must abort the chain walk if there is no lock owner even
 653          * in the dead lock detection case, as we have nothing to
 654          * follow here. This is the end of the chain we are walking.
 655          */
 656         if (!rt_mutex_owner(lock)) {
 657                 /*
 658                  * If the requeue [7] above changed the top waiter,
 659                  * then we need to wake the new top waiter up to try
 660                  * to get the lock.
 661                  */
 662                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
 663                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 664                 raw_spin_unlock(&lock->wait_lock);
 665                 return 0;
 666         }
 667
 668         /* [10] Grab the next task, i.e. the owner of @lock */
 669         task = rt_mutex_owner(lock);
 670         get_task_struct(task);
 671         raw_spin_lock_irqsave(&task->pi_lock, flags);
 672
 673         /* [11] requeue the pi waiters if necessary */
 674         if (waiter == rt_mutex_top_waiter(lock)) {
 675                 /*
 676                  * The waiter became the new top (highest priority)
 677                  * waiter on the lock. Replace the previous top waiter
 678                  * in the owner tasks pi waiters tree with this waiter
 679                  * and adjust the priority of the owner.
 680                  */
 681                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
 682                 rt_mutex_enqueue_pi(task, waiter);
 683                 __rt_mutex_adjust_prio(task);
 684
 685         } else if (prerequeue_top_waiter == waiter) {
 686                 /*
 687                  * The waiter was the top waiter on the lock, but is
 688                  * no longer the top prority waiter. Replace waiter in
 689                  * the owner tasks pi waiters tree with the new top
 690                  * (highest priority) waiter and adjust the priority
 691                  * of the owner.
 692                  * The new top waiter is stored in @waiter so that
 693                  * @waiter == @top_waiter evaluates to true below and
 694                  * we continue to deboost the rest of the chain.
 695                  */
 696                 rt_mutex_dequeue_pi(task, waiter);
 697                 waiter = rt_mutex_top_waiter(lock);
 698                 rt_mutex_enqueue_pi(task, waiter);
 699                 __rt_mutex_adjust_prio(task);
 700         } else {
 701                 /*
 702                  * Nothing changed. No need to do any priority
 703                  * adjustment.
 704                  */
 705         }
 706
 707         /*
 708          * [12] check_exit_conditions_4() protected by task->pi_lock
 709          * and lock->wait_lock. The actual decisions are made after we
 710          * dropped the locks.
 711          *
 712          * Check whether the task which owns the current lock is pi
 713          * blocked itself. If yes we store a pointer to the lock for
 714          * the lock chain change detection above. After we dropped
 715          * task->pi_lock next_lock cannot be dereferenced anymore.
 716          */
 717         next_lock = task_blocked_on_lock(task);
 718         /*
 719          * Store the top waiter of @lock for the end of chain walk
 720          * decision below.
 721          */
 722         top_waiter = rt_mutex_top_waiter(lock);
 723
 724         /* [13] Drop the locks */
 725         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 726         raw_spin_unlock(&lock->wait_lock);
 727
 728         /*
 729          * Make the actual exit decisions [12], based on the stored
 730          * values.
 731          *
 732          * We reached the end of the lock chain. Stop right here. No
 733          * point to go back just to figure that out.
 734          */
 735         if (!next_lock)
 736                 goto out_put_task;
 737
 738         /*
 739          * If the current waiter is not the top waiter on the lock,
 740          * then we can stop the chain walk here if we are not in full
 741          * deadlock detection mode.
 742          */
 743         if (!detect_deadlock && waiter != top_waiter)
 744                 goto out_put_task;
 745
 746         goto again;
 747
 748  out_unlock_pi:
 749         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 750  out_put_task:
 751         put_task_struct(task);
 752
 753         return ret;
 754 }
 755
 756 /*
 757  * Try to take an rt-mutex
 758  *
 759  * Must be called with lock->wait_lock held.
 760  *
 761  * @lock:   The lock to be acquired.
 762  * @task:   The task which wants to acquire the lock
 763  * @waiter: The waiter that is queued to the lock's wait tree if the
 764  *          callsite called task_blocked_on_lock(), otherwise NULL
 765  */
 766 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 767                                 struct rt_mutex_waiter *waiter)
 768 {
 769         unsigned long flags;
 770
 771         /*
 772          * Before testing whether we can acquire @lock, we set the
 773          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
 774          * other tasks which try to modify @lock into the slow path
 775          * and they serialize on @lock->wait_lock.
 776          *
 777          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
 778          * as explained at the top of this file if and only if:
 779          *
 780          * - There is a lock owner. The caller must fixup the
 781          *   transient state if it does a trylock or leaves the lock
 782          *   function due to a signal or timeout.
 783          *
 784          * - @task acquires the lock and there are no other
 785          *   waiters. This is undone in rt_mutex_set_owner(@task) at
 786          *   the end of this function.
 787          */
 788         mark_rt_mutex_waiters(lock);
 789
 790         /*
 791          * If @lock has an owner, give up.
 792          */
 793         if (rt_mutex_owner(lock))
 794                 return 0;
 795
 796         /*
 797          * If @waiter != NULL, @task has already enqueued the waiter
 798          * into @lock waiter tree. If @waiter == NULL then this is a
 799          * trylock attempt.
 800          */
 801         if (waiter) {
 802                 /*
 803                  * If waiter is not the highest priority waiter of
 804                  * @lock, give up.
 805                  */
 806                 if (waiter != rt_mutex_top_waiter(lock))
 807                         return 0;
 808
 809                 /*
 810                  * We can acquire the lock. Remove the waiter from the
 811                  * lock waiters tree.
 812                  */
 813                 rt_mutex_dequeue(lock, waiter);
 814
 815         } else {
 816                 /*
 817                  * If the lock has waiters already we check whether @task is
 818                  * eligible to take over the lock.
 819                  *
 820                  * If there are no other waiters, @task can acquire
 821                  * the lock.  @task->pi_blocked_on is NULL, so it does
 822                  * not need to be dequeued.
 823                  */
 824                 if (rt_mutex_has_waiters(lock)) {
 825                         /*
 826                          * If @task->prio is greater than or equal to
 827                          * the top waiter priority (kernel view),
 828                          * @task lost.
 829                          */
 830                         if (task->prio >= rt_mutex_top_waiter(lock)->prio)
 831                                 return 0;
 832
 833                         /*
 834                          * The current top waiter stays enqueued. We
 835                          * don't have to change anything in the lock
 836                          * waiters order.
 837                          */
 838                 } else {
 839                         /*
 840                          * No waiters. Take the lock without the
 841                          * pi_lock dance.@task->pi_blocked_on is NULL
 842                          * and we have no waiters to enqueue in @task
 843                          * pi waiters tree.
 844                          */
 845                         goto takeit;
 846                 }
 847         }
 848
 849         /*
 850          * Clear @task->pi_blocked_on. Requires protection by
 851          * @task->pi_lock. Redundant operation for the @waiter == NULL
 852          * case, but conditionals are more expensive than a redundant
 853          * store.
 854          */
 855         raw_spin_lock_irqsave(&task->pi_lock, flags);
 856         task->pi_blocked_on = NULL;
 857         /*
 858          * Finish the lock acquisition. @task is the new owner. If
 859          * other waiters exist we have to insert the highest priority
 860          * waiter into @task->pi_waiters tree.
 861          */
 862         if (rt_mutex_has_waiters(lock))
 863                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
 864         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 865
 866 takeit:
 867         /* We got the lock. */
 868         debug_rt_mutex_lock(lock);
 869
 870         /*
 871          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
 872          * are still waiters or clears it.
 873          */
 874         rt_mutex_set_owner(lock, task);
 875
 876         rt_mutex_deadlock_account_lock(lock, task);
 877
 878         return 1;
 879 }
 880
 881 /*
 882  * Task blocks on lock.
 883  *
 884  * Prepare waiter and propagate pi chain
 885  *
 886  * This must be called with lock->wait_lock held.
 887  */
 888 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 889                                    struct rt_mutex_waiter *waiter,
 890                                    struct task_struct *task,
 891                                    enum rtmutex_chainwalk chwalk)
 892 {
 893         struct task_struct *owner = rt_mutex_owner(lock);
 894         struct rt_mutex_waiter *top_waiter = waiter;
 895         struct rt_mutex *next_lock;
 896         int chain_walk = 0, res;
 897         unsigned long flags;
 898
 899         /*
 900          * Early deadlock detection. We really don't want the task to
 901          * enqueue on itself just to untangle the mess later. It's not
 902          * only an optimization. We drop the locks, so another waiter
 903          * can come in before the chain walk detects the deadlock. So
 904          * the other will detect the deadlock and return -EDEADLOCK,
 905          * which is wrong, as the other waiter is not in a deadlock
 906          * situation.
 907          */
 908         if (owner == task)
 909                 return -EDEADLK;
 910
 911         raw_spin_lock_irqsave(&task->pi_lock, flags);
 912         __rt_mutex_adjust_prio(task);
 913         waiter->task = task;
 914         waiter->lock = lock;
 915         waiter->prio = task->prio;
 916
 917         /* Get the top priority waiter on the lock */
 918         if (rt_mutex_has_waiters(lock))
 919                 top_waiter = rt_mutex_top_waiter(lock);
 920         rt_mutex_enqueue(lock, waiter);
 921
 922         task->pi_blocked_on = waiter;
 923
 924         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 925
 926         if (!owner)
 927                 return 0;
 928
 929         raw_spin_lock_irqsave(&owner->pi_lock, flags);
 930         if (waiter == rt_mutex_top_waiter(lock)) {
 931                 rt_mutex_dequeue_pi(owner, top_waiter);
 932                 rt_mutex_enqueue_pi(owner, waiter);
 933
 934                 __rt_mutex_adjust_prio(owner);
 935                 if (owner->pi_blocked_on)
 936                         chain_walk = 1;
 937         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
 938                 chain_walk = 1;
 939         }
 940
 941         /* Store the lock on which owner is blocked or NULL */
 942         next_lock = task_blocked_on_lock(owner);
 943
 944         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 945         /*
 946          * Even if full deadlock detection is on, if the owner is not
 947          * blocked itself, we can avoid finding this out in the chain
 948          * walk.
 949          */
 950         if (!chain_walk || !next_lock)
 951                 return 0;
 952
 953         /*
 954          * The owner can't disappear while holding a lock,
 955          * so the owner struct is protected by wait_lock.
 956          * Gets dropped in rt_mutex_adjust_prio_chain()!
 957          */
 958         get_task_struct(owner);
 959
 960         raw_spin_unlock(&lock->wait_lock);
 961
 962         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
 963                                          next_lock, waiter, task);
 964
 965         raw_spin_lock(&lock->wait_lock);
 966
 967         return res;
 968 }
 969
 970 /*
 971  * Remove the top waiter from the current tasks pi waiter tree and
 972  * queue it up.
 973  *
 974  * Called with lock->wait_lock held.
 975  */
 976 static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
 977                                     struct rt_mutex *lock)
 978 {
 979         struct rt_mutex_waiter *waiter;
 980         unsigned long flags;
 981
 982         raw_spin_lock_irqsave(&current->pi_lock, flags);
 983
 984         waiter = rt_mutex_top_waiter(lock);
 985
 986         /*
 987          * Remove it from current->pi_waiters. We do not adjust a
 988          * possible priority boost right now. We execute wakeup in the
 989          * boosted mode and go back to normal after releasing
 990          * lock->wait_lock.
 991          */
 992         rt_mutex_dequeue_pi(current, waiter);
 993
 994         /*
 995          * As we are waking up the top waiter, and the waiter stays
 996          * queued on the lock until it gets the lock, this lock
 997          * obviously has waiters. Just set the bit here and this has
 998          * the added benefit of forcing all new tasks into the
 999          * slow path making sure no task of lower priority than
1000          * the top waiter can steal this lock.
1001          */
1002         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1003
1004         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
1005
1006         wake_q_add(wake_q, waiter->task);
1007 }
1008
1009 /*
1010  * Remove a waiter from a lock and give up
1011  *
1012  * Must be called with lock->wait_lock held and
1013  * have just failed to try_to_take_rt_mutex().
1014  */
1015 static void remove_waiter(struct rt_mutex *lock,
1016                           struct rt_mutex_waiter *waiter)
1017 {
1018         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1019         struct task_struct *owner = rt_mutex_owner(lock);
1020         struct rt_mutex *next_lock;
1021         unsigned long flags;
1022
1023         raw_spin_lock_irqsave(&current->pi_lock, flags);
1024         rt_mutex_dequeue(lock, waiter);
1025         current->pi_blocked_on = NULL;
1026         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
1027
1028         /*
1029          * Only update priority if the waiter was the highest priority
1030          * waiter of the lock and there is an owner to update.
1031          */
1032         if (!owner || !is_top_waiter)
1033                 return;
1034
1035         raw_spin_lock_irqsave(&owner->pi_lock, flags);
1036
1037         rt_mutex_dequeue_pi(owner, waiter);
1038
1039         if (rt_mutex_has_waiters(lock))
1040                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1041
1042         __rt_mutex_adjust_prio(owner);
1043
1044         /* Store the lock on which owner is blocked or NULL */
1045         next_lock = task_blocked_on_lock(owner);
1046
1047         raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
1048
1049         /*
1050          * Don't walk the chain, if the owner task is not blocked
1051          * itself.
1052          */
1053         if (!next_lock)
1054                 return;
1055
1056         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1057         get_task_struct(owner);
1058
1059         raw_spin_unlock(&lock->wait_lock);
1060
1061         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1062                                    next_lock, NULL, current);
1063
1064         raw_spin_lock(&lock->wait_lock);
1065 }
1066
1067 /*
1068  * Recheck the pi chain, in case we got a priority setting
1069  *
1070  * Called from sched_setscheduler
1071  */
1072 void rt_mutex_adjust_pi(struct task_struct *task)
1073 {
1074         struct rt_mutex_waiter *waiter;
1075         struct rt_mutex *next_lock;
1076         unsigned long flags;
1077
1078         raw_spin_lock_irqsave(&task->pi_lock, flags);
1079
1080         waiter = task->pi_blocked_on;
1081         if (!waiter || (waiter->prio == task->prio &&
1082                         !dl_prio(task->prio))) {
1083                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1084                 return;
1085         }
1086         next_lock = waiter->lock;
1087         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1088
1089         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1090         get_task_struct(task);
1091
1092         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1093                                    next_lock, NULL, task);
1094 }
1095
1096 /**
1097  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1098  * @lock:                the rt_mutex to take
1099  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1100  *                       or TASK_UNINTERRUPTIBLE)
1101  * @timeout:             the pre-initialized and started timer, or NULL for none
1102  * @waiter:              the pre-initialized rt_mutex_waiter
1103  *
1104  * lock->wait_lock must be held by the caller.
1105  */
1106 static int __sched
1107 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1108                     struct hrtimer_sleeper *timeout,
1109                     struct rt_mutex_waiter *waiter)
1110 {
1111         int ret = 0;
1112
1113         for (;;) {
1114                 /* Try to acquire the lock: */
1115                 if (try_to_take_rt_mutex(lock, current, waiter))
1116                         break;
1117
1118                 /*
1119                  * TASK_INTERRUPTIBLE checks for signals and
1120                  * timeout. Ignored otherwise.
1121                  */
1122                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
1123                         /* Signal pending? */
1124                         if (signal_pending(current))
1125                                 ret = -EINTR;
1126                         if (timeout && !timeout->task)
1127                                 ret = -ETIMEDOUT;
1128                         if (ret)
1129                                 break;
1130                 }
1131
1132                 raw_spin_unlock(&lock->wait_lock);
1133
1134                 debug_rt_mutex_print_deadlock(waiter);
1135
1136                 schedule();
1137
1138                 raw_spin_lock(&lock->wait_lock);
1139                 set_current_state(state);
1140         }
1141
1142         __set_current_state(TASK_RUNNING);
1143         return ret;
1144 }
1145
1146 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1147                                      struct rt_mutex_waiter *w)
1148 {
1149         /*
1150          * If the result is not -EDEADLOCK or the caller requested
1151          * deadlock detection, nothing to do here.
1152          */
1153         if (res != -EDEADLOCK || detect_deadlock)
1154                 return;
1155
1156         /*
1157          * Yell lowdly and stop the task right here.
1158          */
1159         rt_mutex_print_deadlock(w);
1160         while (1) {
1161                 set_current_state(TASK_INTERRUPTIBLE);
1162                 schedule();
1163         }
1164 }
1165
1166 /*
1167  * Slow path lock function:
1168  */
1169 static int __sched
1170 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1171                   struct hrtimer_sleeper *timeout,
1172                   enum rtmutex_chainwalk chwalk)
1173 {
1174         struct rt_mutex_waiter waiter;
1175         int ret = 0;
1176
1177         debug_rt_mutex_init_waiter(&waiter);
1178         RB_CLEAR_NODE(&waiter.pi_tree_entry);
1179         RB_CLEAR_NODE(&waiter.tree_entry);
1180
1181         raw_spin_lock(&lock->wait_lock);
1182
1183         /* Try to acquire the lock again: */
1184         if (try_to_take_rt_mutex(lock, current, NULL)) {
1185                 raw_spin_unlock(&lock->wait_lock);
1186                 return 0;
1187         }
1188
1189         set_current_state(state);
1190
1191         /* Setup the timer, when timeout != NULL */
1192         if (unlikely(timeout))
1193                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1194
1195         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1196
1197         if (likely(!ret))
1198                 /* sleep on the mutex */
1199                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1200
1201         if (unlikely(ret)) {
1202                 __set_current_state(TASK_RUNNING);
1203                 if (rt_mutex_has_waiters(lock))
1204                         remove_waiter(lock, &waiter);
1205                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1206         }
1207
1208         /*
1209          * try_to_take_rt_mutex() sets the waiter bit
1210          * unconditionally. We might have to fix that up.
1211          */
1212         fixup_rt_mutex_waiters(lock);
1213
1214         raw_spin_unlock(&lock->wait_lock);
1215
1216         /* Remove pending timer: */
1217         if (unlikely(timeout))
1218                 hrtimer_cancel(&timeout->timer);
1219
1220         debug_rt_mutex_free_waiter(&waiter);
1221
1222         return ret;
1223 }
1224
1225 /*
1226  * Slow path try-lock function:
1227  */
1228 static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1229 {
1230         int ret;
1231
1232         /*
1233          * If the lock already has an owner we fail to get the lock.
1234          * This can be done without taking the @lock->wait_lock as
1235          * it is only being read, and this is a trylock anyway.
1236          */
1237         if (rt_mutex_owner(lock))
1238                 return 0;
1239
1240         /*
1241          * The mutex has currently no owner. Lock the wait lock and
1242          * try to acquire the lock.
1243          */
1244         raw_spin_lock(&lock->wait_lock);
1245
1246         ret = try_to_take_rt_mutex(lock, current, NULL);
1247
1248         /*
1249          * try_to_take_rt_mutex() sets the lock waiters bit
1250          * unconditionally. Clean this up.
1251          */
1252         fixup_rt_mutex_waiters(lock);
1253
1254         raw_spin_unlock(&lock->wait_lock);
1255
1256         return ret;
1257 }
1258
1259 /*
1260  * Slow path to release a rt-mutex.
1261  * Return whether the current task needs to undo a potential priority boosting.
1262  */
1263 static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1264                                         struct wake_q_head *wake_q)
1265 {
1266         raw_spin_lock(&lock->wait_lock);
1267
1268         debug_rt_mutex_unlock(lock);
1269
1270         rt_mutex_deadlock_account_unlock(current);
1271
1272         /*
1273          * We must be careful here if the fast path is enabled. If we
1274          * have no waiters queued we cannot set owner to NULL here
1275          * because of:
1276          *
1277          * foo->lock->owner = NULL;
1278          *                      rtmutex_lock(foo->lock);   <- fast path
1279          *                      free = atomic_dec_and_test(foo->refcnt);
1280          *                      rtmutex_unlock(foo->lock); <- fast path
1281          *                      if (free)
1282          *                              kfree(foo);
1283          * raw_spin_unlock(foo->lock->wait_lock);
1284          *
1285          * So for the fastpath enabled kernel:
1286          *
1287          * Nothing can set the waiters bit as long as we hold
1288          * lock->wait_lock. So we do the following sequence:
1289          *
1290          *      owner = rt_mutex_owner(lock);
1291          *      clear_rt_mutex_waiters(lock);
1292          *      raw_spin_unlock(&lock->wait_lock);
1293          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1294          *              return;
1295          *      goto retry;
1296          *
1297          * The fastpath disabled variant is simple as all access to
1298          * lock->owner is serialized by lock->wait_lock:
1299          *
1300          *      lock->owner = NULL;
1301          *      raw_spin_unlock(&lock->wait_lock);
1302          */
1303         while (!rt_mutex_has_waiters(lock)) {
1304                 /* Drops lock->wait_lock ! */
1305                 if (unlock_rt_mutex_safe(lock) == true)
1306                         return false;
1307                 /* Relock the rtmutex and try again */
1308                 raw_spin_lock(&lock->wait_lock);
1309         }
1310
1311         /*
1312          * The wakeup next waiter path does not suffer from the above
1313          * race. See the comments there.
1314          *
1315          * Queue the next waiter for wakeup once we release the wait_lock.
1316          */
1317         mark_wakeup_next_waiter(wake_q, lock);
1318
1319         raw_spin_unlock(&lock->wait_lock);
1320
1321         /* check PI boosting */
1322         return true;
1323 }
1324
1325 /*
1326  * debug aware fast / slowpath lock,trylock,unlock
1327  *
1328  * The atomic acquire/release ops are compiled away, when either the
1329  * architecture does not support cmpxchg or when debugging is enabled.
1330  */
1331 static inline int
1332 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1333                   int (*slowfn)(struct rt_mutex *lock, int state,
1334                                 struct hrtimer_sleeper *timeout,
1335                                 enum rtmutex_chainwalk chwalk))
1336 {
1337         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current))) {
1338                 rt_mutex_deadlock_account_lock(lock, current);
1339                 return 0;
1340         } else
1341                 return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1342 }
1343
1344 static inline int
1345 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1346                         struct hrtimer_sleeper *timeout,
1347                         enum rtmutex_chainwalk chwalk,
1348                         int (*slowfn)(struct rt_mutex *lock, int state,
1349                                       struct hrtimer_sleeper *timeout,
1350                                       enum rtmutex_chainwalk chwalk))
1351 {
1352         if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1353             likely(rt_mutex_cmpxchg_acquire(lock, NULL, current))) {
1354                 rt_mutex_deadlock_account_lock(lock, current);
1355                 return 0;
1356         } else
1357                 return slowfn(lock, state, timeout, chwalk);
1358 }
1359
1360 static inline int
1361 rt_mutex_fasttrylock(struct rt_mutex *lock,
1362                      int (*slowfn)(struct rt_mutex *lock))
1363 {
1364         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current))) {
1365                 rt_mutex_deadlock_account_lock(lock, current);
1366                 return 1;
1367         }
1368         return slowfn(lock);
1369 }
1370
1371 static inline void
1372 rt_mutex_fastunlock(struct rt_mutex *lock,
1373                     bool (*slowfn)(struct rt_mutex *lock,
1374                                    struct wake_q_head *wqh))
1375 {
1376         WAKE_Q(wake_q);
1377
1378         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL))) {
1379                 rt_mutex_deadlock_account_unlock(current);
1380
1381         } else {
1382                 bool deboost = slowfn(lock, &wake_q);
1383
1384                 wake_up_q(&wake_q);
1385
1386                 /* Undo pi boosting if necessary: */
1387                 if (deboost)
1388                         rt_mutex_adjust_prio(current);
1389         }
1390 }
1391
1392 /**
1393  * rt_mutex_lock - lock a rt_mutex
1394  *
1395  * @lock: the rt_mutex to be locked
1396  */
1397 void __sched rt_mutex_lock(struct rt_mutex *lock)
1398 {
1399         might_sleep();
1400
1401         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1402 }
1403 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1404
1405 /**
1406  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1407  *
1408  * @lock:               the rt_mutex to be locked
1409  *
1410  * Returns:
1411  *  0           on success
1412  * -EINTR       when interrupted by a signal
1413  */
1414 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1415 {
1416         might_sleep();
1417
1418         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1419 }
1420 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1421
1422 /*
1423  * Futex variant with full deadlock detection.
1424  */
1425 int rt_mutex_timed_futex_lock(struct rt_mutex *lock,
1426                               struct hrtimer_sleeper *timeout)
1427 {
1428         might_sleep();
1429
1430         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1431                                        RT_MUTEX_FULL_CHAINWALK,
1432                                        rt_mutex_slowlock);
1433 }
1434
1435 /**
1436  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1437  *                      the timeout structure is provided
1438  *                      by the caller
1439  *
1440  * @lock:               the rt_mutex to be locked
1441  * @timeout:            timeout structure or NULL (no timeout)
1442  *
1443  * Returns:
1444  *  0           on success
1445  * -EINTR       when interrupted by a signal
1446  * -ETIMEDOUT   when the timeout expired
1447  */
1448 int
1449 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1450 {
1451         might_sleep();
1452
1453         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1454                                        RT_MUTEX_MIN_CHAINWALK,
1455                                        rt_mutex_slowlock);
1456 }
1457 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1458
1459 /**
1460  * rt_mutex_trylock - try to lock a rt_mutex
1461  *
1462  * @lock:       the rt_mutex to be locked
1463  *
1464  * This function can only be called in thread context. It's safe to
1465  * call it from atomic regions, but not from hard interrupt or soft
1466  * interrupt context.
1467  *
1468  * Returns 1 on success and 0 on contention
1469  */
1470 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1471 {
1472         if (WARN_ON(in_irq() || in_nmi() || in_serving_softirq()))
1473                 return 0;
1474
1475         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1476 }
1477 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1478
1479 /**
1480  * rt_mutex_unlock - unlock a rt_mutex
1481  *
1482  * @lock: the rt_mutex to be unlocked
1483  */
1484 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1485 {
1486         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1487 }
1488 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1489
1490 /**
1491  * rt_mutex_futex_unlock - Futex variant of rt_mutex_unlock
1492  * @lock: the rt_mutex to be unlocked
1493  *
1494  * Returns: true/false indicating whether priority adjustment is
1495  * required or not.
1496  */
1497 bool __sched rt_mutex_futex_unlock(struct rt_mutex *lock,
1498                                    struct wake_q_head *wqh)
1499 {
1500         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL))) {
1501                 rt_mutex_deadlock_account_unlock(current);
1502                 return false;
1503         }
1504         return rt_mutex_slowunlock(lock, wqh);
1505 }
1506
1507 /**
1508  * rt_mutex_destroy - mark a mutex unusable
1509  * @lock: the mutex to be destroyed
1510  *
1511  * This function marks the mutex uninitialized, and any subsequent
1512  * use of the mutex is forbidden. The mutex must not be locked when
1513  * this function is called.
1514  */
1515 void rt_mutex_destroy(struct rt_mutex *lock)
1516 {
1517         WARN_ON(rt_mutex_is_locked(lock));
1518 #ifdef CONFIG_DEBUG_RT_MUTEXES
1519         lock->magic = NULL;
1520 #endif
1521 }
1522
1523 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1524
1525 /**
1526  * __rt_mutex_init - initialize the rt lock
1527  *
1528  * @lock: the rt lock to be initialized
1529  *
1530  * Initialize the rt lock to unlocked state.
1531  *
1532  * Initializing of a locked rt lock is not allowed
1533  */
1534 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1535 {
1536         lock->owner = NULL;
1537         raw_spin_lock_init(&lock->wait_lock);
1538         lock->waiters = RB_ROOT;
1539         lock->waiters_leftmost = NULL;
1540
1541         debug_rt_mutex_init(lock, name);
1542 }
1543 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1544
1545 /**
1546  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1547  *                              proxy owner
1548  *
1549  * @lock:       the rt_mutex to be locked
1550  * @proxy_owner:the task to set as owner
1551  *
1552  * No locking. Caller has to do serializing itself
1553  * Special API call for PI-futex support
1554  */
1555 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1556                                 struct task_struct *proxy_owner)
1557 {
1558         __rt_mutex_init(lock, NULL);
1559         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1560         rt_mutex_set_owner(lock, proxy_owner);
1561         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1562 }
1563
1564 /**
1565  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1566  *
1567  * @lock:       the rt_mutex to be locked
1568  *
1569  * No locking. Caller has to do serializing itself
1570  * Special API call for PI-futex support
1571  */
1572 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1573                            struct task_struct *proxy_owner)
1574 {
1575         debug_rt_mutex_proxy_unlock(lock);
1576         rt_mutex_set_owner(lock, NULL);
1577         rt_mutex_deadlock_account_unlock(proxy_owner);
1578 }
1579
1580 /**
1581  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1582  * @lock:               the rt_mutex to take
1583  * @waiter:             the pre-initialized rt_mutex_waiter
1584  * @task:               the task to prepare
1585  *
1586  * Returns:
1587  *  0 - task blocked on lock
1588  *  1 - acquired the lock for task, caller should wake it up
1589  * <0 - error
1590  *
1591  * Special API call for FUTEX_REQUEUE_PI support.
1592  */
1593 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1594                               struct rt_mutex_waiter *waiter,
1595                               struct task_struct *task)
1596 {
1597         int ret;
1598
1599         raw_spin_lock(&lock->wait_lock);
1600
1601         if (try_to_take_rt_mutex(lock, task, NULL)) {
1602                 raw_spin_unlock(&lock->wait_lock);
1603                 return 1;
1604         }
1605
1606         /* We enforce deadlock detection for futexes */
1607         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1608                                       RT_MUTEX_FULL_CHAINWALK);
1609
1610         if (ret && !rt_mutex_owner(lock)) {
1611                 /*
1612                  * Reset the return value. We might have
1613                  * returned with -EDEADLK and the owner
1614                  * released the lock while we were walking the
1615                  * pi chain.  Let the waiter sort it out.
1616                  */
1617                 ret = 0;
1618         }
1619
1620         if (unlikely(ret))
1621                 remove_waiter(lock, waiter);
1622
1623         raw_spin_unlock(&lock->wait_lock);
1624
1625         debug_rt_mutex_print_deadlock(waiter);
1626
1627         return ret;
1628 }
1629
1630 /**
1631  * rt_mutex_next_owner - return the next owner of the lock
1632  *
1633  * @lock: the rt lock query
1634  *
1635  * Returns the next owner of the lock or NULL
1636  *
1637  * Caller has to serialize against other accessors to the lock
1638  * itself.
1639  *
1640  * Special API call for PI-futex support
1641  */
1642 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1643 {
1644         if (!rt_mutex_has_waiters(lock))
1645                 return NULL;
1646
1647         return rt_mutex_top_waiter(lock)->task;
1648 }
1649
1650 /**
1651  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1652  * @lock:               the rt_mutex we were woken on
1653  * @to:                 the timeout, null if none. hrtimer should already have
1654  *                      been started.
1655  * @waiter:             the pre-initialized rt_mutex_waiter
1656  *
1657  * Complete the lock acquisition started our behalf by another thread.
1658  *
1659  * Returns:
1660  *  0 - success
1661  * <0 - error, one of -EINTR, -ETIMEDOUT
1662  *
1663  * Special API call for PI-futex requeue support
1664  */
1665 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1666                                struct hrtimer_sleeper *to,
1667                                struct rt_mutex_waiter *waiter)
1668 {
1669         int ret;
1670
1671         raw_spin_lock(&lock->wait_lock);
1672
1673         set_current_state(TASK_INTERRUPTIBLE);
1674
1675         /* sleep on the mutex */
1676         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1677
1678         if (unlikely(ret))
1679                 remove_waiter(lock, waiter);
1680
1681         /*
1682          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1683          * have to fix that up.
1684          */
1685         fixup_rt_mutex_waiters(lock);
1686
1687         raw_spin_unlock(&lock->wait_lock);
1688
1689         return ret;
1690 }