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                                         unsigned long flags)
 104         __releases(lock->wait_lock)
 105 {
 106         struct task_struct *owner = rt_mutex_owner(lock);
 107
 108         clear_rt_mutex_waiters(lock);
 109         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 110         /*
 111          * If a new waiter comes in between the unlock and the cmpxchg
 112          * we have two situations:
 113          *
 114          * unlock(wait_lock);
 115          *                                      lock(wait_lock);
 116          * cmpxchg(p, owner, 0) == owner
 117          *                                      mark_rt_mutex_waiters(lock);
 118          *                                      acquire(lock);
 119          * or:
 120          *
 121          * unlock(wait_lock);
 122          *                                      lock(wait_lock);
 123          *                                      mark_rt_mutex_waiters(lock);
 124          *
 125          * cmpxchg(p, owner, 0) != owner
 126          *                                      enqueue_waiter();
 127          *                                      unlock(wait_lock);
 128          * lock(wait_lock);
 129          * wake waiter();
 130          * unlock(wait_lock);
 131          *                                      lock(wait_lock);
 132          *                                      acquire(lock);
 133          */
 134         return rt_mutex_cmpxchg_release(lock, owner, NULL);
 135 }
 136
 137 #else
 138 # define rt_mutex_cmpxchg_relaxed(l,c,n)        (0)
 139 # define rt_mutex_cmpxchg_acquire(l,c,n)        (0)
 140 # define rt_mutex_cmpxchg_release(l,c,n)        (0)
 141
 142 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 143 {
 144         lock->owner = (struct task_struct *)
 145                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 146 }
 147
 148 /*
 149  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 150  */
 151 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
 152                                         unsigned long flags)
 153         __releases(lock->wait_lock)
 154 {
 155         lock->owner = NULL;
 156         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 157         return true;
 158 }
 159 #endif
 160
 161 static inline int
 162 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
 163                      struct rt_mutex_waiter *right)
 164 {
 165         if (left->prio < right->prio)
 166                 return 1;
 167
 168         /*
 169          * If both waiters have dl_prio(), we check the deadlines of the
 170          * associated tasks.
 171          * If left waiter has a dl_prio(), and we didn't return 1 above,
 172          * then right waiter has a dl_prio() too.
 173          */
 174         if (dl_prio(left->prio))
 175                 return dl_time_before(left->task->dl.deadline,
 176                                       right->task->dl.deadline);
 177
 178         return 0;
 179 }
 180
 181 static void
 182 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 183 {
 184         struct rb_node **link = &lock->waiters.rb_node;
 185         struct rb_node *parent = NULL;
 186         struct rt_mutex_waiter *entry;
 187         int leftmost = 1;
 188
 189         while (*link) {
 190                 parent = *link;
 191                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
 192                 if (rt_mutex_waiter_less(waiter, entry)) {
 193                         link = &parent->rb_left;
 194                 } else {
 195                         link = &parent->rb_right;
 196                         leftmost = 0;
 197                 }
 198         }
 199
 200         if (leftmost)
 201                 lock->waiters_leftmost = &waiter->tree_entry;
 202
 203         rb_link_node(&waiter->tree_entry, parent, link);
 204         rb_insert_color(&waiter->tree_entry, &lock->waiters);
 205 }
 206
 207 static void
 208 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 209 {
 210         if (RB_EMPTY_NODE(&waiter->tree_entry))
 211                 return;
 212
 213         if (lock->waiters_leftmost == &waiter->tree_entry)
 214                 lock->waiters_leftmost = rb_next(&waiter->tree_entry);
 215
 216         rb_erase(&waiter->tree_entry, &lock->waiters);
 217         RB_CLEAR_NODE(&waiter->tree_entry);
 218 }
 219
 220 static void
 221 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 222 {
 223         struct rb_node **link = &task->pi_waiters.rb_node;
 224         struct rb_node *parent = NULL;
 225         struct rt_mutex_waiter *entry;
 226         int leftmost = 1;
 227
 228         while (*link) {
 229                 parent = *link;
 230                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
 231                 if (rt_mutex_waiter_less(waiter, entry)) {
 232                         link = &parent->rb_left;
 233                 } else {
 234                         link = &parent->rb_right;
 235                         leftmost = 0;
 236                 }
 237         }
 238
 239         if (leftmost)
 240                 task->pi_waiters_leftmost = &waiter->pi_tree_entry;
 241
 242         rb_link_node(&waiter->pi_tree_entry, parent, link);
 243         rb_insert_color(&waiter->pi_tree_entry, &task->pi_waiters);
 244 }
 245
 246 static void
 247 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 248 {
 249         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 250                 return;
 251
 252         if (task->pi_waiters_leftmost == &waiter->pi_tree_entry)
 253                 task->pi_waiters_leftmost = rb_next(&waiter->pi_tree_entry);
 254
 255         rb_erase(&waiter->pi_tree_entry, &task->pi_waiters);
 256         RB_CLEAR_NODE(&waiter->pi_tree_entry);
 257 }
 258
 259 /*
 260  * Calculate task priority from the waiter tree priority
 261  *
 262  * Return task->normal_prio when the waiter tree is empty or when
 263  * the waiter is not allowed to do priority boosting
 264  */
 265 int rt_mutex_getprio(struct task_struct *task)
 266 {
 267         if (likely(!task_has_pi_waiters(task)))
 268                 return task->normal_prio;
 269
 270         return min(task_top_pi_waiter(task)->prio,
 271                    task->normal_prio);
 272 }
 273
 274 struct task_struct *rt_mutex_get_top_task(struct task_struct *task)
 275 {
 276         if (likely(!task_has_pi_waiters(task)))
 277                 return NULL;
 278
 279         return task_top_pi_waiter(task)->task;
 280 }
 281
 282 /*
 283  * Called by sched_setscheduler() to get the priority which will be
 284  * effective after the change.
 285  */
 286 int rt_mutex_get_effective_prio(struct task_struct *task, int newprio)
 287 {
 288         if (!task_has_pi_waiters(task))
 289                 return newprio;
 290
 291         if (task_top_pi_waiter(task)->task->prio <= newprio)
 292                 return task_top_pi_waiter(task)->task->prio;
 293         return newprio;
 294 }
 295
 296 /*
 297  * Adjust the priority of a task, after its pi_waiters got modified.
 298  *
 299  * This can be both boosting and unboosting. task->pi_lock must be held.
 300  */
 301 static void __rt_mutex_adjust_prio(struct task_struct *task)
 302 {
 303         int prio = rt_mutex_getprio(task);
 304
 305         if (task->prio != prio || dl_prio(prio))
 306                 rt_mutex_setprio(task, prio);
 307 }
 308
 309 /*
 310  * Adjust task priority (undo boosting). Called from the exit path of
 311  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 312  *
 313  * (Note: We do this outside of the protection of lock->wait_lock to
 314  * allow the lock to be taken while or before we readjust the priority
 315  * of task. We do not use the spin_xx_mutex() variants here as we are
 316  * outside of the debug path.)
 317  */
 318 void rt_mutex_adjust_prio(struct task_struct *task)
 319 {
 320         unsigned long flags;
 321
 322         raw_spin_lock_irqsave(&task->pi_lock, flags);
 323         __rt_mutex_adjust_prio(task);
 324         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 325 }
 326
 327 /*
 328  * Deadlock detection is conditional:
 329  *
 330  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
 331  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
 332  *
 333  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
 334  * conducted independent of the detect argument.
 335  *
 336  * If the waiter argument is NULL this indicates the deboost path and
 337  * deadlock detection is disabled independent of the detect argument
 338  * and the config settings.
 339  */
 340 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
 341                                           enum rtmutex_chainwalk chwalk)
 342 {
 343         /*
 344          * This is just a wrapper function for the following call,
 345          * because debug_rt_mutex_detect_deadlock() smells like a magic
 346          * debug feature and I wanted to keep the cond function in the
 347          * main source file along with the comments instead of having
 348          * two of the same in the headers.
 349          */
 350         return debug_rt_mutex_detect_deadlock(waiter, chwalk);
 351 }
 352
 353 /*
 354  * Max number of times we'll walk the boosting chain:
 355  */
 356 int max_lock_depth = 1024;
 357
 358 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 359 {
 360         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 361 }
 362
 363 /*
 364  * Adjust the priority chain. Also used for deadlock detection.
 365  * Decreases task's usage by one - may thus free the task.
 366  *
 367  * @task:       the task owning the mutex (owner) for which a chain walk is
 368  *              probably needed
 369  * @chwalk:     do we have to carry out deadlock detection?
 370  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 371  *              things for a task that has just got its priority adjusted, and
 372  *              is waiting on a mutex)
 373  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 374  *              we dropped its pi_lock. Is never dereferenced, only used for
 375  *              comparison to detect lock chain changes.
 376  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 377  *              its priority to the mutex owner (can be NULL in the case
 378  *              depicted above or if the top waiter is gone away and we are
 379  *              actually deboosting the owner)
 380  * @top_task:   the current top waiter
 381  *
 382  * Returns 0 or -EDEADLK.
 383  *
 384  * Chain walk basics and protection scope
 385  *
 386  * [R] refcount on task
 387  * [P] task->pi_lock held
 388  * [L] rtmutex->wait_lock held
 389  *
 390  * Step Description                             Protected by
 391  *      function arguments:
 392  *      @task                                   [R]
 393  *      @orig_lock if != NULL                   @top_task is blocked on it
 394  *      @next_lock                              Unprotected. Cannot be
 395  *                                              dereferenced. Only used for
 396  *                                              comparison.
 397  *      @orig_waiter if != NULL                 @top_task is blocked on it
 398  *      @top_task                               current, or in case of proxy
 399  *                                              locking protected by calling
 400  *                                              code
 401  *      again:
 402  *        loop_sanity_check();
 403  *      retry:
 404  * [1]    lock(task->pi_lock);                  [R] acquire [P]
 405  * [2]    waiter = task->pi_blocked_on;         [P]
 406  * [3]    check_exit_conditions_1();            [P]
 407  * [4]    lock = waiter->lock;                  [P]
 408  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
 409  *          unlock(task->pi_lock);              release [P]
 410  *          goto retry;
 411  *        }
 412  * [6]    check_exit_conditions_2();            [P] + [L]
 413  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
 414  * [8]    unlock(task->pi_lock);                release [P]
 415  *        put_task_struct(task);                release [R]
 416  * [9]    check_exit_conditions_3();            [L]
 417  * [10]   task = owner(lock);                   [L]
 418  *        get_task_struct(task);                [L] acquire [R]
 419  *        lock(task->pi_lock);                  [L] acquire [P]
 420  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
 421  * [12]   check_exit_conditions_4();            [P] + [L]
 422  * [13]   unlock(task->pi_lock);                release [P]
 423  *        unlock(lock->wait_lock);              release [L]
 424  *        goto again;
 425  */
 426 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 427                                       enum rtmutex_chainwalk chwalk,
 428                                       struct rt_mutex *orig_lock,
 429                                       struct rt_mutex *next_lock,
 430                                       struct rt_mutex_waiter *orig_waiter,
 431                                       struct task_struct *top_task)
 432 {
 433         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 434         struct rt_mutex_waiter *prerequeue_top_waiter;
 435         int ret = 0, depth = 0;
 436         struct rt_mutex *lock;
 437         bool detect_deadlock;
 438         bool requeue = true;
 439
 440         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
 441
 442         /*
 443          * The (de)boosting is a step by step approach with a lot of
 444          * pitfalls. We want this to be preemptible and we want hold a
 445          * maximum of two locks per step. So we have to check
 446          * carefully whether things change under us.
 447          */
 448  again:
 449         /*
 450          * We limit the lock chain length for each invocation.
 451          */
 452         if (++depth > max_lock_depth) {
 453                 static int prev_max;
 454
 455                 /*
 456                  * Print this only once. If the admin changes the limit,
 457                  * print a new message when reaching the limit again.
 458                  */
 459                 if (prev_max != max_lock_depth) {
 460                         prev_max = max_lock_depth;
 461                         printk(KERN_WARNING "Maximum lock depth %d reached "
 462                                "task: %s (%d)\n", max_lock_depth,
 463                                top_task->comm, task_pid_nr(top_task));
 464                 }
 465                 put_task_struct(task);
 466
 467                 return -EDEADLK;
 468         }
 469
 470         /*
 471          * We are fully preemptible here and only hold the refcount on
 472          * @task. So everything can have changed under us since the
 473          * caller or our own code below (goto retry/again) dropped all
 474          * locks.
 475          */
 476  retry:
 477         /*
 478          * [1] Task cannot go away as we did a get_task() before !
 479          */
 480         raw_spin_lock_irq(&task->pi_lock);
 481
 482         /*
 483          * [2] Get the waiter on which @task is blocked on.
 484          */
 485         waiter = task->pi_blocked_on;
 486
 487         /*
 488          * [3] check_exit_conditions_1() protected by task->pi_lock.
 489          */
 490
 491         /*
 492          * Check whether the end of the boosting chain has been
 493          * reached or the state of the chain has changed while we
 494          * dropped the locks.
 495          */
 496         if (!waiter)
 497                 goto out_unlock_pi;
 498
 499         /*
 500          * Check the orig_waiter state. After we dropped the locks,
 501          * the previous owner of the lock might have released the lock.
 502          */
 503         if (orig_waiter && !rt_mutex_owner(orig_lock))
 504                 goto out_unlock_pi;
 505
 506         /*
 507          * We dropped all locks after taking a refcount on @task, so
 508          * the task might have moved on in the lock chain or even left
 509          * the chain completely and blocks now on an unrelated lock or
 510          * on @orig_lock.
 511          *
 512          * We stored the lock on which @task was blocked in @next_lock,
 513          * so we can detect the chain change.
 514          */
 515         if (next_lock != waiter->lock)
 516                 goto out_unlock_pi;
 517
 518         /*
 519          * Drop out, when the task has no waiters. Note,
 520          * top_waiter can be NULL, when we are in the deboosting
 521          * mode!
 522          */
 523         if (top_waiter) {
 524                 if (!task_has_pi_waiters(task))
 525                         goto out_unlock_pi;
 526                 /*
 527                  * If deadlock detection is off, we stop here if we
 528                  * are not the top pi waiter of the task. If deadlock
 529                  * detection is enabled we continue, but stop the
 530                  * requeueing in the chain walk.
 531                  */
 532                 if (top_waiter != task_top_pi_waiter(task)) {
 533                         if (!detect_deadlock)
 534                                 goto out_unlock_pi;
 535                         else
 536                                 requeue = false;
 537                 }
 538         }
 539
 540         /*
 541          * If the waiter priority is the same as the task priority
 542          * then there is no further priority adjustment necessary.  If
 543          * deadlock detection is off, we stop the chain walk. If its
 544          * enabled we continue, but stop the requeueing in the chain
 545          * walk.
 546          */
 547         if (waiter->prio == task->prio) {
 548                 if (!detect_deadlock)
 549                         goto out_unlock_pi;
 550                 else
 551                         requeue = false;
 552         }
 553
 554         /*
 555          * [4] Get the next lock
 556          */
 557         lock = waiter->lock;
 558         /*
 559          * [5] We need to trylock here as we are holding task->pi_lock,
 560          * which is the reverse lock order versus the other rtmutex
 561          * operations.
 562          */
 563         if (!raw_spin_trylock(&lock->wait_lock)) {
 564                 raw_spin_unlock_irq(&task->pi_lock);
 565                 cpu_relax();
 566                 goto retry;
 567         }
 568
 569         /*
 570          * [6] check_exit_conditions_2() protected by task->pi_lock and
 571          * lock->wait_lock.
 572          *
 573          * Deadlock detection. If the lock is the same as the original
 574          * lock which caused us to walk the lock chain or if the
 575          * current lock is owned by the task which initiated the chain
 576          * walk, we detected a deadlock.
 577          */
 578         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 579                 debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
 580                 raw_spin_unlock(&lock->wait_lock);
 581                 ret = -EDEADLK;
 582                 goto out_unlock_pi;
 583         }
 584
 585         /*
 586          * If we just follow the lock chain for deadlock detection, no
 587          * need to do all the requeue operations. To avoid a truckload
 588          * of conditionals around the various places below, just do the
 589          * minimum chain walk checks.
 590          */
 591         if (!requeue) {
 592                 /*
 593                  * No requeue[7] here. Just release @task [8]
 594                  */
 595                 raw_spin_unlock(&task->pi_lock);
 596                 put_task_struct(task);
 597
 598                 /*
 599                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
 600                  * If there is no owner of the lock, end of chain.
 601                  */
 602                 if (!rt_mutex_owner(lock)) {
 603                         raw_spin_unlock_irq(&lock->wait_lock);
 604                         return 0;
 605                 }
 606
 607                 /* [10] Grab the next task, i.e. owner of @lock */
 608                 task = rt_mutex_owner(lock);
 609                 get_task_struct(task);
 610                 raw_spin_lock(&task->pi_lock);
 611
 612                 /*
 613                  * No requeue [11] here. We just do deadlock detection.
 614                  *
 615                  * [12] Store whether owner is blocked
 616                  * itself. Decision is made after dropping the locks
 617                  */
 618                 next_lock = task_blocked_on_lock(task);
 619                 /*
 620                  * Get the top waiter for the next iteration
 621                  */
 622                 top_waiter = rt_mutex_top_waiter(lock);
 623
 624                 /* [13] Drop locks */
 625                 raw_spin_unlock(&task->pi_lock);
 626                 raw_spin_unlock_irq(&lock->wait_lock);
 627
 628                 /* If owner is not blocked, end of chain. */
 629                 if (!next_lock)
 630                         goto out_put_task;
 631                 goto again;
 632         }
 633
 634         /*
 635          * Store the current top waiter before doing the requeue
 636          * operation on @lock. We need it for the boost/deboost
 637          * decision below.
 638          */
 639         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
 640
 641         /* [7] Requeue the waiter in the lock waiter tree. */
 642         rt_mutex_dequeue(lock, waiter);
 643         waiter->prio = task->prio;
 644         rt_mutex_enqueue(lock, waiter);
 645
 646         /* [8] Release the task */
 647         raw_spin_unlock(&task->pi_lock);
 648         put_task_struct(task);
 649
 650         /*
 651          * [9] check_exit_conditions_3 protected by lock->wait_lock.
 652          *
 653          * We must abort the chain walk if there is no lock owner even
 654          * in the dead lock detection case, as we have nothing to
 655          * follow here. This is the end of the chain we are walking.
 656          */
 657         if (!rt_mutex_owner(lock)) {
 658                 /*
 659                  * If the requeue [7] above changed the top waiter,
 660                  * then we need to wake the new top waiter up to try
 661                  * to get the lock.
 662                  */
 663                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
 664                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 665                 raw_spin_unlock_irq(&lock->wait_lock);
 666                 return 0;
 667         }
 668
 669         /* [10] Grab the next task, i.e. the owner of @lock */
 670         task = rt_mutex_owner(lock);
 671         get_task_struct(task);
 672         raw_spin_lock(&task->pi_lock);
 673
 674         /* [11] requeue the pi waiters if necessary */
 675         if (waiter == rt_mutex_top_waiter(lock)) {
 676                 /*
 677                  * The waiter became the new top (highest priority)
 678                  * waiter on the lock. Replace the previous top waiter
 679                  * in the owner tasks pi waiters tree with this waiter
 680                  * and adjust the priority of the owner.
 681                  */
 682                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
 683                 rt_mutex_enqueue_pi(task, waiter);
 684                 __rt_mutex_adjust_prio(task);
 685
 686         } else if (prerequeue_top_waiter == waiter) {
 687                 /*
 688                  * The waiter was the top waiter on the lock, but is
 689                  * no longer the top prority waiter. Replace waiter in
 690                  * the owner tasks pi waiters tree with the new top
 691                  * (highest priority) waiter and adjust the priority
 692                  * of the owner.
 693                  * The new top waiter is stored in @waiter so that
 694                  * @waiter == @top_waiter evaluates to true below and
 695                  * we continue to deboost the rest of the chain.
 696                  */
 697                 rt_mutex_dequeue_pi(task, waiter);
 698                 waiter = rt_mutex_top_waiter(lock);
 699                 rt_mutex_enqueue_pi(task, waiter);
 700                 __rt_mutex_adjust_prio(task);
 701         } else {
 702                 /*
 703                  * Nothing changed. No need to do any priority
 704                  * adjustment.
 705                  */
 706         }
 707
 708         /*
 709          * [12] check_exit_conditions_4() protected by task->pi_lock
 710          * and lock->wait_lock. The actual decisions are made after we
 711          * dropped the locks.
 712          *
 713          * Check whether the task which owns the current lock is pi
 714          * blocked itself. If yes we store a pointer to the lock for
 715          * the lock chain change detection above. After we dropped
 716          * task->pi_lock next_lock cannot be dereferenced anymore.
 717          */
 718         next_lock = task_blocked_on_lock(task);
 719         /*
 720          * Store the top waiter of @lock for the end of chain walk
 721          * decision below.
 722          */
 723         top_waiter = rt_mutex_top_waiter(lock);
 724
 725         /* [13] Drop the locks */
 726         raw_spin_unlock(&task->pi_lock);
 727         raw_spin_unlock_irq(&lock->wait_lock);
 728
 729         /*
 730          * Make the actual exit decisions [12], based on the stored
 731          * values.
 732          *
 733          * We reached the end of the lock chain. Stop right here. No
 734          * point to go back just to figure that out.
 735          */
 736         if (!next_lock)
 737                 goto out_put_task;
 738
 739         /*
 740          * If the current waiter is not the top waiter on the lock,
 741          * then we can stop the chain walk here if we are not in full
 742          * deadlock detection mode.
 743          */
 744         if (!detect_deadlock && waiter != top_waiter)
 745                 goto out_put_task;
 746
 747         goto again;
 748
 749  out_unlock_pi:
 750         raw_spin_unlock_irq(&task->pi_lock);
 751  out_put_task:
 752         put_task_struct(task);
 753
 754         return ret;
 755 }
 756
 757 /*
 758  * Try to take an rt-mutex
 759  *
 760  * Must be called with lock->wait_lock held and interrupts disabled
 761  *
 762  * @lock:   The lock to be acquired.
 763  * @task:   The task which wants to acquire the lock
 764  * @waiter: The waiter that is queued to the lock's wait tree if the
 765  *          callsite called task_blocked_on_lock(), otherwise NULL
 766  */
 767 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 768                                 struct rt_mutex_waiter *waiter)
 769 {
 770         /*
 771          * Before testing whether we can acquire @lock, we set the
 772          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
 773          * other tasks which try to modify @lock into the slow path
 774          * and they serialize on @lock->wait_lock.
 775          *
 776          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
 777          * as explained at the top of this file if and only if:
 778          *
 779          * - There is a lock owner. The caller must fixup the
 780          *   transient state if it does a trylock or leaves the lock
 781          *   function due to a signal or timeout.
 782          *
 783          * - @task acquires the lock and there are no other
 784          *   waiters. This is undone in rt_mutex_set_owner(@task) at
 785          *   the end of this function.
 786          */
 787         mark_rt_mutex_waiters(lock);
 788
 789         /*
 790          * If @lock has an owner, give up.
 791          */
 792         if (rt_mutex_owner(lock))
 793                 return 0;
 794
 795         /*
 796          * If @waiter != NULL, @task has already enqueued the waiter
 797          * into @lock waiter tree. If @waiter == NULL then this is a
 798          * trylock attempt.
 799          */
 800         if (waiter) {
 801                 /*
 802                  * If waiter is not the highest priority waiter of
 803                  * @lock, give up.
 804                  */
 805                 if (waiter != rt_mutex_top_waiter(lock))
 806                         return 0;
 807
 808                 /*
 809                  * We can acquire the lock. Remove the waiter from the
 810                  * lock waiters tree.
 811                  */
 812                 rt_mutex_dequeue(lock, waiter);
 813
 814         } else {
 815                 /*
 816                  * If the lock has waiters already we check whether @task is
 817                  * eligible to take over the lock.
 818                  *
 819                  * If there are no other waiters, @task can acquire
 820                  * the lock.  @task->pi_blocked_on is NULL, so it does
 821                  * not need to be dequeued.
 822                  */
 823                 if (rt_mutex_has_waiters(lock)) {
 824                         /*
 825                          * If @task->prio is greater than or equal to
 826                          * the top waiter priority (kernel view),
 827                          * @task lost.
 828                          */
 829                         if (task->prio >= rt_mutex_top_waiter(lock)->prio)
 830                                 return 0;
 831
 832                         /*
 833                          * The current top waiter stays enqueued. We
 834                          * don't have to change anything in the lock
 835                          * waiters order.
 836                          */
 837                 } else {
 838                         /*
 839                          * No waiters. Take the lock without the
 840                          * pi_lock dance.@task->pi_blocked_on is NULL
 841                          * and we have no waiters to enqueue in @task
 842                          * pi waiters tree.
 843                          */
 844                         goto takeit;
 845                 }
 846         }
 847
 848         /*
 849          * Clear @task->pi_blocked_on. Requires protection by
 850          * @task->pi_lock. Redundant operation for the @waiter == NULL
 851          * case, but conditionals are more expensive than a redundant
 852          * store.
 853          */
 854         raw_spin_lock(&task->pi_lock);
 855         task->pi_blocked_on = NULL;
 856         /*
 857          * Finish the lock acquisition. @task is the new owner. If
 858          * other waiters exist we have to insert the highest priority
 859          * waiter into @task->pi_waiters tree.
 860          */
 861         if (rt_mutex_has_waiters(lock))
 862                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
 863         raw_spin_unlock(&task->pi_lock);
 864
 865 takeit:
 866         /* We got the lock. */
 867         debug_rt_mutex_lock(lock);
 868
 869         /*
 870          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
 871          * are still waiters or clears it.
 872          */
 873         rt_mutex_set_owner(lock, task);
 874
 875         rt_mutex_deadlock_account_lock(lock, task);
 876
 877         return 1;
 878 }
 879
 880 /*
 881  * Task blocks on lock.
 882  *
 883  * Prepare waiter and propagate pi chain
 884  *
 885  * This must be called with lock->wait_lock held and interrupts disabled
 886  */
 887 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 888                                    struct rt_mutex_waiter *waiter,
 889                                    struct task_struct *task,
 890                                    enum rtmutex_chainwalk chwalk)
 891 {
 892         struct task_struct *owner = rt_mutex_owner(lock);
 893         struct rt_mutex_waiter *top_waiter = waiter;
 894         struct rt_mutex *next_lock;
 895         int chain_walk = 0, res;
 896
 897         /*
 898          * Early deadlock detection. We really don't want the task to
 899          * enqueue on itself just to untangle the mess later. It's not
 900          * only an optimization. We drop the locks, so another waiter
 901          * can come in before the chain walk detects the deadlock. So
 902          * the other will detect the deadlock and return -EDEADLOCK,
 903          * which is wrong, as the other waiter is not in a deadlock
 904          * situation.
 905          */
 906         if (owner == task)
 907                 return -EDEADLK;
 908
 909         raw_spin_lock(&task->pi_lock);
 910         __rt_mutex_adjust_prio(task);
 911         waiter->task = task;
 912         waiter->lock = lock;
 913         waiter->prio = task->prio;
 914
 915         /* Get the top priority waiter on the lock */
 916         if (rt_mutex_has_waiters(lock))
 917                 top_waiter = rt_mutex_top_waiter(lock);
 918         rt_mutex_enqueue(lock, waiter);
 919
 920         task->pi_blocked_on = waiter;
 921
 922         raw_spin_unlock(&task->pi_lock);
 923
 924         if (!owner)
 925                 return 0;
 926
 927         raw_spin_lock(&owner->pi_lock);
 928         if (waiter == rt_mutex_top_waiter(lock)) {
 929                 rt_mutex_dequeue_pi(owner, top_waiter);
 930                 rt_mutex_enqueue_pi(owner, waiter);
 931
 932                 __rt_mutex_adjust_prio(owner);
 933                 if (owner->pi_blocked_on)
 934                         chain_walk = 1;
 935         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
 936                 chain_walk = 1;
 937         }
 938
 939         /* Store the lock on which owner is blocked or NULL */
 940         next_lock = task_blocked_on_lock(owner);
 941
 942         raw_spin_unlock(&owner->pi_lock);
 943         /*
 944          * Even if full deadlock detection is on, if the owner is not
 945          * blocked itself, we can avoid finding this out in the chain
 946          * walk.
 947          */
 948         if (!chain_walk || !next_lock)
 949                 return 0;
 950
 951         /*
 952          * The owner can't disappear while holding a lock,
 953          * so the owner struct is protected by wait_lock.
 954          * Gets dropped in rt_mutex_adjust_prio_chain()!
 955          */
 956         get_task_struct(owner);
 957
 958         raw_spin_unlock_irq(&lock->wait_lock);
 959
 960         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
 961                                          next_lock, waiter, task);
 962
 963         raw_spin_lock_irq(&lock->wait_lock);
 964
 965         return res;
 966 }
 967
 968 /*
 969  * Remove the top waiter from the current tasks pi waiter tree and
 970  * queue it up.
 971  *
 972  * Called with lock->wait_lock held and interrupts disabled.
 973  */
 974 static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
 975                                     struct rt_mutex *lock)
 976 {
 977         struct rt_mutex_waiter *waiter;
 978
 979         raw_spin_lock(&current->pi_lock);
 980
 981         waiter = rt_mutex_top_waiter(lock);
 982
 983         /*
 984          * Remove it from current->pi_waiters. We do not adjust a
 985          * possible priority boost right now. We execute wakeup in the
 986          * boosted mode and go back to normal after releasing
 987          * lock->wait_lock.
 988          */
 989         rt_mutex_dequeue_pi(current, waiter);
 990
 991         /*
 992          * As we are waking up the top waiter, and the waiter stays
 993          * queued on the lock until it gets the lock, this lock
 994          * obviously has waiters. Just set the bit here and this has
 995          * the added benefit of forcing all new tasks into the
 996          * slow path making sure no task of lower priority than
 997          * the top waiter can steal this lock.
 998          */
 999         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1000
1001         raw_spin_unlock(&current->pi_lock);
1002
1003         wake_q_add(wake_q, waiter->task);
1004 }
1005
1006 /*
1007  * Remove a waiter from a lock and give up
1008  *
1009  * Must be called with lock->wait_lock held and interrupts disabled. I must
1010  * have just failed to try_to_take_rt_mutex().
1011  */
1012 static void remove_waiter(struct rt_mutex *lock,
1013                           struct rt_mutex_waiter *waiter)
1014 {
1015         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1016         struct task_struct *owner = rt_mutex_owner(lock);
1017         struct rt_mutex *next_lock;
1018
1019         raw_spin_lock(&current->pi_lock);
1020         rt_mutex_dequeue(lock, waiter);
1021         current->pi_blocked_on = NULL;
1022         raw_spin_unlock(&current->pi_lock);
1023
1024         /*
1025          * Only update priority if the waiter was the highest priority
1026          * waiter of the lock and there is an owner to update.
1027          */
1028         if (!owner || !is_top_waiter)
1029                 return;
1030
1031         raw_spin_lock(&owner->pi_lock);
1032
1033         rt_mutex_dequeue_pi(owner, waiter);
1034
1035         if (rt_mutex_has_waiters(lock))
1036                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1037
1038         __rt_mutex_adjust_prio(owner);
1039
1040         /* Store the lock on which owner is blocked or NULL */
1041         next_lock = task_blocked_on_lock(owner);
1042
1043         raw_spin_unlock(&owner->pi_lock);
1044
1045         /*
1046          * Don't walk the chain, if the owner task is not blocked
1047          * itself.
1048          */
1049         if (!next_lock)
1050                 return;
1051
1052         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1053         get_task_struct(owner);
1054
1055         raw_spin_unlock_irq(&lock->wait_lock);
1056
1057         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1058                                    next_lock, NULL, current);
1059
1060         raw_spin_lock_irq(&lock->wait_lock);
1061 }
1062
1063 /*
1064  * Recheck the pi chain, in case we got a priority setting
1065  *
1066  * Called from sched_setscheduler
1067  */
1068 void rt_mutex_adjust_pi(struct task_struct *task)
1069 {
1070         struct rt_mutex_waiter *waiter;
1071         struct rt_mutex *next_lock;
1072         unsigned long flags;
1073
1074         raw_spin_lock_irqsave(&task->pi_lock, flags);
1075
1076         waiter = task->pi_blocked_on;
1077         if (!waiter || (waiter->prio == task->prio &&
1078                         !dl_prio(task->prio))) {
1079                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1080                 return;
1081         }
1082         next_lock = waiter->lock;
1083         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1084
1085         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1086         get_task_struct(task);
1087
1088         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1089                                    next_lock, NULL, task);
1090 }
1091
1092 /**
1093  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1094  * @lock:                the rt_mutex to take
1095  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1096  *                       or TASK_UNINTERRUPTIBLE)
1097  * @timeout:             the pre-initialized and started timer, or NULL for none
1098  * @waiter:              the pre-initialized rt_mutex_waiter
1099  *
1100  * Must be called with lock->wait_lock held and interrupts disabled
1101  */
1102 static int __sched
1103 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1104                     struct hrtimer_sleeper *timeout,
1105                     struct rt_mutex_waiter *waiter)
1106 {
1107         int ret = 0;
1108
1109         for (;;) {
1110                 /* Try to acquire the lock: */
1111                 if (try_to_take_rt_mutex(lock, current, waiter))
1112                         break;
1113
1114                 /*
1115                  * TASK_INTERRUPTIBLE checks for signals and
1116                  * timeout. Ignored otherwise.
1117                  */
1118                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
1119                         /* Signal pending? */
1120                         if (signal_pending(current))
1121                                 ret = -EINTR;
1122                         if (timeout && !timeout->task)
1123                                 ret = -ETIMEDOUT;
1124                         if (ret)
1125                                 break;
1126                 }
1127
1128                 raw_spin_unlock_irq(&lock->wait_lock);
1129
1130                 debug_rt_mutex_print_deadlock(waiter);
1131
1132                 schedule();
1133
1134                 raw_spin_lock_irq(&lock->wait_lock);
1135                 set_current_state(state);
1136         }
1137
1138         __set_current_state(TASK_RUNNING);
1139         return ret;
1140 }
1141
1142 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1143                                      struct rt_mutex_waiter *w)
1144 {
1145         /*
1146          * If the result is not -EDEADLOCK or the caller requested
1147          * deadlock detection, nothing to do here.
1148          */
1149         if (res != -EDEADLOCK || detect_deadlock)
1150                 return;
1151
1152         /*
1153          * Yell lowdly and stop the task right here.
1154          */
1155         rt_mutex_print_deadlock(w);
1156         while (1) {
1157                 set_current_state(TASK_INTERRUPTIBLE);
1158                 schedule();
1159         }
1160 }
1161
1162 /*
1163  * Slow path lock function:
1164  */
1165 static int __sched
1166 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1167                   struct hrtimer_sleeper *timeout,
1168                   enum rtmutex_chainwalk chwalk)
1169 {
1170         struct rt_mutex_waiter waiter;
1171         unsigned long flags;
1172         int ret = 0;
1173
1174         debug_rt_mutex_init_waiter(&waiter);
1175         RB_CLEAR_NODE(&waiter.pi_tree_entry);
1176         RB_CLEAR_NODE(&waiter.tree_entry);
1177
1178         /*
1179          * Technically we could use raw_spin_[un]lock_irq() here, but this can
1180          * be called in early boot if the cmpxchg() fast path is disabled
1181          * (debug, no architecture support). In this case we will acquire the
1182          * rtmutex with lock->wait_lock held. But we cannot unconditionally
1183          * enable interrupts in that early boot case. So we need to use the
1184          * irqsave/restore variants.
1185          */
1186         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1187
1188         /* Try to acquire the lock again: */
1189         if (try_to_take_rt_mutex(lock, current, NULL)) {
1190                 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1191                 return 0;
1192         }
1193
1194         set_current_state(state);
1195
1196         /* Setup the timer, when timeout != NULL */
1197         if (unlikely(timeout))
1198                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1199
1200         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1201
1202         if (likely(!ret))
1203                 /* sleep on the mutex */
1204                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1205
1206         if (unlikely(ret)) {
1207                 __set_current_state(TASK_RUNNING);
1208                 if (rt_mutex_has_waiters(lock))
1209                         remove_waiter(lock, &waiter);
1210                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1211         }
1212
1213         /*
1214          * try_to_take_rt_mutex() sets the waiter bit
1215          * unconditionally. We might have to fix that up.
1216          */
1217         fixup_rt_mutex_waiters(lock);
1218
1219         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1220
1221         /* Remove pending timer: */
1222         if (unlikely(timeout))
1223                 hrtimer_cancel(&timeout->timer);
1224
1225         debug_rt_mutex_free_waiter(&waiter);
1226
1227         return ret;
1228 }
1229
1230 /*
1231  * Slow path try-lock function:
1232  */
1233 static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1234 {
1235         unsigned long flags;
1236         int ret;
1237
1238         /*
1239          * If the lock already has an owner we fail to get the lock.
1240          * This can be done without taking the @lock->wait_lock as
1241          * it is only being read, and this is a trylock anyway.
1242          */
1243         if (rt_mutex_owner(lock))
1244                 return 0;
1245
1246         /*
1247          * The mutex has currently no owner. Lock the wait lock and try to
1248          * acquire the lock. We use irqsave here to support early boot calls.
1249          */
1250         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1251
1252         ret = try_to_take_rt_mutex(lock, current, NULL);
1253
1254         /*
1255          * try_to_take_rt_mutex() sets the lock waiters bit
1256          * unconditionally. Clean this up.
1257          */
1258         fixup_rt_mutex_waiters(lock);
1259
1260         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1261
1262         return ret;
1263 }
1264
1265 /*
1266  * Slow path to release a rt-mutex.
1267  * Return whether the current task needs to undo a potential priority boosting.
1268  */
1269 static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1270                                         struct wake_q_head *wake_q)
1271 {
1272         unsigned long flags;
1273
1274         /* irqsave required to support early boot calls */
1275         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1276
1277         debug_rt_mutex_unlock(lock);
1278
1279         rt_mutex_deadlock_account_unlock(current);
1280
1281         /*
1282          * We must be careful here if the fast path is enabled. If we
1283          * have no waiters queued we cannot set owner to NULL here
1284          * because of:
1285          *
1286          * foo->lock->owner = NULL;
1287          *                      rtmutex_lock(foo->lock);   <- fast path
1288          *                      free = atomic_dec_and_test(foo->refcnt);
1289          *                      rtmutex_unlock(foo->lock); <- fast path
1290          *                      if (free)
1291          *                              kfree(foo);
1292          * raw_spin_unlock(foo->lock->wait_lock);
1293          *
1294          * So for the fastpath enabled kernel:
1295          *
1296          * Nothing can set the waiters bit as long as we hold
1297          * lock->wait_lock. So we do the following sequence:
1298          *
1299          *      owner = rt_mutex_owner(lock);
1300          *      clear_rt_mutex_waiters(lock);
1301          *      raw_spin_unlock(&lock->wait_lock);
1302          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1303          *              return;
1304          *      goto retry;
1305          *
1306          * The fastpath disabled variant is simple as all access to
1307          * lock->owner is serialized by lock->wait_lock:
1308          *
1309          *      lock->owner = NULL;
1310          *      raw_spin_unlock(&lock->wait_lock);
1311          */
1312         while (!rt_mutex_has_waiters(lock)) {
1313                 /* Drops lock->wait_lock ! */
1314                 if (unlock_rt_mutex_safe(lock, flags) == true)
1315                         return false;
1316                 /* Relock the rtmutex and try again */
1317                 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1318         }
1319
1320         /*
1321          * The wakeup next waiter path does not suffer from the above
1322          * race. See the comments there.
1323          *
1324          * Queue the next waiter for wakeup once we release the wait_lock.
1325          */
1326         mark_wakeup_next_waiter(wake_q, lock);
1327
1328         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1329
1330         /* check PI boosting */
1331         return true;
1332 }
1333
1334 /*
1335  * debug aware fast / slowpath lock,trylock,unlock
1336  *
1337  * The atomic acquire/release ops are compiled away, when either the
1338  * architecture does not support cmpxchg or when debugging is enabled.
1339  */
1340 static inline int
1341 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1342                   int (*slowfn)(struct rt_mutex *lock, int state,
1343                                 struct hrtimer_sleeper *timeout,
1344                                 enum rtmutex_chainwalk chwalk))
1345 {
1346         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current))) {
1347                 rt_mutex_deadlock_account_lock(lock, current);
1348                 return 0;
1349         } else
1350                 return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1351 }
1352
1353 static inline int
1354 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1355                         struct hrtimer_sleeper *timeout,
1356                         enum rtmutex_chainwalk chwalk,
1357                         int (*slowfn)(struct rt_mutex *lock, int state,
1358                                       struct hrtimer_sleeper *timeout,
1359                                       enum rtmutex_chainwalk chwalk))
1360 {
1361         if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1362             likely(rt_mutex_cmpxchg_acquire(lock, NULL, current))) {
1363                 rt_mutex_deadlock_account_lock(lock, current);
1364                 return 0;
1365         } else
1366                 return slowfn(lock, state, timeout, chwalk);
1367 }
1368
1369 static inline int
1370 rt_mutex_fasttrylock(struct rt_mutex *lock,
1371                      int (*slowfn)(struct rt_mutex *lock))
1372 {
1373         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current))) {
1374                 rt_mutex_deadlock_account_lock(lock, current);
1375                 return 1;
1376         }
1377         return slowfn(lock);
1378 }
1379
1380 static inline void
1381 rt_mutex_fastunlock(struct rt_mutex *lock,
1382                     bool (*slowfn)(struct rt_mutex *lock,
1383                                    struct wake_q_head *wqh))
1384 {
1385         WAKE_Q(wake_q);
1386
1387         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL))) {
1388                 rt_mutex_deadlock_account_unlock(current);
1389
1390         } else {
1391                 bool deboost = slowfn(lock, &wake_q);
1392
1393                 wake_up_q(&wake_q);
1394
1395                 /* Undo pi boosting if necessary: */
1396                 if (deboost)
1397                         rt_mutex_adjust_prio(current);
1398         }
1399 }
1400
1401 /**
1402  * rt_mutex_lock - lock a rt_mutex
1403  *
1404  * @lock: the rt_mutex to be locked
1405  */
1406 void __sched rt_mutex_lock(struct rt_mutex *lock)
1407 {
1408         might_sleep();
1409
1410         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1411 }
1412 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1413
1414 /**
1415  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1416  *
1417  * @lock:               the rt_mutex to be locked
1418  *
1419  * Returns:
1420  *  0           on success
1421  * -EINTR       when interrupted by a signal
1422  */
1423 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1424 {
1425         might_sleep();
1426
1427         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1428 }
1429 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1430
1431 /*
1432  * Futex variant with full deadlock detection.
1433  */
1434 int rt_mutex_timed_futex_lock(struct rt_mutex *lock,
1435                               struct hrtimer_sleeper *timeout)
1436 {
1437         might_sleep();
1438
1439         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1440                                        RT_MUTEX_FULL_CHAINWALK,
1441                                        rt_mutex_slowlock);
1442 }
1443
1444 /**
1445  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1446  *                      the timeout structure is provided
1447  *                      by the caller
1448  *
1449  * @lock:               the rt_mutex to be locked
1450  * @timeout:            timeout structure or NULL (no timeout)
1451  *
1452  * Returns:
1453  *  0           on success
1454  * -EINTR       when interrupted by a signal
1455  * -ETIMEDOUT   when the timeout expired
1456  */
1457 int
1458 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1459 {
1460         might_sleep();
1461
1462         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1463                                        RT_MUTEX_MIN_CHAINWALK,
1464                                        rt_mutex_slowlock);
1465 }
1466 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1467
1468 /**
1469  * rt_mutex_trylock - try to lock a rt_mutex
1470  *
1471  * @lock:       the rt_mutex to be locked
1472  *
1473  * This function can only be called in thread context. It's safe to
1474  * call it from atomic regions, but not from hard interrupt or soft
1475  * interrupt context.
1476  *
1477  * Returns 1 on success and 0 on contention
1478  */
1479 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1480 {
1481         if (WARN_ON_ONCE(in_irq() || in_nmi() || in_serving_softirq()))
1482                 return 0;
1483
1484         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1485 }
1486 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1487
1488 /**
1489  * rt_mutex_unlock - unlock a rt_mutex
1490  *
1491  * @lock: the rt_mutex to be unlocked
1492  */
1493 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1494 {
1495         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1496 }
1497 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1498
1499 /**
1500  * rt_mutex_futex_unlock - Futex variant of rt_mutex_unlock
1501  * @lock: the rt_mutex to be unlocked
1502  *
1503  * Returns: true/false indicating whether priority adjustment is
1504  * required or not.
1505  */
1506 bool __sched rt_mutex_futex_unlock(struct rt_mutex *lock,
1507                                    struct wake_q_head *wqh)
1508 {
1509         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL))) {
1510                 rt_mutex_deadlock_account_unlock(current);
1511                 return false;
1512         }
1513         return rt_mutex_slowunlock(lock, wqh);
1514 }
1515
1516 /**
1517  * rt_mutex_destroy - mark a mutex unusable
1518  * @lock: the mutex to be destroyed
1519  *
1520  * This function marks the mutex uninitialized, and any subsequent
1521  * use of the mutex is forbidden. The mutex must not be locked when
1522  * this function is called.
1523  */
1524 void rt_mutex_destroy(struct rt_mutex *lock)
1525 {
1526         WARN_ON(rt_mutex_is_locked(lock));
1527 #ifdef CONFIG_DEBUG_RT_MUTEXES
1528         lock->magic = NULL;
1529 #endif
1530 }
1531
1532 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1533
1534 /**
1535  * __rt_mutex_init - initialize the rt lock
1536  *
1537  * @lock: the rt lock to be initialized
1538  *
1539  * Initialize the rt lock to unlocked state.
1540  *
1541  * Initializing of a locked rt lock is not allowed
1542  */
1543 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
1544 {
1545         lock->owner = NULL;
1546         raw_spin_lock_init(&lock->wait_lock);
1547         lock->waiters = RB_ROOT;
1548         lock->waiters_leftmost = NULL;
1549
1550         debug_rt_mutex_init(lock, name);
1551 }
1552 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1553
1554 /**
1555  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1556  *                              proxy owner
1557  *
1558  * @lock:       the rt_mutex to be locked
1559  * @proxy_owner:the task to set as owner
1560  *
1561  * No locking. Caller has to do serializing itself
1562  * Special API call for PI-futex support
1563  */
1564 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1565                                 struct task_struct *proxy_owner)
1566 {
1567         __rt_mutex_init(lock, NULL);
1568         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1569         rt_mutex_set_owner(lock, proxy_owner);
1570         rt_mutex_deadlock_account_lock(lock, proxy_owner);
1571 }
1572
1573 /**
1574  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1575  *
1576  * @lock:       the rt_mutex to be locked
1577  *
1578  * No locking. Caller has to do serializing itself
1579  * Special API call for PI-futex support
1580  */
1581 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1582                            struct task_struct *proxy_owner)
1583 {
1584         debug_rt_mutex_proxy_unlock(lock);
1585         rt_mutex_set_owner(lock, NULL);
1586         rt_mutex_deadlock_account_unlock(proxy_owner);
1587 }
1588
1589 /**
1590  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1591  * @lock:               the rt_mutex to take
1592  * @waiter:             the pre-initialized rt_mutex_waiter
1593  * @task:               the task to prepare
1594  *
1595  * Returns:
1596  *  0 - task blocked on lock
1597  *  1 - acquired the lock for task, caller should wake it up
1598  * <0 - error
1599  *
1600  * Special API call for FUTEX_REQUEUE_PI support.
1601  */
1602 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1603                               struct rt_mutex_waiter *waiter,
1604                               struct task_struct *task)
1605 {
1606         int ret;
1607
1608         raw_spin_lock_irq(&lock->wait_lock);
1609
1610         if (try_to_take_rt_mutex(lock, task, NULL)) {
1611                 raw_spin_unlock_irq(&lock->wait_lock);
1612                 return 1;
1613         }
1614
1615         /* We enforce deadlock detection for futexes */
1616         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1617                                       RT_MUTEX_FULL_CHAINWALK);
1618
1619         if (ret && !rt_mutex_owner(lock)) {
1620                 /*
1621                  * Reset the return value. We might have
1622                  * returned with -EDEADLK and the owner
1623                  * released the lock while we were walking the
1624                  * pi chain.  Let the waiter sort it out.
1625                  */
1626                 ret = 0;
1627         }
1628
1629         if (unlikely(ret))
1630                 remove_waiter(lock, waiter);
1631
1632         raw_spin_unlock_irq(&lock->wait_lock);
1633
1634         debug_rt_mutex_print_deadlock(waiter);
1635
1636         return ret;
1637 }
1638
1639 /**
1640  * rt_mutex_next_owner - return the next owner of the lock
1641  *
1642  * @lock: the rt lock query
1643  *
1644  * Returns the next owner of the lock or NULL
1645  *
1646  * Caller has to serialize against other accessors to the lock
1647  * itself.
1648  *
1649  * Special API call for PI-futex support
1650  */
1651 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1652 {
1653         if (!rt_mutex_has_waiters(lock))
1654                 return NULL;
1655
1656         return rt_mutex_top_waiter(lock)->task;
1657 }
1658
1659 /**
1660  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1661  * @lock:               the rt_mutex we were woken on
1662  * @to:                 the timeout, null if none. hrtimer should already have
1663  *                      been started.
1664  * @waiter:             the pre-initialized rt_mutex_waiter
1665  *
1666  * Complete the lock acquisition started our behalf by another thread.
1667  *
1668  * Returns:
1669  *  0 - success
1670  * <0 - error, one of -EINTR, -ETIMEDOUT
1671  *
1672  * Special API call for PI-futex requeue support
1673  */
1674 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1675                                struct hrtimer_sleeper *to,
1676                                struct rt_mutex_waiter *waiter)
1677 {
1678         int ret;
1679
1680         raw_spin_lock_irq(&lock->wait_lock);
1681
1682         set_current_state(TASK_INTERRUPTIBLE);
1683
1684         /* sleep on the mutex */
1685         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1686
1687         if (unlikely(ret))
1688                 remove_waiter(lock, waiter);
1689
1690         /*
1691          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1692          * have to fix that up.
1693          */
1694         fixup_rt_mutex_waiters(lock);
1695
1696         raw_spin_unlock_irq(&lock->wait_lock);
1697
1698         return ret;
1699 }