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/signal.h>
  16 #include <linux/sched/rt.h>
  17 #include <linux/sched/deadline.h>
  18 #include <linux/sched/wake_q.h>
  19 #include <linux/sched/debug.h>
  20 #include <linux/timer.h>
  21
  22 #include "rtmutex_common.h"
  23
  24 /*
  25  * lock->owner state tracking:
  26  *
  27  * lock->owner holds the task_struct pointer of the owner. Bit 0
  28  * is used to keep track of the "lock has waiters" state.
  29  *
  30  * owner        bit0
  31  * NULL         0       lock is free (fast acquire possible)
  32  * NULL         1       lock is free and has waiters and the top waiter
  33  *                              is going to take the lock*
  34  * taskpointer  0       lock is held (fast release possible)
  35  * taskpointer  1       lock is held and has waiters**
  36  *
  37  * The fast atomic compare exchange based acquire and release is only
  38  * possible when bit 0 of lock->owner is 0.
  39  *
  40  * (*) It also can be a transitional state when grabbing the lock
  41  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  42  * we need to set the bit0 before looking at the lock, and the owner may be
  43  * NULL in this small time, hence this can be a transitional state.
  44  *
  45  * (**) There is a small time when bit 0 is set but there are no
  46  * waiters. This can happen when grabbing the lock in the slow path.
  47  * To prevent a cmpxchg of the owner releasing the lock, we need to
  48  * set this bit before looking at the lock.
  49  */
  50
  51 static void
  52 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
  53 {
  54         unsigned long val = (unsigned long)owner;
  55
  56         if (rt_mutex_has_waiters(lock))
  57                 val |= RT_MUTEX_HAS_WAITERS;
  58
  59         lock->owner = (struct task_struct *)val;
  60 }
  61
  62 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  63 {
  64         lock->owner = (struct task_struct *)
  65                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  66 }
  67
  68 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  69 {
  70         unsigned long owner, *p = (unsigned long *) &lock->owner;
  71
  72         if (rt_mutex_has_waiters(lock))
  73                 return;
  74
  75         /*
  76          * The rbtree has no waiters enqueued, now make sure that the
  77          * lock->owner still has the waiters bit set, otherwise the
  78          * following can happen:
  79          *
  80          * CPU 0        CPU 1           CPU2
  81          * l->owner=T1
  82          *              rt_mutex_lock(l)
  83          *              lock(l->lock)
  84          *              l->owner = T1 | HAS_WAITERS;
  85          *              enqueue(T2)
  86          *              boost()
  87          *                unlock(l->lock)
  88          *              block()
  89          *
  90          *                              rt_mutex_lock(l)
  91          *                              lock(l->lock)
  92          *                              l->owner = T1 | HAS_WAITERS;
  93          *                              enqueue(T3)
  94          *                              boost()
  95          *                                unlock(l->lock)
  96          *                              block()
  97          *              signal(->T2)    signal(->T3)
  98          *              lock(l->lock)
  99          *              dequeue(T2)
 100          *              deboost()
 101          *                unlock(l->lock)
 102          *                              lock(l->lock)
 103          *                              dequeue(T3)
 104          *                               ==> wait list is empty
 105          *                              deboost()
 106          *                               unlock(l->lock)
 107          *              lock(l->lock)
 108          *              fixup_rt_mutex_waiters()
 109          *                if (wait_list_empty(l) {
 110          *                  l->owner = owner
 111          *                  owner = l->owner & ~HAS_WAITERS;
 112          *                    ==> l->owner = T1
 113          *                }
 114          *                              lock(l->lock)
 115          * rt_mutex_unlock(l)           fixup_rt_mutex_waiters()
 116          *                                if (wait_list_empty(l) {
 117          *                                  owner = l->owner & ~HAS_WAITERS;
 118          * cmpxchg(l->owner, T1, NULL)
 119          *  ===> Success (l->owner = NULL)
 120          *
 121          *                                  l->owner = owner
 122          *                                    ==> l->owner = T1
 123          *                                }
 124          *
 125          * With the check for the waiter bit in place T3 on CPU2 will not
 126          * overwrite. All tasks fiddling with the waiters bit are
 127          * serialized by l->lock, so nothing else can modify the waiters
 128          * bit. If the bit is set then nothing can change l->owner either
 129          * so the simple RMW is safe. The cmpxchg() will simply fail if it
 130          * happens in the middle of the RMW because the waiters bit is
 131          * still set.
 132          */
 133         owner = READ_ONCE(*p);
 134         if (owner & RT_MUTEX_HAS_WAITERS)
 135                 WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
 136 }
 137
 138 /*
 139  * We can speed up the acquire/release, if there's no debugging state to be
 140  * set up.
 141  */
 142 #ifndef CONFIG_DEBUG_RT_MUTEXES
 143 # define rt_mutex_cmpxchg_relaxed(l,c,n) (cmpxchg_relaxed(&l->owner, c, n) == c)
 144 # define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
 145 # define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
 146
 147 /*
 148  * Callers must hold the ->wait_lock -- which is the whole purpose as we force
 149  * all future threads that attempt to [Rmw] the lock to the slowpath. As such
 150  * relaxed semantics suffice.
 151  */
 152 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 153 {
 154         unsigned long owner, *p = (unsigned long *) &lock->owner;
 155
 156         do {
 157                 owner = *p;
 158         } while (cmpxchg_relaxed(p, owner,
 159                                  owner | RT_MUTEX_HAS_WAITERS) != owner);
 160 }
 161
 162 /*
 163  * Safe fastpath aware unlock:
 164  * 1) Clear the waiters bit
 165  * 2) Drop lock->wait_lock
 166  * 3) Try to unlock the lock with cmpxchg
 167  */
 168 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
 169                                         unsigned long flags)
 170         __releases(lock->wait_lock)
 171 {
 172         struct task_struct *owner = rt_mutex_owner(lock);
 173
 174         clear_rt_mutex_waiters(lock);
 175         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 176         /*
 177          * If a new waiter comes in between the unlock and the cmpxchg
 178          * we have two situations:
 179          *
 180          * unlock(wait_lock);
 181          *                                      lock(wait_lock);
 182          * cmpxchg(p, owner, 0) == owner
 183          *                                      mark_rt_mutex_waiters(lock);
 184          *                                      acquire(lock);
 185          * or:
 186          *
 187          * unlock(wait_lock);
 188          *                                      lock(wait_lock);
 189          *                                      mark_rt_mutex_waiters(lock);
 190          *
 191          * cmpxchg(p, owner, 0) != owner
 192          *                                      enqueue_waiter();
 193          *                                      unlock(wait_lock);
 194          * lock(wait_lock);
 195          * wake waiter();
 196          * unlock(wait_lock);
 197          *                                      lock(wait_lock);
 198          *                                      acquire(lock);
 199          */
 200         return rt_mutex_cmpxchg_release(lock, owner, NULL);
 201 }
 202
 203 #else
 204 # define rt_mutex_cmpxchg_relaxed(l,c,n)        (0)
 205 # define rt_mutex_cmpxchg_acquire(l,c,n)        (0)
 206 # define rt_mutex_cmpxchg_release(l,c,n)        (0)
 207
 208 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
 209 {
 210         lock->owner = (struct task_struct *)
 211                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 212 }
 213
 214 /*
 215  * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 216  */
 217 static inline bool unlock_rt_mutex_safe(struct rt_mutex *lock,
 218                                         unsigned long flags)
 219         __releases(lock->wait_lock)
 220 {
 221         lock->owner = NULL;
 222         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 223         return true;
 224 }
 225 #endif
 226
 227 /*
 228  * Only use with rt_mutex_waiter_{less,equal}()
 229  */
 230 #define task_to_waiter(p)       \
 231         &(struct rt_mutex_waiter){ .prio = (p)->prio, .deadline = (p)->dl.deadline }
 232
 233 static inline int
 234 rt_mutex_waiter_less(struct rt_mutex_waiter *left,
 235                      struct rt_mutex_waiter *right)
 236 {
 237         if (left->prio < right->prio)
 238                 return 1;
 239
 240         /*
 241          * If both waiters have dl_prio(), we check the deadlines of the
 242          * associated tasks.
 243          * If left waiter has a dl_prio(), and we didn't return 1 above,
 244          * then right waiter has a dl_prio() too.
 245          */
 246         if (dl_prio(left->prio))
 247                 return dl_time_before(left->deadline, right->deadline);
 248
 249         return 0;
 250 }
 251
 252 static inline int
 253 rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
 254                       struct rt_mutex_waiter *right)
 255 {
 256         if (left->prio != right->prio)
 257                 return 0;
 258
 259         /*
 260          * If both waiters have dl_prio(), we check the deadlines of the
 261          * associated tasks.
 262          * If left waiter has a dl_prio(), and we didn't return 0 above,
 263          * then right waiter has a dl_prio() too.
 264          */
 265         if (dl_prio(left->prio))
 266                 return left->deadline == right->deadline;
 267
 268         return 1;
 269 }
 270
 271 static void
 272 rt_mutex_enqueue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 273 {
 274         struct rb_node **link = &lock->waiters.rb_root.rb_node;
 275         struct rb_node *parent = NULL;
 276         struct rt_mutex_waiter *entry;
 277         bool leftmost = true;
 278
 279         while (*link) {
 280                 parent = *link;
 281                 entry = rb_entry(parent, struct rt_mutex_waiter, tree_entry);
 282                 if (rt_mutex_waiter_less(waiter, entry)) {
 283                         link = &parent->rb_left;
 284                 } else {
 285                         link = &parent->rb_right;
 286                         leftmost = false;
 287                 }
 288         }
 289
 290         rb_link_node(&waiter->tree_entry, parent, link);
 291         rb_insert_color_cached(&waiter->tree_entry, &lock->waiters, leftmost);
 292 }
 293
 294 static void
 295 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 296 {
 297         if (RB_EMPTY_NODE(&waiter->tree_entry))
 298                 return;
 299
 300         rb_erase_cached(&waiter->tree_entry, &lock->waiters);
 301         RB_CLEAR_NODE(&waiter->tree_entry);
 302 }
 303
 304 static void
 305 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 306 {
 307         struct rb_node **link = &task->pi_waiters.rb_root.rb_node;
 308         struct rb_node *parent = NULL;
 309         struct rt_mutex_waiter *entry;
 310         bool leftmost = true;
 311
 312         while (*link) {
 313                 parent = *link;
 314                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
 315                 if (rt_mutex_waiter_less(waiter, entry)) {
 316                         link = &parent->rb_left;
 317                 } else {
 318                         link = &parent->rb_right;
 319                         leftmost = false;
 320                 }
 321         }
 322
 323         rb_link_node(&waiter->pi_tree_entry, parent, link);
 324         rb_insert_color_cached(&waiter->pi_tree_entry, &task->pi_waiters, leftmost);
 325 }
 326
 327 static void
 328 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 329 {
 330         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 331                 return;
 332
 333         rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
 334         RB_CLEAR_NODE(&waiter->pi_tree_entry);
 335 }
 336
 337 static void rt_mutex_adjust_prio(struct task_struct *p)
 338 {
 339         struct task_struct *pi_task = NULL;
 340
 341         lockdep_assert_held(&p->pi_lock);
 342
 343         if (task_has_pi_waiters(p))
 344                 pi_task = task_top_pi_waiter(p)->task;
 345
 346         rt_mutex_setprio(p, pi_task);
 347 }
 348
 349 /*
 350  * Deadlock detection is conditional:
 351  *
 352  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
 353  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
 354  *
 355  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
 356  * conducted independent of the detect argument.
 357  *
 358  * If the waiter argument is NULL this indicates the deboost path and
 359  * deadlock detection is disabled independent of the detect argument
 360  * and the config settings.
 361  */
 362 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
 363                                           enum rtmutex_chainwalk chwalk)
 364 {
 365         /*
 366          * This is just a wrapper function for the following call,
 367          * because debug_rt_mutex_detect_deadlock() smells like a magic
 368          * debug feature and I wanted to keep the cond function in the
 369          * main source file along with the comments instead of having
 370          * two of the same in the headers.
 371          */
 372         return debug_rt_mutex_detect_deadlock(waiter, chwalk);
 373 }
 374
 375 /*
 376  * Max number of times we'll walk the boosting chain:
 377  */
 378 int max_lock_depth = 1024;
 379
 380 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 381 {
 382         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 383 }
 384
 385 /*
 386  * Adjust the priority chain. Also used for deadlock detection.
 387  * Decreases task's usage by one - may thus free the task.
 388  *
 389  * @task:       the task owning the mutex (owner) for which a chain walk is
 390  *              probably needed
 391  * @chwalk:     do we have to carry out deadlock detection?
 392  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 393  *              things for a task that has just got its priority adjusted, and
 394  *              is waiting on a mutex)
 395  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 396  *              we dropped its pi_lock. Is never dereferenced, only used for
 397  *              comparison to detect lock chain changes.
 398  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 399  *              its priority to the mutex owner (can be NULL in the case
 400  *              depicted above or if the top waiter is gone away and we are
 401  *              actually deboosting the owner)
 402  * @top_task:   the current top waiter
 403  *
 404  * Returns 0 or -EDEADLK.
 405  *
 406  * Chain walk basics and protection scope
 407  *
 408  * [R] refcount on task
 409  * [P] task->pi_lock held
 410  * [L] rtmutex->wait_lock held
 411  *
 412  * Step Description                             Protected by
 413  *      function arguments:
 414  *      @task                                   [R]
 415  *      @orig_lock if != NULL                   @top_task is blocked on it
 416  *      @next_lock                              Unprotected. Cannot be
 417  *                                              dereferenced. Only used for
 418  *                                              comparison.
 419  *      @orig_waiter if != NULL                 @top_task is blocked on it
 420  *      @top_task                               current, or in case of proxy
 421  *                                              locking protected by calling
 422  *                                              code
 423  *      again:
 424  *        loop_sanity_check();
 425  *      retry:
 426  * [1]    lock(task->pi_lock);                  [R] acquire [P]
 427  * [2]    waiter = task->pi_blocked_on;         [P]
 428  * [3]    check_exit_conditions_1();            [P]
 429  * [4]    lock = waiter->lock;                  [P]
 430  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
 431  *          unlock(task->pi_lock);              release [P]
 432  *          goto retry;
 433  *        }
 434  * [6]    check_exit_conditions_2();            [P] + [L]
 435  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
 436  * [8]    unlock(task->pi_lock);                release [P]
 437  *        put_task_struct(task);                release [R]
 438  * [9]    check_exit_conditions_3();            [L]
 439  * [10]   task = owner(lock);                   [L]
 440  *        get_task_struct(task);                [L] acquire [R]
 441  *        lock(task->pi_lock);                  [L] acquire [P]
 442  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
 443  * [12]   check_exit_conditions_4();            [P] + [L]
 444  * [13]   unlock(task->pi_lock);                release [P]
 445  *        unlock(lock->wait_lock);              release [L]
 446  *        goto again;
 447  */
 448 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 449                                       enum rtmutex_chainwalk chwalk,
 450                                       struct rt_mutex *orig_lock,
 451                                       struct rt_mutex *next_lock,
 452                                       struct rt_mutex_waiter *orig_waiter,
 453                                       struct task_struct *top_task)
 454 {
 455         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 456         struct rt_mutex_waiter *prerequeue_top_waiter;
 457         int ret = 0, depth = 0;
 458         struct rt_mutex *lock;
 459         bool detect_deadlock;
 460         bool requeue = true;
 461
 462         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
 463
 464         /*
 465          * The (de)boosting is a step by step approach with a lot of
 466          * pitfalls. We want this to be preemptible and we want hold a
 467          * maximum of two locks per step. So we have to check
 468          * carefully whether things change under us.
 469          */
 470  again:
 471         /*
 472          * We limit the lock chain length for each invocation.
 473          */
 474         if (++depth > max_lock_depth) {
 475                 static int prev_max;
 476
 477                 /*
 478                  * Print this only once. If the admin changes the limit,
 479                  * print a new message when reaching the limit again.
 480                  */
 481                 if (prev_max != max_lock_depth) {
 482                         prev_max = max_lock_depth;
 483                         printk(KERN_WARNING "Maximum lock depth %d reached "
 484                                "task: %s (%d)\n", max_lock_depth,
 485                                top_task->comm, task_pid_nr(top_task));
 486                 }
 487                 put_task_struct(task);
 488
 489                 return -EDEADLK;
 490         }
 491
 492         /*
 493          * We are fully preemptible here and only hold the refcount on
 494          * @task. So everything can have changed under us since the
 495          * caller or our own code below (goto retry/again) dropped all
 496          * locks.
 497          */
 498  retry:
 499         /*
 500          * [1] Task cannot go away as we did a get_task() before !
 501          */
 502         raw_spin_lock_irq(&task->pi_lock);
 503
 504         /*
 505          * [2] Get the waiter on which @task is blocked on.
 506          */
 507         waiter = task->pi_blocked_on;
 508
 509         /*
 510          * [3] check_exit_conditions_1() protected by task->pi_lock.
 511          */
 512
 513         /*
 514          * Check whether the end of the boosting chain has been
 515          * reached or the state of the chain has changed while we
 516          * dropped the locks.
 517          */
 518         if (!waiter)
 519                 goto out_unlock_pi;
 520
 521         /*
 522          * Check the orig_waiter state. After we dropped the locks,
 523          * the previous owner of the lock might have released the lock.
 524          */
 525         if (orig_waiter && !rt_mutex_owner(orig_lock))
 526                 goto out_unlock_pi;
 527
 528         /*
 529          * We dropped all locks after taking a refcount on @task, so
 530          * the task might have moved on in the lock chain or even left
 531          * the chain completely and blocks now on an unrelated lock or
 532          * on @orig_lock.
 533          *
 534          * We stored the lock on which @task was blocked in @next_lock,
 535          * so we can detect the chain change.
 536          */
 537         if (next_lock != waiter->lock)
 538                 goto out_unlock_pi;
 539
 540         /*
 541          * Drop out, when the task has no waiters. Note,
 542          * top_waiter can be NULL, when we are in the deboosting
 543          * mode!
 544          */
 545         if (top_waiter) {
 546                 if (!task_has_pi_waiters(task))
 547                         goto out_unlock_pi;
 548                 /*
 549                  * If deadlock detection is off, we stop here if we
 550                  * are not the top pi waiter of the task. If deadlock
 551                  * detection is enabled we continue, but stop the
 552                  * requeueing in the chain walk.
 553                  */
 554                 if (top_waiter != task_top_pi_waiter(task)) {
 555                         if (!detect_deadlock)
 556                                 goto out_unlock_pi;
 557                         else
 558                                 requeue = false;
 559                 }
 560         }
 561
 562         /*
 563          * If the waiter priority is the same as the task priority
 564          * then there is no further priority adjustment necessary.  If
 565          * deadlock detection is off, we stop the chain walk. If its
 566          * enabled we continue, but stop the requeueing in the chain
 567          * walk.
 568          */
 569         if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
 570                 if (!detect_deadlock)
 571                         goto out_unlock_pi;
 572                 else
 573                         requeue = false;
 574         }
 575
 576         /*
 577          * [4] Get the next lock
 578          */
 579         lock = waiter->lock;
 580         /*
 581          * [5] We need to trylock here as we are holding task->pi_lock,
 582          * which is the reverse lock order versus the other rtmutex
 583          * operations.
 584          */
 585         if (!raw_spin_trylock(&lock->wait_lock)) {
 586                 raw_spin_unlock_irq(&task->pi_lock);
 587                 cpu_relax();
 588                 goto retry;
 589         }
 590
 591         /*
 592          * [6] check_exit_conditions_2() protected by task->pi_lock and
 593          * lock->wait_lock.
 594          *
 595          * Deadlock detection. If the lock is the same as the original
 596          * lock which caused us to walk the lock chain or if the
 597          * current lock is owned by the task which initiated the chain
 598          * walk, we detected a deadlock.
 599          */
 600         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 601                 debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
 602                 raw_spin_unlock(&lock->wait_lock);
 603                 ret = -EDEADLK;
 604                 goto out_unlock_pi;
 605         }
 606
 607         /*
 608          * If we just follow the lock chain for deadlock detection, no
 609          * need to do all the requeue operations. To avoid a truckload
 610          * of conditionals around the various places below, just do the
 611          * minimum chain walk checks.
 612          */
 613         if (!requeue) {
 614                 /*
 615                  * No requeue[7] here. Just release @task [8]
 616                  */
 617                 raw_spin_unlock(&task->pi_lock);
 618                 put_task_struct(task);
 619
 620                 /*
 621                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
 622                  * If there is no owner of the lock, end of chain.
 623                  */
 624                 if (!rt_mutex_owner(lock)) {
 625                         raw_spin_unlock_irq(&lock->wait_lock);
 626                         return 0;
 627                 }
 628
 629                 /* [10] Grab the next task, i.e. owner of @lock */
 630                 task = rt_mutex_owner(lock);
 631                 get_task_struct(task);
 632                 raw_spin_lock(&task->pi_lock);
 633
 634                 /*
 635                  * No requeue [11] here. We just do deadlock detection.
 636                  *
 637                  * [12] Store whether owner is blocked
 638                  * itself. Decision is made after dropping the locks
 639                  */
 640                 next_lock = task_blocked_on_lock(task);
 641                 /*
 642                  * Get the top waiter for the next iteration
 643                  */
 644                 top_waiter = rt_mutex_top_waiter(lock);
 645
 646                 /* [13] Drop locks */
 647                 raw_spin_unlock(&task->pi_lock);
 648                 raw_spin_unlock_irq(&lock->wait_lock);
 649
 650                 /* If owner is not blocked, end of chain. */
 651                 if (!next_lock)
 652                         goto out_put_task;
 653                 goto again;
 654         }
 655
 656         /*
 657          * Store the current top waiter before doing the requeue
 658          * operation on @lock. We need it for the boost/deboost
 659          * decision below.
 660          */
 661         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
 662
 663         /* [7] Requeue the waiter in the lock waiter tree. */
 664         rt_mutex_dequeue(lock, waiter);
 665
 666         /*
 667          * Update the waiter prio fields now that we're dequeued.
 668          *
 669          * These values can have changed through either:
 670          *
 671          *   sys_sched_set_scheduler() / sys_sched_setattr()
 672          *
 673          * or
 674          *
 675          *   DL CBS enforcement advancing the effective deadline.
 676          *
 677          * Even though pi_waiters also uses these fields, and that tree is only
 678          * updated in [11], we can do this here, since we hold [L], which
 679          * serializes all pi_waiters access and rb_erase() does not care about
 680          * the values of the node being removed.
 681          */
 682         waiter->prio = task->prio;
 683         waiter->deadline = task->dl.deadline;
 684
 685         rt_mutex_enqueue(lock, waiter);
 686
 687         /* [8] Release the task */
 688         raw_spin_unlock(&task->pi_lock);
 689         put_task_struct(task);
 690
 691         /*
 692          * [9] check_exit_conditions_3 protected by lock->wait_lock.
 693          *
 694          * We must abort the chain walk if there is no lock owner even
 695          * in the dead lock detection case, as we have nothing to
 696          * follow here. This is the end of the chain we are walking.
 697          */
 698         if (!rt_mutex_owner(lock)) {
 699                 /*
 700                  * If the requeue [7] above changed the top waiter,
 701                  * then we need to wake the new top waiter up to try
 702                  * to get the lock.
 703                  */
 704                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
 705                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 706                 raw_spin_unlock_irq(&lock->wait_lock);
 707                 return 0;
 708         }
 709
 710         /* [10] Grab the next task, i.e. the owner of @lock */
 711         task = rt_mutex_owner(lock);
 712         get_task_struct(task);
 713         raw_spin_lock(&task->pi_lock);
 714
 715         /* [11] requeue the pi waiters if necessary */
 716         if (waiter == rt_mutex_top_waiter(lock)) {
 717                 /*
 718                  * The waiter became the new top (highest priority)
 719                  * waiter on the lock. Replace the previous top waiter
 720                  * in the owner tasks pi waiters tree with this waiter
 721                  * and adjust the priority of the owner.
 722                  */
 723                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
 724                 rt_mutex_enqueue_pi(task, waiter);
 725                 rt_mutex_adjust_prio(task);
 726
 727         } else if (prerequeue_top_waiter == waiter) {
 728                 /*
 729                  * The waiter was the top waiter on the lock, but is
 730                  * no longer the top prority waiter. Replace waiter in
 731                  * the owner tasks pi waiters tree with the new top
 732                  * (highest priority) waiter and adjust the priority
 733                  * of the owner.
 734                  * The new top waiter is stored in @waiter so that
 735                  * @waiter == @top_waiter evaluates to true below and
 736                  * we continue to deboost the rest of the chain.
 737                  */
 738                 rt_mutex_dequeue_pi(task, waiter);
 739                 waiter = rt_mutex_top_waiter(lock);
 740                 rt_mutex_enqueue_pi(task, waiter);
 741                 rt_mutex_adjust_prio(task);
 742         } else {
 743                 /*
 744                  * Nothing changed. No need to do any priority
 745                  * adjustment.
 746                  */
 747         }
 748
 749         /*
 750          * [12] check_exit_conditions_4() protected by task->pi_lock
 751          * and lock->wait_lock. The actual decisions are made after we
 752          * dropped the locks.
 753          *
 754          * Check whether the task which owns the current lock is pi
 755          * blocked itself. If yes we store a pointer to the lock for
 756          * the lock chain change detection above. After we dropped
 757          * task->pi_lock next_lock cannot be dereferenced anymore.
 758          */
 759         next_lock = task_blocked_on_lock(task);
 760         /*
 761          * Store the top waiter of @lock for the end of chain walk
 762          * decision below.
 763          */
 764         top_waiter = rt_mutex_top_waiter(lock);
 765
 766         /* [13] Drop the locks */
 767         raw_spin_unlock(&task->pi_lock);
 768         raw_spin_unlock_irq(&lock->wait_lock);
 769
 770         /*
 771          * Make the actual exit decisions [12], based on the stored
 772          * values.
 773          *
 774          * We reached the end of the lock chain. Stop right here. No
 775          * point to go back just to figure that out.
 776          */
 777         if (!next_lock)
 778                 goto out_put_task;
 779
 780         /*
 781          * If the current waiter is not the top waiter on the lock,
 782          * then we can stop the chain walk here if we are not in full
 783          * deadlock detection mode.
 784          */
 785         if (!detect_deadlock && waiter != top_waiter)
 786                 goto out_put_task;
 787
 788         goto again;
 789
 790  out_unlock_pi:
 791         raw_spin_unlock_irq(&task->pi_lock);
 792  out_put_task:
 793         put_task_struct(task);
 794
 795         return ret;
 796 }
 797
 798 /*
 799  * Try to take an rt-mutex
 800  *
 801  * Must be called with lock->wait_lock held and interrupts disabled
 802  *
 803  * @lock:   The lock to be acquired.
 804  * @task:   The task which wants to acquire the lock
 805  * @waiter: The waiter that is queued to the lock's wait tree if the
 806  *          callsite called task_blocked_on_lock(), otherwise NULL
 807  */
 808 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 809                                 struct rt_mutex_waiter *waiter)
 810 {
 811         lockdep_assert_held(&lock->wait_lock);
 812
 813         /*
 814          * Before testing whether we can acquire @lock, we set the
 815          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
 816          * other tasks which try to modify @lock into the slow path
 817          * and they serialize on @lock->wait_lock.
 818          *
 819          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
 820          * as explained at the top of this file if and only if:
 821          *
 822          * - There is a lock owner. The caller must fixup the
 823          *   transient state if it does a trylock or leaves the lock
 824          *   function due to a signal or timeout.
 825          *
 826          * - @task acquires the lock and there are no other
 827          *   waiters. This is undone in rt_mutex_set_owner(@task) at
 828          *   the end of this function.
 829          */
 830         mark_rt_mutex_waiters(lock);
 831
 832         /*
 833          * If @lock has an owner, give up.
 834          */
 835         if (rt_mutex_owner(lock))
 836                 return 0;
 837
 838         /*
 839          * If @waiter != NULL, @task has already enqueued the waiter
 840          * into @lock waiter tree. If @waiter == NULL then this is a
 841          * trylock attempt.
 842          */
 843         if (waiter) {
 844                 /*
 845                  * If waiter is not the highest priority waiter of
 846                  * @lock, give up.
 847                  */
 848                 if (waiter != rt_mutex_top_waiter(lock))
 849                         return 0;
 850
 851                 /*
 852                  * We can acquire the lock. Remove the waiter from the
 853                  * lock waiters tree.
 854                  */
 855                 rt_mutex_dequeue(lock, waiter);
 856
 857         } else {
 858                 /*
 859                  * If the lock has waiters already we check whether @task is
 860                  * eligible to take over the lock.
 861                  *
 862                  * If there are no other waiters, @task can acquire
 863                  * the lock.  @task->pi_blocked_on is NULL, so it does
 864                  * not need to be dequeued.
 865                  */
 866                 if (rt_mutex_has_waiters(lock)) {
 867                         /*
 868                          * If @task->prio is greater than or equal to
 869                          * the top waiter priority (kernel view),
 870                          * @task lost.
 871                          */
 872                         if (!rt_mutex_waiter_less(task_to_waiter(task),
 873                                                   rt_mutex_top_waiter(lock)))
 874                                 return 0;
 875
 876                         /*
 877                          * The current top waiter stays enqueued. We
 878                          * don't have to change anything in the lock
 879                          * waiters order.
 880                          */
 881                 } else {
 882                         /*
 883                          * No waiters. Take the lock without the
 884                          * pi_lock dance.@task->pi_blocked_on is NULL
 885                          * and we have no waiters to enqueue in @task
 886                          * pi waiters tree.
 887                          */
 888                         goto takeit;
 889                 }
 890         }
 891
 892         /*
 893          * Clear @task->pi_blocked_on. Requires protection by
 894          * @task->pi_lock. Redundant operation for the @waiter == NULL
 895          * case, but conditionals are more expensive than a redundant
 896          * store.
 897          */
 898         raw_spin_lock(&task->pi_lock);
 899         task->pi_blocked_on = NULL;
 900         /*
 901          * Finish the lock acquisition. @task is the new owner. If
 902          * other waiters exist we have to insert the highest priority
 903          * waiter into @task->pi_waiters tree.
 904          */
 905         if (rt_mutex_has_waiters(lock))
 906                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
 907         raw_spin_unlock(&task->pi_lock);
 908
 909 takeit:
 910         /* We got the lock. */
 911         debug_rt_mutex_lock(lock);
 912
 913         /*
 914          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
 915          * are still waiters or clears it.
 916          */
 917         rt_mutex_set_owner(lock, task);
 918
 919         return 1;
 920 }
 921
 922 /*
 923  * Task blocks on lock.
 924  *
 925  * Prepare waiter and propagate pi chain
 926  *
 927  * This must be called with lock->wait_lock held and interrupts disabled
 928  */
 929 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 930                                    struct rt_mutex_waiter *waiter,
 931                                    struct task_struct *task,
 932                                    enum rtmutex_chainwalk chwalk)
 933 {
 934         struct task_struct *owner = rt_mutex_owner(lock);
 935         struct rt_mutex_waiter *top_waiter = waiter;
 936         struct rt_mutex *next_lock;
 937         int chain_walk = 0, res;
 938
 939         lockdep_assert_held(&lock->wait_lock);
 940
 941         /*
 942          * Early deadlock detection. We really don't want the task to
 943          * enqueue on itself just to untangle the mess later. It's not
 944          * only an optimization. We drop the locks, so another waiter
 945          * can come in before the chain walk detects the deadlock. So
 946          * the other will detect the deadlock and return -EDEADLOCK,
 947          * which is wrong, as the other waiter is not in a deadlock
 948          * situation.
 949          */
 950         if (owner == task)
 951                 return -EDEADLK;
 952
 953         raw_spin_lock(&task->pi_lock);
 954         waiter->task = task;
 955         waiter->lock = lock;
 956         waiter->prio = task->prio;
 957         waiter->deadline = task->dl.deadline;
 958
 959         /* Get the top priority waiter on the lock */
 960         if (rt_mutex_has_waiters(lock))
 961                 top_waiter = rt_mutex_top_waiter(lock);
 962         rt_mutex_enqueue(lock, waiter);
 963
 964         task->pi_blocked_on = waiter;
 965
 966         raw_spin_unlock(&task->pi_lock);
 967
 968         if (!owner)
 969                 return 0;
 970
 971         raw_spin_lock(&owner->pi_lock);
 972         if (waiter == rt_mutex_top_waiter(lock)) {
 973                 rt_mutex_dequeue_pi(owner, top_waiter);
 974                 rt_mutex_enqueue_pi(owner, waiter);
 975
 976                 rt_mutex_adjust_prio(owner);
 977                 if (owner->pi_blocked_on)
 978                         chain_walk = 1;
 979         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
 980                 chain_walk = 1;
 981         }
 982
 983         /* Store the lock on which owner is blocked or NULL */
 984         next_lock = task_blocked_on_lock(owner);
 985
 986         raw_spin_unlock(&owner->pi_lock);
 987         /*
 988          * Even if full deadlock detection is on, if the owner is not
 989          * blocked itself, we can avoid finding this out in the chain
 990          * walk.
 991          */
 992         if (!chain_walk || !next_lock)
 993                 return 0;
 994
 995         /*
 996          * The owner can't disappear while holding a lock,
 997          * so the owner struct is protected by wait_lock.
 998          * Gets dropped in rt_mutex_adjust_prio_chain()!
 999          */
1000         get_task_struct(owner);
1001
1002         raw_spin_unlock_irq(&lock->wait_lock);
1003
1004         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1005                                          next_lock, waiter, task);
1006
1007         raw_spin_lock_irq(&lock->wait_lock);
1008
1009         return res;
1010 }
1011
1012 /*
1013  * Remove the top waiter from the current tasks pi waiter tree and
1014  * queue it up.
1015  *
1016  * Called with lock->wait_lock held and interrupts disabled.
1017  */
1018 static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
1019                                     struct rt_mutex *lock)
1020 {
1021         struct rt_mutex_waiter *waiter;
1022
1023         raw_spin_lock(&current->pi_lock);
1024
1025         waiter = rt_mutex_top_waiter(lock);
1026
1027         /*
1028          * Remove it from current->pi_waiters and deboost.
1029          *
1030          * We must in fact deboost here in order to ensure we call
1031          * rt_mutex_setprio() to update p->pi_top_task before the
1032          * task unblocks.
1033          */
1034         rt_mutex_dequeue_pi(current, waiter);
1035         rt_mutex_adjust_prio(current);
1036
1037         /*
1038          * As we are waking up the top waiter, and the waiter stays
1039          * queued on the lock until it gets the lock, this lock
1040          * obviously has waiters. Just set the bit here and this has
1041          * the added benefit of forcing all new tasks into the
1042          * slow path making sure no task of lower priority than
1043          * the top waiter can steal this lock.
1044          */
1045         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1046
1047         /*
1048          * We deboosted before waking the top waiter task such that we don't
1049          * run two tasks with the 'same' priority (and ensure the
1050          * p->pi_top_task pointer points to a blocked task). This however can
1051          * lead to priority inversion if we would get preempted after the
1052          * deboost but before waking our donor task, hence the preempt_disable()
1053          * before unlock.
1054          *
1055          * Pairs with preempt_enable() in rt_mutex_postunlock();
1056          */
1057         preempt_disable();
1058         wake_q_add(wake_q, waiter->task);
1059         raw_spin_unlock(&current->pi_lock);
1060 }
1061
1062 /*
1063  * Remove a waiter from a lock and give up
1064  *
1065  * Must be called with lock->wait_lock held and interrupts disabled. I must
1066  * have just failed to try_to_take_rt_mutex().
1067  */
1068 static void remove_waiter(struct rt_mutex *lock,
1069                           struct rt_mutex_waiter *waiter)
1070 {
1071         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1072         struct task_struct *owner = rt_mutex_owner(lock);
1073         struct rt_mutex *next_lock;
1074
1075         lockdep_assert_held(&lock->wait_lock);
1076
1077         raw_spin_lock(&current->pi_lock);
1078         rt_mutex_dequeue(lock, waiter);
1079         current->pi_blocked_on = NULL;
1080         raw_spin_unlock(&current->pi_lock);
1081
1082         /*
1083          * Only update priority if the waiter was the highest priority
1084          * waiter of the lock and there is an owner to update.
1085          */
1086         if (!owner || !is_top_waiter)
1087                 return;
1088
1089         raw_spin_lock(&owner->pi_lock);
1090
1091         rt_mutex_dequeue_pi(owner, waiter);
1092
1093         if (rt_mutex_has_waiters(lock))
1094                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1095
1096         rt_mutex_adjust_prio(owner);
1097
1098         /* Store the lock on which owner is blocked or NULL */
1099         next_lock = task_blocked_on_lock(owner);
1100
1101         raw_spin_unlock(&owner->pi_lock);
1102
1103         /*
1104          * Don't walk the chain, if the owner task is not blocked
1105          * itself.
1106          */
1107         if (!next_lock)
1108                 return;
1109
1110         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1111         get_task_struct(owner);
1112
1113         raw_spin_unlock_irq(&lock->wait_lock);
1114
1115         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1116                                    next_lock, NULL, current);
1117
1118         raw_spin_lock_irq(&lock->wait_lock);
1119 }
1120
1121 /*
1122  * Recheck the pi chain, in case we got a priority setting
1123  *
1124  * Called from sched_setscheduler
1125  */
1126 void rt_mutex_adjust_pi(struct task_struct *task)
1127 {
1128         struct rt_mutex_waiter *waiter;
1129         struct rt_mutex *next_lock;
1130         unsigned long flags;
1131
1132         raw_spin_lock_irqsave(&task->pi_lock, flags);
1133
1134         waiter = task->pi_blocked_on;
1135         if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1136                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1137                 return;
1138         }
1139         next_lock = waiter->lock;
1140         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1141
1142         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1143         get_task_struct(task);
1144
1145         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1146                                    next_lock, NULL, task);
1147 }
1148
1149 void rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1150 {
1151         debug_rt_mutex_init_waiter(waiter);
1152         RB_CLEAR_NODE(&waiter->pi_tree_entry);
1153         RB_CLEAR_NODE(&waiter->tree_entry);
1154         waiter->task = NULL;
1155 }
1156
1157 /**
1158  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1159  * @lock:                the rt_mutex to take
1160  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1161  *                       or TASK_UNINTERRUPTIBLE)
1162  * @timeout:             the pre-initialized and started timer, or NULL for none
1163  * @waiter:              the pre-initialized rt_mutex_waiter
1164  *
1165  * Must be called with lock->wait_lock held and interrupts disabled
1166  */
1167 static int __sched
1168 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1169                     struct hrtimer_sleeper *timeout,
1170                     struct rt_mutex_waiter *waiter)
1171 {
1172         int ret = 0;
1173
1174         for (;;) {
1175                 /* Try to acquire the lock: */
1176                 if (try_to_take_rt_mutex(lock, current, waiter))
1177                         break;
1178
1179                 /*
1180                  * TASK_INTERRUPTIBLE checks for signals and
1181                  * timeout. Ignored otherwise.
1182                  */
1183                 if (likely(state == TASK_INTERRUPTIBLE)) {
1184                         /* Signal pending? */
1185                         if (signal_pending(current))
1186                                 ret = -EINTR;
1187                         if (timeout && !timeout->task)
1188                                 ret = -ETIMEDOUT;
1189                         if (ret)
1190                                 break;
1191                 }
1192
1193                 raw_spin_unlock_irq(&lock->wait_lock);
1194
1195                 debug_rt_mutex_print_deadlock(waiter);
1196
1197                 schedule();
1198
1199                 raw_spin_lock_irq(&lock->wait_lock);
1200                 set_current_state(state);
1201         }
1202
1203         __set_current_state(TASK_RUNNING);
1204         return ret;
1205 }
1206
1207 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1208                                      struct rt_mutex_waiter *w)
1209 {
1210         /*
1211          * If the result is not -EDEADLOCK or the caller requested
1212          * deadlock detection, nothing to do here.
1213          */
1214         if (res != -EDEADLOCK || detect_deadlock)
1215                 return;
1216
1217         /*
1218          * Yell lowdly and stop the task right here.
1219          */
1220         rt_mutex_print_deadlock(w);
1221         while (1) {
1222                 set_current_state(TASK_INTERRUPTIBLE);
1223                 schedule();
1224         }
1225 }
1226
1227 /*
1228  * Slow path lock function:
1229  */
1230 static int __sched
1231 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1232                   struct hrtimer_sleeper *timeout,
1233                   enum rtmutex_chainwalk chwalk)
1234 {
1235         struct rt_mutex_waiter waiter;
1236         unsigned long flags;
1237         int ret = 0;
1238
1239         rt_mutex_init_waiter(&waiter);
1240
1241         /*
1242          * Technically we could use raw_spin_[un]lock_irq() here, but this can
1243          * be called in early boot if the cmpxchg() fast path is disabled
1244          * (debug, no architecture support). In this case we will acquire the
1245          * rtmutex with lock->wait_lock held. But we cannot unconditionally
1246          * enable interrupts in that early boot case. So we need to use the
1247          * irqsave/restore variants.
1248          */
1249         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1250
1251         /* Try to acquire the lock again: */
1252         if (try_to_take_rt_mutex(lock, current, NULL)) {
1253                 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1254                 return 0;
1255         }
1256
1257         set_current_state(state);
1258
1259         /* Setup the timer, when timeout != NULL */
1260         if (unlikely(timeout))
1261                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1262
1263         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1264
1265         if (likely(!ret))
1266                 /* sleep on the mutex */
1267                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1268
1269         if (unlikely(ret)) {
1270                 __set_current_state(TASK_RUNNING);
1271                 if (rt_mutex_has_waiters(lock))
1272                         remove_waiter(lock, &waiter);
1273                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1274         }
1275
1276         /*
1277          * try_to_take_rt_mutex() sets the waiter bit
1278          * unconditionally. We might have to fix that up.
1279          */
1280         fixup_rt_mutex_waiters(lock);
1281
1282         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1283
1284         /* Remove pending timer: */
1285         if (unlikely(timeout))
1286                 hrtimer_cancel(&timeout->timer);
1287
1288         debug_rt_mutex_free_waiter(&waiter);
1289
1290         return ret;
1291 }
1292
1293 static inline int __rt_mutex_slowtrylock(struct rt_mutex *lock)
1294 {
1295         int ret = try_to_take_rt_mutex(lock, current, NULL);
1296
1297         /*
1298          * try_to_take_rt_mutex() sets the lock waiters bit
1299          * unconditionally. Clean this up.
1300          */
1301         fixup_rt_mutex_waiters(lock);
1302
1303         return ret;
1304 }
1305
1306 /*
1307  * Slow path try-lock function:
1308  */
1309 static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1310 {
1311         unsigned long flags;
1312         int ret;
1313
1314         /*
1315          * If the lock already has an owner we fail to get the lock.
1316          * This can be done without taking the @lock->wait_lock as
1317          * it is only being read, and this is a trylock anyway.
1318          */
1319         if (rt_mutex_owner(lock))
1320                 return 0;
1321
1322         /*
1323          * The mutex has currently no owner. Lock the wait lock and try to
1324          * acquire the lock. We use irqsave here to support early boot calls.
1325          */
1326         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1327
1328         ret = __rt_mutex_slowtrylock(lock);
1329
1330         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1331
1332         return ret;
1333 }
1334
1335 /*
1336  * Slow path to release a rt-mutex.
1337  *
1338  * Return whether the current task needs to call rt_mutex_postunlock().
1339  */
1340 static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1341                                         struct wake_q_head *wake_q)
1342 {
1343         unsigned long flags;
1344
1345         /* irqsave required to support early boot calls */
1346         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1347
1348         debug_rt_mutex_unlock(lock);
1349
1350         /*
1351          * We must be careful here if the fast path is enabled. If we
1352          * have no waiters queued we cannot set owner to NULL here
1353          * because of:
1354          *
1355          * foo->lock->owner = NULL;
1356          *                      rtmutex_lock(foo->lock);   <- fast path
1357          *                      free = atomic_dec_and_test(foo->refcnt);
1358          *                      rtmutex_unlock(foo->lock); <- fast path
1359          *                      if (free)
1360          *                              kfree(foo);
1361          * raw_spin_unlock(foo->lock->wait_lock);
1362          *
1363          * So for the fastpath enabled kernel:
1364          *
1365          * Nothing can set the waiters bit as long as we hold
1366          * lock->wait_lock. So we do the following sequence:
1367          *
1368          *      owner = rt_mutex_owner(lock);
1369          *      clear_rt_mutex_waiters(lock);
1370          *      raw_spin_unlock(&lock->wait_lock);
1371          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1372          *              return;
1373          *      goto retry;
1374          *
1375          * The fastpath disabled variant is simple as all access to
1376          * lock->owner is serialized by lock->wait_lock:
1377          *
1378          *      lock->owner = NULL;
1379          *      raw_spin_unlock(&lock->wait_lock);
1380          */
1381         while (!rt_mutex_has_waiters(lock)) {
1382                 /* Drops lock->wait_lock ! */
1383                 if (unlock_rt_mutex_safe(lock, flags) == true)
1384                         return false;
1385                 /* Relock the rtmutex and try again */
1386                 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1387         }
1388
1389         /*
1390          * The wakeup next waiter path does not suffer from the above
1391          * race. See the comments there.
1392          *
1393          * Queue the next waiter for wakeup once we release the wait_lock.
1394          */
1395         mark_wakeup_next_waiter(wake_q, lock);
1396         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1397
1398         return true; /* call rt_mutex_postunlock() */
1399 }
1400
1401 /*
1402  * debug aware fast / slowpath lock,trylock,unlock
1403  *
1404  * The atomic acquire/release ops are compiled away, when either the
1405  * architecture does not support cmpxchg or when debugging is enabled.
1406  */
1407 static inline int
1408 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1409                   int (*slowfn)(struct rt_mutex *lock, int state,
1410                                 struct hrtimer_sleeper *timeout,
1411                                 enum rtmutex_chainwalk chwalk))
1412 {
1413         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1414                 return 0;
1415
1416         return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1417 }
1418
1419 static inline int
1420 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1421                         struct hrtimer_sleeper *timeout,
1422                         enum rtmutex_chainwalk chwalk,
1423                         int (*slowfn)(struct rt_mutex *lock, int state,
1424                                       struct hrtimer_sleeper *timeout,
1425                                       enum rtmutex_chainwalk chwalk))
1426 {
1427         if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1428             likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1429                 return 0;
1430
1431         return slowfn(lock, state, timeout, chwalk);
1432 }
1433
1434 static inline int
1435 rt_mutex_fasttrylock(struct rt_mutex *lock,
1436                      int (*slowfn)(struct rt_mutex *lock))
1437 {
1438         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1439                 return 1;
1440
1441         return slowfn(lock);
1442 }
1443
1444 /*
1445  * Performs the wakeup of the the top-waiter and re-enables preemption.
1446  */
1447 void rt_mutex_postunlock(struct wake_q_head *wake_q)
1448 {
1449         wake_up_q(wake_q);
1450
1451         /* Pairs with preempt_disable() in rt_mutex_slowunlock() */
1452         preempt_enable();
1453 }
1454
1455 static inline void
1456 rt_mutex_fastunlock(struct rt_mutex *lock,
1457                     bool (*slowfn)(struct rt_mutex *lock,
1458                                    struct wake_q_head *wqh))
1459 {
1460         DEFINE_WAKE_Q(wake_q);
1461
1462         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1463                 return;
1464
1465         if (slowfn(lock, &wake_q))
1466                 rt_mutex_postunlock(&wake_q);
1467 }
1468
1469 /**
1470  * rt_mutex_lock - lock a rt_mutex
1471  *
1472  * @lock: the rt_mutex to be locked
1473  */
1474 void __sched rt_mutex_lock(struct rt_mutex *lock)
1475 {
1476         might_sleep();
1477
1478         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1479         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1480 }
1481 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1482
1483 /**
1484  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1485  *
1486  * @lock:               the rt_mutex to be locked
1487  *
1488  * Returns:
1489  *  0           on success
1490  * -EINTR       when interrupted by a signal
1491  */
1492 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1493 {
1494         int ret;
1495
1496         might_sleep();
1497
1498         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1499         ret = rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1500         if (ret)
1501                 mutex_release(&lock->dep_map, 1, _RET_IP_);
1502
1503         return ret;
1504 }
1505 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1506
1507 /*
1508  * Futex variant, must not use fastpath.
1509  */
1510 int __sched rt_mutex_futex_trylock(struct rt_mutex *lock)
1511 {
1512         return rt_mutex_slowtrylock(lock);
1513 }
1514
1515 int __sched __rt_mutex_futex_trylock(struct rt_mutex *lock)
1516 {
1517         return __rt_mutex_slowtrylock(lock);
1518 }
1519
1520 /**
1521  * rt_mutex_timed_lock - lock a rt_mutex interruptible
1522  *                      the timeout structure is provided
1523  *                      by the caller
1524  *
1525  * @lock:               the rt_mutex to be locked
1526  * @timeout:            timeout structure or NULL (no timeout)
1527  *
1528  * Returns:
1529  *  0           on success
1530  * -EINTR       when interrupted by a signal
1531  * -ETIMEDOUT   when the timeout expired
1532  */
1533 int
1534 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout)
1535 {
1536         int ret;
1537
1538         might_sleep();
1539
1540         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1541         ret = rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
1542                                        RT_MUTEX_MIN_CHAINWALK,
1543                                        rt_mutex_slowlock);
1544         if (ret)
1545                 mutex_release(&lock->dep_map, 1, _RET_IP_);
1546
1547         return ret;
1548 }
1549 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
1550
1551 /**
1552  * rt_mutex_trylock - try to lock a rt_mutex
1553  *
1554  * @lock:       the rt_mutex to be locked
1555  *
1556  * This function can only be called in thread context. It's safe to
1557  * call it from atomic regions, but not from hard interrupt or soft
1558  * interrupt context.
1559  *
1560  * Returns 1 on success and 0 on contention
1561  */
1562 int __sched rt_mutex_trylock(struct rt_mutex *lock)
1563 {
1564         int ret;
1565
1566         if (WARN_ON_ONCE(in_irq() || in_nmi() || in_serving_softirq()))
1567                 return 0;
1568
1569         ret = rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
1570         if (ret)
1571                 mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
1572
1573         return ret;
1574 }
1575 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
1576
1577 /**
1578  * rt_mutex_unlock - unlock a rt_mutex
1579  *
1580  * @lock: the rt_mutex to be unlocked
1581  */
1582 void __sched rt_mutex_unlock(struct rt_mutex *lock)
1583 {
1584         mutex_release(&lock->dep_map, 1, _RET_IP_);
1585         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
1586 }
1587 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
1588
1589 /**
1590  * Futex variant, that since futex variants do not use the fast-path, can be
1591  * simple and will not need to retry.
1592  */
1593 bool __sched __rt_mutex_futex_unlock(struct rt_mutex *lock,
1594                                     struct wake_q_head *wake_q)
1595 {
1596         lockdep_assert_held(&lock->wait_lock);
1597
1598         debug_rt_mutex_unlock(lock);
1599
1600         if (!rt_mutex_has_waiters(lock)) {
1601                 lock->owner = NULL;
1602                 return false; /* done */
1603         }
1604
1605         /*
1606          * We've already deboosted, mark_wakeup_next_waiter() will
1607          * retain preempt_disabled when we drop the wait_lock, to
1608          * avoid inversion prior to the wakeup.  preempt_disable()
1609          * therein pairs with rt_mutex_postunlock().
1610          */
1611         mark_wakeup_next_waiter(wake_q, lock);
1612
1613         return true; /* call postunlock() */
1614 }
1615
1616 void __sched rt_mutex_futex_unlock(struct rt_mutex *lock)
1617 {
1618         DEFINE_WAKE_Q(wake_q);
1619         bool postunlock;
1620
1621         raw_spin_lock_irq(&lock->wait_lock);
1622         postunlock = __rt_mutex_futex_unlock(lock, &wake_q);
1623         raw_spin_unlock_irq(&lock->wait_lock);
1624
1625         if (postunlock)
1626                 rt_mutex_postunlock(&wake_q);
1627 }
1628
1629 /**
1630  * rt_mutex_destroy - mark a mutex unusable
1631  * @lock: the mutex to be destroyed
1632  *
1633  * This function marks the mutex uninitialized, and any subsequent
1634  * use of the mutex is forbidden. The mutex must not be locked when
1635  * this function is called.
1636  */
1637 void rt_mutex_destroy(struct rt_mutex *lock)
1638 {
1639         WARN_ON(rt_mutex_is_locked(lock));
1640 #ifdef CONFIG_DEBUG_RT_MUTEXES
1641         lock->magic = NULL;
1642 #endif
1643 }
1644 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
1645
1646 /**
1647  * __rt_mutex_init - initialize the rt lock
1648  *
1649  * @lock: the rt lock to be initialized
1650  *
1651  * Initialize the rt lock to unlocked state.
1652  *
1653  * Initializing of a locked rt lock is not allowed
1654  */
1655 void __rt_mutex_init(struct rt_mutex *lock, const char *name,
1656                      struct lock_class_key *key)
1657 {
1658         lock->owner = NULL;
1659         raw_spin_lock_init(&lock->wait_lock);
1660         lock->waiters = RB_ROOT_CACHED;
1661
1662         if (name && key)
1663                 debug_rt_mutex_init(lock, name, key);
1664 }
1665 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1666
1667 /**
1668  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1669  *                              proxy owner
1670  *
1671  * @lock:       the rt_mutex to be locked
1672  * @proxy_owner:the task to set as owner
1673  *
1674  * No locking. Caller has to do serializing itself
1675  *
1676  * Special API call for PI-futex support. This initializes the rtmutex and
1677  * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1678  * possible at this point because the pi_state which contains the rtmutex
1679  * is not yet visible to other tasks.
1680  */
1681 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1682                                 struct task_struct *proxy_owner)
1683 {
1684         __rt_mutex_init(lock, NULL, NULL);
1685         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1686         rt_mutex_set_owner(lock, proxy_owner);
1687 }
1688
1689 /**
1690  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1691  *
1692  * @lock:       the rt_mutex to be locked
1693  *
1694  * No locking. Caller has to do serializing itself
1695  *
1696  * Special API call for PI-futex support. This merrily cleans up the rtmutex
1697  * (debugging) state. Concurrent operations on this rt_mutex are not
1698  * possible because it belongs to the pi_state which is about to be freed
1699  * and it is not longer visible to other tasks.
1700  */
1701 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1702                            struct task_struct *proxy_owner)
1703 {
1704         debug_rt_mutex_proxy_unlock(lock);
1705         rt_mutex_set_owner(lock, NULL);
1706 }
1707
1708 int __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1709                               struct rt_mutex_waiter *waiter,
1710                               struct task_struct *task)
1711 {
1712         int ret;
1713
1714         if (try_to_take_rt_mutex(lock, task, NULL))
1715                 return 1;
1716
1717         /* We enforce deadlock detection for futexes */
1718         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1719                                       RT_MUTEX_FULL_CHAINWALK);
1720
1721         if (ret && !rt_mutex_owner(lock)) {
1722                 /*
1723                  * Reset the return value. We might have
1724                  * returned with -EDEADLK and the owner
1725                  * released the lock while we were walking the
1726                  * pi chain.  Let the waiter sort it out.
1727                  */
1728                 ret = 0;
1729         }
1730
1731         if (unlikely(ret))
1732                 remove_waiter(lock, waiter);
1733
1734         debug_rt_mutex_print_deadlock(waiter);
1735
1736         return ret;
1737 }
1738
1739 /**
1740  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1741  * @lock:               the rt_mutex to take
1742  * @waiter:             the pre-initialized rt_mutex_waiter
1743  * @task:               the task to prepare
1744  *
1745  * Returns:
1746  *  0 - task blocked on lock
1747  *  1 - acquired the lock for task, caller should wake it up
1748  * <0 - error
1749  *
1750  * Special API call for FUTEX_REQUEUE_PI support.
1751  */
1752 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1753                               struct rt_mutex_waiter *waiter,
1754                               struct task_struct *task)
1755 {
1756         int ret;
1757
1758         raw_spin_lock_irq(&lock->wait_lock);
1759         ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1760         raw_spin_unlock_irq(&lock->wait_lock);
1761
1762         return ret;
1763 }
1764
1765 /**
1766  * rt_mutex_next_owner - return the next owner of the lock
1767  *
1768  * @lock: the rt lock query
1769  *
1770  * Returns the next owner of the lock or NULL
1771  *
1772  * Caller has to serialize against other accessors to the lock
1773  * itself.
1774  *
1775  * Special API call for PI-futex support
1776  */
1777 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1778 {
1779         if (!rt_mutex_has_waiters(lock))
1780                 return NULL;
1781
1782         return rt_mutex_top_waiter(lock)->task;
1783 }
1784
1785 /**
1786  * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1787  * @lock:               the rt_mutex we were woken on
1788  * @to:                 the timeout, null if none. hrtimer should already have
1789  *                      been started.
1790  * @waiter:             the pre-initialized rt_mutex_waiter
1791  *
1792  * Wait for the the lock acquisition started on our behalf by
1793  * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1794  * rt_mutex_cleanup_proxy_lock().
1795  *
1796  * Returns:
1797  *  0 - success
1798  * <0 - error, one of -EINTR, -ETIMEDOUT
1799  *
1800  * Special API call for PI-futex support
1801  */
1802 int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1803                                struct hrtimer_sleeper *to,
1804                                struct rt_mutex_waiter *waiter)
1805 {
1806         int ret;
1807
1808         raw_spin_lock_irq(&lock->wait_lock);
1809         /* sleep on the mutex */
1810         set_current_state(TASK_INTERRUPTIBLE);
1811         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1812         /*
1813          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1814          * have to fix that up.
1815          */
1816         fixup_rt_mutex_waiters(lock);
1817         raw_spin_unlock_irq(&lock->wait_lock);
1818
1819         return ret;
1820 }
1821
1822 /**
1823  * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1824  * @lock:               the rt_mutex we were woken on
1825  * @waiter:             the pre-initialized rt_mutex_waiter
1826  *
1827  * Attempt to clean up after a failed rt_mutex_wait_proxy_lock().
1828  *
1829  * Unless we acquired the lock; we're still enqueued on the wait-list and can
1830  * in fact still be granted ownership until we're removed. Therefore we can
1831  * find we are in fact the owner and must disregard the
1832  * rt_mutex_wait_proxy_lock() failure.
1833  *
1834  * Returns:
1835  *  true  - did the cleanup, we done.
1836  *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1837  *          caller should disregards its return value.
1838  *
1839  * Special API call for PI-futex support
1840  */
1841 bool rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1842                                  struct rt_mutex_waiter *waiter)
1843 {
1844         bool cleanup = false;
1845
1846         raw_spin_lock_irq(&lock->wait_lock);
1847         /*
1848          * Do an unconditional try-lock, this deals with the lock stealing
1849          * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1850          * sets a NULL owner.
1851          *
1852          * We're not interested in the return value, because the subsequent
1853          * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1854          * we will own the lock and it will have removed the waiter. If we
1855          * failed the trylock, we're still not owner and we need to remove
1856          * ourselves.
1857          */
1858         try_to_take_rt_mutex(lock, current, waiter);
1859         /*
1860          * Unless we're the owner; we're still enqueued on the wait_list.
1861          * So check if we became owner, if not, take us off the wait_list.
1862          */
1863         if (rt_mutex_owner(lock) != current) {
1864                 remove_waiter(lock, waiter);
1865                 cleanup = true;
1866         }
1867         /*
1868          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1869          * have to fix that up.
1870          */
1871         fixup_rt_mutex_waiters(lock);
1872
1873         raw_spin_unlock_irq(&lock->wait_lock);
1874
1875         return cleanup;
1876 }