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_node;
 275         struct rb_node *parent = NULL;
 276         struct rt_mutex_waiter *entry;
 277         int leftmost = 1;
 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 = 0;
 287                 }
 288         }
 289
 290         if (leftmost)
 291                 lock->waiters_leftmost = &waiter->tree_entry;
 292
 293         rb_link_node(&waiter->tree_entry, parent, link);
 294         rb_insert_color(&waiter->tree_entry, &lock->waiters);
 295 }
 296
 297 static void
 298 rt_mutex_dequeue(struct rt_mutex *lock, struct rt_mutex_waiter *waiter)
 299 {
 300         if (RB_EMPTY_NODE(&waiter->tree_entry))
 301                 return;
 302
 303         if (lock->waiters_leftmost == &waiter->tree_entry)
 304                 lock->waiters_leftmost = rb_next(&waiter->tree_entry);
 305
 306         rb_erase(&waiter->tree_entry, &lock->waiters);
 307         RB_CLEAR_NODE(&waiter->tree_entry);
 308 }
 309
 310 static void
 311 rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 312 {
 313         struct rb_node **link = &task->pi_waiters.rb_node;
 314         struct rb_node *parent = NULL;
 315         struct rt_mutex_waiter *entry;
 316         int leftmost = 1;
 317
 318         while (*link) {
 319                 parent = *link;
 320                 entry = rb_entry(parent, struct rt_mutex_waiter, pi_tree_entry);
 321                 if (rt_mutex_waiter_less(waiter, entry)) {
 322                         link = &parent->rb_left;
 323                 } else {
 324                         link = &parent->rb_right;
 325                         leftmost = 0;
 326                 }
 327         }
 328
 329         if (leftmost)
 330                 task->pi_waiters_leftmost = &waiter->pi_tree_entry;
 331
 332         rb_link_node(&waiter->pi_tree_entry, parent, link);
 333         rb_insert_color(&waiter->pi_tree_entry, &task->pi_waiters);
 334 }
 335
 336 static void
 337 rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 338 {
 339         if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 340                 return;
 341
 342         if (task->pi_waiters_leftmost == &waiter->pi_tree_entry)
 343                 task->pi_waiters_leftmost = rb_next(&waiter->pi_tree_entry);
 344
 345         rb_erase(&waiter->pi_tree_entry, &task->pi_waiters);
 346         RB_CLEAR_NODE(&waiter->pi_tree_entry);
 347 }
 348
 349 static void rt_mutex_adjust_prio(struct task_struct *p)
 350 {
 351         struct task_struct *pi_task = NULL;
 352
 353         lockdep_assert_held(&p->pi_lock);
 354
 355         if (task_has_pi_waiters(p))
 356                 pi_task = task_top_pi_waiter(p)->task;
 357
 358         rt_mutex_setprio(p, pi_task);
 359 }
 360
 361 /*
 362  * Deadlock detection is conditional:
 363  *
 364  * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
 365  * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
 366  *
 367  * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
 368  * conducted independent of the detect argument.
 369  *
 370  * If the waiter argument is NULL this indicates the deboost path and
 371  * deadlock detection is disabled independent of the detect argument
 372  * and the config settings.
 373  */
 374 static bool rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
 375                                           enum rtmutex_chainwalk chwalk)
 376 {
 377         /*
 378          * This is just a wrapper function for the following call,
 379          * because debug_rt_mutex_detect_deadlock() smells like a magic
 380          * debug feature and I wanted to keep the cond function in the
 381          * main source file along with the comments instead of having
 382          * two of the same in the headers.
 383          */
 384         return debug_rt_mutex_detect_deadlock(waiter, chwalk);
 385 }
 386
 387 /*
 388  * Max number of times we'll walk the boosting chain:
 389  */
 390 int max_lock_depth = 1024;
 391
 392 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 393 {
 394         return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 395 }
 396
 397 /*
 398  * Adjust the priority chain. Also used for deadlock detection.
 399  * Decreases task's usage by one - may thus free the task.
 400  *
 401  * @task:       the task owning the mutex (owner) for which a chain walk is
 402  *              probably needed
 403  * @chwalk:     do we have to carry out deadlock detection?
 404  * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 405  *              things for a task that has just got its priority adjusted, and
 406  *              is waiting on a mutex)
 407  * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 408  *              we dropped its pi_lock. Is never dereferenced, only used for
 409  *              comparison to detect lock chain changes.
 410  * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 411  *              its priority to the mutex owner (can be NULL in the case
 412  *              depicted above or if the top waiter is gone away and we are
 413  *              actually deboosting the owner)
 414  * @top_task:   the current top waiter
 415  *
 416  * Returns 0 or -EDEADLK.
 417  *
 418  * Chain walk basics and protection scope
 419  *
 420  * [R] refcount on task
 421  * [P] task->pi_lock held
 422  * [L] rtmutex->wait_lock held
 423  *
 424  * Step Description                             Protected by
 425  *      function arguments:
 426  *      @task                                   [R]
 427  *      @orig_lock if != NULL                   @top_task is blocked on it
 428  *      @next_lock                              Unprotected. Cannot be
 429  *                                              dereferenced. Only used for
 430  *                                              comparison.
 431  *      @orig_waiter if != NULL                 @top_task is blocked on it
 432  *      @top_task                               current, or in case of proxy
 433  *                                              locking protected by calling
 434  *                                              code
 435  *      again:
 436  *        loop_sanity_check();
 437  *      retry:
 438  * [1]    lock(task->pi_lock);                  [R] acquire [P]
 439  * [2]    waiter = task->pi_blocked_on;         [P]
 440  * [3]    check_exit_conditions_1();            [P]
 441  * [4]    lock = waiter->lock;                  [P]
 442  * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
 443  *          unlock(task->pi_lock);              release [P]
 444  *          goto retry;
 445  *        }
 446  * [6]    check_exit_conditions_2();            [P] + [L]
 447  * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
 448  * [8]    unlock(task->pi_lock);                release [P]
 449  *        put_task_struct(task);                release [R]
 450  * [9]    check_exit_conditions_3();            [L]
 451  * [10]   task = owner(lock);                   [L]
 452  *        get_task_struct(task);                [L] acquire [R]
 453  *        lock(task->pi_lock);                  [L] acquire [P]
 454  * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
 455  * [12]   check_exit_conditions_4();            [P] + [L]
 456  * [13]   unlock(task->pi_lock);                release [P]
 457  *        unlock(lock->wait_lock);              release [L]
 458  *        goto again;
 459  */
 460 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 461                                       enum rtmutex_chainwalk chwalk,
 462                                       struct rt_mutex *orig_lock,
 463                                       struct rt_mutex *next_lock,
 464                                       struct rt_mutex_waiter *orig_waiter,
 465                                       struct task_struct *top_task)
 466 {
 467         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 468         struct rt_mutex_waiter *prerequeue_top_waiter;
 469         int ret = 0, depth = 0;
 470         struct rt_mutex *lock;
 471         bool detect_deadlock;
 472         bool requeue = true;
 473
 474         detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
 475
 476         /*
 477          * The (de)boosting is a step by step approach with a lot of
 478          * pitfalls. We want this to be preemptible and we want hold a
 479          * maximum of two locks per step. So we have to check
 480          * carefully whether things change under us.
 481          */
 482  again:
 483         /*
 484          * We limit the lock chain length for each invocation.
 485          */
 486         if (++depth > max_lock_depth) {
 487                 static int prev_max;
 488
 489                 /*
 490                  * Print this only once. If the admin changes the limit,
 491                  * print a new message when reaching the limit again.
 492                  */
 493                 if (prev_max != max_lock_depth) {
 494                         prev_max = max_lock_depth;
 495                         printk(KERN_WARNING "Maximum lock depth %d reached "
 496                                "task: %s (%d)\n", max_lock_depth,
 497                                top_task->comm, task_pid_nr(top_task));
 498                 }
 499                 put_task_struct(task);
 500
 501                 return -EDEADLK;
 502         }
 503
 504         /*
 505          * We are fully preemptible here and only hold the refcount on
 506          * @task. So everything can have changed under us since the
 507          * caller or our own code below (goto retry/again) dropped all
 508          * locks.
 509          */
 510  retry:
 511         /*
 512          * [1] Task cannot go away as we did a get_task() before !
 513          */
 514         raw_spin_lock_irq(&task->pi_lock);
 515
 516         /*
 517          * [2] Get the waiter on which @task is blocked on.
 518          */
 519         waiter = task->pi_blocked_on;
 520
 521         /*
 522          * [3] check_exit_conditions_1() protected by task->pi_lock.
 523          */
 524
 525         /*
 526          * Check whether the end of the boosting chain has been
 527          * reached or the state of the chain has changed while we
 528          * dropped the locks.
 529          */
 530         if (!waiter)
 531                 goto out_unlock_pi;
 532
 533         /*
 534          * Check the orig_waiter state. After we dropped the locks,
 535          * the previous owner of the lock might have released the lock.
 536          */
 537         if (orig_waiter && !rt_mutex_owner(orig_lock))
 538                 goto out_unlock_pi;
 539
 540         /*
 541          * We dropped all locks after taking a refcount on @task, so
 542          * the task might have moved on in the lock chain or even left
 543          * the chain completely and blocks now on an unrelated lock or
 544          * on @orig_lock.
 545          *
 546          * We stored the lock on which @task was blocked in @next_lock,
 547          * so we can detect the chain change.
 548          */
 549         if (next_lock != waiter->lock)
 550                 goto out_unlock_pi;
 551
 552         /*
 553          * Drop out, when the task has no waiters. Note,
 554          * top_waiter can be NULL, when we are in the deboosting
 555          * mode!
 556          */
 557         if (top_waiter) {
 558                 if (!task_has_pi_waiters(task))
 559                         goto out_unlock_pi;
 560                 /*
 561                  * If deadlock detection is off, we stop here if we
 562                  * are not the top pi waiter of the task. If deadlock
 563                  * detection is enabled we continue, but stop the
 564                  * requeueing in the chain walk.
 565                  */
 566                 if (top_waiter != task_top_pi_waiter(task)) {
 567                         if (!detect_deadlock)
 568                                 goto out_unlock_pi;
 569                         else
 570                                 requeue = false;
 571                 }
 572         }
 573
 574         /*
 575          * If the waiter priority is the same as the task priority
 576          * then there is no further priority adjustment necessary.  If
 577          * deadlock detection is off, we stop the chain walk. If its
 578          * enabled we continue, but stop the requeueing in the chain
 579          * walk.
 580          */
 581         if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
 582                 if (!detect_deadlock)
 583                         goto out_unlock_pi;
 584                 else
 585                         requeue = false;
 586         }
 587
 588         /*
 589          * [4] Get the next lock
 590          */
 591         lock = waiter->lock;
 592         /*
 593          * [5] We need to trylock here as we are holding task->pi_lock,
 594          * which is the reverse lock order versus the other rtmutex
 595          * operations.
 596          */
 597         if (!raw_spin_trylock(&lock->wait_lock)) {
 598                 raw_spin_unlock_irq(&task->pi_lock);
 599                 cpu_relax();
 600                 goto retry;
 601         }
 602
 603         /*
 604          * [6] check_exit_conditions_2() protected by task->pi_lock and
 605          * lock->wait_lock.
 606          *
 607          * Deadlock detection. If the lock is the same as the original
 608          * lock which caused us to walk the lock chain or if the
 609          * current lock is owned by the task which initiated the chain
 610          * walk, we detected a deadlock.
 611          */
 612         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 613                 debug_rt_mutex_deadlock(chwalk, orig_waiter, lock);
 614                 raw_spin_unlock(&lock->wait_lock);
 615                 ret = -EDEADLK;
 616                 goto out_unlock_pi;
 617         }
 618
 619         /*
 620          * If we just follow the lock chain for deadlock detection, no
 621          * need to do all the requeue operations. To avoid a truckload
 622          * of conditionals around the various places below, just do the
 623          * minimum chain walk checks.
 624          */
 625         if (!requeue) {
 626                 /*
 627                  * No requeue[7] here. Just release @task [8]
 628                  */
 629                 raw_spin_unlock(&task->pi_lock);
 630                 put_task_struct(task);
 631
 632                 /*
 633                  * [9] check_exit_conditions_3 protected by lock->wait_lock.
 634                  * If there is no owner of the lock, end of chain.
 635                  */
 636                 if (!rt_mutex_owner(lock)) {
 637                         raw_spin_unlock_irq(&lock->wait_lock);
 638                         return 0;
 639                 }
 640
 641                 /* [10] Grab the next task, i.e. owner of @lock */
 642                 task = rt_mutex_owner(lock);
 643                 get_task_struct(task);
 644                 raw_spin_lock(&task->pi_lock);
 645
 646                 /*
 647                  * No requeue [11] here. We just do deadlock detection.
 648                  *
 649                  * [12] Store whether owner is blocked
 650                  * itself. Decision is made after dropping the locks
 651                  */
 652                 next_lock = task_blocked_on_lock(task);
 653                 /*
 654                  * Get the top waiter for the next iteration
 655                  */
 656                 top_waiter = rt_mutex_top_waiter(lock);
 657
 658                 /* [13] Drop locks */
 659                 raw_spin_unlock(&task->pi_lock);
 660                 raw_spin_unlock_irq(&lock->wait_lock);
 661
 662                 /* If owner is not blocked, end of chain. */
 663                 if (!next_lock)
 664                         goto out_put_task;
 665                 goto again;
 666         }
 667
 668         /*
 669          * Store the current top waiter before doing the requeue
 670          * operation on @lock. We need it for the boost/deboost
 671          * decision below.
 672          */
 673         prerequeue_top_waiter = rt_mutex_top_waiter(lock);
 674
 675         /* [7] Requeue the waiter in the lock waiter tree. */
 676         rt_mutex_dequeue(lock, waiter);
 677
 678         /*
 679          * Update the waiter prio fields now that we're dequeued.
 680          *
 681          * These values can have changed through either:
 682          *
 683          *   sys_sched_set_scheduler() / sys_sched_setattr()
 684          *
 685          * or
 686          *
 687          *   DL CBS enforcement advancing the effective deadline.
 688          *
 689          * Even though pi_waiters also uses these fields, and that tree is only
 690          * updated in [11], we can do this here, since we hold [L], which
 691          * serializes all pi_waiters access and rb_erase() does not care about
 692          * the values of the node being removed.
 693          */
 694         waiter->prio = task->prio;
 695         waiter->deadline = task->dl.deadline;
 696
 697         rt_mutex_enqueue(lock, waiter);
 698
 699         /* [8] Release the task */
 700         raw_spin_unlock(&task->pi_lock);
 701         put_task_struct(task);
 702
 703         /*
 704          * [9] check_exit_conditions_3 protected by lock->wait_lock.
 705          *
 706          * We must abort the chain walk if there is no lock owner even
 707          * in the dead lock detection case, as we have nothing to
 708          * follow here. This is the end of the chain we are walking.
 709          */
 710         if (!rt_mutex_owner(lock)) {
 711                 /*
 712                  * If the requeue [7] above changed the top waiter,
 713                  * then we need to wake the new top waiter up to try
 714                  * to get the lock.
 715                  */
 716                 if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
 717                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 718                 raw_spin_unlock_irq(&lock->wait_lock);
 719                 return 0;
 720         }
 721
 722         /* [10] Grab the next task, i.e. the owner of @lock */
 723         task = rt_mutex_owner(lock);
 724         get_task_struct(task);
 725         raw_spin_lock(&task->pi_lock);
 726
 727         /* [11] requeue the pi waiters if necessary */
 728         if (waiter == rt_mutex_top_waiter(lock)) {
 729                 /*
 730                  * The waiter became the new top (highest priority)
 731                  * waiter on the lock. Replace the previous top waiter
 732                  * in the owner tasks pi waiters tree with this waiter
 733                  * and adjust the priority of the owner.
 734                  */
 735                 rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
 736                 rt_mutex_enqueue_pi(task, waiter);
 737                 rt_mutex_adjust_prio(task);
 738
 739         } else if (prerequeue_top_waiter == waiter) {
 740                 /*
 741                  * The waiter was the top waiter on the lock, but is
 742                  * no longer the top prority waiter. Replace waiter in
 743                  * the owner tasks pi waiters tree with the new top
 744                  * (highest priority) waiter and adjust the priority
 745                  * of the owner.
 746                  * The new top waiter is stored in @waiter so that
 747                  * @waiter == @top_waiter evaluates to true below and
 748                  * we continue to deboost the rest of the chain.
 749                  */
 750                 rt_mutex_dequeue_pi(task, waiter);
 751                 waiter = rt_mutex_top_waiter(lock);
 752                 rt_mutex_enqueue_pi(task, waiter);
 753                 rt_mutex_adjust_prio(task);
 754         } else {
 755                 /*
 756                  * Nothing changed. No need to do any priority
 757                  * adjustment.
 758                  */
 759         }
 760
 761         /*
 762          * [12] check_exit_conditions_4() protected by task->pi_lock
 763          * and lock->wait_lock. The actual decisions are made after we
 764          * dropped the locks.
 765          *
 766          * Check whether the task which owns the current lock is pi
 767          * blocked itself. If yes we store a pointer to the lock for
 768          * the lock chain change detection above. After we dropped
 769          * task->pi_lock next_lock cannot be dereferenced anymore.
 770          */
 771         next_lock = task_blocked_on_lock(task);
 772         /*
 773          * Store the top waiter of @lock for the end of chain walk
 774          * decision below.
 775          */
 776         top_waiter = rt_mutex_top_waiter(lock);
 777
 778         /* [13] Drop the locks */
 779         raw_spin_unlock(&task->pi_lock);
 780         raw_spin_unlock_irq(&lock->wait_lock);
 781
 782         /*
 783          * Make the actual exit decisions [12], based on the stored
 784          * values.
 785          *
 786          * We reached the end of the lock chain. Stop right here. No
 787          * point to go back just to figure that out.
 788          */
 789         if (!next_lock)
 790                 goto out_put_task;
 791
 792         /*
 793          * If the current waiter is not the top waiter on the lock,
 794          * then we can stop the chain walk here if we are not in full
 795          * deadlock detection mode.
 796          */
 797         if (!detect_deadlock && waiter != top_waiter)
 798                 goto out_put_task;
 799
 800         goto again;
 801
 802  out_unlock_pi:
 803         raw_spin_unlock_irq(&task->pi_lock);
 804  out_put_task:
 805         put_task_struct(task);
 806
 807         return ret;
 808 }
 809
 810 /*
 811  * Try to take an rt-mutex
 812  *
 813  * Must be called with lock->wait_lock held and interrupts disabled
 814  *
 815  * @lock:   The lock to be acquired.
 816  * @task:   The task which wants to acquire the lock
 817  * @waiter: The waiter that is queued to the lock's wait tree if the
 818  *          callsite called task_blocked_on_lock(), otherwise NULL
 819  */
 820 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 821                                 struct rt_mutex_waiter *waiter)
 822 {
 823         lockdep_assert_held(&lock->wait_lock);
 824
 825         /*
 826          * Before testing whether we can acquire @lock, we set the
 827          * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
 828          * other tasks which try to modify @lock into the slow path
 829          * and they serialize on @lock->wait_lock.
 830          *
 831          * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
 832          * as explained at the top of this file if and only if:
 833          *
 834          * - There is a lock owner. The caller must fixup the
 835          *   transient state if it does a trylock or leaves the lock
 836          *   function due to a signal or timeout.
 837          *
 838          * - @task acquires the lock and there are no other
 839          *   waiters. This is undone in rt_mutex_set_owner(@task) at
 840          *   the end of this function.
 841          */
 842         mark_rt_mutex_waiters(lock);
 843
 844         /*
 845          * If @lock has an owner, give up.
 846          */
 847         if (rt_mutex_owner(lock))
 848                 return 0;
 849
 850         /*
 851          * If @waiter != NULL, @task has already enqueued the waiter
 852          * into @lock waiter tree. If @waiter == NULL then this is a
 853          * trylock attempt.
 854          */
 855         if (waiter) {
 856                 /*
 857                  * If waiter is not the highest priority waiter of
 858                  * @lock, give up.
 859                  */
 860                 if (waiter != rt_mutex_top_waiter(lock))
 861                         return 0;
 862
 863                 /*
 864                  * We can acquire the lock. Remove the waiter from the
 865                  * lock waiters tree.
 866                  */
 867                 rt_mutex_dequeue(lock, waiter);
 868
 869         } else {
 870                 /*
 871                  * If the lock has waiters already we check whether @task is
 872                  * eligible to take over the lock.
 873                  *
 874                  * If there are no other waiters, @task can acquire
 875                  * the lock.  @task->pi_blocked_on is NULL, so it does
 876                  * not need to be dequeued.
 877                  */
 878                 if (rt_mutex_has_waiters(lock)) {
 879                         /*
 880                          * If @task->prio is greater than or equal to
 881                          * the top waiter priority (kernel view),
 882                          * @task lost.
 883                          */
 884                         if (!rt_mutex_waiter_less(task_to_waiter(task),
 885                                                   rt_mutex_top_waiter(lock)))
 886                                 return 0;
 887
 888                         /*
 889                          * The current top waiter stays enqueued. We
 890                          * don't have to change anything in the lock
 891                          * waiters order.
 892                          */
 893                 } else {
 894                         /*
 895                          * No waiters. Take the lock without the
 896                          * pi_lock dance.@task->pi_blocked_on is NULL
 897                          * and we have no waiters to enqueue in @task
 898                          * pi waiters tree.
 899                          */
 900                         goto takeit;
 901                 }
 902         }
 903
 904         /*
 905          * Clear @task->pi_blocked_on. Requires protection by
 906          * @task->pi_lock. Redundant operation for the @waiter == NULL
 907          * case, but conditionals are more expensive than a redundant
 908          * store.
 909          */
 910         raw_spin_lock(&task->pi_lock);
 911         task->pi_blocked_on = NULL;
 912         /*
 913          * Finish the lock acquisition. @task is the new owner. If
 914          * other waiters exist we have to insert the highest priority
 915          * waiter into @task->pi_waiters tree.
 916          */
 917         if (rt_mutex_has_waiters(lock))
 918                 rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
 919         raw_spin_unlock(&task->pi_lock);
 920
 921 takeit:
 922         /* We got the lock. */
 923         debug_rt_mutex_lock(lock);
 924
 925         /*
 926          * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
 927          * are still waiters or clears it.
 928          */
 929         rt_mutex_set_owner(lock, task);
 930
 931         return 1;
 932 }
 933
 934 /*
 935  * Task blocks on lock.
 936  *
 937  * Prepare waiter and propagate pi chain
 938  *
 939  * This must be called with lock->wait_lock held and interrupts disabled
 940  */
 941 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 942                                    struct rt_mutex_waiter *waiter,
 943                                    struct task_struct *task,
 944                                    enum rtmutex_chainwalk chwalk)
 945 {
 946         struct task_struct *owner = rt_mutex_owner(lock);
 947         struct rt_mutex_waiter *top_waiter = waiter;
 948         struct rt_mutex *next_lock;
 949         int chain_walk = 0, res;
 950
 951         lockdep_assert_held(&lock->wait_lock);
 952
 953         /*
 954          * Early deadlock detection. We really don't want the task to
 955          * enqueue on itself just to untangle the mess later. It's not
 956          * only an optimization. We drop the locks, so another waiter
 957          * can come in before the chain walk detects the deadlock. So
 958          * the other will detect the deadlock and return -EDEADLOCK,
 959          * which is wrong, as the other waiter is not in a deadlock
 960          * situation.
 961          */
 962         if (owner == task)
 963                 return -EDEADLK;
 964
 965         raw_spin_lock(&task->pi_lock);
 966         waiter->task = task;
 967         waiter->lock = lock;
 968         waiter->prio = task->prio;
 969         waiter->deadline = task->dl.deadline;
 970
 971         /* Get the top priority waiter on the lock */
 972         if (rt_mutex_has_waiters(lock))
 973                 top_waiter = rt_mutex_top_waiter(lock);
 974         rt_mutex_enqueue(lock, waiter);
 975
 976         task->pi_blocked_on = waiter;
 977
 978         raw_spin_unlock(&task->pi_lock);
 979
 980         if (!owner)
 981                 return 0;
 982
 983         raw_spin_lock(&owner->pi_lock);
 984         if (waiter == rt_mutex_top_waiter(lock)) {
 985                 rt_mutex_dequeue_pi(owner, top_waiter);
 986                 rt_mutex_enqueue_pi(owner, waiter);
 987
 988                 rt_mutex_adjust_prio(owner);
 989                 if (owner->pi_blocked_on)
 990                         chain_walk = 1;
 991         } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
 992                 chain_walk = 1;
 993         }
 994
 995         /* Store the lock on which owner is blocked or NULL */
 996         next_lock = task_blocked_on_lock(owner);
 997
 998         raw_spin_unlock(&owner->pi_lock);
 999         /*
1000          * Even if full deadlock detection is on, if the owner is not
1001          * blocked itself, we can avoid finding this out in the chain
1002          * walk.
1003          */
1004         if (!chain_walk || !next_lock)
1005                 return 0;
1006
1007         /*
1008          * The owner can't disappear while holding a lock,
1009          * so the owner struct is protected by wait_lock.
1010          * Gets dropped in rt_mutex_adjust_prio_chain()!
1011          */
1012         get_task_struct(owner);
1013
1014         raw_spin_unlock_irq(&lock->wait_lock);
1015
1016         res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1017                                          next_lock, waiter, task);
1018
1019         raw_spin_lock_irq(&lock->wait_lock);
1020
1021         return res;
1022 }
1023
1024 /*
1025  * Remove the top waiter from the current tasks pi waiter tree and
1026  * queue it up.
1027  *
1028  * Called with lock->wait_lock held and interrupts disabled.
1029  */
1030 static void mark_wakeup_next_waiter(struct wake_q_head *wake_q,
1031                                     struct rt_mutex *lock)
1032 {
1033         struct rt_mutex_waiter *waiter;
1034
1035         raw_spin_lock(&current->pi_lock);
1036
1037         waiter = rt_mutex_top_waiter(lock);
1038
1039         /*
1040          * Remove it from current->pi_waiters and deboost.
1041          *
1042          * We must in fact deboost here in order to ensure we call
1043          * rt_mutex_setprio() to update p->pi_top_task before the
1044          * task unblocks.
1045          */
1046         rt_mutex_dequeue_pi(current, waiter);
1047         rt_mutex_adjust_prio(current);
1048
1049         /*
1050          * As we are waking up the top waiter, and the waiter stays
1051          * queued on the lock until it gets the lock, this lock
1052          * obviously has waiters. Just set the bit here and this has
1053          * the added benefit of forcing all new tasks into the
1054          * slow path making sure no task of lower priority than
1055          * the top waiter can steal this lock.
1056          */
1057         lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1058
1059         /*
1060          * We deboosted before waking the top waiter task such that we don't
1061          * run two tasks with the 'same' priority (and ensure the
1062          * p->pi_top_task pointer points to a blocked task). This however can
1063          * lead to priority inversion if we would get preempted after the
1064          * deboost but before waking our donor task, hence the preempt_disable()
1065          * before unlock.
1066          *
1067          * Pairs with preempt_enable() in rt_mutex_postunlock();
1068          */
1069         preempt_disable();
1070         wake_q_add(wake_q, waiter->task);
1071         raw_spin_unlock(&current->pi_lock);
1072 }
1073
1074 /*
1075  * Remove a waiter from a lock and give up
1076  *
1077  * Must be called with lock->wait_lock held and interrupts disabled. I must
1078  * have just failed to try_to_take_rt_mutex().
1079  */
1080 static void remove_waiter(struct rt_mutex *lock,
1081                           struct rt_mutex_waiter *waiter)
1082 {
1083         bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1084         struct task_struct *owner = rt_mutex_owner(lock);
1085         struct rt_mutex *next_lock;
1086
1087         lockdep_assert_held(&lock->wait_lock);
1088
1089         raw_spin_lock(&current->pi_lock);
1090         rt_mutex_dequeue(lock, waiter);
1091         current->pi_blocked_on = NULL;
1092         raw_spin_unlock(&current->pi_lock);
1093
1094         /*
1095          * Only update priority if the waiter was the highest priority
1096          * waiter of the lock and there is an owner to update.
1097          */
1098         if (!owner || !is_top_waiter)
1099                 return;
1100
1101         raw_spin_lock(&owner->pi_lock);
1102
1103         rt_mutex_dequeue_pi(owner, waiter);
1104
1105         if (rt_mutex_has_waiters(lock))
1106                 rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1107
1108         rt_mutex_adjust_prio(owner);
1109
1110         /* Store the lock on which owner is blocked or NULL */
1111         next_lock = task_blocked_on_lock(owner);
1112
1113         raw_spin_unlock(&owner->pi_lock);
1114
1115         /*
1116          * Don't walk the chain, if the owner task is not blocked
1117          * itself.
1118          */
1119         if (!next_lock)
1120                 return;
1121
1122         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1123         get_task_struct(owner);
1124
1125         raw_spin_unlock_irq(&lock->wait_lock);
1126
1127         rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1128                                    next_lock, NULL, current);
1129
1130         raw_spin_lock_irq(&lock->wait_lock);
1131 }
1132
1133 /*
1134  * Recheck the pi chain, in case we got a priority setting
1135  *
1136  * Called from sched_setscheduler
1137  */
1138 void rt_mutex_adjust_pi(struct task_struct *task)
1139 {
1140         struct rt_mutex_waiter *waiter;
1141         struct rt_mutex *next_lock;
1142         unsigned long flags;
1143
1144         raw_spin_lock_irqsave(&task->pi_lock, flags);
1145
1146         waiter = task->pi_blocked_on;
1147         if (!waiter || rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
1148                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1149                 return;
1150         }
1151         next_lock = waiter->lock;
1152         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
1153
1154         /* gets dropped in rt_mutex_adjust_prio_chain()! */
1155         get_task_struct(task);
1156
1157         rt_mutex_adjust_prio_chain(task, RT_MUTEX_MIN_CHAINWALK, NULL,
1158                                    next_lock, NULL, task);
1159 }
1160
1161 void rt_mutex_init_waiter(struct rt_mutex_waiter *waiter)
1162 {
1163         debug_rt_mutex_init_waiter(waiter);
1164         RB_CLEAR_NODE(&waiter->pi_tree_entry);
1165         RB_CLEAR_NODE(&waiter->tree_entry);
1166         waiter->task = NULL;
1167 }
1168
1169 /**
1170  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
1171  * @lock:                the rt_mutex to take
1172  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1173  *                       or TASK_UNINTERRUPTIBLE)
1174  * @timeout:             the pre-initialized and started timer, or NULL for none
1175  * @waiter:              the pre-initialized rt_mutex_waiter
1176  *
1177  * Must be called with lock->wait_lock held and interrupts disabled
1178  */
1179 static int __sched
1180 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
1181                     struct hrtimer_sleeper *timeout,
1182                     struct rt_mutex_waiter *waiter)
1183 {
1184         int ret = 0;
1185
1186         for (;;) {
1187                 /* Try to acquire the lock: */
1188                 if (try_to_take_rt_mutex(lock, current, waiter))
1189                         break;
1190
1191                 /*
1192                  * TASK_INTERRUPTIBLE checks for signals and
1193                  * timeout. Ignored otherwise.
1194                  */
1195                 if (likely(state == TASK_INTERRUPTIBLE)) {
1196                         /* Signal pending? */
1197                         if (signal_pending(current))
1198                                 ret = -EINTR;
1199                         if (timeout && !timeout->task)
1200                                 ret = -ETIMEDOUT;
1201                         if (ret)
1202                                 break;
1203                 }
1204
1205                 raw_spin_unlock_irq(&lock->wait_lock);
1206
1207                 debug_rt_mutex_print_deadlock(waiter);
1208
1209                 schedule();
1210
1211                 raw_spin_lock_irq(&lock->wait_lock);
1212                 set_current_state(state);
1213         }
1214
1215         __set_current_state(TASK_RUNNING);
1216         return ret;
1217 }
1218
1219 static void rt_mutex_handle_deadlock(int res, int detect_deadlock,
1220                                      struct rt_mutex_waiter *w)
1221 {
1222         /*
1223          * If the result is not -EDEADLOCK or the caller requested
1224          * deadlock detection, nothing to do here.
1225          */
1226         if (res != -EDEADLOCK || detect_deadlock)
1227                 return;
1228
1229         /*
1230          * Yell lowdly and stop the task right here.
1231          */
1232         rt_mutex_print_deadlock(w);
1233         while (1) {
1234                 set_current_state(TASK_INTERRUPTIBLE);
1235                 schedule();
1236         }
1237 }
1238
1239 /*
1240  * Slow path lock function:
1241  */
1242 static int __sched
1243 rt_mutex_slowlock(struct rt_mutex *lock, int state,
1244                   struct hrtimer_sleeper *timeout,
1245                   enum rtmutex_chainwalk chwalk)
1246 {
1247         struct rt_mutex_waiter waiter;
1248         unsigned long flags;
1249         int ret = 0;
1250
1251         rt_mutex_init_waiter(&waiter);
1252
1253         /*
1254          * Technically we could use raw_spin_[un]lock_irq() here, but this can
1255          * be called in early boot if the cmpxchg() fast path is disabled
1256          * (debug, no architecture support). In this case we will acquire the
1257          * rtmutex with lock->wait_lock held. But we cannot unconditionally
1258          * enable interrupts in that early boot case. So we need to use the
1259          * irqsave/restore variants.
1260          */
1261         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1262
1263         /* Try to acquire the lock again: */
1264         if (try_to_take_rt_mutex(lock, current, NULL)) {
1265                 raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1266                 return 0;
1267         }
1268
1269         set_current_state(state);
1270
1271         /* Setup the timer, when timeout != NULL */
1272         if (unlikely(timeout))
1273                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1274
1275         ret = task_blocks_on_rt_mutex(lock, &waiter, current, chwalk);
1276
1277         if (likely(!ret))
1278                 /* sleep on the mutex */
1279                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
1280
1281         if (unlikely(ret)) {
1282                 __set_current_state(TASK_RUNNING);
1283                 if (rt_mutex_has_waiters(lock))
1284                         remove_waiter(lock, &waiter);
1285                 rt_mutex_handle_deadlock(ret, chwalk, &waiter);
1286         }
1287
1288         /*
1289          * try_to_take_rt_mutex() sets the waiter bit
1290          * unconditionally. We might have to fix that up.
1291          */
1292         fixup_rt_mutex_waiters(lock);
1293
1294         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1295
1296         /* Remove pending timer: */
1297         if (unlikely(timeout))
1298                 hrtimer_cancel(&timeout->timer);
1299
1300         debug_rt_mutex_free_waiter(&waiter);
1301
1302         return ret;
1303 }
1304
1305 /*
1306  * Slow path try-lock function:
1307  */
1308 static inline int rt_mutex_slowtrylock(struct rt_mutex *lock)
1309 {
1310         unsigned long flags;
1311         int ret;
1312
1313         /*
1314          * If the lock already has an owner we fail to get the lock.
1315          * This can be done without taking the @lock->wait_lock as
1316          * it is only being read, and this is a trylock anyway.
1317          */
1318         if (rt_mutex_owner(lock))
1319                 return 0;
1320
1321         /*
1322          * The mutex has currently no owner. Lock the wait lock and try to
1323          * acquire the lock. We use irqsave here to support early boot calls.
1324          */
1325         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1326
1327         ret = try_to_take_rt_mutex(lock, current, NULL);
1328
1329         /*
1330          * try_to_take_rt_mutex() sets the lock waiters bit
1331          * unconditionally. Clean this up.
1332          */
1333         fixup_rt_mutex_waiters(lock);
1334
1335         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1336
1337         return ret;
1338 }
1339
1340 /*
1341  * Slow path to release a rt-mutex.
1342  *
1343  * Return whether the current task needs to call rt_mutex_postunlock().
1344  */
1345 static bool __sched rt_mutex_slowunlock(struct rt_mutex *lock,
1346                                         struct wake_q_head *wake_q)
1347 {
1348         unsigned long flags;
1349
1350         /* irqsave required to support early boot calls */
1351         raw_spin_lock_irqsave(&lock->wait_lock, flags);
1352
1353         debug_rt_mutex_unlock(lock);
1354
1355         /*
1356          * We must be careful here if the fast path is enabled. If we
1357          * have no waiters queued we cannot set owner to NULL here
1358          * because of:
1359          *
1360          * foo->lock->owner = NULL;
1361          *                      rtmutex_lock(foo->lock);   <- fast path
1362          *                      free = atomic_dec_and_test(foo->refcnt);
1363          *                      rtmutex_unlock(foo->lock); <- fast path
1364          *                      if (free)
1365          *                              kfree(foo);
1366          * raw_spin_unlock(foo->lock->wait_lock);
1367          *
1368          * So for the fastpath enabled kernel:
1369          *
1370          * Nothing can set the waiters bit as long as we hold
1371          * lock->wait_lock. So we do the following sequence:
1372          *
1373          *      owner = rt_mutex_owner(lock);
1374          *      clear_rt_mutex_waiters(lock);
1375          *      raw_spin_unlock(&lock->wait_lock);
1376          *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1377          *              return;
1378          *      goto retry;
1379          *
1380          * The fastpath disabled variant is simple as all access to
1381          * lock->owner is serialized by lock->wait_lock:
1382          *
1383          *      lock->owner = NULL;
1384          *      raw_spin_unlock(&lock->wait_lock);
1385          */
1386         while (!rt_mutex_has_waiters(lock)) {
1387                 /* Drops lock->wait_lock ! */
1388                 if (unlock_rt_mutex_safe(lock, flags) == true)
1389                         return false;
1390                 /* Relock the rtmutex and try again */
1391                 raw_spin_lock_irqsave(&lock->wait_lock, flags);
1392         }
1393
1394         /*
1395          * The wakeup next waiter path does not suffer from the above
1396          * race. See the comments there.
1397          *
1398          * Queue the next waiter for wakeup once we release the wait_lock.
1399          */
1400         mark_wakeup_next_waiter(wake_q, lock);
1401         raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1402
1403         return true; /* call rt_mutex_postunlock() */
1404 }
1405
1406 /*
1407  * debug aware fast / slowpath lock,trylock,unlock
1408  *
1409  * The atomic acquire/release ops are compiled away, when either the
1410  * architecture does not support cmpxchg or when debugging is enabled.
1411  */
1412 static inline int
1413 rt_mutex_fastlock(struct rt_mutex *lock, int state,
1414                   int (*slowfn)(struct rt_mutex *lock, int state,
1415                                 struct hrtimer_sleeper *timeout,
1416                                 enum rtmutex_chainwalk chwalk))
1417 {
1418         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1419                 return 0;
1420
1421         return slowfn(lock, state, NULL, RT_MUTEX_MIN_CHAINWALK);
1422 }
1423
1424 static inline int
1425 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
1426                         struct hrtimer_sleeper *timeout,
1427                         enum rtmutex_chainwalk chwalk,
1428                         int (*slowfn)(struct rt_mutex *lock, int state,
1429                                       struct hrtimer_sleeper *timeout,
1430                                       enum rtmutex_chainwalk chwalk))
1431 {
1432         if (chwalk == RT_MUTEX_MIN_CHAINWALK &&
1433             likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1434                 return 0;
1435
1436         return slowfn(lock, state, timeout, chwalk);
1437 }
1438
1439 static inline int
1440 rt_mutex_fasttrylock(struct rt_mutex *lock,
1441                      int (*slowfn)(struct rt_mutex *lock))
1442 {
1443         if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1444                 return 1;
1445
1446         return slowfn(lock);
1447 }
1448
1449 /*
1450  * Performs the wakeup of the the top-waiter and re-enables preemption.
1451  */
1452 void rt_mutex_postunlock(struct wake_q_head *wake_q)
1453 {
1454         wake_up_q(wake_q);
1455
1456         /* Pairs with preempt_disable() in rt_mutex_slowunlock() */
1457         preempt_enable();
1458 }
1459
1460 static inline void
1461 rt_mutex_fastunlock(struct rt_mutex *lock,
1462                     bool (*slowfn)(struct rt_mutex *lock,
1463                                    struct wake_q_head *wqh))
1464 {
1465         DEFINE_WAKE_Q(wake_q);
1466
1467         if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1468                 return;
1469
1470         if (slowfn(lock, &wake_q))
1471                 rt_mutex_postunlock(&wake_q);
1472 }
1473
1474 /**
1475  * rt_mutex_lock - lock a rt_mutex
1476  *
1477  * @lock: the rt_mutex to be locked
1478  */
1479 void __sched rt_mutex_lock(struct rt_mutex *lock)
1480 {
1481         might_sleep();
1482
1483         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1484         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, rt_mutex_slowlock);
1485 }
1486 EXPORT_SYMBOL_GPL(rt_mutex_lock);
1487
1488 /**
1489  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
1490  *
1491  * @lock:               the rt_mutex to be locked
1492  *
1493  * Returns:
1494  *  0           on success
1495  * -EINTR       when interrupted by a signal
1496  */
1497 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock)
1498 {
1499         int ret;
1500
1501         might_sleep();
1502
1503         mutex_acquire(&lock->dep_map, 0, 0, _RET_IP_);
1504         ret = rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE, rt_mutex_slowlock);
1505         if (ret)
1506                 mutex_release(&lock->dep_map, 1, _RET_IP_);
1507
1508         return ret;
1509 }
1510 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
1511
1512 /*
1513  * Futex variant, must not use fastpath.
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;
1661         lock->waiters_leftmost = NULL;
1662
1663         if (name && key)
1664                 debug_rt_mutex_init(lock, name, key);
1665 }
1666 EXPORT_SYMBOL_GPL(__rt_mutex_init);
1667
1668 /**
1669  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
1670  *                              proxy owner
1671  *
1672  * @lock:       the rt_mutex to be locked
1673  * @proxy_owner:the task to set as owner
1674  *
1675  * No locking. Caller has to do serializing itself
1676  *
1677  * Special API call for PI-futex support. This initializes the rtmutex and
1678  * assigns it to @proxy_owner. Concurrent operations on the rtmutex are not
1679  * possible at this point because the pi_state which contains the rtmutex
1680  * is not yet visible to other tasks.
1681  */
1682 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
1683                                 struct task_struct *proxy_owner)
1684 {
1685         __rt_mutex_init(lock, NULL, NULL);
1686         debug_rt_mutex_proxy_lock(lock, proxy_owner);
1687         rt_mutex_set_owner(lock, proxy_owner);
1688 }
1689
1690 /**
1691  * rt_mutex_proxy_unlock - release a lock on behalf of owner
1692  *
1693  * @lock:       the rt_mutex to be locked
1694  *
1695  * No locking. Caller has to do serializing itself
1696  *
1697  * Special API call for PI-futex support. This merrily cleans up the rtmutex
1698  * (debugging) state. Concurrent operations on this rt_mutex are not
1699  * possible because it belongs to the pi_state which is about to be freed
1700  * and it is not longer visible to other tasks.
1701  */
1702 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
1703                            struct task_struct *proxy_owner)
1704 {
1705         debug_rt_mutex_proxy_unlock(lock);
1706         rt_mutex_set_owner(lock, NULL);
1707 }
1708
1709 int __rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1710                               struct rt_mutex_waiter *waiter,
1711                               struct task_struct *task)
1712 {
1713         int ret;
1714
1715         if (try_to_take_rt_mutex(lock, task, NULL))
1716                 return 1;
1717
1718         /* We enforce deadlock detection for futexes */
1719         ret = task_blocks_on_rt_mutex(lock, waiter, task,
1720                                       RT_MUTEX_FULL_CHAINWALK);
1721
1722         if (ret && !rt_mutex_owner(lock)) {
1723                 /*
1724                  * Reset the return value. We might have
1725                  * returned with -EDEADLK and the owner
1726                  * released the lock while we were walking the
1727                  * pi chain.  Let the waiter sort it out.
1728                  */
1729                 ret = 0;
1730         }
1731
1732         if (unlikely(ret))
1733                 remove_waiter(lock, waiter);
1734
1735         debug_rt_mutex_print_deadlock(waiter);
1736
1737         return ret;
1738 }
1739
1740 /**
1741  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
1742  * @lock:               the rt_mutex to take
1743  * @waiter:             the pre-initialized rt_mutex_waiter
1744  * @task:               the task to prepare
1745  *
1746  * Returns:
1747  *  0 - task blocked on lock
1748  *  1 - acquired the lock for task, caller should wake it up
1749  * <0 - error
1750  *
1751  * Special API call for FUTEX_REQUEUE_PI support.
1752  */
1753 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
1754                               struct rt_mutex_waiter *waiter,
1755                               struct task_struct *task)
1756 {
1757         int ret;
1758
1759         raw_spin_lock_irq(&lock->wait_lock);
1760         ret = __rt_mutex_start_proxy_lock(lock, waiter, task);
1761         raw_spin_unlock_irq(&lock->wait_lock);
1762
1763         return ret;
1764 }
1765
1766 /**
1767  * rt_mutex_next_owner - return the next owner of the lock
1768  *
1769  * @lock: the rt lock query
1770  *
1771  * Returns the next owner of the lock or NULL
1772  *
1773  * Caller has to serialize against other accessors to the lock
1774  * itself.
1775  *
1776  * Special API call for PI-futex support
1777  */
1778 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1779 {
1780         if (!rt_mutex_has_waiters(lock))
1781                 return NULL;
1782
1783         return rt_mutex_top_waiter(lock)->task;
1784 }
1785
1786 /**
1787  * rt_mutex_wait_proxy_lock() - Wait for lock acquisition
1788  * @lock:               the rt_mutex we were woken on
1789  * @to:                 the timeout, null if none. hrtimer should already have
1790  *                      been started.
1791  * @waiter:             the pre-initialized rt_mutex_waiter
1792  *
1793  * Wait for the the lock acquisition started on our behalf by
1794  * rt_mutex_start_proxy_lock(). Upon failure, the caller must call
1795  * rt_mutex_cleanup_proxy_lock().
1796  *
1797  * Returns:
1798  *  0 - success
1799  * <0 - error, one of -EINTR, -ETIMEDOUT
1800  *
1801  * Special API call for PI-futex support
1802  */
1803 int rt_mutex_wait_proxy_lock(struct rt_mutex *lock,
1804                                struct hrtimer_sleeper *to,
1805                                struct rt_mutex_waiter *waiter)
1806 {
1807         int ret;
1808
1809         raw_spin_lock_irq(&lock->wait_lock);
1810         /* sleep on the mutex */
1811         set_current_state(TASK_INTERRUPTIBLE);
1812         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1813         /*
1814          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1815          * have to fix that up.
1816          */
1817         fixup_rt_mutex_waiters(lock);
1818         raw_spin_unlock_irq(&lock->wait_lock);
1819
1820         return ret;
1821 }
1822
1823 /**
1824  * rt_mutex_cleanup_proxy_lock() - Cleanup failed lock acquisition
1825  * @lock:               the rt_mutex we were woken on
1826  * @waiter:             the pre-initialized rt_mutex_waiter
1827  *
1828  * Attempt to clean up after a failed rt_mutex_wait_proxy_lock().
1829  *
1830  * Unless we acquired the lock; we're still enqueued on the wait-list and can
1831  * in fact still be granted ownership until we're removed. Therefore we can
1832  * find we are in fact the owner and must disregard the
1833  * rt_mutex_wait_proxy_lock() failure.
1834  *
1835  * Returns:
1836  *  true  - did the cleanup, we done.
1837  *  false - we acquired the lock after rt_mutex_wait_proxy_lock() returned,
1838  *          caller should disregards its return value.
1839  *
1840  * Special API call for PI-futex support
1841  */
1842 bool rt_mutex_cleanup_proxy_lock(struct rt_mutex *lock,
1843                                  struct rt_mutex_waiter *waiter)
1844 {
1845         bool cleanup = false;
1846
1847         raw_spin_lock_irq(&lock->wait_lock);
1848         /*
1849          * Do an unconditional try-lock, this deals with the lock stealing
1850          * state where __rt_mutex_futex_unlock() -> mark_wakeup_next_waiter()
1851          * sets a NULL owner.
1852          *
1853          * We're not interested in the return value, because the subsequent
1854          * test on rt_mutex_owner() will infer that. If the trylock succeeded,
1855          * we will own the lock and it will have removed the waiter. If we
1856          * failed the trylock, we're still not owner and we need to remove
1857          * ourselves.
1858          */
1859         try_to_take_rt_mutex(lock, current, waiter);
1860         /*
1861          * Unless we're the owner; we're still enqueued on the wait_list.
1862          * So check if we became owner, if not, take us off the wait_list.
1863          */
1864         if (rt_mutex_owner(lock) != current) {
1865                 remove_waiter(lock, waiter);
1866                 cleanup = true;
1867         }
1868         /*
1869          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1870          * have to fix that up.
1871          */
1872         fixup_rt_mutex_waiters(lock);
1873
1874         raw_spin_unlock_irq(&lock->wait_lock);
1875
1876         return cleanup;
1877 }