kernel/rtmutex.c

   1 /*
   2  * RT-Mutexes: simple blocking mutual exclusion locks with PI support
   3  *
   4  * started by Ingo Molnar and Thomas Gleixner.
   5  *
   6  *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   7  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
   8  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
   9  *  Copyright (C) 2006 Esben Nielsen
  10  *
  11  *  See Documentation/rt-mutex-design.txt for details.
  12  */
  13 #include <linux/spinlock.h>
  14 #include <linux/export.h>
  15 #include <linux/sched.h>
  16 #include <linux/sched/rt.h>
  17 #include <linux/timer.h>
  18
  19 #include "rtmutex_common.h"
  20
  21 /*
  22  * lock->owner state tracking:
  23  *
  24  * lock->owner holds the task_struct pointer of the owner. Bit 0
  25  * is used to keep track of the "lock has waiters" state.
  26  *
  27  * owner        bit0
  28  * NULL         0       lock is free (fast acquire possible)
  29  * NULL         1       lock is free and has waiters and the top waiter
  30  *                              is going to take the lock*
  31  * taskpointer  0       lock is held (fast release possible)
  32  * taskpointer  1       lock is held and has waiters**
  33  *
  34  * The fast atomic compare exchange based acquire and release is only
  35  * possible when bit 0 of lock->owner is 0.
  36  *
  37  * (*) It also can be a transitional state when grabbing the lock
  38  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  39  * we need to set the bit0 before looking at the lock, and the owner may be
  40  * NULL in this small time, hence this can be a transitional state.
  41  *
  42  * (**) There is a small time when bit 0 is set but there are no
  43  * waiters. This can happen when grabbing the lock in the slow path.
  44  * To prevent a cmpxchg of the owner releasing the lock, we need to
  45  * set this bit before looking at the lock.
  46  */
  47
  48 static void
  49 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
  50 {
  51         unsigned long val = (unsigned long)owner;
  52
  53         if (rt_mutex_has_waiters(lock))
  54                 val |= RT_MUTEX_HAS_WAITERS;
  55
  56         lock->owner = (struct task_struct *)val;
  57 }
  58
  59 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  60 {
  61         lock->owner = (struct task_struct *)
  62                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  63 }
  64
  65 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  66 {
  67         if (!rt_mutex_has_waiters(lock))
  68                 clear_rt_mutex_waiters(lock);
  69 }
  70
  71 /*
  72  * We can speed up the acquire/release, if the architecture
  73  * supports cmpxchg and if there's no debugging state to be set up
  74  */
  75 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
  76 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
  77 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  78 {
  79         unsigned long owner, *p = (unsigned long *) &lock->owner;
  80
  81         do {
  82                 owner = *p;
  83         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
  84 }
  85 #else
  86 # define rt_mutex_cmpxchg(l,c,n)        (0)
  87 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  88 {
  89         lock->owner = (struct task_struct *)
  90                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
  91 }
  92 #endif
  93
  94 /*
  95  * Calculate task priority from the waiter list priority
  96  *
  97  * Return task->normal_prio when the waiter list is empty or when
  98  * the waiter is not allowed to do priority boosting
  99  */
 100 int rt_mutex_getprio(struct task_struct *task)
 101 {
 102         if (likely(!task_has_pi_waiters(task)))
 103                 return task->normal_prio;
 104
 105         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
 106                    task->normal_prio);
 107 }
 108
 109 /*
 110  * Adjust the priority of a task, after its pi_waiters got modified.
 111  *
 112  * This can be both boosting and unboosting. task->pi_lock must be held.
 113  */
 114 static void __rt_mutex_adjust_prio(struct task_struct *task)
 115 {
 116         int prio = rt_mutex_getprio(task);
 117
 118         if (task->prio != prio)
 119                 rt_mutex_setprio(task, prio);
 120 }
 121
 122 /*
 123  * Adjust task priority (undo boosting). Called from the exit path of
 124  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 125  *
 126  * (Note: We do this outside of the protection of lock->wait_lock to
 127  * allow the lock to be taken while or before we readjust the priority
 128  * of task. We do not use the spin_xx_mutex() variants here as we are
 129  * outside of the debug path.)
 130  */
 131 static void rt_mutex_adjust_prio(struct task_struct *task)
 132 {
 133         unsigned long flags;
 134
 135         raw_spin_lock_irqsave(&task->pi_lock, flags);
 136         __rt_mutex_adjust_prio(task);
 137         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 138 }
 139
 140 /*
 141  * Max number of times we'll walk the boosting chain:
 142  */
 143 int max_lock_depth = 1024;
 144
 145 /*
 146  * Adjust the priority chain. Also used for deadlock detection.
 147  * Decreases task's usage by one - may thus free the task.
 148  * Returns 0 or -EDEADLK.
 149  */
 150 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 151                                       int deadlock_detect,
 152                                       struct rt_mutex *orig_lock,
 153                                       struct rt_mutex_waiter *orig_waiter,
 154                                       struct task_struct *top_task)
 155 {
 156         struct rt_mutex *lock;
 157         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 158         int detect_deadlock, ret = 0, depth = 0;
 159         unsigned long flags;
 160
 161         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 162                                                          deadlock_detect);
 163
 164         /*
 165          * The (de)boosting is a step by step approach with a lot of
 166          * pitfalls. We want this to be preemptible and we want hold a
 167          * maximum of two locks per step. So we have to check
 168          * carefully whether things change under us.
 169          */
 170  again:
 171         if (++depth > max_lock_depth) {
 172                 static int prev_max;
 173
 174                 /*
 175                  * Print this only once. If the admin changes the limit,
 176                  * print a new message when reaching the limit again.
 177                  */
 178                 if (prev_max != max_lock_depth) {
 179                         prev_max = max_lock_depth;
 180                         printk(KERN_WARNING "Maximum lock depth %d reached "
 181                                "task: %s (%d)\n", max_lock_depth,
 182                                top_task->comm, task_pid_nr(top_task));
 183                 }
 184                 put_task_struct(task);
 185
 186                 return deadlock_detect ? -EDEADLK : 0;
 187         }
 188  retry:
 189         /*
 190          * Task can not go away as we did a get_task() before !
 191          */
 192         raw_spin_lock_irqsave(&task->pi_lock, flags);
 193
 194         waiter = task->pi_blocked_on;
 195         /*
 196          * Check whether the end of the boosting chain has been
 197          * reached or the state of the chain has changed while we
 198          * dropped the locks.
 199          */
 200         if (!waiter)
 201                 goto out_unlock_pi;
 202
 203         /*
 204          * Check the orig_waiter state. After we dropped the locks,
 205          * the previous owner of the lock might have released the lock.
 206          */
 207         if (orig_waiter && !rt_mutex_owner(orig_lock))
 208                 goto out_unlock_pi;
 209
 210         /*
 211          * Drop out, when the task has no waiters. Note,
 212          * top_waiter can be NULL, when we are in the deboosting
 213          * mode!
 214          */
 215         if (top_waiter) {
 216                 if (!task_has_pi_waiters(task))
 217                         goto out_unlock_pi;
 218                 /*
 219                  * If deadlock detection is off, we stop here if we
 220                  * are not the top pi waiter of the task.
 221                  */
 222                 if (!detect_deadlock && top_waiter != task_top_pi_waiter(task))
 223                         goto out_unlock_pi;
 224         }
 225
 226         /*
 227          * When deadlock detection is off then we check, if further
 228          * priority adjustment is necessary.
 229          */
 230         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 231                 goto out_unlock_pi;
 232
 233         lock = waiter->lock;
 234         if (!raw_spin_trylock(&lock->wait_lock)) {
 235                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 236                 cpu_relax();
 237                 goto retry;
 238         }
 239
 240         /*
 241          * Deadlock detection. If the lock is the same as the original
 242          * lock which caused us to walk the lock chain or if the
 243          * current lock is owned by the task which initiated the chain
 244          * walk, we detected a deadlock.
 245          */
 246         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 247                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 248                 raw_spin_unlock(&lock->wait_lock);
 249                 ret = deadlock_detect ? -EDEADLK : 0;
 250                 goto out_unlock_pi;
 251         }
 252
 253         top_waiter = rt_mutex_top_waiter(lock);
 254
 255         /* Requeue the waiter */
 256         plist_del(&waiter->list_entry, &lock->wait_list);
 257         waiter->list_entry.prio = task->prio;
 258         plist_add(&waiter->list_entry, &lock->wait_list);
 259
 260         /* Release the task */
 261         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 262         if (!rt_mutex_owner(lock)) {
 263                 /*
 264                  * If the requeue above changed the top waiter, then we need
 265                  * to wake the new top waiter up to try to get the lock.
 266                  */
 267
 268                 if (top_waiter != rt_mutex_top_waiter(lock))
 269                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 270                 raw_spin_unlock(&lock->wait_lock);
 271                 goto out_put_task;
 272         }
 273         put_task_struct(task);
 274
 275         /* Grab the next task */
 276         task = rt_mutex_owner(lock);
 277         get_task_struct(task);
 278         raw_spin_lock_irqsave(&task->pi_lock, flags);
 279
 280         if (waiter == rt_mutex_top_waiter(lock)) {
 281                 /* Boost the owner */
 282                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 283                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 284                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 285                 __rt_mutex_adjust_prio(task);
 286
 287         } else if (top_waiter == waiter) {
 288                 /* Deboost the owner */
 289                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 290                 waiter = rt_mutex_top_waiter(lock);
 291                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 292                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 293                 __rt_mutex_adjust_prio(task);
 294         }
 295
 296         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 297
 298         top_waiter = rt_mutex_top_waiter(lock);
 299         raw_spin_unlock(&lock->wait_lock);
 300
 301         if (!detect_deadlock && waiter != top_waiter)
 302                 goto out_put_task;
 303
 304         goto again;
 305
 306  out_unlock_pi:
 307         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 308  out_put_task:
 309         put_task_struct(task);
 310
 311         return ret;
 312 }
 313
 314 /*
 315  * Try to take an rt-mutex
 316  *
 317  * Must be called with lock->wait_lock held.
 318  *
 319  * @lock:   the lock to be acquired.
 320  * @task:   the task which wants to acquire the lock
 321  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 322  */
 323 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 324                 struct rt_mutex_waiter *waiter)
 325 {
 326         /*
 327          * We have to be careful here if the atomic speedups are
 328          * enabled, such that, when
 329          *  - no other waiter is on the lock
 330          *  - the lock has been released since we did the cmpxchg
 331          * the lock can be released or taken while we are doing the
 332          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 333          *
 334          * The atomic acquire/release aware variant of
 335          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 336          * the WAITERS bit, the atomic release / acquire can not
 337          * happen anymore and lock->wait_lock protects us from the
 338          * non-atomic case.
 339          *
 340          * Note, that this might set lock->owner =
 341          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 342          * any more. This is fixed up when we take the ownership.
 343          * This is the transitional state explained at the top of this file.
 344          */
 345         mark_rt_mutex_waiters(lock);
 346
 347         if (rt_mutex_owner(lock))
 348                 return 0;
 349
 350         /*
 351          * It will get the lock because of one of these conditions:
 352          * 1) there is no waiter
 353          * 2) higher priority than waiters
 354          * 3) it is top waiter
 355          */
 356         if (rt_mutex_has_waiters(lock)) {
 357                 if (task->prio >= rt_mutex_top_waiter(lock)->list_entry.prio) {
 358                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 359                                 return 0;
 360                 }
 361         }
 362
 363         if (waiter || rt_mutex_has_waiters(lock)) {
 364                 unsigned long flags;
 365                 struct rt_mutex_waiter *top;
 366
 367                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 368
 369                 /* remove the queued waiter. */
 370                 if (waiter) {
 371                         plist_del(&waiter->list_entry, &lock->wait_list);
 372                         task->pi_blocked_on = NULL;
 373                 }
 374
 375                 /*
 376                  * We have to enqueue the top waiter(if it exists) into
 377                  * task->pi_waiters list.
 378                  */
 379                 if (rt_mutex_has_waiters(lock)) {
 380                         top = rt_mutex_top_waiter(lock);
 381                         top->pi_list_entry.prio = top->list_entry.prio;
 382                         plist_add(&top->pi_list_entry, &task->pi_waiters);
 383                 }
 384                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 385         }
 386
 387         /* We got the lock. */
 388         debug_rt_mutex_lock(lock);
 389
 390         rt_mutex_set_owner(lock, task);
 391
 392         rt_mutex_deadlock_account_lock(lock, task);
 393
 394         return 1;
 395 }
 396
 397 /*
 398  * Task blocks on lock.
 399  *
 400  * Prepare waiter and propagate pi chain
 401  *
 402  * This must be called with lock->wait_lock held.
 403  */
 404 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 405                                    struct rt_mutex_waiter *waiter,
 406                                    struct task_struct *task,
 407                                    int detect_deadlock)
 408 {
 409         struct task_struct *owner = rt_mutex_owner(lock);
 410         struct rt_mutex_waiter *top_waiter = waiter;
 411         unsigned long flags;
 412         int chain_walk = 0, res;
 413
 414         /*
 415          * Early deadlock detection. We really don't want the task to
 416          * enqueue on itself just to untangle the mess later. It's not
 417          * only an optimization. We drop the locks, so another waiter
 418          * can come in before the chain walk detects the deadlock. So
 419          * the other will detect the deadlock and return -EDEADLOCK,
 420          * which is wrong, as the other waiter is not in a deadlock
 421          * situation.
 422          */
 423         if (detect_deadlock && owner == task)
 424                 return -EDEADLK;
 425
 426         raw_spin_lock_irqsave(&task->pi_lock, flags);
 427         __rt_mutex_adjust_prio(task);
 428         waiter->task = task;
 429         waiter->lock = lock;
 430         plist_node_init(&waiter->list_entry, task->prio);
 431         plist_node_init(&waiter->pi_list_entry, task->prio);
 432
 433         /* Get the top priority waiter on the lock */
 434         if (rt_mutex_has_waiters(lock))
 435                 top_waiter = rt_mutex_top_waiter(lock);
 436         plist_add(&waiter->list_entry, &lock->wait_list);
 437
 438         task->pi_blocked_on = waiter;
 439
 440         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 441
 442         if (!owner)
 443                 return 0;
 444
 445         if (waiter == rt_mutex_top_waiter(lock)) {
 446                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 447                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 448                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 449
 450                 __rt_mutex_adjust_prio(owner);
 451                 if (owner->pi_blocked_on)
 452                         chain_walk = 1;
 453                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 454         }
 455         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 456                 chain_walk = 1;
 457
 458         if (!chain_walk)
 459                 return 0;
 460
 461         /*
 462          * The owner can't disappear while holding a lock,
 463          * so the owner struct is protected by wait_lock.
 464          * Gets dropped in rt_mutex_adjust_prio_chain()!
 465          */
 466         get_task_struct(owner);
 467
 468         raw_spin_unlock(&lock->wait_lock);
 469
 470         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 471                                          task);
 472
 473         raw_spin_lock(&lock->wait_lock);
 474
 475         return res;
 476 }
 477
 478 /*
 479  * Wake up the next waiter on the lock.
 480  *
 481  * Remove the top waiter from the current tasks waiter list and wake it up.
 482  *
 483  * Called with lock->wait_lock held.
 484  */
 485 static void wakeup_next_waiter(struct rt_mutex *lock)
 486 {
 487         struct rt_mutex_waiter *waiter;
 488         unsigned long flags;
 489
 490         raw_spin_lock_irqsave(&current->pi_lock, flags);
 491
 492         waiter = rt_mutex_top_waiter(lock);
 493
 494         /*
 495          * Remove it from current->pi_waiters. We do not adjust a
 496          * possible priority boost right now. We execute wakeup in the
 497          * boosted mode and go back to normal after releasing
 498          * lock->wait_lock.
 499          */
 500         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 501
 502         rt_mutex_set_owner(lock, NULL);
 503
 504         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 505
 506         wake_up_process(waiter->task);
 507 }
 508
 509 /*
 510  * Remove a waiter from a lock and give up
 511  *
 512  * Must be called with lock->wait_lock held and
 513  * have just failed to try_to_take_rt_mutex().
 514  */
 515 static void remove_waiter(struct rt_mutex *lock,
 516                           struct rt_mutex_waiter *waiter)
 517 {
 518         int first = (waiter == rt_mutex_top_waiter(lock));
 519         struct task_struct *owner = rt_mutex_owner(lock);
 520         unsigned long flags;
 521         int chain_walk = 0;
 522
 523         raw_spin_lock_irqsave(&current->pi_lock, flags);
 524         plist_del(&waiter->list_entry, &lock->wait_list);
 525         current->pi_blocked_on = NULL;
 526         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 527
 528         if (!owner)
 529                 return;
 530
 531         if (first) {
 532
 533                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 534
 535                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 536
 537                 if (rt_mutex_has_waiters(lock)) {
 538                         struct rt_mutex_waiter *next;
 539
 540                         next = rt_mutex_top_waiter(lock);
 541                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 542                 }
 543                 __rt_mutex_adjust_prio(owner);
 544
 545                 if (owner->pi_blocked_on)
 546                         chain_walk = 1;
 547
 548                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 549         }
 550
 551         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 552
 553         if (!chain_walk)
 554                 return;
 555
 556         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 557         get_task_struct(owner);
 558
 559         raw_spin_unlock(&lock->wait_lock);
 560
 561         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 562
 563         raw_spin_lock(&lock->wait_lock);
 564 }
 565
 566 /*
 567  * Recheck the pi chain, in case we got a priority setting
 568  *
 569  * Called from sched_setscheduler
 570  */
 571 void rt_mutex_adjust_pi(struct task_struct *task)
 572 {
 573         struct rt_mutex_waiter *waiter;
 574         unsigned long flags;
 575
 576         raw_spin_lock_irqsave(&task->pi_lock, flags);
 577
 578         waiter = task->pi_blocked_on;
 579         if (!waiter || waiter->list_entry.prio == task->prio) {
 580                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 581                 return;
 582         }
 583
 584         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 585
 586         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 587         get_task_struct(task);
 588         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 589 }
 590
 591 /**
 592  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 593  * @lock:                the rt_mutex to take
 594  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 595  *                       or TASK_UNINTERRUPTIBLE)
 596  * @timeout:             the pre-initialized and started timer, or NULL for none
 597  * @waiter:              the pre-initialized rt_mutex_waiter
 598  *
 599  * lock->wait_lock must be held by the caller.
 600  */
 601 static int __sched
 602 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 603                     struct hrtimer_sleeper *timeout,
 604                     struct rt_mutex_waiter *waiter)
 605 {
 606         int ret = 0;
 607
 608         for (;;) {
 609                 /* Try to acquire the lock: */
 610                 if (try_to_take_rt_mutex(lock, current, waiter))
 611                         break;
 612
 613                 /*
 614                  * TASK_INTERRUPTIBLE checks for signals and
 615                  * timeout. Ignored otherwise.
 616                  */
 617                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 618                         /* Signal pending? */
 619                         if (signal_pending(current))
 620                                 ret = -EINTR;
 621                         if (timeout && !timeout->task)
 622                                 ret = -ETIMEDOUT;
 623                         if (ret)
 624                                 break;
 625                 }
 626
 627                 raw_spin_unlock(&lock->wait_lock);
 628
 629                 debug_rt_mutex_print_deadlock(waiter);
 630
 631                 schedule_rt_mutex(lock);
 632
 633                 raw_spin_lock(&lock->wait_lock);
 634                 set_current_state(state);
 635         }
 636
 637         return ret;
 638 }
 639
 640 /*
 641  * Slow path lock function:
 642  */
 643 static int __sched
 644 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 645                   struct hrtimer_sleeper *timeout,
 646                   int detect_deadlock)
 647 {
 648         struct rt_mutex_waiter waiter;
 649         int ret = 0;
 650
 651         debug_rt_mutex_init_waiter(&waiter);
 652
 653         raw_spin_lock(&lock->wait_lock);
 654
 655         /* Try to acquire the lock again: */
 656         if (try_to_take_rt_mutex(lock, current, NULL)) {
 657                 raw_spin_unlock(&lock->wait_lock);
 658                 return 0;
 659         }
 660
 661         set_current_state(state);
 662
 663         /* Setup the timer, when timeout != NULL */
 664         if (unlikely(timeout)) {
 665                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 666                 if (!hrtimer_active(&timeout->timer))
 667                         timeout->task = NULL;
 668         }
 669
 670         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 671
 672         if (likely(!ret))
 673                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
 674
 675         set_current_state(TASK_RUNNING);
 676
 677         if (unlikely(ret))
 678                 remove_waiter(lock, &waiter);
 679
 680         /*
 681          * try_to_take_rt_mutex() sets the waiter bit
 682          * unconditionally. We might have to fix that up.
 683          */
 684         fixup_rt_mutex_waiters(lock);
 685
 686         raw_spin_unlock(&lock->wait_lock);
 687
 688         /* Remove pending timer: */
 689         if (unlikely(timeout))
 690                 hrtimer_cancel(&timeout->timer);
 691
 692         debug_rt_mutex_free_waiter(&waiter);
 693
 694         return ret;
 695 }
 696
 697 /*
 698  * Slow path try-lock function:
 699  */
 700 static inline int
 701 rt_mutex_slowtrylock(struct rt_mutex *lock)
 702 {
 703         int ret = 0;
 704
 705         raw_spin_lock(&lock->wait_lock);
 706
 707         if (likely(rt_mutex_owner(lock) != current)) {
 708
 709                 ret = try_to_take_rt_mutex(lock, current, NULL);
 710                 /*
 711                  * try_to_take_rt_mutex() sets the lock waiters
 712                  * bit unconditionally. Clean this up.
 713                  */
 714                 fixup_rt_mutex_waiters(lock);
 715         }
 716
 717         raw_spin_unlock(&lock->wait_lock);
 718
 719         return ret;
 720 }
 721
 722 /*
 723  * Slow path to release a rt-mutex:
 724  */
 725 static void __sched
 726 rt_mutex_slowunlock(struct rt_mutex *lock)
 727 {
 728         raw_spin_lock(&lock->wait_lock);
 729
 730         debug_rt_mutex_unlock(lock);
 731
 732         rt_mutex_deadlock_account_unlock(current);
 733
 734         if (!rt_mutex_has_waiters(lock)) {
 735                 lock->owner = NULL;
 736                 raw_spin_unlock(&lock->wait_lock);
 737                 return;
 738         }
 739
 740         wakeup_next_waiter(lock);
 741
 742         raw_spin_unlock(&lock->wait_lock);
 743
 744         /* Undo pi boosting if necessary: */
 745         rt_mutex_adjust_prio(current);
 746 }
 747
 748 /*
 749  * debug aware fast / slowpath lock,trylock,unlock
 750  *
 751  * The atomic acquire/release ops are compiled away, when either the
 752  * architecture does not support cmpxchg or when debugging is enabled.
 753  */
 754 static inline int
 755 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 756                   int detect_deadlock,
 757                   int (*slowfn)(struct rt_mutex *lock, int state,
 758                                 struct hrtimer_sleeper *timeout,
 759                                 int detect_deadlock))
 760 {
 761         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 762                 rt_mutex_deadlock_account_lock(lock, current);
 763                 return 0;
 764         } else
 765                 return slowfn(lock, state, NULL, detect_deadlock);
 766 }
 767
 768 static inline int
 769 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 770                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 771                         int (*slowfn)(struct rt_mutex *lock, int state,
 772                                       struct hrtimer_sleeper *timeout,
 773                                       int detect_deadlock))
 774 {
 775         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 776                 rt_mutex_deadlock_account_lock(lock, current);
 777                 return 0;
 778         } else
 779                 return slowfn(lock, state, timeout, detect_deadlock);
 780 }
 781
 782 static inline int
 783 rt_mutex_fasttrylock(struct rt_mutex *lock,
 784                      int (*slowfn)(struct rt_mutex *lock))
 785 {
 786         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 787                 rt_mutex_deadlock_account_lock(lock, current);
 788                 return 1;
 789         }
 790         return slowfn(lock);
 791 }
 792
 793 static inline void
 794 rt_mutex_fastunlock(struct rt_mutex *lock,
 795                     void (*slowfn)(struct rt_mutex *lock))
 796 {
 797         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 798                 rt_mutex_deadlock_account_unlock(current);
 799         else
 800                 slowfn(lock);
 801 }
 802
 803 /**
 804  * rt_mutex_lock - lock a rt_mutex
 805  *
 806  * @lock: the rt_mutex to be locked
 807  */
 808 void __sched rt_mutex_lock(struct rt_mutex *lock)
 809 {
 810         might_sleep();
 811
 812         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 813 }
 814 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 815
 816 /**
 817  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 818  *
 819  * @lock:               the rt_mutex to be locked
 820  * @detect_deadlock:    deadlock detection on/off
 821  *
 822  * Returns:
 823  *  0           on success
 824  * -EINTR       when interrupted by a signal
 825  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 826  */
 827 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 828                                                  int detect_deadlock)
 829 {
 830         might_sleep();
 831
 832         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 833                                  detect_deadlock, rt_mutex_slowlock);
 834 }
 835 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 836
 837 /**
 838  * rt_mutex_timed_lock - lock a rt_mutex interruptible
 839  *                      the timeout structure is provided
 840  *                      by the caller
 841  *
 842  * @lock:               the rt_mutex to be locked
 843  * @timeout:            timeout structure or NULL (no timeout)
 844  * @detect_deadlock:    deadlock detection on/off
 845  *
 846  * Returns:
 847  *  0           on success
 848  * -EINTR       when interrupted by a signal
 849  * -ETIMEDOUT   when the timeout expired
 850  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 851  */
 852 int
 853 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 854                     int detect_deadlock)
 855 {
 856         might_sleep();
 857
 858         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 859                                        detect_deadlock, rt_mutex_slowlock);
 860 }
 861 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 862
 863 /**
 864  * rt_mutex_trylock - try to lock a rt_mutex
 865  *
 866  * @lock:       the rt_mutex to be locked
 867  *
 868  * Returns 1 on success and 0 on contention
 869  */
 870 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 871 {
 872         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 873 }
 874 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 875
 876 /**
 877  * rt_mutex_unlock - unlock a rt_mutex
 878  *
 879  * @lock: the rt_mutex to be unlocked
 880  */
 881 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 882 {
 883         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 884 }
 885 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 886
 887 /**
 888  * rt_mutex_destroy - mark a mutex unusable
 889  * @lock: the mutex to be destroyed
 890  *
 891  * This function marks the mutex uninitialized, and any subsequent
 892  * use of the mutex is forbidden. The mutex must not be locked when
 893  * this function is called.
 894  */
 895 void rt_mutex_destroy(struct rt_mutex *lock)
 896 {
 897         WARN_ON(rt_mutex_is_locked(lock));
 898 #ifdef CONFIG_DEBUG_RT_MUTEXES
 899         lock->magic = NULL;
 900 #endif
 901 }
 902
 903 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 904
 905 /**
 906  * __rt_mutex_init - initialize the rt lock
 907  *
 908  * @lock: the rt lock to be initialized
 909  *
 910  * Initialize the rt lock to unlocked state.
 911  *
 912  * Initializing of a locked rt lock is not allowed
 913  */
 914 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 915 {
 916         lock->owner = NULL;
 917         raw_spin_lock_init(&lock->wait_lock);
 918         plist_head_init(&lock->wait_list);
 919
 920         debug_rt_mutex_init(lock, name);
 921 }
 922 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 923
 924 /**
 925  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 926  *                              proxy owner
 927  *
 928  * @lock:       the rt_mutex to be locked
 929  * @proxy_owner:the task to set as owner
 930  *
 931  * No locking. Caller has to do serializing itself
 932  * Special API call for PI-futex support
 933  */
 934 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 935                                 struct task_struct *proxy_owner)
 936 {
 937         __rt_mutex_init(lock, NULL);
 938         debug_rt_mutex_proxy_lock(lock, proxy_owner);
 939         rt_mutex_set_owner(lock, proxy_owner);
 940         rt_mutex_deadlock_account_lock(lock, proxy_owner);
 941 }
 942
 943 /**
 944  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 945  *
 946  * @lock:       the rt_mutex to be locked
 947  *
 948  * No locking. Caller has to do serializing itself
 949  * Special API call for PI-futex support
 950  */
 951 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 952                            struct task_struct *proxy_owner)
 953 {
 954         debug_rt_mutex_proxy_unlock(lock);
 955         rt_mutex_set_owner(lock, NULL);
 956         rt_mutex_deadlock_account_unlock(proxy_owner);
 957 }
 958
 959 /**
 960  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
 961  * @lock:               the rt_mutex to take
 962  * @waiter:             the pre-initialized rt_mutex_waiter
 963  * @task:               the task to prepare
 964  * @detect_deadlock:    perform deadlock detection (1) or not (0)
 965  *
 966  * Returns:
 967  *  0 - task blocked on lock
 968  *  1 - acquired the lock for task, caller should wake it up
 969  * <0 - error
 970  *
 971  * Special API call for FUTEX_REQUEUE_PI support.
 972  */
 973 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
 974                               struct rt_mutex_waiter *waiter,
 975                               struct task_struct *task, int detect_deadlock)
 976 {
 977         int ret;
 978
 979         raw_spin_lock(&lock->wait_lock);
 980
 981         if (try_to_take_rt_mutex(lock, task, NULL)) {
 982                 raw_spin_unlock(&lock->wait_lock);
 983                 return 1;
 984         }
 985
 986         ret = task_blocks_on_rt_mutex(lock, waiter, task, detect_deadlock);
 987
 988         if (ret && !rt_mutex_owner(lock)) {
 989                 /*
 990                  * Reset the return value. We might have
 991                  * returned with -EDEADLK and the owner
 992                  * released the lock while we were walking the
 993                  * pi chain.  Let the waiter sort it out.
 994                  */
 995                 ret = 0;
 996         }
 997
 998         if (unlikely(ret))
 999                 remove_waiter(lock, waiter);
1000
1001         raw_spin_unlock(&lock->wait_lock);
1002
1003         debug_rt_mutex_print_deadlock(waiter);
1004
1005         return ret;
1006 }
1007
1008 /**
1009  * rt_mutex_next_owner - return the next owner of the lock
1010  *
1011  * @lock: the rt lock query
1012  *
1013  * Returns the next owner of the lock or NULL
1014  *
1015  * Caller has to serialize against other accessors to the lock
1016  * itself.
1017  *
1018  * Special API call for PI-futex support
1019  */
1020 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
1021 {
1022         if (!rt_mutex_has_waiters(lock))
1023                 return NULL;
1024
1025         return rt_mutex_top_waiter(lock)->task;
1026 }
1027
1028 /**
1029  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1030  * @lock:               the rt_mutex we were woken on
1031  * @to:                 the timeout, null if none. hrtimer should already have
1032  *                      been started.
1033  * @waiter:             the pre-initialized rt_mutex_waiter
1034  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1035  *
1036  * Complete the lock acquisition started our behalf by another thread.
1037  *
1038  * Returns:
1039  *  0 - success
1040  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1041  *
1042  * Special API call for PI-futex requeue support
1043  */
1044 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1045                                struct hrtimer_sleeper *to,
1046                                struct rt_mutex_waiter *waiter,
1047                                int detect_deadlock)
1048 {
1049         int ret;
1050
1051         raw_spin_lock(&lock->wait_lock);
1052
1053         set_current_state(TASK_INTERRUPTIBLE);
1054
1055         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1056
1057         set_current_state(TASK_RUNNING);
1058
1059         if (unlikely(ret))
1060                 remove_waiter(lock, waiter);
1061
1062         /*
1063          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1064          * have to fix that up.
1065          */
1066         fixup_rt_mutex_waiters(lock);
1067
1068         raw_spin_unlock(&lock->wait_lock);
1069
1070         return ret;
1071 }