linux/kernel/locking/rtmutex.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * RT-Mutexes: simple blocking mutual exclusion locks with PI support
   4 *
   5 * started by Ingo Molnar and Thomas Gleixner.
   6 *
   7 *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   8 *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
   9 *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
  10 *  Copyright (C) 2006 Esben Nielsen
  11 * Adaptive Spinlocks:
  12 *  Copyright (C) 2008 Novell, Inc., Gregory Haskins, Sven Dietrich,
  13 *                                   and Peter Morreale,
  14 * Adaptive Spinlocks simplification:
  15 *  Copyright (C) 2008 Red Hat, Inc., Steven Rostedt <srostedt@redhat.com>
  16 *
  17 *  See Documentation/locking/rt-mutex-design.rst for details.
  18 */
  19#include <linux/sched.h>
  20#include <linux/sched/debug.h>
  21#include <linux/sched/deadline.h>
  22#include <linux/sched/signal.h>
  23#include <linux/sched/rt.h>
  24#include <linux/sched/wake_q.h>
  25#include <linux/ww_mutex.h>
  26
  27#include "rtmutex_common.h"
  28
  29#ifndef WW_RT
  30# define build_ww_mutex()       (false)
  31# define ww_container_of(rtm)   NULL
  32
  33static inline int __ww_mutex_add_waiter(struct rt_mutex_waiter *waiter,
  34                                        struct rt_mutex *lock,
  35                                        struct ww_acquire_ctx *ww_ctx)
  36{
  37        return 0;
  38}
  39
  40static inline void __ww_mutex_check_waiters(struct rt_mutex *lock,
  41                                            struct ww_acquire_ctx *ww_ctx)
  42{
  43}
  44
  45static inline void ww_mutex_lock_acquired(struct ww_mutex *lock,
  46                                          struct ww_acquire_ctx *ww_ctx)
  47{
  48}
  49
  50static inline int __ww_mutex_check_kill(struct rt_mutex *lock,
  51                                        struct rt_mutex_waiter *waiter,
  52                                        struct ww_acquire_ctx *ww_ctx)
  53{
  54        return 0;
  55}
  56
  57#else
  58# define build_ww_mutex()       (true)
  59# define ww_container_of(rtm)   container_of(rtm, struct ww_mutex, base)
  60# include "ww_mutex.h"
  61#endif
  62
  63/*
  64 * lock->owner state tracking:
  65 *
  66 * lock->owner holds the task_struct pointer of the owner. Bit 0
  67 * is used to keep track of the "lock has waiters" state.
  68 *
  69 * owner        bit0
  70 * NULL         0       lock is free (fast acquire possible)
  71 * NULL         1       lock is free and has waiters and the top waiter
  72 *                              is going to take the lock*
  73 * taskpointer  0       lock is held (fast release possible)
  74 * taskpointer  1       lock is held and has waiters**
  75 *
  76 * The fast atomic compare exchange based acquire and release is only
  77 * possible when bit 0 of lock->owner is 0.
  78 *
  79 * (*) It also can be a transitional state when grabbing the lock
  80 * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  81 * we need to set the bit0 before looking at the lock, and the owner may be
  82 * NULL in this small time, hence this can be a transitional state.
  83 *
  84 * (**) There is a small time when bit 0 is set but there are no
  85 * waiters. This can happen when grabbing the lock in the slow path.
  86 * To prevent a cmpxchg of the owner releasing the lock, we need to
  87 * set this bit before looking at the lock.
  88 */
  89
  90static __always_inline void
  91rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
  92{
  93        unsigned long val = (unsigned long)owner;
  94
  95        if (rt_mutex_has_waiters(lock))
  96                val |= RT_MUTEX_HAS_WAITERS;
  97
  98        WRITE_ONCE(lock->owner, (struct task_struct *)val);
  99}
 100
 101static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
 102{
 103        lock->owner = (struct task_struct *)
 104                        ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
 105}
 106
 107static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex_base *lock)
 108{
 109        unsigned long owner, *p = (unsigned long *) &lock->owner;
 110
 111        if (rt_mutex_has_waiters(lock))
 112                return;
 113
 114        /*
 115         * The rbtree has no waiters enqueued, now make sure that the
 116         * lock->owner still has the waiters bit set, otherwise the
 117         * following can happen:
 118         *
 119         * CPU 0        CPU 1           CPU2
 120         * l->owner=T1
 121         *              rt_mutex_lock(l)
 122         *              lock(l->lock)
 123         *              l->owner = T1 | HAS_WAITERS;
 124         *              enqueue(T2)
 125         *              boost()
 126         *                unlock(l->lock)
 127         *              block()
 128         *
 129         *                              rt_mutex_lock(l)
 130         *                              lock(l->lock)
 131         *                              l->owner = T1 | HAS_WAITERS;
 132         *                              enqueue(T3)
 133         *                              boost()
 134         *                                unlock(l->lock)
 135         *                              block()
 136         *              signal(->T2)    signal(->T3)
 137         *              lock(l->lock)
 138         *              dequeue(T2)
 139         *              deboost()
 140         *                unlock(l->lock)
 141         *                              lock(l->lock)
 142         *                              dequeue(T3)
 143         *                               ==> wait list is empty
 144         *                              deboost()
 145         *                               unlock(l->lock)
 146         *              lock(l->lock)
 147         *              fixup_rt_mutex_waiters()
 148         *                if (wait_list_empty(l) {
 149         *                  l->owner = owner
 150         *                  owner = l->owner & ~HAS_WAITERS;
 151         *                    ==> l->owner = T1
 152         *                }
 153         *                              lock(l->lock)
 154         * rt_mutex_unlock(l)           fixup_rt_mutex_waiters()
 155         *                                if (wait_list_empty(l) {
 156         *                                  owner = l->owner & ~HAS_WAITERS;
 157         * cmpxchg(l->owner, T1, NULL)
 158         *  ===> Success (l->owner = NULL)
 159         *
 160         *                                  l->owner = owner
 161         *                                    ==> l->owner = T1
 162         *                                }
 163         *
 164         * With the check for the waiter bit in place T3 on CPU2 will not
 165         * overwrite. All tasks fiddling with the waiters bit are
 166         * serialized by l->lock, so nothing else can modify the waiters
 167         * bit. If the bit is set then nothing can change l->owner either
 168         * so the simple RMW is safe. The cmpxchg() will simply fail if it
 169         * happens in the middle of the RMW because the waiters bit is
 170         * still set.
 171         */
 172        owner = READ_ONCE(*p);
 173        if (owner & RT_MUTEX_HAS_WAITERS)
 174                WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
 175}
 176
 177/*
 178 * We can speed up the acquire/release, if there's no debugging state to be
 179 * set up.
 180 */
 181#ifndef CONFIG_DEBUG_RT_MUTEXES
 182static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
 183                                                     struct task_struct *old,
 184                                                     struct task_struct *new)
 185{
 186        return try_cmpxchg_acquire(&lock->owner, &old, new);
 187}
 188
 189static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
 190                                                     struct task_struct *old,
 191                                                     struct task_struct *new)
 192{
 193        return try_cmpxchg_release(&lock->owner, &old, new);
 194}
 195
 196/*
 197 * Callers must hold the ->wait_lock -- which is the whole purpose as we force
 198 * all future threads that attempt to [Rmw] the lock to the slowpath. As such
 199 * relaxed semantics suffice.
 200 */
 201static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
 202{
 203        unsigned long owner, *p = (unsigned long *) &lock->owner;
 204
 205        do {
 206                owner = *p;
 207        } while (cmpxchg_relaxed(p, owner,
 208                                 owner | RT_MUTEX_HAS_WAITERS) != owner);
 209}
 210
 211/*
 212 * Safe fastpath aware unlock:
 213 * 1) Clear the waiters bit
 214 * 2) Drop lock->wait_lock
 215 * 3) Try to unlock the lock with cmpxchg
 216 */
 217static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
 218                                                 unsigned long flags)
 219        __releases(lock->wait_lock)
 220{
 221        struct task_struct *owner = rt_mutex_owner(lock);
 222
 223        clear_rt_mutex_waiters(lock);
 224        raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 225        /*
 226         * If a new waiter comes in between the unlock and the cmpxchg
 227         * we have two situations:
 228         *
 229         * unlock(wait_lock);
 230         *                                      lock(wait_lock);
 231         * cmpxchg(p, owner, 0) == owner
 232         *                                      mark_rt_mutex_waiters(lock);
 233         *                                      acquire(lock);
 234         * or:
 235         *
 236         * unlock(wait_lock);
 237         *                                      lock(wait_lock);
 238         *                                      mark_rt_mutex_waiters(lock);
 239         *
 240         * cmpxchg(p, owner, 0) != owner
 241         *                                      enqueue_waiter();
 242         *                                      unlock(wait_lock);
 243         * lock(wait_lock);
 244         * wake waiter();
 245         * unlock(wait_lock);
 246         *                                      lock(wait_lock);
 247         *                                      acquire(lock);
 248         */
 249        return rt_mutex_cmpxchg_release(lock, owner, NULL);
 250}
 251
 252#else
 253static __always_inline bool rt_mutex_cmpxchg_acquire(struct rt_mutex_base *lock,
 254                                                     struct task_struct *old,
 255                                                     struct task_struct *new)
 256{
 257        return false;
 258
 259}
 260
 261static __always_inline bool rt_mutex_cmpxchg_release(struct rt_mutex_base *lock,
 262                                                     struct task_struct *old,
 263                                                     struct task_struct *new)
 264{
 265        return false;
 266}
 267
 268static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
 269{
 270        lock->owner = (struct task_struct *)
 271                        ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
 272}
 273
 274/*
 275 * Simple slow path only version: lock->owner is protected by lock->wait_lock.
 276 */
 277static __always_inline bool unlock_rt_mutex_safe(struct rt_mutex_base *lock,
 278                                                 unsigned long flags)
 279        __releases(lock->wait_lock)
 280{
 281        lock->owner = NULL;
 282        raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
 283        return true;
 284}
 285#endif
 286
 287static __always_inline int __waiter_prio(struct task_struct *task)
 288{
 289        int prio = task->prio;
 290
 291        if (!rt_prio(prio))
 292                return DEFAULT_PRIO;
 293
 294        return prio;
 295}
 296
 297static __always_inline void
 298waiter_update_prio(struct rt_mutex_waiter *waiter, struct task_struct *task)
 299{
 300        waiter->prio = __waiter_prio(task);
 301        waiter->deadline = task->dl.deadline;
 302}
 303
 304/*
 305 * Only use with rt_mutex_waiter_{less,equal}()
 306 */
 307#define task_to_waiter(p)       \
 308        &(struct rt_mutex_waiter){ .prio = __waiter_prio(p), .deadline = (p)->dl.deadline }
 309
 310static __always_inline int rt_mutex_waiter_less(struct rt_mutex_waiter *left,
 311                                                struct rt_mutex_waiter *right)
 312{
 313        if (left->prio < right->prio)
 314                return 1;
 315
 316        /*
 317         * If both waiters have dl_prio(), we check the deadlines of the
 318         * associated tasks.
 319         * If left waiter has a dl_prio(), and we didn't return 1 above,
 320         * then right waiter has a dl_prio() too.
 321         */
 322        if (dl_prio(left->prio))
 323                return dl_time_before(left->deadline, right->deadline);
 324
 325        return 0;
 326}
 327
 328static __always_inline int rt_mutex_waiter_equal(struct rt_mutex_waiter *left,
 329                                                 struct rt_mutex_waiter *right)
 330{
 331        if (left->prio != right->prio)
 332                return 0;
 333
 334        /*
 335         * If both waiters have dl_prio(), we check the deadlines of the
 336         * associated tasks.
 337         * If left waiter has a dl_prio(), and we didn't return 0 above,
 338         * then right waiter has a dl_prio() too.
 339         */
 340        if (dl_prio(left->prio))
 341                return left->deadline == right->deadline;
 342
 343        return 1;
 344}
 345
 346static inline bool rt_mutex_steal(struct rt_mutex_waiter *waiter,
 347                                  struct rt_mutex_waiter *top_waiter)
 348{
 349        if (rt_mutex_waiter_less(waiter, top_waiter))
 350                return true;
 351
 352#ifdef RT_MUTEX_BUILD_SPINLOCKS
 353        /*
 354         * Note that RT tasks are excluded from same priority (lateral)
 355         * steals to prevent the introduction of an unbounded latency.
 356         */
 357        if (rt_prio(waiter->prio) || dl_prio(waiter->prio))
 358                return false;
 359
 360        return rt_mutex_waiter_equal(waiter, top_waiter);
 361#else
 362        return false;
 363#endif
 364}
 365
 366#define __node_2_waiter(node) \
 367        rb_entry((node), struct rt_mutex_waiter, tree_entry)
 368
 369static __always_inline bool __waiter_less(struct rb_node *a, const struct rb_node *b)
 370{
 371        struct rt_mutex_waiter *aw = __node_2_waiter(a);
 372        struct rt_mutex_waiter *bw = __node_2_waiter(b);
 373
 374        if (rt_mutex_waiter_less(aw, bw))
 375                return 1;
 376
 377        if (!build_ww_mutex())
 378                return 0;
 379
 380        if (rt_mutex_waiter_less(bw, aw))
 381                return 0;
 382
 383        /* NOTE: relies on waiter->ww_ctx being set before insertion */
 384        if (aw->ww_ctx) {
 385                if (!bw->ww_ctx)
 386                        return 1;
 387
 388                return (signed long)(aw->ww_ctx->stamp -
 389                                     bw->ww_ctx->stamp) < 0;
 390        }
 391
 392        return 0;
 393}
 394
 395static __always_inline void
 396rt_mutex_enqueue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
 397{
 398        rb_add_cached(&waiter->tree_entry, &lock->waiters, __waiter_less);
 399}
 400
 401static __always_inline void
 402rt_mutex_dequeue(struct rt_mutex_base *lock, struct rt_mutex_waiter *waiter)
 403{
 404        if (RB_EMPTY_NODE(&waiter->tree_entry))
 405                return;
 406
 407        rb_erase_cached(&waiter->tree_entry, &lock->waiters);
 408        RB_CLEAR_NODE(&waiter->tree_entry);
 409}
 410
 411#define __node_2_pi_waiter(node) \
 412        rb_entry((node), struct rt_mutex_waiter, pi_tree_entry)
 413
 414static __always_inline bool
 415__pi_waiter_less(struct rb_node *a, const struct rb_node *b)
 416{
 417        return rt_mutex_waiter_less(__node_2_pi_waiter(a), __node_2_pi_waiter(b));
 418}
 419
 420static __always_inline void
 421rt_mutex_enqueue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 422{
 423        rb_add_cached(&waiter->pi_tree_entry, &task->pi_waiters, __pi_waiter_less);
 424}
 425
 426static __always_inline void
 427rt_mutex_dequeue_pi(struct task_struct *task, struct rt_mutex_waiter *waiter)
 428{
 429        if (RB_EMPTY_NODE(&waiter->pi_tree_entry))
 430                return;
 431
 432        rb_erase_cached(&waiter->pi_tree_entry, &task->pi_waiters);
 433        RB_CLEAR_NODE(&waiter->pi_tree_entry);
 434}
 435
 436static __always_inline void rt_mutex_adjust_prio(struct task_struct *p)
 437{
 438        struct task_struct *pi_task = NULL;
 439
 440        lockdep_assert_held(&p->pi_lock);
 441
 442        if (task_has_pi_waiters(p))
 443                pi_task = task_top_pi_waiter(p)->task;
 444
 445        rt_mutex_setprio(p, pi_task);
 446}
 447
 448/* RT mutex specific wake_q wrappers */
 449static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
 450                                                struct rt_mutex_waiter *w)
 451{
 452        if (IS_ENABLED(CONFIG_PREEMPT_RT) && w->wake_state != TASK_NORMAL) {
 453                if (IS_ENABLED(CONFIG_PROVE_LOCKING))
 454                        WARN_ON_ONCE(wqh->rtlock_task);
 455                get_task_struct(w->task);
 456                wqh->rtlock_task = w->task;
 457        } else {
 458                wake_q_add(&wqh->head, w->task);
 459        }
 460}
 461
 462static __always_inline void rt_mutex_wake_up_q(struct rt_wake_q_head *wqh)
 463{
 464        if (IS_ENABLED(CONFIG_PREEMPT_RT) && wqh->rtlock_task) {
 465                wake_up_state(wqh->rtlock_task, TASK_RTLOCK_WAIT);
 466                put_task_struct(wqh->rtlock_task);
 467                wqh->rtlock_task = NULL;
 468        }
 469
 470        if (!wake_q_empty(&wqh->head))
 471                wake_up_q(&wqh->head);
 472
 473        /* Pairs with preempt_disable() in mark_wakeup_next_waiter() */
 474        preempt_enable();
 475}
 476
 477/*
 478 * Deadlock detection is conditional:
 479 *
 480 * If CONFIG_DEBUG_RT_MUTEXES=n, deadlock detection is only conducted
 481 * if the detect argument is == RT_MUTEX_FULL_CHAINWALK.
 482 *
 483 * If CONFIG_DEBUG_RT_MUTEXES=y, deadlock detection is always
 484 * conducted independent of the detect argument.
 485 *
 486 * If the waiter argument is NULL this indicates the deboost path and
 487 * deadlock detection is disabled independent of the detect argument
 488 * and the config settings.
 489 */
 490static __always_inline bool
 491rt_mutex_cond_detect_deadlock(struct rt_mutex_waiter *waiter,
 492                              enum rtmutex_chainwalk chwalk)
 493{
 494        if (IS_ENABLED(CONFIG_DEBUG_RT_MUTEXES))
 495                return waiter != NULL;
 496        return chwalk == RT_MUTEX_FULL_CHAINWALK;
 497}
 498
 499static __always_inline struct rt_mutex_base *task_blocked_on_lock(struct task_struct *p)
 500{
 501        return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
 502}
 503
 504/*
 505 * Adjust the priority chain. Also used for deadlock detection.
 506 * Decreases task's usage by one - may thus free the task.
 507 *
 508 * @task:       the task owning the mutex (owner) for which a chain walk is
 509 *              probably needed
 510 * @chwalk:     do we have to carry out deadlock detection?
 511 * @orig_lock:  the mutex (can be NULL if we are walking the chain to recheck
 512 *              things for a task that has just got its priority adjusted, and
 513 *              is waiting on a mutex)
 514 * @next_lock:  the mutex on which the owner of @orig_lock was blocked before
 515 *              we dropped its pi_lock. Is never dereferenced, only used for
 516 *              comparison to detect lock chain changes.
 517 * @orig_waiter: rt_mutex_waiter struct for the task that has just donated
 518 *              its priority to the mutex owner (can be NULL in the case
 519 *              depicted above or if the top waiter is gone away and we are
 520 *              actually deboosting the owner)
 521 * @top_task:   the current top waiter
 522 *
 523 * Returns 0 or -EDEADLK.
 524 *
 525 * Chain walk basics and protection scope
 526 *
 527 * [R] refcount on task
 528 * [P] task->pi_lock held
 529 * [L] rtmutex->wait_lock held
 530 *
 531 * Step Description                             Protected by
 532 *      function arguments:
 533 *      @task                                   [R]
 534 *      @orig_lock if != NULL                   @top_task is blocked on it
 535 *      @next_lock                              Unprotected. Cannot be
 536 *                                              dereferenced. Only used for
 537 *                                              comparison.
 538 *      @orig_waiter if != NULL                 @top_task is blocked on it
 539 *      @top_task                               current, or in case of proxy
 540 *                                              locking protected by calling
 541 *                                              code
 542 *      again:
 543 *        loop_sanity_check();
 544 *      retry:
 545 * [1]    lock(task->pi_lock);                  [R] acquire [P]
 546 * [2]    waiter = task->pi_blocked_on;         [P]
 547 * [3]    check_exit_conditions_1();            [P]
 548 * [4]    lock = waiter->lock;                  [P]
 549 * [5]    if (!try_lock(lock->wait_lock)) {     [P] try to acquire [L]
 550 *          unlock(task->pi_lock);              release [P]
 551 *          goto retry;
 552 *        }
 553 * [6]    check_exit_conditions_2();            [P] + [L]
 554 * [7]    requeue_lock_waiter(lock, waiter);    [P] + [L]
 555 * [8]    unlock(task->pi_lock);                release [P]
 556 *        put_task_struct(task);                release [R]
 557 * [9]    check_exit_conditions_3();            [L]
 558 * [10]   task = owner(lock);                   [L]
 559 *        get_task_struct(task);                [L] acquire [R]
 560 *        lock(task->pi_lock);                  [L] acquire [P]
 561 * [11]   requeue_pi_waiter(tsk, waiters(lock));[P] + [L]
 562 * [12]   check_exit_conditions_4();            [P] + [L]
 563 * [13]   unlock(task->pi_lock);                release [P]
 564 *        unlock(lock->wait_lock);              release [L]
 565 *        goto again;
 566 */
 567static int __sched rt_mutex_adjust_prio_chain(struct task_struct *task,
 568                                              enum rtmutex_chainwalk chwalk,
 569                                              struct rt_mutex_base *orig_lock,
 570                                              struct rt_mutex_base *next_lock,
 571                                              struct rt_mutex_waiter *orig_waiter,
 572                                              struct task_struct *top_task)
 573{
 574        struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 575        struct rt_mutex_waiter *prerequeue_top_waiter;
 576        int ret = 0, depth = 0;
 577        struct rt_mutex_base *lock;
 578        bool detect_deadlock;
 579        bool requeue = true;
 580
 581        detect_deadlock = rt_mutex_cond_detect_deadlock(orig_waiter, chwalk);
 582
 583        /*
 584         * The (de)boosting is a step by step approach with a lot of
 585         * pitfalls. We want this to be preemptible and we want hold a
 586         * maximum of two locks per step. So we have to check
 587         * carefully whether things change under us.
 588         */
 589 again:
 590        /*
 591         * We limit the lock chain length for each invocation.
 592         */
 593        if (++depth > max_lock_depth) {
 594                static int prev_max;
 595
 596                /*
 597                 * Print this only once. If the admin changes the limit,
 598                 * print a new message when reaching the limit again.
 599                 */
 600                if (prev_max != max_lock_depth) {
 601                        prev_max = max_lock_depth;
 602                        printk(KERN_WARNING "Maximum lock depth %d reached "
 603                               "task: %s (%d)\n", max_lock_depth,
 604                               top_task->comm, task_pid_nr(top_task));
 605                }
 606                put_task_struct(task);
 607
 608                return -EDEADLK;
 609        }
 610
 611        /*
 612         * We are fully preemptible here and only hold the refcount on
 613         * @task. So everything can have changed under us since the
 614         * caller or our own code below (goto retry/again) dropped all
 615         * locks.
 616         */
 617 retry:
 618        /*
 619         * [1] Task cannot go away as we did a get_task() before !
 620         */
 621        raw_spin_lock_irq(&task->pi_lock);
 622
 623        /*
 624         * [2] Get the waiter on which @task is blocked on.
 625         */
 626        waiter = task->pi_blocked_on;
 627
 628        /*
 629         * [3] check_exit_conditions_1() protected by task->pi_lock.
 630         */
 631
 632        /*
 633         * Check whether the end of the boosting chain has been
 634         * reached or the state of the chain has changed while we
 635         * dropped the locks.
 636         */
 637        if (!waiter)
 638                goto out_unlock_pi;
 639
 640        /*
 641         * Check the orig_waiter state. After we dropped the locks,
 642         * the previous owner of the lock might have released the lock.
 643         */
 644        if (orig_waiter && !rt_mutex_owner(orig_lock))
 645                goto out_unlock_pi;
 646
 647        /*
 648         * We dropped all locks after taking a refcount on @task, so
 649         * the task might have moved on in the lock chain or even left
 650         * the chain completely and blocks now on an unrelated lock or
 651         * on @orig_lock.
 652         *
 653         * We stored the lock on which @task was blocked in @next_lock,
 654         * so we can detect the chain change.
 655         */
 656        if (next_lock != waiter->lock)
 657                goto out_unlock_pi;
 658
 659        /*
 660         * There could be 'spurious' loops in the lock graph due to ww_mutex,
 661         * consider:
 662         *
 663         *   P1: A, ww_A, ww_B
 664         *   P2: ww_B, ww_A
 665         *   P3: A
 666         *
 667         * P3 should not return -EDEADLK because it gets trapped in the cycle
 668         * created by P1 and P2 (which will resolve -- and runs into
 669         * max_lock_depth above). Therefore disable detect_deadlock such that
 670         * the below termination condition can trigger once all relevant tasks
 671         * are boosted.
 672         *
 673         * Even when we start with ww_mutex we can disable deadlock detection,
 674         * since we would supress a ww_mutex induced deadlock at [6] anyway.
 675         * Supressing it here however is not sufficient since we might still
 676         * hit [6] due to adjustment driven iteration.
 677         *
 678         * NOTE: if someone were to create a deadlock between 2 ww_classes we'd
 679         * utterly fail to report it; lockdep should.
 680         */
 681        if (IS_ENABLED(CONFIG_PREEMPT_RT) && waiter->ww_ctx && detect_deadlock)
 682                detect_deadlock = false;
 683
 684        /*
 685         * Drop out, when the task has no waiters. Note,
 686         * top_waiter can be NULL, when we are in the deboosting
 687         * mode!
 688         */
 689        if (top_waiter) {
 690                if (!task_has_pi_waiters(task))
 691                        goto out_unlock_pi;
 692                /*
 693                 * If deadlock detection is off, we stop here if we
 694                 * are not the top pi waiter of the task. If deadlock
 695                 * detection is enabled we continue, but stop the
 696                 * requeueing in the chain walk.
 697                 */
 698                if (top_waiter != task_top_pi_waiter(task)) {
 699                        if (!detect_deadlock)
 700                                goto out_unlock_pi;
 701                        else
 702                                requeue = false;
 703                }
 704        }
 705
 706        /*
 707         * If the waiter priority is the same as the task priority
 708         * then there is no further priority adjustment necessary.  If
 709         * deadlock detection is off, we stop the chain walk. If its
 710         * enabled we continue, but stop the requeueing in the chain
 711         * walk.
 712         */
 713        if (rt_mutex_waiter_equal(waiter, task_to_waiter(task))) {
 714                if (!detect_deadlock)
 715                        goto out_unlock_pi;
 716                else
 717                        requeue = false;
 718        }
 719
 720        /*
 721         * [4] Get the next lock
 722         */
 723        lock = waiter->lock;
 724        /*
 725         * [5] We need to trylock here as we are holding task->pi_lock,
 726         * which is the reverse lock order versus the other rtmutex
 727         * operations.
 728         */
 729        if (!raw_spin_trylock(&lock->wait_lock)) {
 730                raw_spin_unlock_irq(&task->pi_lock);
 731                cpu_relax();
 732                goto retry;
 733        }
 734
 735        /*
 736         * [6] check_exit_conditions_2() protected by task->pi_lock and
 737         * lock->wait_lock.
 738         *
 739         * Deadlock detection. If the lock is the same as the original
 740         * lock which caused us to walk the lock chain or if the
 741         * current lock is owned by the task which initiated the chain
 742         * walk, we detected a deadlock.
 743         */
 744        if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 745                ret = -EDEADLK;
 746
 747                /*
 748                 * When the deadlock is due to ww_mutex; also see above. Don't
 749                 * report the deadlock and instead let the ww_mutex wound/die
 750                 * logic pick which of the contending threads gets -EDEADLK.
 751                 *
 752                 * NOTE: assumes the cycle only contains a single ww_class; any
 753                 * other configuration and we fail to report; also, see
 754                 * lockdep.
 755                 */
 756                if (IS_ENABLED(CONFIG_PREEMPT_RT) && orig_waiter && orig_waiter->ww_ctx)
 757                        ret = 0;
 758
 759                raw_spin_unlock(&lock->wait_lock);
 760                goto out_unlock_pi;
 761        }
 762
 763        /*
 764         * If we just follow the lock chain for deadlock detection, no
 765         * need to do all the requeue operations. To avoid a truckload
 766         * of conditionals around the various places below, just do the
 767         * minimum chain walk checks.
 768         */
 769        if (!requeue) {
 770                /*
 771                 * No requeue[7] here. Just release @task [8]
 772                 */
 773                raw_spin_unlock(&task->pi_lock);
 774                put_task_struct(task);
 775
 776                /*
 777                 * [9] check_exit_conditions_3 protected by lock->wait_lock.
 778                 * If there is no owner of the lock, end of chain.
 779                 */
 780                if (!rt_mutex_owner(lock)) {
 781                        raw_spin_unlock_irq(&lock->wait_lock);
 782                        return 0;
 783                }
 784
 785                /* [10] Grab the next task, i.e. owner of @lock */
 786                task = get_task_struct(rt_mutex_owner(lock));
 787                raw_spin_lock(&task->pi_lock);
 788
 789                /*
 790                 * No requeue [11] here. We just do deadlock detection.
 791                 *
 792                 * [12] Store whether owner is blocked
 793                 * itself. Decision is made after dropping the locks
 794                 */
 795                next_lock = task_blocked_on_lock(task);
 796                /*
 797                 * Get the top waiter for the next iteration
 798                 */
 799                top_waiter = rt_mutex_top_waiter(lock);
 800
 801                /* [13] Drop locks */
 802                raw_spin_unlock(&task->pi_lock);
 803                raw_spin_unlock_irq(&lock->wait_lock);
 804
 805                /* If owner is not blocked, end of chain. */
 806                if (!next_lock)
 807                        goto out_put_task;
 808                goto again;
 809        }
 810
 811        /*
 812         * Store the current top waiter before doing the requeue
 813         * operation on @lock. We need it for the boost/deboost
 814         * decision below.
 815         */
 816        prerequeue_top_waiter = rt_mutex_top_waiter(lock);
 817
 818        /* [7] Requeue the waiter in the lock waiter tree. */
 819        rt_mutex_dequeue(lock, waiter);
 820
 821        /*
 822         * Update the waiter prio fields now that we're dequeued.
 823         *
 824         * These values can have changed through either:
 825         *
 826         *   sys_sched_set_scheduler() / sys_sched_setattr()
 827         *
 828         * or
 829         *
 830         *   DL CBS enforcement advancing the effective deadline.
 831         *
 832         * Even though pi_waiters also uses these fields, and that tree is only
 833         * updated in [11], we can do this here, since we hold [L], which
 834         * serializes all pi_waiters access and rb_erase() does not care about
 835         * the values of the node being removed.
 836         */
 837        waiter_update_prio(waiter, task);
 838
 839        rt_mutex_enqueue(lock, waiter);
 840
 841        /* [8] Release the task */
 842        raw_spin_unlock(&task->pi_lock);
 843        put_task_struct(task);
 844
 845        /*
 846         * [9] check_exit_conditions_3 protected by lock->wait_lock.
 847         *
 848         * We must abort the chain walk if there is no lock owner even
 849         * in the dead lock detection case, as we have nothing to
 850         * follow here. This is the end of the chain we are walking.
 851         */
 852        if (!rt_mutex_owner(lock)) {
 853                /*
 854                 * If the requeue [7] above changed the top waiter,
 855                 * then we need to wake the new top waiter up to try
 856                 * to get the lock.
 857                 */
 858                if (prerequeue_top_waiter != rt_mutex_top_waiter(lock))
 859                        wake_up_state(waiter->task, waiter->wake_state);
 860                raw_spin_unlock_irq(&lock->wait_lock);
 861                return 0;
 862        }
 863
 864        /* [10] Grab the next task, i.e. the owner of @lock */
 865        task = get_task_struct(rt_mutex_owner(lock));
 866        raw_spin_lock(&task->pi_lock);
 867
 868        /* [11] requeue the pi waiters if necessary */
 869        if (waiter == rt_mutex_top_waiter(lock)) {
 870                /*
 871                 * The waiter became the new top (highest priority)
 872                 * waiter on the lock. Replace the previous top waiter
 873                 * in the owner tasks pi waiters tree with this waiter
 874                 * and adjust the priority of the owner.
 875                 */
 876                rt_mutex_dequeue_pi(task, prerequeue_top_waiter);
 877                rt_mutex_enqueue_pi(task, waiter);
 878                rt_mutex_adjust_prio(task);
 879
 880        } else if (prerequeue_top_waiter == waiter) {
 881                /*
 882                 * The waiter was the top waiter on the lock, but is
 883                 * no longer the top priority waiter. Replace waiter in
 884                 * the owner tasks pi waiters tree with the new top
 885                 * (highest priority) waiter and adjust the priority
 886                 * of the owner.
 887                 * The new top waiter is stored in @waiter so that
 888                 * @waiter == @top_waiter evaluates to true below and
 889                 * we continue to deboost the rest of the chain.
 890                 */
 891                rt_mutex_dequeue_pi(task, waiter);
 892                waiter = rt_mutex_top_waiter(lock);
 893                rt_mutex_enqueue_pi(task, waiter);
 894                rt_mutex_adjust_prio(task);
 895        } else {
 896                /*
 897                 * Nothing changed. No need to do any priority
 898                 * adjustment.
 899                 */
 900        }
 901
 902        /*
 903         * [12] check_exit_conditions_4() protected by task->pi_lock
 904         * and lock->wait_lock. The actual decisions are made after we
 905         * dropped the locks.
 906         *
 907         * Check whether the task which owns the current lock is pi
 908         * blocked itself. If yes we store a pointer to the lock for
 909         * the lock chain change detection above. After we dropped
 910         * task->pi_lock next_lock cannot be dereferenced anymore.
 911         */
 912        next_lock = task_blocked_on_lock(task);
 913        /*
 914         * Store the top waiter of @lock for the end of chain walk
 915         * decision below.
 916         */
 917        top_waiter = rt_mutex_top_waiter(lock);
 918
 919        /* [13] Drop the locks */
 920        raw_spin_unlock(&task->pi_lock);
 921        raw_spin_unlock_irq(&lock->wait_lock);
 922
 923        /*
 924         * Make the actual exit decisions [12], based on the stored
 925         * values.
 926         *
 927         * We reached the end of the lock chain. Stop right here. No
 928         * point to go back just to figure that out.
 929         */
 930        if (!next_lock)
 931                goto out_put_task;
 932
 933        /*
 934         * If the current waiter is not the top waiter on the lock,
 935         * then we can stop the chain walk here if we are not in full
 936         * deadlock detection mode.
 937         */
 938        if (!detect_deadlock && waiter != top_waiter)
 939                goto out_put_task;
 940
 941        goto again;
 942
 943 out_unlock_pi:
 944        raw_spin_unlock_irq(&task->pi_lock);
 945 out_put_task:
 946        put_task_struct(task);
 947
 948        return ret;
 949}
 950
 951/*
 952 * Try to take an rt-mutex
 953 *
 954 * Must be called with lock->wait_lock held and interrupts disabled
 955 *
 956 * @lock:   The lock to be acquired.
 957 * @task:   The task which wants to acquire the lock
 958 * @waiter: The waiter that is queued to the lock's wait tree if the
 959 *          callsite called task_blocked_on_lock(), otherwise NULL
 960 */
 961static int __sched
 962try_to_take_rt_mutex(struct rt_mutex_base *lock, struct task_struct *task,
 963                     struct rt_mutex_waiter *waiter)
 964{
 965        lockdep_assert_held(&lock->wait_lock);
 966
 967        /*
 968         * Before testing whether we can acquire @lock, we set the
 969         * RT_MUTEX_HAS_WAITERS bit in @lock->owner. This forces all
 970         * other tasks which try to modify @lock into the slow path
 971         * and they serialize on @lock->wait_lock.
 972         *
 973         * The RT_MUTEX_HAS_WAITERS bit can have a transitional state
 974         * as explained at the top of this file if and only if:
 975         *
 976         * - There is a lock owner. The caller must fixup the
 977         *   transient state if it does a trylock or leaves the lock
 978         *   function due to a signal or timeout.
 979         *
 980         * - @task acquires the lock and there are no other
 981         *   waiters. This is undone in rt_mutex_set_owner(@task) at
 982         *   the end of this function.
 983         */
 984        mark_rt_mutex_waiters(lock);
 985
 986        /*
 987         * If @lock has an owner, give up.
 988         */
 989        if (rt_mutex_owner(lock))
 990                return 0;
 991
 992        /*
 993         * If @waiter != NULL, @task has already enqueued the waiter
 994         * into @lock waiter tree. If @waiter == NULL then this is a
 995         * trylock attempt.
 996         */
 997        if (waiter) {
 998                struct rt_mutex_waiter *top_waiter = rt_mutex_top_waiter(lock);
 999
1000                /*
1001                 * If waiter is the highest priority waiter of @lock,
1002                 * or allowed to steal it, take it over.
1003                 */
1004                if (waiter == top_waiter || rt_mutex_steal(waiter, top_waiter)) {
1005                        /*
1006                         * We can acquire the lock. Remove the waiter from the
1007                         * lock waiters tree.
1008                         */
1009                        rt_mutex_dequeue(lock, waiter);
1010                } else {
1011                        return 0;
1012                }
1013        } else {
1014                /*
1015                 * If the lock has waiters already we check whether @task is
1016                 * eligible to take over the lock.
1017                 *
1018                 * If there are no other waiters, @task can acquire
1019                 * the lock.  @task->pi_blocked_on is NULL, so it does
1020                 * not need to be dequeued.
1021                 */
1022                if (rt_mutex_has_waiters(lock)) {
1023                        /* Check whether the trylock can steal it. */
1024                        if (!rt_mutex_steal(task_to_waiter(task),
1025                                            rt_mutex_top_waiter(lock)))
1026                                return 0;
1027
1028                        /*
1029                         * The current top waiter stays enqueued. We
1030                         * don't have to change anything in the lock
1031                         * waiters order.
1032                         */
1033                } else {
1034                        /*
1035                         * No waiters. Take the lock without the
1036                         * pi_lock dance.@task->pi_blocked_on is NULL
1037                         * and we have no waiters to enqueue in @task
1038                         * pi waiters tree.
1039                         */
1040                        goto takeit;
1041                }
1042        }
1043
1044        /*
1045         * Clear @task->pi_blocked_on. Requires protection by
1046         * @task->pi_lock. Redundant operation for the @waiter == NULL
1047         * case, but conditionals are more expensive than a redundant
1048         * store.
1049         */
1050        raw_spin_lock(&task->pi_lock);
1051        task->pi_blocked_on = NULL;
1052        /*
1053         * Finish the lock acquisition. @task is the new owner. If
1054         * other waiters exist we have to insert the highest priority
1055         * waiter into @task->pi_waiters tree.
1056         */
1057        if (rt_mutex_has_waiters(lock))
1058                rt_mutex_enqueue_pi(task, rt_mutex_top_waiter(lock));
1059        raw_spin_unlock(&task->pi_lock);
1060
1061takeit:
1062        /*
1063         * This either preserves the RT_MUTEX_HAS_WAITERS bit if there
1064         * are still waiters or clears it.
1065         */
1066        rt_mutex_set_owner(lock, task);
1067
1068        return 1;
1069}
1070
1071/*
1072 * Task blocks on lock.
1073 *
1074 * Prepare waiter and propagate pi chain
1075 *
1076 * This must be called with lock->wait_lock held and interrupts disabled
1077 */
1078static int __sched task_blocks_on_rt_mutex(struct rt_mutex_base *lock,
1079                                           struct rt_mutex_waiter *waiter,
1080                                           struct task_struct *task,
1081                                           struct ww_acquire_ctx *ww_ctx,
1082                                           enum rtmutex_chainwalk chwalk)
1083{
1084        struct task_struct *owner = rt_mutex_owner(lock);
1085        struct rt_mutex_waiter *top_waiter = waiter;
1086        struct rt_mutex_base *next_lock;
1087        int chain_walk = 0, res;
1088
1089        lockdep_assert_held(&lock->wait_lock);
1090
1091        /*
1092         * Early deadlock detection. We really don't want the task to
1093         * enqueue on itself just to untangle the mess later. It's not
1094         * only an optimization. We drop the locks, so another waiter
1095         * can come in before the chain walk detects the deadlock. So
1096         * the other will detect the deadlock and return -EDEADLOCK,
1097         * which is wrong, as the other waiter is not in a deadlock
1098         * situation.
1099         */
1100        if (owner == task)
1101                return -EDEADLK;
1102
1103        raw_spin_lock(&task->pi_lock);
1104        waiter->task = task;
1105        waiter->lock = lock;
1106        waiter_update_prio(waiter, task);
1107
1108        /* Get the top priority waiter on the lock */
1109        if (rt_mutex_has_waiters(lock))
1110                top_waiter = rt_mutex_top_waiter(lock);
1111        rt_mutex_enqueue(lock, waiter);
1112
1113        task->pi_blocked_on = waiter;
1114
1115        raw_spin_unlock(&task->pi_lock);
1116
1117        if (build_ww_mutex() && ww_ctx) {
1118                struct rt_mutex *rtm;
1119
1120                /* Check whether the waiter should back out immediately */
1121                rtm = container_of(lock, struct rt_mutex, rtmutex);
1122                res = __ww_mutex_add_waiter(waiter, rtm, ww_ctx);
1123                if (res) {
1124                        raw_spin_lock(&task->pi_lock);
1125                        rt_mutex_dequeue(lock, waiter);
1126                        task->pi_blocked_on = NULL;
1127                        raw_spin_unlock(&task->pi_lock);
1128                        return res;
1129                }
1130        }
1131
1132        if (!owner)
1133                return 0;
1134
1135        raw_spin_lock(&owner->pi_lock);
1136        if (waiter == rt_mutex_top_waiter(lock)) {
1137                rt_mutex_dequeue_pi(owner, top_waiter);
1138                rt_mutex_enqueue_pi(owner, waiter);
1139
1140                rt_mutex_adjust_prio(owner);
1141                if (owner->pi_blocked_on)
1142                        chain_walk = 1;
1143        } else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
1144                chain_walk = 1;
1145        }
1146
1147        /* Store the lock on which owner is blocked or NULL */
1148        next_lock = task_blocked_on_lock(owner);
1149
1150        raw_spin_unlock(&owner->pi_lock);
1151        /*
1152         * Even if full deadlock detection is on, if the owner is not
1153         * blocked itself, we can avoid finding this out in the chain
1154         * walk.
1155         */
1156        if (!chain_walk || !next_lock)
1157                return 0;
1158
1159        /*
1160         * The owner can't disappear while holding a lock,
1161         * so the owner struct is protected by wait_lock.
1162         * Gets dropped in rt_mutex_adjust_prio_chain()!
1163         */
1164        get_task_struct(owner);
1165
1166        raw_spin_unlock_irq(&lock->wait_lock);
1167
1168        res = rt_mutex_adjust_prio_chain(owner, chwalk, lock,
1169                                         next_lock, waiter, task);
1170
1171        raw_spin_lock_irq(&lock->wait_lock);
1172
1173        return res;
1174}
1175
1176/*
1177 * Remove the top waiter from the current tasks pi waiter tree and
1178 * queue it up.
1179 *
1180 * Called with lock->wait_lock held and interrupts disabled.
1181 */
1182static void __sched mark_wakeup_next_waiter(struct rt_wake_q_head *wqh,
1183                                            struct rt_mutex_base *lock)
1184{
1185        struct rt_mutex_waiter *waiter;
1186
1187        raw_spin_lock(&current->pi_lock);
1188
1189        waiter = rt_mutex_top_waiter(lock);
1190
1191        /*
1192         * Remove it from current->pi_waiters and deboost.
1193         *
1194         * We must in fact deboost here in order to ensure we call
1195         * rt_mutex_setprio() to update p->pi_top_task before the
1196         * task unblocks.
1197         */
1198        rt_mutex_dequeue_pi(current, waiter);
1199        rt_mutex_adjust_prio(current);
1200
1201        /*
1202         * As we are waking up the top waiter, and the waiter stays
1203         * queued on the lock until it gets the lock, this lock
1204         * obviously has waiters. Just set the bit here and this has
1205         * the added benefit of forcing all new tasks into the
1206         * slow path making sure no task of lower priority than
1207         * the top waiter can steal this lock.
1208         */
1209        lock->owner = (void *) RT_MUTEX_HAS_WAITERS;
1210
1211        /*
1212         * We deboosted before waking the top waiter task such that we don't
1213         * run two tasks with the 'same' priority (and ensure the
1214         * p->pi_top_task pointer points to a blocked task). This however can
1215         * lead to priority inversion if we would get preempted after the
1216         * deboost but before waking our donor task, hence the preempt_disable()
1217         * before unlock.
1218         *
1219         * Pairs with preempt_enable() in rt_mutex_wake_up_q();
1220         */
1221        preempt_disable();
1222        rt_mutex_wake_q_add(wqh, waiter);
1223        raw_spin_unlock(&current->pi_lock);
1224}
1225
1226static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1227{
1228        int ret = try_to_take_rt_mutex(lock, current, NULL);
1229
1230        /*
1231         * try_to_take_rt_mutex() sets the lock waiters bit
1232         * unconditionally. Clean this up.
1233         */
1234        fixup_rt_mutex_waiters(lock);
1235
1236        return ret;
1237}
1238
1239/*
1240 * Slow path try-lock function:
1241 */
1242static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock)
1243{
1244        unsigned long flags;
1245        int ret;
1246
1247        /*
1248         * If the lock already has an owner we fail to get the lock.
1249         * This can be done without taking the @lock->wait_lock as
1250         * it is only being read, and this is a trylock anyway.
1251         */
1252        if (rt_mutex_owner(lock))
1253                return 0;
1254
1255        /*
1256         * The mutex has currently no owner. Lock the wait lock and try to
1257         * acquire the lock. We use irqsave here to support early boot calls.
1258         */
1259        raw_spin_lock_irqsave(&lock->wait_lock, flags);
1260
1261        ret = __rt_mutex_slowtrylock(lock);
1262
1263        raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1264
1265        return ret;
1266}
1267
1268static __always_inline int __rt_mutex_trylock(struct rt_mutex_base *lock)
1269{
1270        if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1271                return 1;
1272
1273        return rt_mutex_slowtrylock(lock);
1274}
1275
1276/*
1277 * Slow path to release a rt-mutex.
1278 */
1279static void __sched rt_mutex_slowunlock(struct rt_mutex_base *lock)
1280{
1281        DEFINE_RT_WAKE_Q(wqh);
1282        unsigned long flags;
1283
1284        /* irqsave required to support early boot calls */
1285        raw_spin_lock_irqsave(&lock->wait_lock, flags);
1286
1287        debug_rt_mutex_unlock(lock);
1288
1289        /*
1290         * We must be careful here if the fast path is enabled. If we
1291         * have no waiters queued we cannot set owner to NULL here
1292         * because of:
1293         *
1294         * foo->lock->owner = NULL;
1295         *                      rtmutex_lock(foo->lock);   <- fast path
1296         *                      free = atomic_dec_and_test(foo->refcnt);
1297         *                      rtmutex_unlock(foo->lock); <- fast path
1298         *                      if (free)
1299         *                              kfree(foo);
1300         * raw_spin_unlock(foo->lock->wait_lock);
1301         *
1302         * So for the fastpath enabled kernel:
1303         *
1304         * Nothing can set the waiters bit as long as we hold
1305         * lock->wait_lock. So we do the following sequence:
1306         *
1307         *      owner = rt_mutex_owner(lock);
1308         *      clear_rt_mutex_waiters(lock);
1309         *      raw_spin_unlock(&lock->wait_lock);
1310         *      if (cmpxchg(&lock->owner, owner, 0) == owner)
1311         *              return;
1312         *      goto retry;
1313         *
1314         * The fastpath disabled variant is simple as all access to
1315         * lock->owner is serialized by lock->wait_lock:
1316         *
1317         *      lock->owner = NULL;
1318         *      raw_spin_unlock(&lock->wait_lock);
1319         */
1320        while (!rt_mutex_has_waiters(lock)) {
1321                /* Drops lock->wait_lock ! */
1322                if (unlock_rt_mutex_safe(lock, flags) == true)
1323                        return;
1324                /* Relock the rtmutex and try again */
1325                raw_spin_lock_irqsave(&lock->wait_lock, flags);
1326        }
1327
1328        /*
1329         * The wakeup next waiter path does not suffer from the above
1330         * race. See the comments there.
1331         *
1332         * Queue the next waiter for wakeup once we release the wait_lock.
1333         */
1334        mark_wakeup_next_waiter(&wqh, lock);
1335        raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1336
1337        rt_mutex_wake_up_q(&wqh);
1338}
1339
1340static __always_inline void __rt_mutex_unlock(struct rt_mutex_base *lock)
1341{
1342        if (likely(rt_mutex_cmpxchg_release(lock, current, NULL)))
1343                return;
1344
1345        rt_mutex_slowunlock(lock);
1346}
1347
1348#ifdef CONFIG_SMP
1349static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1350                                  struct rt_mutex_waiter *waiter,
1351                                  struct task_struct *owner)
1352{
1353        bool res = true;
1354
1355        rcu_read_lock();
1356        for (;;) {
1357                /* If owner changed, trylock again. */
1358                if (owner != rt_mutex_owner(lock))
1359                        break;
1360                /*
1361                 * Ensure that @owner is dereferenced after checking that
1362                 * the lock owner still matches @owner. If that fails,
1363                 * @owner might point to freed memory. If it still matches,
1364                 * the rcu_read_lock() ensures the memory stays valid.
1365                 */
1366                barrier();
1367                /*
1368                 * Stop spinning when:
1369                 *  - the lock owner has been scheduled out
1370                 *  - current is not longer the top waiter
1371                 *  - current is requested to reschedule (redundant
1372                 *    for CONFIG_PREEMPT_RCU=y)
1373                 *  - the VCPU on which owner runs is preempted
1374                 */
1375                if (!owner->on_cpu || need_resched() ||
1376                    rt_mutex_waiter_is_top_waiter(lock, waiter) ||
1377                    vcpu_is_preempted(task_cpu(owner))) {
1378                        res = false;
1379                        break;
1380                }
1381                cpu_relax();
1382        }
1383        rcu_read_unlock();
1384        return res;
1385}
1386#else
1387static bool rtmutex_spin_on_owner(struct rt_mutex_base *lock,
1388                                  struct rt_mutex_waiter *waiter,
1389                                  struct task_struct *owner)
1390{
1391        return false;
1392}
1393#endif
1394
1395#ifdef RT_MUTEX_BUILD_MUTEX
1396/*
1397 * Functions required for:
1398 *      - rtmutex, futex on all kernels
1399 *      - mutex and rwsem substitutions on RT kernels
1400 */
1401
1402/*
1403 * Remove a waiter from a lock and give up
1404 *
1405 * Must be called with lock->wait_lock held and interrupts disabled. It must
1406 * have just failed to try_to_take_rt_mutex().
1407 */
1408static void __sched remove_waiter(struct rt_mutex_base *lock,
1409                                  struct rt_mutex_waiter *waiter)
1410{
1411        bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
1412        struct task_struct *owner = rt_mutex_owner(lock);
1413        struct rt_mutex_base *next_lock;
1414
1415        lockdep_assert_held(&lock->wait_lock);
1416
1417        raw_spin_lock(&current->pi_lock);
1418        rt_mutex_dequeue(lock, waiter);
1419        current->pi_blocked_on = NULL;
1420        raw_spin_unlock(&current->pi_lock);
1421
1422        /*
1423         * Only update priority if the waiter was the highest priority
1424         * waiter of the lock and there is an owner to update.
1425         */
1426        if (!owner || !is_top_waiter)
1427                return;
1428
1429        raw_spin_lock(&owner->pi_lock);
1430
1431        rt_mutex_dequeue_pi(owner, waiter);
1432
1433        if (rt_mutex_has_waiters(lock))
1434                rt_mutex_enqueue_pi(owner, rt_mutex_top_waiter(lock));
1435
1436        rt_mutex_adjust_prio(owner);
1437
1438        /* Store the lock on which owner is blocked or NULL */
1439        next_lock = task_blocked_on_lock(owner);
1440
1441        raw_spin_unlock(&owner->pi_lock);
1442
1443        /*
1444         * Don't walk the chain, if the owner task is not blocked
1445         * itself.
1446         */
1447        if (!next_lock)
1448                return;
1449
1450        /* gets dropped in rt_mutex_adjust_prio_chain()! */
1451        get_task_struct(owner);
1452
1453        raw_spin_unlock_irq(&lock->wait_lock);
1454
1455        rt_mutex_adjust_prio_chain(owner, RT_MUTEX_MIN_CHAINWALK, lock,
1456                                   next_lock, NULL, current);
1457
1458        raw_spin_lock_irq(&lock->wait_lock);
1459}
1460
1461/**
1462 * rt_mutex_slowlock_block() - Perform the wait-wake-try-to-take loop
1463 * @lock:                the rt_mutex to take
1464 * @ww_ctx:              WW mutex context pointer
1465 * @state:               the state the task should block in (TASK_INTERRUPTIBLE
1466 *                       or TASK_UNINTERRUPTIBLE)
1467 * @timeout:             the pre-initialized and started timer, or NULL for none
1468 * @waiter:              the pre-initialized rt_mutex_waiter
1469 *
1470 * Must be called with lock->wait_lock held and interrupts disabled
1471 */
1472static int __sched rt_mutex_slowlock_block(struct rt_mutex_base *lock,
1473                                           struct ww_acquire_ctx *ww_ctx,
1474                                           unsigned int state,
1475                                           struct hrtimer_sleeper *timeout,
1476                                           struct rt_mutex_waiter *waiter)
1477{
1478        struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1479        struct task_struct *owner;
1480        int ret = 0;
1481
1482        for (;;) {
1483                /* Try to acquire the lock: */
1484                if (try_to_take_rt_mutex(lock, current, waiter))
1485                        break;
1486
1487                if (timeout && !timeout->task) {
1488                        ret = -ETIMEDOUT;
1489                        break;
1490                }
1491                if (signal_pending_state(state, current)) {
1492                        ret = -EINTR;
1493                        break;
1494                }
1495
1496                if (build_ww_mutex() && ww_ctx) {
1497                        ret = __ww_mutex_check_kill(rtm, waiter, ww_ctx);
1498                        if (ret)
1499                                break;
1500                }
1501
1502                if (waiter == rt_mutex_top_waiter(lock))
1503                        owner = rt_mutex_owner(lock);
1504                else
1505                        owner = NULL;
1506                raw_spin_unlock_irq(&lock->wait_lock);
1507
1508                if (!owner || !rtmutex_spin_on_owner(lock, waiter, owner))
1509                        schedule();
1510
1511                raw_spin_lock_irq(&lock->wait_lock);
1512                set_current_state(state);
1513        }
1514
1515        __set_current_state(TASK_RUNNING);
1516        return ret;
1517}
1518
1519static void __sched rt_mutex_handle_deadlock(int res, int detect_deadlock,
1520                                             struct rt_mutex_waiter *w)
1521{
1522        /*
1523         * If the result is not -EDEADLOCK or the caller requested
1524         * deadlock detection, nothing to do here.
1525         */
1526        if (res != -EDEADLOCK || detect_deadlock)
1527                return;
1528
1529        if (build_ww_mutex() && w->ww_ctx)
1530                return;
1531
1532        /*
1533         * Yell loudly and stop the task right here.
1534         */
1535        WARN(1, "rtmutex deadlock detected\n");
1536        while (1) {
1537                set_current_state(TASK_INTERRUPTIBLE);
1538                schedule();
1539        }
1540}
1541
1542/**
1543 * __rt_mutex_slowlock - Locking slowpath invoked with lock::wait_lock held
1544 * @lock:       The rtmutex to block lock
1545 * @ww_ctx:     WW mutex context pointer
1546 * @state:      The task state for sleeping
1547 * @chwalk:     Indicator whether full or partial chainwalk is requested
1548 * @waiter:     Initializer waiter for blocking
1549 */
1550static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
1551                                       struct ww_acquire_ctx *ww_ctx,
1552                                       unsigned int state,
1553                                       enum rtmutex_chainwalk chwalk,
1554                                       struct rt_mutex_waiter *waiter)
1555{
1556        struct rt_mutex *rtm = container_of(lock, struct rt_mutex, rtmutex);
1557        struct ww_mutex *ww = ww_container_of(rtm);
1558        int ret;
1559
1560        lockdep_assert_held(&lock->wait_lock);
1561
1562        /* Try to acquire the lock again: */
1563        if (try_to_take_rt_mutex(lock, current, NULL)) {
1564                if (build_ww_mutex() && ww_ctx) {
1565                        __ww_mutex_check_waiters(rtm, ww_ctx);
1566                        ww_mutex_lock_acquired(ww, ww_ctx);
1567                }
1568                return 0;
1569        }
1570
1571        set_current_state(state);
1572
1573        ret = task_blocks_on_rt_mutex(lock, waiter, current, ww_ctx, chwalk);
1574        if (likely(!ret))
1575                ret = rt_mutex_slowlock_block(lock, ww_ctx, state, NULL, waiter);
1576
1577        if (likely(!ret)) {
1578                /* acquired the lock */
1579                if (build_ww_mutex() && ww_ctx) {
1580                        if (!ww_ctx->is_wait_die)
1581                                __ww_mutex_check_waiters(rtm, ww_ctx);
1582                        ww_mutex_lock_acquired(ww, ww_ctx);
1583                }
1584        } else {
1585                __set_current_state(TASK_RUNNING);
1586                remove_waiter(lock, waiter);
1587                rt_mutex_handle_deadlock(ret, chwalk, waiter);
1588        }
1589
1590        /*
1591         * try_to_take_rt_mutex() sets the waiter bit
1592         * unconditionally. We might have to fix that up.
1593         */
1594        fixup_rt_mutex_waiters(lock);
1595        return ret;
1596}
1597
1598static inline int __rt_mutex_slowlock_locked(struct rt_mutex_base *lock,
1599                                             struct ww_acquire_ctx *ww_ctx,
1600                                             unsigned int state)
1601{
1602        struct rt_mutex_waiter waiter;
1603        int ret;
1604
1605        rt_mutex_init_waiter(&waiter);
1606        waiter.ww_ctx = ww_ctx;
1607
1608        ret = __rt_mutex_slowlock(lock, ww_ctx, state, RT_MUTEX_MIN_CHAINWALK,
1609                                  &waiter);
1610
1611        debug_rt_mutex_free_waiter(&waiter);
1612        return ret;
1613}
1614
1615/*
1616 * rt_mutex_slowlock - Locking slowpath invoked when fast path fails
1617 * @lock:       The rtmutex to block lock
1618 * @ww_ctx:     WW mutex context pointer
1619 * @state:      The task state for sleeping
1620 */
1621static int __sched rt_mutex_slowlock(struct rt_mutex_base *lock,
1622                                     struct ww_acquire_ctx *ww_ctx,
1623                                     unsigned int state)
1624{
1625        unsigned long flags;
1626        int ret;
1627
1628        /*
1629         * Technically we could use raw_spin_[un]lock_irq() here, but this can
1630         * be called in early boot if the cmpxchg() fast path is disabled
1631         * (debug, no architecture support). In this case we will acquire the
1632         * rtmutex with lock->wait_lock held. But we cannot unconditionally
1633         * enable interrupts in that early boot case. So we need to use the
1634         * irqsave/restore variants.
1635         */
1636        raw_spin_lock_irqsave(&lock->wait_lock, flags);
1637        ret = __rt_mutex_slowlock_locked(lock, ww_ctx, state);
1638        raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1639
1640        return ret;
1641}
1642
1643static __always_inline int __rt_mutex_lock(struct rt_mutex_base *lock,
1644                                           unsigned int state)
1645{
1646        if (likely(rt_mutex_cmpxchg_acquire(lock, NULL, current)))
1647                return 0;
1648
1649        return rt_mutex_slowlock(lock, NULL, state);
1650}
1651#endif /* RT_MUTEX_BUILD_MUTEX */
1652
1653#ifdef RT_MUTEX_BUILD_SPINLOCKS
1654/*
1655 * Functions required for spin/rw_lock substitution on RT kernels
1656 */
1657
1658/**
1659 * rtlock_slowlock_locked - Slow path lock acquisition for RT locks
1660 * @lock:       The underlying RT mutex
1661 */
1662static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock)
1663{
1664        struct rt_mutex_waiter waiter;
1665        struct task_struct *owner;
1666
1667        lockdep_assert_held(&lock->wait_lock);
1668
1669        if (try_to_take_rt_mutex(lock, current, NULL))
1670                return;
1671
1672        rt_mutex_init_rtlock_waiter(&waiter);
1673
1674        /* Save current state and set state to TASK_RTLOCK_WAIT */
1675        current_save_and_set_rtlock_wait_state();
1676
1677        task_blocks_on_rt_mutex(lock, &waiter, current, NULL, RT_MUTEX_MIN_CHAINWALK);
1678
1679        for (;;) {
1680                /* Try to acquire the lock again */
1681                if (try_to_take_rt_mutex(lock, current, &waiter))
1682                        break;
1683
1684                if (&waiter == rt_mutex_top_waiter(lock))
1685                        owner = rt_mutex_owner(lock);
1686                else
1687                        owner = NULL;
1688                raw_spin_unlock_irq(&lock->wait_lock);
1689
1690                if (!owner || !rtmutex_spin_on_owner(lock, &waiter, owner))
1691                        schedule_rtlock();
1692
1693                raw_spin_lock_irq(&lock->wait_lock);
1694                set_current_state(TASK_RTLOCK_WAIT);
1695        }
1696
1697        /* Restore the task state */
1698        current_restore_rtlock_saved_state();
1699
1700        /*
1701         * try_to_take_rt_mutex() sets the waiter bit unconditionally.
1702         * We might have to fix that up:
1703         */
1704        fixup_rt_mutex_waiters(lock);
1705        debug_rt_mutex_free_waiter(&waiter);
1706}
1707
1708static __always_inline void __sched rtlock_slowlock(struct rt_mutex_base *lock)
1709{
1710        unsigned long flags;
1711
1712        raw_spin_lock_irqsave(&lock->wait_lock, flags);
1713        rtlock_slowlock_locked(lock);
1714        raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
1715}
1716
1717#endif /* RT_MUTEX_BUILD_SPINLOCKS */
1718