kernel/locking/rwsem.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* kernel/rwsem.c: R/W semaphores, public implementation
   3  *
   4  * Written by David Howells (dhowells@redhat.com).
   5  * Derived from asm-i386/semaphore.h
   6  *
   7  * Writer lock-stealing by Alex Shi <alex.shi@intel.com>
   8  * and Michel Lespinasse <walken@google.com>
   9  *
  10  * Optimistic spinning by Tim Chen <tim.c.chen@intel.com>
  11  * and Davidlohr Bueso <davidlohr@hp.com>. Based on mutexes.
  12  *
  13  * Rwsem count bit fields re-definition and rwsem rearchitecture
  14  * by Waiman Long <longman@redhat.com>.
  15  */
  16
  17 #include <linux/types.h>
  18 #include <linux/kernel.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/rt.h>
  21 #include <linux/sched/task.h>
  22 #include <linux/sched/debug.h>
  23 #include <linux/sched/wake_q.h>
  24 #include <linux/sched/signal.h>
  25 #include <linux/export.h>
  26 #include <linux/rwsem.h>
  27 #include <linux/atomic.h>
  28
  29 #include "rwsem.h"
  30 #include "lock_events.h"
  31
  32 /*
  33  * The least significant 2 bits of the owner value has the following
  34  * meanings when set.
  35  *  - RWSEM_READER_OWNED (bit 0): The rwsem is owned by readers
  36  *  - RWSEM_ANONYMOUSLY_OWNED (bit 1): The rwsem is anonymously owned,
  37  *    i.e. the owner(s) cannot be readily determined. It can be reader
  38  *    owned or the owning writer is indeterminate.
  39  *
  40  * When a writer acquires a rwsem, it puts its task_struct pointer
  41  * into the owner field. It is cleared after an unlock.
  42  *
  43  * When a reader acquires a rwsem, it will also puts its task_struct
  44  * pointer into the owner field with both the RWSEM_READER_OWNED and
  45  * RWSEM_ANONYMOUSLY_OWNED bits set. On unlock, the owner field will
  46  * largely be left untouched. So for a free or reader-owned rwsem,
  47  * the owner value may contain information about the last reader that
  48  * acquires the rwsem. The anonymous bit is set because that particular
  49  * reader may or may not still own the lock.
  50  *
  51  * That information may be helpful in debugging cases where the system
  52  * seems to hang on a reader owned rwsem especially if only one reader
  53  * is involved. Ideally we would like to track all the readers that own
  54  * a rwsem, but the overhead is simply too big.
  55  */
  56 #define RWSEM_READER_OWNED      (1UL << 0)
  57 #define RWSEM_ANONYMOUSLY_OWNED (1UL << 1)
  58
  59 #ifdef CONFIG_DEBUG_RWSEMS
  60 # define DEBUG_RWSEMS_WARN_ON(c, sem)   do {                    \
  61         if (!debug_locks_silent &&                              \
  62             WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
  63                 #c, atomic_long_read(&(sem)->count),            \
  64                 (long)((sem)->owner), (long)current,            \
  65                 list_empty(&(sem)->wait_list) ? "" : "not "))   \
  66                         debug_locks_off();                      \
  67         } while (0)
  68 #else
  69 # define DEBUG_RWSEMS_WARN_ON(c, sem)
  70 #endif
  71
  72 /*
  73  * The definition of the atomic counter in the semaphore:
  74  *
  75  * Bit  0   - writer locked bit
  76  * Bit  1   - waiters present bit
  77  * Bits 2-7 - reserved
  78  * Bits 8-X - 24-bit (32-bit) or 56-bit reader count
  79  *
  80  * atomic_long_fetch_add() is used to obtain reader lock, whereas
  81  * atomic_long_cmpxchg() will be used to obtain writer lock.
  82  */
  83 #define RWSEM_WRITER_LOCKED     (1UL << 0)
  84 #define RWSEM_FLAG_WAITERS      (1UL << 1)
  85 #define RWSEM_READER_SHIFT      8
  86 #define RWSEM_READER_BIAS       (1UL << RWSEM_READER_SHIFT)
  87 #define RWSEM_READER_MASK       (~(RWSEM_READER_BIAS - 1))
  88 #define RWSEM_WRITER_MASK       RWSEM_WRITER_LOCKED
  89 #define RWSEM_LOCK_MASK         (RWSEM_WRITER_MASK|RWSEM_READER_MASK)
  90 #define RWSEM_READ_FAILED_MASK  (RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS)
  91
  92 /*
  93  * All writes to owner are protected by WRITE_ONCE() to make sure that
  94  * store tearing can't happen as optimistic spinners may read and use
  95  * the owner value concurrently without lock. Read from owner, however,
  96  * may not need READ_ONCE() as long as the pointer value is only used
  97  * for comparison and isn't being dereferenced.
  98  */
  99 static inline void rwsem_set_owner(struct rw_semaphore *sem)
 100 {
 101         WRITE_ONCE(sem->owner, current);
 102 }
 103
 104 static inline void rwsem_clear_owner(struct rw_semaphore *sem)
 105 {
 106         WRITE_ONCE(sem->owner, NULL);
 107 }
 108
 109 /*
 110  * The task_struct pointer of the last owning reader will be left in
 111  * the owner field.
 112  *
 113  * Note that the owner value just indicates the task has owned the rwsem
 114  * previously, it may not be the real owner or one of the real owners
 115  * anymore when that field is examined, so take it with a grain of salt.
 116  */
 117 static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
 118                                             struct task_struct *owner)
 119 {
 120         unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED
 121                                                  | RWSEM_ANONYMOUSLY_OWNED;
 122
 123         WRITE_ONCE(sem->owner, (struct task_struct *)val);
 124 }
 125
 126 static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
 127 {
 128         __rwsem_set_reader_owned(sem, current);
 129 }
 130
 131 /*
 132  * Return true if the a rwsem waiter can spin on the rwsem's owner
 133  * and steal the lock, i.e. the lock is not anonymously owned.
 134  * N.B. !owner is considered spinnable.
 135  */
 136 static inline bool is_rwsem_owner_spinnable(struct task_struct *owner)
 137 {
 138         return !((unsigned long)owner & RWSEM_ANONYMOUSLY_OWNED);
 139 }
 140
 141 /*
 142  * Return true if rwsem is owned by an anonymous writer or readers.
 143  */
 144 static inline bool rwsem_has_anonymous_owner(struct task_struct *owner)
 145 {
 146         return (unsigned long)owner & RWSEM_ANONYMOUSLY_OWNED;
 147 }
 148
 149 #ifdef CONFIG_DEBUG_RWSEMS
 150 /*
 151  * With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
 152  * is a task pointer in owner of a reader-owned rwsem, it will be the
 153  * real owner or one of the real owners. The only exception is when the
 154  * unlock is done by up_read_non_owner().
 155  */
 156 static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
 157 {
 158         unsigned long val = (unsigned long)current | RWSEM_READER_OWNED
 159                                                    | RWSEM_ANONYMOUSLY_OWNED;
 160         if (READ_ONCE(sem->owner) == (struct task_struct *)val)
 161                 cmpxchg_relaxed((unsigned long *)&sem->owner, val,
 162                                 RWSEM_READER_OWNED | RWSEM_ANONYMOUSLY_OWNED);
 163 }
 164 #else
 165 static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
 166 {
 167 }
 168 #endif
 169
 170 /*
 171  * Guide to the rw_semaphore's count field.
 172  *
 173  * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
 174  * by a writer.
 175  *
 176  * The lock is owned by readers when
 177  * (1) the RWSEM_WRITER_LOCKED isn't set in count,
 178  * (2) some of the reader bits are set in count, and
 179  * (3) the owner field has RWSEM_READ_OWNED bit set.
 180  *
 181  * Having some reader bits set is not enough to guarantee a readers owned
 182  * lock as the readers may be in the process of backing out from the count
 183  * and a writer has just released the lock. So another writer may steal
 184  * the lock immediately after that.
 185  */
 186
 187 /*
 188  * Initialize an rwsem:
 189  */
 190 void __init_rwsem(struct rw_semaphore *sem, const char *name,
 191                   struct lock_class_key *key)
 192 {
 193 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 194         /*
 195          * Make sure we are not reinitializing a held semaphore:
 196          */
 197         debug_check_no_locks_freed((void *)sem, sizeof(*sem));
 198         lockdep_init_map(&sem->dep_map, name, key, 0);
 199 #endif
 200         atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
 201         raw_spin_lock_init(&sem->wait_lock);
 202         INIT_LIST_HEAD(&sem->wait_list);
 203         sem->owner = NULL;
 204 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
 205         osq_lock_init(&sem->osq);
 206 #endif
 207 }
 208 EXPORT_SYMBOL(__init_rwsem);
 209
 210 enum rwsem_waiter_type {
 211         RWSEM_WAITING_FOR_WRITE,
 212         RWSEM_WAITING_FOR_READ
 213 };
 214
 215 struct rwsem_waiter {
 216         struct list_head list;
 217         struct task_struct *task;
 218         enum rwsem_waiter_type type;
 219 };
 220
 221 enum rwsem_wake_type {
 222         RWSEM_WAKE_ANY,         /* Wake whatever's at head of wait list */
 223         RWSEM_WAKE_READERS,     /* Wake readers only */
 224         RWSEM_WAKE_READ_OWNED   /* Waker thread holds the read lock */
 225 };
 226
 227 /*
 228  * handle the lock release when processes blocked on it that can now run
 229  * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
 230  *   have been set.
 231  * - there must be someone on the queue
 232  * - the wait_lock must be held by the caller
 233  * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
 234  *   to actually wakeup the blocked task(s) and drop the reference count,
 235  *   preferably when the wait_lock is released
 236  * - woken process blocks are discarded from the list after having task zeroed
 237  * - writers are only marked woken if downgrading is false
 238  */
 239 static void rwsem_mark_wake(struct rw_semaphore *sem,
 240                             enum rwsem_wake_type wake_type,
 241                             struct wake_q_head *wake_q)
 242 {
 243         struct rwsem_waiter *waiter, *tmp;
 244         long oldcount, woken = 0, adjustment = 0;
 245         struct list_head wlist;
 246
 247         /*
 248          * Take a peek at the queue head waiter such that we can determine
 249          * the wakeup(s) to perform.
 250          */
 251         waiter = list_first_entry(&sem->wait_list, struct rwsem_waiter, list);
 252
 253         if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
 254                 if (wake_type == RWSEM_WAKE_ANY) {
 255                         /*
 256                          * Mark writer at the front of the queue for wakeup.
 257                          * Until the task is actually later awoken later by
 258                          * the caller, other writers are able to steal it.
 259                          * Readers, on the other hand, will block as they
 260                          * will notice the queued writer.
 261                          */
 262                         wake_q_add(wake_q, waiter->task);
 263                         lockevent_inc(rwsem_wake_writer);
 264                 }
 265
 266                 return;
 267         }
 268
 269         /*
 270          * Writers might steal the lock before we grant it to the next reader.
 271          * We prefer to do the first reader grant before counting readers
 272          * so we can bail out early if a writer stole the lock.
 273          */
 274         if (wake_type != RWSEM_WAKE_READ_OWNED) {
 275                 adjustment = RWSEM_READER_BIAS;
 276                 oldcount = atomic_long_fetch_add(adjustment, &sem->count);
 277                 if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
 278                         atomic_long_sub(adjustment, &sem->count);
 279                         return;
 280                 }
 281                 /*
 282                  * Set it to reader-owned to give spinners an early
 283                  * indication that readers now have the lock.
 284                  */
 285                 __rwsem_set_reader_owned(sem, waiter->task);
 286         }
 287
 288         /*
 289          * Grant an infinite number of read locks to the readers at the front
 290          * of the queue. We know that woken will be at least 1 as we accounted
 291          * for above. Note we increment the 'active part' of the count by the
 292          * number of readers before waking any processes up.
 293          *
 294          * We have to do wakeup in 2 passes to prevent the possibility that
 295          * the reader count may be decremented before it is incremented. It
 296          * is because the to-be-woken waiter may not have slept yet. So it
 297          * may see waiter->task got cleared, finish its critical section and
 298          * do an unlock before the reader count increment.
 299          *
 300          * 1) Collect the read-waiters in a separate list, count them and
 301          *    fully increment the reader count in rwsem.
 302          * 2) For each waiters in the new list, clear waiter->task and
 303          *    put them into wake_q to be woken up later.
 304          */
 305         list_for_each_entry(waiter, &sem->wait_list, list) {
 306                 if (waiter->type == RWSEM_WAITING_FOR_WRITE)
 307                         break;
 308
 309                 woken++;
 310         }
 311         list_cut_before(&wlist, &sem->wait_list, &waiter->list);
 312
 313         adjustment = woken * RWSEM_READER_BIAS - adjustment;
 314         lockevent_cond_inc(rwsem_wake_reader, woken);
 315         if (list_empty(&sem->wait_list)) {
 316                 /* hit end of list above */
 317                 adjustment -= RWSEM_FLAG_WAITERS;
 318         }
 319
 320         if (adjustment)
 321                 atomic_long_add(adjustment, &sem->count);
 322
 323         /* 2nd pass */
 324         list_for_each_entry_safe(waiter, tmp, &wlist, list) {
 325                 struct task_struct *tsk;
 326
 327                 tsk = waiter->task;
 328                 get_task_struct(tsk);
 329
 330                 /*
 331                  * Ensure calling get_task_struct() before setting the reader
 332                  * waiter to nil such that rwsem_down_read_slowpath() cannot
 333                  * race with do_exit() by always holding a reference count
 334                  * to the task to wakeup.
 335                  */
 336                 smp_store_release(&waiter->task, NULL);
 337                 /*
 338                  * Ensure issuing the wakeup (either by us or someone else)
 339                  * after setting the reader waiter to nil.
 340                  */
 341                 wake_q_add_safe(wake_q, tsk);
 342         }
 343 }
 344
 345 /*
 346  * This function must be called with the sem->wait_lock held to prevent
 347  * race conditions between checking the rwsem wait list and setting the
 348  * sem->count accordingly.
 349  */
 350 static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem)
 351 {
 352         long new;
 353
 354         if (count & RWSEM_LOCK_MASK)
 355                 return false;
 356
 357         new = count + RWSEM_WRITER_LOCKED -
 358              (list_is_singular(&sem->wait_list) ? RWSEM_FLAG_WAITERS : 0);
 359
 360         if (atomic_long_try_cmpxchg_acquire(&sem->count, &count, new)) {
 361                 rwsem_set_owner(sem);
 362                 return true;
 363         }
 364
 365         return false;
 366 }
 367
 368 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
 369 /*
 370  * Try to acquire write lock before the writer has been put on wait queue.
 371  */
 372 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
 373 {
 374         long count = atomic_long_read(&sem->count);
 375
 376         while (!(count & RWSEM_LOCK_MASK)) {
 377                 if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
 378                                         count + RWSEM_WRITER_LOCKED)) {
 379                         rwsem_set_owner(sem);
 380                         lockevent_inc(rwsem_opt_wlock);
 381                         return true;
 382                 }
 383         }
 384         return false;
 385 }
 386
 387 static inline bool owner_on_cpu(struct task_struct *owner)
 388 {
 389         /*
 390          * As lock holder preemption issue, we both skip spinning if
 391          * task is not on cpu or its cpu is preempted
 392          */
 393         return owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
 394 }
 395
 396 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
 397 {
 398         struct task_struct *owner;
 399         bool ret = true;
 400
 401         BUILD_BUG_ON(!rwsem_has_anonymous_owner(RWSEM_OWNER_UNKNOWN));
 402
 403         if (need_resched())
 404                 return false;
 405
 406         rcu_read_lock();
 407         owner = READ_ONCE(sem->owner);
 408         if (owner) {
 409                 ret = is_rwsem_owner_spinnable(owner) &&
 410                       owner_on_cpu(owner);
 411         }
 412         rcu_read_unlock();
 413         return ret;
 414 }
 415
 416 /*
 417  * The rwsem_spin_on_owner() function returns the folowing 4 values
 418  * depending on the lock owner state.
 419  *   OWNER_NULL  : owner is currently NULL
 420  *   OWNER_WRITER: when owner changes and is a writer
 421  *   OWNER_READER: when owner changes and the new owner may be a reader.
 422  *   OWNER_NONSPINNABLE:
 423  *                 when optimistic spinning has to stop because either the
 424  *                 owner stops running, is unknown, or its timeslice has
 425  *                 been used up.
 426  */
 427 enum owner_state {
 428         OWNER_NULL              = 1 << 0,
 429         OWNER_WRITER            = 1 << 1,
 430         OWNER_READER            = 1 << 2,
 431         OWNER_NONSPINNABLE      = 1 << 3,
 432 };
 433 #define OWNER_SPINNABLE         (OWNER_NULL | OWNER_WRITER)
 434
 435 static inline enum owner_state rwsem_owner_state(unsigned long owner)
 436 {
 437         if (!owner)
 438                 return OWNER_NULL;
 439
 440         if (owner & RWSEM_ANONYMOUSLY_OWNED)
 441                 return OWNER_NONSPINNABLE;
 442
 443         if (owner & RWSEM_READER_OWNED)
 444                 return OWNER_READER;
 445
 446         return OWNER_WRITER;
 447 }
 448
 449 static noinline enum owner_state rwsem_spin_on_owner(struct rw_semaphore *sem)
 450 {
 451         struct task_struct *tmp, *owner = READ_ONCE(sem->owner);
 452         enum owner_state state = rwsem_owner_state((unsigned long)owner);
 453
 454         if (state != OWNER_WRITER)
 455                 return state;
 456
 457         rcu_read_lock();
 458         for (;;) {
 459                 tmp = READ_ONCE(sem->owner);
 460                 if (tmp != owner) {
 461                         state = rwsem_owner_state((unsigned long)tmp);
 462                         break;
 463                 }
 464
 465                 /*
 466                  * Ensure we emit the owner->on_cpu, dereference _after_
 467                  * checking sem->owner still matches owner, if that fails,
 468                  * owner might point to free()d memory, if it still matches,
 469                  * the rcu_read_lock() ensures the memory stays valid.
 470                  */
 471                 barrier();
 472
 473                 if (need_resched() || !owner_on_cpu(owner)) {
 474                         state = OWNER_NONSPINNABLE;
 475                         break;
 476                 }
 477
 478                 cpu_relax();
 479         }
 480         rcu_read_unlock();
 481
 482         return state;
 483 }
 484
 485 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
 486 {
 487         bool taken = false;
 488
 489         preempt_disable();
 490
 491         /* sem->wait_lock should not be held when doing optimistic spinning */
 492         if (!rwsem_can_spin_on_owner(sem))
 493                 goto done;
 494
 495         if (!osq_lock(&sem->osq))
 496                 goto done;
 497
 498         /*
 499          * Optimistically spin on the owner field and attempt to acquire the
 500          * lock whenever the owner changes. Spinning will be stopped when:
 501          *  1) the owning writer isn't running; or
 502          *  2) readers own the lock as we can't determine if they are
 503          *     actively running or not.
 504          */
 505         while (rwsem_spin_on_owner(sem) & OWNER_SPINNABLE) {
 506                 /*
 507                  * Try to acquire the lock
 508                  */
 509                 if (rwsem_try_write_lock_unqueued(sem)) {
 510                         taken = true;
 511                         break;
 512                 }
 513
 514                 /*
 515                  * When there's no owner, we might have preempted between the
 516                  * owner acquiring the lock and setting the owner field. If
 517                  * we're an RT task that will live-lock because we won't let
 518                  * the owner complete.
 519                  */
 520                 if (!sem->owner && (need_resched() || rt_task(current)))
 521                         break;
 522
 523                 /*
 524                  * The cpu_relax() call is a compiler barrier which forces
 525                  * everything in this loop to be re-loaded. We don't need
 526                  * memory barriers as we'll eventually observe the right
 527                  * values at the cost of a few extra spins.
 528                  */
 529                 cpu_relax();
 530         }
 531         osq_unlock(&sem->osq);
 532 done:
 533         preempt_enable();
 534         lockevent_cond_inc(rwsem_opt_fail, !taken);
 535         return taken;
 536 }
 537 #else
 538 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
 539 {
 540         return false;
 541 }
 542 #endif
 543
 544 /*
 545  * Wait for the read lock to be granted
 546  */
 547 static struct rw_semaphore __sched *
 548 rwsem_down_read_slowpath(struct rw_semaphore *sem, int state)
 549 {
 550         long count, adjustment = -RWSEM_READER_BIAS;
 551         struct rwsem_waiter waiter;
 552         DEFINE_WAKE_Q(wake_q);
 553
 554         waiter.task = current;
 555         waiter.type = RWSEM_WAITING_FOR_READ;
 556
 557         raw_spin_lock_irq(&sem->wait_lock);
 558         if (list_empty(&sem->wait_list)) {
 559                 /*
 560                  * In case the wait queue is empty and the lock isn't owned
 561                  * by a writer, this reader can exit the slowpath and return
 562                  * immediately as its RWSEM_READER_BIAS has already been
 563                  * set in the count.
 564                  */
 565                 if (!(atomic_long_read(&sem->count) & RWSEM_WRITER_MASK)) {
 566                         raw_spin_unlock_irq(&sem->wait_lock);
 567                         rwsem_set_reader_owned(sem);
 568                         lockevent_inc(rwsem_rlock_fast);
 569                         return sem;
 570                 }
 571                 adjustment += RWSEM_FLAG_WAITERS;
 572         }
 573         list_add_tail(&waiter.list, &sem->wait_list);
 574
 575         /* we're now waiting on the lock, but no longer actively locking */
 576         count = atomic_long_add_return(adjustment, &sem->count);
 577
 578         /*
 579          * If there are no active locks, wake the front queued process(es).
 580          *
 581          * If there are no writers and we are first in the queue,
 582          * wake our own waiter to join the existing active readers !
 583          */
 584         if (!(count & RWSEM_LOCK_MASK) ||
 585            (!(count & RWSEM_WRITER_MASK) && (adjustment & RWSEM_FLAG_WAITERS)))
 586                 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
 587
 588         raw_spin_unlock_irq(&sem->wait_lock);
 589         wake_up_q(&wake_q);
 590
 591         /* wait to be given the lock */
 592         while (true) {
 593                 set_current_state(state);
 594                 if (!waiter.task)
 595                         break;
 596                 if (signal_pending_state(state, current)) {
 597                         raw_spin_lock_irq(&sem->wait_lock);
 598                         if (waiter.task)
 599                                 goto out_nolock;
 600                         raw_spin_unlock_irq(&sem->wait_lock);
 601                         break;
 602                 }
 603                 schedule();
 604                 lockevent_inc(rwsem_sleep_reader);
 605         }
 606
 607         __set_current_state(TASK_RUNNING);
 608         lockevent_inc(rwsem_rlock);
 609         return sem;
 610 out_nolock:
 611         list_del(&waiter.list);
 612         if (list_empty(&sem->wait_list))
 613                 atomic_long_andnot(RWSEM_FLAG_WAITERS, &sem->count);
 614         raw_spin_unlock_irq(&sem->wait_lock);
 615         __set_current_state(TASK_RUNNING);
 616         lockevent_inc(rwsem_rlock_fail);
 617         return ERR_PTR(-EINTR);
 618 }
 619
 620 /*
 621  * Wait until we successfully acquire the write lock
 622  */
 623 static struct rw_semaphore *
 624 rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
 625 {
 626         long count;
 627         bool waiting = true; /* any queued threads before us */
 628         struct rwsem_waiter waiter;
 629         struct rw_semaphore *ret = sem;
 630         DEFINE_WAKE_Q(wake_q);
 631
 632         /* do optimistic spinning and steal lock if possible */
 633         if (rwsem_optimistic_spin(sem))
 634                 return sem;
 635
 636         /*
 637          * Optimistic spinning failed, proceed to the slowpath
 638          * and block until we can acquire the sem.
 639          */
 640         waiter.task = current;
 641         waiter.type = RWSEM_WAITING_FOR_WRITE;
 642
 643         raw_spin_lock_irq(&sem->wait_lock);
 644
 645         /* account for this before adding a new element to the list */
 646         if (list_empty(&sem->wait_list))
 647                 waiting = false;
 648
 649         list_add_tail(&waiter.list, &sem->wait_list);
 650
 651         /* we're now waiting on the lock */
 652         if (waiting) {
 653                 count = atomic_long_read(&sem->count);
 654
 655                 /*
 656                  * If there were already threads queued before us and there are
 657                  * no active writers and some readers, the lock must be read
 658                  * owned; so we try to  any read locks that were queued ahead
 659                  * of us.
 660                  */
 661                 if (!(count & RWSEM_WRITER_MASK) &&
 662                      (count & RWSEM_READER_MASK)) {
 663                         rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q);
 664                         /*
 665                          * The wakeup is normally called _after_ the wait_lock
 666                          * is released, but given that we are proactively waking
 667                          * readers we can deal with the wake_q overhead as it is
 668                          * similar to releasing and taking the wait_lock again
 669                          * for attempting rwsem_try_write_lock().
 670                          */
 671                         wake_up_q(&wake_q);
 672
 673                         /*
 674                          * Reinitialize wake_q after use.
 675                          */
 676                         wake_q_init(&wake_q);
 677                 }
 678
 679         } else {
 680                 count = atomic_long_add_return(RWSEM_FLAG_WAITERS, &sem->count);
 681         }
 682
 683         /* wait until we successfully acquire the lock */
 684         set_current_state(state);
 685         while (true) {
 686                 if (rwsem_try_write_lock(count, sem))
 687                         break;
 688                 raw_spin_unlock_irq(&sem->wait_lock);
 689
 690                 /* Block until there are no active lockers. */
 691                 do {
 692                         if (signal_pending_state(state, current))
 693                                 goto out_nolock;
 694
 695                         schedule();
 696                         lockevent_inc(rwsem_sleep_writer);
 697                         set_current_state(state);
 698                         count = atomic_long_read(&sem->count);
 699                 } while (count & RWSEM_LOCK_MASK);
 700
 701                 raw_spin_lock_irq(&sem->wait_lock);
 702         }
 703         __set_current_state(TASK_RUNNING);
 704         list_del(&waiter.list);
 705         raw_spin_unlock_irq(&sem->wait_lock);
 706         lockevent_inc(rwsem_wlock);
 707
 708         return ret;
 709
 710 out_nolock:
 711         __set_current_state(TASK_RUNNING);
 712         raw_spin_lock_irq(&sem->wait_lock);
 713         list_del(&waiter.list);
 714         if (list_empty(&sem->wait_list))
 715                 atomic_long_andnot(RWSEM_FLAG_WAITERS, &sem->count);
 716         else
 717                 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
 718         raw_spin_unlock_irq(&sem->wait_lock);
 719         wake_up_q(&wake_q);
 720         lockevent_inc(rwsem_wlock_fail);
 721
 722         return ERR_PTR(-EINTR);
 723 }
 724
 725 /*
 726  * handle waking up a waiter on the semaphore
 727  * - up_read/up_write has decremented the active part of count if we come here
 728  */
 729 static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
 730 {
 731         unsigned long flags;
 732         DEFINE_WAKE_Q(wake_q);
 733
 734         raw_spin_lock_irqsave(&sem->wait_lock, flags);
 735
 736         if (!list_empty(&sem->wait_list))
 737                 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
 738
 739         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
 740         wake_up_q(&wake_q);
 741
 742         return sem;
 743 }
 744
 745 /*
 746  * downgrade a write lock into a read lock
 747  * - caller incremented waiting part of count and discovered it still negative
 748  * - just wake up any readers at the front of the queue
 749  */
 750 static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
 751 {
 752         unsigned long flags;
 753         DEFINE_WAKE_Q(wake_q);
 754
 755         raw_spin_lock_irqsave(&sem->wait_lock, flags);
 756
 757         if (!list_empty(&sem->wait_list))
 758                 rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
 759
 760         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
 761         wake_up_q(&wake_q);
 762
 763         return sem;
 764 }
 765
 766 /*
 767  * lock for reading
 768  */
 769 inline void __down_read(struct rw_semaphore *sem)
 770 {
 771         if (unlikely(atomic_long_fetch_add_acquire(RWSEM_READER_BIAS,
 772                         &sem->count) & RWSEM_READ_FAILED_MASK)) {
 773                 rwsem_down_read_slowpath(sem, TASK_UNINTERRUPTIBLE);
 774                 DEBUG_RWSEMS_WARN_ON(!((unsigned long)sem->owner &
 775                                         RWSEM_READER_OWNED), sem);
 776         } else {
 777                 rwsem_set_reader_owned(sem);
 778         }
 779 }
 780
 781 static inline int __down_read_killable(struct rw_semaphore *sem)
 782 {
 783         if (unlikely(atomic_long_fetch_add_acquire(RWSEM_READER_BIAS,
 784                         &sem->count) & RWSEM_READ_FAILED_MASK)) {
 785                 if (IS_ERR(rwsem_down_read_slowpath(sem, TASK_KILLABLE)))
 786                         return -EINTR;
 787                 DEBUG_RWSEMS_WARN_ON(!((unsigned long)sem->owner &
 788                                         RWSEM_READER_OWNED), sem);
 789         } else {
 790                 rwsem_set_reader_owned(sem);
 791         }
 792         return 0;
 793 }
 794
 795 static inline int __down_read_trylock(struct rw_semaphore *sem)
 796 {
 797         /*
 798          * Optimize for the case when the rwsem is not locked at all.
 799          */
 800         long tmp = RWSEM_UNLOCKED_VALUE;
 801
 802         do {
 803                 if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
 804                                         tmp + RWSEM_READER_BIAS)) {
 805                         rwsem_set_reader_owned(sem);
 806                         return 1;
 807                 }
 808         } while (!(tmp & RWSEM_READ_FAILED_MASK));
 809         return 0;
 810 }
 811
 812 /*
 813  * lock for writing
 814  */
 815 static inline void __down_write(struct rw_semaphore *sem)
 816 {
 817         long tmp = RWSEM_UNLOCKED_VALUE;
 818
 819         if (unlikely(!atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
 820                                                       RWSEM_WRITER_LOCKED)))
 821                 rwsem_down_write_slowpath(sem, TASK_UNINTERRUPTIBLE);
 822         rwsem_set_owner(sem);
 823 }
 824
 825 static inline int __down_write_killable(struct rw_semaphore *sem)
 826 {
 827         long tmp = RWSEM_UNLOCKED_VALUE;
 828
 829         if (unlikely(!atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
 830                                                       RWSEM_WRITER_LOCKED))) {
 831                 if (IS_ERR(rwsem_down_write_slowpath(sem, TASK_KILLABLE)))
 832                         return -EINTR;
 833         }
 834         rwsem_set_owner(sem);
 835         return 0;
 836 }
 837
 838 static inline int __down_write_trylock(struct rw_semaphore *sem)
 839 {
 840         long tmp = RWSEM_UNLOCKED_VALUE;
 841
 842         if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
 843                                             RWSEM_WRITER_LOCKED)) {
 844                 rwsem_set_owner(sem);
 845                 return true;
 846         }
 847         return false;
 848 }
 849
 850 /*
 851  * unlock after reading
 852  */
 853 inline void __up_read(struct rw_semaphore *sem)
 854 {
 855         long tmp;
 856
 857         DEBUG_RWSEMS_WARN_ON(!((unsigned long)sem->owner & RWSEM_READER_OWNED), sem);
 858         rwsem_clear_reader_owned(sem);
 859         tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
 860         if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
 861                       RWSEM_FLAG_WAITERS))
 862                 rwsem_wake(sem);
 863 }
 864
 865 /*
 866  * unlock after writing
 867  */
 868 static inline void __up_write(struct rw_semaphore *sem)
 869 {
 870         long tmp;
 871
 872         DEBUG_RWSEMS_WARN_ON(sem->owner != current, sem);
 873         rwsem_clear_owner(sem);
 874         tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
 875         if (unlikely(tmp & RWSEM_FLAG_WAITERS))
 876                 rwsem_wake(sem);
 877 }
 878
 879 /*
 880  * downgrade write lock to read lock
 881  */
 882 static inline void __downgrade_write(struct rw_semaphore *sem)
 883 {
 884         long tmp;
 885
 886         /*
 887          * When downgrading from exclusive to shared ownership,
 888          * anything inside the write-locked region cannot leak
 889          * into the read side. In contrast, anything in the
 890          * read-locked region is ok to be re-ordered into the
 891          * write side. As such, rely on RELEASE semantics.
 892          */
 893         DEBUG_RWSEMS_WARN_ON(sem->owner != current, sem);
 894         tmp = atomic_long_fetch_add_release(
 895                 -RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
 896         rwsem_set_reader_owned(sem);
 897         if (tmp & RWSEM_FLAG_WAITERS)
 898                 rwsem_downgrade_wake(sem);
 899 }
 900
 901 /*
 902  * lock for reading
 903  */
 904 void __sched down_read(struct rw_semaphore *sem)
 905 {
 906         might_sleep();
 907         rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
 908
 909         LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
 910 }
 911 EXPORT_SYMBOL(down_read);
 912
 913 int __sched down_read_killable(struct rw_semaphore *sem)
 914 {
 915         might_sleep();
 916         rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
 917
 918         if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
 919                 rwsem_release(&sem->dep_map, 1, _RET_IP_);
 920                 return -EINTR;
 921         }
 922
 923         return 0;
 924 }
 925 EXPORT_SYMBOL(down_read_killable);
 926
 927 /*
 928  * trylock for reading -- returns 1 if successful, 0 if contention
 929  */
 930 int down_read_trylock(struct rw_semaphore *sem)
 931 {
 932         int ret = __down_read_trylock(sem);
 933
 934         if (ret == 1)
 935                 rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
 936         return ret;
 937 }
 938 EXPORT_SYMBOL(down_read_trylock);
 939
 940 /*
 941  * lock for writing
 942  */
 943 void __sched down_write(struct rw_semaphore *sem)
 944 {
 945         might_sleep();
 946         rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
 947         LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
 948 }
 949 EXPORT_SYMBOL(down_write);
 950
 951 /*
 952  * lock for writing
 953  */
 954 int __sched down_write_killable(struct rw_semaphore *sem)
 955 {
 956         might_sleep();
 957         rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
 958
 959         if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
 960                                   __down_write_killable)) {
 961                 rwsem_release(&sem->dep_map, 1, _RET_IP_);
 962                 return -EINTR;
 963         }
 964
 965         return 0;
 966 }
 967 EXPORT_SYMBOL(down_write_killable);
 968
 969 /*
 970  * trylock for writing -- returns 1 if successful, 0 if contention
 971  */
 972 int down_write_trylock(struct rw_semaphore *sem)
 973 {
 974         int ret = __down_write_trylock(sem);
 975
 976         if (ret == 1)
 977                 rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
 978
 979         return ret;
 980 }
 981 EXPORT_SYMBOL(down_write_trylock);
 982
 983 /*
 984  * release a read lock
 985  */
 986 void up_read(struct rw_semaphore *sem)
 987 {
 988         rwsem_release(&sem->dep_map, 1, _RET_IP_);
 989         __up_read(sem);
 990 }
 991 EXPORT_SYMBOL(up_read);
 992
 993 /*
 994  * release a write lock
 995  */
 996 void up_write(struct rw_semaphore *sem)
 997 {
 998         rwsem_release(&sem->dep_map, 1, _RET_IP_);
 999         __up_write(sem);
1000 }
1001 EXPORT_SYMBOL(up_write);
1002
1003 /*
1004  * downgrade write lock to read lock
1005  */
1006 void downgrade_write(struct rw_semaphore *sem)
1007 {
1008         lock_downgrade(&sem->dep_map, _RET_IP_);
1009         __downgrade_write(sem);
1010 }
1011 EXPORT_SYMBOL(downgrade_write);
1012
1013 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1014
1015 void down_read_nested(struct rw_semaphore *sem, int subclass)
1016 {
1017         might_sleep();
1018         rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1019         LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1020 }
1021 EXPORT_SYMBOL(down_read_nested);
1022
1023 void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1024 {
1025         might_sleep();
1026         rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1027         LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1028 }
1029 EXPORT_SYMBOL(_down_write_nest_lock);
1030
1031 void down_read_non_owner(struct rw_semaphore *sem)
1032 {
1033         might_sleep();
1034         __down_read(sem);
1035         __rwsem_set_reader_owned(sem, NULL);
1036 }
1037 EXPORT_SYMBOL(down_read_non_owner);
1038
1039 void down_write_nested(struct rw_semaphore *sem, int subclass)
1040 {
1041         might_sleep();
1042         rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1043         LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1044 }
1045 EXPORT_SYMBOL(down_write_nested);
1046
1047 int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1048 {
1049         might_sleep();
1050         rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1051
1052         if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1053                                   __down_write_killable)) {
1054                 rwsem_release(&sem->dep_map, 1, _RET_IP_);
1055                 return -EINTR;
1056         }
1057
1058         return 0;
1059 }
1060 EXPORT_SYMBOL(down_write_killable_nested);
1061
1062 void up_read_non_owner(struct rw_semaphore *sem)
1063 {
1064         DEBUG_RWSEMS_WARN_ON(!((unsigned long)sem->owner & RWSEM_READER_OWNED),
1065                                 sem);
1066         __up_read(sem);
1067 }
1068 EXPORT_SYMBOL(up_read_non_owner);
1069
1070 #endif