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 by
  14 * Waiman Long <longman@redhat.com> and
  15 * Peter Zijlstra <peterz@infradead.org>.
  16 */
  17
  18#include <linux/types.h>
  19#include <linux/kernel.h>
  20#include <linux/sched.h>
  21#include <linux/sched/rt.h>
  22#include <linux/sched/task.h>
  23#include <linux/sched/debug.h>
  24#include <linux/sched/wake_q.h>
  25#include <linux/sched/signal.h>
  26#include <linux/sched/clock.h>
  27#include <linux/export.h>
  28#include <linux/rwsem.h>
  29#include <linux/atomic.h>
  30#include <trace/events/lock.h>
  31
  32#ifndef CONFIG_PREEMPT_RT
  33#include "lock_events.h"
  34
  35/*
  36 * The least significant 2 bits of the owner value has the following
  37 * meanings when set.
  38 *  - Bit 0: RWSEM_READER_OWNED - rwsem may be owned by readers (just a hint)
  39 *  - Bit 1: RWSEM_NONSPINNABLE - Cannot spin on a reader-owned lock
  40 *
  41 * When the rwsem is reader-owned and a spinning writer has timed out,
  42 * the nonspinnable bit will be set to disable optimistic spinning.
  43
  44 * When a writer acquires a rwsem, it puts its task_struct pointer
  45 * into the owner field. It is cleared after an unlock.
  46 *
  47 * When a reader acquires a rwsem, it will also puts its task_struct
  48 * pointer into the owner field with the RWSEM_READER_OWNED bit set.
  49 * On unlock, the owner field will largely be left untouched. So
  50 * for a free or reader-owned rwsem, the owner value may contain
  51 * information about the last reader that acquires the rwsem.
  52 *
  53 * That information may be helpful in debugging cases where the system
  54 * seems to hang on a reader owned rwsem especially if only one reader
  55 * is involved. Ideally we would like to track all the readers that own
  56 * a rwsem, but the overhead is simply too big.
  57 *
  58 * A fast path reader optimistic lock stealing is supported when the rwsem
  59 * is previously owned by a writer and the following conditions are met:
  60 *  - rwsem is not currently writer owned
  61 *  - the handoff isn't set.
  62 */
  63#define RWSEM_READER_OWNED	(1UL << 0)
  64#define RWSEM_NONSPINNABLE	(1UL << 1)
  65#define RWSEM_OWNER_FLAGS_MASK	(RWSEM_READER_OWNED | RWSEM_NONSPINNABLE)
  66
  67#ifdef CONFIG_DEBUG_RWSEMS
  68# define DEBUG_RWSEMS_WARN_ON(c, sem)	do {			\
  69	if (!debug_locks_silent &&				\
  70	    WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, magic = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
  71		#c, atomic_long_read(&(sem)->count),		\
  72		(unsigned long) sem->magic,			\
  73		atomic_long_read(&(sem)->owner), (long)current,	\
  74		list_empty(&(sem)->wait_list) ? "" : "not "))	\
  75			debug_locks_off();			\
  76	} while (0)
  77#else
  78# define DEBUG_RWSEMS_WARN_ON(c, sem)
  79#endif
  80
  81/*
  82 * On 64-bit architectures, the bit definitions of the count are:
  83 *
  84 * Bit  0    - writer locked bit
  85 * Bit  1    - waiters present bit
  86 * Bit  2    - lock handoff bit
  87 * Bits 3-7  - reserved
  88 * Bits 8-62 - 55-bit reader count
  89 * Bit  63   - read fail bit
  90 *
  91 * On 32-bit architectures, the bit definitions of the count are:
  92 *
  93 * Bit  0    - writer locked bit
  94 * Bit  1    - waiters present bit
  95 * Bit  2    - lock handoff bit
  96 * Bits 3-7  - reserved
  97 * Bits 8-30 - 23-bit reader count
  98 * Bit  31   - read fail bit
  99 *
 100 * It is not likely that the most significant bit (read fail bit) will ever
 101 * be set. This guard bit is still checked anyway in the down_read() fastpath
 102 * just in case we need to use up more of the reader bits for other purpose
 103 * in the future.
 104 *
 105 * atomic_long_fetch_add() is used to obtain reader lock, whereas
 106 * atomic_long_cmpxchg() will be used to obtain writer lock.
 107 *
 108 * There are three places where the lock handoff bit may be set or cleared.
 109 * 1) rwsem_mark_wake() for readers		-- set, clear
 110 * 2) rwsem_try_write_lock() for writers	-- set, clear
 111 * 3) rwsem_del_waiter()			-- clear
 112 *
 113 * For all the above cases, wait_lock will be held. A writer must also
 114 * be the first one in the wait_list to be eligible for setting the handoff
 115 * bit. So concurrent setting/clearing of handoff bit is not possible.
 116 */
 117#define RWSEM_WRITER_LOCKED	(1UL << 0)
 118#define RWSEM_FLAG_WAITERS	(1UL << 1)
 119#define RWSEM_FLAG_HANDOFF	(1UL << 2)
 120#define RWSEM_FLAG_READFAIL	(1UL << (BITS_PER_LONG - 1))
 121
 122#define RWSEM_READER_SHIFT	8
 123#define RWSEM_READER_BIAS	(1UL << RWSEM_READER_SHIFT)
 124#define RWSEM_READER_MASK	(~(RWSEM_READER_BIAS - 1))
 125#define RWSEM_WRITER_MASK	RWSEM_WRITER_LOCKED
 126#define RWSEM_LOCK_MASK		(RWSEM_WRITER_MASK|RWSEM_READER_MASK)
 127#define RWSEM_READ_FAILED_MASK	(RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
 128				 RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
 129
 130/*
 131 * All writes to owner are protected by WRITE_ONCE() to make sure that
 132 * store tearing can't happen as optimistic spinners may read and use
 133 * the owner value concurrently without lock. Read from owner, however,
 134 * may not need READ_ONCE() as long as the pointer value is only used
 135 * for comparison and isn't being dereferenced.
 136 *
 137 * Both rwsem_{set,clear}_owner() functions should be in the same
 138 * preempt disable section as the atomic op that changes sem->count.
 139 */
 140static inline void rwsem_set_owner(struct rw_semaphore *sem)
 141{
 142	lockdep_assert_preemption_disabled();
 143	atomic_long_set(&sem->owner, (long)current);
 144}
 145
 146static inline void rwsem_clear_owner(struct rw_semaphore *sem)
 147{
 148	lockdep_assert_preemption_disabled();
 149	atomic_long_set(&sem->owner, 0);
 150}
 151
 152/*
 153 * Test the flags in the owner field.
 154 */
 155static inline bool rwsem_test_oflags(struct rw_semaphore *sem, long flags)
 156{
 157	return atomic_long_read(&sem->owner) & flags;
 158}
 159
 160/*
 161 * The task_struct pointer of the last owning reader will be left in
 162 * the owner field.
 163 *
 164 * Note that the owner value just indicates the task has owned the rwsem
 165 * previously, it may not be the real owner or one of the real owners
 166 * anymore when that field is examined, so take it with a grain of salt.
 167 *
 168 * The reader non-spinnable bit is preserved.
 169 */
 170static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
 171					    struct task_struct *owner)
 172{
 173	unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED |
 174		(atomic_long_read(&sem->owner) & RWSEM_NONSPINNABLE);
 175
 176	atomic_long_set(&sem->owner, val);
 177}
 178
 179static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
 180{
 181	__rwsem_set_reader_owned(sem, current);
 182}
 183
 184#ifdef CONFIG_DEBUG_RWSEMS
 185/*
 186 * Return just the real task structure pointer of the owner
 187 */
 188static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
 189{
 190	return (struct task_struct *)
 191		(atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
 192}
 193
 194/*
 195 * Return true if the rwsem is owned by a reader.
 196 */
 197static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
 198{
 
 199	/*
 200	 * Check the count to see if it is write-locked.
 201	 */
 202	long count = atomic_long_read(&sem->count);
 203
 204	if (count & RWSEM_WRITER_MASK)
 205		return false;
 
 206	return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
 207}
 208
 
 209/*
 210 * With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
 211 * is a task pointer in owner of a reader-owned rwsem, it will be the
 212 * real owner or one of the real owners. The only exception is when the
 213 * unlock is done by up_read_non_owner().
 214 */
 215static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
 216{
 217	unsigned long val = atomic_long_read(&sem->owner);
 218
 219	while ((val & ~RWSEM_OWNER_FLAGS_MASK) == (unsigned long)current) {
 220		if (atomic_long_try_cmpxchg(&sem->owner, &val,
 221					    val & RWSEM_OWNER_FLAGS_MASK))
 222			return;
 223	}
 224}
 225#else
 226static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
 227{
 228}
 229#endif
 230
 231/*
 232 * Set the RWSEM_NONSPINNABLE bits if the RWSEM_READER_OWNED flag
 233 * remains set. Otherwise, the operation will be aborted.
 234 */
 235static inline void rwsem_set_nonspinnable(struct rw_semaphore *sem)
 236{
 237	unsigned long owner = atomic_long_read(&sem->owner);
 238
 239	do {
 240		if (!(owner & RWSEM_READER_OWNED))
 241			break;
 242		if (owner & RWSEM_NONSPINNABLE)
 243			break;
 244	} while (!atomic_long_try_cmpxchg(&sem->owner, &owner,
 245					  owner | RWSEM_NONSPINNABLE));
 246}
 247
 248static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cntp)
 249{
 250	*cntp = atomic_long_add_return_acquire(RWSEM_READER_BIAS, &sem->count);
 251
 252	if (WARN_ON_ONCE(*cntp < 0))
 253		rwsem_set_nonspinnable(sem);
 254
 255	if (!(*cntp & RWSEM_READ_FAILED_MASK)) {
 256		rwsem_set_reader_owned(sem);
 257		return true;
 258	}
 259
 260	return false;
 261}
 262
 263static inline bool rwsem_write_trylock(struct rw_semaphore *sem)
 264{
 265	long tmp = RWSEM_UNLOCKED_VALUE;
 266
 267	if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp, RWSEM_WRITER_LOCKED)) {
 268		rwsem_set_owner(sem);
 269		return true;
 270	}
 271
 272	return false;
 273}
 274
 275/*
 
 
 
 
 
 
 
 
 
 276 * Return the real task structure pointer of the owner and the embedded
 277 * flags in the owner. pflags must be non-NULL.
 278 */
 279static inline struct task_struct *
 280rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
 281{
 282	unsigned long owner = atomic_long_read(&sem->owner);
 283
 284	*pflags = owner & RWSEM_OWNER_FLAGS_MASK;
 285	return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
 286}
 287
 288/*
 289 * Guide to the rw_semaphore's count field.
 290 *
 291 * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
 292 * by a writer.
 293 *
 294 * The lock is owned by readers when
 295 * (1) the RWSEM_WRITER_LOCKED isn't set in count,
 296 * (2) some of the reader bits are set in count, and
 297 * (3) the owner field has RWSEM_READ_OWNED bit set.
 298 *
 299 * Having some reader bits set is not enough to guarantee a readers owned
 300 * lock as the readers may be in the process of backing out from the count
 301 * and a writer has just released the lock. So another writer may steal
 302 * the lock immediately after that.
 303 */
 304
 305/*
 306 * Initialize an rwsem:
 307 */
 308void __init_rwsem(struct rw_semaphore *sem, const char *name,
 309		  struct lock_class_key *key)
 310{
 311#ifdef CONFIG_DEBUG_LOCK_ALLOC
 312	/*
 313	 * Make sure we are not reinitializing a held semaphore:
 314	 */
 315	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
 316	lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
 317#endif
 318#ifdef CONFIG_DEBUG_RWSEMS
 319	sem->magic = sem;
 320#endif
 321	atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
 322	raw_spin_lock_init(&sem->wait_lock);
 323	INIT_LIST_HEAD(&sem->wait_list);
 324	atomic_long_set(&sem->owner, 0L);
 325#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
 326	osq_lock_init(&sem->osq);
 327#endif
 328}
 329EXPORT_SYMBOL(__init_rwsem);
 330
 331enum rwsem_waiter_type {
 332	RWSEM_WAITING_FOR_WRITE,
 333	RWSEM_WAITING_FOR_READ
 334};
 335
 336struct rwsem_waiter {
 337	struct list_head list;
 338	struct task_struct *task;
 339	enum rwsem_waiter_type type;
 340	unsigned long timeout;
 341	bool handoff_set;
 342};
 343#define rwsem_first_waiter(sem) \
 344	list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
 345
 346enum rwsem_wake_type {
 347	RWSEM_WAKE_ANY,		/* Wake whatever's at head of wait list */
 348	RWSEM_WAKE_READERS,	/* Wake readers only */
 349	RWSEM_WAKE_READ_OWNED	/* Waker thread holds the read lock */
 350};
 351
 352/*
 353 * The typical HZ value is either 250 or 1000. So set the minimum waiting
 354 * time to at least 4ms or 1 jiffy (if it is higher than 4ms) in the wait
 355 * queue before initiating the handoff protocol.
 356 */
 357#define RWSEM_WAIT_TIMEOUT	DIV_ROUND_UP(HZ, 250)
 358
 359/*
 360 * Magic number to batch-wakeup waiting readers, even when writers are
 361 * also present in the queue. This both limits the amount of work the
 362 * waking thread must do and also prevents any potential counter overflow,
 363 * however unlikely.
 364 */
 365#define MAX_READERS_WAKEUP	0x100
 366
 367static inline void
 368rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
 369{
 370	lockdep_assert_held(&sem->wait_lock);
 371	list_add_tail(&waiter->list, &sem->wait_list);
 372	/* caller will set RWSEM_FLAG_WAITERS */
 373}
 374
 375/*
 376 * Remove a waiter from the wait_list and clear flags.
 377 *
 378 * Both rwsem_mark_wake() and rwsem_try_write_lock() contain a full 'copy' of
 379 * this function. Modify with care.
 380 *
 381 * Return: true if wait_list isn't empty and false otherwise
 382 */
 383static inline bool
 384rwsem_del_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
 385{
 386	lockdep_assert_held(&sem->wait_lock);
 387	list_del(&waiter->list);
 388	if (likely(!list_empty(&sem->wait_list)))
 389		return true;
 390
 391	atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS, &sem->count);
 392	return false;
 393}
 394
 395/*
 396 * handle the lock release when processes blocked on it that can now run
 397 * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
 398 *   have been set.
 399 * - there must be someone on the queue
 400 * - the wait_lock must be held by the caller
 401 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
 402 *   to actually wakeup the blocked task(s) and drop the reference count,
 403 *   preferably when the wait_lock is released
 404 * - woken process blocks are discarded from the list after having task zeroed
 405 * - writers are only marked woken if downgrading is false
 406 *
 407 * Implies rwsem_del_waiter() for all woken readers.
 408 */
 409static void rwsem_mark_wake(struct rw_semaphore *sem,
 410			    enum rwsem_wake_type wake_type,
 411			    struct wake_q_head *wake_q)
 412{
 413	struct rwsem_waiter *waiter, *tmp;
 414	long oldcount, woken = 0, adjustment = 0;
 415	struct list_head wlist;
 416
 417	lockdep_assert_held(&sem->wait_lock);
 418
 419	/*
 420	 * Take a peek at the queue head waiter such that we can determine
 421	 * the wakeup(s) to perform.
 422	 */
 423	waiter = rwsem_first_waiter(sem);
 424
 425	if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
 426		if (wake_type == RWSEM_WAKE_ANY) {
 427			/*
 428			 * Mark writer at the front of the queue for wakeup.
 429			 * Until the task is actually later awoken later by
 430			 * the caller, other writers are able to steal it.
 431			 * Readers, on the other hand, will block as they
 432			 * will notice the queued writer.
 433			 */
 434			wake_q_add(wake_q, waiter->task);
 435			lockevent_inc(rwsem_wake_writer);
 436		}
 437
 438		return;
 439	}
 440
 441	/*
 442	 * No reader wakeup if there are too many of them already.
 443	 */
 444	if (unlikely(atomic_long_read(&sem->count) < 0))
 445		return;
 446
 447	/*
 448	 * Writers might steal the lock before we grant it to the next reader.
 449	 * We prefer to do the first reader grant before counting readers
 450	 * so we can bail out early if a writer stole the lock.
 451	 */
 452	if (wake_type != RWSEM_WAKE_READ_OWNED) {
 453		struct task_struct *owner;
 454
 455		adjustment = RWSEM_READER_BIAS;
 456		oldcount = atomic_long_fetch_add(adjustment, &sem->count);
 457		if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
 458			/*
 459			 * When we've been waiting "too" long (for writers
 460			 * to give up the lock), request a HANDOFF to
 461			 * force the issue.
 462			 */
 463			if (time_after(jiffies, waiter->timeout)) {
 464				if (!(oldcount & RWSEM_FLAG_HANDOFF)) {
 465					adjustment -= RWSEM_FLAG_HANDOFF;
 466					lockevent_inc(rwsem_rlock_handoff);
 467				}
 468				waiter->handoff_set = true;
 469			}
 470
 471			atomic_long_add(-adjustment, &sem->count);
 472			return;
 473		}
 474		/*
 475		 * Set it to reader-owned to give spinners an early
 476		 * indication that readers now have the lock.
 477		 * The reader nonspinnable bit seen at slowpath entry of
 478		 * the reader is copied over.
 479		 */
 480		owner = waiter->task;
 481		__rwsem_set_reader_owned(sem, owner);
 482	}
 483
 484	/*
 485	 * Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
 486	 * queue. We know that the woken will be at least 1 as we accounted
 487	 * for above. Note we increment the 'active part' of the count by the
 488	 * number of readers before waking any processes up.
 489	 *
 490	 * This is an adaptation of the phase-fair R/W locks where at the
 491	 * reader phase (first waiter is a reader), all readers are eligible
 492	 * to acquire the lock at the same time irrespective of their order
 493	 * in the queue. The writers acquire the lock according to their
 494	 * order in the queue.
 495	 *
 496	 * We have to do wakeup in 2 passes to prevent the possibility that
 497	 * the reader count may be decremented before it is incremented. It
 498	 * is because the to-be-woken waiter may not have slept yet. So it
 499	 * may see waiter->task got cleared, finish its critical section and
 500	 * do an unlock before the reader count increment.
 501	 *
 502	 * 1) Collect the read-waiters in a separate list, count them and
 503	 *    fully increment the reader count in rwsem.
 504	 * 2) For each waiters in the new list, clear waiter->task and
 505	 *    put them into wake_q to be woken up later.
 506	 */
 507	INIT_LIST_HEAD(&wlist);
 508	list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
 509		if (waiter->type == RWSEM_WAITING_FOR_WRITE)
 510			continue;
 511
 512		woken++;
 513		list_move_tail(&waiter->list, &wlist);
 514
 515		/*
 516		 * Limit # of readers that can be woken up per wakeup call.
 517		 */
 518		if (unlikely(woken >= MAX_READERS_WAKEUP))
 519			break;
 520	}
 521
 522	adjustment = woken * RWSEM_READER_BIAS - adjustment;
 523	lockevent_cond_inc(rwsem_wake_reader, woken);
 524
 525	oldcount = atomic_long_read(&sem->count);
 526	if (list_empty(&sem->wait_list)) {
 527		/*
 528		 * Combined with list_move_tail() above, this implies
 529		 * rwsem_del_waiter().
 530		 */
 531		adjustment -= RWSEM_FLAG_WAITERS;
 532		if (oldcount & RWSEM_FLAG_HANDOFF)
 533			adjustment -= RWSEM_FLAG_HANDOFF;
 534	} else if (woken) {
 535		/*
 536		 * When we've woken a reader, we no longer need to force
 537		 * writers to give up the lock and we can clear HANDOFF.
 538		 */
 539		if (oldcount & RWSEM_FLAG_HANDOFF)
 540			adjustment -= RWSEM_FLAG_HANDOFF;
 541	}
 542
 543	if (adjustment)
 544		atomic_long_add(adjustment, &sem->count);
 545
 546	/* 2nd pass */
 547	list_for_each_entry_safe(waiter, tmp, &wlist, list) {
 548		struct task_struct *tsk;
 549
 550		tsk = waiter->task;
 551		get_task_struct(tsk);
 552
 553		/*
 554		 * Ensure calling get_task_struct() before setting the reader
 555		 * waiter to nil such that rwsem_down_read_slowpath() cannot
 556		 * race with do_exit() by always holding a reference count
 557		 * to the task to wakeup.
 558		 */
 559		smp_store_release(&waiter->task, NULL);
 560		/*
 561		 * Ensure issuing the wakeup (either by us or someone else)
 562		 * after setting the reader waiter to nil.
 563		 */
 564		wake_q_add_safe(wake_q, tsk);
 565	}
 566}
 567
 568/*
 569 * Remove a waiter and try to wake up other waiters in the wait queue
 570 * This function is called from the out_nolock path of both the reader and
 571 * writer slowpaths with wait_lock held. It releases the wait_lock and
 572 * optionally wake up waiters before it returns.
 573 */
 574static inline void
 575rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,
 576		      struct wake_q_head *wake_q)
 577		      __releases(&sem->wait_lock)
 578{
 579	bool first = rwsem_first_waiter(sem) == waiter;
 580
 581	wake_q_init(wake_q);
 582
 583	/*
 584	 * If the wait_list isn't empty and the waiter to be deleted is
 585	 * the first waiter, we wake up the remaining waiters as they may
 586	 * be eligible to acquire or spin on the lock.
 587	 */
 588	if (rwsem_del_waiter(sem, waiter) && first)
 589		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, wake_q);
 590	raw_spin_unlock_irq(&sem->wait_lock);
 591	if (!wake_q_empty(wake_q))
 592		wake_up_q(wake_q);
 593}
 594
 595/*
 596 * This function must be called with the sem->wait_lock held to prevent
 597 * race conditions between checking the rwsem wait list and setting the
 598 * sem->count accordingly.
 599 *
 600 * Implies rwsem_del_waiter() on success.
 601 */
 602static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
 603					struct rwsem_waiter *waiter)
 604{
 605	struct rwsem_waiter *first = rwsem_first_waiter(sem);
 606	long count, new;
 607
 608	lockdep_assert_held(&sem->wait_lock);
 609
 610	count = atomic_long_read(&sem->count);
 611	do {
 612		bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
 613
 614		if (has_handoff) {
 615			/*
 616			 * Honor handoff bit and yield only when the first
 617			 * waiter is the one that set it. Otherwisee, we
 618			 * still try to acquire the rwsem.
 619			 */
 620			if (first->handoff_set && (waiter != first))
 621				return false;
 622		}
 623
 624		new = count;
 625
 626		if (count & RWSEM_LOCK_MASK) {
 627			/*
 628			 * A waiter (first or not) can set the handoff bit
 629			 * if it is an RT task or wait in the wait queue
 630			 * for too long.
 631			 */
 632			if (has_handoff || (!rt_or_dl_task(waiter->task) &&
 633					    !time_after(jiffies, waiter->timeout)))
 634				return false;
 635
 636			new |= RWSEM_FLAG_HANDOFF;
 637		} else {
 638			new |= RWSEM_WRITER_LOCKED;
 639			new &= ~RWSEM_FLAG_HANDOFF;
 640
 641			if (list_is_singular(&sem->wait_list))
 642				new &= ~RWSEM_FLAG_WAITERS;
 643		}
 644	} while (!atomic_long_try_cmpxchg_acquire(&sem->count, &count, new));
 645
 646	/*
 647	 * We have either acquired the lock with handoff bit cleared or set
 648	 * the handoff bit. Only the first waiter can have its handoff_set
 649	 * set here to enable optimistic spinning in slowpath loop.
 650	 */
 651	if (new & RWSEM_FLAG_HANDOFF) {
 652		first->handoff_set = true;
 653		lockevent_inc(rwsem_wlock_handoff);
 654		return false;
 655	}
 656
 657	/*
 658	 * Have rwsem_try_write_lock() fully imply rwsem_del_waiter() on
 659	 * success.
 660	 */
 661	list_del(&waiter->list);
 662	rwsem_set_owner(sem);
 663	return true;
 664}
 665
 666/*
 667 * The rwsem_spin_on_owner() function returns the following 4 values
 668 * depending on the lock owner state.
 669 *   OWNER_NULL  : owner is currently NULL
 670 *   OWNER_WRITER: when owner changes and is a writer
 671 *   OWNER_READER: when owner changes and the new owner may be a reader.
 672 *   OWNER_NONSPINNABLE:
 673 *		   when optimistic spinning has to stop because either the
 674 *		   owner stops running, is unknown, or its timeslice has
 675 *		   been used up.
 676 */
 677enum owner_state {
 678	OWNER_NULL		= 1 << 0,
 679	OWNER_WRITER		= 1 << 1,
 680	OWNER_READER		= 1 << 2,
 681	OWNER_NONSPINNABLE	= 1 << 3,
 682};
 683
 684#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
 685/*
 686 * Try to acquire write lock before the writer has been put on wait queue.
 687 */
 688static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
 689{
 690	long count = atomic_long_read(&sem->count);
 691
 692	while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
 693		if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
 694					count | RWSEM_WRITER_LOCKED)) {
 695			rwsem_set_owner(sem);
 696			lockevent_inc(rwsem_opt_lock);
 697			return true;
 698		}
 699	}
 700	return false;
 701}
 702
 703static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
 704{
 705	struct task_struct *owner;
 706	unsigned long flags;
 707	bool ret = true;
 708
 709	if (need_resched()) {
 710		lockevent_inc(rwsem_opt_fail);
 711		return false;
 712	}
 713
 714	/*
 715	 * Disable preemption is equal to the RCU read-side crital section,
 716	 * thus the task_strcut structure won't go away.
 717	 */
 718	owner = rwsem_owner_flags(sem, &flags);
 719	/*
 720	 * Don't check the read-owner as the entry may be stale.
 721	 */
 722	if ((flags & RWSEM_NONSPINNABLE) ||
 723	    (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
 724		ret = false;
 725
 726	lockevent_cond_inc(rwsem_opt_fail, !ret);
 727	return ret;
 728}
 729
 730#define OWNER_SPINNABLE		(OWNER_NULL | OWNER_WRITER | OWNER_READER)
 731
 732static inline enum owner_state
 733rwsem_owner_state(struct task_struct *owner, unsigned long flags)
 734{
 735	if (flags & RWSEM_NONSPINNABLE)
 736		return OWNER_NONSPINNABLE;
 737
 738	if (flags & RWSEM_READER_OWNED)
 739		return OWNER_READER;
 740
 741	return owner ? OWNER_WRITER : OWNER_NULL;
 742}
 743
 744static noinline enum owner_state
 745rwsem_spin_on_owner(struct rw_semaphore *sem)
 746{
 747	struct task_struct *new, *owner;
 748	unsigned long flags, new_flags;
 749	enum owner_state state;
 750
 751	lockdep_assert_preemption_disabled();
 752
 753	owner = rwsem_owner_flags(sem, &flags);
 754	state = rwsem_owner_state(owner, flags);
 755	if (state != OWNER_WRITER)
 756		return state;
 757
 758	for (;;) {
 759		/*
 760		 * When a waiting writer set the handoff flag, it may spin
 761		 * on the owner as well. Once that writer acquires the lock,
 762		 * we can spin on it. So we don't need to quit even when the
 763		 * handoff bit is set.
 764		 */
 765		new = rwsem_owner_flags(sem, &new_flags);
 766		if ((new != owner) || (new_flags != flags)) {
 767			state = rwsem_owner_state(new, new_flags);
 768			break;
 769		}
 770
 771		/*
 772		 * Ensure we emit the owner->on_cpu, dereference _after_
 773		 * checking sem->owner still matches owner, if that fails,
 774		 * owner might point to free()d memory, if it still matches,
 775		 * our spinning context already disabled preemption which is
 776		 * equal to RCU read-side crital section ensures the memory
 777		 * stays valid.
 778		 */
 779		barrier();
 780
 781		if (need_resched() || !owner_on_cpu(owner)) {
 782			state = OWNER_NONSPINNABLE;
 783			break;
 784		}
 785
 786		cpu_relax();
 787	}
 788
 789	return state;
 790}
 791
 792/*
 793 * Calculate reader-owned rwsem spinning threshold for writer
 794 *
 795 * The more readers own the rwsem, the longer it will take for them to
 796 * wind down and free the rwsem. So the empirical formula used to
 797 * determine the actual spinning time limit here is:
 798 *
 799 *   Spinning threshold = (10 + nr_readers/2)us
 800 *
 801 * The limit is capped to a maximum of 25us (30 readers). This is just
 802 * a heuristic and is subjected to change in the future.
 803 */
 804static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
 805{
 806	long count = atomic_long_read(&sem->count);
 807	int readers = count >> RWSEM_READER_SHIFT;
 808	u64 delta;
 809
 810	if (readers > 30)
 811		readers = 30;
 812	delta = (20 + readers) * NSEC_PER_USEC / 2;
 813
 814	return sched_clock() + delta;
 815}
 816
 817static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
 818{
 819	bool taken = false;
 820	int prev_owner_state = OWNER_NULL;
 821	int loop = 0;
 822	u64 rspin_threshold = 0;
 823
 824	/* sem->wait_lock should not be held when doing optimistic spinning */
 825	if (!osq_lock(&sem->osq))
 826		goto done;
 827
 828	/*
 829	 * Optimistically spin on the owner field and attempt to acquire the
 830	 * lock whenever the owner changes. Spinning will be stopped when:
 831	 *  1) the owning writer isn't running; or
 832	 *  2) readers own the lock and spinning time has exceeded limit.
 833	 */
 834	for (;;) {
 835		enum owner_state owner_state;
 836
 837		owner_state = rwsem_spin_on_owner(sem);
 838		if (!(owner_state & OWNER_SPINNABLE))
 839			break;
 840
 841		/*
 842		 * Try to acquire the lock
 843		 */
 844		taken = rwsem_try_write_lock_unqueued(sem);
 845
 846		if (taken)
 847			break;
 848
 849		/*
 850		 * Time-based reader-owned rwsem optimistic spinning
 851		 */
 852		if (owner_state == OWNER_READER) {
 853			/*
 854			 * Re-initialize rspin_threshold every time when
 855			 * the owner state changes from non-reader to reader.
 856			 * This allows a writer to steal the lock in between
 857			 * 2 reader phases and have the threshold reset at
 858			 * the beginning of the 2nd reader phase.
 859			 */
 860			if (prev_owner_state != OWNER_READER) {
 861				if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
 862					break;
 863				rspin_threshold = rwsem_rspin_threshold(sem);
 864				loop = 0;
 865			}
 866
 867			/*
 868			 * Check time threshold once every 16 iterations to
 869			 * avoid calling sched_clock() too frequently so
 870			 * as to reduce the average latency between the times
 871			 * when the lock becomes free and when the spinner
 872			 * is ready to do a trylock.
 873			 */
 874			else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
 875				rwsem_set_nonspinnable(sem);
 876				lockevent_inc(rwsem_opt_nospin);
 877				break;
 878			}
 879		}
 880
 881		/*
 882		 * An RT task cannot do optimistic spinning if it cannot
 883		 * be sure the lock holder is running or live-lock may
 884		 * happen if the current task and the lock holder happen
 885		 * to run in the same CPU. However, aborting optimistic
 886		 * spinning while a NULL owner is detected may miss some
 887		 * opportunity where spinning can continue without causing
 888		 * problem.
 889		 *
 890		 * There are 2 possible cases where an RT task may be able
 891		 * to continue spinning.
 892		 *
 893		 * 1) The lock owner is in the process of releasing the
 894		 *    lock, sem->owner is cleared but the lock has not
 895		 *    been released yet.
 896		 * 2) The lock was free and owner cleared, but another
 897		 *    task just comes in and acquire the lock before
 898		 *    we try to get it. The new owner may be a spinnable
 899		 *    writer.
 900		 *
 901		 * To take advantage of two scenarios listed above, the RT
 902		 * task is made to retry one more time to see if it can
 903		 * acquire the lock or continue spinning on the new owning
 904		 * writer. Of course, if the time lag is long enough or the
 905		 * new owner is not a writer or spinnable, the RT task will
 906		 * quit spinning.
 907		 *
 908		 * If the owner is a writer, the need_resched() check is
 909		 * done inside rwsem_spin_on_owner(). If the owner is not
 910		 * a writer, need_resched() check needs to be done here.
 911		 */
 912		if (owner_state != OWNER_WRITER) {
 913			if (need_resched())
 914				break;
 915			if (rt_or_dl_task(current) &&
 916			   (prev_owner_state != OWNER_WRITER))
 917				break;
 918		}
 919		prev_owner_state = owner_state;
 920
 921		/*
 922		 * The cpu_relax() call is a compiler barrier which forces
 923		 * everything in this loop to be re-loaded. We don't need
 924		 * memory barriers as we'll eventually observe the right
 925		 * values at the cost of a few extra spins.
 926		 */
 927		cpu_relax();
 928	}
 929	osq_unlock(&sem->osq);
 930done:
 931	lockevent_cond_inc(rwsem_opt_fail, !taken);
 932	return taken;
 933}
 934
 935/*
 936 * Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
 937 * only be called when the reader count reaches 0.
 938 */
 939static inline void clear_nonspinnable(struct rw_semaphore *sem)
 940{
 941	if (unlikely(rwsem_test_oflags(sem, RWSEM_NONSPINNABLE)))
 942		atomic_long_andnot(RWSEM_NONSPINNABLE, &sem->owner);
 943}
 944
 945#else
 946static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
 947{
 948	return false;
 949}
 950
 951static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
 952{
 953	return false;
 954}
 955
 956static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
 957
 958static inline enum owner_state
 959rwsem_spin_on_owner(struct rw_semaphore *sem)
 960{
 961	return OWNER_NONSPINNABLE;
 962}
 963#endif
 964
 965/*
 966 * Prepare to wake up waiter(s) in the wait queue by putting them into the
 967 * given wake_q if the rwsem lock owner isn't a writer. If rwsem is likely
 968 * reader-owned, wake up read lock waiters in queue front or wake up any
 969 * front waiter otherwise.
 970
 971 * This is being called from both reader and writer slow paths.
 972 */
 973static inline void rwsem_cond_wake_waiter(struct rw_semaphore *sem, long count,
 974					  struct wake_q_head *wake_q)
 975{
 976	enum rwsem_wake_type wake_type;
 977
 978	if (count & RWSEM_WRITER_MASK)
 979		return;
 980
 981	if (count & RWSEM_READER_MASK) {
 982		wake_type = RWSEM_WAKE_READERS;
 983	} else {
 984		wake_type = RWSEM_WAKE_ANY;
 985		clear_nonspinnable(sem);
 986	}
 987	rwsem_mark_wake(sem, wake_type, wake_q);
 988}
 989
 990/*
 991 * Wait for the read lock to be granted
 992 */
 993static struct rw_semaphore __sched *
 994rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
 995{
 996	long adjustment = -RWSEM_READER_BIAS;
 997	long rcnt = (count >> RWSEM_READER_SHIFT);
 998	struct rwsem_waiter waiter;
 999	DEFINE_WAKE_Q(wake_q);
1000
1001	/*
1002	 * To prevent a constant stream of readers from starving a sleeping
1003	 * writer, don't attempt optimistic lock stealing if the lock is
1004	 * very likely owned by readers.
1005	 */
1006	if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
1007	    (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
1008		goto queue;
1009
1010	/*
1011	 * Reader optimistic lock stealing.
1012	 */
1013	if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
1014		rwsem_set_reader_owned(sem);
1015		lockevent_inc(rwsem_rlock_steal);
1016
1017		/*
1018		 * Wake up other readers in the wait queue if it is
1019		 * the first reader.
1020		 */
1021		if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
1022			raw_spin_lock_irq(&sem->wait_lock);
1023			if (!list_empty(&sem->wait_list))
1024				rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
1025						&wake_q);
1026			raw_spin_unlock_irq(&sem->wait_lock);
1027			wake_up_q(&wake_q);
1028		}
1029		return sem;
1030	}
1031
1032queue:
1033	waiter.task = current;
1034	waiter.type = RWSEM_WAITING_FOR_READ;
1035	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1036	waiter.handoff_set = false;
1037
1038	raw_spin_lock_irq(&sem->wait_lock);
1039	if (list_empty(&sem->wait_list)) {
1040		/*
1041		 * In case the wait queue is empty and the lock isn't owned
1042		 * by a writer, this reader can exit the slowpath and return
1043		 * immediately as its RWSEM_READER_BIAS has already been set
1044		 * in the count.
1045		 */
1046		if (!(atomic_long_read(&sem->count) & RWSEM_WRITER_MASK)) {
1047			/* Provide lock ACQUIRE */
1048			smp_acquire__after_ctrl_dep();
1049			raw_spin_unlock_irq(&sem->wait_lock);
1050			rwsem_set_reader_owned(sem);
1051			lockevent_inc(rwsem_rlock_fast);
1052			return sem;
1053		}
1054		adjustment += RWSEM_FLAG_WAITERS;
1055	}
1056	rwsem_add_waiter(sem, &waiter);
1057
1058	/* we're now waiting on the lock, but no longer actively locking */
1059	count = atomic_long_add_return(adjustment, &sem->count);
1060
1061	rwsem_cond_wake_waiter(sem, count, &wake_q);
1062	raw_spin_unlock_irq(&sem->wait_lock);
1063
1064	if (!wake_q_empty(&wake_q))
1065		wake_up_q(&wake_q);
1066
1067	trace_contention_begin(sem, LCB_F_READ);
1068
1069	/* wait to be given the lock */
1070	for (;;) {
1071		set_current_state(state);
1072		if (!smp_load_acquire(&waiter.task)) {
1073			/* Matches rwsem_mark_wake()'s smp_store_release(). */
1074			break;
1075		}
1076		if (signal_pending_state(state, current)) {
1077			raw_spin_lock_irq(&sem->wait_lock);
1078			if (waiter.task)
1079				goto out_nolock;
1080			raw_spin_unlock_irq(&sem->wait_lock);
1081			/* Ordered by sem->wait_lock against rwsem_mark_wake(). */
1082			break;
1083		}
1084		schedule_preempt_disabled();
1085		lockevent_inc(rwsem_sleep_reader);
1086	}
1087
1088	__set_current_state(TASK_RUNNING);
1089	lockevent_inc(rwsem_rlock);
1090	trace_contention_end(sem, 0);
1091	return sem;
1092
1093out_nolock:
1094	rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1095	__set_current_state(TASK_RUNNING);
1096	lockevent_inc(rwsem_rlock_fail);
1097	trace_contention_end(sem, -EINTR);
1098	return ERR_PTR(-EINTR);
1099}
1100
1101/*
1102 * Wait until we successfully acquire the write lock
1103 */
1104static struct rw_semaphore __sched *
1105rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1106{
1107	struct rwsem_waiter waiter;
1108	DEFINE_WAKE_Q(wake_q);
1109
1110	/* do optimistic spinning and steal lock if possible */
1111	if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
1112		/* rwsem_optimistic_spin() implies ACQUIRE on success */
1113		return sem;
1114	}
1115
1116	/*
1117	 * Optimistic spinning failed, proceed to the slowpath
1118	 * and block until we can acquire the sem.
1119	 */
1120	waiter.task = current;
1121	waiter.type = RWSEM_WAITING_FOR_WRITE;
1122	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1123	waiter.handoff_set = false;
1124
1125	raw_spin_lock_irq(&sem->wait_lock);
1126	rwsem_add_waiter(sem, &waiter);
1127
1128	/* we're now waiting on the lock */
1129	if (rwsem_first_waiter(sem) != &waiter) {
1130		rwsem_cond_wake_waiter(sem, atomic_long_read(&sem->count),
1131				       &wake_q);
1132		if (!wake_q_empty(&wake_q)) {
1133			/*
1134			 * We want to minimize wait_lock hold time especially
1135			 * when a large number of readers are to be woken up.
1136			 */
1137			raw_spin_unlock_irq(&sem->wait_lock);
1138			wake_up_q(&wake_q);
1139			raw_spin_lock_irq(&sem->wait_lock);
1140		}
1141	} else {
1142		atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
1143	}
1144
1145	/* wait until we successfully acquire the lock */
1146	set_current_state(state);
1147	trace_contention_begin(sem, LCB_F_WRITE);
1148
1149	for (;;) {
1150		if (rwsem_try_write_lock(sem, &waiter)) {
1151			/* rwsem_try_write_lock() implies ACQUIRE on success */
1152			break;
1153		}
1154
1155		raw_spin_unlock_irq(&sem->wait_lock);
1156
1157		if (signal_pending_state(state, current))
1158			goto out_nolock;
1159
1160		/*
1161		 * After setting the handoff bit and failing to acquire
1162		 * the lock, attempt to spin on owner to accelerate lock
1163		 * transfer. If the previous owner is a on-cpu writer and it
1164		 * has just released the lock, OWNER_NULL will be returned.
1165		 * In this case, we attempt to acquire the lock again
1166		 * without sleeping.
1167		 */
1168		if (waiter.handoff_set) {
1169			enum owner_state owner_state;
1170
1171			owner_state = rwsem_spin_on_owner(sem);
1172			if (owner_state == OWNER_NULL)
1173				goto trylock_again;
1174		}
1175
1176		schedule_preempt_disabled();
1177		lockevent_inc(rwsem_sleep_writer);
1178		set_current_state(state);
1179trylock_again:
1180		raw_spin_lock_irq(&sem->wait_lock);
1181	}
1182	__set_current_state(TASK_RUNNING);
1183	raw_spin_unlock_irq(&sem->wait_lock);
1184	lockevent_inc(rwsem_wlock);
1185	trace_contention_end(sem, 0);
1186	return sem;
1187
1188out_nolock:
1189	__set_current_state(TASK_RUNNING);
1190	raw_spin_lock_irq(&sem->wait_lock);
1191	rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1192	lockevent_inc(rwsem_wlock_fail);
1193	trace_contention_end(sem, -EINTR);
1194	return ERR_PTR(-EINTR);
1195}
1196
1197/*
1198 * handle waking up a waiter on the semaphore
1199 * - up_read/up_write has decremented the active part of count if we come here
1200 */
1201static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
1202{
1203	unsigned long flags;
1204	DEFINE_WAKE_Q(wake_q);
1205
1206	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1207
1208	if (!list_empty(&sem->wait_list))
1209		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1210
1211	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1212	wake_up_q(&wake_q);
1213
1214	return sem;
1215}
1216
1217/*
1218 * downgrade a write lock into a read lock
1219 * - caller incremented waiting part of count and discovered it still negative
1220 * - just wake up any readers at the front of the queue
1221 */
1222static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1223{
1224	unsigned long flags;
1225	DEFINE_WAKE_Q(wake_q);
1226
1227	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1228
1229	if (!list_empty(&sem->wait_list))
1230		rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
1231
1232	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1233	wake_up_q(&wake_q);
1234
1235	return sem;
1236}
1237
1238/*
1239 * lock for reading
1240 */
1241static __always_inline int __down_read_common(struct rw_semaphore *sem, int state)
1242{
1243	int ret = 0;
1244	long count;
1245
1246	preempt_disable();
1247	if (!rwsem_read_trylock(sem, &count)) {
1248		if (IS_ERR(rwsem_down_read_slowpath(sem, count, state))) {
1249			ret = -EINTR;
1250			goto out;
1251		}
1252		DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1253	}
1254out:
1255	preempt_enable();
1256	return ret;
1257}
1258
1259static __always_inline void __down_read(struct rw_semaphore *sem)
1260{
1261	__down_read_common(sem, TASK_UNINTERRUPTIBLE);
1262}
1263
1264static __always_inline int __down_read_interruptible(struct rw_semaphore *sem)
1265{
1266	return __down_read_common(sem, TASK_INTERRUPTIBLE);
1267}
1268
1269static __always_inline int __down_read_killable(struct rw_semaphore *sem)
1270{
1271	return __down_read_common(sem, TASK_KILLABLE);
1272}
1273
1274static inline int __down_read_trylock(struct rw_semaphore *sem)
1275{
1276	int ret = 0;
1277	long tmp;
1278
1279	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1280
1281	preempt_disable();
1282	tmp = atomic_long_read(&sem->count);
1283	while (!(tmp & RWSEM_READ_FAILED_MASK)) {
1284		if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1285						    tmp + RWSEM_READER_BIAS)) {
1286			rwsem_set_reader_owned(sem);
1287			ret = 1;
1288			break;
1289		}
1290	}
1291	preempt_enable();
1292	return ret;
1293}
1294
1295/*
1296 * lock for writing
1297 */
1298static __always_inline int __down_write_common(struct rw_semaphore *sem, int state)
1299{
1300	int ret = 0;
1301
1302	preempt_disable();
1303	if (unlikely(!rwsem_write_trylock(sem))) {
1304		if (IS_ERR(rwsem_down_write_slowpath(sem, state)))
1305			ret = -EINTR;
1306	}
1307	preempt_enable();
1308	return ret;
1309}
1310
1311static __always_inline void __down_write(struct rw_semaphore *sem)
1312{
1313	__down_write_common(sem, TASK_UNINTERRUPTIBLE);
1314}
1315
1316static __always_inline int __down_write_killable(struct rw_semaphore *sem)
1317{
1318	return __down_write_common(sem, TASK_KILLABLE);
1319}
1320
1321static inline int __down_write_trylock(struct rw_semaphore *sem)
1322{
1323	int ret;
1324
1325	preempt_disable();
1326	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1327	ret = rwsem_write_trylock(sem);
1328	preempt_enable();
1329
1330	return ret;
1331}
1332
1333/*
1334 * unlock after reading
1335 */
1336static inline void __up_read(struct rw_semaphore *sem)
1337{
1338	long tmp;
1339
1340	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1341	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1342
1343	preempt_disable();
1344	rwsem_clear_reader_owned(sem);
1345	tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
1346	DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1347	if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1348		      RWSEM_FLAG_WAITERS)) {
1349		clear_nonspinnable(sem);
1350		rwsem_wake(sem);
1351	}
1352	preempt_enable();
1353}
1354
1355/*
1356 * unlock after writing
1357 */
1358static inline void __up_write(struct rw_semaphore *sem)
1359{
1360	long tmp;
1361
1362	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1363	/*
1364	 * sem->owner may differ from current if the ownership is transferred
1365	 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1366	 */
1367	DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1368			    !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1369
1370	preempt_disable();
1371	rwsem_clear_owner(sem);
1372	tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
1373	if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1374		rwsem_wake(sem);
1375	preempt_enable();
1376}
1377
1378/*
1379 * downgrade write lock to read lock
1380 */
1381static inline void __downgrade_write(struct rw_semaphore *sem)
1382{
1383	long tmp;
1384
1385	/*
1386	 * When downgrading from exclusive to shared ownership,
1387	 * anything inside the write-locked region cannot leak
1388	 * into the read side. In contrast, anything in the
1389	 * read-locked region is ok to be re-ordered into the
1390	 * write side. As such, rely on RELEASE semantics.
1391	 */
1392	DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1393	preempt_disable();
1394	tmp = atomic_long_fetch_add_release(
1395		-RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
1396	rwsem_set_reader_owned(sem);
1397	if (tmp & RWSEM_FLAG_WAITERS)
1398		rwsem_downgrade_wake(sem);
1399	preempt_enable();
1400}
1401
1402#else /* !CONFIG_PREEMPT_RT */
1403
1404#define RT_MUTEX_BUILD_MUTEX
1405#include "rtmutex.c"
1406
1407#define rwbase_set_and_save_current_state(state)	\
1408	set_current_state(state)
1409
1410#define rwbase_restore_current_state()			\
1411	__set_current_state(TASK_RUNNING)
1412
1413#define rwbase_rtmutex_lock_state(rtm, state)		\
1414	__rt_mutex_lock(rtm, state)
1415
1416#define rwbase_rtmutex_slowlock_locked(rtm, state, wq)	\
1417	__rt_mutex_slowlock_locked(rtm, NULL, state, wq)
1418
1419#define rwbase_rtmutex_unlock(rtm)			\
1420	__rt_mutex_unlock(rtm)
1421
1422#define rwbase_rtmutex_trylock(rtm)			\
1423	__rt_mutex_trylock(rtm)
1424
1425#define rwbase_signal_pending_state(state, current)	\
1426	signal_pending_state(state, current)
1427
1428#define rwbase_pre_schedule()				\
1429	rt_mutex_pre_schedule()
1430
1431#define rwbase_schedule()				\
1432	rt_mutex_schedule()
1433
1434#define rwbase_post_schedule()				\
1435	rt_mutex_post_schedule()
1436
1437#include "rwbase_rt.c"
1438
1439void __init_rwsem(struct rw_semaphore *sem, const char *name,
1440		  struct lock_class_key *key)
1441{
1442	init_rwbase_rt(&(sem)->rwbase);
1443
1444#ifdef CONFIG_DEBUG_LOCK_ALLOC
1445	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
1446	lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
1447#endif
1448}
1449EXPORT_SYMBOL(__init_rwsem);
1450
1451static inline void __down_read(struct rw_semaphore *sem)
1452{
1453	rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1454}
1455
1456static inline int __down_read_interruptible(struct rw_semaphore *sem)
1457{
1458	return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
1459}
1460
1461static inline int __down_read_killable(struct rw_semaphore *sem)
1462{
1463	return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
1464}
1465
1466static inline int __down_read_trylock(struct rw_semaphore *sem)
1467{
1468	return rwbase_read_trylock(&sem->rwbase);
1469}
1470
1471static inline void __up_read(struct rw_semaphore *sem)
1472{
1473	rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
1474}
1475
1476static inline void __sched __down_write(struct rw_semaphore *sem)
1477{
1478	rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1479}
1480
1481static inline int __sched __down_write_killable(struct rw_semaphore *sem)
1482{
1483	return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
1484}
1485
1486static inline int __down_write_trylock(struct rw_semaphore *sem)
1487{
1488	return rwbase_write_trylock(&sem->rwbase);
1489}
1490
1491static inline void __up_write(struct rw_semaphore *sem)
1492{
1493	rwbase_write_unlock(&sem->rwbase);
1494}
1495
1496static inline void __downgrade_write(struct rw_semaphore *sem)
1497{
1498	rwbase_write_downgrade(&sem->rwbase);
1499}
1500
1501/* Debug stubs for the common API */
1502#define DEBUG_RWSEMS_WARN_ON(c, sem)
1503
1504static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
1505					    struct task_struct *owner)
1506{
1507}
1508
1509static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
1510{
1511	int count = atomic_read(&sem->rwbase.readers);
1512
1513	return count < 0 && count != READER_BIAS;
1514}
1515
1516#endif /* CONFIG_PREEMPT_RT */
1517
1518/*
1519 * lock for reading
1520 */
1521void __sched down_read(struct rw_semaphore *sem)
1522{
1523	might_sleep();
1524	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1525
1526	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1527}
1528EXPORT_SYMBOL(down_read);
1529
1530int __sched down_read_interruptible(struct rw_semaphore *sem)
1531{
1532	might_sleep();
1533	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1534
1535	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
1536		rwsem_release(&sem->dep_map, _RET_IP_);
1537		return -EINTR;
1538	}
1539
1540	return 0;
1541}
1542EXPORT_SYMBOL(down_read_interruptible);
1543
1544int __sched down_read_killable(struct rw_semaphore *sem)
1545{
1546	might_sleep();
1547	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1548
1549	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1550		rwsem_release(&sem->dep_map, _RET_IP_);
1551		return -EINTR;
1552	}
1553
1554	return 0;
1555}
1556EXPORT_SYMBOL(down_read_killable);
1557
1558/*
1559 * trylock for reading -- returns 1 if successful, 0 if contention
1560 */
1561int down_read_trylock(struct rw_semaphore *sem)
1562{
1563	int ret = __down_read_trylock(sem);
1564
1565	if (ret == 1)
1566		rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1567	return ret;
1568}
1569EXPORT_SYMBOL(down_read_trylock);
1570
1571/*
1572 * lock for writing
1573 */
1574void __sched down_write(struct rw_semaphore *sem)
1575{
1576	might_sleep();
1577	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1578	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1579}
1580EXPORT_SYMBOL(down_write);
1581
1582/*
1583 * lock for writing
1584 */
1585int __sched down_write_killable(struct rw_semaphore *sem)
1586{
1587	might_sleep();
1588	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1589
1590	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1591				  __down_write_killable)) {
1592		rwsem_release(&sem->dep_map, _RET_IP_);
1593		return -EINTR;
1594	}
1595
1596	return 0;
1597}
1598EXPORT_SYMBOL(down_write_killable);
1599
1600/*
1601 * trylock for writing -- returns 1 if successful, 0 if contention
1602 */
1603int down_write_trylock(struct rw_semaphore *sem)
1604{
1605	int ret = __down_write_trylock(sem);
1606
1607	if (ret == 1)
1608		rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1609
1610	return ret;
1611}
1612EXPORT_SYMBOL(down_write_trylock);
1613
1614/*
1615 * release a read lock
1616 */
1617void up_read(struct rw_semaphore *sem)
1618{
1619	rwsem_release(&sem->dep_map, _RET_IP_);
1620	__up_read(sem);
1621}
1622EXPORT_SYMBOL(up_read);
1623
1624/*
1625 * release a write lock
1626 */
1627void up_write(struct rw_semaphore *sem)
1628{
1629	rwsem_release(&sem->dep_map, _RET_IP_);
1630	__up_write(sem);
1631}
1632EXPORT_SYMBOL(up_write);
1633
1634/*
1635 * downgrade write lock to read lock
1636 */
1637void downgrade_write(struct rw_semaphore *sem)
1638{
1639	lock_downgrade(&sem->dep_map, _RET_IP_);
1640	__downgrade_write(sem);
1641}
1642EXPORT_SYMBOL(downgrade_write);
1643
1644#ifdef CONFIG_DEBUG_LOCK_ALLOC
1645
1646void down_read_nested(struct rw_semaphore *sem, int subclass)
1647{
1648	might_sleep();
1649	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1650	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1651}
1652EXPORT_SYMBOL(down_read_nested);
1653
1654int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
1655{
1656	might_sleep();
1657	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1658
1659	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1660		rwsem_release(&sem->dep_map, _RET_IP_);
1661		return -EINTR;
1662	}
1663
1664	return 0;
1665}
1666EXPORT_SYMBOL(down_read_killable_nested);
1667
1668void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1669{
1670	might_sleep();
1671	rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1672	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1673}
1674EXPORT_SYMBOL(_down_write_nest_lock);
1675
1676void down_read_non_owner(struct rw_semaphore *sem)
1677{
1678	might_sleep();
1679	__down_read(sem);
1680	/*
1681	 * The owner value for a reader-owned lock is mostly for debugging
1682	 * purpose only and is not critical to the correct functioning of
1683	 * rwsem. So it is perfectly fine to set it in a preempt-enabled
1684	 * context here.
1685	 */
1686	__rwsem_set_reader_owned(sem, NULL);
1687}
1688EXPORT_SYMBOL(down_read_non_owner);
1689
1690void down_write_nested(struct rw_semaphore *sem, int subclass)
1691{
1692	might_sleep();
1693	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1694	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1695}
1696EXPORT_SYMBOL(down_write_nested);
1697
1698int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1699{
1700	might_sleep();
1701	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1702
1703	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1704				  __down_write_killable)) {
1705		rwsem_release(&sem->dep_map, _RET_IP_);
1706		return -EINTR;
1707	}
1708
1709	return 0;
1710}
1711EXPORT_SYMBOL(down_write_killable_nested);
1712
1713void up_read_non_owner(struct rw_semaphore *sem)
1714{
1715	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1716	__up_read(sem);
1717}
1718EXPORT_SYMBOL(up_read_non_owner);
1719
1720#endif