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
  31#include "rwsem.h"
  32#include "lock_events.h"
  33
  34/*
  35 * The least significant 3 bits of the owner value has the following
  36 * meanings when set.
  37 *  - Bit 0: RWSEM_READER_OWNED - The rwsem is owned by readers
  38 *  - Bit 1: RWSEM_RD_NONSPINNABLE - Readers cannot spin on this lock.
  39 *  - Bit 2: RWSEM_WR_NONSPINNABLE - Writers cannot spin on this lock.
  40 *
  41 * When the rwsem is either owned by an anonymous writer, or it is
  42 * reader-owned, but a spinning writer has timed out, both nonspinnable
  43 * bits will be set to disable optimistic spinning by readers and writers.
  44 * In the later case, the last unlocking reader should then check the
  45 * writer nonspinnable bit and clear it only to give writers preference
  46 * to acquire the lock via optimistic spinning, but not readers. Similar
  47 * action is also done in the reader slowpath.
  48
  49 * When a writer acquires a rwsem, it puts its task_struct pointer
  50 * into the owner field. It is cleared after an unlock.
  51 *
  52 * When a reader acquires a rwsem, it will also puts its task_struct
  53 * pointer into the owner field with the RWSEM_READER_OWNED bit set.
  54 * On unlock, the owner field will largely be left untouched. So
  55 * for a free or reader-owned rwsem, the owner value may contain
  56 * information about the last reader that acquires the rwsem.
  57 *
  58 * That information may be helpful in debugging cases where the system
  59 * seems to hang on a reader owned rwsem especially if only one reader
  60 * is involved. Ideally we would like to track all the readers that own
  61 * a rwsem, but the overhead is simply too big.
  62 *
  63 * Reader optimistic spinning is helpful when the reader critical section
  64 * is short and there aren't that many readers around. It makes readers
  65 * relatively more preferred than writers. When a writer times out spinning
  66 * on a reader-owned lock and set the nospinnable bits, there are two main
  67 * reasons for that.
  68 *
  69 *  1) The reader critical section is long, perhaps the task sleeps after
  70 *     acquiring the read lock.
  71 *  2) There are just too many readers contending the lock causing it to
  72 *     take a while to service all of them.
  73 *
  74 * In the former case, long reader critical section will impede the progress
  75 * of writers which is usually more important for system performance. In
  76 * the later case, reader optimistic spinning tends to make the reader
  77 * groups that contain readers that acquire the lock together smaller
  78 * leading to more of them. That may hurt performance in some cases. In
  79 * other words, the setting of nonspinnable bits indicates that reader
  80 * optimistic spinning may not be helpful for those workloads that cause
  81 * it.
  82 *
  83 * Therefore, any writers that had observed the setting of the writer
  84 * nonspinnable bit for a given rwsem after they fail to acquire the lock
  85 * via optimistic spinning will set the reader nonspinnable bit once they
  86 * acquire the write lock. Similarly, readers that observe the setting
  87 * of reader nonspinnable bit at slowpath entry will set the reader
  88 * nonspinnable bits when they acquire the read lock via the wakeup path.
  89 *
  90 * Once the reader nonspinnable bit is on, it will only be reset when
  91 * a writer is able to acquire the rwsem in the fast path or somehow a
  92 * reader or writer in the slowpath doesn't observe the nonspinable bit.
  93 *
  94 * This is to discourage reader optmistic spinning on that particular
  95 * rwsem and make writers more preferred. This adaptive disabling of reader
  96 * optimistic spinning will alleviate the negative side effect of this
  97 * feature.
  98 */
  99#define RWSEM_READER_OWNED	(1UL << 0)
 100#define RWSEM_RD_NONSPINNABLE	(1UL << 1)
 101#define RWSEM_WR_NONSPINNABLE	(1UL << 2)
 102#define RWSEM_NONSPINNABLE	(RWSEM_RD_NONSPINNABLE | RWSEM_WR_NONSPINNABLE)
 103#define RWSEM_OWNER_FLAGS_MASK	(RWSEM_READER_OWNED | RWSEM_NONSPINNABLE)
 104
 105#ifdef CONFIG_DEBUG_RWSEMS
 106# define DEBUG_RWSEMS_WARN_ON(c, sem)	do {			\
 107	if (!debug_locks_silent &&				\
 108	    WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, magic = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
 109		#c, atomic_long_read(&(sem)->count),		\
 110		(unsigned long) sem->magic,			\
 111		atomic_long_read(&(sem)->owner), (long)current,	\
 112		list_empty(&(sem)->wait_list) ? "" : "not "))	\
 113			debug_locks_off();			\
 114	} while (0)
 115#else
 116# define DEBUG_RWSEMS_WARN_ON(c, sem)
 117#endif
 118
 119/*
 120 * On 64-bit architectures, the bit definitions of the count are:
 121 *
 122 * Bit  0    - writer locked bit
 123 * Bit  1    - waiters present bit
 124 * Bit  2    - lock handoff bit
 125 * Bits 3-7  - reserved
 126 * Bits 8-62 - 55-bit reader count
 127 * Bit  63   - read fail bit
 128 *
 129 * On 32-bit architectures, the bit definitions of the count are:
 130 *
 131 * Bit  0    - writer locked bit
 132 * Bit  1    - waiters present bit
 133 * Bit  2    - lock handoff bit
 134 * Bits 3-7  - reserved
 135 * Bits 8-30 - 23-bit reader count
 136 * Bit  31   - read fail bit
 137 *
 138 * It is not likely that the most significant bit (read fail bit) will ever
 139 * be set. This guard bit is still checked anyway in the down_read() fastpath
 140 * just in case we need to use up more of the reader bits for other purpose
 141 * in the future.
 142 *
 143 * atomic_long_fetch_add() is used to obtain reader lock, whereas
 144 * atomic_long_cmpxchg() will be used to obtain writer lock.
 145 *
 146 * There are three places where the lock handoff bit may be set or cleared.
 147 * 1) rwsem_mark_wake() for readers.
 148 * 2) rwsem_try_write_lock() for writers.
 149 * 3) Error path of rwsem_down_write_slowpath().
 150 *
 151 * For all the above cases, wait_lock will be held. A writer must also
 152 * be the first one in the wait_list to be eligible for setting the handoff
 153 * bit. So concurrent setting/clearing of handoff bit is not possible.
 154 */
 155#define RWSEM_WRITER_LOCKED	(1UL << 0)
 156#define RWSEM_FLAG_WAITERS	(1UL << 1)
 157#define RWSEM_FLAG_HANDOFF	(1UL << 2)
 158#define RWSEM_FLAG_READFAIL	(1UL << (BITS_PER_LONG - 1))
 159
 160#define RWSEM_READER_SHIFT	8
 161#define RWSEM_READER_BIAS	(1UL << RWSEM_READER_SHIFT)
 162#define RWSEM_READER_MASK	(~(RWSEM_READER_BIAS - 1))
 163#define RWSEM_WRITER_MASK	RWSEM_WRITER_LOCKED
 164#define RWSEM_LOCK_MASK		(RWSEM_WRITER_MASK|RWSEM_READER_MASK)
 165#define RWSEM_READ_FAILED_MASK	(RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
 166				 RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
 167
 168/*
 169 * All writes to owner are protected by WRITE_ONCE() to make sure that
 170 * store tearing can't happen as optimistic spinners may read and use
 171 * the owner value concurrently without lock. Read from owner, however,
 172 * may not need READ_ONCE() as long as the pointer value is only used
 173 * for comparison and isn't being dereferenced.
 174 */
 175static inline void rwsem_set_owner(struct rw_semaphore *sem)
 176{
 177	atomic_long_set(&sem->owner, (long)current);
 178}
 179
 180static inline void rwsem_clear_owner(struct rw_semaphore *sem)
 181{
 182	atomic_long_set(&sem->owner, 0);
 183}
 184
 185/*
 186 * Test the flags in the owner field.
 187 */
 188static inline bool rwsem_test_oflags(struct rw_semaphore *sem, long flags)
 189{
 190	return atomic_long_read(&sem->owner) & flags;
 191}
 192
 193/*
 194 * The task_struct pointer of the last owning reader will be left in
 195 * the owner field.
 196 *
 197 * Note that the owner value just indicates the task has owned the rwsem
 198 * previously, it may not be the real owner or one of the real owners
 199 * anymore when that field is examined, so take it with a grain of salt.
 200 *
 201 * The reader non-spinnable bit is preserved.
 202 */
 203static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
 204					    struct task_struct *owner)
 205{
 206	unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED |
 207		(atomic_long_read(&sem->owner) & RWSEM_RD_NONSPINNABLE);
 208
 209	atomic_long_set(&sem->owner, val);
 210}
 211
 212static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
 213{
 214	__rwsem_set_reader_owned(sem, current);
 215}
 216
 217/*
 218 * Return true if the rwsem is owned by a reader.
 219 */
 220static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
 221{
 222#ifdef CONFIG_DEBUG_RWSEMS
 223	/*
 224	 * Check the count to see if it is write-locked.
 225	 */
 226	long count = atomic_long_read(&sem->count);
 227
 228	if (count & RWSEM_WRITER_MASK)
 229		return false;
 230#endif
 231	return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
 232}
 233
 234#ifdef CONFIG_DEBUG_RWSEMS
 235/*
 236 * With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
 237 * is a task pointer in owner of a reader-owned rwsem, it will be the
 238 * real owner or one of the real owners. The only exception is when the
 239 * unlock is done by up_read_non_owner().
 240 */
 241static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
 242{
 243	unsigned long val = atomic_long_read(&sem->owner);
 244
 245	while ((val & ~RWSEM_OWNER_FLAGS_MASK) == (unsigned long)current) {
 246		if (atomic_long_try_cmpxchg(&sem->owner, &val,
 247					    val & RWSEM_OWNER_FLAGS_MASK))
 248			return;
 249	}
 250}
 251#else
 252static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
 253{
 254}
 255#endif
 256
 257/*
 258 * Set the RWSEM_NONSPINNABLE bits if the RWSEM_READER_OWNED flag
 259 * remains set. Otherwise, the operation will be aborted.
 260 */
 261static inline void rwsem_set_nonspinnable(struct rw_semaphore *sem)
 262{
 263	unsigned long owner = atomic_long_read(&sem->owner);
 264
 265	do {
 266		if (!(owner & RWSEM_READER_OWNED))
 267			break;
 268		if (owner & RWSEM_NONSPINNABLE)
 269			break;
 270	} while (!atomic_long_try_cmpxchg(&sem->owner, &owner,
 271					  owner | RWSEM_NONSPINNABLE));
 272}
 273
 274static inline bool rwsem_read_trylock(struct rw_semaphore *sem)
 275{
 276	long cnt = atomic_long_add_return_acquire(RWSEM_READER_BIAS, &sem->count);
 277	if (WARN_ON_ONCE(cnt < 0))
 278		rwsem_set_nonspinnable(sem);
 279	return !(cnt & RWSEM_READ_FAILED_MASK);
 280}
 281
 282/*
 283 * Return just the real task structure pointer of the owner
 284 */
 285static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
 286{
 287	return (struct task_struct *)
 288		(atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
 289}
 290
 291/*
 292 * Return the real task structure pointer of the owner and the embedded
 293 * flags in the owner. pflags must be non-NULL.
 294 */
 295static inline struct task_struct *
 296rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
 297{
 298	unsigned long owner = atomic_long_read(&sem->owner);
 299
 300	*pflags = owner & RWSEM_OWNER_FLAGS_MASK;
 301	return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
 302}
 303
 304/*
 305 * Guide to the rw_semaphore's count field.
 306 *
 307 * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
 308 * by a writer.
 309 *
 310 * The lock is owned by readers when
 311 * (1) the RWSEM_WRITER_LOCKED isn't set in count,
 312 * (2) some of the reader bits are set in count, and
 313 * (3) the owner field has RWSEM_READ_OWNED bit set.
 314 *
 315 * Having some reader bits set is not enough to guarantee a readers owned
 316 * lock as the readers may be in the process of backing out from the count
 317 * and a writer has just released the lock. So another writer may steal
 318 * the lock immediately after that.
 319 */
 320
 321/*
 322 * Initialize an rwsem:
 323 */
 324void __init_rwsem(struct rw_semaphore *sem, const char *name,
 325		  struct lock_class_key *key)
 326{
 327#ifdef CONFIG_DEBUG_LOCK_ALLOC
 328	/*
 329	 * Make sure we are not reinitializing a held semaphore:
 330	 */
 331	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
 332	lockdep_init_map(&sem->dep_map, name, key, 0);
 333#endif
 334#ifdef CONFIG_DEBUG_RWSEMS
 335	sem->magic = sem;
 336#endif
 337	atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
 338	raw_spin_lock_init(&sem->wait_lock);
 339	INIT_LIST_HEAD(&sem->wait_list);
 340	atomic_long_set(&sem->owner, 0L);
 341#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
 342	osq_lock_init(&sem->osq);
 343#endif
 344}
 345EXPORT_SYMBOL(__init_rwsem);
 346
 347enum rwsem_waiter_type {
 348	RWSEM_WAITING_FOR_WRITE,
 349	RWSEM_WAITING_FOR_READ
 350};
 351
 352struct rwsem_waiter {
 353	struct list_head list;
 354	struct task_struct *task;
 355	enum rwsem_waiter_type type;
 356	unsigned long timeout;
 357	unsigned long last_rowner;
 358};
 359#define rwsem_first_waiter(sem) \
 360	list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
 361
 362enum rwsem_wake_type {
 363	RWSEM_WAKE_ANY,		/* Wake whatever's at head of wait list */
 364	RWSEM_WAKE_READERS,	/* Wake readers only */
 365	RWSEM_WAKE_READ_OWNED	/* Waker thread holds the read lock */
 366};
 367
 368enum writer_wait_state {
 369	WRITER_NOT_FIRST,	/* Writer is not first in wait list */
 370	WRITER_FIRST,		/* Writer is first in wait list     */
 371	WRITER_HANDOFF		/* Writer is first & handoff needed */
 372};
 373
 374/*
 375 * The typical HZ value is either 250 or 1000. So set the minimum waiting
 376 * time to at least 4ms or 1 jiffy (if it is higher than 4ms) in the wait
 377 * queue before initiating the handoff protocol.
 378 */
 379#define RWSEM_WAIT_TIMEOUT	DIV_ROUND_UP(HZ, 250)
 380
 381/*
 382 * Magic number to batch-wakeup waiting readers, even when writers are
 383 * also present in the queue. This both limits the amount of work the
 384 * waking thread must do and also prevents any potential counter overflow,
 385 * however unlikely.
 386 */
 387#define MAX_READERS_WAKEUP	0x100
 388
 389/*
 390 * handle the lock release when processes blocked on it that can now run
 391 * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
 392 *   have been set.
 393 * - there must be someone on the queue
 394 * - the wait_lock must be held by the caller
 395 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
 396 *   to actually wakeup the blocked task(s) and drop the reference count,
 397 *   preferably when the wait_lock is released
 398 * - woken process blocks are discarded from the list after having task zeroed
 399 * - writers are only marked woken if downgrading is false
 400 */
 401static void rwsem_mark_wake(struct rw_semaphore *sem,
 402			    enum rwsem_wake_type wake_type,
 403			    struct wake_q_head *wake_q)
 404{
 405	struct rwsem_waiter *waiter, *tmp;
 406	long oldcount, woken = 0, adjustment = 0;
 407	struct list_head wlist;
 408
 409	lockdep_assert_held(&sem->wait_lock);
 410
 411	/*
 412	 * Take a peek at the queue head waiter such that we can determine
 413	 * the wakeup(s) to perform.
 414	 */
 415	waiter = rwsem_first_waiter(sem);
 416
 417	if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
 418		if (wake_type == RWSEM_WAKE_ANY) {
 419			/*
 420			 * Mark writer at the front of the queue for wakeup.
 421			 * Until the task is actually later awoken later by
 422			 * the caller, other writers are able to steal it.
 423			 * Readers, on the other hand, will block as they
 424			 * will notice the queued writer.
 425			 */
 426			wake_q_add(wake_q, waiter->task);
 427			lockevent_inc(rwsem_wake_writer);
 428		}
 429
 430		return;
 431	}
 432
 433	/*
 434	 * No reader wakeup if there are too many of them already.
 435	 */
 436	if (unlikely(atomic_long_read(&sem->count) < 0))
 437		return;
 438
 439	/*
 440	 * Writers might steal the lock before we grant it to the next reader.
 441	 * We prefer to do the first reader grant before counting readers
 442	 * so we can bail out early if a writer stole the lock.
 443	 */
 444	if (wake_type != RWSEM_WAKE_READ_OWNED) {
 445		struct task_struct *owner;
 446
 447		adjustment = RWSEM_READER_BIAS;
 448		oldcount = atomic_long_fetch_add(adjustment, &sem->count);
 449		if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
 450			/*
 451			 * When we've been waiting "too" long (for writers
 452			 * to give up the lock), request a HANDOFF to
 453			 * force the issue.
 454			 */
 455			if (!(oldcount & RWSEM_FLAG_HANDOFF) &&
 456			    time_after(jiffies, waiter->timeout)) {
 457				adjustment -= RWSEM_FLAG_HANDOFF;
 458				lockevent_inc(rwsem_rlock_handoff);
 459			}
 460
 461			atomic_long_add(-adjustment, &sem->count);
 462			return;
 463		}
 464		/*
 465		 * Set it to reader-owned to give spinners an early
 466		 * indication that readers now have the lock.
 467		 * The reader nonspinnable bit seen at slowpath entry of
 468		 * the reader is copied over.
 469		 */
 470		owner = waiter->task;
 471		if (waiter->last_rowner & RWSEM_RD_NONSPINNABLE) {
 472			owner = (void *)((unsigned long)owner | RWSEM_RD_NONSPINNABLE);
 473			lockevent_inc(rwsem_opt_norspin);
 474		}
 475		__rwsem_set_reader_owned(sem, owner);
 476	}
 477
 478	/*
 479	 * Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
 480	 * queue. We know that the woken will be at least 1 as we accounted
 481	 * for above. Note we increment the 'active part' of the count by the
 482	 * number of readers before waking any processes up.
 483	 *
 484	 * This is an adaptation of the phase-fair R/W locks where at the
 485	 * reader phase (first waiter is a reader), all readers are eligible
 486	 * to acquire the lock at the same time irrespective of their order
 487	 * in the queue. The writers acquire the lock according to their
 488	 * order in the queue.
 489	 *
 490	 * We have to do wakeup in 2 passes to prevent the possibility that
 491	 * the reader count may be decremented before it is incremented. It
 492	 * is because the to-be-woken waiter may not have slept yet. So it
 493	 * may see waiter->task got cleared, finish its critical section and
 494	 * do an unlock before the reader count increment.
 495	 *
 496	 * 1) Collect the read-waiters in a separate list, count them and
 497	 *    fully increment the reader count in rwsem.
 498	 * 2) For each waiters in the new list, clear waiter->task and
 499	 *    put them into wake_q to be woken up later.
 500	 */
 501	INIT_LIST_HEAD(&wlist);
 502	list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
 503		if (waiter->type == RWSEM_WAITING_FOR_WRITE)
 504			continue;
 505
 506		woken++;
 507		list_move_tail(&waiter->list, &wlist);
 508
 509		/*
 510		 * Limit # of readers that can be woken up per wakeup call.
 511		 */
 512		if (woken >= MAX_READERS_WAKEUP)
 513			break;
 514	}
 515
 516	adjustment = woken * RWSEM_READER_BIAS - adjustment;
 517	lockevent_cond_inc(rwsem_wake_reader, woken);
 518	if (list_empty(&sem->wait_list)) {
 519		/* hit end of list above */
 520		adjustment -= RWSEM_FLAG_WAITERS;
 521	}
 522
 523	/*
 524	 * When we've woken a reader, we no longer need to force writers
 525	 * to give up the lock and we can clear HANDOFF.
 526	 */
 527	if (woken && (atomic_long_read(&sem->count) & RWSEM_FLAG_HANDOFF))
 528		adjustment -= RWSEM_FLAG_HANDOFF;
 529
 530	if (adjustment)
 531		atomic_long_add(adjustment, &sem->count);
 532
 533	/* 2nd pass */
 534	list_for_each_entry_safe(waiter, tmp, &wlist, list) {
 535		struct task_struct *tsk;
 536
 537		tsk = waiter->task;
 538		get_task_struct(tsk);
 539
 540		/*
 541		 * Ensure calling get_task_struct() before setting the reader
 542		 * waiter to nil such that rwsem_down_read_slowpath() cannot
 543		 * race with do_exit() by always holding a reference count
 544		 * to the task to wakeup.
 545		 */
 546		smp_store_release(&waiter->task, NULL);
 547		/*
 548		 * Ensure issuing the wakeup (either by us or someone else)
 549		 * after setting the reader waiter to nil.
 550		 */
 551		wake_q_add_safe(wake_q, tsk);
 552	}
 553}
 554
 555/*
 556 * This function must be called with the sem->wait_lock held to prevent
 557 * race conditions between checking the rwsem wait list and setting the
 558 * sem->count accordingly.
 559 *
 560 * If wstate is WRITER_HANDOFF, it will make sure that either the handoff
 561 * bit is set or the lock is acquired with handoff bit cleared.
 562 */
 563static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
 564					enum writer_wait_state wstate)
 565{
 566	long count, new;
 567
 568	lockdep_assert_held(&sem->wait_lock);
 569
 570	count = atomic_long_read(&sem->count);
 571	do {
 572		bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
 573
 574		if (has_handoff && wstate == WRITER_NOT_FIRST)
 575			return false;
 576
 577		new = count;
 578
 579		if (count & RWSEM_LOCK_MASK) {
 580			if (has_handoff || (wstate != WRITER_HANDOFF))
 581				return false;
 582
 583			new |= RWSEM_FLAG_HANDOFF;
 584		} else {
 585			new |= RWSEM_WRITER_LOCKED;
 586			new &= ~RWSEM_FLAG_HANDOFF;
 587
 588			if (list_is_singular(&sem->wait_list))
 589				new &= ~RWSEM_FLAG_WAITERS;
 590		}
 591	} while (!atomic_long_try_cmpxchg_acquire(&sem->count, &count, new));
 592
 593	/*
 594	 * We have either acquired the lock with handoff bit cleared or
 595	 * set the handoff bit.
 596	 */
 597	if (new & RWSEM_FLAG_HANDOFF)
 598		return false;
 599
 600	rwsem_set_owner(sem);
 601	return true;
 602}
 603
 604#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
 605/*
 606 * Try to acquire read lock before the reader is put on wait queue.
 607 * Lock acquisition isn't allowed if the rwsem is locked or a writer handoff
 608 * is ongoing.
 609 */
 610static inline bool rwsem_try_read_lock_unqueued(struct rw_semaphore *sem)
 611{
 612	long count = atomic_long_read(&sem->count);
 613
 614	if (count & (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))
 615		return false;
 616
 617	count = atomic_long_fetch_add_acquire(RWSEM_READER_BIAS, &sem->count);
 618	if (!(count & (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
 619		rwsem_set_reader_owned(sem);
 620		lockevent_inc(rwsem_opt_rlock);
 621		return true;
 622	}
 623
 624	/* Back out the change */
 625	atomic_long_add(-RWSEM_READER_BIAS, &sem->count);
 626	return false;
 627}
 628
 629/*
 630 * Try to acquire write lock before the writer has been put on wait queue.
 631 */
 632static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
 633{
 634	long count = atomic_long_read(&sem->count);
 635
 636	while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
 637		if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
 638					count | RWSEM_WRITER_LOCKED)) {
 639			rwsem_set_owner(sem);
 640			lockevent_inc(rwsem_opt_wlock);
 641			return true;
 642		}
 643	}
 644	return false;
 645}
 646
 647static inline bool owner_on_cpu(struct task_struct *owner)
 648{
 649	/*
 650	 * As lock holder preemption issue, we both skip spinning if
 651	 * task is not on cpu or its cpu is preempted
 652	 */
 653	return owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
 654}
 655
 656static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem,
 657					   unsigned long nonspinnable)
 658{
 659	struct task_struct *owner;
 660	unsigned long flags;
 661	bool ret = true;
 662
 663	BUILD_BUG_ON(!(RWSEM_OWNER_UNKNOWN & RWSEM_NONSPINNABLE));
 664
 665	if (need_resched()) {
 666		lockevent_inc(rwsem_opt_fail);
 667		return false;
 668	}
 669
 670	preempt_disable();
 671	rcu_read_lock();
 672	owner = rwsem_owner_flags(sem, &flags);
 673	/*
 674	 * Don't check the read-owner as the entry may be stale.
 675	 */
 676	if ((flags & nonspinnable) ||
 677	    (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
 678		ret = false;
 679	rcu_read_unlock();
 680	preempt_enable();
 681
 682	lockevent_cond_inc(rwsem_opt_fail, !ret);
 683	return ret;
 684}
 685
 686/*
 687 * The rwsem_spin_on_owner() function returns the folowing 4 values
 688 * depending on the lock owner state.
 689 *   OWNER_NULL  : owner is currently NULL
 690 *   OWNER_WRITER: when owner changes and is a writer
 691 *   OWNER_READER: when owner changes and the new owner may be a reader.
 692 *   OWNER_NONSPINNABLE:
 693 *		   when optimistic spinning has to stop because either the
 694 *		   owner stops running, is unknown, or its timeslice has
 695 *		   been used up.
 696 */
 697enum owner_state {
 698	OWNER_NULL		= 1 << 0,
 699	OWNER_WRITER		= 1 << 1,
 700	OWNER_READER		= 1 << 2,
 701	OWNER_NONSPINNABLE	= 1 << 3,
 702};
 703#define OWNER_SPINNABLE		(OWNER_NULL | OWNER_WRITER | OWNER_READER)
 704
 705static inline enum owner_state
 706rwsem_owner_state(struct task_struct *owner, unsigned long flags, unsigned long nonspinnable)
 707{
 708	if (flags & nonspinnable)
 709		return OWNER_NONSPINNABLE;
 710
 711	if (flags & RWSEM_READER_OWNED)
 712		return OWNER_READER;
 713
 714	return owner ? OWNER_WRITER : OWNER_NULL;
 715}
 716
 717static noinline enum owner_state
 718rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
 719{
 720	struct task_struct *new, *owner;
 721	unsigned long flags, new_flags;
 722	enum owner_state state;
 723
 724	owner = rwsem_owner_flags(sem, &flags);
 725	state = rwsem_owner_state(owner, flags, nonspinnable);
 726	if (state != OWNER_WRITER)
 727		return state;
 728
 729	rcu_read_lock();
 730	for (;;) {
 731		/*
 732		 * When a waiting writer set the handoff flag, it may spin
 733		 * on the owner as well. Once that writer acquires the lock,
 734		 * we can spin on it. So we don't need to quit even when the
 735		 * handoff bit is set.
 736		 */
 737		new = rwsem_owner_flags(sem, &new_flags);
 738		if ((new != owner) || (new_flags != flags)) {
 739			state = rwsem_owner_state(new, new_flags, nonspinnable);
 740			break;
 741		}
 742
 743		/*
 744		 * Ensure we emit the owner->on_cpu, dereference _after_
 745		 * checking sem->owner still matches owner, if that fails,
 746		 * owner might point to free()d memory, if it still matches,
 747		 * the rcu_read_lock() ensures the memory stays valid.
 748		 */
 749		barrier();
 750
 751		if (need_resched() || !owner_on_cpu(owner)) {
 752			state = OWNER_NONSPINNABLE;
 753			break;
 754		}
 755
 756		cpu_relax();
 757	}
 758	rcu_read_unlock();
 759
 760	return state;
 761}
 762
 763/*
 764 * Calculate reader-owned rwsem spinning threshold for writer
 765 *
 766 * The more readers own the rwsem, the longer it will take for them to
 767 * wind down and free the rwsem. So the empirical formula used to
 768 * determine the actual spinning time limit here is:
 769 *
 770 *   Spinning threshold = (10 + nr_readers/2)us
 771 *
 772 * The limit is capped to a maximum of 25us (30 readers). This is just
 773 * a heuristic and is subjected to change in the future.
 774 */
 775static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
 776{
 777	long count = atomic_long_read(&sem->count);
 778	int readers = count >> RWSEM_READER_SHIFT;
 779	u64 delta;
 780
 781	if (readers > 30)
 782		readers = 30;
 783	delta = (20 + readers) * NSEC_PER_USEC / 2;
 784
 785	return sched_clock() + delta;
 786}
 787
 788static bool rwsem_optimistic_spin(struct rw_semaphore *sem, bool wlock)
 789{
 790	bool taken = false;
 791	int prev_owner_state = OWNER_NULL;
 792	int loop = 0;
 793	u64 rspin_threshold = 0;
 794	unsigned long nonspinnable = wlock ? RWSEM_WR_NONSPINNABLE
 795					   : RWSEM_RD_NONSPINNABLE;
 796
 797	preempt_disable();
 798
 799	/* sem->wait_lock should not be held when doing optimistic spinning */
 800	if (!osq_lock(&sem->osq))
 801		goto done;
 802
 803	/*
 804	 * Optimistically spin on the owner field and attempt to acquire the
 805	 * lock whenever the owner changes. Spinning will be stopped when:
 806	 *  1) the owning writer isn't running; or
 807	 *  2) readers own the lock and spinning time has exceeded limit.
 808	 */
 809	for (;;) {
 810		enum owner_state owner_state;
 811
 812		owner_state = rwsem_spin_on_owner(sem, nonspinnable);
 813		if (!(owner_state & OWNER_SPINNABLE))
 814			break;
 815
 816		/*
 817		 * Try to acquire the lock
 818		 */
 819		taken = wlock ? rwsem_try_write_lock_unqueued(sem)
 820			      : rwsem_try_read_lock_unqueued(sem);
 821
 822		if (taken)
 823			break;
 824
 825		/*
 826		 * Time-based reader-owned rwsem optimistic spinning
 827		 */
 828		if (wlock && (owner_state == OWNER_READER)) {
 829			/*
 830			 * Re-initialize rspin_threshold every time when
 831			 * the owner state changes from non-reader to reader.
 832			 * This allows a writer to steal the lock in between
 833			 * 2 reader phases and have the threshold reset at
 834			 * the beginning of the 2nd reader phase.
 835			 */
 836			if (prev_owner_state != OWNER_READER) {
 837				if (rwsem_test_oflags(sem, nonspinnable))
 838					break;
 839				rspin_threshold = rwsem_rspin_threshold(sem);
 840				loop = 0;
 841			}
 842
 843			/*
 844			 * Check time threshold once every 16 iterations to
 845			 * avoid calling sched_clock() too frequently so
 846			 * as to reduce the average latency between the times
 847			 * when the lock becomes free and when the spinner
 848			 * is ready to do a trylock.
 849			 */
 850			else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
 851				rwsem_set_nonspinnable(sem);
 852				lockevent_inc(rwsem_opt_nospin);
 853				break;
 854			}
 855		}
 856
 857		/*
 858		 * An RT task cannot do optimistic spinning if it cannot
 859		 * be sure the lock holder is running or live-lock may
 860		 * happen if the current task and the lock holder happen
 861		 * to run in the same CPU. However, aborting optimistic
 862		 * spinning while a NULL owner is detected may miss some
 863		 * opportunity where spinning can continue without causing
 864		 * problem.
 865		 *
 866		 * There are 2 possible cases where an RT task may be able
 867		 * to continue spinning.
 868		 *
 869		 * 1) The lock owner is in the process of releasing the
 870		 *    lock, sem->owner is cleared but the lock has not
 871		 *    been released yet.
 872		 * 2) The lock was free and owner cleared, but another
 873		 *    task just comes in and acquire the lock before
 874		 *    we try to get it. The new owner may be a spinnable
 875		 *    writer.
 876		 *
 877		 * To take advantage of two scenarios listed agove, the RT
 878		 * task is made to retry one more time to see if it can
 879		 * acquire the lock or continue spinning on the new owning
 880		 * writer. Of course, if the time lag is long enough or the
 881		 * new owner is not a writer or spinnable, the RT task will
 882		 * quit spinning.
 883		 *
 884		 * If the owner is a writer, the need_resched() check is
 885		 * done inside rwsem_spin_on_owner(). If the owner is not
 886		 * a writer, need_resched() check needs to be done here.
 887		 */
 888		if (owner_state != OWNER_WRITER) {
 889			if (need_resched())
 890				break;
 891			if (rt_task(current) &&
 892			   (prev_owner_state != OWNER_WRITER))
 893				break;
 894		}
 895		prev_owner_state = owner_state;
 896
 897		/*
 898		 * The cpu_relax() call is a compiler barrier which forces
 899		 * everything in this loop to be re-loaded. We don't need
 900		 * memory barriers as we'll eventually observe the right
 901		 * values at the cost of a few extra spins.
 902		 */
 903		cpu_relax();
 904	}
 905	osq_unlock(&sem->osq);
 906done:
 907	preempt_enable();
 908	lockevent_cond_inc(rwsem_opt_fail, !taken);
 909	return taken;
 910}
 911
 912/*
 913 * Clear the owner's RWSEM_WR_NONSPINNABLE bit if it is set. This should
 914 * only be called when the reader count reaches 0.
 915 *
 916 * This give writers better chance to acquire the rwsem first before
 917 * readers when the rwsem was being held by readers for a relatively long
 918 * period of time. Race can happen that an optimistic spinner may have
 919 * just stolen the rwsem and set the owner, but just clearing the
 920 * RWSEM_WR_NONSPINNABLE bit will do no harm anyway.
 921 */
 922static inline void clear_wr_nonspinnable(struct rw_semaphore *sem)
 923{
 924	if (rwsem_test_oflags(sem, RWSEM_WR_NONSPINNABLE))
 925		atomic_long_andnot(RWSEM_WR_NONSPINNABLE, &sem->owner);
 926}
 927
 928/*
 929 * This function is called when the reader fails to acquire the lock via
 930 * optimistic spinning. In this case we will still attempt to do a trylock
 931 * when comparing the rwsem state right now with the state when entering
 932 * the slowpath indicates that the reader is still in a valid reader phase.
 933 * This happens when the following conditions are true:
 934 *
 935 * 1) The lock is currently reader owned, and
 936 * 2) The lock is previously not reader-owned or the last read owner changes.
 937 *
 938 * In the former case, we have transitioned from a writer phase to a
 939 * reader-phase while spinning. In the latter case, it means the reader
 940 * phase hasn't ended when we entered the optimistic spinning loop. In
 941 * both cases, the reader is eligible to acquire the lock. This is the
 942 * secondary path where a read lock is acquired optimistically.
 943 *
 944 * The reader non-spinnable bit wasn't set at time of entry or it will
 945 * not be here at all.
 946 */
 947static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
 948					      unsigned long last_rowner)
 949{
 950	unsigned long owner = atomic_long_read(&sem->owner);
 951
 952	if (!(owner & RWSEM_READER_OWNED))
 953		return false;
 954
 955	if (((owner ^ last_rowner) & ~RWSEM_OWNER_FLAGS_MASK) &&
 956	    rwsem_try_read_lock_unqueued(sem)) {
 957		lockevent_inc(rwsem_opt_rlock2);
 958		lockevent_add(rwsem_opt_fail, -1);
 959		return true;
 960	}
 961	return false;
 962}
 963#else
 964static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem,
 965					   unsigned long nonspinnable)
 966{
 967	return false;
 968}
 969
 970static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem, bool wlock)
 971{
 972	return false;
 973}
 974
 975static inline void clear_wr_nonspinnable(struct rw_semaphore *sem) { }
 976
 977static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
 978					      unsigned long last_rowner)
 979{
 980	return false;
 981}
 982
 983static inline int
 984rwsem_spin_on_owner(struct rw_semaphore *sem, unsigned long nonspinnable)
 985{
 986	return 0;
 987}
 988#define OWNER_NULL	1
 989#endif
 990
 991/*
 992 * Wait for the read lock to be granted
 993 */
 994static struct rw_semaphore __sched *
 995rwsem_down_read_slowpath(struct rw_semaphore *sem, int state)
 996{
 997	long count, adjustment = -RWSEM_READER_BIAS;
 998	struct rwsem_waiter waiter;
 999	DEFINE_WAKE_Q(wake_q);
1000	bool wake = false;
1001
1002	/*
1003	 * Save the current read-owner of rwsem, if available, and the
1004	 * reader nonspinnable bit.
1005	 */
1006	waiter.last_rowner = atomic_long_read(&sem->owner);
1007	if (!(waiter.last_rowner & RWSEM_READER_OWNED))
1008		waiter.last_rowner &= RWSEM_RD_NONSPINNABLE;
1009
1010	if (!rwsem_can_spin_on_owner(sem, RWSEM_RD_NONSPINNABLE))
1011		goto queue;
1012
1013	/*
1014	 * Undo read bias from down_read() and do optimistic spinning.
1015	 */
1016	atomic_long_add(-RWSEM_READER_BIAS, &sem->count);
1017	adjustment = 0;
1018	if (rwsem_optimistic_spin(sem, false)) {
1019		/* rwsem_optimistic_spin() implies ACQUIRE on success */
1020		/*
1021		 * Wake up other readers in the wait list if the front
1022		 * waiter is a reader.
1023		 */
1024		if ((atomic_long_read(&sem->count) & RWSEM_FLAG_WAITERS)) {
1025			raw_spin_lock_irq(&sem->wait_lock);
1026			if (!list_empty(&sem->wait_list))
1027				rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
1028						&wake_q);
1029			raw_spin_unlock_irq(&sem->wait_lock);
1030			wake_up_q(&wake_q);
1031		}
1032		return sem;
1033	} else if (rwsem_reader_phase_trylock(sem, waiter.last_rowner)) {
1034		/* rwsem_reader_phase_trylock() implies ACQUIRE on success */
1035		return sem;
1036	}
1037
1038queue:
1039	waiter.task = current;
1040	waiter.type = RWSEM_WAITING_FOR_READ;
1041	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1042
1043	raw_spin_lock_irq(&sem->wait_lock);
1044	if (list_empty(&sem->wait_list)) {
1045		/*
1046		 * In case the wait queue is empty and the lock isn't owned
1047		 * by a writer or has the handoff bit set, this reader can
1048		 * exit the slowpath and return immediately as its
1049		 * RWSEM_READER_BIAS has already been set in the count.
1050		 */
1051		if (adjustment && !(atomic_long_read(&sem->count) &
1052		     (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
1053			/* Provide lock ACQUIRE */
1054			smp_acquire__after_ctrl_dep();
1055			raw_spin_unlock_irq(&sem->wait_lock);
1056			rwsem_set_reader_owned(sem);
1057			lockevent_inc(rwsem_rlock_fast);
1058			return sem;
1059		}
1060		adjustment += RWSEM_FLAG_WAITERS;
1061	}
1062	list_add_tail(&waiter.list, &sem->wait_list);
1063
1064	/* we're now waiting on the lock, but no longer actively locking */
1065	if (adjustment)
1066		count = atomic_long_add_return(adjustment, &sem->count);
1067	else
1068		count = atomic_long_read(&sem->count);
1069
1070	/*
1071	 * If there are no active locks, wake the front queued process(es).
1072	 *
1073	 * If there are no writers and we are first in the queue,
1074	 * wake our own waiter to join the existing active readers !
1075	 */
1076	if (!(count & RWSEM_LOCK_MASK)) {
1077		clear_wr_nonspinnable(sem);
1078		wake = true;
1079	}
1080	if (wake || (!(count & RWSEM_WRITER_MASK) &&
1081		    (adjustment & RWSEM_FLAG_WAITERS)))
1082		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1083
1084	raw_spin_unlock_irq(&sem->wait_lock);
1085	wake_up_q(&wake_q);
1086
1087	/* wait to be given the lock */
1088	for (;;) {
1089		set_current_state(state);
1090		if (!smp_load_acquire(&waiter.task)) {
1091			/* Matches rwsem_mark_wake()'s smp_store_release(). */
1092			break;
1093		}
1094		if (signal_pending_state(state, current)) {
1095			raw_spin_lock_irq(&sem->wait_lock);
1096			if (waiter.task)
1097				goto out_nolock;
1098			raw_spin_unlock_irq(&sem->wait_lock);
1099			/* Ordered by sem->wait_lock against rwsem_mark_wake(). */
1100			break;
1101		}
1102		schedule();
1103		lockevent_inc(rwsem_sleep_reader);
1104	}
1105
1106	__set_current_state(TASK_RUNNING);
1107	lockevent_inc(rwsem_rlock);
1108	return sem;
1109
1110out_nolock:
1111	list_del(&waiter.list);
1112	if (list_empty(&sem->wait_list)) {
1113		atomic_long_andnot(RWSEM_FLAG_WAITERS|RWSEM_FLAG_HANDOFF,
1114				   &sem->count);
1115	}
1116	raw_spin_unlock_irq(&sem->wait_lock);
1117	__set_current_state(TASK_RUNNING);
1118	lockevent_inc(rwsem_rlock_fail);
1119	return ERR_PTR(-EINTR);
1120}
1121
1122/*
1123 * This function is called by the a write lock owner. So the owner value
1124 * won't get changed by others.
1125 */
1126static inline void rwsem_disable_reader_optspin(struct rw_semaphore *sem,
1127						bool disable)
1128{
1129	if (unlikely(disable)) {
1130		atomic_long_or(RWSEM_RD_NONSPINNABLE, &sem->owner);
1131		lockevent_inc(rwsem_opt_norspin);
1132	}
1133}
1134
1135/*
1136 * Wait until we successfully acquire the write lock
1137 */
1138static struct rw_semaphore *
1139rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1140{
1141	long count;
1142	bool disable_rspin;
1143	enum writer_wait_state wstate;
1144	struct rwsem_waiter waiter;
1145	struct rw_semaphore *ret = sem;
1146	DEFINE_WAKE_Q(wake_q);
1147
1148	/* do optimistic spinning and steal lock if possible */
1149	if (rwsem_can_spin_on_owner(sem, RWSEM_WR_NONSPINNABLE) &&
1150	    rwsem_optimistic_spin(sem, true)) {
1151		/* rwsem_optimistic_spin() implies ACQUIRE on success */
1152		return sem;
1153	}
1154
1155	/*
1156	 * Disable reader optimistic spinning for this rwsem after
1157	 * acquiring the write lock when the setting of the nonspinnable
1158	 * bits are observed.
1159	 */
1160	disable_rspin = atomic_long_read(&sem->owner) & RWSEM_NONSPINNABLE;
1161
1162	/*
1163	 * Optimistic spinning failed, proceed to the slowpath
1164	 * and block until we can acquire the sem.
1165	 */
1166	waiter.task = current;
1167	waiter.type = RWSEM_WAITING_FOR_WRITE;
1168	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1169
1170	raw_spin_lock_irq(&sem->wait_lock);
1171
1172	/* account for this before adding a new element to the list */
1173	wstate = list_empty(&sem->wait_list) ? WRITER_FIRST : WRITER_NOT_FIRST;
1174
1175	list_add_tail(&waiter.list, &sem->wait_list);
1176
1177	/* we're now waiting on the lock */
1178	if (wstate == WRITER_NOT_FIRST) {
1179		count = atomic_long_read(&sem->count);
1180
1181		/*
1182		 * If there were already threads queued before us and:
1183		 *  1) there are no no active locks, wake the front
1184		 *     queued process(es) as the handoff bit might be set.
1185		 *  2) there are no active writers and some readers, the lock
1186		 *     must be read owned; so we try to wake any read lock
1187		 *     waiters that were queued ahead of us.
1188		 */
1189		if (count & RWSEM_WRITER_MASK)
1190			goto wait;
1191
1192		rwsem_mark_wake(sem, (count & RWSEM_READER_MASK)
1193					? RWSEM_WAKE_READERS
1194					: RWSEM_WAKE_ANY, &wake_q);
1195
1196		if (!wake_q_empty(&wake_q)) {
1197			/*
1198			 * We want to minimize wait_lock hold time especially
1199			 * when a large number of readers are to be woken up.
1200			 */
1201			raw_spin_unlock_irq(&sem->wait_lock);
1202			wake_up_q(&wake_q);
1203			wake_q_init(&wake_q);	/* Used again, reinit */
1204			raw_spin_lock_irq(&sem->wait_lock);
1205		}
1206	} else {
1207		atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
1208	}
1209
1210wait:
1211	/* wait until we successfully acquire the lock */
1212	set_current_state(state);
1213	for (;;) {
1214		if (rwsem_try_write_lock(sem, wstate)) {
1215			/* rwsem_try_write_lock() implies ACQUIRE on success */
1216			break;
1217		}
1218
1219		raw_spin_unlock_irq(&sem->wait_lock);
1220
1221		/*
1222		 * After setting the handoff bit and failing to acquire
1223		 * the lock, attempt to spin on owner to accelerate lock
1224		 * transfer. If the previous owner is a on-cpu writer and it
1225		 * has just released the lock, OWNER_NULL will be returned.
1226		 * In this case, we attempt to acquire the lock again
1227		 * without sleeping.
1228		 */
1229		if ((wstate == WRITER_HANDOFF) &&
1230		    (rwsem_spin_on_owner(sem, 0) == OWNER_NULL))
1231			goto trylock_again;
1232
1233		/* Block until there are no active lockers. */
1234		for (;;) {
1235			if (signal_pending_state(state, current))
1236				goto out_nolock;
1237
1238			schedule();
1239			lockevent_inc(rwsem_sleep_writer);
1240			set_current_state(state);
1241			/*
1242			 * If HANDOFF bit is set, unconditionally do
1243			 * a trylock.
1244			 */
1245			if (wstate == WRITER_HANDOFF)
1246				break;
1247
1248			if ((wstate == WRITER_NOT_FIRST) &&
1249			    (rwsem_first_waiter(sem) == &waiter))
1250				wstate = WRITER_FIRST;
1251
1252			count = atomic_long_read(&sem->count);
1253			if (!(count & RWSEM_LOCK_MASK))
1254				break;
1255
1256			/*
1257			 * The setting of the handoff bit is deferred
1258			 * until rwsem_try_write_lock() is called.
1259			 */
1260			if ((wstate == WRITER_FIRST) && (rt_task(current) ||
1261			    time_after(jiffies, waiter.timeout))) {
1262				wstate = WRITER_HANDOFF;
1263				lockevent_inc(rwsem_wlock_handoff);
1264				break;
1265			}
1266		}
1267trylock_again:
1268		raw_spin_lock_irq(&sem->wait_lock);
1269	}
1270	__set_current_state(TASK_RUNNING);
1271	list_del(&waiter.list);
1272	rwsem_disable_reader_optspin(sem, disable_rspin);
1273	raw_spin_unlock_irq(&sem->wait_lock);
1274	lockevent_inc(rwsem_wlock);
1275
1276	return ret;
1277
1278out_nolock:
1279	__set_current_state(TASK_RUNNING);
1280	raw_spin_lock_irq(&sem->wait_lock);
1281	list_del(&waiter.list);
1282
1283	if (unlikely(wstate == WRITER_HANDOFF))
1284		atomic_long_add(-RWSEM_FLAG_HANDOFF,  &sem->count);
1285
1286	if (list_empty(&sem->wait_list))
1287		atomic_long_andnot(RWSEM_FLAG_WAITERS, &sem->count);
1288	else
1289		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1290	raw_spin_unlock_irq(&sem->wait_lock);
1291	wake_up_q(&wake_q);
1292	lockevent_inc(rwsem_wlock_fail);
1293
1294	return ERR_PTR(-EINTR);
1295}
1296
1297/*
1298 * handle waking up a waiter on the semaphore
1299 * - up_read/up_write has decremented the active part of count if we come here
1300 */
1301static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem, long count)
1302{
1303	unsigned long flags;
1304	DEFINE_WAKE_Q(wake_q);
1305
1306	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1307
1308	if (!list_empty(&sem->wait_list))
1309		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1310
1311	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1312	wake_up_q(&wake_q);
1313
1314	return sem;
1315}
1316
1317/*
1318 * downgrade a write lock into a read lock
1319 * - caller incremented waiting part of count and discovered it still negative
1320 * - just wake up any readers at the front of the queue
1321 */
1322static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1323{
1324	unsigned long flags;
1325	DEFINE_WAKE_Q(wake_q);
1326
1327	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1328
1329	if (!list_empty(&sem->wait_list))
1330		rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
1331
1332	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1333	wake_up_q(&wake_q);
1334
1335	return sem;
1336}
1337
1338/*
1339 * lock for reading
1340 */
1341inline void __down_read(struct rw_semaphore *sem)
1342{
1343	if (!rwsem_read_trylock(sem)) {
1344		rwsem_down_read_slowpath(sem, TASK_UNINTERRUPTIBLE);
1345		DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1346	} else {
1347		rwsem_set_reader_owned(sem);
1348	}
1349}
1350
1351static inline int __down_read_killable(struct rw_semaphore *sem)
1352{
1353	if (!rwsem_read_trylock(sem)) {
1354		if (IS_ERR(rwsem_down_read_slowpath(sem, TASK_KILLABLE)))
1355			return -EINTR;
1356		DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1357	} else {
1358		rwsem_set_reader_owned(sem);
1359	}
1360	return 0;
1361}
1362
1363static inline int __down_read_trylock(struct rw_semaphore *sem)
1364{
1365	long tmp;
1366
1367	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1368
1369	/*
1370	 * Optimize for the case when the rwsem is not locked at all.
1371	 */
1372	tmp = RWSEM_UNLOCKED_VALUE;
1373	do {
1374		if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1375					tmp + RWSEM_READER_BIAS)) {
1376			rwsem_set_reader_owned(sem);
1377			return 1;
1378		}
1379	} while (!(tmp & RWSEM_READ_FAILED_MASK));
1380	return 0;
1381}
1382
1383/*
1384 * lock for writing
1385 */
1386static inline void __down_write(struct rw_semaphore *sem)
1387{
1388	long tmp = RWSEM_UNLOCKED_VALUE;
1389
1390	if (unlikely(!atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1391						      RWSEM_WRITER_LOCKED)))
1392		rwsem_down_write_slowpath(sem, TASK_UNINTERRUPTIBLE);
1393	else
1394		rwsem_set_owner(sem);
1395}
1396
1397static inline int __down_write_killable(struct rw_semaphore *sem)
1398{
1399	long tmp = RWSEM_UNLOCKED_VALUE;
1400
1401	if (unlikely(!atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1402						      RWSEM_WRITER_LOCKED))) {
1403		if (IS_ERR(rwsem_down_write_slowpath(sem, TASK_KILLABLE)))
1404			return -EINTR;
1405	} else {
1406		rwsem_set_owner(sem);
1407	}
1408	return 0;
1409}
1410
1411static inline int __down_write_trylock(struct rw_semaphore *sem)
1412{
1413	long tmp;
1414
1415	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1416
1417	tmp  = RWSEM_UNLOCKED_VALUE;
1418	if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1419					    RWSEM_WRITER_LOCKED)) {
1420		rwsem_set_owner(sem);
1421		return true;
1422	}
1423	return false;
1424}
1425
1426/*
1427 * unlock after reading
1428 */
1429inline void __up_read(struct rw_semaphore *sem)
1430{
1431	long tmp;
1432
1433	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1434	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1435
1436	rwsem_clear_reader_owned(sem);
1437	tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
1438	DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1439	if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1440		      RWSEM_FLAG_WAITERS)) {
1441		clear_wr_nonspinnable(sem);
1442		rwsem_wake(sem, tmp);
1443	}
1444}
1445
1446/*
1447 * unlock after writing
1448 */
1449static inline void __up_write(struct rw_semaphore *sem)
1450{
1451	long tmp;
1452
1453	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1454	/*
1455	 * sem->owner may differ from current if the ownership is transferred
1456	 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1457	 */
1458	DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1459			    !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1460
1461	rwsem_clear_owner(sem);
1462	tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
1463	if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1464		rwsem_wake(sem, tmp);
1465}
1466
1467/*
1468 * downgrade write lock to read lock
1469 */
1470static inline void __downgrade_write(struct rw_semaphore *sem)
1471{
1472	long tmp;
1473
1474	/*
1475	 * When downgrading from exclusive to shared ownership,
1476	 * anything inside the write-locked region cannot leak
1477	 * into the read side. In contrast, anything in the
1478	 * read-locked region is ok to be re-ordered into the
1479	 * write side. As such, rely on RELEASE semantics.
1480	 */
1481	DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1482	tmp = atomic_long_fetch_add_release(
1483		-RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
1484	rwsem_set_reader_owned(sem);
1485	if (tmp & RWSEM_FLAG_WAITERS)
1486		rwsem_downgrade_wake(sem);
1487}
1488
1489/*
1490 * lock for reading
1491 */
1492void __sched down_read(struct rw_semaphore *sem)
1493{
1494	might_sleep();
1495	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1496
1497	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1498}
1499EXPORT_SYMBOL(down_read);
1500
1501int __sched down_read_killable(struct rw_semaphore *sem)
1502{
1503	might_sleep();
1504	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1505
1506	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1507		rwsem_release(&sem->dep_map, 1, _RET_IP_);
1508		return -EINTR;
1509	}
1510
1511	return 0;
1512}
1513EXPORT_SYMBOL(down_read_killable);
1514
1515/*
1516 * trylock for reading -- returns 1 if successful, 0 if contention
1517 */
1518int down_read_trylock(struct rw_semaphore *sem)
1519{
1520	int ret = __down_read_trylock(sem);
1521
1522	if (ret == 1)
1523		rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1524	return ret;
1525}
1526EXPORT_SYMBOL(down_read_trylock);
1527
1528/*
1529 * lock for writing
1530 */
1531void __sched down_write(struct rw_semaphore *sem)
1532{
1533	might_sleep();
1534	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1535	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1536}
1537EXPORT_SYMBOL(down_write);
1538
1539/*
1540 * lock for writing
1541 */
1542int __sched down_write_killable(struct rw_semaphore *sem)
1543{
1544	might_sleep();
1545	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1546
1547	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1548				  __down_write_killable)) {
1549		rwsem_release(&sem->dep_map, 1, _RET_IP_);
1550		return -EINTR;
1551	}
1552
1553	return 0;
1554}
1555EXPORT_SYMBOL(down_write_killable);
1556
1557/*
1558 * trylock for writing -- returns 1 if successful, 0 if contention
1559 */
1560int down_write_trylock(struct rw_semaphore *sem)
1561{
1562	int ret = __down_write_trylock(sem);
1563
1564	if (ret == 1)
1565		rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1566
1567	return ret;
1568}
1569EXPORT_SYMBOL(down_write_trylock);
1570
1571/*
1572 * release a read lock
1573 */
1574void up_read(struct rw_semaphore *sem)
1575{
1576	rwsem_release(&sem->dep_map, 1, _RET_IP_);
1577	__up_read(sem);
1578}
1579EXPORT_SYMBOL(up_read);
1580
1581/*
1582 * release a write lock
1583 */
1584void up_write(struct rw_semaphore *sem)
1585{
1586	rwsem_release(&sem->dep_map, 1, _RET_IP_);
1587	__up_write(sem);
1588}
1589EXPORT_SYMBOL(up_write);
1590
1591/*
1592 * downgrade write lock to read lock
1593 */
1594void downgrade_write(struct rw_semaphore *sem)
1595{
1596	lock_downgrade(&sem->dep_map, _RET_IP_);
1597	__downgrade_write(sem);
1598}
1599EXPORT_SYMBOL(downgrade_write);
1600
1601#ifdef CONFIG_DEBUG_LOCK_ALLOC
1602
1603void down_read_nested(struct rw_semaphore *sem, int subclass)
1604{
1605	might_sleep();
1606	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1607	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1608}
1609EXPORT_SYMBOL(down_read_nested);
1610
1611void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1612{
1613	might_sleep();
1614	rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1615	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1616}
1617EXPORT_SYMBOL(_down_write_nest_lock);
1618
1619void down_read_non_owner(struct rw_semaphore *sem)
1620{
1621	might_sleep();
1622	__down_read(sem);
1623	__rwsem_set_reader_owned(sem, NULL);
1624}
1625EXPORT_SYMBOL(down_read_non_owner);
1626
1627void down_write_nested(struct rw_semaphore *sem, int subclass)
1628{
1629	might_sleep();
1630	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1631	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1632}
1633EXPORT_SYMBOL(down_write_nested);
1634
1635int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1636{
1637	might_sleep();
1638	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1639
1640	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1641				  __down_write_killable)) {
1642		rwsem_release(&sem->dep_map, 1, _RET_IP_);
1643		return -EINTR;
1644	}
1645
1646	return 0;
1647}
1648EXPORT_SYMBOL(down_write_killable_nested);
1649
1650void up_read_non_owner(struct rw_semaphore *sem)
1651{
1652	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1653	__up_read(sem);
1654}
1655EXPORT_SYMBOL(up_read_non_owner);
1656
1657#endif