Linux Audio

Check our new training course

Loading...
v4.10.11
   1/*
   2 *  Fast Userspace Mutexes (which I call "Futexes!").
   3 *  (C) Rusty Russell, IBM 2002
   4 *
   5 *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
   6 *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
   7 *
   8 *  Removed page pinning, fix privately mapped COW pages and other cleanups
   9 *  (C) Copyright 2003, 2004 Jamie Lokier
  10 *
  11 *  Robust futex support started by Ingo Molnar
  12 *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
  13 *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
  14 *
  15 *  PI-futex support started by Ingo Molnar and Thomas Gleixner
  16 *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  17 *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
  18 *
  19 *  PRIVATE futexes by Eric Dumazet
  20 *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
  21 *
  22 *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
  23 *  Copyright (C) IBM Corporation, 2009
  24 *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
  25 *
  26 *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
  27 *  enough at me, Linus for the original (flawed) idea, Matthew
  28 *  Kirkwood for proof-of-concept implementation.
  29 *
  30 *  "The futexes are also cursed."
  31 *  "But they come in a choice of three flavours!"
  32 *
  33 *  This program is free software; you can redistribute it and/or modify
  34 *  it under the terms of the GNU General Public License as published by
  35 *  the Free Software Foundation; either version 2 of the License, or
  36 *  (at your option) any later version.
  37 *
  38 *  This program is distributed in the hope that it will be useful,
  39 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  40 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  41 *  GNU General Public License for more details.
  42 *
  43 *  You should have received a copy of the GNU General Public License
  44 *  along with this program; if not, write to the Free Software
  45 *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  46 */
  47#include <linux/slab.h>
  48#include <linux/poll.h>
  49#include <linux/fs.h>
  50#include <linux/file.h>
  51#include <linux/jhash.h>
  52#include <linux/init.h>
  53#include <linux/futex.h>
  54#include <linux/mount.h>
  55#include <linux/pagemap.h>
  56#include <linux/syscalls.h>
  57#include <linux/signal.h>
  58#include <linux/export.h>
  59#include <linux/magic.h>
  60#include <linux/pid.h>
  61#include <linux/nsproxy.h>
  62#include <linux/ptrace.h>
  63#include <linux/sched/rt.h>
  64#include <linux/hugetlb.h>
  65#include <linux/freezer.h>
  66#include <linux/bootmem.h>
  67#include <linux/fault-inject.h>
  68
  69#include <asm/futex.h>
  70
  71#include "locking/rtmutex_common.h"
  72
  73/*
  74 * READ this before attempting to hack on futexes!
  75 *
  76 * Basic futex operation and ordering guarantees
  77 * =============================================
  78 *
  79 * The waiter reads the futex value in user space and calls
  80 * futex_wait(). This function computes the hash bucket and acquires
  81 * the hash bucket lock. After that it reads the futex user space value
  82 * again and verifies that the data has not changed. If it has not changed
  83 * it enqueues itself into the hash bucket, releases the hash bucket lock
  84 * and schedules.
  85 *
  86 * The waker side modifies the user space value of the futex and calls
  87 * futex_wake(). This function computes the hash bucket and acquires the
  88 * hash bucket lock. Then it looks for waiters on that futex in the hash
  89 * bucket and wakes them.
  90 *
  91 * In futex wake up scenarios where no tasks are blocked on a futex, taking
  92 * the hb spinlock can be avoided and simply return. In order for this
  93 * optimization to work, ordering guarantees must exist so that the waiter
  94 * being added to the list is acknowledged when the list is concurrently being
  95 * checked by the waker, avoiding scenarios like the following:
  96 *
  97 * CPU 0                               CPU 1
  98 * val = *futex;
  99 * sys_futex(WAIT, futex, val);
 100 *   futex_wait(futex, val);
 101 *   uval = *futex;
 102 *                                     *futex = newval;
 103 *                                     sys_futex(WAKE, futex);
 104 *                                       futex_wake(futex);
 105 *                                       if (queue_empty())
 106 *                                         return;
 107 *   if (uval == val)
 108 *      lock(hash_bucket(futex));
 109 *      queue();
 110 *     unlock(hash_bucket(futex));
 111 *     schedule();
 112 *
 113 * This would cause the waiter on CPU 0 to wait forever because it
 114 * missed the transition of the user space value from val to newval
 115 * and the waker did not find the waiter in the hash bucket queue.
 116 *
 117 * The correct serialization ensures that a waiter either observes
 118 * the changed user space value before blocking or is woken by a
 119 * concurrent waker:
 120 *
 121 * CPU 0                                 CPU 1
 122 * val = *futex;
 123 * sys_futex(WAIT, futex, val);
 124 *   futex_wait(futex, val);
 125 *
 126 *   waiters++; (a)
 127 *   smp_mb(); (A) <-- paired with -.
 128 *                                  |
 129 *   lock(hash_bucket(futex));      |
 130 *                                  |
 131 *   uval = *futex;                 |
 132 *                                  |        *futex = newval;
 133 *                                  |        sys_futex(WAKE, futex);
 134 *                                  |          futex_wake(futex);
 135 *                                  |
 136 *                                  `--------> smp_mb(); (B)
 137 *   if (uval == val)
 138 *     queue();
 139 *     unlock(hash_bucket(futex));
 140 *     schedule();                         if (waiters)
 141 *                                           lock(hash_bucket(futex));
 142 *   else                                    wake_waiters(futex);
 143 *     waiters--; (b)                        unlock(hash_bucket(futex));
 144 *
 145 * Where (A) orders the waiters increment and the futex value read through
 146 * atomic operations (see hb_waiters_inc) and where (B) orders the write
 147 * to futex and the waiters read -- this is done by the barriers for both
 148 * shared and private futexes in get_futex_key_refs().
 149 *
 150 * This yields the following case (where X:=waiters, Y:=futex):
 151 *
 152 *	X = Y = 0
 153 *
 154 *	w[X]=1		w[Y]=1
 155 *	MB		MB
 156 *	r[Y]=y		r[X]=x
 157 *
 158 * Which guarantees that x==0 && y==0 is impossible; which translates back into
 159 * the guarantee that we cannot both miss the futex variable change and the
 160 * enqueue.
 161 *
 162 * Note that a new waiter is accounted for in (a) even when it is possible that
 163 * the wait call can return error, in which case we backtrack from it in (b).
 164 * Refer to the comment in queue_lock().
 165 *
 166 * Similarly, in order to account for waiters being requeued on another
 167 * address we always increment the waiters for the destination bucket before
 168 * acquiring the lock. It then decrements them again  after releasing it -
 169 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
 170 * will do the additional required waiter count housekeeping. This is done for
 171 * double_lock_hb() and double_unlock_hb(), respectively.
 172 */
 173
 174#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
 175int __read_mostly futex_cmpxchg_enabled;
 176#endif
 177
 178/*
 179 * Futex flags used to encode options to functions and preserve them across
 180 * restarts.
 181 */
 182#ifdef CONFIG_MMU
 183# define FLAGS_SHARED		0x01
 184#else
 185/*
 186 * NOMMU does not have per process address space. Let the compiler optimize
 187 * code away.
 188 */
 189# define FLAGS_SHARED		0x00
 190#endif
 191#define FLAGS_CLOCKRT		0x02
 192#define FLAGS_HAS_TIMEOUT	0x04
 193
 194/*
 195 * Priority Inheritance state:
 196 */
 197struct futex_pi_state {
 198	/*
 199	 * list of 'owned' pi_state instances - these have to be
 200	 * cleaned up in do_exit() if the task exits prematurely:
 201	 */
 202	struct list_head list;
 203
 204	/*
 205	 * The PI object:
 206	 */
 207	struct rt_mutex pi_mutex;
 208
 209	struct task_struct *owner;
 210	atomic_t refcount;
 211
 212	union futex_key key;
 213};
 214
 215/**
 216 * struct futex_q - The hashed futex queue entry, one per waiting task
 217 * @list:		priority-sorted list of tasks waiting on this futex
 218 * @task:		the task waiting on the futex
 219 * @lock_ptr:		the hash bucket lock
 220 * @key:		the key the futex is hashed on
 221 * @pi_state:		optional priority inheritance state
 222 * @rt_waiter:		rt_waiter storage for use with requeue_pi
 223 * @requeue_pi_key:	the requeue_pi target futex key
 224 * @bitset:		bitset for the optional bitmasked wakeup
 225 *
 226 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
 227 * we can wake only the relevant ones (hashed queues may be shared).
 228 *
 229 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
 230 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
 231 * The order of wakeup is always to make the first condition true, then
 232 * the second.
 233 *
 234 * PI futexes are typically woken before they are removed from the hash list via
 235 * the rt_mutex code. See unqueue_me_pi().
 236 */
 237struct futex_q {
 238	struct plist_node list;
 239
 240	struct task_struct *task;
 241	spinlock_t *lock_ptr;
 242	union futex_key key;
 243	struct futex_pi_state *pi_state;
 244	struct rt_mutex_waiter *rt_waiter;
 245	union futex_key *requeue_pi_key;
 246	u32 bitset;
 247};
 248
 249static const struct futex_q futex_q_init = {
 250	/* list gets initialized in queue_me()*/
 251	.key = FUTEX_KEY_INIT,
 252	.bitset = FUTEX_BITSET_MATCH_ANY
 253};
 254
 255/*
 256 * Hash buckets are shared by all the futex_keys that hash to the same
 257 * location.  Each key may have multiple futex_q structures, one for each task
 258 * waiting on a futex.
 259 */
 260struct futex_hash_bucket {
 261	atomic_t waiters;
 262	spinlock_t lock;
 263	struct plist_head chain;
 264} ____cacheline_aligned_in_smp;
 265
 266/*
 267 * The base of the bucket array and its size are always used together
 268 * (after initialization only in hash_futex()), so ensure that they
 269 * reside in the same cacheline.
 270 */
 271static struct {
 272	struct futex_hash_bucket *queues;
 273	unsigned long            hashsize;
 274} __futex_data __read_mostly __aligned(2*sizeof(long));
 275#define futex_queues   (__futex_data.queues)
 276#define futex_hashsize (__futex_data.hashsize)
 277
 278
 279/*
 280 * Fault injections for futexes.
 281 */
 282#ifdef CONFIG_FAIL_FUTEX
 283
 284static struct {
 285	struct fault_attr attr;
 286
 287	bool ignore_private;
 288} fail_futex = {
 289	.attr = FAULT_ATTR_INITIALIZER,
 290	.ignore_private = false,
 291};
 292
 293static int __init setup_fail_futex(char *str)
 294{
 295	return setup_fault_attr(&fail_futex.attr, str);
 296}
 297__setup("fail_futex=", setup_fail_futex);
 298
 299static bool should_fail_futex(bool fshared)
 300{
 301	if (fail_futex.ignore_private && !fshared)
 302		return false;
 303
 304	return should_fail(&fail_futex.attr, 1);
 305}
 306
 307#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 308
 309static int __init fail_futex_debugfs(void)
 310{
 311	umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
 312	struct dentry *dir;
 313
 314	dir = fault_create_debugfs_attr("fail_futex", NULL,
 315					&fail_futex.attr);
 316	if (IS_ERR(dir))
 317		return PTR_ERR(dir);
 318
 319	if (!debugfs_create_bool("ignore-private", mode, dir,
 320				 &fail_futex.ignore_private)) {
 321		debugfs_remove_recursive(dir);
 322		return -ENOMEM;
 323	}
 324
 325	return 0;
 326}
 327
 328late_initcall(fail_futex_debugfs);
 329
 330#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
 331
 332#else
 333static inline bool should_fail_futex(bool fshared)
 334{
 335	return false;
 336}
 337#endif /* CONFIG_FAIL_FUTEX */
 338
 339static inline void futex_get_mm(union futex_key *key)
 340{
 341	atomic_inc(&key->private.mm->mm_count);
 342	/*
 343	 * Ensure futex_get_mm() implies a full barrier such that
 344	 * get_futex_key() implies a full barrier. This is relied upon
 345	 * as smp_mb(); (B), see the ordering comment above.
 346	 */
 347	smp_mb__after_atomic();
 348}
 349
 350/*
 351 * Reflects a new waiter being added to the waitqueue.
 352 */
 353static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
 354{
 355#ifdef CONFIG_SMP
 356	atomic_inc(&hb->waiters);
 357	/*
 358	 * Full barrier (A), see the ordering comment above.
 359	 */
 360	smp_mb__after_atomic();
 361#endif
 362}
 363
 364/*
 365 * Reflects a waiter being removed from the waitqueue by wakeup
 366 * paths.
 367 */
 368static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
 369{
 370#ifdef CONFIG_SMP
 371	atomic_dec(&hb->waiters);
 372#endif
 373}
 374
 375static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
 376{
 377#ifdef CONFIG_SMP
 378	return atomic_read(&hb->waiters);
 379#else
 380	return 1;
 381#endif
 382}
 383
 384/**
 385 * hash_futex - Return the hash bucket in the global hash
 386 * @key:	Pointer to the futex key for which the hash is calculated
 387 *
 388 * We hash on the keys returned from get_futex_key (see below) and return the
 389 * corresponding hash bucket in the global hash.
 390 */
 391static struct futex_hash_bucket *hash_futex(union futex_key *key)
 392{
 393	u32 hash = jhash2((u32*)&key->both.word,
 394			  (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
 395			  key->both.offset);
 396	return &futex_queues[hash & (futex_hashsize - 1)];
 397}
 398
 399
 400/**
 401 * match_futex - Check whether two futex keys are equal
 402 * @key1:	Pointer to key1
 403 * @key2:	Pointer to key2
 404 *
 405 * Return 1 if two futex_keys are equal, 0 otherwise.
 406 */
 407static inline int match_futex(union futex_key *key1, union futex_key *key2)
 408{
 409	return (key1 && key2
 410		&& key1->both.word == key2->both.word
 411		&& key1->both.ptr == key2->both.ptr
 412		&& key1->both.offset == key2->both.offset);
 413}
 414
 415/*
 416 * Take a reference to the resource addressed by a key.
 417 * Can be called while holding spinlocks.
 418 *
 419 */
 420static void get_futex_key_refs(union futex_key *key)
 421{
 422	if (!key->both.ptr)
 423		return;
 424
 425	/*
 426	 * On MMU less systems futexes are always "private" as there is no per
 427	 * process address space. We need the smp wmb nevertheless - yes,
 428	 * arch/blackfin has MMU less SMP ...
 429	 */
 430	if (!IS_ENABLED(CONFIG_MMU)) {
 431		smp_mb(); /* explicit smp_mb(); (B) */
 432		return;
 433	}
 434
 435	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
 436	case FUT_OFF_INODE:
 437		ihold(key->shared.inode); /* implies smp_mb(); (B) */
 438		break;
 439	case FUT_OFF_MMSHARED:
 440		futex_get_mm(key); /* implies smp_mb(); (B) */
 441		break;
 442	default:
 443		/*
 444		 * Private futexes do not hold reference on an inode or
 445		 * mm, therefore the only purpose of calling get_futex_key_refs
 446		 * is because we need the barrier for the lockless waiter check.
 447		 */
 448		smp_mb(); /* explicit smp_mb(); (B) */
 449	}
 450}
 451
 452/*
 453 * Drop a reference to the resource addressed by a key.
 454 * The hash bucket spinlock must not be held. This is
 455 * a no-op for private futexes, see comment in the get
 456 * counterpart.
 457 */
 458static void drop_futex_key_refs(union futex_key *key)
 459{
 460	if (!key->both.ptr) {
 461		/* If we're here then we tried to put a key we failed to get */
 462		WARN_ON_ONCE(1);
 463		return;
 464	}
 465
 466	if (!IS_ENABLED(CONFIG_MMU))
 467		return;
 468
 469	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
 470	case FUT_OFF_INODE:
 471		iput(key->shared.inode);
 472		break;
 473	case FUT_OFF_MMSHARED:
 474		mmdrop(key->private.mm);
 475		break;
 476	}
 477}
 478
 479/**
 480 * get_futex_key() - Get parameters which are the keys for a futex
 481 * @uaddr:	virtual address of the futex
 482 * @fshared:	0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
 483 * @key:	address where result is stored.
 484 * @rw:		mapping needs to be read/write (values: VERIFY_READ,
 485 *              VERIFY_WRITE)
 486 *
 487 * Return: a negative error code or 0
 488 *
 489 * The key words are stored in *key on success.
 490 *
 491 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
 492 * offset_within_page).  For private mappings, it's (uaddr, current->mm).
 493 * We can usually work out the index without swapping in the page.
 494 *
 495 * lock_page() might sleep, the caller should not hold a spinlock.
 496 */
 497static int
 498get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
 499{
 500	unsigned long address = (unsigned long)uaddr;
 501	struct mm_struct *mm = current->mm;
 502	struct page *page, *tail;
 503	struct address_space *mapping;
 504	int err, ro = 0;
 505
 506	/*
 507	 * The futex address must be "naturally" aligned.
 508	 */
 509	key->both.offset = address % PAGE_SIZE;
 510	if (unlikely((address % sizeof(u32)) != 0))
 511		return -EINVAL;
 512	address -= key->both.offset;
 513
 514	if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
 515		return -EFAULT;
 516
 517	if (unlikely(should_fail_futex(fshared)))
 518		return -EFAULT;
 519
 520	/*
 521	 * PROCESS_PRIVATE futexes are fast.
 522	 * As the mm cannot disappear under us and the 'key' only needs
 523	 * virtual address, we dont even have to find the underlying vma.
 524	 * Note : We do have to check 'uaddr' is a valid user address,
 525	 *        but access_ok() should be faster than find_vma()
 526	 */
 527	if (!fshared) {
 528		key->private.mm = mm;
 529		key->private.address = address;
 530		get_futex_key_refs(key);  /* implies smp_mb(); (B) */
 531		return 0;
 532	}
 533
 534again:
 535	/* Ignore any VERIFY_READ mapping (futex common case) */
 536	if (unlikely(should_fail_futex(fshared)))
 537		return -EFAULT;
 538
 539	err = get_user_pages_fast(address, 1, 1, &page);
 540	/*
 541	 * If write access is not required (eg. FUTEX_WAIT), try
 542	 * and get read-only access.
 543	 */
 544	if (err == -EFAULT && rw == VERIFY_READ) {
 545		err = get_user_pages_fast(address, 1, 0, &page);
 546		ro = 1;
 547	}
 548	if (err < 0)
 549		return err;
 550	else
 551		err = 0;
 552
 553	/*
 554	 * The treatment of mapping from this point on is critical. The page
 555	 * lock protects many things but in this context the page lock
 556	 * stabilizes mapping, prevents inode freeing in the shared
 557	 * file-backed region case and guards against movement to swap cache.
 558	 *
 559	 * Strictly speaking the page lock is not needed in all cases being
 560	 * considered here and page lock forces unnecessarily serialization
 561	 * From this point on, mapping will be re-verified if necessary and
 562	 * page lock will be acquired only if it is unavoidable
 563	 *
 564	 * Mapping checks require the head page for any compound page so the
 565	 * head page and mapping is looked up now. For anonymous pages, it
 566	 * does not matter if the page splits in the future as the key is
 567	 * based on the address. For filesystem-backed pages, the tail is
 568	 * required as the index of the page determines the key. For
 569	 * base pages, there is no tail page and tail == page.
 570	 */
 571	tail = page;
 572	page = compound_head(page);
 573	mapping = READ_ONCE(page->mapping);
 574
 575	/*
 576	 * If page->mapping is NULL, then it cannot be a PageAnon
 577	 * page; but it might be the ZERO_PAGE or in the gate area or
 578	 * in a special mapping (all cases which we are happy to fail);
 579	 * or it may have been a good file page when get_user_pages_fast
 580	 * found it, but truncated or holepunched or subjected to
 581	 * invalidate_complete_page2 before we got the page lock (also
 582	 * cases which we are happy to fail).  And we hold a reference,
 583	 * so refcount care in invalidate_complete_page's remove_mapping
 584	 * prevents drop_caches from setting mapping to NULL beneath us.
 585	 *
 586	 * The case we do have to guard against is when memory pressure made
 587	 * shmem_writepage move it from filecache to swapcache beneath us:
 588	 * an unlikely race, but we do need to retry for page->mapping.
 589	 */
 590	if (unlikely(!mapping)) {
 591		int shmem_swizzled;
 592
 593		/*
 594		 * Page lock is required to identify which special case above
 595		 * applies. If this is really a shmem page then the page lock
 596		 * will prevent unexpected transitions.
 597		 */
 598		lock_page(page);
 599		shmem_swizzled = PageSwapCache(page) || page->mapping;
 600		unlock_page(page);
 601		put_page(page);
 602
 603		if (shmem_swizzled)
 604			goto again;
 605
 606		return -EFAULT;
 607	}
 608
 609	/*
 610	 * Private mappings are handled in a simple way.
 611	 *
 612	 * If the futex key is stored on an anonymous page, then the associated
 613	 * object is the mm which is implicitly pinned by the calling process.
 614	 *
 615	 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
 616	 * it's a read-only handle, it's expected that futexes attach to
 617	 * the object not the particular process.
 618	 */
 619	if (PageAnon(page)) {
 620		/*
 621		 * A RO anonymous page will never change and thus doesn't make
 622		 * sense for futex operations.
 623		 */
 624		if (unlikely(should_fail_futex(fshared)) || ro) {
 625			err = -EFAULT;
 626			goto out;
 627		}
 628
 629		key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
 630		key->private.mm = mm;
 631		key->private.address = address;
 632
 633		get_futex_key_refs(key); /* implies smp_mb(); (B) */
 634
 635	} else {
 636		struct inode *inode;
 637
 638		/*
 639		 * The associated futex object in this case is the inode and
 640		 * the page->mapping must be traversed. Ordinarily this should
 641		 * be stabilised under page lock but it's not strictly
 642		 * necessary in this case as we just want to pin the inode, not
 643		 * update the radix tree or anything like that.
 644		 *
 645		 * The RCU read lock is taken as the inode is finally freed
 646		 * under RCU. If the mapping still matches expectations then the
 647		 * mapping->host can be safely accessed as being a valid inode.
 648		 */
 649		rcu_read_lock();
 650
 651		if (READ_ONCE(page->mapping) != mapping) {
 652			rcu_read_unlock();
 653			put_page(page);
 654
 655			goto again;
 656		}
 657
 658		inode = READ_ONCE(mapping->host);
 659		if (!inode) {
 660			rcu_read_unlock();
 661			put_page(page);
 662
 663			goto again;
 664		}
 665
 666		/*
 667		 * Take a reference unless it is about to be freed. Previously
 668		 * this reference was taken by ihold under the page lock
 669		 * pinning the inode in place so i_lock was unnecessary. The
 670		 * only way for this check to fail is if the inode was
 671		 * truncated in parallel so warn for now if this happens.
 672		 *
 673		 * We are not calling into get_futex_key_refs() in file-backed
 674		 * cases, therefore a successful atomic_inc return below will
 675		 * guarantee that get_futex_key() will still imply smp_mb(); (B).
 676		 */
 677		if (WARN_ON_ONCE(!atomic_inc_not_zero(&inode->i_count))) {
 678			rcu_read_unlock();
 679			put_page(page);
 680
 681			goto again;
 682		}
 683
 684		/* Should be impossible but lets be paranoid for now */
 685		if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
 686			err = -EFAULT;
 687			rcu_read_unlock();
 688			iput(inode);
 689
 690			goto out;
 691		}
 692
 693		key->both.offset |= FUT_OFF_INODE; /* inode-based key */
 694		key->shared.inode = inode;
 695		key->shared.pgoff = basepage_index(tail);
 696		rcu_read_unlock();
 697	}
 698
 699out:
 700	put_page(page);
 701	return err;
 702}
 703
 704static inline void put_futex_key(union futex_key *key)
 705{
 706	drop_futex_key_refs(key);
 707}
 708
 709/**
 710 * fault_in_user_writeable() - Fault in user address and verify RW access
 711 * @uaddr:	pointer to faulting user space address
 712 *
 713 * Slow path to fixup the fault we just took in the atomic write
 714 * access to @uaddr.
 715 *
 716 * We have no generic implementation of a non-destructive write to the
 717 * user address. We know that we faulted in the atomic pagefault
 718 * disabled section so we can as well avoid the #PF overhead by
 719 * calling get_user_pages() right away.
 720 */
 721static int fault_in_user_writeable(u32 __user *uaddr)
 722{
 723	struct mm_struct *mm = current->mm;
 724	int ret;
 725
 726	down_read(&mm->mmap_sem);
 727	ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
 728			       FAULT_FLAG_WRITE, NULL);
 729	up_read(&mm->mmap_sem);
 730
 731	return ret < 0 ? ret : 0;
 732}
 733
 734/**
 735 * futex_top_waiter() - Return the highest priority waiter on a futex
 736 * @hb:		the hash bucket the futex_q's reside in
 737 * @key:	the futex key (to distinguish it from other futex futex_q's)
 738 *
 739 * Must be called with the hb lock held.
 740 */
 741static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
 742					union futex_key *key)
 743{
 744	struct futex_q *this;
 745
 746	plist_for_each_entry(this, &hb->chain, list) {
 747		if (match_futex(&this->key, key))
 748			return this;
 749	}
 750	return NULL;
 751}
 752
 753static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
 754				      u32 uval, u32 newval)
 755{
 756	int ret;
 757
 758	pagefault_disable();
 759	ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
 760	pagefault_enable();
 761
 762	return ret;
 763}
 764
 765static int get_futex_value_locked(u32 *dest, u32 __user *from)
 766{
 767	int ret;
 768
 769	pagefault_disable();
 770	ret = __get_user(*dest, from);
 771	pagefault_enable();
 772
 773	return ret ? -EFAULT : 0;
 774}
 775
 776
 777/*
 778 * PI code:
 779 */
 780static int refill_pi_state_cache(void)
 781{
 782	struct futex_pi_state *pi_state;
 783
 784	if (likely(current->pi_state_cache))
 785		return 0;
 786
 787	pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
 788
 789	if (!pi_state)
 790		return -ENOMEM;
 791
 792	INIT_LIST_HEAD(&pi_state->list);
 793	/* pi_mutex gets initialized later */
 794	pi_state->owner = NULL;
 795	atomic_set(&pi_state->refcount, 1);
 796	pi_state->key = FUTEX_KEY_INIT;
 797
 798	current->pi_state_cache = pi_state;
 799
 800	return 0;
 801}
 802
 803static struct futex_pi_state * alloc_pi_state(void)
 804{
 805	struct futex_pi_state *pi_state = current->pi_state_cache;
 806
 807	WARN_ON(!pi_state);
 808	current->pi_state_cache = NULL;
 809
 810	return pi_state;
 811}
 812
 813/*
 814 * Drops a reference to the pi_state object and frees or caches it
 815 * when the last reference is gone.
 816 *
 817 * Must be called with the hb lock held.
 818 */
 819static void put_pi_state(struct futex_pi_state *pi_state)
 820{
 821	if (!pi_state)
 822		return;
 823
 824	if (!atomic_dec_and_test(&pi_state->refcount))
 825		return;
 826
 827	/*
 828	 * If pi_state->owner is NULL, the owner is most probably dying
 829	 * and has cleaned up the pi_state already
 830	 */
 831	if (pi_state->owner) {
 832		raw_spin_lock_irq(&pi_state->owner->pi_lock);
 833		list_del_init(&pi_state->list);
 834		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
 835
 836		rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
 837	}
 838
 839	if (current->pi_state_cache)
 840		kfree(pi_state);
 841	else {
 842		/*
 843		 * pi_state->list is already empty.
 844		 * clear pi_state->owner.
 845		 * refcount is at 0 - put it back to 1.
 846		 */
 847		pi_state->owner = NULL;
 848		atomic_set(&pi_state->refcount, 1);
 849		current->pi_state_cache = pi_state;
 850	}
 851}
 852
 853/*
 854 * Look up the task based on what TID userspace gave us.
 855 * We dont trust it.
 856 */
 857static struct task_struct * futex_find_get_task(pid_t pid)
 858{
 859	struct task_struct *p;
 860
 861	rcu_read_lock();
 862	p = find_task_by_vpid(pid);
 863	if (p)
 864		get_task_struct(p);
 865
 866	rcu_read_unlock();
 867
 868	return p;
 869}
 870
 871/*
 872 * This task is holding PI mutexes at exit time => bad.
 873 * Kernel cleans up PI-state, but userspace is likely hosed.
 874 * (Robust-futex cleanup is separate and might save the day for userspace.)
 875 */
 876void exit_pi_state_list(struct task_struct *curr)
 877{
 878	struct list_head *next, *head = &curr->pi_state_list;
 879	struct futex_pi_state *pi_state;
 880	struct futex_hash_bucket *hb;
 881	union futex_key key = FUTEX_KEY_INIT;
 882
 883	if (!futex_cmpxchg_enabled)
 884		return;
 885	/*
 886	 * We are a ZOMBIE and nobody can enqueue itself on
 887	 * pi_state_list anymore, but we have to be careful
 888	 * versus waiters unqueueing themselves:
 889	 */
 890	raw_spin_lock_irq(&curr->pi_lock);
 891	while (!list_empty(head)) {
 892
 893		next = head->next;
 894		pi_state = list_entry(next, struct futex_pi_state, list);
 895		key = pi_state->key;
 896		hb = hash_futex(&key);
 897		raw_spin_unlock_irq(&curr->pi_lock);
 898
 899		spin_lock(&hb->lock);
 900
 901		raw_spin_lock_irq(&curr->pi_lock);
 902		/*
 903		 * We dropped the pi-lock, so re-check whether this
 904		 * task still owns the PI-state:
 905		 */
 906		if (head->next != next) {
 907			spin_unlock(&hb->lock);
 908			continue;
 909		}
 910
 911		WARN_ON(pi_state->owner != curr);
 912		WARN_ON(list_empty(&pi_state->list));
 913		list_del_init(&pi_state->list);
 914		pi_state->owner = NULL;
 915		raw_spin_unlock_irq(&curr->pi_lock);
 916
 917		rt_mutex_unlock(&pi_state->pi_mutex);
 918
 919		spin_unlock(&hb->lock);
 920
 921		raw_spin_lock_irq(&curr->pi_lock);
 922	}
 923	raw_spin_unlock_irq(&curr->pi_lock);
 924}
 925
 926/*
 927 * We need to check the following states:
 928 *
 929 *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
 930 *
 931 * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
 932 * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
 933 *
 934 * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
 935 *
 936 * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
 937 * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
 938 *
 939 * [6]  Found  | Found    | task      | 0         | 1      | Valid
 940 *
 941 * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
 942 *
 943 * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
 944 * [9]  Found  | Found    | task      | 0         | 0      | Invalid
 945 * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
 946 *
 947 * [1]	Indicates that the kernel can acquire the futex atomically. We
 948 *	came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
 949 *
 950 * [2]	Valid, if TID does not belong to a kernel thread. If no matching
 951 *      thread is found then it indicates that the owner TID has died.
 952 *
 953 * [3]	Invalid. The waiter is queued on a non PI futex
 954 *
 955 * [4]	Valid state after exit_robust_list(), which sets the user space
 956 *	value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
 957 *
 958 * [5]	The user space value got manipulated between exit_robust_list()
 959 *	and exit_pi_state_list()
 960 *
 961 * [6]	Valid state after exit_pi_state_list() which sets the new owner in
 962 *	the pi_state but cannot access the user space value.
 963 *
 964 * [7]	pi_state->owner can only be NULL when the OWNER_DIED bit is set.
 965 *
 966 * [8]	Owner and user space value match
 967 *
 968 * [9]	There is no transient state which sets the user space TID to 0
 969 *	except exit_robust_list(), but this is indicated by the
 970 *	FUTEX_OWNER_DIED bit. See [4]
 971 *
 972 * [10] There is no transient state which leaves owner and user space
 973 *	TID out of sync.
 974 */
 975
 976/*
 977 * Validate that the existing waiter has a pi_state and sanity check
 978 * the pi_state against the user space value. If correct, attach to
 979 * it.
 980 */
 981static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
 982			      struct futex_pi_state **ps)
 983{
 984	pid_t pid = uval & FUTEX_TID_MASK;
 985
 986	/*
 987	 * Userspace might have messed up non-PI and PI futexes [3]
 988	 */
 989	if (unlikely(!pi_state))
 990		return -EINVAL;
 991
 992	WARN_ON(!atomic_read(&pi_state->refcount));
 993
 994	/*
 995	 * Handle the owner died case:
 996	 */
 997	if (uval & FUTEX_OWNER_DIED) {
 998		/*
 999		 * exit_pi_state_list sets owner to NULL and wakes the
1000		 * topmost waiter. The task which acquires the
1001		 * pi_state->rt_mutex will fixup owner.
1002		 */
1003		if (!pi_state->owner) {
1004			/*
1005			 * No pi state owner, but the user space TID
1006			 * is not 0. Inconsistent state. [5]
1007			 */
1008			if (pid)
1009				return -EINVAL;
1010			/*
1011			 * Take a ref on the state and return success. [4]
1012			 */
1013			goto out_state;
1014		}
1015
1016		/*
1017		 * If TID is 0, then either the dying owner has not
1018		 * yet executed exit_pi_state_list() or some waiter
1019		 * acquired the rtmutex in the pi state, but did not
1020		 * yet fixup the TID in user space.
1021		 *
1022		 * Take a ref on the state and return success. [6]
1023		 */
1024		if (!pid)
1025			goto out_state;
1026	} else {
1027		/*
1028		 * If the owner died bit is not set, then the pi_state
1029		 * must have an owner. [7]
1030		 */
1031		if (!pi_state->owner)
1032			return -EINVAL;
1033	}
1034
1035	/*
1036	 * Bail out if user space manipulated the futex value. If pi
1037	 * state exists then the owner TID must be the same as the
1038	 * user space TID. [9/10]
1039	 */
1040	if (pid != task_pid_vnr(pi_state->owner))
1041		return -EINVAL;
1042out_state:
1043	atomic_inc(&pi_state->refcount);
1044	*ps = pi_state;
1045	return 0;
1046}
1047
1048/*
1049 * Lookup the task for the TID provided from user space and attach to
1050 * it after doing proper sanity checks.
1051 */
1052static int attach_to_pi_owner(u32 uval, union futex_key *key,
1053			      struct futex_pi_state **ps)
1054{
1055	pid_t pid = uval & FUTEX_TID_MASK;
1056	struct futex_pi_state *pi_state;
1057	struct task_struct *p;
1058
1059	/*
1060	 * We are the first waiter - try to look up the real owner and attach
1061	 * the new pi_state to it, but bail out when TID = 0 [1]
1062	 */
1063	if (!pid)
1064		return -ESRCH;
1065	p = futex_find_get_task(pid);
1066	if (!p)
1067		return -ESRCH;
1068
1069	if (unlikely(p->flags & PF_KTHREAD)) {
1070		put_task_struct(p);
1071		return -EPERM;
1072	}
1073
1074	/*
1075	 * We need to look at the task state flags to figure out,
1076	 * whether the task is exiting. To protect against the do_exit
1077	 * change of the task flags, we do this protected by
1078	 * p->pi_lock:
1079	 */
1080	raw_spin_lock_irq(&p->pi_lock);
1081	if (unlikely(p->flags & PF_EXITING)) {
1082		/*
1083		 * The task is on the way out. When PF_EXITPIDONE is
1084		 * set, we know that the task has finished the
1085		 * cleanup:
1086		 */
1087		int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1088
1089		raw_spin_unlock_irq(&p->pi_lock);
1090		put_task_struct(p);
1091		return ret;
1092	}
1093
1094	/*
1095	 * No existing pi state. First waiter. [2]
1096	 */
1097	pi_state = alloc_pi_state();
1098
1099	/*
1100	 * Initialize the pi_mutex in locked state and make @p
1101	 * the owner of it:
1102	 */
1103	rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1104
1105	/* Store the key for possible exit cleanups: */
1106	pi_state->key = *key;
1107
1108	WARN_ON(!list_empty(&pi_state->list));
1109	list_add(&pi_state->list, &p->pi_state_list);
1110	pi_state->owner = p;
1111	raw_spin_unlock_irq(&p->pi_lock);
1112
1113	put_task_struct(p);
1114
1115	*ps = pi_state;
1116
1117	return 0;
1118}
1119
1120static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
1121			   union futex_key *key, struct futex_pi_state **ps)
1122{
1123	struct futex_q *match = futex_top_waiter(hb, key);
1124
1125	/*
1126	 * If there is a waiter on that futex, validate it and
1127	 * attach to the pi_state when the validation succeeds.
1128	 */
1129	if (match)
1130		return attach_to_pi_state(uval, match->pi_state, ps);
1131
1132	/*
1133	 * We are the first waiter - try to look up the owner based on
1134	 * @uval and attach to it.
1135	 */
1136	return attach_to_pi_owner(uval, key, ps);
1137}
1138
1139static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1140{
1141	u32 uninitialized_var(curval);
1142
1143	if (unlikely(should_fail_futex(true)))
1144		return -EFAULT;
1145
1146	if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1147		return -EFAULT;
1148
1149	/*If user space value changed, let the caller retry */
1150	return curval != uval ? -EAGAIN : 0;
1151}
1152
1153/**
1154 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1155 * @uaddr:		the pi futex user address
1156 * @hb:			the pi futex hash bucket
1157 * @key:		the futex key associated with uaddr and hb
1158 * @ps:			the pi_state pointer where we store the result of the
1159 *			lookup
1160 * @task:		the task to perform the atomic lock work for.  This will
1161 *			be "current" except in the case of requeue pi.
1162 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1163 *
1164 * Return:
1165 *  0 - ready to wait;
1166 *  1 - acquired the lock;
1167 * <0 - error
1168 *
1169 * The hb->lock and futex_key refs shall be held by the caller.
1170 */
1171static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1172				union futex_key *key,
1173				struct futex_pi_state **ps,
1174				struct task_struct *task, int set_waiters)
1175{
1176	u32 uval, newval, vpid = task_pid_vnr(task);
1177	struct futex_q *match;
1178	int ret;
1179
1180	/*
1181	 * Read the user space value first so we can validate a few
1182	 * things before proceeding further.
1183	 */
1184	if (get_futex_value_locked(&uval, uaddr))
1185		return -EFAULT;
1186
1187	if (unlikely(should_fail_futex(true)))
1188		return -EFAULT;
1189
1190	/*
1191	 * Detect deadlocks.
1192	 */
1193	if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1194		return -EDEADLK;
1195
1196	if ((unlikely(should_fail_futex(true))))
1197		return -EDEADLK;
1198
1199	/*
1200	 * Lookup existing state first. If it exists, try to attach to
1201	 * its pi_state.
1202	 */
1203	match = futex_top_waiter(hb, key);
1204	if (match)
1205		return attach_to_pi_state(uval, match->pi_state, ps);
1206
1207	/*
1208	 * No waiter and user TID is 0. We are here because the
1209	 * waiters or the owner died bit is set or called from
1210	 * requeue_cmp_pi or for whatever reason something took the
1211	 * syscall.
1212	 */
1213	if (!(uval & FUTEX_TID_MASK)) {
1214		/*
1215		 * We take over the futex. No other waiters and the user space
1216		 * TID is 0. We preserve the owner died bit.
1217		 */
1218		newval = uval & FUTEX_OWNER_DIED;
1219		newval |= vpid;
1220
1221		/* The futex requeue_pi code can enforce the waiters bit */
1222		if (set_waiters)
1223			newval |= FUTEX_WAITERS;
1224
1225		ret = lock_pi_update_atomic(uaddr, uval, newval);
1226		/* If the take over worked, return 1 */
1227		return ret < 0 ? ret : 1;
1228	}
1229
1230	/*
1231	 * First waiter. Set the waiters bit before attaching ourself to
1232	 * the owner. If owner tries to unlock, it will be forced into
1233	 * the kernel and blocked on hb->lock.
1234	 */
1235	newval = uval | FUTEX_WAITERS;
1236	ret = lock_pi_update_atomic(uaddr, uval, newval);
1237	if (ret)
1238		return ret;
1239	/*
1240	 * If the update of the user space value succeeded, we try to
1241	 * attach to the owner. If that fails, no harm done, we only
1242	 * set the FUTEX_WAITERS bit in the user space variable.
1243	 */
1244	return attach_to_pi_owner(uval, key, ps);
1245}
1246
1247/**
1248 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1249 * @q:	The futex_q to unqueue
1250 *
1251 * The q->lock_ptr must not be NULL and must be held by the caller.
1252 */
1253static void __unqueue_futex(struct futex_q *q)
1254{
1255	struct futex_hash_bucket *hb;
1256
1257	if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1258	    || WARN_ON(plist_node_empty(&q->list)))
1259		return;
1260
1261	hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1262	plist_del(&q->list, &hb->chain);
1263	hb_waiters_dec(hb);
1264}
1265
1266/*
1267 * The hash bucket lock must be held when this is called.
1268 * Afterwards, the futex_q must not be accessed. Callers
1269 * must ensure to later call wake_up_q() for the actual
1270 * wakeups to occur.
1271 */
1272static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1273{
1274	struct task_struct *p = q->task;
1275
1276	if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1277		return;
1278
1279	/*
1280	 * Queue the task for later wakeup for after we've released
1281	 * the hb->lock. wake_q_add() grabs reference to p.
1282	 */
1283	wake_q_add(wake_q, p);
1284	__unqueue_futex(q);
1285	/*
1286	 * The waiting task can free the futex_q as soon as
1287	 * q->lock_ptr = NULL is written, without taking any locks. A
1288	 * memory barrier is required here to prevent the following
1289	 * store to lock_ptr from getting ahead of the plist_del.
1290	 */
1291	smp_wmb();
1292	q->lock_ptr = NULL;
1293}
1294
1295static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1296			 struct futex_hash_bucket *hb)
1297{
1298	struct task_struct *new_owner;
1299	struct futex_pi_state *pi_state = this->pi_state;
1300	u32 uninitialized_var(curval), newval;
1301	DEFINE_WAKE_Q(wake_q);
1302	bool deboost;
1303	int ret = 0;
1304
1305	if (!pi_state)
1306		return -EINVAL;
1307
1308	/*
1309	 * If current does not own the pi_state then the futex is
1310	 * inconsistent and user space fiddled with the futex value.
1311	 */
1312	if (pi_state->owner != current)
1313		return -EINVAL;
1314
1315	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1316	new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1317
1318	/*
1319	 * It is possible that the next waiter (the one that brought
1320	 * this owner to the kernel) timed out and is no longer
1321	 * waiting on the lock.
1322	 */
1323	if (!new_owner)
1324		new_owner = this->task;
1325
1326	/*
1327	 * We pass it to the next owner. The WAITERS bit is always
1328	 * kept enabled while there is PI state around. We cleanup the
1329	 * owner died bit, because we are the owner.
1330	 */
1331	newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1332
1333	if (unlikely(should_fail_futex(true)))
1334		ret = -EFAULT;
1335
1336	if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1337		ret = -EFAULT;
1338	} else if (curval != uval) {
1339		/*
1340		 * If a unconditional UNLOCK_PI operation (user space did not
1341		 * try the TID->0 transition) raced with a waiter setting the
1342		 * FUTEX_WAITERS flag between get_user() and locking the hash
1343		 * bucket lock, retry the operation.
1344		 */
1345		if ((FUTEX_TID_MASK & curval) == uval)
1346			ret = -EAGAIN;
1347		else
1348			ret = -EINVAL;
1349	}
1350	if (ret) {
1351		raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1352		return ret;
1353	}
1354
1355	raw_spin_lock(&pi_state->owner->pi_lock);
1356	WARN_ON(list_empty(&pi_state->list));
1357	list_del_init(&pi_state->list);
1358	raw_spin_unlock(&pi_state->owner->pi_lock);
1359
1360	raw_spin_lock(&new_owner->pi_lock);
1361	WARN_ON(!list_empty(&pi_state->list));
1362	list_add(&pi_state->list, &new_owner->pi_state_list);
1363	pi_state->owner = new_owner;
1364	raw_spin_unlock(&new_owner->pi_lock);
1365
1366	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1367
1368	deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1369
1370	/*
1371	 * First unlock HB so the waiter does not spin on it once he got woken
1372	 * up. Second wake up the waiter before the priority is adjusted. If we
1373	 * deboost first (and lose our higher priority), then the task might get
1374	 * scheduled away before the wake up can take place.
1375	 */
1376	spin_unlock(&hb->lock);
1377	wake_up_q(&wake_q);
1378	if (deboost)
1379		rt_mutex_adjust_prio(current);
1380
1381	return 0;
1382}
1383
1384/*
1385 * Express the locking dependencies for lockdep:
1386 */
1387static inline void
1388double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1389{
1390	if (hb1 <= hb2) {
1391		spin_lock(&hb1->lock);
1392		if (hb1 < hb2)
1393			spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1394	} else { /* hb1 > hb2 */
1395		spin_lock(&hb2->lock);
1396		spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1397	}
1398}
1399
1400static inline void
1401double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1402{
1403	spin_unlock(&hb1->lock);
1404	if (hb1 != hb2)
1405		spin_unlock(&hb2->lock);
1406}
1407
1408/*
1409 * Wake up waiters matching bitset queued on this futex (uaddr).
1410 */
1411static int
1412futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1413{
1414	struct futex_hash_bucket *hb;
1415	struct futex_q *this, *next;
1416	union futex_key key = FUTEX_KEY_INIT;
1417	int ret;
1418	DEFINE_WAKE_Q(wake_q);
1419
1420	if (!bitset)
1421		return -EINVAL;
1422
1423	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1424	if (unlikely(ret != 0))
1425		goto out;
1426
1427	hb = hash_futex(&key);
1428
1429	/* Make sure we really have tasks to wakeup */
1430	if (!hb_waiters_pending(hb))
1431		goto out_put_key;
1432
1433	spin_lock(&hb->lock);
1434
1435	plist_for_each_entry_safe(this, next, &hb->chain, list) {
1436		if (match_futex (&this->key, &key)) {
1437			if (this->pi_state || this->rt_waiter) {
1438				ret = -EINVAL;
1439				break;
1440			}
1441
1442			/* Check if one of the bits is set in both bitsets */
1443			if (!(this->bitset & bitset))
1444				continue;
1445
1446			mark_wake_futex(&wake_q, this);
1447			if (++ret >= nr_wake)
1448				break;
1449		}
1450	}
1451
1452	spin_unlock(&hb->lock);
1453	wake_up_q(&wake_q);
1454out_put_key:
1455	put_futex_key(&key);
1456out:
1457	return ret;
1458}
1459
1460/*
1461 * Wake up all waiters hashed on the physical page that is mapped
1462 * to this virtual address:
1463 */
1464static int
1465futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1466	      int nr_wake, int nr_wake2, int op)
1467{
1468	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1469	struct futex_hash_bucket *hb1, *hb2;
1470	struct futex_q *this, *next;
1471	int ret, op_ret;
1472	DEFINE_WAKE_Q(wake_q);
1473
1474retry:
1475	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1476	if (unlikely(ret != 0))
1477		goto out;
1478	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1479	if (unlikely(ret != 0))
1480		goto out_put_key1;
1481
1482	hb1 = hash_futex(&key1);
1483	hb2 = hash_futex(&key2);
1484
1485retry_private:
1486	double_lock_hb(hb1, hb2);
1487	op_ret = futex_atomic_op_inuser(op, uaddr2);
1488	if (unlikely(op_ret < 0)) {
1489
1490		double_unlock_hb(hb1, hb2);
1491
1492#ifndef CONFIG_MMU
1493		/*
1494		 * we don't get EFAULT from MMU faults if we don't have an MMU,
1495		 * but we might get them from range checking
1496		 */
1497		ret = op_ret;
1498		goto out_put_keys;
1499#endif
1500
1501		if (unlikely(op_ret != -EFAULT)) {
1502			ret = op_ret;
1503			goto out_put_keys;
1504		}
1505
1506		ret = fault_in_user_writeable(uaddr2);
1507		if (ret)
1508			goto out_put_keys;
1509
1510		if (!(flags & FLAGS_SHARED))
1511			goto retry_private;
1512
1513		put_futex_key(&key2);
1514		put_futex_key(&key1);
1515		goto retry;
1516	}
1517
1518	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1519		if (match_futex (&this->key, &key1)) {
1520			if (this->pi_state || this->rt_waiter) {
1521				ret = -EINVAL;
1522				goto out_unlock;
1523			}
1524			mark_wake_futex(&wake_q, this);
1525			if (++ret >= nr_wake)
1526				break;
1527		}
1528	}
1529
1530	if (op_ret > 0) {
1531		op_ret = 0;
1532		plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1533			if (match_futex (&this->key, &key2)) {
1534				if (this->pi_state || this->rt_waiter) {
1535					ret = -EINVAL;
1536					goto out_unlock;
1537				}
1538				mark_wake_futex(&wake_q, this);
1539				if (++op_ret >= nr_wake2)
1540					break;
1541			}
1542		}
1543		ret += op_ret;
1544	}
1545
1546out_unlock:
1547	double_unlock_hb(hb1, hb2);
1548	wake_up_q(&wake_q);
1549out_put_keys:
1550	put_futex_key(&key2);
1551out_put_key1:
1552	put_futex_key(&key1);
1553out:
1554	return ret;
1555}
1556
1557/**
1558 * requeue_futex() - Requeue a futex_q from one hb to another
1559 * @q:		the futex_q to requeue
1560 * @hb1:	the source hash_bucket
1561 * @hb2:	the target hash_bucket
1562 * @key2:	the new key for the requeued futex_q
1563 */
1564static inline
1565void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1566		   struct futex_hash_bucket *hb2, union futex_key *key2)
1567{
1568
1569	/*
1570	 * If key1 and key2 hash to the same bucket, no need to
1571	 * requeue.
1572	 */
1573	if (likely(&hb1->chain != &hb2->chain)) {
1574		plist_del(&q->list, &hb1->chain);
1575		hb_waiters_dec(hb1);
1576		hb_waiters_inc(hb2);
1577		plist_add(&q->list, &hb2->chain);
1578		q->lock_ptr = &hb2->lock;
1579	}
1580	get_futex_key_refs(key2);
1581	q->key = *key2;
1582}
1583
1584/**
1585 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1586 * @q:		the futex_q
1587 * @key:	the key of the requeue target futex
1588 * @hb:		the hash_bucket of the requeue target futex
1589 *
1590 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1591 * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1592 * to the requeue target futex so the waiter can detect the wakeup on the right
1593 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1594 * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1595 * to protect access to the pi_state to fixup the owner later.  Must be called
1596 * with both q->lock_ptr and hb->lock held.
1597 */
1598static inline
1599void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1600			   struct futex_hash_bucket *hb)
1601{
1602	get_futex_key_refs(key);
1603	q->key = *key;
1604
1605	__unqueue_futex(q);
1606
1607	WARN_ON(!q->rt_waiter);
1608	q->rt_waiter = NULL;
1609
1610	q->lock_ptr = &hb->lock;
1611
1612	wake_up_state(q->task, TASK_NORMAL);
1613}
1614
1615/**
1616 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1617 * @pifutex:		the user address of the to futex
1618 * @hb1:		the from futex hash bucket, must be locked by the caller
1619 * @hb2:		the to futex hash bucket, must be locked by the caller
1620 * @key1:		the from futex key
1621 * @key2:		the to futex key
1622 * @ps:			address to store the pi_state pointer
1623 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1624 *
1625 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1626 * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1627 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1628 * hb1 and hb2 must be held by the caller.
1629 *
1630 * Return:
1631 *  0 - failed to acquire the lock atomically;
1632 * >0 - acquired the lock, return value is vpid of the top_waiter
1633 * <0 - error
1634 */
1635static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1636				 struct futex_hash_bucket *hb1,
1637				 struct futex_hash_bucket *hb2,
1638				 union futex_key *key1, union futex_key *key2,
1639				 struct futex_pi_state **ps, int set_waiters)
1640{
1641	struct futex_q *top_waiter = NULL;
1642	u32 curval;
1643	int ret, vpid;
1644
1645	if (get_futex_value_locked(&curval, pifutex))
1646		return -EFAULT;
1647
1648	if (unlikely(should_fail_futex(true)))
1649		return -EFAULT;
1650
1651	/*
1652	 * Find the top_waiter and determine if there are additional waiters.
1653	 * If the caller intends to requeue more than 1 waiter to pifutex,
1654	 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1655	 * as we have means to handle the possible fault.  If not, don't set
1656	 * the bit unecessarily as it will force the subsequent unlock to enter
1657	 * the kernel.
1658	 */
1659	top_waiter = futex_top_waiter(hb1, key1);
1660
1661	/* There are no waiters, nothing for us to do. */
1662	if (!top_waiter)
1663		return 0;
1664
1665	/* Ensure we requeue to the expected futex. */
1666	if (!match_futex(top_waiter->requeue_pi_key, key2))
1667		return -EINVAL;
1668
1669	/*
1670	 * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1671	 * the contended case or if set_waiters is 1.  The pi_state is returned
1672	 * in ps in contended cases.
1673	 */
1674	vpid = task_pid_vnr(top_waiter->task);
1675	ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1676				   set_waiters);
1677	if (ret == 1) {
1678		requeue_pi_wake_futex(top_waiter, key2, hb2);
1679		return vpid;
1680	}
1681	return ret;
1682}
1683
1684/**
1685 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1686 * @uaddr1:	source futex user address
1687 * @flags:	futex flags (FLAGS_SHARED, etc.)
1688 * @uaddr2:	target futex user address
1689 * @nr_wake:	number of waiters to wake (must be 1 for requeue_pi)
1690 * @nr_requeue:	number of waiters to requeue (0-INT_MAX)
1691 * @cmpval:	@uaddr1 expected value (or %NULL)
1692 * @requeue_pi:	if we are attempting to requeue from a non-pi futex to a
1693 *		pi futex (pi to pi requeue is not supported)
1694 *
1695 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1696 * uaddr2 atomically on behalf of the top waiter.
1697 *
1698 * Return:
1699 * >=0 - on success, the number of tasks requeued or woken;
1700 *  <0 - on error
1701 */
1702static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1703			 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1704			 u32 *cmpval, int requeue_pi)
1705{
1706	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1707	int drop_count = 0, task_count = 0, ret;
1708	struct futex_pi_state *pi_state = NULL;
1709	struct futex_hash_bucket *hb1, *hb2;
1710	struct futex_q *this, *next;
1711	DEFINE_WAKE_Q(wake_q);
1712
1713	if (requeue_pi) {
1714		/*
1715		 * Requeue PI only works on two distinct uaddrs. This
1716		 * check is only valid for private futexes. See below.
1717		 */
1718		if (uaddr1 == uaddr2)
1719			return -EINVAL;
1720
1721		/*
1722		 * requeue_pi requires a pi_state, try to allocate it now
1723		 * without any locks in case it fails.
1724		 */
1725		if (refill_pi_state_cache())
1726			return -ENOMEM;
1727		/*
1728		 * requeue_pi must wake as many tasks as it can, up to nr_wake
1729		 * + nr_requeue, since it acquires the rt_mutex prior to
1730		 * returning to userspace, so as to not leave the rt_mutex with
1731		 * waiters and no owner.  However, second and third wake-ups
1732		 * cannot be predicted as they involve race conditions with the
1733		 * first wake and a fault while looking up the pi_state.  Both
1734		 * pthread_cond_signal() and pthread_cond_broadcast() should
1735		 * use nr_wake=1.
1736		 */
1737		if (nr_wake != 1)
1738			return -EINVAL;
1739	}
1740
1741retry:
1742	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1743	if (unlikely(ret != 0))
1744		goto out;
1745	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1746			    requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1747	if (unlikely(ret != 0))
1748		goto out_put_key1;
1749
1750	/*
1751	 * The check above which compares uaddrs is not sufficient for
1752	 * shared futexes. We need to compare the keys:
1753	 */
1754	if (requeue_pi && match_futex(&key1, &key2)) {
1755		ret = -EINVAL;
1756		goto out_put_keys;
1757	}
1758
1759	hb1 = hash_futex(&key1);
1760	hb2 = hash_futex(&key2);
1761
1762retry_private:
1763	hb_waiters_inc(hb2);
1764	double_lock_hb(hb1, hb2);
1765
1766	if (likely(cmpval != NULL)) {
1767		u32 curval;
1768
1769		ret = get_futex_value_locked(&curval, uaddr1);
1770
1771		if (unlikely(ret)) {
1772			double_unlock_hb(hb1, hb2);
1773			hb_waiters_dec(hb2);
1774
1775			ret = get_user(curval, uaddr1);
1776			if (ret)
1777				goto out_put_keys;
1778
1779			if (!(flags & FLAGS_SHARED))
1780				goto retry_private;
1781
1782			put_futex_key(&key2);
1783			put_futex_key(&key1);
1784			goto retry;
1785		}
1786		if (curval != *cmpval) {
1787			ret = -EAGAIN;
1788			goto out_unlock;
1789		}
1790	}
1791
1792	if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1793		/*
1794		 * Attempt to acquire uaddr2 and wake the top waiter. If we
1795		 * intend to requeue waiters, force setting the FUTEX_WAITERS
1796		 * bit.  We force this here where we are able to easily handle
1797		 * faults rather in the requeue loop below.
1798		 */
1799		ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1800						 &key2, &pi_state, nr_requeue);
1801
1802		/*
1803		 * At this point the top_waiter has either taken uaddr2 or is
1804		 * waiting on it.  If the former, then the pi_state will not
1805		 * exist yet, look it up one more time to ensure we have a
1806		 * reference to it. If the lock was taken, ret contains the
1807		 * vpid of the top waiter task.
1808		 * If the lock was not taken, we have pi_state and an initial
1809		 * refcount on it. In case of an error we have nothing.
1810		 */
1811		if (ret > 0) {
1812			WARN_ON(pi_state);
1813			drop_count++;
1814			task_count++;
1815			/*
1816			 * If we acquired the lock, then the user space value
1817			 * of uaddr2 should be vpid. It cannot be changed by
1818			 * the top waiter as it is blocked on hb2 lock if it
1819			 * tries to do so. If something fiddled with it behind
1820			 * our back the pi state lookup might unearth it. So
1821			 * we rather use the known value than rereading and
1822			 * handing potential crap to lookup_pi_state.
1823			 *
1824			 * If that call succeeds then we have pi_state and an
1825			 * initial refcount on it.
1826			 */
1827			ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1828		}
1829
1830		switch (ret) {
1831		case 0:
1832			/* We hold a reference on the pi state. */
1833			break;
1834
1835			/* If the above failed, then pi_state is NULL */
1836		case -EFAULT:
1837			double_unlock_hb(hb1, hb2);
1838			hb_waiters_dec(hb2);
1839			put_futex_key(&key2);
1840			put_futex_key(&key1);
1841			ret = fault_in_user_writeable(uaddr2);
1842			if (!ret)
1843				goto retry;
1844			goto out;
1845		case -EAGAIN:
1846			/*
1847			 * Two reasons for this:
1848			 * - Owner is exiting and we just wait for the
1849			 *   exit to complete.
1850			 * - The user space value changed.
1851			 */
1852			double_unlock_hb(hb1, hb2);
1853			hb_waiters_dec(hb2);
1854			put_futex_key(&key2);
1855			put_futex_key(&key1);
1856			cond_resched();
1857			goto retry;
1858		default:
1859			goto out_unlock;
1860		}
1861	}
1862
1863	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1864		if (task_count - nr_wake >= nr_requeue)
1865			break;
1866
1867		if (!match_futex(&this->key, &key1))
1868			continue;
1869
1870		/*
1871		 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1872		 * be paired with each other and no other futex ops.
1873		 *
1874		 * We should never be requeueing a futex_q with a pi_state,
1875		 * which is awaiting a futex_unlock_pi().
1876		 */
1877		if ((requeue_pi && !this->rt_waiter) ||
1878		    (!requeue_pi && this->rt_waiter) ||
1879		    this->pi_state) {
1880			ret = -EINVAL;
1881			break;
1882		}
1883
1884		/*
1885		 * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1886		 * lock, we already woke the top_waiter.  If not, it will be
1887		 * woken by futex_unlock_pi().
1888		 */
1889		if (++task_count <= nr_wake && !requeue_pi) {
1890			mark_wake_futex(&wake_q, this);
1891			continue;
1892		}
1893
1894		/* Ensure we requeue to the expected futex for requeue_pi. */
1895		if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1896			ret = -EINVAL;
1897			break;
1898		}
1899
1900		/*
1901		 * Requeue nr_requeue waiters and possibly one more in the case
1902		 * of requeue_pi if we couldn't acquire the lock atomically.
1903		 */
1904		if (requeue_pi) {
1905			/*
1906			 * Prepare the waiter to take the rt_mutex. Take a
1907			 * refcount on the pi_state and store the pointer in
1908			 * the futex_q object of the waiter.
1909			 */
1910			atomic_inc(&pi_state->refcount);
1911			this->pi_state = pi_state;
1912			ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1913							this->rt_waiter,
1914							this->task);
1915			if (ret == 1) {
1916				/*
1917				 * We got the lock. We do neither drop the
1918				 * refcount on pi_state nor clear
1919				 * this->pi_state because the waiter needs the
1920				 * pi_state for cleaning up the user space
1921				 * value. It will drop the refcount after
1922				 * doing so.
1923				 */
1924				requeue_pi_wake_futex(this, &key2, hb2);
1925				drop_count++;
1926				continue;
1927			} else if (ret) {
1928				/*
1929				 * rt_mutex_start_proxy_lock() detected a
1930				 * potential deadlock when we tried to queue
1931				 * that waiter. Drop the pi_state reference
1932				 * which we took above and remove the pointer
1933				 * to the state from the waiters futex_q
1934				 * object.
1935				 */
1936				this->pi_state = NULL;
1937				put_pi_state(pi_state);
1938				/*
1939				 * We stop queueing more waiters and let user
1940				 * space deal with the mess.
1941				 */
1942				break;
1943			}
1944		}
1945		requeue_futex(this, hb1, hb2, &key2);
1946		drop_count++;
1947	}
1948
1949	/*
1950	 * We took an extra initial reference to the pi_state either
1951	 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
1952	 * need to drop it here again.
1953	 */
1954	put_pi_state(pi_state);
1955
1956out_unlock:
1957	double_unlock_hb(hb1, hb2);
1958	wake_up_q(&wake_q);
1959	hb_waiters_dec(hb2);
1960
1961	/*
1962	 * drop_futex_key_refs() must be called outside the spinlocks. During
1963	 * the requeue we moved futex_q's from the hash bucket at key1 to the
1964	 * one at key2 and updated their key pointer.  We no longer need to
1965	 * hold the references to key1.
1966	 */
1967	while (--drop_count >= 0)
1968		drop_futex_key_refs(&key1);
1969
1970out_put_keys:
1971	put_futex_key(&key2);
1972out_put_key1:
1973	put_futex_key(&key1);
1974out:
1975	return ret ? ret : task_count;
1976}
1977
1978/* The key must be already stored in q->key. */
1979static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1980	__acquires(&hb->lock)
1981{
1982	struct futex_hash_bucket *hb;
1983
1984	hb = hash_futex(&q->key);
1985
1986	/*
1987	 * Increment the counter before taking the lock so that
1988	 * a potential waker won't miss a to-be-slept task that is
1989	 * waiting for the spinlock. This is safe as all queue_lock()
1990	 * users end up calling queue_me(). Similarly, for housekeeping,
1991	 * decrement the counter at queue_unlock() when some error has
1992	 * occurred and we don't end up adding the task to the list.
1993	 */
1994	hb_waiters_inc(hb);
1995
1996	q->lock_ptr = &hb->lock;
1997
1998	spin_lock(&hb->lock); /* implies smp_mb(); (A) */
1999	return hb;
2000}
2001
2002static inline void
2003queue_unlock(struct futex_hash_bucket *hb)
2004	__releases(&hb->lock)
2005{
2006	spin_unlock(&hb->lock);
2007	hb_waiters_dec(hb);
2008}
2009
2010/**
2011 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2012 * @q:	The futex_q to enqueue
2013 * @hb:	The destination hash bucket
2014 *
2015 * The hb->lock must be held by the caller, and is released here. A call to
2016 * queue_me() is typically paired with exactly one call to unqueue_me().  The
2017 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2018 * or nothing if the unqueue is done as part of the wake process and the unqueue
2019 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2020 * an example).
2021 */
2022static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2023	__releases(&hb->lock)
2024{
2025	int prio;
2026
2027	/*
2028	 * The priority used to register this element is
2029	 * - either the real thread-priority for the real-time threads
2030	 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2031	 * - or MAX_RT_PRIO for non-RT threads.
2032	 * Thus, all RT-threads are woken first in priority order, and
2033	 * the others are woken last, in FIFO order.
2034	 */
2035	prio = min(current->normal_prio, MAX_RT_PRIO);
2036
2037	plist_node_init(&q->list, prio);
2038	plist_add(&q->list, &hb->chain);
2039	q->task = current;
2040	spin_unlock(&hb->lock);
2041}
2042
2043/**
2044 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2045 * @q:	The futex_q to unqueue
2046 *
2047 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2048 * be paired with exactly one earlier call to queue_me().
2049 *
2050 * Return:
2051 *   1 - if the futex_q was still queued (and we removed unqueued it);
2052 *   0 - if the futex_q was already removed by the waking thread
2053 */
2054static int unqueue_me(struct futex_q *q)
2055{
2056	spinlock_t *lock_ptr;
2057	int ret = 0;
2058
2059	/* In the common case we don't take the spinlock, which is nice. */
2060retry:
2061	/*
2062	 * q->lock_ptr can change between this read and the following spin_lock.
2063	 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2064	 * optimizing lock_ptr out of the logic below.
2065	 */
2066	lock_ptr = READ_ONCE(q->lock_ptr);
2067	if (lock_ptr != NULL) {
2068		spin_lock(lock_ptr);
2069		/*
2070		 * q->lock_ptr can change between reading it and
2071		 * spin_lock(), causing us to take the wrong lock.  This
2072		 * corrects the race condition.
2073		 *
2074		 * Reasoning goes like this: if we have the wrong lock,
2075		 * q->lock_ptr must have changed (maybe several times)
2076		 * between reading it and the spin_lock().  It can
2077		 * change again after the spin_lock() but only if it was
2078		 * already changed before the spin_lock().  It cannot,
2079		 * however, change back to the original value.  Therefore
2080		 * we can detect whether we acquired the correct lock.
2081		 */
2082		if (unlikely(lock_ptr != q->lock_ptr)) {
2083			spin_unlock(lock_ptr);
2084			goto retry;
2085		}
2086		__unqueue_futex(q);
2087
2088		BUG_ON(q->pi_state);
2089
2090		spin_unlock(lock_ptr);
2091		ret = 1;
2092	}
2093
2094	drop_futex_key_refs(&q->key);
2095	return ret;
2096}
2097
2098/*
2099 * PI futexes can not be requeued and must remove themself from the
2100 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2101 * and dropped here.
2102 */
2103static void unqueue_me_pi(struct futex_q *q)
2104	__releases(q->lock_ptr)
2105{
2106	__unqueue_futex(q);
2107
2108	BUG_ON(!q->pi_state);
2109	put_pi_state(q->pi_state);
2110	q->pi_state = NULL;
2111
2112	spin_unlock(q->lock_ptr);
2113}
2114
2115/*
2116 * Fixup the pi_state owner with the new owner.
2117 *
2118 * Must be called with hash bucket lock held and mm->sem held for non
2119 * private futexes.
2120 */
2121static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2122				struct task_struct *newowner)
2123{
2124	u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2125	struct futex_pi_state *pi_state = q->pi_state;
2126	struct task_struct *oldowner = pi_state->owner;
2127	u32 uval, uninitialized_var(curval), newval;
2128	int ret;
2129
2130	/* Owner died? */
2131	if (!pi_state->owner)
2132		newtid |= FUTEX_OWNER_DIED;
2133
2134	/*
2135	 * We are here either because we stole the rtmutex from the
2136	 * previous highest priority waiter or we are the highest priority
2137	 * waiter but failed to get the rtmutex the first time.
2138	 * We have to replace the newowner TID in the user space variable.
2139	 * This must be atomic as we have to preserve the owner died bit here.
2140	 *
2141	 * Note: We write the user space value _before_ changing the pi_state
2142	 * because we can fault here. Imagine swapped out pages or a fork
2143	 * that marked all the anonymous memory readonly for cow.
2144	 *
2145	 * Modifying pi_state _before_ the user space value would
2146	 * leave the pi_state in an inconsistent state when we fault
2147	 * here, because we need to drop the hash bucket lock to
2148	 * handle the fault. This might be observed in the PID check
2149	 * in lookup_pi_state.
2150	 */
2151retry:
2152	if (get_futex_value_locked(&uval, uaddr))
2153		goto handle_fault;
2154
2155	while (1) {
2156		newval = (uval & FUTEX_OWNER_DIED) | newtid;
2157
2158		if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2159			goto handle_fault;
2160		if (curval == uval)
2161			break;
2162		uval = curval;
2163	}
2164
2165	/*
2166	 * We fixed up user space. Now we need to fix the pi_state
2167	 * itself.
2168	 */
2169	if (pi_state->owner != NULL) {
2170		raw_spin_lock_irq(&pi_state->owner->pi_lock);
2171		WARN_ON(list_empty(&pi_state->list));
2172		list_del_init(&pi_state->list);
2173		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2174	}
2175
2176	pi_state->owner = newowner;
2177
2178	raw_spin_lock_irq(&newowner->pi_lock);
2179	WARN_ON(!list_empty(&pi_state->list));
2180	list_add(&pi_state->list, &newowner->pi_state_list);
2181	raw_spin_unlock_irq(&newowner->pi_lock);
2182	return 0;
2183
2184	/*
2185	 * To handle the page fault we need to drop the hash bucket
2186	 * lock here. That gives the other task (either the highest priority
2187	 * waiter itself or the task which stole the rtmutex) the
2188	 * chance to try the fixup of the pi_state. So once we are
2189	 * back from handling the fault we need to check the pi_state
2190	 * after reacquiring the hash bucket lock and before trying to
2191	 * do another fixup. When the fixup has been done already we
2192	 * simply return.
2193	 */
2194handle_fault:
2195	spin_unlock(q->lock_ptr);
2196
2197	ret = fault_in_user_writeable(uaddr);
2198
2199	spin_lock(q->lock_ptr);
2200
2201	/*
2202	 * Check if someone else fixed it for us:
2203	 */
2204	if (pi_state->owner != oldowner)
2205		return 0;
2206
2207	if (ret)
2208		return ret;
2209
2210	goto retry;
2211}
2212
2213static long futex_wait_restart(struct restart_block *restart);
2214
2215/**
2216 * fixup_owner() - Post lock pi_state and corner case management
2217 * @uaddr:	user address of the futex
2218 * @q:		futex_q (contains pi_state and access to the rt_mutex)
2219 * @locked:	if the attempt to take the rt_mutex succeeded (1) or not (0)
2220 *
2221 * After attempting to lock an rt_mutex, this function is called to cleanup
2222 * the pi_state owner as well as handle race conditions that may allow us to
2223 * acquire the lock. Must be called with the hb lock held.
2224 *
2225 * Return:
2226 *  1 - success, lock taken;
2227 *  0 - success, lock not taken;
2228 * <0 - on error (-EFAULT)
2229 */
2230static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2231{
2232	struct task_struct *owner;
2233	int ret = 0;
2234
2235	if (locked) {
2236		/*
2237		 * Got the lock. We might not be the anticipated owner if we
2238		 * did a lock-steal - fix up the PI-state in that case:
2239		 */
2240		if (q->pi_state->owner != current)
2241			ret = fixup_pi_state_owner(uaddr, q, current);
2242		goto out;
2243	}
2244
2245	/*
2246	 * Catch the rare case, where the lock was released when we were on the
2247	 * way back before we locked the hash bucket.
2248	 */
2249	if (q->pi_state->owner == current) {
2250		/*
2251		 * Try to get the rt_mutex now. This might fail as some other
2252		 * task acquired the rt_mutex after we removed ourself from the
2253		 * rt_mutex waiters list.
2254		 */
2255		if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2256			locked = 1;
2257			goto out;
2258		}
2259
2260		/*
2261		 * pi_state is incorrect, some other task did a lock steal and
2262		 * we returned due to timeout or signal without taking the
2263		 * rt_mutex. Too late.
2264		 */
2265		raw_spin_lock_irq(&q->pi_state->pi_mutex.wait_lock);
2266		owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2267		if (!owner)
2268			owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2269		raw_spin_unlock_irq(&q->pi_state->pi_mutex.wait_lock);
2270		ret = fixup_pi_state_owner(uaddr, q, owner);
2271		goto out;
2272	}
2273
2274	/*
2275	 * Paranoia check. If we did not take the lock, then we should not be
2276	 * the owner of the rt_mutex.
2277	 */
2278	if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2279		printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2280				"pi-state %p\n", ret,
2281				q->pi_state->pi_mutex.owner,
2282				q->pi_state->owner);
2283
2284out:
2285	return ret ? ret : locked;
2286}
2287
2288/**
2289 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2290 * @hb:		the futex hash bucket, must be locked by the caller
2291 * @q:		the futex_q to queue up on
2292 * @timeout:	the prepared hrtimer_sleeper, or null for no timeout
2293 */
2294static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2295				struct hrtimer_sleeper *timeout)
2296{
2297	/*
2298	 * The task state is guaranteed to be set before another task can
2299	 * wake it. set_current_state() is implemented using smp_store_mb() and
2300	 * queue_me() calls spin_unlock() upon completion, both serializing
2301	 * access to the hash list and forcing another memory barrier.
2302	 */
2303	set_current_state(TASK_INTERRUPTIBLE);
2304	queue_me(q, hb);
2305
2306	/* Arm the timer */
2307	if (timeout)
2308		hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2309
2310	/*
2311	 * If we have been removed from the hash list, then another task
2312	 * has tried to wake us, and we can skip the call to schedule().
2313	 */
2314	if (likely(!plist_node_empty(&q->list))) {
2315		/*
2316		 * If the timer has already expired, current will already be
2317		 * flagged for rescheduling. Only call schedule if there
2318		 * is no timeout, or if it has yet to expire.
2319		 */
2320		if (!timeout || timeout->task)
2321			freezable_schedule();
2322	}
2323	__set_current_state(TASK_RUNNING);
2324}
2325
2326/**
2327 * futex_wait_setup() - Prepare to wait on a futex
2328 * @uaddr:	the futex userspace address
2329 * @val:	the expected value
2330 * @flags:	futex flags (FLAGS_SHARED, etc.)
2331 * @q:		the associated futex_q
2332 * @hb:		storage for hash_bucket pointer to be returned to caller
2333 *
2334 * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2335 * compare it with the expected value.  Handle atomic faults internally.
2336 * Return with the hb lock held and a q.key reference on success, and unlocked
2337 * with no q.key reference on failure.
2338 *
2339 * Return:
2340 *  0 - uaddr contains val and hb has been locked;
2341 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2342 */
2343static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2344			   struct futex_q *q, struct futex_hash_bucket **hb)
2345{
2346	u32 uval;
2347	int ret;
2348
2349	/*
2350	 * Access the page AFTER the hash-bucket is locked.
2351	 * Order is important:
2352	 *
2353	 *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2354	 *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2355	 *
2356	 * The basic logical guarantee of a futex is that it blocks ONLY
2357	 * if cond(var) is known to be true at the time of blocking, for
2358	 * any cond.  If we locked the hash-bucket after testing *uaddr, that
2359	 * would open a race condition where we could block indefinitely with
2360	 * cond(var) false, which would violate the guarantee.
2361	 *
2362	 * On the other hand, we insert q and release the hash-bucket only
2363	 * after testing *uaddr.  This guarantees that futex_wait() will NOT
2364	 * absorb a wakeup if *uaddr does not match the desired values
2365	 * while the syscall executes.
2366	 */
2367retry:
2368	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2369	if (unlikely(ret != 0))
2370		return ret;
2371
2372retry_private:
2373	*hb = queue_lock(q);
2374
2375	ret = get_futex_value_locked(&uval, uaddr);
2376
2377	if (ret) {
2378		queue_unlock(*hb);
2379
2380		ret = get_user(uval, uaddr);
2381		if (ret)
2382			goto out;
2383
2384		if (!(flags & FLAGS_SHARED))
2385			goto retry_private;
2386
2387		put_futex_key(&q->key);
2388		goto retry;
2389	}
2390
2391	if (uval != val) {
2392		queue_unlock(*hb);
2393		ret = -EWOULDBLOCK;
2394	}
2395
2396out:
2397	if (ret)
2398		put_futex_key(&q->key);
2399	return ret;
2400}
2401
2402static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2403		      ktime_t *abs_time, u32 bitset)
2404{
2405	struct hrtimer_sleeper timeout, *to = NULL;
2406	struct restart_block *restart;
2407	struct futex_hash_bucket *hb;
2408	struct futex_q q = futex_q_init;
2409	int ret;
2410
2411	if (!bitset)
2412		return -EINVAL;
2413	q.bitset = bitset;
2414
2415	if (abs_time) {
2416		to = &timeout;
2417
2418		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2419				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2420				      HRTIMER_MODE_ABS);
2421		hrtimer_init_sleeper(to, current);
2422		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2423					     current->timer_slack_ns);
2424	}
2425
2426retry:
2427	/*
2428	 * Prepare to wait on uaddr. On success, holds hb lock and increments
2429	 * q.key refs.
2430	 */
2431	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2432	if (ret)
2433		goto out;
2434
2435	/* queue_me and wait for wakeup, timeout, or a signal. */
2436	futex_wait_queue_me(hb, &q, to);
2437
2438	/* If we were woken (and unqueued), we succeeded, whatever. */
2439	ret = 0;
2440	/* unqueue_me() drops q.key ref */
2441	if (!unqueue_me(&q))
2442		goto out;
2443	ret = -ETIMEDOUT;
2444	if (to && !to->task)
2445		goto out;
2446
2447	/*
2448	 * We expect signal_pending(current), but we might be the
2449	 * victim of a spurious wakeup as well.
2450	 */
2451	if (!signal_pending(current))
2452		goto retry;
2453
2454	ret = -ERESTARTSYS;
2455	if (!abs_time)
2456		goto out;
2457
2458	restart = &current->restart_block;
2459	restart->fn = futex_wait_restart;
2460	restart->futex.uaddr = uaddr;
2461	restart->futex.val = val;
2462	restart->futex.time = *abs_time;
2463	restart->futex.bitset = bitset;
2464	restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2465
2466	ret = -ERESTART_RESTARTBLOCK;
2467
2468out:
2469	if (to) {
2470		hrtimer_cancel(&to->timer);
2471		destroy_hrtimer_on_stack(&to->timer);
2472	}
2473	return ret;
2474}
2475
2476
2477static long futex_wait_restart(struct restart_block *restart)
2478{
2479	u32 __user *uaddr = restart->futex.uaddr;
2480	ktime_t t, *tp = NULL;
2481
2482	if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2483		t = restart->futex.time;
2484		tp = &t;
2485	}
2486	restart->fn = do_no_restart_syscall;
2487
2488	return (long)futex_wait(uaddr, restart->futex.flags,
2489				restart->futex.val, tp, restart->futex.bitset);
2490}
2491
2492
2493/*
2494 * Userspace tried a 0 -> TID atomic transition of the futex value
2495 * and failed. The kernel side here does the whole locking operation:
2496 * if there are waiters then it will block as a consequence of relying
2497 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2498 * a 0 value of the futex too.).
2499 *
2500 * Also serves as futex trylock_pi()'ing, and due semantics.
2501 */
2502static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2503			 ktime_t *time, int trylock)
2504{
2505	struct hrtimer_sleeper timeout, *to = NULL;
2506	struct futex_hash_bucket *hb;
2507	struct futex_q q = futex_q_init;
2508	int res, ret;
2509
2510	if (refill_pi_state_cache())
2511		return -ENOMEM;
2512
2513	if (time) {
2514		to = &timeout;
2515		hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2516				      HRTIMER_MODE_ABS);
2517		hrtimer_init_sleeper(to, current);
2518		hrtimer_set_expires(&to->timer, *time);
2519	}
2520
2521retry:
2522	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2523	if (unlikely(ret != 0))
2524		goto out;
2525
2526retry_private:
2527	hb = queue_lock(&q);
2528
2529	ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2530	if (unlikely(ret)) {
2531		/*
2532		 * Atomic work succeeded and we got the lock,
2533		 * or failed. Either way, we do _not_ block.
2534		 */
2535		switch (ret) {
2536		case 1:
2537			/* We got the lock. */
2538			ret = 0;
2539			goto out_unlock_put_key;
2540		case -EFAULT:
2541			goto uaddr_faulted;
2542		case -EAGAIN:
2543			/*
2544			 * Two reasons for this:
2545			 * - Task is exiting and we just wait for the
2546			 *   exit to complete.
2547			 * - The user space value changed.
2548			 */
2549			queue_unlock(hb);
2550			put_futex_key(&q.key);
2551			cond_resched();
2552			goto retry;
2553		default:
2554			goto out_unlock_put_key;
2555		}
2556	}
2557
2558	/*
2559	 * Only actually queue now that the atomic ops are done:
2560	 */
2561	queue_me(&q, hb);
2562
2563	WARN_ON(!q.pi_state);
2564	/*
2565	 * Block on the PI mutex:
2566	 */
2567	if (!trylock) {
2568		ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2569	} else {
2570		ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2571		/* Fixup the trylock return value: */
2572		ret = ret ? 0 : -EWOULDBLOCK;
2573	}
2574
2575	spin_lock(q.lock_ptr);
2576	/*
2577	 * Fixup the pi_state owner and possibly acquire the lock if we
2578	 * haven't already.
2579	 */
2580	res = fixup_owner(uaddr, &q, !ret);
2581	/*
2582	 * If fixup_owner() returned an error, proprogate that.  If it acquired
2583	 * the lock, clear our -ETIMEDOUT or -EINTR.
2584	 */
2585	if (res)
2586		ret = (res < 0) ? res : 0;
2587
2588	/*
2589	 * If fixup_owner() faulted and was unable to handle the fault, unlock
2590	 * it and return the fault to userspace.
2591	 */
2592	if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2593		rt_mutex_unlock(&q.pi_state->pi_mutex);
2594
2595	/* Unqueue and drop the lock */
2596	unqueue_me_pi(&q);
2597
2598	goto out_put_key;
2599
2600out_unlock_put_key:
2601	queue_unlock(hb);
2602
2603out_put_key:
2604	put_futex_key(&q.key);
2605out:
2606	if (to)
2607		destroy_hrtimer_on_stack(&to->timer);
2608	return ret != -EINTR ? ret : -ERESTARTNOINTR;
2609
2610uaddr_faulted:
2611	queue_unlock(hb);
2612
2613	ret = fault_in_user_writeable(uaddr);
2614	if (ret)
2615		goto out_put_key;
2616
2617	if (!(flags & FLAGS_SHARED))
2618		goto retry_private;
2619
2620	put_futex_key(&q.key);
2621	goto retry;
2622}
2623
2624/*
2625 * Userspace attempted a TID -> 0 atomic transition, and failed.
2626 * This is the in-kernel slowpath: we look up the PI state (if any),
2627 * and do the rt-mutex unlock.
2628 */
2629static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2630{
2631	u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2632	union futex_key key = FUTEX_KEY_INIT;
2633	struct futex_hash_bucket *hb;
2634	struct futex_q *match;
2635	int ret;
2636
2637retry:
2638	if (get_user(uval, uaddr))
2639		return -EFAULT;
2640	/*
2641	 * We release only a lock we actually own:
2642	 */
2643	if ((uval & FUTEX_TID_MASK) != vpid)
2644		return -EPERM;
2645
2646	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2647	if (ret)
2648		return ret;
2649
2650	hb = hash_futex(&key);
2651	spin_lock(&hb->lock);
2652
2653	/*
2654	 * Check waiters first. We do not trust user space values at
2655	 * all and we at least want to know if user space fiddled
2656	 * with the futex value instead of blindly unlocking.
2657	 */
2658	match = futex_top_waiter(hb, &key);
2659	if (match) {
2660		ret = wake_futex_pi(uaddr, uval, match, hb);
2661		/*
2662		 * In case of success wake_futex_pi dropped the hash
2663		 * bucket lock.
2664		 */
2665		if (!ret)
2666			goto out_putkey;
2667		/*
2668		 * The atomic access to the futex value generated a
2669		 * pagefault, so retry the user-access and the wakeup:
2670		 */
2671		if (ret == -EFAULT)
2672			goto pi_faulted;
2673		/*
2674		 * A unconditional UNLOCK_PI op raced against a waiter
2675		 * setting the FUTEX_WAITERS bit. Try again.
2676		 */
2677		if (ret == -EAGAIN) {
2678			spin_unlock(&hb->lock);
2679			put_futex_key(&key);
2680			goto retry;
2681		}
2682		/*
2683		 * wake_futex_pi has detected invalid state. Tell user
2684		 * space.
2685		 */
2686		goto out_unlock;
2687	}
2688
2689	/*
2690	 * We have no kernel internal state, i.e. no waiters in the
2691	 * kernel. Waiters which are about to queue themselves are stuck
2692	 * on hb->lock. So we can safely ignore them. We do neither
2693	 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2694	 * owner.
2695	 */
2696	if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2697		goto pi_faulted;
2698
2699	/*
2700	 * If uval has changed, let user space handle it.
2701	 */
2702	ret = (curval == uval) ? 0 : -EAGAIN;
2703
2704out_unlock:
2705	spin_unlock(&hb->lock);
2706out_putkey:
2707	put_futex_key(&key);
2708	return ret;
2709
2710pi_faulted:
2711	spin_unlock(&hb->lock);
2712	put_futex_key(&key);
2713
2714	ret = fault_in_user_writeable(uaddr);
2715	if (!ret)
2716		goto retry;
2717
2718	return ret;
2719}
2720
2721/**
2722 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2723 * @hb:		the hash_bucket futex_q was original enqueued on
2724 * @q:		the futex_q woken while waiting to be requeued
2725 * @key2:	the futex_key of the requeue target futex
2726 * @timeout:	the timeout associated with the wait (NULL if none)
2727 *
2728 * Detect if the task was woken on the initial futex as opposed to the requeue
2729 * target futex.  If so, determine if it was a timeout or a signal that caused
2730 * the wakeup and return the appropriate error code to the caller.  Must be
2731 * called with the hb lock held.
2732 *
2733 * Return:
2734 *  0 = no early wakeup detected;
2735 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2736 */
2737static inline
2738int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2739				   struct futex_q *q, union futex_key *key2,
2740				   struct hrtimer_sleeper *timeout)
2741{
2742	int ret = 0;
2743
2744	/*
2745	 * With the hb lock held, we avoid races while we process the wakeup.
2746	 * We only need to hold hb (and not hb2) to ensure atomicity as the
2747	 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2748	 * It can't be requeued from uaddr2 to something else since we don't
2749	 * support a PI aware source futex for requeue.
2750	 */
2751	if (!match_futex(&q->key, key2)) {
2752		WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2753		/*
2754		 * We were woken prior to requeue by a timeout or a signal.
2755		 * Unqueue the futex_q and determine which it was.
2756		 */
2757		plist_del(&q->list, &hb->chain);
2758		hb_waiters_dec(hb);
2759
2760		/* Handle spurious wakeups gracefully */
2761		ret = -EWOULDBLOCK;
2762		if (timeout && !timeout->task)
2763			ret = -ETIMEDOUT;
2764		else if (signal_pending(current))
2765			ret = -ERESTARTNOINTR;
2766	}
2767	return ret;
2768}
2769
2770/**
2771 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2772 * @uaddr:	the futex we initially wait on (non-pi)
2773 * @flags:	futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2774 *		the same type, no requeueing from private to shared, etc.
2775 * @val:	the expected value of uaddr
2776 * @abs_time:	absolute timeout
2777 * @bitset:	32 bit wakeup bitset set by userspace, defaults to all
2778 * @uaddr2:	the pi futex we will take prior to returning to user-space
2779 *
2780 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2781 * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2782 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2783 * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2784 * without one, the pi logic would not know which task to boost/deboost, if
2785 * there was a need to.
2786 *
2787 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2788 * via the following--
2789 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2790 * 2) wakeup on uaddr2 after a requeue
2791 * 3) signal
2792 * 4) timeout
2793 *
2794 * If 3, cleanup and return -ERESTARTNOINTR.
2795 *
2796 * If 2, we may then block on trying to take the rt_mutex and return via:
2797 * 5) successful lock
2798 * 6) signal
2799 * 7) timeout
2800 * 8) other lock acquisition failure
2801 *
2802 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2803 *
2804 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2805 *
2806 * Return:
2807 *  0 - On success;
2808 * <0 - On error
2809 */
2810static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2811				 u32 val, ktime_t *abs_time, u32 bitset,
2812				 u32 __user *uaddr2)
2813{
2814	struct hrtimer_sleeper timeout, *to = NULL;
2815	struct rt_mutex_waiter rt_waiter;
 
2816	struct futex_hash_bucket *hb;
2817	union futex_key key2 = FUTEX_KEY_INIT;
2818	struct futex_q q = futex_q_init;
2819	int res, ret;
2820
2821	if (uaddr == uaddr2)
2822		return -EINVAL;
2823
2824	if (!bitset)
2825		return -EINVAL;
2826
2827	if (abs_time) {
2828		to = &timeout;
2829		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2830				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2831				      HRTIMER_MODE_ABS);
2832		hrtimer_init_sleeper(to, current);
2833		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2834					     current->timer_slack_ns);
2835	}
2836
2837	/*
2838	 * The waiter is allocated on our stack, manipulated by the requeue
2839	 * code while we sleep on uaddr.
2840	 */
2841	debug_rt_mutex_init_waiter(&rt_waiter);
2842	RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2843	RB_CLEAR_NODE(&rt_waiter.tree_entry);
2844	rt_waiter.task = NULL;
2845
2846	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2847	if (unlikely(ret != 0))
2848		goto out;
2849
2850	q.bitset = bitset;
2851	q.rt_waiter = &rt_waiter;
2852	q.requeue_pi_key = &key2;
2853
2854	/*
2855	 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2856	 * count.
2857	 */
2858	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2859	if (ret)
2860		goto out_key2;
2861
2862	/*
2863	 * The check above which compares uaddrs is not sufficient for
2864	 * shared futexes. We need to compare the keys:
2865	 */
2866	if (match_futex(&q.key, &key2)) {
2867		queue_unlock(hb);
2868		ret = -EINVAL;
2869		goto out_put_keys;
2870	}
2871
2872	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
2873	futex_wait_queue_me(hb, &q, to);
2874
2875	spin_lock(&hb->lock);
2876	ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2877	spin_unlock(&hb->lock);
2878	if (ret)
2879		goto out_put_keys;
2880
2881	/*
2882	 * In order for us to be here, we know our q.key == key2, and since
2883	 * we took the hb->lock above, we also know that futex_requeue() has
2884	 * completed and we no longer have to concern ourselves with a wakeup
2885	 * race with the atomic proxy lock acquisition by the requeue code. The
2886	 * futex_requeue dropped our key1 reference and incremented our key2
2887	 * reference count.
2888	 */
2889
2890	/* Check if the requeue code acquired the second futex for us. */
2891	if (!q.rt_waiter) {
2892		/*
2893		 * Got the lock. We might not be the anticipated owner if we
2894		 * did a lock-steal - fix up the PI-state in that case.
2895		 */
2896		if (q.pi_state && (q.pi_state->owner != current)) {
2897			spin_lock(q.lock_ptr);
2898			ret = fixup_pi_state_owner(uaddr2, &q, current);
2899			if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current)
2900				rt_mutex_unlock(&q.pi_state->pi_mutex);
2901			/*
2902			 * Drop the reference to the pi state which
2903			 * the requeue_pi() code acquired for us.
2904			 */
2905			put_pi_state(q.pi_state);
2906			spin_unlock(q.lock_ptr);
2907		}
2908	} else {
2909		struct rt_mutex *pi_mutex;
2910
2911		/*
2912		 * We have been woken up by futex_unlock_pi(), a timeout, or a
2913		 * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2914		 * the pi_state.
2915		 */
2916		WARN_ON(!q.pi_state);
2917		pi_mutex = &q.pi_state->pi_mutex;
2918		ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2919		debug_rt_mutex_free_waiter(&rt_waiter);
2920
2921		spin_lock(q.lock_ptr);
2922		/*
2923		 * Fixup the pi_state owner and possibly acquire the lock if we
2924		 * haven't already.
2925		 */
2926		res = fixup_owner(uaddr2, &q, !ret);
2927		/*
2928		 * If fixup_owner() returned an error, proprogate that.  If it
2929		 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2930		 */
2931		if (res)
2932			ret = (res < 0) ? res : 0;
2933
2934		/*
2935		 * If fixup_pi_state_owner() faulted and was unable to handle
2936		 * the fault, unlock the rt_mutex and return the fault to
2937		 * userspace.
2938		 */
2939		if (ret && rt_mutex_owner(pi_mutex) == current)
2940			rt_mutex_unlock(pi_mutex);
2941
2942		/* Unqueue and drop the lock. */
2943		unqueue_me_pi(&q);
2944	}
2945
2946	if (ret == -EINTR) {
 
 
 
 
 
 
 
2947		/*
2948		 * We've already been requeued, but cannot restart by calling
2949		 * futex_lock_pi() directly. We could restart this syscall, but
2950		 * it would detect that the user space "val" changed and return
2951		 * -EWOULDBLOCK.  Save the overhead of the restart and return
2952		 * -EWOULDBLOCK directly.
2953		 */
2954		ret = -EWOULDBLOCK;
2955	}
2956
2957out_put_keys:
2958	put_futex_key(&q.key);
2959out_key2:
2960	put_futex_key(&key2);
2961
2962out:
2963	if (to) {
2964		hrtimer_cancel(&to->timer);
2965		destroy_hrtimer_on_stack(&to->timer);
2966	}
2967	return ret;
2968}
2969
2970/*
2971 * Support for robust futexes: the kernel cleans up held futexes at
2972 * thread exit time.
2973 *
2974 * Implementation: user-space maintains a per-thread list of locks it
2975 * is holding. Upon do_exit(), the kernel carefully walks this list,
2976 * and marks all locks that are owned by this thread with the
2977 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2978 * always manipulated with the lock held, so the list is private and
2979 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2980 * field, to allow the kernel to clean up if the thread dies after
2981 * acquiring the lock, but just before it could have added itself to
2982 * the list. There can only be one such pending lock.
2983 */
2984
2985/**
2986 * sys_set_robust_list() - Set the robust-futex list head of a task
2987 * @head:	pointer to the list-head
2988 * @len:	length of the list-head, as userspace expects
2989 */
2990SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2991		size_t, len)
2992{
2993	if (!futex_cmpxchg_enabled)
2994		return -ENOSYS;
2995	/*
2996	 * The kernel knows only one size for now:
2997	 */
2998	if (unlikely(len != sizeof(*head)))
2999		return -EINVAL;
3000
3001	current->robust_list = head;
3002
3003	return 0;
3004}
3005
3006/**
3007 * sys_get_robust_list() - Get the robust-futex list head of a task
3008 * @pid:	pid of the process [zero for current task]
3009 * @head_ptr:	pointer to a list-head pointer, the kernel fills it in
3010 * @len_ptr:	pointer to a length field, the kernel fills in the header size
3011 */
3012SYSCALL_DEFINE3(get_robust_list, int, pid,
3013		struct robust_list_head __user * __user *, head_ptr,
3014		size_t __user *, len_ptr)
3015{
3016	struct robust_list_head __user *head;
3017	unsigned long ret;
3018	struct task_struct *p;
3019
3020	if (!futex_cmpxchg_enabled)
3021		return -ENOSYS;
3022
3023	rcu_read_lock();
3024
3025	ret = -ESRCH;
3026	if (!pid)
3027		p = current;
3028	else {
3029		p = find_task_by_vpid(pid);
3030		if (!p)
3031			goto err_unlock;
3032	}
3033
3034	ret = -EPERM;
3035	if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3036		goto err_unlock;
3037
3038	head = p->robust_list;
3039	rcu_read_unlock();
3040
3041	if (put_user(sizeof(*head), len_ptr))
3042		return -EFAULT;
3043	return put_user(head, head_ptr);
3044
3045err_unlock:
3046	rcu_read_unlock();
3047
3048	return ret;
3049}
3050
3051/*
3052 * Process a futex-list entry, check whether it's owned by the
3053 * dying task, and do notification if so:
3054 */
3055int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3056{
3057	u32 uval, uninitialized_var(nval), mval;
3058
3059retry:
3060	if (get_user(uval, uaddr))
3061		return -1;
3062
3063	if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3064		/*
3065		 * Ok, this dying thread is truly holding a futex
3066		 * of interest. Set the OWNER_DIED bit atomically
3067		 * via cmpxchg, and if the value had FUTEX_WAITERS
3068		 * set, wake up a waiter (if any). (We have to do a
3069		 * futex_wake() even if OWNER_DIED is already set -
3070		 * to handle the rare but possible case of recursive
3071		 * thread-death.) The rest of the cleanup is done in
3072		 * userspace.
3073		 */
3074		mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3075		/*
3076		 * We are not holding a lock here, but we want to have
3077		 * the pagefault_disable/enable() protection because
3078		 * we want to handle the fault gracefully. If the
3079		 * access fails we try to fault in the futex with R/W
3080		 * verification via get_user_pages. get_user() above
3081		 * does not guarantee R/W access. If that fails we
3082		 * give up and leave the futex locked.
3083		 */
3084		if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3085			if (fault_in_user_writeable(uaddr))
3086				return -1;
3087			goto retry;
3088		}
3089		if (nval != uval)
3090			goto retry;
3091
3092		/*
3093		 * Wake robust non-PI futexes here. The wakeup of
3094		 * PI futexes happens in exit_pi_state():
3095		 */
3096		if (!pi && (uval & FUTEX_WAITERS))
3097			futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3098	}
3099	return 0;
3100}
3101
3102/*
3103 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3104 */
3105static inline int fetch_robust_entry(struct robust_list __user **entry,
3106				     struct robust_list __user * __user *head,
3107				     unsigned int *pi)
3108{
3109	unsigned long uentry;
3110
3111	if (get_user(uentry, (unsigned long __user *)head))
3112		return -EFAULT;
3113
3114	*entry = (void __user *)(uentry & ~1UL);
3115	*pi = uentry & 1;
3116
3117	return 0;
3118}
3119
3120/*
3121 * Walk curr->robust_list (very carefully, it's a userspace list!)
3122 * and mark any locks found there dead, and notify any waiters.
3123 *
3124 * We silently return on any sign of list-walking problem.
3125 */
3126void exit_robust_list(struct task_struct *curr)
3127{
3128	struct robust_list_head __user *head = curr->robust_list;
3129	struct robust_list __user *entry, *next_entry, *pending;
3130	unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3131	unsigned int uninitialized_var(next_pi);
3132	unsigned long futex_offset;
3133	int rc;
3134
3135	if (!futex_cmpxchg_enabled)
3136		return;
3137
3138	/*
3139	 * Fetch the list head (which was registered earlier, via
3140	 * sys_set_robust_list()):
3141	 */
3142	if (fetch_robust_entry(&entry, &head->list.next, &pi))
3143		return;
3144	/*
3145	 * Fetch the relative futex offset:
3146	 */
3147	if (get_user(futex_offset, &head->futex_offset))
3148		return;
3149	/*
3150	 * Fetch any possibly pending lock-add first, and handle it
3151	 * if it exists:
3152	 */
3153	if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3154		return;
3155
3156	next_entry = NULL;	/* avoid warning with gcc */
3157	while (entry != &head->list) {
3158		/*
3159		 * Fetch the next entry in the list before calling
3160		 * handle_futex_death:
3161		 */
3162		rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3163		/*
3164		 * A pending lock might already be on the list, so
3165		 * don't process it twice:
3166		 */
3167		if (entry != pending)
3168			if (handle_futex_death((void __user *)entry + futex_offset,
3169						curr, pi))
3170				return;
3171		if (rc)
3172			return;
3173		entry = next_entry;
3174		pi = next_pi;
3175		/*
3176		 * Avoid excessively long or circular lists:
3177		 */
3178		if (!--limit)
3179			break;
3180
3181		cond_resched();
3182	}
3183
3184	if (pending)
3185		handle_futex_death((void __user *)pending + futex_offset,
3186				   curr, pip);
3187}
3188
3189long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3190		u32 __user *uaddr2, u32 val2, u32 val3)
3191{
3192	int cmd = op & FUTEX_CMD_MASK;
3193	unsigned int flags = 0;
3194
3195	if (!(op & FUTEX_PRIVATE_FLAG))
3196		flags |= FLAGS_SHARED;
3197
3198	if (op & FUTEX_CLOCK_REALTIME) {
3199		flags |= FLAGS_CLOCKRT;
3200		if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3201		    cmd != FUTEX_WAIT_REQUEUE_PI)
3202			return -ENOSYS;
3203	}
3204
3205	switch (cmd) {
3206	case FUTEX_LOCK_PI:
3207	case FUTEX_UNLOCK_PI:
3208	case FUTEX_TRYLOCK_PI:
3209	case FUTEX_WAIT_REQUEUE_PI:
3210	case FUTEX_CMP_REQUEUE_PI:
3211		if (!futex_cmpxchg_enabled)
3212			return -ENOSYS;
3213	}
3214
3215	switch (cmd) {
3216	case FUTEX_WAIT:
3217		val3 = FUTEX_BITSET_MATCH_ANY;
3218	case FUTEX_WAIT_BITSET:
3219		return futex_wait(uaddr, flags, val, timeout, val3);
3220	case FUTEX_WAKE:
3221		val3 = FUTEX_BITSET_MATCH_ANY;
3222	case FUTEX_WAKE_BITSET:
3223		return futex_wake(uaddr, flags, val, val3);
3224	case FUTEX_REQUEUE:
3225		return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3226	case FUTEX_CMP_REQUEUE:
3227		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3228	case FUTEX_WAKE_OP:
3229		return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3230	case FUTEX_LOCK_PI:
3231		return futex_lock_pi(uaddr, flags, timeout, 0);
3232	case FUTEX_UNLOCK_PI:
3233		return futex_unlock_pi(uaddr, flags);
3234	case FUTEX_TRYLOCK_PI:
3235		return futex_lock_pi(uaddr, flags, NULL, 1);
3236	case FUTEX_WAIT_REQUEUE_PI:
3237		val3 = FUTEX_BITSET_MATCH_ANY;
3238		return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3239					     uaddr2);
3240	case FUTEX_CMP_REQUEUE_PI:
3241		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3242	}
3243	return -ENOSYS;
3244}
3245
3246
3247SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3248		struct timespec __user *, utime, u32 __user *, uaddr2,
3249		u32, val3)
3250{
3251	struct timespec ts;
3252	ktime_t t, *tp = NULL;
3253	u32 val2 = 0;
3254	int cmd = op & FUTEX_CMD_MASK;
3255
3256	if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3257		      cmd == FUTEX_WAIT_BITSET ||
3258		      cmd == FUTEX_WAIT_REQUEUE_PI)) {
3259		if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3260			return -EFAULT;
3261		if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3262			return -EFAULT;
3263		if (!timespec_valid(&ts))
3264			return -EINVAL;
3265
3266		t = timespec_to_ktime(ts);
3267		if (cmd == FUTEX_WAIT)
3268			t = ktime_add_safe(ktime_get(), t);
3269		tp = &t;
3270	}
3271	/*
3272	 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3273	 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3274	 */
3275	if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3276	    cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3277		val2 = (u32) (unsigned long) utime;
3278
3279	return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3280}
3281
3282static void __init futex_detect_cmpxchg(void)
3283{
3284#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3285	u32 curval;
3286
3287	/*
3288	 * This will fail and we want it. Some arch implementations do
3289	 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3290	 * functionality. We want to know that before we call in any
3291	 * of the complex code paths. Also we want to prevent
3292	 * registration of robust lists in that case. NULL is
3293	 * guaranteed to fault and we get -EFAULT on functional
3294	 * implementation, the non-functional ones will return
3295	 * -ENOSYS.
3296	 */
3297	if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3298		futex_cmpxchg_enabled = 1;
3299#endif
3300}
3301
3302static int __init futex_init(void)
3303{
3304	unsigned int futex_shift;
3305	unsigned long i;
3306
3307#if CONFIG_BASE_SMALL
3308	futex_hashsize = 16;
3309#else
3310	futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3311#endif
3312
3313	futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3314					       futex_hashsize, 0,
3315					       futex_hashsize < 256 ? HASH_SMALL : 0,
3316					       &futex_shift, NULL,
3317					       futex_hashsize, futex_hashsize);
3318	futex_hashsize = 1UL << futex_shift;
3319
3320	futex_detect_cmpxchg();
3321
3322	for (i = 0; i < futex_hashsize; i++) {
3323		atomic_set(&futex_queues[i].waiters, 0);
3324		plist_head_init(&futex_queues[i].chain);
3325		spin_lock_init(&futex_queues[i].lock);
3326	}
3327
3328	return 0;
3329}
3330core_initcall(futex_init);
v4.6
   1/*
   2 *  Fast Userspace Mutexes (which I call "Futexes!").
   3 *  (C) Rusty Russell, IBM 2002
   4 *
   5 *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
   6 *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
   7 *
   8 *  Removed page pinning, fix privately mapped COW pages and other cleanups
   9 *  (C) Copyright 2003, 2004 Jamie Lokier
  10 *
  11 *  Robust futex support started by Ingo Molnar
  12 *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
  13 *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
  14 *
  15 *  PI-futex support started by Ingo Molnar and Thomas Gleixner
  16 *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  17 *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
  18 *
  19 *  PRIVATE futexes by Eric Dumazet
  20 *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
  21 *
  22 *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
  23 *  Copyright (C) IBM Corporation, 2009
  24 *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
  25 *
  26 *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
  27 *  enough at me, Linus for the original (flawed) idea, Matthew
  28 *  Kirkwood for proof-of-concept implementation.
  29 *
  30 *  "The futexes are also cursed."
  31 *  "But they come in a choice of three flavours!"
  32 *
  33 *  This program is free software; you can redistribute it and/or modify
  34 *  it under the terms of the GNU General Public License as published by
  35 *  the Free Software Foundation; either version 2 of the License, or
  36 *  (at your option) any later version.
  37 *
  38 *  This program is distributed in the hope that it will be useful,
  39 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  40 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  41 *  GNU General Public License for more details.
  42 *
  43 *  You should have received a copy of the GNU General Public License
  44 *  along with this program; if not, write to the Free Software
  45 *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  46 */
  47#include <linux/slab.h>
  48#include <linux/poll.h>
  49#include <linux/fs.h>
  50#include <linux/file.h>
  51#include <linux/jhash.h>
  52#include <linux/init.h>
  53#include <linux/futex.h>
  54#include <linux/mount.h>
  55#include <linux/pagemap.h>
  56#include <linux/syscalls.h>
  57#include <linux/signal.h>
  58#include <linux/export.h>
  59#include <linux/magic.h>
  60#include <linux/pid.h>
  61#include <linux/nsproxy.h>
  62#include <linux/ptrace.h>
  63#include <linux/sched/rt.h>
  64#include <linux/hugetlb.h>
  65#include <linux/freezer.h>
  66#include <linux/bootmem.h>
  67#include <linux/fault-inject.h>
  68
  69#include <asm/futex.h>
  70
  71#include "locking/rtmutex_common.h"
  72
  73/*
  74 * READ this before attempting to hack on futexes!
  75 *
  76 * Basic futex operation and ordering guarantees
  77 * =============================================
  78 *
  79 * The waiter reads the futex value in user space and calls
  80 * futex_wait(). This function computes the hash bucket and acquires
  81 * the hash bucket lock. After that it reads the futex user space value
  82 * again and verifies that the data has not changed. If it has not changed
  83 * it enqueues itself into the hash bucket, releases the hash bucket lock
  84 * and schedules.
  85 *
  86 * The waker side modifies the user space value of the futex and calls
  87 * futex_wake(). This function computes the hash bucket and acquires the
  88 * hash bucket lock. Then it looks for waiters on that futex in the hash
  89 * bucket and wakes them.
  90 *
  91 * In futex wake up scenarios where no tasks are blocked on a futex, taking
  92 * the hb spinlock can be avoided and simply return. In order for this
  93 * optimization to work, ordering guarantees must exist so that the waiter
  94 * being added to the list is acknowledged when the list is concurrently being
  95 * checked by the waker, avoiding scenarios like the following:
  96 *
  97 * CPU 0                               CPU 1
  98 * val = *futex;
  99 * sys_futex(WAIT, futex, val);
 100 *   futex_wait(futex, val);
 101 *   uval = *futex;
 102 *                                     *futex = newval;
 103 *                                     sys_futex(WAKE, futex);
 104 *                                       futex_wake(futex);
 105 *                                       if (queue_empty())
 106 *                                         return;
 107 *   if (uval == val)
 108 *      lock(hash_bucket(futex));
 109 *      queue();
 110 *     unlock(hash_bucket(futex));
 111 *     schedule();
 112 *
 113 * This would cause the waiter on CPU 0 to wait forever because it
 114 * missed the transition of the user space value from val to newval
 115 * and the waker did not find the waiter in the hash bucket queue.
 116 *
 117 * The correct serialization ensures that a waiter either observes
 118 * the changed user space value before blocking or is woken by a
 119 * concurrent waker:
 120 *
 121 * CPU 0                                 CPU 1
 122 * val = *futex;
 123 * sys_futex(WAIT, futex, val);
 124 *   futex_wait(futex, val);
 125 *
 126 *   waiters++; (a)
 127 *   smp_mb(); (A) <-- paired with -.
 128 *                                  |
 129 *   lock(hash_bucket(futex));      |
 130 *                                  |
 131 *   uval = *futex;                 |
 132 *                                  |        *futex = newval;
 133 *                                  |        sys_futex(WAKE, futex);
 134 *                                  |          futex_wake(futex);
 135 *                                  |
 136 *                                  `--------> smp_mb(); (B)
 137 *   if (uval == val)
 138 *     queue();
 139 *     unlock(hash_bucket(futex));
 140 *     schedule();                         if (waiters)
 141 *                                           lock(hash_bucket(futex));
 142 *   else                                    wake_waiters(futex);
 143 *     waiters--; (b)                        unlock(hash_bucket(futex));
 144 *
 145 * Where (A) orders the waiters increment and the futex value read through
 146 * atomic operations (see hb_waiters_inc) and where (B) orders the write
 147 * to futex and the waiters read -- this is done by the barriers for both
 148 * shared and private futexes in get_futex_key_refs().
 149 *
 150 * This yields the following case (where X:=waiters, Y:=futex):
 151 *
 152 *	X = Y = 0
 153 *
 154 *	w[X]=1		w[Y]=1
 155 *	MB		MB
 156 *	r[Y]=y		r[X]=x
 157 *
 158 * Which guarantees that x==0 && y==0 is impossible; which translates back into
 159 * the guarantee that we cannot both miss the futex variable change and the
 160 * enqueue.
 161 *
 162 * Note that a new waiter is accounted for in (a) even when it is possible that
 163 * the wait call can return error, in which case we backtrack from it in (b).
 164 * Refer to the comment in queue_lock().
 165 *
 166 * Similarly, in order to account for waiters being requeued on another
 167 * address we always increment the waiters for the destination bucket before
 168 * acquiring the lock. It then decrements them again  after releasing it -
 169 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
 170 * will do the additional required waiter count housekeeping. This is done for
 171 * double_lock_hb() and double_unlock_hb(), respectively.
 172 */
 173
 174#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
 175int __read_mostly futex_cmpxchg_enabled;
 176#endif
 177
 178/*
 179 * Futex flags used to encode options to functions and preserve them across
 180 * restarts.
 181 */
 182#define FLAGS_SHARED		0x01
 
 
 
 
 
 
 
 
 183#define FLAGS_CLOCKRT		0x02
 184#define FLAGS_HAS_TIMEOUT	0x04
 185
 186/*
 187 * Priority Inheritance state:
 188 */
 189struct futex_pi_state {
 190	/*
 191	 * list of 'owned' pi_state instances - these have to be
 192	 * cleaned up in do_exit() if the task exits prematurely:
 193	 */
 194	struct list_head list;
 195
 196	/*
 197	 * The PI object:
 198	 */
 199	struct rt_mutex pi_mutex;
 200
 201	struct task_struct *owner;
 202	atomic_t refcount;
 203
 204	union futex_key key;
 205};
 206
 207/**
 208 * struct futex_q - The hashed futex queue entry, one per waiting task
 209 * @list:		priority-sorted list of tasks waiting on this futex
 210 * @task:		the task waiting on the futex
 211 * @lock_ptr:		the hash bucket lock
 212 * @key:		the key the futex is hashed on
 213 * @pi_state:		optional priority inheritance state
 214 * @rt_waiter:		rt_waiter storage for use with requeue_pi
 215 * @requeue_pi_key:	the requeue_pi target futex key
 216 * @bitset:		bitset for the optional bitmasked wakeup
 217 *
 218 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
 219 * we can wake only the relevant ones (hashed queues may be shared).
 220 *
 221 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
 222 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
 223 * The order of wakeup is always to make the first condition true, then
 224 * the second.
 225 *
 226 * PI futexes are typically woken before they are removed from the hash list via
 227 * the rt_mutex code. See unqueue_me_pi().
 228 */
 229struct futex_q {
 230	struct plist_node list;
 231
 232	struct task_struct *task;
 233	spinlock_t *lock_ptr;
 234	union futex_key key;
 235	struct futex_pi_state *pi_state;
 236	struct rt_mutex_waiter *rt_waiter;
 237	union futex_key *requeue_pi_key;
 238	u32 bitset;
 239};
 240
 241static const struct futex_q futex_q_init = {
 242	/* list gets initialized in queue_me()*/
 243	.key = FUTEX_KEY_INIT,
 244	.bitset = FUTEX_BITSET_MATCH_ANY
 245};
 246
 247/*
 248 * Hash buckets are shared by all the futex_keys that hash to the same
 249 * location.  Each key may have multiple futex_q structures, one for each task
 250 * waiting on a futex.
 251 */
 252struct futex_hash_bucket {
 253	atomic_t waiters;
 254	spinlock_t lock;
 255	struct plist_head chain;
 256} ____cacheline_aligned_in_smp;
 257
 258/*
 259 * The base of the bucket array and its size are always used together
 260 * (after initialization only in hash_futex()), so ensure that they
 261 * reside in the same cacheline.
 262 */
 263static struct {
 264	struct futex_hash_bucket *queues;
 265	unsigned long            hashsize;
 266} __futex_data __read_mostly __aligned(2*sizeof(long));
 267#define futex_queues   (__futex_data.queues)
 268#define futex_hashsize (__futex_data.hashsize)
 269
 270
 271/*
 272 * Fault injections for futexes.
 273 */
 274#ifdef CONFIG_FAIL_FUTEX
 275
 276static struct {
 277	struct fault_attr attr;
 278
 279	bool ignore_private;
 280} fail_futex = {
 281	.attr = FAULT_ATTR_INITIALIZER,
 282	.ignore_private = false,
 283};
 284
 285static int __init setup_fail_futex(char *str)
 286{
 287	return setup_fault_attr(&fail_futex.attr, str);
 288}
 289__setup("fail_futex=", setup_fail_futex);
 290
 291static bool should_fail_futex(bool fshared)
 292{
 293	if (fail_futex.ignore_private && !fshared)
 294		return false;
 295
 296	return should_fail(&fail_futex.attr, 1);
 297}
 298
 299#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 300
 301static int __init fail_futex_debugfs(void)
 302{
 303	umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
 304	struct dentry *dir;
 305
 306	dir = fault_create_debugfs_attr("fail_futex", NULL,
 307					&fail_futex.attr);
 308	if (IS_ERR(dir))
 309		return PTR_ERR(dir);
 310
 311	if (!debugfs_create_bool("ignore-private", mode, dir,
 312				 &fail_futex.ignore_private)) {
 313		debugfs_remove_recursive(dir);
 314		return -ENOMEM;
 315	}
 316
 317	return 0;
 318}
 319
 320late_initcall(fail_futex_debugfs);
 321
 322#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
 323
 324#else
 325static inline bool should_fail_futex(bool fshared)
 326{
 327	return false;
 328}
 329#endif /* CONFIG_FAIL_FUTEX */
 330
 331static inline void futex_get_mm(union futex_key *key)
 332{
 333	atomic_inc(&key->private.mm->mm_count);
 334	/*
 335	 * Ensure futex_get_mm() implies a full barrier such that
 336	 * get_futex_key() implies a full barrier. This is relied upon
 337	 * as smp_mb(); (B), see the ordering comment above.
 338	 */
 339	smp_mb__after_atomic();
 340}
 341
 342/*
 343 * Reflects a new waiter being added to the waitqueue.
 344 */
 345static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
 346{
 347#ifdef CONFIG_SMP
 348	atomic_inc(&hb->waiters);
 349	/*
 350	 * Full barrier (A), see the ordering comment above.
 351	 */
 352	smp_mb__after_atomic();
 353#endif
 354}
 355
 356/*
 357 * Reflects a waiter being removed from the waitqueue by wakeup
 358 * paths.
 359 */
 360static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
 361{
 362#ifdef CONFIG_SMP
 363	atomic_dec(&hb->waiters);
 364#endif
 365}
 366
 367static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
 368{
 369#ifdef CONFIG_SMP
 370	return atomic_read(&hb->waiters);
 371#else
 372	return 1;
 373#endif
 374}
 375
 376/*
 377 * We hash on the keys returned from get_futex_key (see below).
 
 
 
 
 378 */
 379static struct futex_hash_bucket *hash_futex(union futex_key *key)
 380{
 381	u32 hash = jhash2((u32*)&key->both.word,
 382			  (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
 383			  key->both.offset);
 384	return &futex_queues[hash & (futex_hashsize - 1)];
 385}
 386
 387/*
 
 
 
 
 
 388 * Return 1 if two futex_keys are equal, 0 otherwise.
 389 */
 390static inline int match_futex(union futex_key *key1, union futex_key *key2)
 391{
 392	return (key1 && key2
 393		&& key1->both.word == key2->both.word
 394		&& key1->both.ptr == key2->both.ptr
 395		&& key1->both.offset == key2->both.offset);
 396}
 397
 398/*
 399 * Take a reference to the resource addressed by a key.
 400 * Can be called while holding spinlocks.
 401 *
 402 */
 403static void get_futex_key_refs(union futex_key *key)
 404{
 405	if (!key->both.ptr)
 406		return;
 407
 
 
 
 
 
 
 
 
 
 
 408	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
 409	case FUT_OFF_INODE:
 410		ihold(key->shared.inode); /* implies smp_mb(); (B) */
 411		break;
 412	case FUT_OFF_MMSHARED:
 413		futex_get_mm(key); /* implies smp_mb(); (B) */
 414		break;
 415	default:
 416		/*
 417		 * Private futexes do not hold reference on an inode or
 418		 * mm, therefore the only purpose of calling get_futex_key_refs
 419		 * is because we need the barrier for the lockless waiter check.
 420		 */
 421		smp_mb(); /* explicit smp_mb(); (B) */
 422	}
 423}
 424
 425/*
 426 * Drop a reference to the resource addressed by a key.
 427 * The hash bucket spinlock must not be held. This is
 428 * a no-op for private futexes, see comment in the get
 429 * counterpart.
 430 */
 431static void drop_futex_key_refs(union futex_key *key)
 432{
 433	if (!key->both.ptr) {
 434		/* If we're here then we tried to put a key we failed to get */
 435		WARN_ON_ONCE(1);
 436		return;
 437	}
 438
 
 
 
 439	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
 440	case FUT_OFF_INODE:
 441		iput(key->shared.inode);
 442		break;
 443	case FUT_OFF_MMSHARED:
 444		mmdrop(key->private.mm);
 445		break;
 446	}
 447}
 448
 449/**
 450 * get_futex_key() - Get parameters which are the keys for a futex
 451 * @uaddr:	virtual address of the futex
 452 * @fshared:	0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
 453 * @key:	address where result is stored.
 454 * @rw:		mapping needs to be read/write (values: VERIFY_READ,
 455 *              VERIFY_WRITE)
 456 *
 457 * Return: a negative error code or 0
 458 *
 459 * The key words are stored in *key on success.
 460 *
 461 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
 462 * offset_within_page).  For private mappings, it's (uaddr, current->mm).
 463 * We can usually work out the index without swapping in the page.
 464 *
 465 * lock_page() might sleep, the caller should not hold a spinlock.
 466 */
 467static int
 468get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
 469{
 470	unsigned long address = (unsigned long)uaddr;
 471	struct mm_struct *mm = current->mm;
 472	struct page *page;
 473	struct address_space *mapping;
 474	int err, ro = 0;
 475
 476	/*
 477	 * The futex address must be "naturally" aligned.
 478	 */
 479	key->both.offset = address % PAGE_SIZE;
 480	if (unlikely((address % sizeof(u32)) != 0))
 481		return -EINVAL;
 482	address -= key->both.offset;
 483
 484	if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
 485		return -EFAULT;
 486
 487	if (unlikely(should_fail_futex(fshared)))
 488		return -EFAULT;
 489
 490	/*
 491	 * PROCESS_PRIVATE futexes are fast.
 492	 * As the mm cannot disappear under us and the 'key' only needs
 493	 * virtual address, we dont even have to find the underlying vma.
 494	 * Note : We do have to check 'uaddr' is a valid user address,
 495	 *        but access_ok() should be faster than find_vma()
 496	 */
 497	if (!fshared) {
 498		key->private.mm = mm;
 499		key->private.address = address;
 500		get_futex_key_refs(key);  /* implies smp_mb(); (B) */
 501		return 0;
 502	}
 503
 504again:
 505	/* Ignore any VERIFY_READ mapping (futex common case) */
 506	if (unlikely(should_fail_futex(fshared)))
 507		return -EFAULT;
 508
 509	err = get_user_pages_fast(address, 1, 1, &page);
 510	/*
 511	 * If write access is not required (eg. FUTEX_WAIT), try
 512	 * and get read-only access.
 513	 */
 514	if (err == -EFAULT && rw == VERIFY_READ) {
 515		err = get_user_pages_fast(address, 1, 0, &page);
 516		ro = 1;
 517	}
 518	if (err < 0)
 519		return err;
 520	else
 521		err = 0;
 522
 523	/*
 524	 * The treatment of mapping from this point on is critical. The page
 525	 * lock protects many things but in this context the page lock
 526	 * stabilizes mapping, prevents inode freeing in the shared
 527	 * file-backed region case and guards against movement to swap cache.
 528	 *
 529	 * Strictly speaking the page lock is not needed in all cases being
 530	 * considered here and page lock forces unnecessarily serialization
 531	 * From this point on, mapping will be re-verified if necessary and
 532	 * page lock will be acquired only if it is unavoidable
 
 
 
 
 
 
 
 533	 */
 
 534	page = compound_head(page);
 535	mapping = READ_ONCE(page->mapping);
 536
 537	/*
 538	 * If page->mapping is NULL, then it cannot be a PageAnon
 539	 * page; but it might be the ZERO_PAGE or in the gate area or
 540	 * in a special mapping (all cases which we are happy to fail);
 541	 * or it may have been a good file page when get_user_pages_fast
 542	 * found it, but truncated or holepunched or subjected to
 543	 * invalidate_complete_page2 before we got the page lock (also
 544	 * cases which we are happy to fail).  And we hold a reference,
 545	 * so refcount care in invalidate_complete_page's remove_mapping
 546	 * prevents drop_caches from setting mapping to NULL beneath us.
 547	 *
 548	 * The case we do have to guard against is when memory pressure made
 549	 * shmem_writepage move it from filecache to swapcache beneath us:
 550	 * an unlikely race, but we do need to retry for page->mapping.
 551	 */
 552	if (unlikely(!mapping)) {
 553		int shmem_swizzled;
 554
 555		/*
 556		 * Page lock is required to identify which special case above
 557		 * applies. If this is really a shmem page then the page lock
 558		 * will prevent unexpected transitions.
 559		 */
 560		lock_page(page);
 561		shmem_swizzled = PageSwapCache(page) || page->mapping;
 562		unlock_page(page);
 563		put_page(page);
 564
 565		if (shmem_swizzled)
 566			goto again;
 567
 568		return -EFAULT;
 569	}
 570
 571	/*
 572	 * Private mappings are handled in a simple way.
 573	 *
 574	 * If the futex key is stored on an anonymous page, then the associated
 575	 * object is the mm which is implicitly pinned by the calling process.
 576	 *
 577	 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
 578	 * it's a read-only handle, it's expected that futexes attach to
 579	 * the object not the particular process.
 580	 */
 581	if (PageAnon(page)) {
 582		/*
 583		 * A RO anonymous page will never change and thus doesn't make
 584		 * sense for futex operations.
 585		 */
 586		if (unlikely(should_fail_futex(fshared)) || ro) {
 587			err = -EFAULT;
 588			goto out;
 589		}
 590
 591		key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
 592		key->private.mm = mm;
 593		key->private.address = address;
 594
 595		get_futex_key_refs(key); /* implies smp_mb(); (B) */
 596
 597	} else {
 598		struct inode *inode;
 599
 600		/*
 601		 * The associated futex object in this case is the inode and
 602		 * the page->mapping must be traversed. Ordinarily this should
 603		 * be stabilised under page lock but it's not strictly
 604		 * necessary in this case as we just want to pin the inode, not
 605		 * update the radix tree or anything like that.
 606		 *
 607		 * The RCU read lock is taken as the inode is finally freed
 608		 * under RCU. If the mapping still matches expectations then the
 609		 * mapping->host can be safely accessed as being a valid inode.
 610		 */
 611		rcu_read_lock();
 612
 613		if (READ_ONCE(page->mapping) != mapping) {
 614			rcu_read_unlock();
 615			put_page(page);
 616
 617			goto again;
 618		}
 619
 620		inode = READ_ONCE(mapping->host);
 621		if (!inode) {
 622			rcu_read_unlock();
 623			put_page(page);
 624
 625			goto again;
 626		}
 627
 628		/*
 629		 * Take a reference unless it is about to be freed. Previously
 630		 * this reference was taken by ihold under the page lock
 631		 * pinning the inode in place so i_lock was unnecessary. The
 632		 * only way for this check to fail is if the inode was
 633		 * truncated in parallel so warn for now if this happens.
 634		 *
 635		 * We are not calling into get_futex_key_refs() in file-backed
 636		 * cases, therefore a successful atomic_inc return below will
 637		 * guarantee that get_futex_key() will still imply smp_mb(); (B).
 638		 */
 639		if (WARN_ON_ONCE(!atomic_inc_not_zero(&inode->i_count))) {
 640			rcu_read_unlock();
 641			put_page(page);
 642
 643			goto again;
 644		}
 645
 646		/* Should be impossible but lets be paranoid for now */
 647		if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
 648			err = -EFAULT;
 649			rcu_read_unlock();
 650			iput(inode);
 651
 652			goto out;
 653		}
 654
 655		key->both.offset |= FUT_OFF_INODE; /* inode-based key */
 656		key->shared.inode = inode;
 657		key->shared.pgoff = basepage_index(page);
 658		rcu_read_unlock();
 659	}
 660
 661out:
 662	put_page(page);
 663	return err;
 664}
 665
 666static inline void put_futex_key(union futex_key *key)
 667{
 668	drop_futex_key_refs(key);
 669}
 670
 671/**
 672 * fault_in_user_writeable() - Fault in user address and verify RW access
 673 * @uaddr:	pointer to faulting user space address
 674 *
 675 * Slow path to fixup the fault we just took in the atomic write
 676 * access to @uaddr.
 677 *
 678 * We have no generic implementation of a non-destructive write to the
 679 * user address. We know that we faulted in the atomic pagefault
 680 * disabled section so we can as well avoid the #PF overhead by
 681 * calling get_user_pages() right away.
 682 */
 683static int fault_in_user_writeable(u32 __user *uaddr)
 684{
 685	struct mm_struct *mm = current->mm;
 686	int ret;
 687
 688	down_read(&mm->mmap_sem);
 689	ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
 690			       FAULT_FLAG_WRITE, NULL);
 691	up_read(&mm->mmap_sem);
 692
 693	return ret < 0 ? ret : 0;
 694}
 695
 696/**
 697 * futex_top_waiter() - Return the highest priority waiter on a futex
 698 * @hb:		the hash bucket the futex_q's reside in
 699 * @key:	the futex key (to distinguish it from other futex futex_q's)
 700 *
 701 * Must be called with the hb lock held.
 702 */
 703static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
 704					union futex_key *key)
 705{
 706	struct futex_q *this;
 707
 708	plist_for_each_entry(this, &hb->chain, list) {
 709		if (match_futex(&this->key, key))
 710			return this;
 711	}
 712	return NULL;
 713}
 714
 715static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
 716				      u32 uval, u32 newval)
 717{
 718	int ret;
 719
 720	pagefault_disable();
 721	ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
 722	pagefault_enable();
 723
 724	return ret;
 725}
 726
 727static int get_futex_value_locked(u32 *dest, u32 __user *from)
 728{
 729	int ret;
 730
 731	pagefault_disable();
 732	ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
 733	pagefault_enable();
 734
 735	return ret ? -EFAULT : 0;
 736}
 737
 738
 739/*
 740 * PI code:
 741 */
 742static int refill_pi_state_cache(void)
 743{
 744	struct futex_pi_state *pi_state;
 745
 746	if (likely(current->pi_state_cache))
 747		return 0;
 748
 749	pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
 750
 751	if (!pi_state)
 752		return -ENOMEM;
 753
 754	INIT_LIST_HEAD(&pi_state->list);
 755	/* pi_mutex gets initialized later */
 756	pi_state->owner = NULL;
 757	atomic_set(&pi_state->refcount, 1);
 758	pi_state->key = FUTEX_KEY_INIT;
 759
 760	current->pi_state_cache = pi_state;
 761
 762	return 0;
 763}
 764
 765static struct futex_pi_state * alloc_pi_state(void)
 766{
 767	struct futex_pi_state *pi_state = current->pi_state_cache;
 768
 769	WARN_ON(!pi_state);
 770	current->pi_state_cache = NULL;
 771
 772	return pi_state;
 773}
 774
 775/*
 776 * Drops a reference to the pi_state object and frees or caches it
 777 * when the last reference is gone.
 778 *
 779 * Must be called with the hb lock held.
 780 */
 781static void put_pi_state(struct futex_pi_state *pi_state)
 782{
 783	if (!pi_state)
 784		return;
 785
 786	if (!atomic_dec_and_test(&pi_state->refcount))
 787		return;
 788
 789	/*
 790	 * If pi_state->owner is NULL, the owner is most probably dying
 791	 * and has cleaned up the pi_state already
 792	 */
 793	if (pi_state->owner) {
 794		raw_spin_lock_irq(&pi_state->owner->pi_lock);
 795		list_del_init(&pi_state->list);
 796		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
 797
 798		rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
 799	}
 800
 801	if (current->pi_state_cache)
 802		kfree(pi_state);
 803	else {
 804		/*
 805		 * pi_state->list is already empty.
 806		 * clear pi_state->owner.
 807		 * refcount is at 0 - put it back to 1.
 808		 */
 809		pi_state->owner = NULL;
 810		atomic_set(&pi_state->refcount, 1);
 811		current->pi_state_cache = pi_state;
 812	}
 813}
 814
 815/*
 816 * Look up the task based on what TID userspace gave us.
 817 * We dont trust it.
 818 */
 819static struct task_struct * futex_find_get_task(pid_t pid)
 820{
 821	struct task_struct *p;
 822
 823	rcu_read_lock();
 824	p = find_task_by_vpid(pid);
 825	if (p)
 826		get_task_struct(p);
 827
 828	rcu_read_unlock();
 829
 830	return p;
 831}
 832
 833/*
 834 * This task is holding PI mutexes at exit time => bad.
 835 * Kernel cleans up PI-state, but userspace is likely hosed.
 836 * (Robust-futex cleanup is separate and might save the day for userspace.)
 837 */
 838void exit_pi_state_list(struct task_struct *curr)
 839{
 840	struct list_head *next, *head = &curr->pi_state_list;
 841	struct futex_pi_state *pi_state;
 842	struct futex_hash_bucket *hb;
 843	union futex_key key = FUTEX_KEY_INIT;
 844
 845	if (!futex_cmpxchg_enabled)
 846		return;
 847	/*
 848	 * We are a ZOMBIE and nobody can enqueue itself on
 849	 * pi_state_list anymore, but we have to be careful
 850	 * versus waiters unqueueing themselves:
 851	 */
 852	raw_spin_lock_irq(&curr->pi_lock);
 853	while (!list_empty(head)) {
 854
 855		next = head->next;
 856		pi_state = list_entry(next, struct futex_pi_state, list);
 857		key = pi_state->key;
 858		hb = hash_futex(&key);
 859		raw_spin_unlock_irq(&curr->pi_lock);
 860
 861		spin_lock(&hb->lock);
 862
 863		raw_spin_lock_irq(&curr->pi_lock);
 864		/*
 865		 * We dropped the pi-lock, so re-check whether this
 866		 * task still owns the PI-state:
 867		 */
 868		if (head->next != next) {
 869			spin_unlock(&hb->lock);
 870			continue;
 871		}
 872
 873		WARN_ON(pi_state->owner != curr);
 874		WARN_ON(list_empty(&pi_state->list));
 875		list_del_init(&pi_state->list);
 876		pi_state->owner = NULL;
 877		raw_spin_unlock_irq(&curr->pi_lock);
 878
 879		rt_mutex_unlock(&pi_state->pi_mutex);
 880
 881		spin_unlock(&hb->lock);
 882
 883		raw_spin_lock_irq(&curr->pi_lock);
 884	}
 885	raw_spin_unlock_irq(&curr->pi_lock);
 886}
 887
 888/*
 889 * We need to check the following states:
 890 *
 891 *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
 892 *
 893 * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
 894 * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
 895 *
 896 * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
 897 *
 898 * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
 899 * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
 900 *
 901 * [6]  Found  | Found    | task      | 0         | 1      | Valid
 902 *
 903 * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
 904 *
 905 * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
 906 * [9]  Found  | Found    | task      | 0         | 0      | Invalid
 907 * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
 908 *
 909 * [1]	Indicates that the kernel can acquire the futex atomically. We
 910 *	came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
 911 *
 912 * [2]	Valid, if TID does not belong to a kernel thread. If no matching
 913 *      thread is found then it indicates that the owner TID has died.
 914 *
 915 * [3]	Invalid. The waiter is queued on a non PI futex
 916 *
 917 * [4]	Valid state after exit_robust_list(), which sets the user space
 918 *	value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
 919 *
 920 * [5]	The user space value got manipulated between exit_robust_list()
 921 *	and exit_pi_state_list()
 922 *
 923 * [6]	Valid state after exit_pi_state_list() which sets the new owner in
 924 *	the pi_state but cannot access the user space value.
 925 *
 926 * [7]	pi_state->owner can only be NULL when the OWNER_DIED bit is set.
 927 *
 928 * [8]	Owner and user space value match
 929 *
 930 * [9]	There is no transient state which sets the user space TID to 0
 931 *	except exit_robust_list(), but this is indicated by the
 932 *	FUTEX_OWNER_DIED bit. See [4]
 933 *
 934 * [10] There is no transient state which leaves owner and user space
 935 *	TID out of sync.
 936 */
 937
 938/*
 939 * Validate that the existing waiter has a pi_state and sanity check
 940 * the pi_state against the user space value. If correct, attach to
 941 * it.
 942 */
 943static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
 944			      struct futex_pi_state **ps)
 945{
 946	pid_t pid = uval & FUTEX_TID_MASK;
 947
 948	/*
 949	 * Userspace might have messed up non-PI and PI futexes [3]
 950	 */
 951	if (unlikely(!pi_state))
 952		return -EINVAL;
 953
 954	WARN_ON(!atomic_read(&pi_state->refcount));
 955
 956	/*
 957	 * Handle the owner died case:
 958	 */
 959	if (uval & FUTEX_OWNER_DIED) {
 960		/*
 961		 * exit_pi_state_list sets owner to NULL and wakes the
 962		 * topmost waiter. The task which acquires the
 963		 * pi_state->rt_mutex will fixup owner.
 964		 */
 965		if (!pi_state->owner) {
 966			/*
 967			 * No pi state owner, but the user space TID
 968			 * is not 0. Inconsistent state. [5]
 969			 */
 970			if (pid)
 971				return -EINVAL;
 972			/*
 973			 * Take a ref on the state and return success. [4]
 974			 */
 975			goto out_state;
 976		}
 977
 978		/*
 979		 * If TID is 0, then either the dying owner has not
 980		 * yet executed exit_pi_state_list() or some waiter
 981		 * acquired the rtmutex in the pi state, but did not
 982		 * yet fixup the TID in user space.
 983		 *
 984		 * Take a ref on the state and return success. [6]
 985		 */
 986		if (!pid)
 987			goto out_state;
 988	} else {
 989		/*
 990		 * If the owner died bit is not set, then the pi_state
 991		 * must have an owner. [7]
 992		 */
 993		if (!pi_state->owner)
 994			return -EINVAL;
 995	}
 996
 997	/*
 998	 * Bail out if user space manipulated the futex value. If pi
 999	 * state exists then the owner TID must be the same as the
1000	 * user space TID. [9/10]
1001	 */
1002	if (pid != task_pid_vnr(pi_state->owner))
1003		return -EINVAL;
1004out_state:
1005	atomic_inc(&pi_state->refcount);
1006	*ps = pi_state;
1007	return 0;
1008}
1009
1010/*
1011 * Lookup the task for the TID provided from user space and attach to
1012 * it after doing proper sanity checks.
1013 */
1014static int attach_to_pi_owner(u32 uval, union futex_key *key,
1015			      struct futex_pi_state **ps)
1016{
1017	pid_t pid = uval & FUTEX_TID_MASK;
1018	struct futex_pi_state *pi_state;
1019	struct task_struct *p;
1020
1021	/*
1022	 * We are the first waiter - try to look up the real owner and attach
1023	 * the new pi_state to it, but bail out when TID = 0 [1]
1024	 */
1025	if (!pid)
1026		return -ESRCH;
1027	p = futex_find_get_task(pid);
1028	if (!p)
1029		return -ESRCH;
1030
1031	if (unlikely(p->flags & PF_KTHREAD)) {
1032		put_task_struct(p);
1033		return -EPERM;
1034	}
1035
1036	/*
1037	 * We need to look at the task state flags to figure out,
1038	 * whether the task is exiting. To protect against the do_exit
1039	 * change of the task flags, we do this protected by
1040	 * p->pi_lock:
1041	 */
1042	raw_spin_lock_irq(&p->pi_lock);
1043	if (unlikely(p->flags & PF_EXITING)) {
1044		/*
1045		 * The task is on the way out. When PF_EXITPIDONE is
1046		 * set, we know that the task has finished the
1047		 * cleanup:
1048		 */
1049		int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1050
1051		raw_spin_unlock_irq(&p->pi_lock);
1052		put_task_struct(p);
1053		return ret;
1054	}
1055
1056	/*
1057	 * No existing pi state. First waiter. [2]
1058	 */
1059	pi_state = alloc_pi_state();
1060
1061	/*
1062	 * Initialize the pi_mutex in locked state and make @p
1063	 * the owner of it:
1064	 */
1065	rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1066
1067	/* Store the key for possible exit cleanups: */
1068	pi_state->key = *key;
1069
1070	WARN_ON(!list_empty(&pi_state->list));
1071	list_add(&pi_state->list, &p->pi_state_list);
1072	pi_state->owner = p;
1073	raw_spin_unlock_irq(&p->pi_lock);
1074
1075	put_task_struct(p);
1076
1077	*ps = pi_state;
1078
1079	return 0;
1080}
1081
1082static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
1083			   union futex_key *key, struct futex_pi_state **ps)
1084{
1085	struct futex_q *match = futex_top_waiter(hb, key);
1086
1087	/*
1088	 * If there is a waiter on that futex, validate it and
1089	 * attach to the pi_state when the validation succeeds.
1090	 */
1091	if (match)
1092		return attach_to_pi_state(uval, match->pi_state, ps);
1093
1094	/*
1095	 * We are the first waiter - try to look up the owner based on
1096	 * @uval and attach to it.
1097	 */
1098	return attach_to_pi_owner(uval, key, ps);
1099}
1100
1101static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1102{
1103	u32 uninitialized_var(curval);
1104
1105	if (unlikely(should_fail_futex(true)))
1106		return -EFAULT;
1107
1108	if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1109		return -EFAULT;
1110
1111	/*If user space value changed, let the caller retry */
1112	return curval != uval ? -EAGAIN : 0;
1113}
1114
1115/**
1116 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1117 * @uaddr:		the pi futex user address
1118 * @hb:			the pi futex hash bucket
1119 * @key:		the futex key associated with uaddr and hb
1120 * @ps:			the pi_state pointer where we store the result of the
1121 *			lookup
1122 * @task:		the task to perform the atomic lock work for.  This will
1123 *			be "current" except in the case of requeue pi.
1124 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1125 *
1126 * Return:
1127 *  0 - ready to wait;
1128 *  1 - acquired the lock;
1129 * <0 - error
1130 *
1131 * The hb->lock and futex_key refs shall be held by the caller.
1132 */
1133static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1134				union futex_key *key,
1135				struct futex_pi_state **ps,
1136				struct task_struct *task, int set_waiters)
1137{
1138	u32 uval, newval, vpid = task_pid_vnr(task);
1139	struct futex_q *match;
1140	int ret;
1141
1142	/*
1143	 * Read the user space value first so we can validate a few
1144	 * things before proceeding further.
1145	 */
1146	if (get_futex_value_locked(&uval, uaddr))
1147		return -EFAULT;
1148
1149	if (unlikely(should_fail_futex(true)))
1150		return -EFAULT;
1151
1152	/*
1153	 * Detect deadlocks.
1154	 */
1155	if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1156		return -EDEADLK;
1157
1158	if ((unlikely(should_fail_futex(true))))
1159		return -EDEADLK;
1160
1161	/*
1162	 * Lookup existing state first. If it exists, try to attach to
1163	 * its pi_state.
1164	 */
1165	match = futex_top_waiter(hb, key);
1166	if (match)
1167		return attach_to_pi_state(uval, match->pi_state, ps);
1168
1169	/*
1170	 * No waiter and user TID is 0. We are here because the
1171	 * waiters or the owner died bit is set or called from
1172	 * requeue_cmp_pi or for whatever reason something took the
1173	 * syscall.
1174	 */
1175	if (!(uval & FUTEX_TID_MASK)) {
1176		/*
1177		 * We take over the futex. No other waiters and the user space
1178		 * TID is 0. We preserve the owner died bit.
1179		 */
1180		newval = uval & FUTEX_OWNER_DIED;
1181		newval |= vpid;
1182
1183		/* The futex requeue_pi code can enforce the waiters bit */
1184		if (set_waiters)
1185			newval |= FUTEX_WAITERS;
1186
1187		ret = lock_pi_update_atomic(uaddr, uval, newval);
1188		/* If the take over worked, return 1 */
1189		return ret < 0 ? ret : 1;
1190	}
1191
1192	/*
1193	 * First waiter. Set the waiters bit before attaching ourself to
1194	 * the owner. If owner tries to unlock, it will be forced into
1195	 * the kernel and blocked on hb->lock.
1196	 */
1197	newval = uval | FUTEX_WAITERS;
1198	ret = lock_pi_update_atomic(uaddr, uval, newval);
1199	if (ret)
1200		return ret;
1201	/*
1202	 * If the update of the user space value succeeded, we try to
1203	 * attach to the owner. If that fails, no harm done, we only
1204	 * set the FUTEX_WAITERS bit in the user space variable.
1205	 */
1206	return attach_to_pi_owner(uval, key, ps);
1207}
1208
1209/**
1210 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1211 * @q:	The futex_q to unqueue
1212 *
1213 * The q->lock_ptr must not be NULL and must be held by the caller.
1214 */
1215static void __unqueue_futex(struct futex_q *q)
1216{
1217	struct futex_hash_bucket *hb;
1218
1219	if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1220	    || WARN_ON(plist_node_empty(&q->list)))
1221		return;
1222
1223	hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1224	plist_del(&q->list, &hb->chain);
1225	hb_waiters_dec(hb);
1226}
1227
1228/*
1229 * The hash bucket lock must be held when this is called.
1230 * Afterwards, the futex_q must not be accessed. Callers
1231 * must ensure to later call wake_up_q() for the actual
1232 * wakeups to occur.
1233 */
1234static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1235{
1236	struct task_struct *p = q->task;
1237
1238	if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1239		return;
1240
1241	/*
1242	 * Queue the task for later wakeup for after we've released
1243	 * the hb->lock. wake_q_add() grabs reference to p.
1244	 */
1245	wake_q_add(wake_q, p);
1246	__unqueue_futex(q);
1247	/*
1248	 * The waiting task can free the futex_q as soon as
1249	 * q->lock_ptr = NULL is written, without taking any locks. A
1250	 * memory barrier is required here to prevent the following
1251	 * store to lock_ptr from getting ahead of the plist_del.
1252	 */
1253	smp_wmb();
1254	q->lock_ptr = NULL;
1255}
1256
1257static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
1258			 struct futex_hash_bucket *hb)
1259{
1260	struct task_struct *new_owner;
1261	struct futex_pi_state *pi_state = this->pi_state;
1262	u32 uninitialized_var(curval), newval;
1263	WAKE_Q(wake_q);
1264	bool deboost;
1265	int ret = 0;
1266
1267	if (!pi_state)
1268		return -EINVAL;
1269
1270	/*
1271	 * If current does not own the pi_state then the futex is
1272	 * inconsistent and user space fiddled with the futex value.
1273	 */
1274	if (pi_state->owner != current)
1275		return -EINVAL;
1276
1277	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1278	new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1279
1280	/*
1281	 * It is possible that the next waiter (the one that brought
1282	 * this owner to the kernel) timed out and is no longer
1283	 * waiting on the lock.
1284	 */
1285	if (!new_owner)
1286		new_owner = this->task;
1287
1288	/*
1289	 * We pass it to the next owner. The WAITERS bit is always
1290	 * kept enabled while there is PI state around. We cleanup the
1291	 * owner died bit, because we are the owner.
1292	 */
1293	newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1294
1295	if (unlikely(should_fail_futex(true)))
1296		ret = -EFAULT;
1297
1298	if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1299		ret = -EFAULT;
1300	} else if (curval != uval) {
1301		/*
1302		 * If a unconditional UNLOCK_PI operation (user space did not
1303		 * try the TID->0 transition) raced with a waiter setting the
1304		 * FUTEX_WAITERS flag between get_user() and locking the hash
1305		 * bucket lock, retry the operation.
1306		 */
1307		if ((FUTEX_TID_MASK & curval) == uval)
1308			ret = -EAGAIN;
1309		else
1310			ret = -EINVAL;
1311	}
1312	if (ret) {
1313		raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1314		return ret;
1315	}
1316
1317	raw_spin_lock(&pi_state->owner->pi_lock);
1318	WARN_ON(list_empty(&pi_state->list));
1319	list_del_init(&pi_state->list);
1320	raw_spin_unlock(&pi_state->owner->pi_lock);
1321
1322	raw_spin_lock(&new_owner->pi_lock);
1323	WARN_ON(!list_empty(&pi_state->list));
1324	list_add(&pi_state->list, &new_owner->pi_state_list);
1325	pi_state->owner = new_owner;
1326	raw_spin_unlock(&new_owner->pi_lock);
1327
1328	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1329
1330	deboost = rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1331
1332	/*
1333	 * First unlock HB so the waiter does not spin on it once he got woken
1334	 * up. Second wake up the waiter before the priority is adjusted. If we
1335	 * deboost first (and lose our higher priority), then the task might get
1336	 * scheduled away before the wake up can take place.
1337	 */
1338	spin_unlock(&hb->lock);
1339	wake_up_q(&wake_q);
1340	if (deboost)
1341		rt_mutex_adjust_prio(current);
1342
1343	return 0;
1344}
1345
1346/*
1347 * Express the locking dependencies for lockdep:
1348 */
1349static inline void
1350double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1351{
1352	if (hb1 <= hb2) {
1353		spin_lock(&hb1->lock);
1354		if (hb1 < hb2)
1355			spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1356	} else { /* hb1 > hb2 */
1357		spin_lock(&hb2->lock);
1358		spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1359	}
1360}
1361
1362static inline void
1363double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1364{
1365	spin_unlock(&hb1->lock);
1366	if (hb1 != hb2)
1367		spin_unlock(&hb2->lock);
1368}
1369
1370/*
1371 * Wake up waiters matching bitset queued on this futex (uaddr).
1372 */
1373static int
1374futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1375{
1376	struct futex_hash_bucket *hb;
1377	struct futex_q *this, *next;
1378	union futex_key key = FUTEX_KEY_INIT;
1379	int ret;
1380	WAKE_Q(wake_q);
1381
1382	if (!bitset)
1383		return -EINVAL;
1384
1385	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1386	if (unlikely(ret != 0))
1387		goto out;
1388
1389	hb = hash_futex(&key);
1390
1391	/* Make sure we really have tasks to wakeup */
1392	if (!hb_waiters_pending(hb))
1393		goto out_put_key;
1394
1395	spin_lock(&hb->lock);
1396
1397	plist_for_each_entry_safe(this, next, &hb->chain, list) {
1398		if (match_futex (&this->key, &key)) {
1399			if (this->pi_state || this->rt_waiter) {
1400				ret = -EINVAL;
1401				break;
1402			}
1403
1404			/* Check if one of the bits is set in both bitsets */
1405			if (!(this->bitset & bitset))
1406				continue;
1407
1408			mark_wake_futex(&wake_q, this);
1409			if (++ret >= nr_wake)
1410				break;
1411		}
1412	}
1413
1414	spin_unlock(&hb->lock);
1415	wake_up_q(&wake_q);
1416out_put_key:
1417	put_futex_key(&key);
1418out:
1419	return ret;
1420}
1421
1422/*
1423 * Wake up all waiters hashed on the physical page that is mapped
1424 * to this virtual address:
1425 */
1426static int
1427futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1428	      int nr_wake, int nr_wake2, int op)
1429{
1430	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1431	struct futex_hash_bucket *hb1, *hb2;
1432	struct futex_q *this, *next;
1433	int ret, op_ret;
1434	WAKE_Q(wake_q);
1435
1436retry:
1437	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1438	if (unlikely(ret != 0))
1439		goto out;
1440	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1441	if (unlikely(ret != 0))
1442		goto out_put_key1;
1443
1444	hb1 = hash_futex(&key1);
1445	hb2 = hash_futex(&key2);
1446
1447retry_private:
1448	double_lock_hb(hb1, hb2);
1449	op_ret = futex_atomic_op_inuser(op, uaddr2);
1450	if (unlikely(op_ret < 0)) {
1451
1452		double_unlock_hb(hb1, hb2);
1453
1454#ifndef CONFIG_MMU
1455		/*
1456		 * we don't get EFAULT from MMU faults if we don't have an MMU,
1457		 * but we might get them from range checking
1458		 */
1459		ret = op_ret;
1460		goto out_put_keys;
1461#endif
1462
1463		if (unlikely(op_ret != -EFAULT)) {
1464			ret = op_ret;
1465			goto out_put_keys;
1466		}
1467
1468		ret = fault_in_user_writeable(uaddr2);
1469		if (ret)
1470			goto out_put_keys;
1471
1472		if (!(flags & FLAGS_SHARED))
1473			goto retry_private;
1474
1475		put_futex_key(&key2);
1476		put_futex_key(&key1);
1477		goto retry;
1478	}
1479
1480	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1481		if (match_futex (&this->key, &key1)) {
1482			if (this->pi_state || this->rt_waiter) {
1483				ret = -EINVAL;
1484				goto out_unlock;
1485			}
1486			mark_wake_futex(&wake_q, this);
1487			if (++ret >= nr_wake)
1488				break;
1489		}
1490	}
1491
1492	if (op_ret > 0) {
1493		op_ret = 0;
1494		plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1495			if (match_futex (&this->key, &key2)) {
1496				if (this->pi_state || this->rt_waiter) {
1497					ret = -EINVAL;
1498					goto out_unlock;
1499				}
1500				mark_wake_futex(&wake_q, this);
1501				if (++op_ret >= nr_wake2)
1502					break;
1503			}
1504		}
1505		ret += op_ret;
1506	}
1507
1508out_unlock:
1509	double_unlock_hb(hb1, hb2);
1510	wake_up_q(&wake_q);
1511out_put_keys:
1512	put_futex_key(&key2);
1513out_put_key1:
1514	put_futex_key(&key1);
1515out:
1516	return ret;
1517}
1518
1519/**
1520 * requeue_futex() - Requeue a futex_q from one hb to another
1521 * @q:		the futex_q to requeue
1522 * @hb1:	the source hash_bucket
1523 * @hb2:	the target hash_bucket
1524 * @key2:	the new key for the requeued futex_q
1525 */
1526static inline
1527void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1528		   struct futex_hash_bucket *hb2, union futex_key *key2)
1529{
1530
1531	/*
1532	 * If key1 and key2 hash to the same bucket, no need to
1533	 * requeue.
1534	 */
1535	if (likely(&hb1->chain != &hb2->chain)) {
1536		plist_del(&q->list, &hb1->chain);
1537		hb_waiters_dec(hb1);
1538		hb_waiters_inc(hb2);
1539		plist_add(&q->list, &hb2->chain);
1540		q->lock_ptr = &hb2->lock;
1541	}
1542	get_futex_key_refs(key2);
1543	q->key = *key2;
1544}
1545
1546/**
1547 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1548 * @q:		the futex_q
1549 * @key:	the key of the requeue target futex
1550 * @hb:		the hash_bucket of the requeue target futex
1551 *
1552 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1553 * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1554 * to the requeue target futex so the waiter can detect the wakeup on the right
1555 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1556 * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1557 * to protect access to the pi_state to fixup the owner later.  Must be called
1558 * with both q->lock_ptr and hb->lock held.
1559 */
1560static inline
1561void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1562			   struct futex_hash_bucket *hb)
1563{
1564	get_futex_key_refs(key);
1565	q->key = *key;
1566
1567	__unqueue_futex(q);
1568
1569	WARN_ON(!q->rt_waiter);
1570	q->rt_waiter = NULL;
1571
1572	q->lock_ptr = &hb->lock;
1573
1574	wake_up_state(q->task, TASK_NORMAL);
1575}
1576
1577/**
1578 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1579 * @pifutex:		the user address of the to futex
1580 * @hb1:		the from futex hash bucket, must be locked by the caller
1581 * @hb2:		the to futex hash bucket, must be locked by the caller
1582 * @key1:		the from futex key
1583 * @key2:		the to futex key
1584 * @ps:			address to store the pi_state pointer
1585 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1586 *
1587 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1588 * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1589 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1590 * hb1 and hb2 must be held by the caller.
1591 *
1592 * Return:
1593 *  0 - failed to acquire the lock atomically;
1594 * >0 - acquired the lock, return value is vpid of the top_waiter
1595 * <0 - error
1596 */
1597static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1598				 struct futex_hash_bucket *hb1,
1599				 struct futex_hash_bucket *hb2,
1600				 union futex_key *key1, union futex_key *key2,
1601				 struct futex_pi_state **ps, int set_waiters)
1602{
1603	struct futex_q *top_waiter = NULL;
1604	u32 curval;
1605	int ret, vpid;
1606
1607	if (get_futex_value_locked(&curval, pifutex))
1608		return -EFAULT;
1609
1610	if (unlikely(should_fail_futex(true)))
1611		return -EFAULT;
1612
1613	/*
1614	 * Find the top_waiter and determine if there are additional waiters.
1615	 * If the caller intends to requeue more than 1 waiter to pifutex,
1616	 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1617	 * as we have means to handle the possible fault.  If not, don't set
1618	 * the bit unecessarily as it will force the subsequent unlock to enter
1619	 * the kernel.
1620	 */
1621	top_waiter = futex_top_waiter(hb1, key1);
1622
1623	/* There are no waiters, nothing for us to do. */
1624	if (!top_waiter)
1625		return 0;
1626
1627	/* Ensure we requeue to the expected futex. */
1628	if (!match_futex(top_waiter->requeue_pi_key, key2))
1629		return -EINVAL;
1630
1631	/*
1632	 * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1633	 * the contended case or if set_waiters is 1.  The pi_state is returned
1634	 * in ps in contended cases.
1635	 */
1636	vpid = task_pid_vnr(top_waiter->task);
1637	ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1638				   set_waiters);
1639	if (ret == 1) {
1640		requeue_pi_wake_futex(top_waiter, key2, hb2);
1641		return vpid;
1642	}
1643	return ret;
1644}
1645
1646/**
1647 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1648 * @uaddr1:	source futex user address
1649 * @flags:	futex flags (FLAGS_SHARED, etc.)
1650 * @uaddr2:	target futex user address
1651 * @nr_wake:	number of waiters to wake (must be 1 for requeue_pi)
1652 * @nr_requeue:	number of waiters to requeue (0-INT_MAX)
1653 * @cmpval:	@uaddr1 expected value (or %NULL)
1654 * @requeue_pi:	if we are attempting to requeue from a non-pi futex to a
1655 *		pi futex (pi to pi requeue is not supported)
1656 *
1657 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1658 * uaddr2 atomically on behalf of the top waiter.
1659 *
1660 * Return:
1661 * >=0 - on success, the number of tasks requeued or woken;
1662 *  <0 - on error
1663 */
1664static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1665			 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1666			 u32 *cmpval, int requeue_pi)
1667{
1668	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1669	int drop_count = 0, task_count = 0, ret;
1670	struct futex_pi_state *pi_state = NULL;
1671	struct futex_hash_bucket *hb1, *hb2;
1672	struct futex_q *this, *next;
1673	WAKE_Q(wake_q);
1674
1675	if (requeue_pi) {
1676		/*
1677		 * Requeue PI only works on two distinct uaddrs. This
1678		 * check is only valid for private futexes. See below.
1679		 */
1680		if (uaddr1 == uaddr2)
1681			return -EINVAL;
1682
1683		/*
1684		 * requeue_pi requires a pi_state, try to allocate it now
1685		 * without any locks in case it fails.
1686		 */
1687		if (refill_pi_state_cache())
1688			return -ENOMEM;
1689		/*
1690		 * requeue_pi must wake as many tasks as it can, up to nr_wake
1691		 * + nr_requeue, since it acquires the rt_mutex prior to
1692		 * returning to userspace, so as to not leave the rt_mutex with
1693		 * waiters and no owner.  However, second and third wake-ups
1694		 * cannot be predicted as they involve race conditions with the
1695		 * first wake and a fault while looking up the pi_state.  Both
1696		 * pthread_cond_signal() and pthread_cond_broadcast() should
1697		 * use nr_wake=1.
1698		 */
1699		if (nr_wake != 1)
1700			return -EINVAL;
1701	}
1702
1703retry:
1704	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1705	if (unlikely(ret != 0))
1706		goto out;
1707	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1708			    requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1709	if (unlikely(ret != 0))
1710		goto out_put_key1;
1711
1712	/*
1713	 * The check above which compares uaddrs is not sufficient for
1714	 * shared futexes. We need to compare the keys:
1715	 */
1716	if (requeue_pi && match_futex(&key1, &key2)) {
1717		ret = -EINVAL;
1718		goto out_put_keys;
1719	}
1720
1721	hb1 = hash_futex(&key1);
1722	hb2 = hash_futex(&key2);
1723
1724retry_private:
1725	hb_waiters_inc(hb2);
1726	double_lock_hb(hb1, hb2);
1727
1728	if (likely(cmpval != NULL)) {
1729		u32 curval;
1730
1731		ret = get_futex_value_locked(&curval, uaddr1);
1732
1733		if (unlikely(ret)) {
1734			double_unlock_hb(hb1, hb2);
1735			hb_waiters_dec(hb2);
1736
1737			ret = get_user(curval, uaddr1);
1738			if (ret)
1739				goto out_put_keys;
1740
1741			if (!(flags & FLAGS_SHARED))
1742				goto retry_private;
1743
1744			put_futex_key(&key2);
1745			put_futex_key(&key1);
1746			goto retry;
1747		}
1748		if (curval != *cmpval) {
1749			ret = -EAGAIN;
1750			goto out_unlock;
1751		}
1752	}
1753
1754	if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1755		/*
1756		 * Attempt to acquire uaddr2 and wake the top waiter. If we
1757		 * intend to requeue waiters, force setting the FUTEX_WAITERS
1758		 * bit.  We force this here where we are able to easily handle
1759		 * faults rather in the requeue loop below.
1760		 */
1761		ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1762						 &key2, &pi_state, nr_requeue);
1763
1764		/*
1765		 * At this point the top_waiter has either taken uaddr2 or is
1766		 * waiting on it.  If the former, then the pi_state will not
1767		 * exist yet, look it up one more time to ensure we have a
1768		 * reference to it. If the lock was taken, ret contains the
1769		 * vpid of the top waiter task.
1770		 * If the lock was not taken, we have pi_state and an initial
1771		 * refcount on it. In case of an error we have nothing.
1772		 */
1773		if (ret > 0) {
1774			WARN_ON(pi_state);
1775			drop_count++;
1776			task_count++;
1777			/*
1778			 * If we acquired the lock, then the user space value
1779			 * of uaddr2 should be vpid. It cannot be changed by
1780			 * the top waiter as it is blocked on hb2 lock if it
1781			 * tries to do so. If something fiddled with it behind
1782			 * our back the pi state lookup might unearth it. So
1783			 * we rather use the known value than rereading and
1784			 * handing potential crap to lookup_pi_state.
1785			 *
1786			 * If that call succeeds then we have pi_state and an
1787			 * initial refcount on it.
1788			 */
1789			ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1790		}
1791
1792		switch (ret) {
1793		case 0:
1794			/* We hold a reference on the pi state. */
1795			break;
1796
1797			/* If the above failed, then pi_state is NULL */
1798		case -EFAULT:
1799			double_unlock_hb(hb1, hb2);
1800			hb_waiters_dec(hb2);
1801			put_futex_key(&key2);
1802			put_futex_key(&key1);
1803			ret = fault_in_user_writeable(uaddr2);
1804			if (!ret)
1805				goto retry;
1806			goto out;
1807		case -EAGAIN:
1808			/*
1809			 * Two reasons for this:
1810			 * - Owner is exiting and we just wait for the
1811			 *   exit to complete.
1812			 * - The user space value changed.
1813			 */
1814			double_unlock_hb(hb1, hb2);
1815			hb_waiters_dec(hb2);
1816			put_futex_key(&key2);
1817			put_futex_key(&key1);
1818			cond_resched();
1819			goto retry;
1820		default:
1821			goto out_unlock;
1822		}
1823	}
1824
1825	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1826		if (task_count - nr_wake >= nr_requeue)
1827			break;
1828
1829		if (!match_futex(&this->key, &key1))
1830			continue;
1831
1832		/*
1833		 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1834		 * be paired with each other and no other futex ops.
1835		 *
1836		 * We should never be requeueing a futex_q with a pi_state,
1837		 * which is awaiting a futex_unlock_pi().
1838		 */
1839		if ((requeue_pi && !this->rt_waiter) ||
1840		    (!requeue_pi && this->rt_waiter) ||
1841		    this->pi_state) {
1842			ret = -EINVAL;
1843			break;
1844		}
1845
1846		/*
1847		 * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1848		 * lock, we already woke the top_waiter.  If not, it will be
1849		 * woken by futex_unlock_pi().
1850		 */
1851		if (++task_count <= nr_wake && !requeue_pi) {
1852			mark_wake_futex(&wake_q, this);
1853			continue;
1854		}
1855
1856		/* Ensure we requeue to the expected futex for requeue_pi. */
1857		if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1858			ret = -EINVAL;
1859			break;
1860		}
1861
1862		/*
1863		 * Requeue nr_requeue waiters and possibly one more in the case
1864		 * of requeue_pi if we couldn't acquire the lock atomically.
1865		 */
1866		if (requeue_pi) {
1867			/*
1868			 * Prepare the waiter to take the rt_mutex. Take a
1869			 * refcount on the pi_state and store the pointer in
1870			 * the futex_q object of the waiter.
1871			 */
1872			atomic_inc(&pi_state->refcount);
1873			this->pi_state = pi_state;
1874			ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1875							this->rt_waiter,
1876							this->task);
1877			if (ret == 1) {
1878				/*
1879				 * We got the lock. We do neither drop the
1880				 * refcount on pi_state nor clear
1881				 * this->pi_state because the waiter needs the
1882				 * pi_state for cleaning up the user space
1883				 * value. It will drop the refcount after
1884				 * doing so.
1885				 */
1886				requeue_pi_wake_futex(this, &key2, hb2);
1887				drop_count++;
1888				continue;
1889			} else if (ret) {
1890				/*
1891				 * rt_mutex_start_proxy_lock() detected a
1892				 * potential deadlock when we tried to queue
1893				 * that waiter. Drop the pi_state reference
1894				 * which we took above and remove the pointer
1895				 * to the state from the waiters futex_q
1896				 * object.
1897				 */
1898				this->pi_state = NULL;
1899				put_pi_state(pi_state);
1900				/*
1901				 * We stop queueing more waiters and let user
1902				 * space deal with the mess.
1903				 */
1904				break;
1905			}
1906		}
1907		requeue_futex(this, hb1, hb2, &key2);
1908		drop_count++;
1909	}
1910
1911	/*
1912	 * We took an extra initial reference to the pi_state either
1913	 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
1914	 * need to drop it here again.
1915	 */
1916	put_pi_state(pi_state);
1917
1918out_unlock:
1919	double_unlock_hb(hb1, hb2);
1920	wake_up_q(&wake_q);
1921	hb_waiters_dec(hb2);
1922
1923	/*
1924	 * drop_futex_key_refs() must be called outside the spinlocks. During
1925	 * the requeue we moved futex_q's from the hash bucket at key1 to the
1926	 * one at key2 and updated their key pointer.  We no longer need to
1927	 * hold the references to key1.
1928	 */
1929	while (--drop_count >= 0)
1930		drop_futex_key_refs(&key1);
1931
1932out_put_keys:
1933	put_futex_key(&key2);
1934out_put_key1:
1935	put_futex_key(&key1);
1936out:
1937	return ret ? ret : task_count;
1938}
1939
1940/* The key must be already stored in q->key. */
1941static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1942	__acquires(&hb->lock)
1943{
1944	struct futex_hash_bucket *hb;
1945
1946	hb = hash_futex(&q->key);
1947
1948	/*
1949	 * Increment the counter before taking the lock so that
1950	 * a potential waker won't miss a to-be-slept task that is
1951	 * waiting for the spinlock. This is safe as all queue_lock()
1952	 * users end up calling queue_me(). Similarly, for housekeeping,
1953	 * decrement the counter at queue_unlock() when some error has
1954	 * occurred and we don't end up adding the task to the list.
1955	 */
1956	hb_waiters_inc(hb);
1957
1958	q->lock_ptr = &hb->lock;
1959
1960	spin_lock(&hb->lock); /* implies smp_mb(); (A) */
1961	return hb;
1962}
1963
1964static inline void
1965queue_unlock(struct futex_hash_bucket *hb)
1966	__releases(&hb->lock)
1967{
1968	spin_unlock(&hb->lock);
1969	hb_waiters_dec(hb);
1970}
1971
1972/**
1973 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1974 * @q:	The futex_q to enqueue
1975 * @hb:	The destination hash bucket
1976 *
1977 * The hb->lock must be held by the caller, and is released here. A call to
1978 * queue_me() is typically paired with exactly one call to unqueue_me().  The
1979 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1980 * or nothing if the unqueue is done as part of the wake process and the unqueue
1981 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1982 * an example).
1983 */
1984static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1985	__releases(&hb->lock)
1986{
1987	int prio;
1988
1989	/*
1990	 * The priority used to register this element is
1991	 * - either the real thread-priority for the real-time threads
1992	 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1993	 * - or MAX_RT_PRIO for non-RT threads.
1994	 * Thus, all RT-threads are woken first in priority order, and
1995	 * the others are woken last, in FIFO order.
1996	 */
1997	prio = min(current->normal_prio, MAX_RT_PRIO);
1998
1999	plist_node_init(&q->list, prio);
2000	plist_add(&q->list, &hb->chain);
2001	q->task = current;
2002	spin_unlock(&hb->lock);
2003}
2004
2005/**
2006 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2007 * @q:	The futex_q to unqueue
2008 *
2009 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2010 * be paired with exactly one earlier call to queue_me().
2011 *
2012 * Return:
2013 *   1 - if the futex_q was still queued (and we removed unqueued it);
2014 *   0 - if the futex_q was already removed by the waking thread
2015 */
2016static int unqueue_me(struct futex_q *q)
2017{
2018	spinlock_t *lock_ptr;
2019	int ret = 0;
2020
2021	/* In the common case we don't take the spinlock, which is nice. */
2022retry:
2023	/*
2024	 * q->lock_ptr can change between this read and the following spin_lock.
2025	 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2026	 * optimizing lock_ptr out of the logic below.
2027	 */
2028	lock_ptr = READ_ONCE(q->lock_ptr);
2029	if (lock_ptr != NULL) {
2030		spin_lock(lock_ptr);
2031		/*
2032		 * q->lock_ptr can change between reading it and
2033		 * spin_lock(), causing us to take the wrong lock.  This
2034		 * corrects the race condition.
2035		 *
2036		 * Reasoning goes like this: if we have the wrong lock,
2037		 * q->lock_ptr must have changed (maybe several times)
2038		 * between reading it and the spin_lock().  It can
2039		 * change again after the spin_lock() but only if it was
2040		 * already changed before the spin_lock().  It cannot,
2041		 * however, change back to the original value.  Therefore
2042		 * we can detect whether we acquired the correct lock.
2043		 */
2044		if (unlikely(lock_ptr != q->lock_ptr)) {
2045			spin_unlock(lock_ptr);
2046			goto retry;
2047		}
2048		__unqueue_futex(q);
2049
2050		BUG_ON(q->pi_state);
2051
2052		spin_unlock(lock_ptr);
2053		ret = 1;
2054	}
2055
2056	drop_futex_key_refs(&q->key);
2057	return ret;
2058}
2059
2060/*
2061 * PI futexes can not be requeued and must remove themself from the
2062 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2063 * and dropped here.
2064 */
2065static void unqueue_me_pi(struct futex_q *q)
2066	__releases(q->lock_ptr)
2067{
2068	__unqueue_futex(q);
2069
2070	BUG_ON(!q->pi_state);
2071	put_pi_state(q->pi_state);
2072	q->pi_state = NULL;
2073
2074	spin_unlock(q->lock_ptr);
2075}
2076
2077/*
2078 * Fixup the pi_state owner with the new owner.
2079 *
2080 * Must be called with hash bucket lock held and mm->sem held for non
2081 * private futexes.
2082 */
2083static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2084				struct task_struct *newowner)
2085{
2086	u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2087	struct futex_pi_state *pi_state = q->pi_state;
2088	struct task_struct *oldowner = pi_state->owner;
2089	u32 uval, uninitialized_var(curval), newval;
2090	int ret;
2091
2092	/* Owner died? */
2093	if (!pi_state->owner)
2094		newtid |= FUTEX_OWNER_DIED;
2095
2096	/*
2097	 * We are here either because we stole the rtmutex from the
2098	 * previous highest priority waiter or we are the highest priority
2099	 * waiter but failed to get the rtmutex the first time.
2100	 * We have to replace the newowner TID in the user space variable.
2101	 * This must be atomic as we have to preserve the owner died bit here.
2102	 *
2103	 * Note: We write the user space value _before_ changing the pi_state
2104	 * because we can fault here. Imagine swapped out pages or a fork
2105	 * that marked all the anonymous memory readonly for cow.
2106	 *
2107	 * Modifying pi_state _before_ the user space value would
2108	 * leave the pi_state in an inconsistent state when we fault
2109	 * here, because we need to drop the hash bucket lock to
2110	 * handle the fault. This might be observed in the PID check
2111	 * in lookup_pi_state.
2112	 */
2113retry:
2114	if (get_futex_value_locked(&uval, uaddr))
2115		goto handle_fault;
2116
2117	while (1) {
2118		newval = (uval & FUTEX_OWNER_DIED) | newtid;
2119
2120		if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
2121			goto handle_fault;
2122		if (curval == uval)
2123			break;
2124		uval = curval;
2125	}
2126
2127	/*
2128	 * We fixed up user space. Now we need to fix the pi_state
2129	 * itself.
2130	 */
2131	if (pi_state->owner != NULL) {
2132		raw_spin_lock_irq(&pi_state->owner->pi_lock);
2133		WARN_ON(list_empty(&pi_state->list));
2134		list_del_init(&pi_state->list);
2135		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
2136	}
2137
2138	pi_state->owner = newowner;
2139
2140	raw_spin_lock_irq(&newowner->pi_lock);
2141	WARN_ON(!list_empty(&pi_state->list));
2142	list_add(&pi_state->list, &newowner->pi_state_list);
2143	raw_spin_unlock_irq(&newowner->pi_lock);
2144	return 0;
2145
2146	/*
2147	 * To handle the page fault we need to drop the hash bucket
2148	 * lock here. That gives the other task (either the highest priority
2149	 * waiter itself or the task which stole the rtmutex) the
2150	 * chance to try the fixup of the pi_state. So once we are
2151	 * back from handling the fault we need to check the pi_state
2152	 * after reacquiring the hash bucket lock and before trying to
2153	 * do another fixup. When the fixup has been done already we
2154	 * simply return.
2155	 */
2156handle_fault:
2157	spin_unlock(q->lock_ptr);
2158
2159	ret = fault_in_user_writeable(uaddr);
2160
2161	spin_lock(q->lock_ptr);
2162
2163	/*
2164	 * Check if someone else fixed it for us:
2165	 */
2166	if (pi_state->owner != oldowner)
2167		return 0;
2168
2169	if (ret)
2170		return ret;
2171
2172	goto retry;
2173}
2174
2175static long futex_wait_restart(struct restart_block *restart);
2176
2177/**
2178 * fixup_owner() - Post lock pi_state and corner case management
2179 * @uaddr:	user address of the futex
2180 * @q:		futex_q (contains pi_state and access to the rt_mutex)
2181 * @locked:	if the attempt to take the rt_mutex succeeded (1) or not (0)
2182 *
2183 * After attempting to lock an rt_mutex, this function is called to cleanup
2184 * the pi_state owner as well as handle race conditions that may allow us to
2185 * acquire the lock. Must be called with the hb lock held.
2186 *
2187 * Return:
2188 *  1 - success, lock taken;
2189 *  0 - success, lock not taken;
2190 * <0 - on error (-EFAULT)
2191 */
2192static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2193{
2194	struct task_struct *owner;
2195	int ret = 0;
2196
2197	if (locked) {
2198		/*
2199		 * Got the lock. We might not be the anticipated owner if we
2200		 * did a lock-steal - fix up the PI-state in that case:
2201		 */
2202		if (q->pi_state->owner != current)
2203			ret = fixup_pi_state_owner(uaddr, q, current);
2204		goto out;
2205	}
2206
2207	/*
2208	 * Catch the rare case, where the lock was released when we were on the
2209	 * way back before we locked the hash bucket.
2210	 */
2211	if (q->pi_state->owner == current) {
2212		/*
2213		 * Try to get the rt_mutex now. This might fail as some other
2214		 * task acquired the rt_mutex after we removed ourself from the
2215		 * rt_mutex waiters list.
2216		 */
2217		if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2218			locked = 1;
2219			goto out;
2220		}
2221
2222		/*
2223		 * pi_state is incorrect, some other task did a lock steal and
2224		 * we returned due to timeout or signal without taking the
2225		 * rt_mutex. Too late.
2226		 */
2227		raw_spin_lock_irq(&q->pi_state->pi_mutex.wait_lock);
2228		owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2229		if (!owner)
2230			owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2231		raw_spin_unlock_irq(&q->pi_state->pi_mutex.wait_lock);
2232		ret = fixup_pi_state_owner(uaddr, q, owner);
2233		goto out;
2234	}
2235
2236	/*
2237	 * Paranoia check. If we did not take the lock, then we should not be
2238	 * the owner of the rt_mutex.
2239	 */
2240	if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2241		printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2242				"pi-state %p\n", ret,
2243				q->pi_state->pi_mutex.owner,
2244				q->pi_state->owner);
2245
2246out:
2247	return ret ? ret : locked;
2248}
2249
2250/**
2251 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2252 * @hb:		the futex hash bucket, must be locked by the caller
2253 * @q:		the futex_q to queue up on
2254 * @timeout:	the prepared hrtimer_sleeper, or null for no timeout
2255 */
2256static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2257				struct hrtimer_sleeper *timeout)
2258{
2259	/*
2260	 * The task state is guaranteed to be set before another task can
2261	 * wake it. set_current_state() is implemented using smp_store_mb() and
2262	 * queue_me() calls spin_unlock() upon completion, both serializing
2263	 * access to the hash list and forcing another memory barrier.
2264	 */
2265	set_current_state(TASK_INTERRUPTIBLE);
2266	queue_me(q, hb);
2267
2268	/* Arm the timer */
2269	if (timeout)
2270		hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2271
2272	/*
2273	 * If we have been removed from the hash list, then another task
2274	 * has tried to wake us, and we can skip the call to schedule().
2275	 */
2276	if (likely(!plist_node_empty(&q->list))) {
2277		/*
2278		 * If the timer has already expired, current will already be
2279		 * flagged for rescheduling. Only call schedule if there
2280		 * is no timeout, or if it has yet to expire.
2281		 */
2282		if (!timeout || timeout->task)
2283			freezable_schedule();
2284	}
2285	__set_current_state(TASK_RUNNING);
2286}
2287
2288/**
2289 * futex_wait_setup() - Prepare to wait on a futex
2290 * @uaddr:	the futex userspace address
2291 * @val:	the expected value
2292 * @flags:	futex flags (FLAGS_SHARED, etc.)
2293 * @q:		the associated futex_q
2294 * @hb:		storage for hash_bucket pointer to be returned to caller
2295 *
2296 * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2297 * compare it with the expected value.  Handle atomic faults internally.
2298 * Return with the hb lock held and a q.key reference on success, and unlocked
2299 * with no q.key reference on failure.
2300 *
2301 * Return:
2302 *  0 - uaddr contains val and hb has been locked;
2303 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2304 */
2305static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2306			   struct futex_q *q, struct futex_hash_bucket **hb)
2307{
2308	u32 uval;
2309	int ret;
2310
2311	/*
2312	 * Access the page AFTER the hash-bucket is locked.
2313	 * Order is important:
2314	 *
2315	 *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2316	 *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2317	 *
2318	 * The basic logical guarantee of a futex is that it blocks ONLY
2319	 * if cond(var) is known to be true at the time of blocking, for
2320	 * any cond.  If we locked the hash-bucket after testing *uaddr, that
2321	 * would open a race condition where we could block indefinitely with
2322	 * cond(var) false, which would violate the guarantee.
2323	 *
2324	 * On the other hand, we insert q and release the hash-bucket only
2325	 * after testing *uaddr.  This guarantees that futex_wait() will NOT
2326	 * absorb a wakeup if *uaddr does not match the desired values
2327	 * while the syscall executes.
2328	 */
2329retry:
2330	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2331	if (unlikely(ret != 0))
2332		return ret;
2333
2334retry_private:
2335	*hb = queue_lock(q);
2336
2337	ret = get_futex_value_locked(&uval, uaddr);
2338
2339	if (ret) {
2340		queue_unlock(*hb);
2341
2342		ret = get_user(uval, uaddr);
2343		if (ret)
2344			goto out;
2345
2346		if (!(flags & FLAGS_SHARED))
2347			goto retry_private;
2348
2349		put_futex_key(&q->key);
2350		goto retry;
2351	}
2352
2353	if (uval != val) {
2354		queue_unlock(*hb);
2355		ret = -EWOULDBLOCK;
2356	}
2357
2358out:
2359	if (ret)
2360		put_futex_key(&q->key);
2361	return ret;
2362}
2363
2364static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2365		      ktime_t *abs_time, u32 bitset)
2366{
2367	struct hrtimer_sleeper timeout, *to = NULL;
2368	struct restart_block *restart;
2369	struct futex_hash_bucket *hb;
2370	struct futex_q q = futex_q_init;
2371	int ret;
2372
2373	if (!bitset)
2374		return -EINVAL;
2375	q.bitset = bitset;
2376
2377	if (abs_time) {
2378		to = &timeout;
2379
2380		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2381				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2382				      HRTIMER_MODE_ABS);
2383		hrtimer_init_sleeper(to, current);
2384		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2385					     current->timer_slack_ns);
2386	}
2387
2388retry:
2389	/*
2390	 * Prepare to wait on uaddr. On success, holds hb lock and increments
2391	 * q.key refs.
2392	 */
2393	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2394	if (ret)
2395		goto out;
2396
2397	/* queue_me and wait for wakeup, timeout, or a signal. */
2398	futex_wait_queue_me(hb, &q, to);
2399
2400	/* If we were woken (and unqueued), we succeeded, whatever. */
2401	ret = 0;
2402	/* unqueue_me() drops q.key ref */
2403	if (!unqueue_me(&q))
2404		goto out;
2405	ret = -ETIMEDOUT;
2406	if (to && !to->task)
2407		goto out;
2408
2409	/*
2410	 * We expect signal_pending(current), but we might be the
2411	 * victim of a spurious wakeup as well.
2412	 */
2413	if (!signal_pending(current))
2414		goto retry;
2415
2416	ret = -ERESTARTSYS;
2417	if (!abs_time)
2418		goto out;
2419
2420	restart = &current->restart_block;
2421	restart->fn = futex_wait_restart;
2422	restart->futex.uaddr = uaddr;
2423	restart->futex.val = val;
2424	restart->futex.time = abs_time->tv64;
2425	restart->futex.bitset = bitset;
2426	restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2427
2428	ret = -ERESTART_RESTARTBLOCK;
2429
2430out:
2431	if (to) {
2432		hrtimer_cancel(&to->timer);
2433		destroy_hrtimer_on_stack(&to->timer);
2434	}
2435	return ret;
2436}
2437
2438
2439static long futex_wait_restart(struct restart_block *restart)
2440{
2441	u32 __user *uaddr = restart->futex.uaddr;
2442	ktime_t t, *tp = NULL;
2443
2444	if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2445		t.tv64 = restart->futex.time;
2446		tp = &t;
2447	}
2448	restart->fn = do_no_restart_syscall;
2449
2450	return (long)futex_wait(uaddr, restart->futex.flags,
2451				restart->futex.val, tp, restart->futex.bitset);
2452}
2453
2454
2455/*
2456 * Userspace tried a 0 -> TID atomic transition of the futex value
2457 * and failed. The kernel side here does the whole locking operation:
2458 * if there are waiters then it will block as a consequence of relying
2459 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2460 * a 0 value of the futex too.).
2461 *
2462 * Also serves as futex trylock_pi()'ing, and due semantics.
2463 */
2464static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2465			 ktime_t *time, int trylock)
2466{
2467	struct hrtimer_sleeper timeout, *to = NULL;
2468	struct futex_hash_bucket *hb;
2469	struct futex_q q = futex_q_init;
2470	int res, ret;
2471
2472	if (refill_pi_state_cache())
2473		return -ENOMEM;
2474
2475	if (time) {
2476		to = &timeout;
2477		hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2478				      HRTIMER_MODE_ABS);
2479		hrtimer_init_sleeper(to, current);
2480		hrtimer_set_expires(&to->timer, *time);
2481	}
2482
2483retry:
2484	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2485	if (unlikely(ret != 0))
2486		goto out;
2487
2488retry_private:
2489	hb = queue_lock(&q);
2490
2491	ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2492	if (unlikely(ret)) {
2493		/*
2494		 * Atomic work succeeded and we got the lock,
2495		 * or failed. Either way, we do _not_ block.
2496		 */
2497		switch (ret) {
2498		case 1:
2499			/* We got the lock. */
2500			ret = 0;
2501			goto out_unlock_put_key;
2502		case -EFAULT:
2503			goto uaddr_faulted;
2504		case -EAGAIN:
2505			/*
2506			 * Two reasons for this:
2507			 * - Task is exiting and we just wait for the
2508			 *   exit to complete.
2509			 * - The user space value changed.
2510			 */
2511			queue_unlock(hb);
2512			put_futex_key(&q.key);
2513			cond_resched();
2514			goto retry;
2515		default:
2516			goto out_unlock_put_key;
2517		}
2518	}
2519
2520	/*
2521	 * Only actually queue now that the atomic ops are done:
2522	 */
2523	queue_me(&q, hb);
2524
2525	WARN_ON(!q.pi_state);
2526	/*
2527	 * Block on the PI mutex:
2528	 */
2529	if (!trylock) {
2530		ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2531	} else {
2532		ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2533		/* Fixup the trylock return value: */
2534		ret = ret ? 0 : -EWOULDBLOCK;
2535	}
2536
2537	spin_lock(q.lock_ptr);
2538	/*
2539	 * Fixup the pi_state owner and possibly acquire the lock if we
2540	 * haven't already.
2541	 */
2542	res = fixup_owner(uaddr, &q, !ret);
2543	/*
2544	 * If fixup_owner() returned an error, proprogate that.  If it acquired
2545	 * the lock, clear our -ETIMEDOUT or -EINTR.
2546	 */
2547	if (res)
2548		ret = (res < 0) ? res : 0;
2549
2550	/*
2551	 * If fixup_owner() faulted and was unable to handle the fault, unlock
2552	 * it and return the fault to userspace.
2553	 */
2554	if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2555		rt_mutex_unlock(&q.pi_state->pi_mutex);
2556
2557	/* Unqueue and drop the lock */
2558	unqueue_me_pi(&q);
2559
2560	goto out_put_key;
2561
2562out_unlock_put_key:
2563	queue_unlock(hb);
2564
2565out_put_key:
2566	put_futex_key(&q.key);
2567out:
2568	if (to)
2569		destroy_hrtimer_on_stack(&to->timer);
2570	return ret != -EINTR ? ret : -ERESTARTNOINTR;
2571
2572uaddr_faulted:
2573	queue_unlock(hb);
2574
2575	ret = fault_in_user_writeable(uaddr);
2576	if (ret)
2577		goto out_put_key;
2578
2579	if (!(flags & FLAGS_SHARED))
2580		goto retry_private;
2581
2582	put_futex_key(&q.key);
2583	goto retry;
2584}
2585
2586/*
2587 * Userspace attempted a TID -> 0 atomic transition, and failed.
2588 * This is the in-kernel slowpath: we look up the PI state (if any),
2589 * and do the rt-mutex unlock.
2590 */
2591static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2592{
2593	u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2594	union futex_key key = FUTEX_KEY_INIT;
2595	struct futex_hash_bucket *hb;
2596	struct futex_q *match;
2597	int ret;
2598
2599retry:
2600	if (get_user(uval, uaddr))
2601		return -EFAULT;
2602	/*
2603	 * We release only a lock we actually own:
2604	 */
2605	if ((uval & FUTEX_TID_MASK) != vpid)
2606		return -EPERM;
2607
2608	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2609	if (ret)
2610		return ret;
2611
2612	hb = hash_futex(&key);
2613	spin_lock(&hb->lock);
2614
2615	/*
2616	 * Check waiters first. We do not trust user space values at
2617	 * all and we at least want to know if user space fiddled
2618	 * with the futex value instead of blindly unlocking.
2619	 */
2620	match = futex_top_waiter(hb, &key);
2621	if (match) {
2622		ret = wake_futex_pi(uaddr, uval, match, hb);
2623		/*
2624		 * In case of success wake_futex_pi dropped the hash
2625		 * bucket lock.
2626		 */
2627		if (!ret)
2628			goto out_putkey;
2629		/*
2630		 * The atomic access to the futex value generated a
2631		 * pagefault, so retry the user-access and the wakeup:
2632		 */
2633		if (ret == -EFAULT)
2634			goto pi_faulted;
2635		/*
2636		 * A unconditional UNLOCK_PI op raced against a waiter
2637		 * setting the FUTEX_WAITERS bit. Try again.
2638		 */
2639		if (ret == -EAGAIN) {
2640			spin_unlock(&hb->lock);
2641			put_futex_key(&key);
2642			goto retry;
2643		}
2644		/*
2645		 * wake_futex_pi has detected invalid state. Tell user
2646		 * space.
2647		 */
2648		goto out_unlock;
2649	}
2650
2651	/*
2652	 * We have no kernel internal state, i.e. no waiters in the
2653	 * kernel. Waiters which are about to queue themselves are stuck
2654	 * on hb->lock. So we can safely ignore them. We do neither
2655	 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2656	 * owner.
2657	 */
2658	if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2659		goto pi_faulted;
2660
2661	/*
2662	 * If uval has changed, let user space handle it.
2663	 */
2664	ret = (curval == uval) ? 0 : -EAGAIN;
2665
2666out_unlock:
2667	spin_unlock(&hb->lock);
2668out_putkey:
2669	put_futex_key(&key);
2670	return ret;
2671
2672pi_faulted:
2673	spin_unlock(&hb->lock);
2674	put_futex_key(&key);
2675
2676	ret = fault_in_user_writeable(uaddr);
2677	if (!ret)
2678		goto retry;
2679
2680	return ret;
2681}
2682
2683/**
2684 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2685 * @hb:		the hash_bucket futex_q was original enqueued on
2686 * @q:		the futex_q woken while waiting to be requeued
2687 * @key2:	the futex_key of the requeue target futex
2688 * @timeout:	the timeout associated with the wait (NULL if none)
2689 *
2690 * Detect if the task was woken on the initial futex as opposed to the requeue
2691 * target futex.  If so, determine if it was a timeout or a signal that caused
2692 * the wakeup and return the appropriate error code to the caller.  Must be
2693 * called with the hb lock held.
2694 *
2695 * Return:
2696 *  0 = no early wakeup detected;
2697 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2698 */
2699static inline
2700int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2701				   struct futex_q *q, union futex_key *key2,
2702				   struct hrtimer_sleeper *timeout)
2703{
2704	int ret = 0;
2705
2706	/*
2707	 * With the hb lock held, we avoid races while we process the wakeup.
2708	 * We only need to hold hb (and not hb2) to ensure atomicity as the
2709	 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2710	 * It can't be requeued from uaddr2 to something else since we don't
2711	 * support a PI aware source futex for requeue.
2712	 */
2713	if (!match_futex(&q->key, key2)) {
2714		WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2715		/*
2716		 * We were woken prior to requeue by a timeout or a signal.
2717		 * Unqueue the futex_q and determine which it was.
2718		 */
2719		plist_del(&q->list, &hb->chain);
2720		hb_waiters_dec(hb);
2721
2722		/* Handle spurious wakeups gracefully */
2723		ret = -EWOULDBLOCK;
2724		if (timeout && !timeout->task)
2725			ret = -ETIMEDOUT;
2726		else if (signal_pending(current))
2727			ret = -ERESTARTNOINTR;
2728	}
2729	return ret;
2730}
2731
2732/**
2733 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2734 * @uaddr:	the futex we initially wait on (non-pi)
2735 * @flags:	futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2736 *		the same type, no requeueing from private to shared, etc.
2737 * @val:	the expected value of uaddr
2738 * @abs_time:	absolute timeout
2739 * @bitset:	32 bit wakeup bitset set by userspace, defaults to all
2740 * @uaddr2:	the pi futex we will take prior to returning to user-space
2741 *
2742 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2743 * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2744 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2745 * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2746 * without one, the pi logic would not know which task to boost/deboost, if
2747 * there was a need to.
2748 *
2749 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2750 * via the following--
2751 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2752 * 2) wakeup on uaddr2 after a requeue
2753 * 3) signal
2754 * 4) timeout
2755 *
2756 * If 3, cleanup and return -ERESTARTNOINTR.
2757 *
2758 * If 2, we may then block on trying to take the rt_mutex and return via:
2759 * 5) successful lock
2760 * 6) signal
2761 * 7) timeout
2762 * 8) other lock acquisition failure
2763 *
2764 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2765 *
2766 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2767 *
2768 * Return:
2769 *  0 - On success;
2770 * <0 - On error
2771 */
2772static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2773				 u32 val, ktime_t *abs_time, u32 bitset,
2774				 u32 __user *uaddr2)
2775{
2776	struct hrtimer_sleeper timeout, *to = NULL;
2777	struct rt_mutex_waiter rt_waiter;
2778	struct rt_mutex *pi_mutex = NULL;
2779	struct futex_hash_bucket *hb;
2780	union futex_key key2 = FUTEX_KEY_INIT;
2781	struct futex_q q = futex_q_init;
2782	int res, ret;
2783
2784	if (uaddr == uaddr2)
2785		return -EINVAL;
2786
2787	if (!bitset)
2788		return -EINVAL;
2789
2790	if (abs_time) {
2791		to = &timeout;
2792		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2793				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2794				      HRTIMER_MODE_ABS);
2795		hrtimer_init_sleeper(to, current);
2796		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2797					     current->timer_slack_ns);
2798	}
2799
2800	/*
2801	 * The waiter is allocated on our stack, manipulated by the requeue
2802	 * code while we sleep on uaddr.
2803	 */
2804	debug_rt_mutex_init_waiter(&rt_waiter);
2805	RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2806	RB_CLEAR_NODE(&rt_waiter.tree_entry);
2807	rt_waiter.task = NULL;
2808
2809	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2810	if (unlikely(ret != 0))
2811		goto out;
2812
2813	q.bitset = bitset;
2814	q.rt_waiter = &rt_waiter;
2815	q.requeue_pi_key = &key2;
2816
2817	/*
2818	 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2819	 * count.
2820	 */
2821	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2822	if (ret)
2823		goto out_key2;
2824
2825	/*
2826	 * The check above which compares uaddrs is not sufficient for
2827	 * shared futexes. We need to compare the keys:
2828	 */
2829	if (match_futex(&q.key, &key2)) {
2830		queue_unlock(hb);
2831		ret = -EINVAL;
2832		goto out_put_keys;
2833	}
2834
2835	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
2836	futex_wait_queue_me(hb, &q, to);
2837
2838	spin_lock(&hb->lock);
2839	ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2840	spin_unlock(&hb->lock);
2841	if (ret)
2842		goto out_put_keys;
2843
2844	/*
2845	 * In order for us to be here, we know our q.key == key2, and since
2846	 * we took the hb->lock above, we also know that futex_requeue() has
2847	 * completed and we no longer have to concern ourselves with a wakeup
2848	 * race with the atomic proxy lock acquisition by the requeue code. The
2849	 * futex_requeue dropped our key1 reference and incremented our key2
2850	 * reference count.
2851	 */
2852
2853	/* Check if the requeue code acquired the second futex for us. */
2854	if (!q.rt_waiter) {
2855		/*
2856		 * Got the lock. We might not be the anticipated owner if we
2857		 * did a lock-steal - fix up the PI-state in that case.
2858		 */
2859		if (q.pi_state && (q.pi_state->owner != current)) {
2860			spin_lock(q.lock_ptr);
2861			ret = fixup_pi_state_owner(uaddr2, &q, current);
 
 
2862			/*
2863			 * Drop the reference to the pi state which
2864			 * the requeue_pi() code acquired for us.
2865			 */
2866			put_pi_state(q.pi_state);
2867			spin_unlock(q.lock_ptr);
2868		}
2869	} else {
 
 
2870		/*
2871		 * We have been woken up by futex_unlock_pi(), a timeout, or a
2872		 * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2873		 * the pi_state.
2874		 */
2875		WARN_ON(!q.pi_state);
2876		pi_mutex = &q.pi_state->pi_mutex;
2877		ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2878		debug_rt_mutex_free_waiter(&rt_waiter);
2879
2880		spin_lock(q.lock_ptr);
2881		/*
2882		 * Fixup the pi_state owner and possibly acquire the lock if we
2883		 * haven't already.
2884		 */
2885		res = fixup_owner(uaddr2, &q, !ret);
2886		/*
2887		 * If fixup_owner() returned an error, proprogate that.  If it
2888		 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2889		 */
2890		if (res)
2891			ret = (res < 0) ? res : 0;
2892
 
 
 
 
 
 
 
 
2893		/* Unqueue and drop the lock. */
2894		unqueue_me_pi(&q);
2895	}
2896
2897	/*
2898	 * If fixup_pi_state_owner() faulted and was unable to handle the
2899	 * fault, unlock the rt_mutex and return the fault to userspace.
2900	 */
2901	if (ret == -EFAULT) {
2902		if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2903			rt_mutex_unlock(pi_mutex);
2904	} else if (ret == -EINTR) {
2905		/*
2906		 * We've already been requeued, but cannot restart by calling
2907		 * futex_lock_pi() directly. We could restart this syscall, but
2908		 * it would detect that the user space "val" changed and return
2909		 * -EWOULDBLOCK.  Save the overhead of the restart and return
2910		 * -EWOULDBLOCK directly.
2911		 */
2912		ret = -EWOULDBLOCK;
2913	}
2914
2915out_put_keys:
2916	put_futex_key(&q.key);
2917out_key2:
2918	put_futex_key(&key2);
2919
2920out:
2921	if (to) {
2922		hrtimer_cancel(&to->timer);
2923		destroy_hrtimer_on_stack(&to->timer);
2924	}
2925	return ret;
2926}
2927
2928/*
2929 * Support for robust futexes: the kernel cleans up held futexes at
2930 * thread exit time.
2931 *
2932 * Implementation: user-space maintains a per-thread list of locks it
2933 * is holding. Upon do_exit(), the kernel carefully walks this list,
2934 * and marks all locks that are owned by this thread with the
2935 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2936 * always manipulated with the lock held, so the list is private and
2937 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2938 * field, to allow the kernel to clean up if the thread dies after
2939 * acquiring the lock, but just before it could have added itself to
2940 * the list. There can only be one such pending lock.
2941 */
2942
2943/**
2944 * sys_set_robust_list() - Set the robust-futex list head of a task
2945 * @head:	pointer to the list-head
2946 * @len:	length of the list-head, as userspace expects
2947 */
2948SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2949		size_t, len)
2950{
2951	if (!futex_cmpxchg_enabled)
2952		return -ENOSYS;
2953	/*
2954	 * The kernel knows only one size for now:
2955	 */
2956	if (unlikely(len != sizeof(*head)))
2957		return -EINVAL;
2958
2959	current->robust_list = head;
2960
2961	return 0;
2962}
2963
2964/**
2965 * sys_get_robust_list() - Get the robust-futex list head of a task
2966 * @pid:	pid of the process [zero for current task]
2967 * @head_ptr:	pointer to a list-head pointer, the kernel fills it in
2968 * @len_ptr:	pointer to a length field, the kernel fills in the header size
2969 */
2970SYSCALL_DEFINE3(get_robust_list, int, pid,
2971		struct robust_list_head __user * __user *, head_ptr,
2972		size_t __user *, len_ptr)
2973{
2974	struct robust_list_head __user *head;
2975	unsigned long ret;
2976	struct task_struct *p;
2977
2978	if (!futex_cmpxchg_enabled)
2979		return -ENOSYS;
2980
2981	rcu_read_lock();
2982
2983	ret = -ESRCH;
2984	if (!pid)
2985		p = current;
2986	else {
2987		p = find_task_by_vpid(pid);
2988		if (!p)
2989			goto err_unlock;
2990	}
2991
2992	ret = -EPERM;
2993	if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
2994		goto err_unlock;
2995
2996	head = p->robust_list;
2997	rcu_read_unlock();
2998
2999	if (put_user(sizeof(*head), len_ptr))
3000		return -EFAULT;
3001	return put_user(head, head_ptr);
3002
3003err_unlock:
3004	rcu_read_unlock();
3005
3006	return ret;
3007}
3008
3009/*
3010 * Process a futex-list entry, check whether it's owned by the
3011 * dying task, and do notification if so:
3012 */
3013int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
3014{
3015	u32 uval, uninitialized_var(nval), mval;
3016
3017retry:
3018	if (get_user(uval, uaddr))
3019		return -1;
3020
3021	if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
3022		/*
3023		 * Ok, this dying thread is truly holding a futex
3024		 * of interest. Set the OWNER_DIED bit atomically
3025		 * via cmpxchg, and if the value had FUTEX_WAITERS
3026		 * set, wake up a waiter (if any). (We have to do a
3027		 * futex_wake() even if OWNER_DIED is already set -
3028		 * to handle the rare but possible case of recursive
3029		 * thread-death.) The rest of the cleanup is done in
3030		 * userspace.
3031		 */
3032		mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3033		/*
3034		 * We are not holding a lock here, but we want to have
3035		 * the pagefault_disable/enable() protection because
3036		 * we want to handle the fault gracefully. If the
3037		 * access fails we try to fault in the futex with R/W
3038		 * verification via get_user_pages. get_user() above
3039		 * does not guarantee R/W access. If that fails we
3040		 * give up and leave the futex locked.
3041		 */
3042		if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3043			if (fault_in_user_writeable(uaddr))
3044				return -1;
3045			goto retry;
3046		}
3047		if (nval != uval)
3048			goto retry;
3049
3050		/*
3051		 * Wake robust non-PI futexes here. The wakeup of
3052		 * PI futexes happens in exit_pi_state():
3053		 */
3054		if (!pi && (uval & FUTEX_WAITERS))
3055			futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3056	}
3057	return 0;
3058}
3059
3060/*
3061 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3062 */
3063static inline int fetch_robust_entry(struct robust_list __user **entry,
3064				     struct robust_list __user * __user *head,
3065				     unsigned int *pi)
3066{
3067	unsigned long uentry;
3068
3069	if (get_user(uentry, (unsigned long __user *)head))
3070		return -EFAULT;
3071
3072	*entry = (void __user *)(uentry & ~1UL);
3073	*pi = uentry & 1;
3074
3075	return 0;
3076}
3077
3078/*
3079 * Walk curr->robust_list (very carefully, it's a userspace list!)
3080 * and mark any locks found there dead, and notify any waiters.
3081 *
3082 * We silently return on any sign of list-walking problem.
3083 */
3084void exit_robust_list(struct task_struct *curr)
3085{
3086	struct robust_list_head __user *head = curr->robust_list;
3087	struct robust_list __user *entry, *next_entry, *pending;
3088	unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3089	unsigned int uninitialized_var(next_pi);
3090	unsigned long futex_offset;
3091	int rc;
3092
3093	if (!futex_cmpxchg_enabled)
3094		return;
3095
3096	/*
3097	 * Fetch the list head (which was registered earlier, via
3098	 * sys_set_robust_list()):
3099	 */
3100	if (fetch_robust_entry(&entry, &head->list.next, &pi))
3101		return;
3102	/*
3103	 * Fetch the relative futex offset:
3104	 */
3105	if (get_user(futex_offset, &head->futex_offset))
3106		return;
3107	/*
3108	 * Fetch any possibly pending lock-add first, and handle it
3109	 * if it exists:
3110	 */
3111	if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3112		return;
3113
3114	next_entry = NULL;	/* avoid warning with gcc */
3115	while (entry != &head->list) {
3116		/*
3117		 * Fetch the next entry in the list before calling
3118		 * handle_futex_death:
3119		 */
3120		rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3121		/*
3122		 * A pending lock might already be on the list, so
3123		 * don't process it twice:
3124		 */
3125		if (entry != pending)
3126			if (handle_futex_death((void __user *)entry + futex_offset,
3127						curr, pi))
3128				return;
3129		if (rc)
3130			return;
3131		entry = next_entry;
3132		pi = next_pi;
3133		/*
3134		 * Avoid excessively long or circular lists:
3135		 */
3136		if (!--limit)
3137			break;
3138
3139		cond_resched();
3140	}
3141
3142	if (pending)
3143		handle_futex_death((void __user *)pending + futex_offset,
3144				   curr, pip);
3145}
3146
3147long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3148		u32 __user *uaddr2, u32 val2, u32 val3)
3149{
3150	int cmd = op & FUTEX_CMD_MASK;
3151	unsigned int flags = 0;
3152
3153	if (!(op & FUTEX_PRIVATE_FLAG))
3154		flags |= FLAGS_SHARED;
3155
3156	if (op & FUTEX_CLOCK_REALTIME) {
3157		flags |= FLAGS_CLOCKRT;
3158		if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3159		    cmd != FUTEX_WAIT_REQUEUE_PI)
3160			return -ENOSYS;
3161	}
3162
3163	switch (cmd) {
3164	case FUTEX_LOCK_PI:
3165	case FUTEX_UNLOCK_PI:
3166	case FUTEX_TRYLOCK_PI:
3167	case FUTEX_WAIT_REQUEUE_PI:
3168	case FUTEX_CMP_REQUEUE_PI:
3169		if (!futex_cmpxchg_enabled)
3170			return -ENOSYS;
3171	}
3172
3173	switch (cmd) {
3174	case FUTEX_WAIT:
3175		val3 = FUTEX_BITSET_MATCH_ANY;
3176	case FUTEX_WAIT_BITSET:
3177		return futex_wait(uaddr, flags, val, timeout, val3);
3178	case FUTEX_WAKE:
3179		val3 = FUTEX_BITSET_MATCH_ANY;
3180	case FUTEX_WAKE_BITSET:
3181		return futex_wake(uaddr, flags, val, val3);
3182	case FUTEX_REQUEUE:
3183		return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3184	case FUTEX_CMP_REQUEUE:
3185		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3186	case FUTEX_WAKE_OP:
3187		return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3188	case FUTEX_LOCK_PI:
3189		return futex_lock_pi(uaddr, flags, timeout, 0);
3190	case FUTEX_UNLOCK_PI:
3191		return futex_unlock_pi(uaddr, flags);
3192	case FUTEX_TRYLOCK_PI:
3193		return futex_lock_pi(uaddr, flags, NULL, 1);
3194	case FUTEX_WAIT_REQUEUE_PI:
3195		val3 = FUTEX_BITSET_MATCH_ANY;
3196		return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3197					     uaddr2);
3198	case FUTEX_CMP_REQUEUE_PI:
3199		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3200	}
3201	return -ENOSYS;
3202}
3203
3204
3205SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3206		struct timespec __user *, utime, u32 __user *, uaddr2,
3207		u32, val3)
3208{
3209	struct timespec ts;
3210	ktime_t t, *tp = NULL;
3211	u32 val2 = 0;
3212	int cmd = op & FUTEX_CMD_MASK;
3213
3214	if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3215		      cmd == FUTEX_WAIT_BITSET ||
3216		      cmd == FUTEX_WAIT_REQUEUE_PI)) {
3217		if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3218			return -EFAULT;
3219		if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3220			return -EFAULT;
3221		if (!timespec_valid(&ts))
3222			return -EINVAL;
3223
3224		t = timespec_to_ktime(ts);
3225		if (cmd == FUTEX_WAIT)
3226			t = ktime_add_safe(ktime_get(), t);
3227		tp = &t;
3228	}
3229	/*
3230	 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3231	 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3232	 */
3233	if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3234	    cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3235		val2 = (u32) (unsigned long) utime;
3236
3237	return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3238}
3239
3240static void __init futex_detect_cmpxchg(void)
3241{
3242#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3243	u32 curval;
3244
3245	/*
3246	 * This will fail and we want it. Some arch implementations do
3247	 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3248	 * functionality. We want to know that before we call in any
3249	 * of the complex code paths. Also we want to prevent
3250	 * registration of robust lists in that case. NULL is
3251	 * guaranteed to fault and we get -EFAULT on functional
3252	 * implementation, the non-functional ones will return
3253	 * -ENOSYS.
3254	 */
3255	if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3256		futex_cmpxchg_enabled = 1;
3257#endif
3258}
3259
3260static int __init futex_init(void)
3261{
3262	unsigned int futex_shift;
3263	unsigned long i;
3264
3265#if CONFIG_BASE_SMALL
3266	futex_hashsize = 16;
3267#else
3268	futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3269#endif
3270
3271	futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3272					       futex_hashsize, 0,
3273					       futex_hashsize < 256 ? HASH_SMALL : 0,
3274					       &futex_shift, NULL,
3275					       futex_hashsize, futex_hashsize);
3276	futex_hashsize = 1UL << futex_shift;
3277
3278	futex_detect_cmpxchg();
3279
3280	for (i = 0; i < futex_hashsize; i++) {
3281		atomic_set(&futex_queues[i].waiters, 0);
3282		plist_head_init(&futex_queues[i].chain);
3283		spin_lock_init(&futex_queues[i].lock);
3284	}
3285
3286	return 0;
3287}
3288__initcall(futex_init);