Linux Audio

Check our new training course

Loading...
v3.15
 
   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
  68#include <asm/futex.h>
  69
  70#include "locking/rtmutex_common.h"
  71
  72/*
  73 * READ this before attempting to hack on futexes!
  74 *
  75 * Basic futex operation and ordering guarantees
  76 * =============================================
  77 *
  78 * The waiter reads the futex value in user space and calls
  79 * futex_wait(). This function computes the hash bucket and acquires
  80 * the hash bucket lock. After that it reads the futex user space value
  81 * again and verifies that the data has not changed. If it has not changed
  82 * it enqueues itself into the hash bucket, releases the hash bucket lock
  83 * and schedules.
  84 *
  85 * The waker side modifies the user space value of the futex and calls
  86 * futex_wake(). This function computes the hash bucket and acquires the
  87 * hash bucket lock. Then it looks for waiters on that futex in the hash
  88 * bucket and wakes them.
  89 *
  90 * In futex wake up scenarios where no tasks are blocked on a futex, taking
  91 * the hb spinlock can be avoided and simply return. In order for this
  92 * optimization to work, ordering guarantees must exist so that the waiter
  93 * being added to the list is acknowledged when the list is concurrently being
  94 * checked by the waker, avoiding scenarios like the following:
  95 *
  96 * CPU 0                               CPU 1
  97 * val = *futex;
  98 * sys_futex(WAIT, futex, val);
  99 *   futex_wait(futex, val);
 100 *   uval = *futex;
 101 *                                     *futex = newval;
 102 *                                     sys_futex(WAKE, futex);
 103 *                                       futex_wake(futex);
 104 *                                       if (queue_empty())
 105 *                                         return;
 106 *   if (uval == val)
 107 *      lock(hash_bucket(futex));
 108 *      queue();
 109 *     unlock(hash_bucket(futex));
 110 *     schedule();
 111 *
 112 * This would cause the waiter on CPU 0 to wait forever because it
 113 * missed the transition of the user space value from val to newval
 114 * and the waker did not find the waiter in the hash bucket queue.
 115 *
 116 * The correct serialization ensures that a waiter either observes
 117 * the changed user space value before blocking or is woken by a
 118 * concurrent waker:
 119 *
 120 * CPU 0                                 CPU 1
 121 * val = *futex;
 122 * sys_futex(WAIT, futex, val);
 123 *   futex_wait(futex, val);
 124 *
 125 *   waiters++; (a)
 126 *   mb(); (A) <-- paired with -.
 127 *                              |
 128 *   lock(hash_bucket(futex));  |
 129 *                              |
 130 *   uval = *futex;             |
 131 *                              |        *futex = newval;
 132 *                              |        sys_futex(WAKE, futex);
 133 *                              |          futex_wake(futex);
 134 *                              |
 135 *                              `------->  mb(); (B)
 136 *   if (uval == val)
 137 *     queue();
 138 *     unlock(hash_bucket(futex));
 139 *     schedule();                         if (waiters)
 140 *                                           lock(hash_bucket(futex));
 141 *   else                                    wake_waiters(futex);
 142 *     waiters--; (b)                        unlock(hash_bucket(futex));
 143 *
 144 * Where (A) orders the waiters increment and the futex value read through
 145 * atomic operations (see hb_waiters_inc) and where (B) orders the write
 146 * to futex and the waiters read -- this is done by the barriers in
 147 * get_futex_key_refs(), through either ihold or atomic_inc, depending on the
 148 * futex type.
 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
 258static unsigned long __read_mostly futex_hashsize;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 259
 260static struct futex_hash_bucket *futex_queues;
 
 
 
 
 261
 262static inline void futex_get_mm(union futex_key *key)
 263{
 264	atomic_inc(&key->private.mm->mm_count);
 265	/*
 266	 * Ensure futex_get_mm() implies a full barrier such that
 267	 * get_futex_key() implies a full barrier. This is relied upon
 268	 * as full barrier (B), see the ordering comment above.
 269	 */
 270	smp_mb__after_atomic_inc();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 271}
 
 
 
 
 
 272
 273/*
 274 * Reflects a new waiter being added to the waitqueue.
 275 */
 276static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
 277{
 278#ifdef CONFIG_SMP
 279	atomic_inc(&hb->waiters);
 280	/*
 281	 * Full barrier (A), see the ordering comment above.
 282	 */
 283	smp_mb__after_atomic_inc();
 284#endif
 285}
 286
 287/*
 288 * Reflects a waiter being removed from the waitqueue by wakeup
 289 * paths.
 290 */
 291static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
 292{
 293#ifdef CONFIG_SMP
 294	atomic_dec(&hb->waiters);
 295#endif
 296}
 297
 298static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
 299{
 300#ifdef CONFIG_SMP
 
 
 
 
 301	return atomic_read(&hb->waiters);
 302#else
 303	return 1;
 304#endif
 305}
 306
 307/*
 308 * We hash on the keys returned from get_futex_key (see below).
 
 
 
 
 309 */
 310static struct futex_hash_bucket *hash_futex(union futex_key *key)
 311{
 312	u32 hash = jhash2((u32*)&key->both.word,
 313			  (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
 314			  key->both.offset);
 
 315	return &futex_queues[hash & (futex_hashsize - 1)];
 316}
 317
 318/*
 
 
 
 
 
 319 * Return 1 if two futex_keys are equal, 0 otherwise.
 320 */
 321static inline int match_futex(union futex_key *key1, union futex_key *key2)
 322{
 323	return (key1 && key2
 324		&& key1->both.word == key2->both.word
 325		&& key1->both.ptr == key2->both.ptr
 326		&& key1->both.offset == key2->both.offset);
 327}
 328
 329/*
 330 * Take a reference to the resource addressed by a key.
 331 * Can be called while holding spinlocks.
 332 *
 333 */
 334static void get_futex_key_refs(union futex_key *key)
 
 
 
 
 
 
 
 
 
 
 
 
 335{
 336	if (!key->both.ptr)
 337		return;
 338
 339	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
 340	case FUT_OFF_INODE:
 341		ihold(key->shared.inode); /* implies MB (B) */
 342		break;
 343	case FUT_OFF_MMSHARED:
 344		futex_get_mm(key); /* implies MB (B) */
 345		break;
 346	}
 
 
 347}
 348
 349/*
 350 * Drop a reference to the resource addressed by a key.
 351 * The hash bucket spinlock must not be held.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 352 */
 353static void drop_futex_key_refs(union futex_key *key)
 354{
 355	if (!key->both.ptr) {
 356		/* If we're here then we tried to put a key we failed to get */
 357		WARN_ON_ONCE(1);
 358		return;
 359	}
 360
 361	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
 362	case FUT_OFF_INODE:
 363		iput(key->shared.inode);
 364		break;
 365	case FUT_OFF_MMSHARED:
 366		mmdrop(key->private.mm);
 367		break;
 
 
 
 
 
 
 
 368	}
 369}
 370
 371/**
 372 * get_futex_key() - Get parameters which are the keys for a futex
 373 * @uaddr:	virtual address of the futex
 374 * @fshared:	0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
 375 * @key:	address where result is stored.
 376 * @rw:		mapping needs to be read/write (values: VERIFY_READ,
 377 *              VERIFY_WRITE)
 378 *
 379 * Return: a negative error code or 0
 380 *
 381 * The key words are stored in *key on success.
 
 
 
 
 
 
 382 *
 383 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
 384 * offset_within_page).  For private mappings, it's (uaddr, current->mm).
 385 * We can usually work out the index without swapping in the page.
 
 
 
 386 *
 387 * lock_page() might sleep, the caller should not hold a spinlock.
 388 */
 389static int
 390get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
 391{
 392	unsigned long address = (unsigned long)uaddr;
 393	struct mm_struct *mm = current->mm;
 394	struct page *page, *page_head;
 
 395	int err, ro = 0;
 396
 397	/*
 398	 * The futex address must be "naturally" aligned.
 399	 */
 400	key->both.offset = address % PAGE_SIZE;
 401	if (unlikely((address % sizeof(u32)) != 0))
 402		return -EINVAL;
 403	address -= key->both.offset;
 404
 405	if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
 
 
 
 406		return -EFAULT;
 407
 408	/*
 409	 * PROCESS_PRIVATE futexes are fast.
 410	 * As the mm cannot disappear under us and the 'key' only needs
 411	 * virtual address, we dont even have to find the underlying vma.
 412	 * Note : We do have to check 'uaddr' is a valid user address,
 413	 *        but access_ok() should be faster than find_vma()
 414	 */
 415	if (!fshared) {
 416		key->private.mm = mm;
 417		key->private.address = address;
 418		get_futex_key_refs(key);  /* implies MB (B) */
 419		return 0;
 420	}
 421
 422again:
 423	err = get_user_pages_fast(address, 1, 1, &page);
 
 
 
 
 424	/*
 425	 * If write access is not required (eg. FUTEX_WAIT), try
 426	 * and get read-only access.
 427	 */
 428	if (err == -EFAULT && rw == VERIFY_READ) {
 429		err = get_user_pages_fast(address, 1, 0, &page);
 430		ro = 1;
 431	}
 432	if (err < 0)
 433		return err;
 434	else
 435		err = 0;
 436
 437#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 438	page_head = page;
 439	if (unlikely(PageTail(page))) {
 440		put_page(page);
 441		/* serialize against __split_huge_page_splitting() */
 442		local_irq_disable();
 443		if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
 444			page_head = compound_head(page);
 445			/*
 446			 * page_head is valid pointer but we must pin
 447			 * it before taking the PG_lock and/or
 448			 * PG_compound_lock. The moment we re-enable
 449			 * irqs __split_huge_page_splitting() can
 450			 * return and the head page can be freed from
 451			 * under us. We can't take the PG_lock and/or
 452			 * PG_compound_lock on a page that could be
 453			 * freed from under us.
 454			 */
 455			if (page != page_head) {
 456				get_page(page_head);
 457				put_page(page);
 458			}
 459			local_irq_enable();
 460		} else {
 461			local_irq_enable();
 462			goto again;
 463		}
 464	}
 465#else
 466	page_head = compound_head(page);
 467	if (page != page_head) {
 468		get_page(page_head);
 469		put_page(page);
 470	}
 471#endif
 472
 473	lock_page(page_head);
 474
 475	/*
 476	 * If page_head->mapping is NULL, then it cannot be a PageAnon
 477	 * page; but it might be the ZERO_PAGE or in the gate area or
 478	 * in a special mapping (all cases which we are happy to fail);
 479	 * or it may have been a good file page when get_user_pages_fast
 480	 * found it, but truncated or holepunched or subjected to
 481	 * invalidate_complete_page2 before we got the page lock (also
 482	 * cases which we are happy to fail).  And we hold a reference,
 483	 * so refcount care in invalidate_complete_page's remove_mapping
 484	 * prevents drop_caches from setting mapping to NULL beneath us.
 485	 *
 486	 * The case we do have to guard against is when memory pressure made
 487	 * shmem_writepage move it from filecache to swapcache beneath us:
 488	 * an unlikely race, but we do need to retry for page_head->mapping.
 489	 */
 490	if (!page_head->mapping) {
 491		int shmem_swizzled = PageSwapCache(page_head);
 492		unlock_page(page_head);
 493		put_page(page_head);
 
 
 
 
 
 
 
 
 
 494		if (shmem_swizzled)
 495			goto again;
 
 496		return -EFAULT;
 497	}
 498
 499	/*
 500	 * Private mappings are handled in a simple way.
 501	 *
 
 
 
 502	 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
 503	 * it's a read-only handle, it's expected that futexes attach to
 504	 * the object not the particular process.
 505	 */
 506	if (PageAnon(page_head)) {
 507		/*
 508		 * A RO anonymous page will never change and thus doesn't make
 509		 * sense for futex operations.
 510		 */
 511		if (ro) {
 512			err = -EFAULT;
 513			goto out;
 514		}
 515
 516		key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
 517		key->private.mm = mm;
 518		key->private.address = address;
 
 519	} else {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 520		key->both.offset |= FUT_OFF_INODE; /* inode-based key */
 521		key->shared.inode = page_head->mapping->host;
 522		key->shared.pgoff = basepage_index(page);
 
 523	}
 524
 525	get_futex_key_refs(key); /* implies MB (B) */
 526
 527out:
 528	unlock_page(page_head);
 529	put_page(page_head);
 530	return err;
 531}
 532
 533static inline void put_futex_key(union futex_key *key)
 534{
 535	drop_futex_key_refs(key);
 536}
 537
 538/**
 539 * fault_in_user_writeable() - Fault in user address and verify RW access
 540 * @uaddr:	pointer to faulting user space address
 541 *
 542 * Slow path to fixup the fault we just took in the atomic write
 543 * access to @uaddr.
 544 *
 545 * We have no generic implementation of a non-destructive write to the
 546 * user address. We know that we faulted in the atomic pagefault
 547 * disabled section so we can as well avoid the #PF overhead by
 548 * calling get_user_pages() right away.
 549 */
 550static int fault_in_user_writeable(u32 __user *uaddr)
 551{
 552	struct mm_struct *mm = current->mm;
 553	int ret;
 554
 555	down_read(&mm->mmap_sem);
 556	ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
 557			       FAULT_FLAG_WRITE);
 558	up_read(&mm->mmap_sem);
 559
 560	return ret < 0 ? ret : 0;
 561}
 562
 563/**
 564 * futex_top_waiter() - Return the highest priority waiter on a futex
 565 * @hb:		the hash bucket the futex_q's reside in
 566 * @key:	the futex key (to distinguish it from other futex futex_q's)
 567 *
 568 * Must be called with the hb lock held.
 569 */
 570static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
 571					union futex_key *key)
 572{
 573	struct futex_q *this;
 574
 575	plist_for_each_entry(this, &hb->chain, list) {
 576		if (match_futex(&this->key, key))
 577			return this;
 578	}
 579	return NULL;
 580}
 581
 582static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
 583				      u32 uval, u32 newval)
 584{
 585	int ret;
 586
 587	pagefault_disable();
 588	ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
 589	pagefault_enable();
 590
 591	return ret;
 592}
 593
 594static int get_futex_value_locked(u32 *dest, u32 __user *from)
 595{
 596	int ret;
 597
 598	pagefault_disable();
 599	ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
 600	pagefault_enable();
 601
 602	return ret ? -EFAULT : 0;
 603}
 604
 605
 606/*
 607 * PI code:
 608 */
 609static int refill_pi_state_cache(void)
 610{
 611	struct futex_pi_state *pi_state;
 612
 613	if (likely(current->pi_state_cache))
 614		return 0;
 615
 616	pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
 617
 618	if (!pi_state)
 619		return -ENOMEM;
 620
 621	INIT_LIST_HEAD(&pi_state->list);
 622	/* pi_mutex gets initialized later */
 623	pi_state->owner = NULL;
 624	atomic_set(&pi_state->refcount, 1);
 625	pi_state->key = FUTEX_KEY_INIT;
 626
 627	current->pi_state_cache = pi_state;
 628
 629	return 0;
 630}
 631
 632static struct futex_pi_state * alloc_pi_state(void)
 633{
 634	struct futex_pi_state *pi_state = current->pi_state_cache;
 635
 636	WARN_ON(!pi_state);
 637	current->pi_state_cache = NULL;
 638
 639	return pi_state;
 640}
 641
 642static void free_pi_state(struct futex_pi_state *pi_state)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 643{
 644	if (!atomic_dec_and_test(&pi_state->refcount))
 
 
 
 645		return;
 646
 647	/*
 648	 * If pi_state->owner is NULL, the owner is most probably dying
 649	 * and has cleaned up the pi_state already
 650	 */
 651	if (pi_state->owner) {
 652		raw_spin_lock_irq(&pi_state->owner->pi_lock);
 653		list_del_init(&pi_state->list);
 654		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
 655
 656		rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
 
 
 
 657	}
 658
 659	if (current->pi_state_cache)
 660		kfree(pi_state);
 661	else {
 662		/*
 663		 * pi_state->list is already empty.
 664		 * clear pi_state->owner.
 665		 * refcount is at 0 - put it back to 1.
 666		 */
 667		pi_state->owner = NULL;
 668		atomic_set(&pi_state->refcount, 1);
 669		current->pi_state_cache = pi_state;
 670	}
 671}
 672
 673/*
 674 * Look up the task based on what TID userspace gave us.
 675 * We dont trust it.
 676 */
 677static struct task_struct * futex_find_get_task(pid_t pid)
 678{
 679	struct task_struct *p;
 680
 681	rcu_read_lock();
 682	p = find_task_by_vpid(pid);
 683	if (p)
 684		get_task_struct(p);
 685
 686	rcu_read_unlock();
 687
 688	return p;
 689}
 690
 691/*
 692 * This task is holding PI mutexes at exit time => bad.
 693 * Kernel cleans up PI-state, but userspace is likely hosed.
 694 * (Robust-futex cleanup is separate and might save the day for userspace.)
 695 */
 696void exit_pi_state_list(struct task_struct *curr)
 697{
 698	struct list_head *next, *head = &curr->pi_state_list;
 699	struct futex_pi_state *pi_state;
 700	struct futex_hash_bucket *hb;
 701	union futex_key key = FUTEX_KEY_INIT;
 702
 703	if (!futex_cmpxchg_enabled)
 704		return;
 705	/*
 706	 * We are a ZOMBIE and nobody can enqueue itself on
 707	 * pi_state_list anymore, but we have to be careful
 708	 * versus waiters unqueueing themselves:
 709	 */
 710	raw_spin_lock_irq(&curr->pi_lock);
 711	while (!list_empty(head)) {
 712
 713		next = head->next;
 714		pi_state = list_entry(next, struct futex_pi_state, list);
 715		key = pi_state->key;
 716		hb = hash_futex(&key);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 717		raw_spin_unlock_irq(&curr->pi_lock);
 718
 719		spin_lock(&hb->lock);
 720
 721		raw_spin_lock_irq(&curr->pi_lock);
 722		/*
 723		 * We dropped the pi-lock, so re-check whether this
 724		 * task still owns the PI-state:
 725		 */
 726		if (head->next != next) {
 
 
 727			spin_unlock(&hb->lock);
 
 728			continue;
 729		}
 730
 731		WARN_ON(pi_state->owner != curr);
 732		WARN_ON(list_empty(&pi_state->list));
 733		list_del_init(&pi_state->list);
 734		pi_state->owner = NULL;
 735		raw_spin_unlock_irq(&curr->pi_lock);
 736
 737		rt_mutex_unlock(&pi_state->pi_mutex);
 738
 
 
 739		spin_unlock(&hb->lock);
 740
 
 
 
 741		raw_spin_lock_irq(&curr->pi_lock);
 742	}
 743	raw_spin_unlock_irq(&curr->pi_lock);
 744}
 
 
 
 745
 746/*
 747 * We need to check the following states:
 748 *
 749 *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
 750 *
 751 * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
 752 * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
 753 *
 754 * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
 755 *
 756 * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
 757 * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
 758 *
 759 * [6]  Found  | Found    | task      | 0         | 1      | Valid
 760 *
 761 * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
 762 *
 763 * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
 764 * [9]  Found  | Found    | task      | 0         | 0      | Invalid
 765 * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
 766 *
 767 * [1]	Indicates that the kernel can acquire the futex atomically. We
 768 *	came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
 769 *
 770 * [2]	Valid, if TID does not belong to a kernel thread. If no matching
 771 *      thread is found then it indicates that the owner TID has died.
 772 *
 773 * [3]	Invalid. The waiter is queued on a non PI futex
 774 *
 775 * [4]	Valid state after exit_robust_list(), which sets the user space
 776 *	value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
 777 *
 778 * [5]	The user space value got manipulated between exit_robust_list()
 779 *	and exit_pi_state_list()
 780 *
 781 * [6]	Valid state after exit_pi_state_list() which sets the new owner in
 782 *	the pi_state but cannot access the user space value.
 783 *
 784 * [7]	pi_state->owner can only be NULL when the OWNER_DIED bit is set.
 785 *
 786 * [8]	Owner and user space value match
 787 *
 788 * [9]	There is no transient state which sets the user space TID to 0
 789 *	except exit_robust_list(), but this is indicated by the
 790 *	FUTEX_OWNER_DIED bit. See [4]
 791 *
 792 * [10] There is no transient state which leaves owner and user space
 793 *	TID out of sync.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 794 */
 795static int
 796lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
 797		union futex_key *key, struct futex_pi_state **ps)
 
 
 
 
 
 
 798{
 799	struct futex_pi_state *pi_state = NULL;
 800	struct futex_q *this, *next;
 801	struct task_struct *p;
 802	pid_t pid = uval & FUTEX_TID_MASK;
 
 
 803
 804	plist_for_each_entry_safe(this, next, &hb->chain, list) {
 805		if (match_futex(&this->key, key)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 806			/*
 807			 * Sanity check the waiter before increasing
 808			 * the refcount and attaching to it.
 809			 */
 810			pi_state = this->pi_state;
 
 811			/*
 812			 * Userspace might have messed up non-PI and
 813			 * PI futexes [3]
 814			 */
 815			if (unlikely(!pi_state))
 816				return -EINVAL;
 817
 818			WARN_ON(!atomic_read(&pi_state->refcount));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 819
 820			/*
 821			 * Handle the owner died case:
 822			 */
 823			if (uval & FUTEX_OWNER_DIED) {
 824				/*
 825				 * exit_pi_state_list sets owner to NULL and
 826				 * wakes the topmost waiter. The task which
 827				 * acquires the pi_state->rt_mutex will fixup
 828				 * owner.
 829				 */
 830				if (!pi_state->owner) {
 831					/*
 832					 * No pi state owner, but the user
 833					 * space TID is not 0. Inconsistent
 834					 * state. [5]
 835					 */
 836					if (pid)
 837						return -EINVAL;
 838					/*
 839					 * Take a ref on the state and
 840					 * return. [4]
 841					 */
 842					goto out_state;
 843				}
 844
 845				/*
 846				 * If TID is 0, then either the dying owner
 847				 * has not yet executed exit_pi_state_list()
 848				 * or some waiter acquired the rtmutex in the
 849				 * pi state, but did not yet fixup the TID in
 850				 * user space.
 851				 *
 852				 * Take a ref on the state and return. [6]
 853				 */
 854				if (!pid)
 855					goto out_state;
 856			} else {
 857				/*
 858				 * If the owner died bit is not set,
 859				 * then the pi_state must have an
 860				 * owner. [7]
 861				 */
 862				if (!pi_state->owner)
 863					return -EINVAL;
 864			}
 865
 866			/*
 867			 * Bail out if user space manipulated the
 868			 * futex value. If pi state exists then the
 869			 * owner TID must be the same as the user
 870			 * space TID. [9/10]
 871			 */
 872			if (pid != task_pid_vnr(pi_state->owner))
 873				return -EINVAL;
 
 
 
 874
 875		out_state:
 876			atomic_inc(&pi_state->refcount);
 877			*ps = pi_state;
 878			return 0;
 879		}
 
 
 
 
 
 
 
 
 
 
 
 
 880	}
 881
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 882	/*
 883	 * We are the first waiter - try to look up the real owner and attach
 884	 * the new pi_state to it, but bail out when TID = 0 [1]
 
 
 
 885	 */
 886	if (!pid)
 887		return -ESRCH;
 888	p = futex_find_get_task(pid);
 889	if (!p)
 890		return -ESRCH;
 891
 892	if (!p->mm) {
 893		put_task_struct(p);
 894		return -EPERM;
 895	}
 896
 897	/*
 898	 * We need to look at the task state flags to figure out,
 899	 * whether the task is exiting. To protect against the do_exit
 900	 * change of the task flags, we do this protected by
 901	 * p->pi_lock:
 902	 */
 903	raw_spin_lock_irq(&p->pi_lock);
 904	if (unlikely(p->flags & PF_EXITING)) {
 905		/*
 906		 * The task is on the way out. When PF_EXITPIDONE is
 907		 * set, we know that the task has finished the
 908		 * cleanup:
 909		 */
 910		int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
 911
 912		raw_spin_unlock_irq(&p->pi_lock);
 913		put_task_struct(p);
 
 
 
 
 
 
 
 
 
 
 
 
 914		return ret;
 915	}
 916
 917	/*
 918	 * No existing pi state. First waiter. [2]
 
 
 
 919	 */
 920	pi_state = alloc_pi_state();
 921
 922	/*
 923	 * Initialize the pi_mutex in locked state and make 'p'
 924	 * the owner of it:
 925	 */
 926	rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
 927
 928	/* Store the key for possible exit cleanups: */
 929	pi_state->key = *key;
 930
 931	WARN_ON(!list_empty(&pi_state->list));
 932	list_add(&pi_state->list, &p->pi_state_list);
 
 
 
 
 933	pi_state->owner = p;
 934	raw_spin_unlock_irq(&p->pi_lock);
 935
 936	put_task_struct(p);
 937
 938	*ps = pi_state;
 939
 940	return 0;
 941}
 942
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 943/**
 944 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
 945 * @uaddr:		the pi futex user address
 946 * @hb:			the pi futex hash bucket
 947 * @key:		the futex key associated with uaddr and hb
 948 * @ps:			the pi_state pointer where we store the result of the
 949 *			lookup
 950 * @task:		the task to perform the atomic lock work for.  This will
 951 *			be "current" except in the case of requeue pi.
 
 
 952 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
 953 *
 954 * Return:
 955 *  0 - ready to wait;
 956 *  1 - acquired the lock;
 957 * <0 - error
 958 *
 959 * The hb->lock and futex_key refs shall be held by the caller.
 
 
 
 
 960 */
 961static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
 962				union futex_key *key,
 963				struct futex_pi_state **ps,
 964				struct task_struct *task, int set_waiters)
 
 
 965{
 966	int lock_taken, ret, force_take = 0;
 967	u32 uval, newval, curval, vpid = task_pid_vnr(task);
 968
 969retry:
 970	ret = lock_taken = 0;
 971
 972	/*
 973	 * To avoid races, we attempt to take the lock here again
 974	 * (by doing a 0 -> TID atomic cmpxchg), while holding all
 975	 * the locks. It will most likely not succeed.
 976	 */
 977	newval = vpid;
 978	if (set_waiters)
 979		newval |= FUTEX_WAITERS;
 980
 981	if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, 0, newval)))
 982		return -EFAULT;
 983
 984	/*
 985	 * Detect deadlocks.
 986	 */
 987	if ((unlikely((curval & FUTEX_TID_MASK) == vpid)))
 988		return -EDEADLK;
 989
 990	/*
 991	 * Surprise - we got the lock, but we do not trust user space at all.
 992	 */
 993	if (unlikely(!curval)) {
 994		/*
 995		 * We verify whether there is kernel state for this
 996		 * futex. If not, we can safely assume, that the 0 ->
 997		 * TID transition is correct. If state exists, we do
 998		 * not bother to fixup the user space state as it was
 999		 * corrupted already.
1000		 */
1001		return futex_top_waiter(hb, key) ? -EINVAL : 1;
1002	}
1003
1004	uval = curval;
1005
1006	/*
1007	 * Set the FUTEX_WAITERS flag, so the owner will know it has someone
1008	 * to wake at the next unlock.
1009	 */
1010	newval = curval | FUTEX_WAITERS;
 
 
1011
1012	/*
1013	 * Should we force take the futex? See below.
 
 
 
1014	 */
1015	if (unlikely(force_take)) {
1016		/*
1017		 * Keep the OWNER_DIED and the WAITERS bit and set the
1018		 * new TID value.
1019		 */
1020		newval = (curval & ~FUTEX_TID_MASK) | vpid;
1021		force_take = 0;
1022		lock_taken = 1;
1023	}
1024
1025	if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1026		return -EFAULT;
1027	if (unlikely(curval != uval))
1028		goto retry;
 
 
 
 
1029
1030	/*
1031	 * We took the lock due to forced take over.
 
 
1032	 */
1033	if (unlikely(lock_taken))
1034		return 1;
1035
 
1036	/*
1037	 * We dont have the lock. Look up the PI state (or create it if
1038	 * we are the first waiter):
 
1039	 */
1040	ret = lookup_pi_state(uval, hb, key, ps);
1041
1042	if (unlikely(ret)) {
1043		switch (ret) {
1044		case -ESRCH:
1045			/*
1046			 * We failed to find an owner for this
1047			 * futex. So we have no pi_state to block
1048			 * on. This can happen in two cases:
1049			 *
1050			 * 1) The owner died
1051			 * 2) A stale FUTEX_WAITERS bit
1052			 *
1053			 * Re-read the futex value.
1054			 */
1055			if (get_futex_value_locked(&curval, uaddr))
1056				return -EFAULT;
1057
1058			/*
1059			 * If the owner died or we have a stale
1060			 * WAITERS bit the owner TID in the user space
1061			 * futex is 0.
1062			 */
1063			if (!(curval & FUTEX_TID_MASK)) {
1064				force_take = 1;
1065				goto retry;
1066			}
1067		default:
1068			break;
1069		}
1070	}
1071
1072	return ret;
1073}
1074
1075/**
1076 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1077 * @q:	The futex_q to unqueue
1078 *
1079 * The q->lock_ptr must not be NULL and must be held by the caller.
1080 */
1081static void __unqueue_futex(struct futex_q *q)
1082{
1083	struct futex_hash_bucket *hb;
1084
1085	if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1086	    || WARN_ON(plist_node_empty(&q->list)))
1087		return;
 
1088
1089	hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1090	plist_del(&q->list, &hb->chain);
1091	hb_waiters_dec(hb);
1092}
1093
1094/*
1095 * The hash bucket lock must be held when this is called.
1096 * Afterwards, the futex_q must not be accessed.
 
 
1097 */
1098static void wake_futex(struct futex_q *q)
1099{
1100	struct task_struct *p = q->task;
1101
1102	if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1103		return;
1104
1105	/*
1106	 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1107	 * a non-futex wake up happens on another CPU then the task
1108	 * might exit and p would dereference a non-existing task
1109	 * struct. Prevent this by holding a reference on p across the
1110	 * wake up.
1111	 */
1112	get_task_struct(p);
1113
1114	__unqueue_futex(q);
1115	/*
1116	 * The waiting task can free the futex_q as soon as
1117	 * q->lock_ptr = NULL is written, without taking any locks. A
1118	 * memory barrier is required here to prevent the following
1119	 * store to lock_ptr from getting ahead of the plist_del.
 
1120	 */
1121	smp_wmb();
1122	q->lock_ptr = NULL;
1123
1124	wake_up_state(p, TASK_NORMAL);
1125	put_task_struct(p);
 
 
 
1126}
1127
1128static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
 
 
 
1129{
 
 
1130	struct task_struct *new_owner;
1131	struct futex_pi_state *pi_state = this->pi_state;
1132	u32 uninitialized_var(curval), newval;
1133	int ret = 0;
1134
1135	if (!pi_state)
1136		return -EINVAL;
1137
1138	/*
1139	 * If current does not own the pi_state then the futex is
1140	 * inconsistent and user space fiddled with the futex value.
1141	 */
1142	if (pi_state->owner != current)
1143		return -EINVAL;
1144
1145	raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1146	new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
 
1147
1148	/*
1149	 * It is possible that the next waiter (the one that brought
1150	 * this owner to the kernel) timed out and is no longer
1151	 * waiting on the lock.
1152	 */
1153	if (!new_owner)
1154		new_owner = this->task;
1155
1156	/*
1157	 * We pass it to the next owner. The WAITERS bit is always
1158	 * kept enabled while there is PI state around. We cleanup the
1159	 * owner died bit, because we are the owner.
1160	 */
1161	newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1162
1163	if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1164		ret = -EFAULT;
1165	else if (curval != uval)
1166		ret = -EINVAL;
1167	if (ret) {
1168		raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1169		return ret;
1170	}
1171
1172	raw_spin_lock_irq(&pi_state->owner->pi_lock);
1173	WARN_ON(list_empty(&pi_state->list));
1174	list_del_init(&pi_state->list);
1175	raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1176
1177	raw_spin_lock_irq(&new_owner->pi_lock);
1178	WARN_ON(!list_empty(&pi_state->list));
1179	list_add(&pi_state->list, &new_owner->pi_state_list);
1180	pi_state->owner = new_owner;
1181	raw_spin_unlock_irq(&new_owner->pi_lock);
1182
1183	raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1184	rt_mutex_unlock(&pi_state->pi_mutex);
1185
1186	return 0;
1187}
 
 
 
 
 
 
 
1188
1189static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
1190{
1191	u32 uninitialized_var(oldval);
1192
1193	/*
1194	 * There is no waiter, so we unlock the futex. The owner died
1195	 * bit has not to be preserved here. We are the owner:
1196	 */
1197	if (cmpxchg_futex_value_locked(&oldval, uaddr, uval, 0))
1198		return -EFAULT;
1199	if (oldval != uval)
1200		return -EAGAIN;
1201
1202	return 0;
1203}
1204
1205/*
1206 * Express the locking dependencies for lockdep:
1207 */
1208static inline void
1209double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1210{
1211	if (hb1 <= hb2) {
1212		spin_lock(&hb1->lock);
1213		if (hb1 < hb2)
1214			spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1215	} else { /* hb1 > hb2 */
1216		spin_lock(&hb2->lock);
1217		spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1218	}
1219}
1220
1221static inline void
1222double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1223{
1224	spin_unlock(&hb1->lock);
1225	if (hb1 != hb2)
1226		spin_unlock(&hb2->lock);
1227}
1228
1229/*
1230 * Wake up waiters matching bitset queued on this futex (uaddr).
1231 */
1232static int
1233futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1234{
1235	struct futex_hash_bucket *hb;
1236	struct futex_q *this, *next;
1237	union futex_key key = FUTEX_KEY_INIT;
1238	int ret;
 
1239
1240	if (!bitset)
1241		return -EINVAL;
1242
1243	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1244	if (unlikely(ret != 0))
1245		goto out;
1246
1247	hb = hash_futex(&key);
1248
1249	/* Make sure we really have tasks to wakeup */
1250	if (!hb_waiters_pending(hb))
1251		goto out_put_key;
1252
1253	spin_lock(&hb->lock);
1254
1255	plist_for_each_entry_safe(this, next, &hb->chain, list) {
1256		if (match_futex (&this->key, &key)) {
1257			if (this->pi_state || this->rt_waiter) {
1258				ret = -EINVAL;
1259				break;
1260			}
1261
1262			/* Check if one of the bits is set in both bitsets */
1263			if (!(this->bitset & bitset))
1264				continue;
1265
1266			wake_futex(this);
1267			if (++ret >= nr_wake)
1268				break;
1269		}
1270	}
1271
1272	spin_unlock(&hb->lock);
1273out_put_key:
1274	put_futex_key(&key);
1275out:
1276	return ret;
1277}
1278
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1279/*
1280 * Wake up all waiters hashed on the physical page that is mapped
1281 * to this virtual address:
1282 */
1283static int
1284futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1285	      int nr_wake, int nr_wake2, int op)
1286{
1287	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1288	struct futex_hash_bucket *hb1, *hb2;
1289	struct futex_q *this, *next;
1290	int ret, op_ret;
 
1291
1292retry:
1293	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1294	if (unlikely(ret != 0))
1295		goto out;
1296	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1297	if (unlikely(ret != 0))
1298		goto out_put_key1;
1299
1300	hb1 = hash_futex(&key1);
1301	hb2 = hash_futex(&key2);
1302
1303retry_private:
1304	double_lock_hb(hb1, hb2);
1305	op_ret = futex_atomic_op_inuser(op, uaddr2);
1306	if (unlikely(op_ret < 0)) {
1307
1308		double_unlock_hb(hb1, hb2);
1309
1310#ifndef CONFIG_MMU
1311		/*
1312		 * we don't get EFAULT from MMU faults if we don't have an MMU,
1313		 * but we might get them from range checking
1314		 */
1315		ret = op_ret;
1316		goto out_put_keys;
1317#endif
1318
1319		if (unlikely(op_ret != -EFAULT)) {
1320			ret = op_ret;
1321			goto out_put_keys;
1322		}
1323
1324		ret = fault_in_user_writeable(uaddr2);
1325		if (ret)
1326			goto out_put_keys;
 
 
1327
 
1328		if (!(flags & FLAGS_SHARED))
1329			goto retry_private;
1330
1331		put_futex_key(&key2);
1332		put_futex_key(&key1);
1333		goto retry;
1334	}
1335
1336	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1337		if (match_futex (&this->key, &key1)) {
1338			if (this->pi_state || this->rt_waiter) {
1339				ret = -EINVAL;
1340				goto out_unlock;
1341			}
1342			wake_futex(this);
1343			if (++ret >= nr_wake)
1344				break;
1345		}
1346	}
1347
1348	if (op_ret > 0) {
1349		op_ret = 0;
1350		plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1351			if (match_futex (&this->key, &key2)) {
1352				if (this->pi_state || this->rt_waiter) {
1353					ret = -EINVAL;
1354					goto out_unlock;
1355				}
1356				wake_futex(this);
1357				if (++op_ret >= nr_wake2)
1358					break;
1359			}
1360		}
1361		ret += op_ret;
1362	}
1363
1364out_unlock:
1365	double_unlock_hb(hb1, hb2);
1366out_put_keys:
1367	put_futex_key(&key2);
1368out_put_key1:
1369	put_futex_key(&key1);
1370out:
1371	return ret;
1372}
1373
1374/**
1375 * requeue_futex() - Requeue a futex_q from one hb to another
1376 * @q:		the futex_q to requeue
1377 * @hb1:	the source hash_bucket
1378 * @hb2:	the target hash_bucket
1379 * @key2:	the new key for the requeued futex_q
1380 */
1381static inline
1382void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1383		   struct futex_hash_bucket *hb2, union futex_key *key2)
1384{
1385
1386	/*
1387	 * If key1 and key2 hash to the same bucket, no need to
1388	 * requeue.
1389	 */
1390	if (likely(&hb1->chain != &hb2->chain)) {
1391		plist_del(&q->list, &hb1->chain);
1392		hb_waiters_dec(hb1);
1393		plist_add(&q->list, &hb2->chain);
1394		hb_waiters_inc(hb2);
 
1395		q->lock_ptr = &hb2->lock;
1396	}
1397	get_futex_key_refs(key2);
1398	q->key = *key2;
1399}
1400
1401/**
1402 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1403 * @q:		the futex_q
1404 * @key:	the key of the requeue target futex
1405 * @hb:		the hash_bucket of the requeue target futex
1406 *
1407 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1408 * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1409 * to the requeue target futex so the waiter can detect the wakeup on the right
1410 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1411 * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1412 * to protect access to the pi_state to fixup the owner later.  Must be called
1413 * with both q->lock_ptr and hb->lock held.
1414 */
1415static inline
1416void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1417			   struct futex_hash_bucket *hb)
1418{
1419	get_futex_key_refs(key);
1420	q->key = *key;
1421
1422	__unqueue_futex(q);
1423
1424	WARN_ON(!q->rt_waiter);
1425	q->rt_waiter = NULL;
1426
1427	q->lock_ptr = &hb->lock;
1428
1429	wake_up_state(q->task, TASK_NORMAL);
1430}
1431
1432/**
1433 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1434 * @pifutex:		the user address of the to futex
1435 * @hb1:		the from futex hash bucket, must be locked by the caller
1436 * @hb2:		the to futex hash bucket, must be locked by the caller
1437 * @key1:		the from futex key
1438 * @key2:		the to futex key
1439 * @ps:			address to store the pi_state pointer
 
 
1440 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1441 *
1442 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1443 * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1444 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1445 * hb1 and hb2 must be held by the caller.
1446 *
 
 
 
 
1447 * Return:
1448 *  0 - failed to acquire the lock atomically;
1449 * >0 - acquired the lock, return value is vpid of the top_waiter
1450 * <0 - error
1451 */
1452static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1453				 struct futex_hash_bucket *hb1,
1454				 struct futex_hash_bucket *hb2,
1455				 union futex_key *key1, union futex_key *key2,
1456				 struct futex_pi_state **ps, int set_waiters)
1457{
1458	struct futex_q *top_waiter = NULL;
1459	u32 curval;
1460	int ret, vpid;
1461
1462	if (get_futex_value_locked(&curval, pifutex))
1463		return -EFAULT;
1464
 
 
 
1465	/*
1466	 * Find the top_waiter and determine if there are additional waiters.
1467	 * If the caller intends to requeue more than 1 waiter to pifutex,
1468	 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1469	 * as we have means to handle the possible fault.  If not, don't set
1470	 * the bit unecessarily as it will force the subsequent unlock to enter
1471	 * the kernel.
1472	 */
1473	top_waiter = futex_top_waiter(hb1, key1);
1474
1475	/* There are no waiters, nothing for us to do. */
1476	if (!top_waiter)
1477		return 0;
1478
1479	/* Ensure we requeue to the expected futex. */
1480	if (!match_futex(top_waiter->requeue_pi_key, key2))
1481		return -EINVAL;
1482
1483	/*
1484	 * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1485	 * the contended case or if set_waiters is 1.  The pi_state is returned
1486	 * in ps in contended cases.
1487	 */
1488	vpid = task_pid_vnr(top_waiter->task);
1489	ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1490				   set_waiters);
1491	if (ret == 1) {
1492		requeue_pi_wake_futex(top_waiter, key2, hb2);
1493		return vpid;
1494	}
1495	return ret;
1496}
1497
1498/**
1499 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1500 * @uaddr1:	source futex user address
1501 * @flags:	futex flags (FLAGS_SHARED, etc.)
1502 * @uaddr2:	target futex user address
1503 * @nr_wake:	number of waiters to wake (must be 1 for requeue_pi)
1504 * @nr_requeue:	number of waiters to requeue (0-INT_MAX)
1505 * @cmpval:	@uaddr1 expected value (or %NULL)
1506 * @requeue_pi:	if we are attempting to requeue from a non-pi futex to a
1507 *		pi futex (pi to pi requeue is not supported)
1508 *
1509 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1510 * uaddr2 atomically on behalf of the top waiter.
1511 *
1512 * Return:
1513 * >=0 - on success, the number of tasks requeued or woken;
1514 *  <0 - on error
1515 */
1516static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1517			 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1518			 u32 *cmpval, int requeue_pi)
1519{
1520	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1521	int drop_count = 0, task_count = 0, ret;
1522	struct futex_pi_state *pi_state = NULL;
1523	struct futex_hash_bucket *hb1, *hb2;
1524	struct futex_q *this, *next;
 
 
 
 
 
 
 
 
 
 
 
 
 
1525
1526	if (requeue_pi) {
1527		/*
1528		 * Requeue PI only works on two distinct uaddrs. This
1529		 * check is only valid for private futexes. See below.
1530		 */
1531		if (uaddr1 == uaddr2)
1532			return -EINVAL;
1533
1534		/*
1535		 * requeue_pi requires a pi_state, try to allocate it now
1536		 * without any locks in case it fails.
1537		 */
1538		if (refill_pi_state_cache())
1539			return -ENOMEM;
1540		/*
1541		 * requeue_pi must wake as many tasks as it can, up to nr_wake
1542		 * + nr_requeue, since it acquires the rt_mutex prior to
1543		 * returning to userspace, so as to not leave the rt_mutex with
1544		 * waiters and no owner.  However, second and third wake-ups
1545		 * cannot be predicted as they involve race conditions with the
1546		 * first wake and a fault while looking up the pi_state.  Both
1547		 * pthread_cond_signal() and pthread_cond_broadcast() should
1548		 * use nr_wake=1.
1549		 */
1550		if (nr_wake != 1)
1551			return -EINVAL;
1552	}
1553
1554retry:
1555	if (pi_state != NULL) {
1556		/*
1557		 * We will have to lookup the pi_state again, so free this one
1558		 * to keep the accounting correct.
1559		 */
1560		free_pi_state(pi_state);
1561		pi_state = NULL;
1562	}
1563
1564	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1565	if (unlikely(ret != 0))
1566		goto out;
1567	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1568			    requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1569	if (unlikely(ret != 0))
1570		goto out_put_key1;
1571
1572	/*
1573	 * The check above which compares uaddrs is not sufficient for
1574	 * shared futexes. We need to compare the keys:
1575	 */
1576	if (requeue_pi && match_futex(&key1, &key2)) {
1577		ret = -EINVAL;
1578		goto out_put_keys;
1579	}
1580
1581	hb1 = hash_futex(&key1);
1582	hb2 = hash_futex(&key2);
1583
1584retry_private:
1585	hb_waiters_inc(hb2);
1586	double_lock_hb(hb1, hb2);
1587
1588	if (likely(cmpval != NULL)) {
1589		u32 curval;
1590
1591		ret = get_futex_value_locked(&curval, uaddr1);
1592
1593		if (unlikely(ret)) {
1594			double_unlock_hb(hb1, hb2);
1595			hb_waiters_dec(hb2);
1596
1597			ret = get_user(curval, uaddr1);
1598			if (ret)
1599				goto out_put_keys;
1600
1601			if (!(flags & FLAGS_SHARED))
1602				goto retry_private;
1603
1604			put_futex_key(&key2);
1605			put_futex_key(&key1);
1606			goto retry;
1607		}
1608		if (curval != *cmpval) {
1609			ret = -EAGAIN;
1610			goto out_unlock;
1611		}
1612	}
1613
1614	if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
 
 
1615		/*
1616		 * Attempt to acquire uaddr2 and wake the top waiter. If we
1617		 * intend to requeue waiters, force setting the FUTEX_WAITERS
1618		 * bit.  We force this here where we are able to easily handle
1619		 * faults rather in the requeue loop below.
1620		 */
1621		ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1622						 &key2, &pi_state, nr_requeue);
 
1623
1624		/*
1625		 * At this point the top_waiter has either taken uaddr2 or is
1626		 * waiting on it.  If the former, then the pi_state will not
1627		 * exist yet, look it up one more time to ensure we have a
1628		 * reference to it. If the lock was taken, ret contains the
1629		 * vpid of the top waiter task.
 
 
1630		 */
1631		if (ret > 0) {
1632			WARN_ON(pi_state);
1633			drop_count++;
1634			task_count++;
1635			/*
1636			 * If we acquired the lock, then the user
1637			 * space value of uaddr2 should be vpid. It
1638			 * cannot be changed by the top waiter as it
1639			 * is blocked on hb2 lock if it tries to do
1640			 * so. If something fiddled with it behind our
1641			 * back the pi state lookup might unearth
1642			 * it. So we rather use the known value than
1643			 * rereading and handing potential crap to
1644			 * lookup_pi_state.
 
1645			 */
1646			ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
 
1647		}
1648
1649		switch (ret) {
1650		case 0:
 
1651			break;
 
 
1652		case -EFAULT:
1653			double_unlock_hb(hb1, hb2);
1654			hb_waiters_dec(hb2);
1655			put_futex_key(&key2);
1656			put_futex_key(&key1);
1657			ret = fault_in_user_writeable(uaddr2);
1658			if (!ret)
1659				goto retry;
1660			goto out;
 
1661		case -EAGAIN:
1662			/* The owner was exiting, try again. */
 
 
 
 
 
1663			double_unlock_hb(hb1, hb2);
1664			hb_waiters_dec(hb2);
1665			put_futex_key(&key2);
1666			put_futex_key(&key1);
 
 
 
 
1667			cond_resched();
1668			goto retry;
1669		default:
1670			goto out_unlock;
1671		}
1672	}
1673
1674	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1675		if (task_count - nr_wake >= nr_requeue)
1676			break;
1677
1678		if (!match_futex(&this->key, &key1))
1679			continue;
1680
1681		/*
1682		 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1683		 * be paired with each other and no other futex ops.
1684		 *
1685		 * We should never be requeueing a futex_q with a pi_state,
1686		 * which is awaiting a futex_unlock_pi().
1687		 */
1688		if ((requeue_pi && !this->rt_waiter) ||
1689		    (!requeue_pi && this->rt_waiter) ||
1690		    this->pi_state) {
1691			ret = -EINVAL;
1692			break;
1693		}
1694
1695		/*
1696		 * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1697		 * lock, we already woke the top_waiter.  If not, it will be
1698		 * woken by futex_unlock_pi().
1699		 */
1700		if (++task_count <= nr_wake && !requeue_pi) {
1701			wake_futex(this);
1702			continue;
1703		}
1704
1705		/* Ensure we requeue to the expected futex for requeue_pi. */
1706		if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1707			ret = -EINVAL;
1708			break;
1709		}
1710
1711		/*
1712		 * Requeue nr_requeue waiters and possibly one more in the case
1713		 * of requeue_pi if we couldn't acquire the lock atomically.
1714		 */
1715		if (requeue_pi) {
1716			/* Prepare the waiter to take the rt_mutex. */
1717			atomic_inc(&pi_state->refcount);
 
 
 
 
1718			this->pi_state = pi_state;
1719			ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1720							this->rt_waiter,
1721							this->task, 1);
1722			if (ret == 1) {
1723				/* We got the lock. */
 
 
 
 
 
 
 
1724				requeue_pi_wake_futex(this, &key2, hb2);
1725				drop_count++;
1726				continue;
1727			} else if (ret) {
1728				/* -EDEADLK */
 
 
 
 
 
 
 
1729				this->pi_state = NULL;
1730				free_pi_state(pi_state);
1731				goto out_unlock;
 
 
 
 
1732			}
1733		}
1734		requeue_futex(this, hb1, hb2, &key2);
1735		drop_count++;
1736	}
1737
 
 
 
 
 
 
 
1738out_unlock:
1739	double_unlock_hb(hb1, hb2);
 
1740	hb_waiters_dec(hb2);
1741
1742	/*
1743	 * drop_futex_key_refs() must be called outside the spinlocks. During
1744	 * the requeue we moved futex_q's from the hash bucket at key1 to the
1745	 * one at key2 and updated their key pointer.  We no longer need to
1746	 * hold the references to key1.
1747	 */
1748	while (--drop_count >= 0)
1749		drop_futex_key_refs(&key1);
1750
1751out_put_keys:
1752	put_futex_key(&key2);
1753out_put_key1:
1754	put_futex_key(&key1);
1755out:
1756	if (pi_state != NULL)
1757		free_pi_state(pi_state);
1758	return ret ? ret : task_count;
1759}
1760
1761/* The key must be already stored in q->key. */
1762static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1763	__acquires(&hb->lock)
1764{
1765	struct futex_hash_bucket *hb;
1766
1767	hb = hash_futex(&q->key);
1768
1769	/*
1770	 * Increment the counter before taking the lock so that
1771	 * a potential waker won't miss a to-be-slept task that is
1772	 * waiting for the spinlock. This is safe as all queue_lock()
1773	 * users end up calling queue_me(). Similarly, for housekeeping,
1774	 * decrement the counter at queue_unlock() when some error has
1775	 * occurred and we don't end up adding the task to the list.
1776	 */
1777	hb_waiters_inc(hb);
1778
1779	q->lock_ptr = &hb->lock;
1780
1781	spin_lock(&hb->lock); /* implies MB (A) */
1782	return hb;
1783}
1784
1785static inline void
1786queue_unlock(struct futex_hash_bucket *hb)
1787	__releases(&hb->lock)
1788{
1789	spin_unlock(&hb->lock);
1790	hb_waiters_dec(hb);
1791}
1792
1793/**
1794 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1795 * @q:	The futex_q to enqueue
1796 * @hb:	The destination hash bucket
1797 *
1798 * The hb->lock must be held by the caller, and is released here. A call to
1799 * queue_me() is typically paired with exactly one call to unqueue_me().  The
1800 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1801 * or nothing if the unqueue is done as part of the wake process and the unqueue
1802 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1803 * an example).
1804 */
1805static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1806	__releases(&hb->lock)
1807{
1808	int prio;
1809
1810	/*
1811	 * The priority used to register this element is
1812	 * - either the real thread-priority for the real-time threads
1813	 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1814	 * - or MAX_RT_PRIO for non-RT threads.
1815	 * Thus, all RT-threads are woken first in priority order, and
1816	 * the others are woken last, in FIFO order.
1817	 */
1818	prio = min(current->normal_prio, MAX_RT_PRIO);
1819
1820	plist_node_init(&q->list, prio);
1821	plist_add(&q->list, &hb->chain);
1822	q->task = current;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1823	spin_unlock(&hb->lock);
1824}
1825
1826/**
1827 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1828 * @q:	The futex_q to unqueue
1829 *
1830 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1831 * be paired with exactly one earlier call to queue_me().
1832 *
1833 * Return:
1834 *   1 - if the futex_q was still queued (and we removed unqueued it);
1835 *   0 - if the futex_q was already removed by the waking thread
1836 */
1837static int unqueue_me(struct futex_q *q)
1838{
1839	spinlock_t *lock_ptr;
1840	int ret = 0;
1841
1842	/* In the common case we don't take the spinlock, which is nice. */
1843retry:
1844	lock_ptr = q->lock_ptr;
1845	barrier();
 
 
 
 
1846	if (lock_ptr != NULL) {
1847		spin_lock(lock_ptr);
1848		/*
1849		 * q->lock_ptr can change between reading it and
1850		 * spin_lock(), causing us to take the wrong lock.  This
1851		 * corrects the race condition.
1852		 *
1853		 * Reasoning goes like this: if we have the wrong lock,
1854		 * q->lock_ptr must have changed (maybe several times)
1855		 * between reading it and the spin_lock().  It can
1856		 * change again after the spin_lock() but only if it was
1857		 * already changed before the spin_lock().  It cannot,
1858		 * however, change back to the original value.  Therefore
1859		 * we can detect whether we acquired the correct lock.
1860		 */
1861		if (unlikely(lock_ptr != q->lock_ptr)) {
1862			spin_unlock(lock_ptr);
1863			goto retry;
1864		}
1865		__unqueue_futex(q);
1866
1867		BUG_ON(q->pi_state);
1868
1869		spin_unlock(lock_ptr);
1870		ret = 1;
1871	}
1872
1873	drop_futex_key_refs(&q->key);
1874	return ret;
1875}
1876
1877/*
1878 * PI futexes can not be requeued and must remove themself from the
1879 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1880 * and dropped here.
1881 */
1882static void unqueue_me_pi(struct futex_q *q)
1883	__releases(q->lock_ptr)
1884{
1885	__unqueue_futex(q);
1886
1887	BUG_ON(!q->pi_state);
1888	free_pi_state(q->pi_state);
1889	q->pi_state = NULL;
1890
1891	spin_unlock(q->lock_ptr);
1892}
1893
1894/*
1895 * Fixup the pi_state owner with the new owner.
1896 *
1897 * Must be called with hash bucket lock held and mm->sem held for non
1898 * private futexes.
1899 */
1900static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1901				struct task_struct *newowner)
1902{
1903	u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1904	struct futex_pi_state *pi_state = q->pi_state;
1905	struct task_struct *oldowner = pi_state->owner;
1906	u32 uval, uninitialized_var(curval), newval;
1907	int ret;
1908
1909	/* Owner died? */
1910	if (!pi_state->owner)
1911		newtid |= FUTEX_OWNER_DIED;
1912
1913	/*
1914	 * We are here either because we stole the rtmutex from the
1915	 * previous highest priority waiter or we are the highest priority
1916	 * waiter but failed to get the rtmutex the first time.
1917	 * We have to replace the newowner TID in the user space variable.
 
 
 
 
 
 
 
1918	 * This must be atomic as we have to preserve the owner died bit here.
1919	 *
1920	 * Note: We write the user space value _before_ changing the pi_state
1921	 * because we can fault here. Imagine swapped out pages or a fork
1922	 * that marked all the anonymous memory readonly for cow.
1923	 *
1924	 * Modifying pi_state _before_ the user space value would
1925	 * leave the pi_state in an inconsistent state when we fault
1926	 * here, because we need to drop the hash bucket lock to
1927	 * handle the fault. This might be observed in the PID check
1928	 * in lookup_pi_state.
1929	 */
1930retry:
1931	if (get_futex_value_locked(&uval, uaddr))
1932		goto handle_fault;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1933
1934	while (1) {
1935		newval = (uval & FUTEX_OWNER_DIED) | newtid;
1936
1937		if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1938			goto handle_fault;
 
 
1939		if (curval == uval)
1940			break;
1941		uval = curval;
1942	}
1943
1944	/*
1945	 * We fixed up user space. Now we need to fix the pi_state
1946	 * itself.
1947	 */
1948	if (pi_state->owner != NULL) {
1949		raw_spin_lock_irq(&pi_state->owner->pi_lock);
1950		WARN_ON(list_empty(&pi_state->list));
1951		list_del_init(&pi_state->list);
1952		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1953	}
1954
1955	pi_state->owner = newowner;
1956
1957	raw_spin_lock_irq(&newowner->pi_lock);
1958	WARN_ON(!list_empty(&pi_state->list));
1959	list_add(&pi_state->list, &newowner->pi_state_list);
1960	raw_spin_unlock_irq(&newowner->pi_lock);
1961	return 0;
1962
1963	/*
1964	 * To handle the page fault we need to drop the hash bucket
1965	 * lock here. That gives the other task (either the highest priority
1966	 * waiter itself or the task which stole the rtmutex) the
1967	 * chance to try the fixup of the pi_state. So once we are
1968	 * back from handling the fault we need to check the pi_state
1969	 * after reacquiring the hash bucket lock and before trying to
1970	 * do another fixup. When the fixup has been done already we
1971	 * simply return.
 
 
 
1972	 */
1973handle_fault:
 
1974	spin_unlock(q->lock_ptr);
1975
1976	ret = fault_in_user_writeable(uaddr);
 
 
 
 
 
 
 
 
 
 
 
 
 
1977
1978	spin_lock(q->lock_ptr);
 
1979
1980	/*
1981	 * Check if someone else fixed it for us:
1982	 */
1983	if (pi_state->owner != oldowner)
1984		return 0;
1985
1986	if (ret)
1987		return ret;
 
1988
1989	goto retry;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1990}
1991
1992static long futex_wait_restart(struct restart_block *restart);
1993
1994/**
1995 * fixup_owner() - Post lock pi_state and corner case management
1996 * @uaddr:	user address of the futex
1997 * @q:		futex_q (contains pi_state and access to the rt_mutex)
1998 * @locked:	if the attempt to take the rt_mutex succeeded (1) or not (0)
1999 *
2000 * After attempting to lock an rt_mutex, this function is called to cleanup
2001 * the pi_state owner as well as handle race conditions that may allow us to
2002 * acquire the lock. Must be called with the hb lock held.
2003 *
2004 * Return:
2005 *  1 - success, lock taken;
2006 *  0 - success, lock not taken;
2007 * <0 - on error (-EFAULT)
2008 */
2009static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2010{
2011	struct task_struct *owner;
2012	int ret = 0;
2013
2014	if (locked) {
2015		/*
2016		 * Got the lock. We might not be the anticipated owner if we
2017		 * did a lock-steal - fix up the PI-state in that case:
 
 
 
 
2018		 */
2019		if (q->pi_state->owner != current)
2020			ret = fixup_pi_state_owner(uaddr, q, current);
2021		goto out;
2022	}
2023
2024	/*
2025	 * Catch the rare case, where the lock was released when we were on the
2026	 * way back before we locked the hash bucket.
 
 
 
 
2027	 */
2028	if (q->pi_state->owner == current) {
2029		/*
2030		 * Try to get the rt_mutex now. This might fail as some other
2031		 * task acquired the rt_mutex after we removed ourself from the
2032		 * rt_mutex waiters list.
2033		 */
2034		if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2035			locked = 1;
2036			goto out;
2037		}
2038
2039		/*
2040		 * pi_state is incorrect, some other task did a lock steal and
2041		 * we returned due to timeout or signal without taking the
2042		 * rt_mutex. Too late.
2043		 */
2044		raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2045		owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2046		if (!owner)
2047			owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2048		raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2049		ret = fixup_pi_state_owner(uaddr, q, owner);
2050		goto out;
2051	}
2052
2053	/*
2054	 * Paranoia check. If we did not take the lock, then we should not be
2055	 * the owner of the rt_mutex.
2056	 */
2057	if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2058		printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2059				"pi-state %p\n", ret,
2060				q->pi_state->pi_mutex.owner,
2061				q->pi_state->owner);
2062
2063out:
2064	return ret ? ret : locked;
2065}
2066
2067/**
2068 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2069 * @hb:		the futex hash bucket, must be locked by the caller
2070 * @q:		the futex_q to queue up on
2071 * @timeout:	the prepared hrtimer_sleeper, or null for no timeout
2072 */
2073static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2074				struct hrtimer_sleeper *timeout)
2075{
2076	/*
2077	 * The task state is guaranteed to be set before another task can
2078	 * wake it. set_current_state() is implemented using set_mb() and
2079	 * queue_me() calls spin_unlock() upon completion, both serializing
2080	 * access to the hash list and forcing another memory barrier.
2081	 */
2082	set_current_state(TASK_INTERRUPTIBLE);
2083	queue_me(q, hb);
2084
2085	/* Arm the timer */
2086	if (timeout) {
2087		hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2088		if (!hrtimer_active(&timeout->timer))
2089			timeout->task = NULL;
2090	}
2091
2092	/*
2093	 * If we have been removed from the hash list, then another task
2094	 * has tried to wake us, and we can skip the call to schedule().
2095	 */
2096	if (likely(!plist_node_empty(&q->list))) {
2097		/*
2098		 * If the timer has already expired, current will already be
2099		 * flagged for rescheduling. Only call schedule if there
2100		 * is no timeout, or if it has yet to expire.
2101		 */
2102		if (!timeout || timeout->task)
2103			freezable_schedule();
2104	}
2105	__set_current_state(TASK_RUNNING);
2106}
2107
2108/**
2109 * futex_wait_setup() - Prepare to wait on a futex
2110 * @uaddr:	the futex userspace address
2111 * @val:	the expected value
2112 * @flags:	futex flags (FLAGS_SHARED, etc.)
2113 * @q:		the associated futex_q
2114 * @hb:		storage for hash_bucket pointer to be returned to caller
2115 *
2116 * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2117 * compare it with the expected value.  Handle atomic faults internally.
2118 * Return with the hb lock held and a q.key reference on success, and unlocked
2119 * with no q.key reference on failure.
2120 *
2121 * Return:
2122 *  0 - uaddr contains val and hb has been locked;
2123 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2124 */
2125static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2126			   struct futex_q *q, struct futex_hash_bucket **hb)
2127{
2128	u32 uval;
2129	int ret;
2130
2131	/*
2132	 * Access the page AFTER the hash-bucket is locked.
2133	 * Order is important:
2134	 *
2135	 *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2136	 *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2137	 *
2138	 * The basic logical guarantee of a futex is that it blocks ONLY
2139	 * if cond(var) is known to be true at the time of blocking, for
2140	 * any cond.  If we locked the hash-bucket after testing *uaddr, that
2141	 * would open a race condition where we could block indefinitely with
2142	 * cond(var) false, which would violate the guarantee.
2143	 *
2144	 * On the other hand, we insert q and release the hash-bucket only
2145	 * after testing *uaddr.  This guarantees that futex_wait() will NOT
2146	 * absorb a wakeup if *uaddr does not match the desired values
2147	 * while the syscall executes.
2148	 */
2149retry:
2150	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2151	if (unlikely(ret != 0))
2152		return ret;
2153
2154retry_private:
2155	*hb = queue_lock(q);
2156
2157	ret = get_futex_value_locked(&uval, uaddr);
2158
2159	if (ret) {
2160		queue_unlock(*hb);
2161
2162		ret = get_user(uval, uaddr);
2163		if (ret)
2164			goto out;
2165
2166		if (!(flags & FLAGS_SHARED))
2167			goto retry_private;
2168
2169		put_futex_key(&q->key);
2170		goto retry;
2171	}
2172
2173	if (uval != val) {
2174		queue_unlock(*hb);
2175		ret = -EWOULDBLOCK;
2176	}
2177
2178out:
2179	if (ret)
2180		put_futex_key(&q->key);
2181	return ret;
2182}
2183
2184static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2185		      ktime_t *abs_time, u32 bitset)
2186{
2187	struct hrtimer_sleeper timeout, *to = NULL;
2188	struct restart_block *restart;
2189	struct futex_hash_bucket *hb;
2190	struct futex_q q = futex_q_init;
2191	int ret;
2192
2193	if (!bitset)
2194		return -EINVAL;
2195	q.bitset = bitset;
2196
2197	if (abs_time) {
2198		to = &timeout;
2199
2200		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2201				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2202				      HRTIMER_MODE_ABS);
2203		hrtimer_init_sleeper(to, current);
2204		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2205					     current->timer_slack_ns);
2206	}
2207
2208retry:
2209	/*
2210	 * Prepare to wait on uaddr. On success, holds hb lock and increments
2211	 * q.key refs.
2212	 */
2213	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2214	if (ret)
2215		goto out;
2216
2217	/* queue_me and wait for wakeup, timeout, or a signal. */
2218	futex_wait_queue_me(hb, &q, to);
2219
2220	/* If we were woken (and unqueued), we succeeded, whatever. */
2221	ret = 0;
2222	/* unqueue_me() drops q.key ref */
2223	if (!unqueue_me(&q))
2224		goto out;
2225	ret = -ETIMEDOUT;
2226	if (to && !to->task)
2227		goto out;
2228
2229	/*
2230	 * We expect signal_pending(current), but we might be the
2231	 * victim of a spurious wakeup as well.
2232	 */
2233	if (!signal_pending(current))
2234		goto retry;
2235
2236	ret = -ERESTARTSYS;
2237	if (!abs_time)
2238		goto out;
2239
2240	restart = &current_thread_info()->restart_block;
2241	restart->fn = futex_wait_restart;
2242	restart->futex.uaddr = uaddr;
2243	restart->futex.val = val;
2244	restart->futex.time = abs_time->tv64;
2245	restart->futex.bitset = bitset;
2246	restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2247
2248	ret = -ERESTART_RESTARTBLOCK;
2249
2250out:
2251	if (to) {
2252		hrtimer_cancel(&to->timer);
2253		destroy_hrtimer_on_stack(&to->timer);
2254	}
2255	return ret;
2256}
2257
2258
2259static long futex_wait_restart(struct restart_block *restart)
2260{
2261	u32 __user *uaddr = restart->futex.uaddr;
2262	ktime_t t, *tp = NULL;
2263
2264	if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2265		t.tv64 = restart->futex.time;
2266		tp = &t;
2267	}
2268	restart->fn = do_no_restart_syscall;
2269
2270	return (long)futex_wait(uaddr, restart->futex.flags,
2271				restart->futex.val, tp, restart->futex.bitset);
2272}
2273
2274
2275/*
2276 * Userspace tried a 0 -> TID atomic transition of the futex value
2277 * and failed. The kernel side here does the whole locking operation:
2278 * if there are waiters then it will block, it does PI, etc. (Due to
2279 * races the kernel might see a 0 value of the futex too.)
 
 
 
2280 */
2281static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2282			 ktime_t *time, int trylock)
2283{
2284	struct hrtimer_sleeper timeout, *to = NULL;
 
 
2285	struct futex_hash_bucket *hb;
2286	struct futex_q q = futex_q_init;
2287	int res, ret;
2288
 
 
 
2289	if (refill_pi_state_cache())
2290		return -ENOMEM;
2291
2292	if (time) {
2293		to = &timeout;
2294		hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2295				      HRTIMER_MODE_ABS);
2296		hrtimer_init_sleeper(to, current);
2297		hrtimer_set_expires(&to->timer, *time);
2298	}
2299
2300retry:
2301	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2302	if (unlikely(ret != 0))
2303		goto out;
2304
2305retry_private:
2306	hb = queue_lock(&q);
2307
2308	ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
 
2309	if (unlikely(ret)) {
 
 
 
 
2310		switch (ret) {
2311		case 1:
2312			/* We got the lock. */
2313			ret = 0;
2314			goto out_unlock_put_key;
2315		case -EFAULT:
2316			goto uaddr_faulted;
 
2317		case -EAGAIN:
2318			/*
2319			 * Task is exiting and we just wait for the
2320			 * exit to complete.
 
 
2321			 */
2322			queue_unlock(hb);
2323			put_futex_key(&q.key);
 
 
 
 
 
2324			cond_resched();
2325			goto retry;
2326		default:
2327			goto out_unlock_put_key;
2328		}
2329	}
2330
 
 
2331	/*
2332	 * Only actually queue now that the atomic ops are done:
2333	 */
2334	queue_me(&q, hb);
2335
2336	WARN_ON(!q.pi_state);
2337	/*
2338	 * Block on the PI mutex:
2339	 */
2340	if (!trylock)
2341		ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
2342	else {
2343		ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2344		/* Fixup the trylock return value: */
2345		ret = ret ? 0 : -EWOULDBLOCK;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2346	}
2347
 
 
 
 
 
 
2348	spin_lock(q.lock_ptr);
2349	/*
 
 
 
 
 
 
 
 
 
 
 
 
 
2350	 * Fixup the pi_state owner and possibly acquire the lock if we
2351	 * haven't already.
2352	 */
2353	res = fixup_owner(uaddr, &q, !ret);
2354	/*
2355	 * If fixup_owner() returned an error, proprogate that.  If it acquired
2356	 * the lock, clear our -ETIMEDOUT or -EINTR.
2357	 */
2358	if (res)
2359		ret = (res < 0) ? res : 0;
2360
2361	/*
2362	 * If fixup_owner() faulted and was unable to handle the fault, unlock
2363	 * it and return the fault to userspace.
2364	 */
2365	if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2366		rt_mutex_unlock(&q.pi_state->pi_mutex);
2367
2368	/* Unqueue and drop the lock */
2369	unqueue_me_pi(&q);
2370
2371	goto out_put_key;
2372
2373out_unlock_put_key:
2374	queue_unlock(hb);
2375
2376out_put_key:
2377	put_futex_key(&q.key);
2378out:
2379	if (to)
 
2380		destroy_hrtimer_on_stack(&to->timer);
 
2381	return ret != -EINTR ? ret : -ERESTARTNOINTR;
2382
2383uaddr_faulted:
2384	queue_unlock(hb);
2385
2386	ret = fault_in_user_writeable(uaddr);
2387	if (ret)
2388		goto out_put_key;
2389
2390	if (!(flags & FLAGS_SHARED))
2391		goto retry_private;
2392
2393	put_futex_key(&q.key);
2394	goto retry;
2395}
2396
2397/*
2398 * Userspace attempted a TID -> 0 atomic transition, and failed.
2399 * This is the in-kernel slowpath: we look up the PI state (if any),
2400 * and do the rt-mutex unlock.
2401 */
2402static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2403{
2404	struct futex_hash_bucket *hb;
2405	struct futex_q *this, *next;
2406	union futex_key key = FUTEX_KEY_INIT;
2407	u32 uval, vpid = task_pid_vnr(current);
 
2408	int ret;
2409
 
 
 
2410retry:
2411	if (get_user(uval, uaddr))
2412		return -EFAULT;
2413	/*
2414	 * We release only a lock we actually own:
2415	 */
2416	if ((uval & FUTEX_TID_MASK) != vpid)
2417		return -EPERM;
2418
2419	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2420	if (unlikely(ret != 0))
2421		goto out;
2422
2423	hb = hash_futex(&key);
2424	spin_lock(&hb->lock);
2425
2426	/*
2427	 * To avoid races, try to do the TID -> 0 atomic transition
2428	 * again. If it succeeds then we can return without waking
2429	 * anyone else up. We only try this if neither the waiters nor
2430	 * the owner died bit are set.
2431	 */
2432	if (!(uval & ~FUTEX_TID_MASK) &&
2433	    cmpxchg_futex_value_locked(&uval, uaddr, vpid, 0))
2434		goto pi_faulted;
2435	/*
2436	 * Rare case: we managed to release the lock atomically,
2437	 * no need to wake anyone else up:
2438	 */
2439	if (unlikely(uval == vpid))
2440		goto out_unlock;
2441
2442	/*
2443	 * Ok, other tasks may need to be woken up - check waiters
2444	 * and do the wakeup if necessary:
2445	 */
2446	plist_for_each_entry_safe(this, next, &hb->chain, list) {
2447		if (!match_futex (&this->key, &key))
2448			continue;
2449		ret = wake_futex_pi(uaddr, uval, this);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2450		/*
2451		 * The atomic access to the futex value
2452		 * generated a pagefault, so retry the
2453		 * user-access and the wakeup:
2454		 */
2455		if (ret == -EFAULT)
2456			goto pi_faulted;
2457		goto out_unlock;
 
 
 
 
 
 
 
 
 
 
2458	}
 
2459	/*
2460	 * No waiters - kernel unlocks the futex:
 
 
 
 
2461	 */
2462	ret = unlock_futex_pi(uaddr, uval);
2463	if (ret == -EFAULT)
2464		goto pi_faulted;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2465
2466out_unlock:
2467	spin_unlock(&hb->lock);
2468	put_futex_key(&key);
2469
2470out:
2471	return ret;
2472
 
 
 
 
2473pi_faulted:
2474	spin_unlock(&hb->lock);
2475	put_futex_key(&key);
2476
2477	ret = fault_in_user_writeable(uaddr);
2478	if (!ret)
2479		goto retry;
2480
2481	return ret;
2482}
2483
2484/**
2485 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2486 * @hb:		the hash_bucket futex_q was original enqueued on
2487 * @q:		the futex_q woken while waiting to be requeued
2488 * @key2:	the futex_key of the requeue target futex
2489 * @timeout:	the timeout associated with the wait (NULL if none)
2490 *
2491 * Detect if the task was woken on the initial futex as opposed to the requeue
2492 * target futex.  If so, determine if it was a timeout or a signal that caused
2493 * the wakeup and return the appropriate error code to the caller.  Must be
2494 * called with the hb lock held.
2495 *
2496 * Return:
2497 *  0 = no early wakeup detected;
2498 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2499 */
2500static inline
2501int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2502				   struct futex_q *q, union futex_key *key2,
2503				   struct hrtimer_sleeper *timeout)
2504{
2505	int ret = 0;
2506
2507	/*
2508	 * With the hb lock held, we avoid races while we process the wakeup.
2509	 * We only need to hold hb (and not hb2) to ensure atomicity as the
2510	 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2511	 * It can't be requeued from uaddr2 to something else since we don't
2512	 * support a PI aware source futex for requeue.
2513	 */
2514	if (!match_futex(&q->key, key2)) {
2515		WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2516		/*
2517		 * We were woken prior to requeue by a timeout or a signal.
2518		 * Unqueue the futex_q and determine which it was.
2519		 */
2520		plist_del(&q->list, &hb->chain);
2521		hb_waiters_dec(hb);
2522
2523		/* Handle spurious wakeups gracefully */
2524		ret = -EWOULDBLOCK;
2525		if (timeout && !timeout->task)
2526			ret = -ETIMEDOUT;
2527		else if (signal_pending(current))
2528			ret = -ERESTARTNOINTR;
2529	}
2530	return ret;
2531}
2532
2533/**
2534 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2535 * @uaddr:	the futex we initially wait on (non-pi)
2536 * @flags:	futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2537 * 		the same type, no requeueing from private to shared, etc.
2538 * @val:	the expected value of uaddr
2539 * @abs_time:	absolute timeout
2540 * @bitset:	32 bit wakeup bitset set by userspace, defaults to all
2541 * @uaddr2:	the pi futex we will take prior to returning to user-space
2542 *
2543 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2544 * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2545 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2546 * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2547 * without one, the pi logic would not know which task to boost/deboost, if
2548 * there was a need to.
2549 *
2550 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2551 * via the following--
2552 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2553 * 2) wakeup on uaddr2 after a requeue
2554 * 3) signal
2555 * 4) timeout
2556 *
2557 * If 3, cleanup and return -ERESTARTNOINTR.
2558 *
2559 * If 2, we may then block on trying to take the rt_mutex and return via:
2560 * 5) successful lock
2561 * 6) signal
2562 * 7) timeout
2563 * 8) other lock acquisition failure
2564 *
2565 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2566 *
2567 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2568 *
2569 * Return:
2570 *  0 - On success;
2571 * <0 - On error
2572 */
2573static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2574				 u32 val, ktime_t *abs_time, u32 bitset,
2575				 u32 __user *uaddr2)
2576{
2577	struct hrtimer_sleeper timeout, *to = NULL;
2578	struct rt_mutex_waiter rt_waiter;
2579	struct rt_mutex *pi_mutex = NULL;
2580	struct futex_hash_bucket *hb;
2581	union futex_key key2 = FUTEX_KEY_INIT;
2582	struct futex_q q = futex_q_init;
2583	int res, ret;
2584
 
 
 
2585	if (uaddr == uaddr2)
2586		return -EINVAL;
2587
2588	if (!bitset)
2589		return -EINVAL;
2590
2591	if (abs_time) {
2592		to = &timeout;
2593		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2594				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2595				      HRTIMER_MODE_ABS);
2596		hrtimer_init_sleeper(to, current);
2597		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2598					     current->timer_slack_ns);
2599	}
2600
2601	/*
2602	 * The waiter is allocated on our stack, manipulated by the requeue
2603	 * code while we sleep on uaddr.
2604	 */
2605	debug_rt_mutex_init_waiter(&rt_waiter);
2606	RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2607	RB_CLEAR_NODE(&rt_waiter.tree_entry);
2608	rt_waiter.task = NULL;
2609
2610	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2611	if (unlikely(ret != 0))
2612		goto out;
2613
2614	q.bitset = bitset;
2615	q.rt_waiter = &rt_waiter;
2616	q.requeue_pi_key = &key2;
2617
2618	/*
2619	 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2620	 * count.
2621	 */
2622	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2623	if (ret)
2624		goto out_key2;
2625
2626	/*
2627	 * The check above which compares uaddrs is not sufficient for
2628	 * shared futexes. We need to compare the keys:
2629	 */
2630	if (match_futex(&q.key, &key2)) {
 
2631		ret = -EINVAL;
2632		goto out_put_keys;
2633	}
2634
2635	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
2636	futex_wait_queue_me(hb, &q, to);
2637
2638	spin_lock(&hb->lock);
2639	ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2640	spin_unlock(&hb->lock);
2641	if (ret)
2642		goto out_put_keys;
2643
2644	/*
2645	 * In order for us to be here, we know our q.key == key2, and since
2646	 * we took the hb->lock above, we also know that futex_requeue() has
2647	 * completed and we no longer have to concern ourselves with a wakeup
2648	 * race with the atomic proxy lock acquisition by the requeue code. The
2649	 * futex_requeue dropped our key1 reference and incremented our key2
2650	 * reference count.
2651	 */
2652
2653	/* Check if the requeue code acquired the second futex for us. */
 
 
 
2654	if (!q.rt_waiter) {
2655		/*
2656		 * Got the lock. We might not be the anticipated owner if we
2657		 * did a lock-steal - fix up the PI-state in that case.
2658		 */
2659		if (q.pi_state && (q.pi_state->owner != current)) {
2660			spin_lock(q.lock_ptr);
2661			ret = fixup_pi_state_owner(uaddr2, &q, current);
 
 
 
 
 
2662			spin_unlock(q.lock_ptr);
 
 
 
 
 
2663		}
2664	} else {
 
 
2665		/*
2666		 * We have been woken up by futex_unlock_pi(), a timeout, or a
2667		 * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2668		 * the pi_state.
2669		 */
2670		WARN_ON(!q.pi_state);
2671		pi_mutex = &q.pi_state->pi_mutex;
2672		ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2673		debug_rt_mutex_free_waiter(&rt_waiter);
2674
2675		spin_lock(q.lock_ptr);
 
 
 
 
2676		/*
2677		 * Fixup the pi_state owner and possibly acquire the lock if we
2678		 * haven't already.
2679		 */
2680		res = fixup_owner(uaddr2, &q, !ret);
2681		/*
2682		 * If fixup_owner() returned an error, proprogate that.  If it
2683		 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2684		 */
2685		if (res)
2686			ret = (res < 0) ? res : 0;
2687
2688		/* Unqueue and drop the lock. */
2689		unqueue_me_pi(&q);
 
2690	}
2691
2692	/*
2693	 * If fixup_pi_state_owner() faulted and was unable to handle the
2694	 * fault, unlock the rt_mutex and return the fault to userspace.
2695	 */
2696	if (ret == -EFAULT) {
2697		if (pi_mutex && rt_mutex_owner(pi_mutex) == current)
2698			rt_mutex_unlock(pi_mutex);
2699	} else if (ret == -EINTR) {
2700		/*
2701		 * We've already been requeued, but cannot restart by calling
2702		 * futex_lock_pi() directly. We could restart this syscall, but
2703		 * it would detect that the user space "val" changed and return
2704		 * -EWOULDBLOCK.  Save the overhead of the restart and return
2705		 * -EWOULDBLOCK directly.
2706		 */
2707		ret = -EWOULDBLOCK;
2708	}
2709
2710out_put_keys:
2711	put_futex_key(&q.key);
2712out_key2:
2713	put_futex_key(&key2);
2714
2715out:
2716	if (to) {
2717		hrtimer_cancel(&to->timer);
2718		destroy_hrtimer_on_stack(&to->timer);
2719	}
2720	return ret;
2721}
2722
2723/*
2724 * Support for robust futexes: the kernel cleans up held futexes at
2725 * thread exit time.
2726 *
2727 * Implementation: user-space maintains a per-thread list of locks it
2728 * is holding. Upon do_exit(), the kernel carefully walks this list,
2729 * and marks all locks that are owned by this thread with the
2730 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2731 * always manipulated with the lock held, so the list is private and
2732 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2733 * field, to allow the kernel to clean up if the thread dies after
2734 * acquiring the lock, but just before it could have added itself to
2735 * the list. There can only be one such pending lock.
2736 */
2737
2738/**
2739 * sys_set_robust_list() - Set the robust-futex list head of a task
2740 * @head:	pointer to the list-head
2741 * @len:	length of the list-head, as userspace expects
2742 */
2743SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2744		size_t, len)
2745{
2746	if (!futex_cmpxchg_enabled)
2747		return -ENOSYS;
2748	/*
2749	 * The kernel knows only one size for now:
2750	 */
2751	if (unlikely(len != sizeof(*head)))
2752		return -EINVAL;
2753
2754	current->robust_list = head;
2755
2756	return 0;
2757}
2758
2759/**
2760 * sys_get_robust_list() - Get the robust-futex list head of a task
2761 * @pid:	pid of the process [zero for current task]
2762 * @head_ptr:	pointer to a list-head pointer, the kernel fills it in
2763 * @len_ptr:	pointer to a length field, the kernel fills in the header size
2764 */
2765SYSCALL_DEFINE3(get_robust_list, int, pid,
2766		struct robust_list_head __user * __user *, head_ptr,
2767		size_t __user *, len_ptr)
2768{
2769	struct robust_list_head __user *head;
2770	unsigned long ret;
2771	struct task_struct *p;
2772
2773	if (!futex_cmpxchg_enabled)
2774		return -ENOSYS;
2775
2776	rcu_read_lock();
2777
2778	ret = -ESRCH;
2779	if (!pid)
2780		p = current;
2781	else {
2782		p = find_task_by_vpid(pid);
2783		if (!p)
2784			goto err_unlock;
2785	}
2786
2787	ret = -EPERM;
2788	if (!ptrace_may_access(p, PTRACE_MODE_READ))
2789		goto err_unlock;
2790
2791	head = p->robust_list;
2792	rcu_read_unlock();
2793
2794	if (put_user(sizeof(*head), len_ptr))
2795		return -EFAULT;
2796	return put_user(head, head_ptr);
2797
2798err_unlock:
2799	rcu_read_unlock();
2800
2801	return ret;
2802}
2803
 
 
 
 
2804/*
2805 * Process a futex-list entry, check whether it's owned by the
2806 * dying task, and do notification if so:
2807 */
2808int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
 
2809{
2810	u32 uval, uninitialized_var(nval), mval;
 
 
 
 
 
2811
2812retry:
2813	if (get_user(uval, uaddr))
2814		return -1;
2815
2816	if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2817		/*
2818		 * Ok, this dying thread is truly holding a futex
2819		 * of interest. Set the OWNER_DIED bit atomically
2820		 * via cmpxchg, and if the value had FUTEX_WAITERS
2821		 * set, wake up a waiter (if any). (We have to do a
2822		 * futex_wake() even if OWNER_DIED is already set -
2823		 * to handle the rare but possible case of recursive
2824		 * thread-death.) The rest of the cleanup is done in
2825		 * userspace.
2826		 */
2827		mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2828		/*
2829		 * We are not holding a lock here, but we want to have
2830		 * the pagefault_disable/enable() protection because
2831		 * we want to handle the fault gracefully. If the
2832		 * access fails we try to fault in the futex with R/W
2833		 * verification via get_user_pages. get_user() above
2834		 * does not guarantee R/W access. If that fails we
2835		 * give up and leave the futex locked.
2836		 */
2837		if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2838			if (fault_in_user_writeable(uaddr))
2839				return -1;
2840			goto retry;
2841		}
2842		if (nval != uval)
 
2843			goto retry;
2844
2845		/*
2846		 * Wake robust non-PI futexes here. The wakeup of
2847		 * PI futexes happens in exit_pi_state():
2848		 */
2849		if (!pi && (uval & FUTEX_WAITERS))
2850			futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2851	}
 
 
 
 
 
 
 
 
 
 
 
2852	return 0;
2853}
2854
2855/*
2856 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2857 */
2858static inline int fetch_robust_entry(struct robust_list __user **entry,
2859				     struct robust_list __user * __user *head,
2860				     unsigned int *pi)
2861{
2862	unsigned long uentry;
2863
2864	if (get_user(uentry, (unsigned long __user *)head))
2865		return -EFAULT;
2866
2867	*entry = (void __user *)(uentry & ~1UL);
2868	*pi = uentry & 1;
2869
2870	return 0;
2871}
2872
2873/*
2874 * Walk curr->robust_list (very carefully, it's a userspace list!)
2875 * and mark any locks found there dead, and notify any waiters.
2876 *
2877 * We silently return on any sign of list-walking problem.
2878 */
2879void exit_robust_list(struct task_struct *curr)
2880{
2881	struct robust_list_head __user *head = curr->robust_list;
2882	struct robust_list __user *entry, *next_entry, *pending;
2883	unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2884	unsigned int uninitialized_var(next_pi);
2885	unsigned long futex_offset;
2886	int rc;
2887
2888	if (!futex_cmpxchg_enabled)
2889		return;
2890
2891	/*
2892	 * Fetch the list head (which was registered earlier, via
2893	 * sys_set_robust_list()):
2894	 */
2895	if (fetch_robust_entry(&entry, &head->list.next, &pi))
2896		return;
2897	/*
2898	 * Fetch the relative futex offset:
2899	 */
2900	if (get_user(futex_offset, &head->futex_offset))
2901		return;
2902	/*
2903	 * Fetch any possibly pending lock-add first, and handle it
2904	 * if it exists:
2905	 */
2906	if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2907		return;
2908
2909	next_entry = NULL;	/* avoid warning with gcc */
2910	while (entry != &head->list) {
2911		/*
2912		 * Fetch the next entry in the list before calling
2913		 * handle_futex_death:
2914		 */
2915		rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2916		/*
2917		 * A pending lock might already be on the list, so
2918		 * don't process it twice:
2919		 */
2920		if (entry != pending)
2921			if (handle_futex_death((void __user *)entry + futex_offset,
2922						curr, pi))
2923				return;
 
2924		if (rc)
2925			return;
2926		entry = next_entry;
2927		pi = next_pi;
2928		/*
2929		 * Avoid excessively long or circular lists:
2930		 */
2931		if (!--limit)
2932			break;
2933
2934		cond_resched();
2935	}
2936
2937	if (pending)
2938		handle_futex_death((void __user *)pending + futex_offset,
2939				   curr, pip);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2940}
2941
2942long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2943		u32 __user *uaddr2, u32 val2, u32 val3)
2944{
2945	int cmd = op & FUTEX_CMD_MASK;
2946	unsigned int flags = 0;
2947
2948	if (!(op & FUTEX_PRIVATE_FLAG))
2949		flags |= FLAGS_SHARED;
2950
2951	if (op & FUTEX_CLOCK_REALTIME) {
2952		flags |= FLAGS_CLOCKRT;
2953		if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
 
2954			return -ENOSYS;
2955	}
2956
2957	switch (cmd) {
2958	case FUTEX_LOCK_PI:
 
2959	case FUTEX_UNLOCK_PI:
2960	case FUTEX_TRYLOCK_PI:
2961	case FUTEX_WAIT_REQUEUE_PI:
2962	case FUTEX_CMP_REQUEUE_PI:
2963		if (!futex_cmpxchg_enabled)
2964			return -ENOSYS;
2965	}
2966
2967	switch (cmd) {
2968	case FUTEX_WAIT:
2969		val3 = FUTEX_BITSET_MATCH_ANY;
 
2970	case FUTEX_WAIT_BITSET:
2971		return futex_wait(uaddr, flags, val, timeout, val3);
2972	case FUTEX_WAKE:
2973		val3 = FUTEX_BITSET_MATCH_ANY;
 
2974	case FUTEX_WAKE_BITSET:
2975		return futex_wake(uaddr, flags, val, val3);
2976	case FUTEX_REQUEUE:
2977		return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2978	case FUTEX_CMP_REQUEUE:
2979		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2980	case FUTEX_WAKE_OP:
2981		return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2982	case FUTEX_LOCK_PI:
2983		return futex_lock_pi(uaddr, flags, val, timeout, 0);
 
 
 
2984	case FUTEX_UNLOCK_PI:
2985		return futex_unlock_pi(uaddr, flags);
2986	case FUTEX_TRYLOCK_PI:
2987		return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2988	case FUTEX_WAIT_REQUEUE_PI:
2989		val3 = FUTEX_BITSET_MATCH_ANY;
2990		return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2991					     uaddr2);
2992	case FUTEX_CMP_REQUEUE_PI:
2993		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2994	}
2995	return -ENOSYS;
2996}
2997
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2998
2999SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3000		struct timespec __user *, utime, u32 __user *, uaddr2,
3001		u32, val3)
3002{
3003	struct timespec ts;
3004	ktime_t t, *tp = NULL;
3005	u32 val2 = 0;
3006	int cmd = op & FUTEX_CMD_MASK;
3007
3008	if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3009		      cmd == FUTEX_WAIT_BITSET ||
3010		      cmd == FUTEX_WAIT_REQUEUE_PI)) {
3011		if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3012			return -EFAULT;
3013		if (!timespec_valid(&ts))
3014			return -EINVAL;
3015
3016		t = timespec_to_ktime(ts);
3017		if (cmd == FUTEX_WAIT)
3018			t = ktime_add_safe(ktime_get(), t);
3019		tp = &t;
3020	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3021	/*
3022	 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3023	 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3024	 */
3025	if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3026	    cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3027		val2 = (u32) (unsigned long) utime;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3028
3029	return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3030}
 
3031
3032static void __init futex_detect_cmpxchg(void)
3033{
3034#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3035	u32 curval;
3036
3037	/*
3038	 * This will fail and we want it. Some arch implementations do
3039	 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3040	 * functionality. We want to know that before we call in any
3041	 * of the complex code paths. Also we want to prevent
3042	 * registration of robust lists in that case. NULL is
3043	 * guaranteed to fault and we get -EFAULT on functional
3044	 * implementation, the non-functional ones will return
3045	 * -ENOSYS.
3046	 */
3047	if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3048		futex_cmpxchg_enabled = 1;
3049#endif
3050}
3051
3052static int __init futex_init(void)
3053{
3054	unsigned int futex_shift;
3055	unsigned long i;
3056
3057#if CONFIG_BASE_SMALL
3058	futex_hashsize = 16;
3059#else
3060	futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3061#endif
3062
3063	futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3064					       futex_hashsize, 0,
3065					       futex_hashsize < 256 ? HASH_SMALL : 0,
3066					       &futex_shift, NULL,
3067					       futex_hashsize, futex_hashsize);
3068	futex_hashsize = 1UL << futex_shift;
3069
3070	futex_detect_cmpxchg();
3071
3072	for (i = 0; i < futex_hashsize; i++) {
3073		atomic_set(&futex_queues[i].waiters, 0);
3074		plist_head_init(&futex_queues[i].chain);
3075		spin_lock_init(&futex_queues[i].lock);
3076	}
3077
3078	return 0;
3079}
3080__initcall(futex_init);
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 *  Fast Userspace Mutexes (which I call "Futexes!").
   4 *  (C) Rusty Russell, IBM 2002
   5 *
   6 *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
   7 *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
   8 *
   9 *  Removed page pinning, fix privately mapped COW pages and other cleanups
  10 *  (C) Copyright 2003, 2004 Jamie Lokier
  11 *
  12 *  Robust futex support started by Ingo Molnar
  13 *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
  14 *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
  15 *
  16 *  PI-futex support started by Ingo Molnar and Thomas Gleixner
  17 *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  18 *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
  19 *
  20 *  PRIVATE futexes by Eric Dumazet
  21 *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
  22 *
  23 *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
  24 *  Copyright (C) IBM Corporation, 2009
  25 *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
  26 *
  27 *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
  28 *  enough at me, Linus for the original (flawed) idea, Matthew
  29 *  Kirkwood for proof-of-concept implementation.
  30 *
  31 *  "The futexes are also cursed."
  32 *  "But they come in a choice of three flavours!"
  33 */
  34#include <linux/compat.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  35#include <linux/jhash.h>
 
 
 
  36#include <linux/pagemap.h>
  37#include <linux/syscalls.h>
 
 
 
 
 
 
 
 
  38#include <linux/freezer.h>
  39#include <linux/memblock.h>
  40#include <linux/fault-inject.h>
  41#include <linux/time_namespace.h>
  42
  43#include <asm/futex.h>
  44
  45#include "locking/rtmutex_common.h"
  46
  47/*
  48 * READ this before attempting to hack on futexes!
  49 *
  50 * Basic futex operation and ordering guarantees
  51 * =============================================
  52 *
  53 * The waiter reads the futex value in user space and calls
  54 * futex_wait(). This function computes the hash bucket and acquires
  55 * the hash bucket lock. After that it reads the futex user space value
  56 * again and verifies that the data has not changed. If it has not changed
  57 * it enqueues itself into the hash bucket, releases the hash bucket lock
  58 * and schedules.
  59 *
  60 * The waker side modifies the user space value of the futex and calls
  61 * futex_wake(). This function computes the hash bucket and acquires the
  62 * hash bucket lock. Then it looks for waiters on that futex in the hash
  63 * bucket and wakes them.
  64 *
  65 * In futex wake up scenarios where no tasks are blocked on a futex, taking
  66 * the hb spinlock can be avoided and simply return. In order for this
  67 * optimization to work, ordering guarantees must exist so that the waiter
  68 * being added to the list is acknowledged when the list is concurrently being
  69 * checked by the waker, avoiding scenarios like the following:
  70 *
  71 * CPU 0                               CPU 1
  72 * val = *futex;
  73 * sys_futex(WAIT, futex, val);
  74 *   futex_wait(futex, val);
  75 *   uval = *futex;
  76 *                                     *futex = newval;
  77 *                                     sys_futex(WAKE, futex);
  78 *                                       futex_wake(futex);
  79 *                                       if (queue_empty())
  80 *                                         return;
  81 *   if (uval == val)
  82 *      lock(hash_bucket(futex));
  83 *      queue();
  84 *     unlock(hash_bucket(futex));
  85 *     schedule();
  86 *
  87 * This would cause the waiter on CPU 0 to wait forever because it
  88 * missed the transition of the user space value from val to newval
  89 * and the waker did not find the waiter in the hash bucket queue.
  90 *
  91 * The correct serialization ensures that a waiter either observes
  92 * the changed user space value before blocking or is woken by a
  93 * concurrent waker:
  94 *
  95 * CPU 0                                 CPU 1
  96 * val = *futex;
  97 * sys_futex(WAIT, futex, val);
  98 *   futex_wait(futex, val);
  99 *
 100 *   waiters++; (a)
 101 *   smp_mb(); (A) <-- paired with -.
 102 *                                  |
 103 *   lock(hash_bucket(futex));      |
 104 *                                  |
 105 *   uval = *futex;                 |
 106 *                                  |        *futex = newval;
 107 *                                  |        sys_futex(WAKE, futex);
 108 *                                  |          futex_wake(futex);
 109 *                                  |
 110 *                                  `--------> smp_mb(); (B)
 111 *   if (uval == val)
 112 *     queue();
 113 *     unlock(hash_bucket(futex));
 114 *     schedule();                         if (waiters)
 115 *                                           lock(hash_bucket(futex));
 116 *   else                                    wake_waiters(futex);
 117 *     waiters--; (b)                        unlock(hash_bucket(futex));
 118 *
 119 * Where (A) orders the waiters increment and the futex value read through
 120 * atomic operations (see hb_waiters_inc) and where (B) orders the write
 121 * to futex and the waiters read (see hb_waiters_pending()).
 
 
 122 *
 123 * This yields the following case (where X:=waiters, Y:=futex):
 124 *
 125 *	X = Y = 0
 126 *
 127 *	w[X]=1		w[Y]=1
 128 *	MB		MB
 129 *	r[Y]=y		r[X]=x
 130 *
 131 * Which guarantees that x==0 && y==0 is impossible; which translates back into
 132 * the guarantee that we cannot both miss the futex variable change and the
 133 * enqueue.
 134 *
 135 * Note that a new waiter is accounted for in (a) even when it is possible that
 136 * the wait call can return error, in which case we backtrack from it in (b).
 137 * Refer to the comment in queue_lock().
 138 *
 139 * Similarly, in order to account for waiters being requeued on another
 140 * address we always increment the waiters for the destination bucket before
 141 * acquiring the lock. It then decrements them again  after releasing it -
 142 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
 143 * will do the additional required waiter count housekeeping. This is done for
 144 * double_lock_hb() and double_unlock_hb(), respectively.
 145 */
 146
 147#ifdef CONFIG_HAVE_FUTEX_CMPXCHG
 148#define futex_cmpxchg_enabled 1
 149#else
 150static int  __read_mostly futex_cmpxchg_enabled;
 151#endif
 152
 153/*
 154 * Futex flags used to encode options to functions and preserve them across
 155 * restarts.
 156 */
 157#ifdef CONFIG_MMU
 158# define FLAGS_SHARED		0x01
 159#else
 160/*
 161 * NOMMU does not have per process address space. Let the compiler optimize
 162 * code away.
 163 */
 164# define FLAGS_SHARED		0x00
 165#endif
 166#define FLAGS_CLOCKRT		0x02
 167#define FLAGS_HAS_TIMEOUT	0x04
 168
 169/*
 170 * Priority Inheritance state:
 171 */
 172struct futex_pi_state {
 173	/*
 174	 * list of 'owned' pi_state instances - these have to be
 175	 * cleaned up in do_exit() if the task exits prematurely:
 176	 */
 177	struct list_head list;
 178
 179	/*
 180	 * The PI object:
 181	 */
 182	struct rt_mutex pi_mutex;
 183
 184	struct task_struct *owner;
 185	refcount_t refcount;
 186
 187	union futex_key key;
 188} __randomize_layout;
 189
 190/**
 191 * struct futex_q - The hashed futex queue entry, one per waiting task
 192 * @list:		priority-sorted list of tasks waiting on this futex
 193 * @task:		the task waiting on the futex
 194 * @lock_ptr:		the hash bucket lock
 195 * @key:		the key the futex is hashed on
 196 * @pi_state:		optional priority inheritance state
 197 * @rt_waiter:		rt_waiter storage for use with requeue_pi
 198 * @requeue_pi_key:	the requeue_pi target futex key
 199 * @bitset:		bitset for the optional bitmasked wakeup
 200 *
 201 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
 202 * we can wake only the relevant ones (hashed queues may be shared).
 203 *
 204 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
 205 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
 206 * The order of wakeup is always to make the first condition true, then
 207 * the second.
 208 *
 209 * PI futexes are typically woken before they are removed from the hash list via
 210 * the rt_mutex code. See unqueue_me_pi().
 211 */
 212struct futex_q {
 213	struct plist_node list;
 214
 215	struct task_struct *task;
 216	spinlock_t *lock_ptr;
 217	union futex_key key;
 218	struct futex_pi_state *pi_state;
 219	struct rt_mutex_waiter *rt_waiter;
 220	union futex_key *requeue_pi_key;
 221	u32 bitset;
 222} __randomize_layout;
 223
 224static const struct futex_q futex_q_init = {
 225	/* list gets initialized in queue_me()*/
 226	.key = FUTEX_KEY_INIT,
 227	.bitset = FUTEX_BITSET_MATCH_ANY
 228};
 229
 230/*
 231 * Hash buckets are shared by all the futex_keys that hash to the same
 232 * location.  Each key may have multiple futex_q structures, one for each task
 233 * waiting on a futex.
 234 */
 235struct futex_hash_bucket {
 236	atomic_t waiters;
 237	spinlock_t lock;
 238	struct plist_head chain;
 239} ____cacheline_aligned_in_smp;
 240
 241/*
 242 * The base of the bucket array and its size are always used together
 243 * (after initialization only in hash_futex()), so ensure that they
 244 * reside in the same cacheline.
 245 */
 246static struct {
 247	struct futex_hash_bucket *queues;
 248	unsigned long            hashsize;
 249} __futex_data __read_mostly __aligned(2*sizeof(long));
 250#define futex_queues   (__futex_data.queues)
 251#define futex_hashsize (__futex_data.hashsize)
 252
 253
 254/*
 255 * Fault injections for futexes.
 256 */
 257#ifdef CONFIG_FAIL_FUTEX
 258
 259static struct {
 260	struct fault_attr attr;
 261
 262	bool ignore_private;
 263} fail_futex = {
 264	.attr = FAULT_ATTR_INITIALIZER,
 265	.ignore_private = false,
 266};
 267
 268static int __init setup_fail_futex(char *str)
 269{
 270	return setup_fault_attr(&fail_futex.attr, str);
 271}
 272__setup("fail_futex=", setup_fail_futex);
 273
 274static bool should_fail_futex(bool fshared)
 275{
 276	if (fail_futex.ignore_private && !fshared)
 277		return false;
 278
 279	return should_fail(&fail_futex.attr, 1);
 280}
 281
 282#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 283
 284static int __init fail_futex_debugfs(void)
 285{
 286	umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
 287	struct dentry *dir;
 288
 289	dir = fault_create_debugfs_attr("fail_futex", NULL,
 290					&fail_futex.attr);
 291	if (IS_ERR(dir))
 292		return PTR_ERR(dir);
 293
 294	debugfs_create_bool("ignore-private", mode, dir,
 295			    &fail_futex.ignore_private);
 296	return 0;
 297}
 298
 299late_initcall(fail_futex_debugfs);
 300
 301#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
 302
 303#else
 304static inline bool should_fail_futex(bool fshared)
 305{
 306	return false;
 307}
 308#endif /* CONFIG_FAIL_FUTEX */
 309
 310#ifdef CONFIG_COMPAT
 311static void compat_exit_robust_list(struct task_struct *curr);
 312#endif
 313
 314/*
 315 * Reflects a new waiter being added to the waitqueue.
 316 */
 317static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
 318{
 319#ifdef CONFIG_SMP
 320	atomic_inc(&hb->waiters);
 321	/*
 322	 * Full barrier (A), see the ordering comment above.
 323	 */
 324	smp_mb__after_atomic();
 325#endif
 326}
 327
 328/*
 329 * Reflects a waiter being removed from the waitqueue by wakeup
 330 * paths.
 331 */
 332static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
 333{
 334#ifdef CONFIG_SMP
 335	atomic_dec(&hb->waiters);
 336#endif
 337}
 338
 339static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
 340{
 341#ifdef CONFIG_SMP
 342	/*
 343	 * Full barrier (B), see the ordering comment above.
 344	 */
 345	smp_mb();
 346	return atomic_read(&hb->waiters);
 347#else
 348	return 1;
 349#endif
 350}
 351
 352/**
 353 * hash_futex - Return the hash bucket in the global hash
 354 * @key:	Pointer to the futex key for which the hash is calculated
 355 *
 356 * We hash on the keys returned from get_futex_key (see below) and return the
 357 * corresponding hash bucket in the global hash.
 358 */
 359static struct futex_hash_bucket *hash_futex(union futex_key *key)
 360{
 361	u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
 
 362			  key->both.offset);
 363
 364	return &futex_queues[hash & (futex_hashsize - 1)];
 365}
 366
 367
 368/**
 369 * match_futex - Check whether two futex keys are equal
 370 * @key1:	Pointer to key1
 371 * @key2:	Pointer to key2
 372 *
 373 * Return 1 if two futex_keys are equal, 0 otherwise.
 374 */
 375static inline int match_futex(union futex_key *key1, union futex_key *key2)
 376{
 377	return (key1 && key2
 378		&& key1->both.word == key2->both.word
 379		&& key1->both.ptr == key2->both.ptr
 380		&& key1->both.offset == key2->both.offset);
 381}
 382
 383enum futex_access {
 384	FUTEX_READ,
 385	FUTEX_WRITE
 386};
 387
 388/**
 389 * futex_setup_timer - set up the sleeping hrtimer.
 390 * @time:	ptr to the given timeout value
 391 * @timeout:	the hrtimer_sleeper structure to be set up
 392 * @flags:	futex flags
 393 * @range_ns:	optional range in ns
 394 *
 395 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
 396 *	   value given
 397 */
 398static inline struct hrtimer_sleeper *
 399futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
 400		  int flags, u64 range_ns)
 401{
 402	if (!time)
 403		return NULL;
 404
 405	hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
 406				      CLOCK_REALTIME : CLOCK_MONOTONIC,
 407				      HRTIMER_MODE_ABS);
 408	/*
 409	 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
 410	 * effectively the same as calling hrtimer_set_expires().
 411	 */
 412	hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
 413
 414	return timeout;
 415}
 416
 417/*
 418 * Generate a machine wide unique identifier for this inode.
 419 *
 420 * This relies on u64 not wrapping in the life-time of the machine; which with
 421 * 1ns resolution means almost 585 years.
 422 *
 423 * This further relies on the fact that a well formed program will not unmap
 424 * the file while it has a (shared) futex waiting on it. This mapping will have
 425 * a file reference which pins the mount and inode.
 426 *
 427 * If for some reason an inode gets evicted and read back in again, it will get
 428 * a new sequence number and will _NOT_ match, even though it is the exact same
 429 * file.
 430 *
 431 * It is important that match_futex() will never have a false-positive, esp.
 432 * for PI futexes that can mess up the state. The above argues that false-negatives
 433 * are only possible for malformed programs.
 434 */
 435static u64 get_inode_sequence_number(struct inode *inode)
 436{
 437	static atomic64_t i_seq;
 438	u64 old;
 
 
 
 439
 440	/* Does the inode already have a sequence number? */
 441	old = atomic64_read(&inode->i_sequence);
 442	if (likely(old))
 443		return old;
 444
 445	for (;;) {
 446		u64 new = atomic64_add_return(1, &i_seq);
 447		if (WARN_ON_ONCE(!new))
 448			continue;
 449
 450		old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
 451		if (old)
 452			return old;
 453		return new;
 454	}
 455}
 456
 457/**
 458 * get_futex_key() - Get parameters which are the keys for a futex
 459 * @uaddr:	virtual address of the futex
 460 * @fshared:	false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
 461 * @key:	address where result is stored.
 462 * @rw:		mapping needs to be read/write (values: FUTEX_READ,
 463 *              FUTEX_WRITE)
 464 *
 465 * Return: a negative error code or 0
 466 *
 467 * The key words are stored in @key on success.
 468 *
 469 * For shared mappings (when @fshared), the key is:
 470 *
 471 *   ( inode->i_sequence, page->index, offset_within_page )
 472 *
 473 * [ also see get_inode_sequence_number() ]
 474 *
 475 * For private mappings (or when !@fshared), the key is:
 476 *
 477 *   ( current->mm, address, 0 )
 478 *
 479 * This allows (cross process, where applicable) identification of the futex
 480 * without keeping the page pinned for the duration of the FUTEX_WAIT.
 481 *
 482 * lock_page() might sleep, the caller should not hold a spinlock.
 483 */
 484static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
 485			 enum futex_access rw)
 486{
 487	unsigned long address = (unsigned long)uaddr;
 488	struct mm_struct *mm = current->mm;
 489	struct page *page, *tail;
 490	struct address_space *mapping;
 491	int err, ro = 0;
 492
 493	/*
 494	 * The futex address must be "naturally" aligned.
 495	 */
 496	key->both.offset = address % PAGE_SIZE;
 497	if (unlikely((address % sizeof(u32)) != 0))
 498		return -EINVAL;
 499	address -= key->both.offset;
 500
 501	if (unlikely(!access_ok(uaddr, sizeof(u32))))
 502		return -EFAULT;
 503
 504	if (unlikely(should_fail_futex(fshared)))
 505		return -EFAULT;
 506
 507	/*
 508	 * PROCESS_PRIVATE futexes are fast.
 509	 * As the mm cannot disappear under us and the 'key' only needs
 510	 * virtual address, we dont even have to find the underlying vma.
 511	 * Note : We do have to check 'uaddr' is a valid user address,
 512	 *        but access_ok() should be faster than find_vma()
 513	 */
 514	if (!fshared) {
 515		key->private.mm = mm;
 516		key->private.address = address;
 
 517		return 0;
 518	}
 519
 520again:
 521	/* Ignore any VERIFY_READ mapping (futex common case) */
 522	if (unlikely(should_fail_futex(true)))
 523		return -EFAULT;
 524
 525	err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
 526	/*
 527	 * If write access is not required (eg. FUTEX_WAIT), try
 528	 * and get read-only access.
 529	 */
 530	if (err == -EFAULT && rw == FUTEX_READ) {
 531		err = get_user_pages_fast(address, 1, 0, &page);
 532		ro = 1;
 533	}
 534	if (err < 0)
 535		return err;
 536	else
 537		err = 0;
 538
 539	/*
 540	 * The treatment of mapping from this point on is critical. The page
 541	 * lock protects many things but in this context the page lock
 542	 * stabilizes mapping, prevents inode freeing in the shared
 543	 * file-backed region case and guards against movement to swap cache.
 544	 *
 545	 * Strictly speaking the page lock is not needed in all cases being
 546	 * considered here and page lock forces unnecessarily serialization
 547	 * From this point on, mapping will be re-verified if necessary and
 548	 * page lock will be acquired only if it is unavoidable
 549	 *
 550	 * Mapping checks require the head page for any compound page so the
 551	 * head page and mapping is looked up now. For anonymous pages, it
 552	 * does not matter if the page splits in the future as the key is
 553	 * based on the address. For filesystem-backed pages, the tail is
 554	 * required as the index of the page determines the key. For
 555	 * base pages, there is no tail page and tail == page.
 556	 */
 557	tail = page;
 558	page = compound_head(page);
 559	mapping = READ_ONCE(page->mapping);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 560
 561	/*
 562	 * If page->mapping is NULL, then it cannot be a PageAnon
 563	 * page; but it might be the ZERO_PAGE or in the gate area or
 564	 * in a special mapping (all cases which we are happy to fail);
 565	 * or it may have been a good file page when get_user_pages_fast
 566	 * found it, but truncated or holepunched or subjected to
 567	 * invalidate_complete_page2 before we got the page lock (also
 568	 * cases which we are happy to fail).  And we hold a reference,
 569	 * so refcount care in invalidate_complete_page's remove_mapping
 570	 * prevents drop_caches from setting mapping to NULL beneath us.
 571	 *
 572	 * The case we do have to guard against is when memory pressure made
 573	 * shmem_writepage move it from filecache to swapcache beneath us:
 574	 * an unlikely race, but we do need to retry for page->mapping.
 575	 */
 576	if (unlikely(!mapping)) {
 577		int shmem_swizzled;
 578
 579		/*
 580		 * Page lock is required to identify which special case above
 581		 * applies. If this is really a shmem page then the page lock
 582		 * will prevent unexpected transitions.
 583		 */
 584		lock_page(page);
 585		shmem_swizzled = PageSwapCache(page) || page->mapping;
 586		unlock_page(page);
 587		put_page(page);
 588
 589		if (shmem_swizzled)
 590			goto again;
 591
 592		return -EFAULT;
 593	}
 594
 595	/*
 596	 * Private mappings are handled in a simple way.
 597	 *
 598	 * If the futex key is stored on an anonymous page, then the associated
 599	 * object is the mm which is implicitly pinned by the calling process.
 600	 *
 601	 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
 602	 * it's a read-only handle, it's expected that futexes attach to
 603	 * the object not the particular process.
 604	 */
 605	if (PageAnon(page)) {
 606		/*
 607		 * A RO anonymous page will never change and thus doesn't make
 608		 * sense for futex operations.
 609		 */
 610		if (unlikely(should_fail_futex(true)) || ro) {
 611			err = -EFAULT;
 612			goto out;
 613		}
 614
 615		key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
 616		key->private.mm = mm;
 617		key->private.address = address;
 618
 619	} else {
 620		struct inode *inode;
 621
 622		/*
 623		 * The associated futex object in this case is the inode and
 624		 * the page->mapping must be traversed. Ordinarily this should
 625		 * be stabilised under page lock but it's not strictly
 626		 * necessary in this case as we just want to pin the inode, not
 627		 * update the radix tree or anything like that.
 628		 *
 629		 * The RCU read lock is taken as the inode is finally freed
 630		 * under RCU. If the mapping still matches expectations then the
 631		 * mapping->host can be safely accessed as being a valid inode.
 632		 */
 633		rcu_read_lock();
 634
 635		if (READ_ONCE(page->mapping) != mapping) {
 636			rcu_read_unlock();
 637			put_page(page);
 638
 639			goto again;
 640		}
 641
 642		inode = READ_ONCE(mapping->host);
 643		if (!inode) {
 644			rcu_read_unlock();
 645			put_page(page);
 646
 647			goto again;
 648		}
 649
 650		key->both.offset |= FUT_OFF_INODE; /* inode-based key */
 651		key->shared.i_seq = get_inode_sequence_number(inode);
 652		key->shared.pgoff = page_to_pgoff(tail);
 653		rcu_read_unlock();
 654	}
 655
 
 
 656out:
 657	put_page(page);
 
 658	return err;
 659}
 660
 
 
 
 
 
 661/**
 662 * fault_in_user_writeable() - Fault in user address and verify RW access
 663 * @uaddr:	pointer to faulting user space address
 664 *
 665 * Slow path to fixup the fault we just took in the atomic write
 666 * access to @uaddr.
 667 *
 668 * We have no generic implementation of a non-destructive write to the
 669 * user address. We know that we faulted in the atomic pagefault
 670 * disabled section so we can as well avoid the #PF overhead by
 671 * calling get_user_pages() right away.
 672 */
 673static int fault_in_user_writeable(u32 __user *uaddr)
 674{
 675	struct mm_struct *mm = current->mm;
 676	int ret;
 677
 678	mmap_read_lock(mm);
 679	ret = fixup_user_fault(mm, (unsigned long)uaddr,
 680			       FAULT_FLAG_WRITE, NULL);
 681	mmap_read_unlock(mm);
 682
 683	return ret < 0 ? ret : 0;
 684}
 685
 686/**
 687 * futex_top_waiter() - Return the highest priority waiter on a futex
 688 * @hb:		the hash bucket the futex_q's reside in
 689 * @key:	the futex key (to distinguish it from other futex futex_q's)
 690 *
 691 * Must be called with the hb lock held.
 692 */
 693static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
 694					union futex_key *key)
 695{
 696	struct futex_q *this;
 697
 698	plist_for_each_entry(this, &hb->chain, list) {
 699		if (match_futex(&this->key, key))
 700			return this;
 701	}
 702	return NULL;
 703}
 704
 705static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
 706				      u32 uval, u32 newval)
 707{
 708	int ret;
 709
 710	pagefault_disable();
 711	ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
 712	pagefault_enable();
 713
 714	return ret;
 715}
 716
 717static int get_futex_value_locked(u32 *dest, u32 __user *from)
 718{
 719	int ret;
 720
 721	pagefault_disable();
 722	ret = __get_user(*dest, from);
 723	pagefault_enable();
 724
 725	return ret ? -EFAULT : 0;
 726}
 727
 728
 729/*
 730 * PI code:
 731 */
 732static int refill_pi_state_cache(void)
 733{
 734	struct futex_pi_state *pi_state;
 735
 736	if (likely(current->pi_state_cache))
 737		return 0;
 738
 739	pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
 740
 741	if (!pi_state)
 742		return -ENOMEM;
 743
 744	INIT_LIST_HEAD(&pi_state->list);
 745	/* pi_mutex gets initialized later */
 746	pi_state->owner = NULL;
 747	refcount_set(&pi_state->refcount, 1);
 748	pi_state->key = FUTEX_KEY_INIT;
 749
 750	current->pi_state_cache = pi_state;
 751
 752	return 0;
 753}
 754
 755static struct futex_pi_state *alloc_pi_state(void)
 756{
 757	struct futex_pi_state *pi_state = current->pi_state_cache;
 758
 759	WARN_ON(!pi_state);
 760	current->pi_state_cache = NULL;
 761
 762	return pi_state;
 763}
 764
 765static void pi_state_update_owner(struct futex_pi_state *pi_state,
 766				  struct task_struct *new_owner)
 767{
 768	struct task_struct *old_owner = pi_state->owner;
 769
 770	lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
 771
 772	if (old_owner) {
 773		raw_spin_lock(&old_owner->pi_lock);
 774		WARN_ON(list_empty(&pi_state->list));
 775		list_del_init(&pi_state->list);
 776		raw_spin_unlock(&old_owner->pi_lock);
 777	}
 778
 779	if (new_owner) {
 780		raw_spin_lock(&new_owner->pi_lock);
 781		WARN_ON(!list_empty(&pi_state->list));
 782		list_add(&pi_state->list, &new_owner->pi_state_list);
 783		pi_state->owner = new_owner;
 784		raw_spin_unlock(&new_owner->pi_lock);
 785	}
 786}
 787
 788static void get_pi_state(struct futex_pi_state *pi_state)
 789{
 790	WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
 791}
 792
 793/*
 794 * Drops a reference to the pi_state object and frees or caches it
 795 * when the last reference is gone.
 796 */
 797static void put_pi_state(struct futex_pi_state *pi_state)
 798{
 799	if (!pi_state)
 800		return;
 801
 802	if (!refcount_dec_and_test(&pi_state->refcount))
 803		return;
 804
 805	/*
 806	 * If pi_state->owner is NULL, the owner is most probably dying
 807	 * and has cleaned up the pi_state already
 808	 */
 809	if (pi_state->owner) {
 810		unsigned long flags;
 
 
 811
 812		raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
 813		pi_state_update_owner(pi_state, NULL);
 814		rt_mutex_proxy_unlock(&pi_state->pi_mutex);
 815		raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
 816	}
 817
 818	if (current->pi_state_cache) {
 819		kfree(pi_state);
 820	} else {
 821		/*
 822		 * pi_state->list is already empty.
 823		 * clear pi_state->owner.
 824		 * refcount is at 0 - put it back to 1.
 825		 */
 826		pi_state->owner = NULL;
 827		refcount_set(&pi_state->refcount, 1);
 828		current->pi_state_cache = pi_state;
 829	}
 830}
 831
 832#ifdef CONFIG_FUTEX_PI
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 833
 834/*
 835 * This task is holding PI mutexes at exit time => bad.
 836 * Kernel cleans up PI-state, but userspace is likely hosed.
 837 * (Robust-futex cleanup is separate and might save the day for userspace.)
 838 */
 839static void exit_pi_state_list(struct task_struct *curr)
 840{
 841	struct list_head *next, *head = &curr->pi_state_list;
 842	struct futex_pi_state *pi_state;
 843	struct futex_hash_bucket *hb;
 844	union futex_key key = FUTEX_KEY_INIT;
 845
 846	if (!futex_cmpxchg_enabled)
 847		return;
 848	/*
 849	 * We are a ZOMBIE and nobody can enqueue itself on
 850	 * pi_state_list anymore, but we have to be careful
 851	 * versus waiters unqueueing themselves:
 852	 */
 853	raw_spin_lock_irq(&curr->pi_lock);
 854	while (!list_empty(head)) {
 
 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
 860		/*
 861		 * We can race against put_pi_state() removing itself from the
 862		 * list (a waiter going away). put_pi_state() will first
 863		 * decrement the reference count and then modify the list, so
 864		 * its possible to see the list entry but fail this reference
 865		 * acquire.
 866		 *
 867		 * In that case; drop the locks to let put_pi_state() make
 868		 * progress and retry the loop.
 869		 */
 870		if (!refcount_inc_not_zero(&pi_state->refcount)) {
 871			raw_spin_unlock_irq(&curr->pi_lock);
 872			cpu_relax();
 873			raw_spin_lock_irq(&curr->pi_lock);
 874			continue;
 875		}
 876		raw_spin_unlock_irq(&curr->pi_lock);
 877
 878		spin_lock(&hb->lock);
 879		raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
 880		raw_spin_lock(&curr->pi_lock);
 881		/*
 882		 * We dropped the pi-lock, so re-check whether this
 883		 * task still owns the PI-state:
 884		 */
 885		if (head->next != next) {
 886			/* retain curr->pi_lock for the loop invariant */
 887			raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
 888			spin_unlock(&hb->lock);
 889			put_pi_state(pi_state);
 890			continue;
 891		}
 892
 893		WARN_ON(pi_state->owner != curr);
 894		WARN_ON(list_empty(&pi_state->list));
 895		list_del_init(&pi_state->list);
 896		pi_state->owner = NULL;
 
 
 
 897
 898		raw_spin_unlock(&curr->pi_lock);
 899		raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
 900		spin_unlock(&hb->lock);
 901
 902		rt_mutex_futex_unlock(&pi_state->pi_mutex);
 903		put_pi_state(pi_state);
 904
 905		raw_spin_lock_irq(&curr->pi_lock);
 906	}
 907	raw_spin_unlock_irq(&curr->pi_lock);
 908}
 909#else
 910static inline void exit_pi_state_list(struct task_struct *curr) { }
 911#endif
 912
 913/*
 914 * We need to check the following states:
 915 *
 916 *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
 917 *
 918 * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
 919 * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
 920 *
 921 * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
 922 *
 923 * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
 924 * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
 925 *
 926 * [6]  Found  | Found    | task      | 0         | 1      | Valid
 927 *
 928 * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
 929 *
 930 * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
 931 * [9]  Found  | Found    | task      | 0         | 0      | Invalid
 932 * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
 933 *
 934 * [1]	Indicates that the kernel can acquire the futex atomically. We
 935 *	came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
 936 *
 937 * [2]	Valid, if TID does not belong to a kernel thread. If no matching
 938 *      thread is found then it indicates that the owner TID has died.
 939 *
 940 * [3]	Invalid. The waiter is queued on a non PI futex
 941 *
 942 * [4]	Valid state after exit_robust_list(), which sets the user space
 943 *	value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
 944 *
 945 * [5]	The user space value got manipulated between exit_robust_list()
 946 *	and exit_pi_state_list()
 947 *
 948 * [6]	Valid state after exit_pi_state_list() which sets the new owner in
 949 *	the pi_state but cannot access the user space value.
 950 *
 951 * [7]	pi_state->owner can only be NULL when the OWNER_DIED bit is set.
 952 *
 953 * [8]	Owner and user space value match
 954 *
 955 * [9]	There is no transient state which sets the user space TID to 0
 956 *	except exit_robust_list(), but this is indicated by the
 957 *	FUTEX_OWNER_DIED bit. See [4]
 958 *
 959 * [10] There is no transient state which leaves owner and user space
 960 *	TID out of sync. Except one error case where the kernel is denied
 961 *	write access to the user address, see fixup_pi_state_owner().
 962 *
 963 *
 964 * Serialization and lifetime rules:
 965 *
 966 * hb->lock:
 967 *
 968 *	hb -> futex_q, relation
 969 *	futex_q -> pi_state, relation
 970 *
 971 *	(cannot be raw because hb can contain arbitrary amount
 972 *	 of futex_q's)
 973 *
 974 * pi_mutex->wait_lock:
 975 *
 976 *	{uval, pi_state}
 977 *
 978 *	(and pi_mutex 'obviously')
 979 *
 980 * p->pi_lock:
 981 *
 982 *	p->pi_state_list -> pi_state->list, relation
 983 *	pi_mutex->owner -> pi_state->owner, relation
 984 *
 985 * pi_state->refcount:
 986 *
 987 *	pi_state lifetime
 988 *
 989 *
 990 * Lock order:
 991 *
 992 *   hb->lock
 993 *     pi_mutex->wait_lock
 994 *       p->pi_lock
 995 *
 996 */
 997
 998/*
 999 * Validate that the existing waiter has a pi_state and sanity check
1000 * the pi_state against the user space value. If correct, attach to
1001 * it.
1002 */
1003static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1004			      struct futex_pi_state *pi_state,
1005			      struct futex_pi_state **ps)
1006{
 
 
 
1007	pid_t pid = uval & FUTEX_TID_MASK;
1008	u32 uval2;
1009	int ret;
1010
1011	/*
1012	 * Userspace might have messed up non-PI and PI futexes [3]
1013	 */
1014	if (unlikely(!pi_state))
1015		return -EINVAL;
1016
1017	/*
1018	 * We get here with hb->lock held, and having found a
1019	 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1020	 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1021	 * which in turn means that futex_lock_pi() still has a reference on
1022	 * our pi_state.
1023	 *
1024	 * The waiter holding a reference on @pi_state also protects against
1025	 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1026	 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1027	 * free pi_state before we can take a reference ourselves.
1028	 */
1029	WARN_ON(!refcount_read(&pi_state->refcount));
1030
1031	/*
1032	 * Now that we have a pi_state, we can acquire wait_lock
1033	 * and do the state validation.
1034	 */
1035	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1036
1037	/*
1038	 * Since {uval, pi_state} is serialized by wait_lock, and our current
1039	 * uval was read without holding it, it can have changed. Verify it
1040	 * still is what we expect it to be, otherwise retry the entire
1041	 * operation.
1042	 */
1043	if (get_futex_value_locked(&uval2, uaddr))
1044		goto out_efault;
1045
1046	if (uval != uval2)
1047		goto out_eagain;
1048
1049	/*
1050	 * Handle the owner died case:
1051	 */
1052	if (uval & FUTEX_OWNER_DIED) {
1053		/*
1054		 * exit_pi_state_list sets owner to NULL and wakes the
1055		 * topmost waiter. The task which acquires the
1056		 * pi_state->rt_mutex will fixup owner.
1057		 */
1058		if (!pi_state->owner) {
1059			/*
1060			 * No pi state owner, but the user space TID
1061			 * is not 0. Inconsistent state. [5]
1062			 */
1063			if (pid)
1064				goto out_einval;
1065			/*
1066			 * Take a ref on the state and return success. [4]
 
1067			 */
1068			goto out_attach;
1069		}
1070
1071		/*
1072		 * If TID is 0, then either the dying owner has not
1073		 * yet executed exit_pi_state_list() or some waiter
1074		 * acquired the rtmutex in the pi state, but did not
1075		 * yet fixup the TID in user space.
1076		 *
1077		 * Take a ref on the state and return success. [6]
1078		 */
1079		if (!pid)
1080			goto out_attach;
1081	} else {
1082		/*
1083		 * If the owner died bit is not set, then the pi_state
1084		 * must have an owner. [7]
1085		 */
1086		if (!pi_state->owner)
1087			goto out_einval;
1088	}
1089
1090	/*
1091	 * Bail out if user space manipulated the futex value. If pi
1092	 * state exists then the owner TID must be the same as the
1093	 * user space TID. [9/10]
1094	 */
1095	if (pid != task_pid_vnr(pi_state->owner))
1096		goto out_einval;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1097
1098out_attach:
1099	get_pi_state(pi_state);
1100	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1101	*ps = pi_state;
1102	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1103
1104out_einval:
1105	ret = -EINVAL;
1106	goto out_error;
1107
1108out_eagain:
1109	ret = -EAGAIN;
1110	goto out_error;
1111
1112out_efault:
1113	ret = -EFAULT;
1114	goto out_error;
1115
1116out_error:
1117	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1118	return ret;
1119}
1120
1121/**
1122 * wait_for_owner_exiting - Block until the owner has exited
1123 * @ret: owner's current futex lock status
1124 * @exiting:	Pointer to the exiting task
1125 *
1126 * Caller must hold a refcount on @exiting.
1127 */
1128static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1129{
1130	if (ret != -EBUSY) {
1131		WARN_ON_ONCE(exiting);
1132		return;
1133	}
1134
1135	if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1136		return;
1137
1138	mutex_lock(&exiting->futex_exit_mutex);
1139	/*
1140	 * No point in doing state checking here. If the waiter got here
1141	 * while the task was in exec()->exec_futex_release() then it can
1142	 * have any FUTEX_STATE_* value when the waiter has acquired the
1143	 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1144	 * already. Highly unlikely and not a problem. Just one more round
1145	 * through the futex maze.
1146	 */
1147	mutex_unlock(&exiting->futex_exit_mutex);
1148
1149	put_task_struct(exiting);
1150}
1151
1152static int handle_exit_race(u32 __user *uaddr, u32 uval,
1153			    struct task_struct *tsk)
1154{
1155	u32 uval2;
1156
1157	/*
1158	 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1159	 * caller that the alleged owner is busy.
1160	 */
1161	if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1162		return -EBUSY;
1163
1164	/*
1165	 * Reread the user space value to handle the following situation:
1166	 *
1167	 * CPU0				CPU1
1168	 *
1169	 * sys_exit()			sys_futex()
1170	 *  do_exit()			 futex_lock_pi()
1171	 *                                futex_lock_pi_atomic()
1172	 *   exit_signals(tsk)		    No waiters:
1173	 *    tsk->flags |= PF_EXITING;	    *uaddr == 0x00000PID
1174	 *  mm_release(tsk)		    Set waiter bit
1175	 *   exit_robust_list(tsk) {	    *uaddr = 0x80000PID;
1176	 *      Set owner died		    attach_to_pi_owner() {
1177	 *    *uaddr = 0xC0000000;	     tsk = get_task(PID);
1178	 *   }				     if (!tsk->flags & PF_EXITING) {
1179	 *  ...				       attach();
1180	 *  tsk->futex_state =               } else {
1181	 *	FUTEX_STATE_DEAD;              if (tsk->futex_state !=
1182	 *					  FUTEX_STATE_DEAD)
1183	 *				         return -EAGAIN;
1184	 *				       return -ESRCH; <--- FAIL
1185	 *				     }
1186	 *
1187	 * Returning ESRCH unconditionally is wrong here because the
1188	 * user space value has been changed by the exiting task.
1189	 *
1190	 * The same logic applies to the case where the exiting task is
1191	 * already gone.
1192	 */
1193	if (get_futex_value_locked(&uval2, uaddr))
1194		return -EFAULT;
1195
1196	/* If the user space value has changed, try again. */
1197	if (uval2 != uval)
1198		return -EAGAIN;
1199
1200	/*
1201	 * The exiting task did not have a robust list, the robust list was
1202	 * corrupted or the user space value in *uaddr is simply bogus.
1203	 * Give up and tell user space.
1204	 */
1205	return -ESRCH;
1206}
1207
1208/*
1209 * Lookup the task for the TID provided from user space and attach to
1210 * it after doing proper sanity checks.
1211 */
1212static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1213			      struct futex_pi_state **ps,
1214			      struct task_struct **exiting)
1215{
1216	pid_t pid = uval & FUTEX_TID_MASK;
1217	struct futex_pi_state *pi_state;
1218	struct task_struct *p;
1219
1220	/*
1221	 * We are the first waiter - try to look up the real owner and attach
1222	 * the new pi_state to it, but bail out when TID = 0 [1]
1223	 *
1224	 * The !pid check is paranoid. None of the call sites should end up
1225	 * with pid == 0, but better safe than sorry. Let the caller retry
1226	 */
1227	if (!pid)
1228		return -EAGAIN;
1229	p = find_get_task_by_vpid(pid);
1230	if (!p)
1231		return handle_exit_race(uaddr, uval, NULL);
1232
1233	if (unlikely(p->flags & PF_KTHREAD)) {
1234		put_task_struct(p);
1235		return -EPERM;
1236	}
1237
1238	/*
1239	 * We need to look at the task state to figure out, whether the
1240	 * task is exiting. To protect against the change of the task state
1241	 * in futex_exit_release(), we do this protected by p->pi_lock:
 
1242	 */
1243	raw_spin_lock_irq(&p->pi_lock);
1244	if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1245		/*
1246		 * The task is on the way out. When the futex state is
1247		 * FUTEX_STATE_DEAD, we know that the task has finished
1248		 * the cleanup:
1249		 */
1250		int ret = handle_exit_race(uaddr, uval, p);
1251
1252		raw_spin_unlock_irq(&p->pi_lock);
1253		/*
1254		 * If the owner task is between FUTEX_STATE_EXITING and
1255		 * FUTEX_STATE_DEAD then store the task pointer and keep
1256		 * the reference on the task struct. The calling code will
1257		 * drop all locks, wait for the task to reach
1258		 * FUTEX_STATE_DEAD and then drop the refcount. This is
1259		 * required to prevent a live lock when the current task
1260		 * preempted the exiting task between the two states.
1261		 */
1262		if (ret == -EBUSY)
1263			*exiting = p;
1264		else
1265			put_task_struct(p);
1266		return ret;
1267	}
1268
1269	/*
1270	 * No existing pi state. First waiter. [2]
1271	 *
1272	 * This creates pi_state, we have hb->lock held, this means nothing can
1273	 * observe this state, wait_lock is irrelevant.
1274	 */
1275	pi_state = alloc_pi_state();
1276
1277	/*
1278	 * Initialize the pi_mutex in locked state and make @p
1279	 * the owner of it:
1280	 */
1281	rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1282
1283	/* Store the key for possible exit cleanups: */
1284	pi_state->key = *key;
1285
1286	WARN_ON(!list_empty(&pi_state->list));
1287	list_add(&pi_state->list, &p->pi_state_list);
1288	/*
1289	 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1290	 * because there is no concurrency as the object is not published yet.
1291	 */
1292	pi_state->owner = p;
1293	raw_spin_unlock_irq(&p->pi_lock);
1294
1295	put_task_struct(p);
1296
1297	*ps = pi_state;
1298
1299	return 0;
1300}
1301
1302static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1303			   struct futex_hash_bucket *hb,
1304			   union futex_key *key, struct futex_pi_state **ps,
1305			   struct task_struct **exiting)
1306{
1307	struct futex_q *top_waiter = futex_top_waiter(hb, key);
1308
1309	/*
1310	 * If there is a waiter on that futex, validate it and
1311	 * attach to the pi_state when the validation succeeds.
1312	 */
1313	if (top_waiter)
1314		return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1315
1316	/*
1317	 * We are the first waiter - try to look up the owner based on
1318	 * @uval and attach to it.
1319	 */
1320	return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1321}
1322
1323static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1324{
1325	int err;
1326	u32 curval;
1327
1328	if (unlikely(should_fail_futex(true)))
1329		return -EFAULT;
1330
1331	err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1332	if (unlikely(err))
1333		return err;
1334
1335	/* If user space value changed, let the caller retry */
1336	return curval != uval ? -EAGAIN : 0;
1337}
1338
1339/**
1340 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1341 * @uaddr:		the pi futex user address
1342 * @hb:			the pi futex hash bucket
1343 * @key:		the futex key associated with uaddr and hb
1344 * @ps:			the pi_state pointer where we store the result of the
1345 *			lookup
1346 * @task:		the task to perform the atomic lock work for.  This will
1347 *			be "current" except in the case of requeue pi.
1348 * @exiting:		Pointer to store the task pointer of the owner task
1349 *			which is in the middle of exiting
1350 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1351 *
1352 * Return:
1353 *  -  0 - ready to wait;
1354 *  -  1 - acquired the lock;
1355 *  - <0 - error
1356 *
1357 * The hb->lock and futex_key refs shall be held by the caller.
1358 *
1359 * @exiting is only set when the return value is -EBUSY. If so, this holds
1360 * a refcount on the exiting task on return and the caller needs to drop it
1361 * after waiting for the exit to complete.
1362 */
1363static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1364				union futex_key *key,
1365				struct futex_pi_state **ps,
1366				struct task_struct *task,
1367				struct task_struct **exiting,
1368				int set_waiters)
1369{
1370	u32 uval, newval, vpid = task_pid_vnr(task);
1371	struct futex_q *top_waiter;
1372	int ret;
 
 
1373
1374	/*
1375	 * Read the user space value first so we can validate a few
1376	 * things before proceeding further.
 
1377	 */
1378	if (get_futex_value_locked(&uval, uaddr))
1379		return -EFAULT;
 
1380
1381	if (unlikely(should_fail_futex(true)))
1382		return -EFAULT;
1383
1384	/*
1385	 * Detect deadlocks.
1386	 */
1387	if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1388		return -EDEADLK;
1389
1390	if ((unlikely(should_fail_futex(true))))
1391		return -EDEADLK;
 
 
 
 
 
 
 
 
 
 
 
 
 
1392
1393	/*
1394	 * Lookup existing state first. If it exists, try to attach to
1395	 * its pi_state.
1396	 */
1397	top_waiter = futex_top_waiter(hb, key);
1398	if (top_waiter)
1399		return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1400
1401	/*
1402	 * No waiter and user TID is 0. We are here because the
1403	 * waiters or the owner died bit is set or called from
1404	 * requeue_cmp_pi or for whatever reason something took the
1405	 * syscall.
1406	 */
1407	if (!(uval & FUTEX_TID_MASK)) {
1408		/*
1409		 * We take over the futex. No other waiters and the user space
1410		 * TID is 0. We preserve the owner died bit.
1411		 */
1412		newval = uval & FUTEX_OWNER_DIED;
1413		newval |= vpid;
 
 
1414
1415		/* The futex requeue_pi code can enforce the waiters bit */
1416		if (set_waiters)
1417			newval |= FUTEX_WAITERS;
1418
1419		ret = lock_pi_update_atomic(uaddr, uval, newval);
1420		/* If the take over worked, return 1 */
1421		return ret < 0 ? ret : 1;
1422	}
1423
1424	/*
1425	 * First waiter. Set the waiters bit before attaching ourself to
1426	 * the owner. If owner tries to unlock, it will be forced into
1427	 * the kernel and blocked on hb->lock.
1428	 */
1429	newval = uval | FUTEX_WAITERS;
1430	ret = lock_pi_update_atomic(uaddr, uval, newval);
1431	if (ret)
1432		return ret;
1433	/*
1434	 * If the update of the user space value succeeded, we try to
1435	 * attach to the owner. If that fails, no harm done, we only
1436	 * set the FUTEX_WAITERS bit in the user space variable.
1437	 */
1438	return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1439}
1440
1441/**
1442 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1443 * @q:	The futex_q to unqueue
1444 *
1445 * The q->lock_ptr must not be NULL and must be held by the caller.
1446 */
1447static void __unqueue_futex(struct futex_q *q)
1448{
1449	struct futex_hash_bucket *hb;
1450
1451	if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
 
1452		return;
1453	lockdep_assert_held(q->lock_ptr);
1454
1455	hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1456	plist_del(&q->list, &hb->chain);
1457	hb_waiters_dec(hb);
1458}
1459
1460/*
1461 * The hash bucket lock must be held when this is called.
1462 * Afterwards, the futex_q must not be accessed. Callers
1463 * must ensure to later call wake_up_q() for the actual
1464 * wakeups to occur.
1465 */
1466static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1467{
1468	struct task_struct *p = q->task;
1469
1470	if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1471		return;
1472
 
 
 
 
 
 
 
1473	get_task_struct(p);
 
1474	__unqueue_futex(q);
1475	/*
1476	 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1477	 * is written, without taking any locks. This is possible in the event
1478	 * of a spurious wakeup, for example. A memory barrier is required here
1479	 * to prevent the following store to lock_ptr from getting ahead of the
1480	 * plist_del in __unqueue_futex().
1481	 */
1482	smp_store_release(&q->lock_ptr, NULL);
 
1483
1484	/*
1485	 * Queue the task for later wakeup for after we've released
1486	 * the hb->lock.
1487	 */
1488	wake_q_add_safe(wake_q, p);
1489}
1490
1491/*
1492 * Caller must hold a reference on @pi_state.
1493 */
1494static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1495{
1496	u32 curval, newval;
1497	struct rt_mutex_waiter *top_waiter;
1498	struct task_struct *new_owner;
1499	bool postunlock = false;
1500	DEFINE_WAKE_Q(wake_q);
1501	int ret = 0;
1502
1503	top_waiter = rt_mutex_top_waiter(&pi_state->pi_mutex);
1504	if (WARN_ON_ONCE(!top_waiter)) {
1505		/*
1506		 * As per the comment in futex_unlock_pi() this should not happen.
1507		 *
1508		 * When this happens, give up our locks and try again, giving
1509		 * the futex_lock_pi() instance time to complete, either by
1510		 * waiting on the rtmutex or removing itself from the futex
1511		 * queue.
1512		 */
1513		ret = -EAGAIN;
1514		goto out_unlock;
1515	}
1516
1517	new_owner = top_waiter->task;
 
 
 
 
 
 
1518
1519	/*
1520	 * We pass it to the next owner. The WAITERS bit is always kept
1521	 * enabled while there is PI state around. We cleanup the owner
1522	 * died bit, because we are the owner.
1523	 */
1524	newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1525
1526	if (unlikely(should_fail_futex(true))) {
1527		ret = -EFAULT;
1528		goto out_unlock;
 
 
 
 
1529	}
1530
1531	ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1532	if (!ret && (curval != uval)) {
1533		/*
1534		 * If a unconditional UNLOCK_PI operation (user space did not
1535		 * try the TID->0 transition) raced with a waiter setting the
1536		 * FUTEX_WAITERS flag between get_user() and locking the hash
1537		 * bucket lock, retry the operation.
1538		 */
1539		if ((FUTEX_TID_MASK & curval) == uval)
1540			ret = -EAGAIN;
1541		else
1542			ret = -EINVAL;
1543	}
1544
1545	if (!ret) {
1546		/*
1547		 * This is a point of no return; once we modified the uval
1548		 * there is no going back and subsequent operations must
1549		 * not fail.
1550		 */
1551		pi_state_update_owner(pi_state, new_owner);
1552		postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1553	}
1554
1555out_unlock:
1556	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
 
1557
1558	if (postunlock)
1559		rt_mutex_postunlock(&wake_q);
 
 
 
 
 
 
1560
1561	return ret;
1562}
1563
1564/*
1565 * Express the locking dependencies for lockdep:
1566 */
1567static inline void
1568double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1569{
1570	if (hb1 <= hb2) {
1571		spin_lock(&hb1->lock);
1572		if (hb1 < hb2)
1573			spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1574	} else { /* hb1 > hb2 */
1575		spin_lock(&hb2->lock);
1576		spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1577	}
1578}
1579
1580static inline void
1581double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1582{
1583	spin_unlock(&hb1->lock);
1584	if (hb1 != hb2)
1585		spin_unlock(&hb2->lock);
1586}
1587
1588/*
1589 * Wake up waiters matching bitset queued on this futex (uaddr).
1590 */
1591static int
1592futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1593{
1594	struct futex_hash_bucket *hb;
1595	struct futex_q *this, *next;
1596	union futex_key key = FUTEX_KEY_INIT;
1597	int ret;
1598	DEFINE_WAKE_Q(wake_q);
1599
1600	if (!bitset)
1601		return -EINVAL;
1602
1603	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1604	if (unlikely(ret != 0))
1605		return ret;
1606
1607	hb = hash_futex(&key);
1608
1609	/* Make sure we really have tasks to wakeup */
1610	if (!hb_waiters_pending(hb))
1611		return ret;
1612
1613	spin_lock(&hb->lock);
1614
1615	plist_for_each_entry_safe(this, next, &hb->chain, list) {
1616		if (match_futex (&this->key, &key)) {
1617			if (this->pi_state || this->rt_waiter) {
1618				ret = -EINVAL;
1619				break;
1620			}
1621
1622			/* Check if one of the bits is set in both bitsets */
1623			if (!(this->bitset & bitset))
1624				continue;
1625
1626			mark_wake_futex(&wake_q, this);
1627			if (++ret >= nr_wake)
1628				break;
1629		}
1630	}
1631
1632	spin_unlock(&hb->lock);
1633	wake_up_q(&wake_q);
 
 
1634	return ret;
1635}
1636
1637static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1638{
1639	unsigned int op =	  (encoded_op & 0x70000000) >> 28;
1640	unsigned int cmp =	  (encoded_op & 0x0f000000) >> 24;
1641	int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1642	int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1643	int oldval, ret;
1644
1645	if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1646		if (oparg < 0 || oparg > 31) {
1647			char comm[sizeof(current->comm)];
1648			/*
1649			 * kill this print and return -EINVAL when userspace
1650			 * is sane again
1651			 */
1652			pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1653					get_task_comm(comm, current), oparg);
1654			oparg &= 31;
1655		}
1656		oparg = 1 << oparg;
1657	}
1658
1659	pagefault_disable();
1660	ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1661	pagefault_enable();
1662	if (ret)
1663		return ret;
1664
1665	switch (cmp) {
1666	case FUTEX_OP_CMP_EQ:
1667		return oldval == cmparg;
1668	case FUTEX_OP_CMP_NE:
1669		return oldval != cmparg;
1670	case FUTEX_OP_CMP_LT:
1671		return oldval < cmparg;
1672	case FUTEX_OP_CMP_GE:
1673		return oldval >= cmparg;
1674	case FUTEX_OP_CMP_LE:
1675		return oldval <= cmparg;
1676	case FUTEX_OP_CMP_GT:
1677		return oldval > cmparg;
1678	default:
1679		return -ENOSYS;
1680	}
1681}
1682
1683/*
1684 * Wake up all waiters hashed on the physical page that is mapped
1685 * to this virtual address:
1686 */
1687static int
1688futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1689	      int nr_wake, int nr_wake2, int op)
1690{
1691	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1692	struct futex_hash_bucket *hb1, *hb2;
1693	struct futex_q *this, *next;
1694	int ret, op_ret;
1695	DEFINE_WAKE_Q(wake_q);
1696
1697retry:
1698	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1699	if (unlikely(ret != 0))
1700		return ret;
1701	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1702	if (unlikely(ret != 0))
1703		return ret;
1704
1705	hb1 = hash_futex(&key1);
1706	hb2 = hash_futex(&key2);
1707
1708retry_private:
1709	double_lock_hb(hb1, hb2);
1710	op_ret = futex_atomic_op_inuser(op, uaddr2);
1711	if (unlikely(op_ret < 0)) {
 
1712		double_unlock_hb(hb1, hb2);
1713
1714		if (!IS_ENABLED(CONFIG_MMU) ||
1715		    unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1716			/*
1717			 * we don't get EFAULT from MMU faults if we don't have
1718			 * an MMU, but we might get them from range checking
1719			 */
 
 
 
 
1720			ret = op_ret;
1721			return ret;
1722		}
1723
1724		if (op_ret == -EFAULT) {
1725			ret = fault_in_user_writeable(uaddr2);
1726			if (ret)
1727				return ret;
1728		}
1729
1730		cond_resched();
1731		if (!(flags & FLAGS_SHARED))
1732			goto retry_private;
 
 
 
1733		goto retry;
1734	}
1735
1736	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1737		if (match_futex (&this->key, &key1)) {
1738			if (this->pi_state || this->rt_waiter) {
1739				ret = -EINVAL;
1740				goto out_unlock;
1741			}
1742			mark_wake_futex(&wake_q, this);
1743			if (++ret >= nr_wake)
1744				break;
1745		}
1746	}
1747
1748	if (op_ret > 0) {
1749		op_ret = 0;
1750		plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1751			if (match_futex (&this->key, &key2)) {
1752				if (this->pi_state || this->rt_waiter) {
1753					ret = -EINVAL;
1754					goto out_unlock;
1755				}
1756				mark_wake_futex(&wake_q, this);
1757				if (++op_ret >= nr_wake2)
1758					break;
1759			}
1760		}
1761		ret += op_ret;
1762	}
1763
1764out_unlock:
1765	double_unlock_hb(hb1, hb2);
1766	wake_up_q(&wake_q);
 
 
 
 
1767	return ret;
1768}
1769
1770/**
1771 * requeue_futex() - Requeue a futex_q from one hb to another
1772 * @q:		the futex_q to requeue
1773 * @hb1:	the source hash_bucket
1774 * @hb2:	the target hash_bucket
1775 * @key2:	the new key for the requeued futex_q
1776 */
1777static inline
1778void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1779		   struct futex_hash_bucket *hb2, union futex_key *key2)
1780{
1781
1782	/*
1783	 * If key1 and key2 hash to the same bucket, no need to
1784	 * requeue.
1785	 */
1786	if (likely(&hb1->chain != &hb2->chain)) {
1787		plist_del(&q->list, &hb1->chain);
1788		hb_waiters_dec(hb1);
 
1789		hb_waiters_inc(hb2);
1790		plist_add(&q->list, &hb2->chain);
1791		q->lock_ptr = &hb2->lock;
1792	}
 
1793	q->key = *key2;
1794}
1795
1796/**
1797 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1798 * @q:		the futex_q
1799 * @key:	the key of the requeue target futex
1800 * @hb:		the hash_bucket of the requeue target futex
1801 *
1802 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1803 * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1804 * to the requeue target futex so the waiter can detect the wakeup on the right
1805 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1806 * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1807 * to protect access to the pi_state to fixup the owner later.  Must be called
1808 * with both q->lock_ptr and hb->lock held.
1809 */
1810static inline
1811void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1812			   struct futex_hash_bucket *hb)
1813{
 
1814	q->key = *key;
1815
1816	__unqueue_futex(q);
1817
1818	WARN_ON(!q->rt_waiter);
1819	q->rt_waiter = NULL;
1820
1821	q->lock_ptr = &hb->lock;
1822
1823	wake_up_state(q->task, TASK_NORMAL);
1824}
1825
1826/**
1827 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1828 * @pifutex:		the user address of the to futex
1829 * @hb1:		the from futex hash bucket, must be locked by the caller
1830 * @hb2:		the to futex hash bucket, must be locked by the caller
1831 * @key1:		the from futex key
1832 * @key2:		the to futex key
1833 * @ps:			address to store the pi_state pointer
1834 * @exiting:		Pointer to store the task pointer of the owner task
1835 *			which is in the middle of exiting
1836 * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1837 *
1838 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1839 * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1840 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1841 * hb1 and hb2 must be held by the caller.
1842 *
1843 * @exiting is only set when the return value is -EBUSY. If so, this holds
1844 * a refcount on the exiting task on return and the caller needs to drop it
1845 * after waiting for the exit to complete.
1846 *
1847 * Return:
1848 *  -  0 - failed to acquire the lock atomically;
1849 *  - >0 - acquired the lock, return value is vpid of the top_waiter
1850 *  - <0 - error
1851 */
1852static int
1853futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1854			   struct futex_hash_bucket *hb2, union futex_key *key1,
1855			   union futex_key *key2, struct futex_pi_state **ps,
1856			   struct task_struct **exiting, int set_waiters)
1857{
1858	struct futex_q *top_waiter = NULL;
1859	u32 curval;
1860	int ret, vpid;
1861
1862	if (get_futex_value_locked(&curval, pifutex))
1863		return -EFAULT;
1864
1865	if (unlikely(should_fail_futex(true)))
1866		return -EFAULT;
1867
1868	/*
1869	 * Find the top_waiter and determine if there are additional waiters.
1870	 * If the caller intends to requeue more than 1 waiter to pifutex,
1871	 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1872	 * as we have means to handle the possible fault.  If not, don't set
1873	 * the bit unnecessarily as it will force the subsequent unlock to enter
1874	 * the kernel.
1875	 */
1876	top_waiter = futex_top_waiter(hb1, key1);
1877
1878	/* There are no waiters, nothing for us to do. */
1879	if (!top_waiter)
1880		return 0;
1881
1882	/* Ensure we requeue to the expected futex. */
1883	if (!match_futex(top_waiter->requeue_pi_key, key2))
1884		return -EINVAL;
1885
1886	/*
1887	 * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1888	 * the contended case or if set_waiters is 1.  The pi_state is returned
1889	 * in ps in contended cases.
1890	 */
1891	vpid = task_pid_vnr(top_waiter->task);
1892	ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1893				   exiting, set_waiters);
1894	if (ret == 1) {
1895		requeue_pi_wake_futex(top_waiter, key2, hb2);
1896		return vpid;
1897	}
1898	return ret;
1899}
1900
1901/**
1902 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1903 * @uaddr1:	source futex user address
1904 * @flags:	futex flags (FLAGS_SHARED, etc.)
1905 * @uaddr2:	target futex user address
1906 * @nr_wake:	number of waiters to wake (must be 1 for requeue_pi)
1907 * @nr_requeue:	number of waiters to requeue (0-INT_MAX)
1908 * @cmpval:	@uaddr1 expected value (or %NULL)
1909 * @requeue_pi:	if we are attempting to requeue from a non-pi futex to a
1910 *		pi futex (pi to pi requeue is not supported)
1911 *
1912 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1913 * uaddr2 atomically on behalf of the top waiter.
1914 *
1915 * Return:
1916 *  - >=0 - on success, the number of tasks requeued or woken;
1917 *  -  <0 - on error
1918 */
1919static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1920			 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1921			 u32 *cmpval, int requeue_pi)
1922{
1923	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1924	int task_count = 0, ret;
1925	struct futex_pi_state *pi_state = NULL;
1926	struct futex_hash_bucket *hb1, *hb2;
1927	struct futex_q *this, *next;
1928	DEFINE_WAKE_Q(wake_q);
1929
1930	if (nr_wake < 0 || nr_requeue < 0)
1931		return -EINVAL;
1932
1933	/*
1934	 * When PI not supported: return -ENOSYS if requeue_pi is true,
1935	 * consequently the compiler knows requeue_pi is always false past
1936	 * this point which will optimize away all the conditional code
1937	 * further down.
1938	 */
1939	if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1940		return -ENOSYS;
1941
1942	if (requeue_pi) {
1943		/*
1944		 * Requeue PI only works on two distinct uaddrs. This
1945		 * check is only valid for private futexes. See below.
1946		 */
1947		if (uaddr1 == uaddr2)
1948			return -EINVAL;
1949
1950		/*
1951		 * requeue_pi requires a pi_state, try to allocate it now
1952		 * without any locks in case it fails.
1953		 */
1954		if (refill_pi_state_cache())
1955			return -ENOMEM;
1956		/*
1957		 * requeue_pi must wake as many tasks as it can, up to nr_wake
1958		 * + nr_requeue, since it acquires the rt_mutex prior to
1959		 * returning to userspace, so as to not leave the rt_mutex with
1960		 * waiters and no owner.  However, second and third wake-ups
1961		 * cannot be predicted as they involve race conditions with the
1962		 * first wake and a fault while looking up the pi_state.  Both
1963		 * pthread_cond_signal() and pthread_cond_broadcast() should
1964		 * use nr_wake=1.
1965		 */
1966		if (nr_wake != 1)
1967			return -EINVAL;
1968	}
1969
1970retry:
1971	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
 
 
 
 
 
 
 
 
 
1972	if (unlikely(ret != 0))
1973		return ret;
1974	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1975			    requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1976	if (unlikely(ret != 0))
1977		return ret;
1978
1979	/*
1980	 * The check above which compares uaddrs is not sufficient for
1981	 * shared futexes. We need to compare the keys:
1982	 */
1983	if (requeue_pi && match_futex(&key1, &key2))
1984		return -EINVAL;
 
 
1985
1986	hb1 = hash_futex(&key1);
1987	hb2 = hash_futex(&key2);
1988
1989retry_private:
1990	hb_waiters_inc(hb2);
1991	double_lock_hb(hb1, hb2);
1992
1993	if (likely(cmpval != NULL)) {
1994		u32 curval;
1995
1996		ret = get_futex_value_locked(&curval, uaddr1);
1997
1998		if (unlikely(ret)) {
1999			double_unlock_hb(hb1, hb2);
2000			hb_waiters_dec(hb2);
2001
2002			ret = get_user(curval, uaddr1);
2003			if (ret)
2004				return ret;
2005
2006			if (!(flags & FLAGS_SHARED))
2007				goto retry_private;
2008
 
 
2009			goto retry;
2010		}
2011		if (curval != *cmpval) {
2012			ret = -EAGAIN;
2013			goto out_unlock;
2014		}
2015	}
2016
2017	if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2018		struct task_struct *exiting = NULL;
2019
2020		/*
2021		 * Attempt to acquire uaddr2 and wake the top waiter. If we
2022		 * intend to requeue waiters, force setting the FUTEX_WAITERS
2023		 * bit.  We force this here where we are able to easily handle
2024		 * faults rather in the requeue loop below.
2025		 */
2026		ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2027						 &key2, &pi_state,
2028						 &exiting, nr_requeue);
2029
2030		/*
2031		 * At this point the top_waiter has either taken uaddr2 or is
2032		 * waiting on it.  If the former, then the pi_state will not
2033		 * exist yet, look it up one more time to ensure we have a
2034		 * reference to it. If the lock was taken, ret contains the
2035		 * vpid of the top waiter task.
2036		 * If the lock was not taken, we have pi_state and an initial
2037		 * refcount on it. In case of an error we have nothing.
2038		 */
2039		if (ret > 0) {
2040			WARN_ON(pi_state);
 
2041			task_count++;
2042			/*
2043			 * If we acquired the lock, then the user space value
2044			 * of uaddr2 should be vpid. It cannot be changed by
2045			 * the top waiter as it is blocked on hb2 lock if it
2046			 * tries to do so. If something fiddled with it behind
2047			 * our back the pi state lookup might unearth it. So
2048			 * we rather use the known value than rereading and
2049			 * handing potential crap to lookup_pi_state.
2050			 *
2051			 * If that call succeeds then we have pi_state and an
2052			 * initial refcount on it.
2053			 */
2054			ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2055					      &pi_state, &exiting);
2056		}
2057
2058		switch (ret) {
2059		case 0:
2060			/* We hold a reference on the pi state. */
2061			break;
2062
2063			/* If the above failed, then pi_state is NULL */
2064		case -EFAULT:
2065			double_unlock_hb(hb1, hb2);
2066			hb_waiters_dec(hb2);
 
 
2067			ret = fault_in_user_writeable(uaddr2);
2068			if (!ret)
2069				goto retry;
2070			return ret;
2071		case -EBUSY:
2072		case -EAGAIN:
2073			/*
2074			 * Two reasons for this:
2075			 * - EBUSY: Owner is exiting and we just wait for the
2076			 *   exit to complete.
2077			 * - EAGAIN: The user space value changed.
2078			 */
2079			double_unlock_hb(hb1, hb2);
2080			hb_waiters_dec(hb2);
2081			/*
2082			 * Handle the case where the owner is in the middle of
2083			 * exiting. Wait for the exit to complete otherwise
2084			 * this task might loop forever, aka. live lock.
2085			 */
2086			wait_for_owner_exiting(ret, exiting);
2087			cond_resched();
2088			goto retry;
2089		default:
2090			goto out_unlock;
2091		}
2092	}
2093
2094	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2095		if (task_count - nr_wake >= nr_requeue)
2096			break;
2097
2098		if (!match_futex(&this->key, &key1))
2099			continue;
2100
2101		/*
2102		 * FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2103		 * be paired with each other and no other futex ops.
2104		 *
2105		 * We should never be requeueing a futex_q with a pi_state,
2106		 * which is awaiting a futex_unlock_pi().
2107		 */
2108		if ((requeue_pi && !this->rt_waiter) ||
2109		    (!requeue_pi && this->rt_waiter) ||
2110		    this->pi_state) {
2111			ret = -EINVAL;
2112			break;
2113		}
2114
2115		/*
2116		 * Wake nr_wake waiters.  For requeue_pi, if we acquired the
2117		 * lock, we already woke the top_waiter.  If not, it will be
2118		 * woken by futex_unlock_pi().
2119		 */
2120		if (++task_count <= nr_wake && !requeue_pi) {
2121			mark_wake_futex(&wake_q, this);
2122			continue;
2123		}
2124
2125		/* Ensure we requeue to the expected futex for requeue_pi. */
2126		if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2127			ret = -EINVAL;
2128			break;
2129		}
2130
2131		/*
2132		 * Requeue nr_requeue waiters and possibly one more in the case
2133		 * of requeue_pi if we couldn't acquire the lock atomically.
2134		 */
2135		if (requeue_pi) {
2136			/*
2137			 * Prepare the waiter to take the rt_mutex. Take a
2138			 * refcount on the pi_state and store the pointer in
2139			 * the futex_q object of the waiter.
2140			 */
2141			get_pi_state(pi_state);
2142			this->pi_state = pi_state;
2143			ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2144							this->rt_waiter,
2145							this->task);
2146			if (ret == 1) {
2147				/*
2148				 * We got the lock. We do neither drop the
2149				 * refcount on pi_state nor clear
2150				 * this->pi_state because the waiter needs the
2151				 * pi_state for cleaning up the user space
2152				 * value. It will drop the refcount after
2153				 * doing so.
2154				 */
2155				requeue_pi_wake_futex(this, &key2, hb2);
 
2156				continue;
2157			} else if (ret) {
2158				/*
2159				 * rt_mutex_start_proxy_lock() detected a
2160				 * potential deadlock when we tried to queue
2161				 * that waiter. Drop the pi_state reference
2162				 * which we took above and remove the pointer
2163				 * to the state from the waiters futex_q
2164				 * object.
2165				 */
2166				this->pi_state = NULL;
2167				put_pi_state(pi_state);
2168				/*
2169				 * We stop queueing more waiters and let user
2170				 * space deal with the mess.
2171				 */
2172				break;
2173			}
2174		}
2175		requeue_futex(this, hb1, hb2, &key2);
 
2176	}
2177
2178	/*
2179	 * We took an extra initial reference to the pi_state either
2180	 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2181	 * need to drop it here again.
2182	 */
2183	put_pi_state(pi_state);
2184
2185out_unlock:
2186	double_unlock_hb(hb1, hb2);
2187	wake_up_q(&wake_q);
2188	hb_waiters_dec(hb2);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2189	return ret ? ret : task_count;
2190}
2191
2192/* The key must be already stored in q->key. */
2193static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2194	__acquires(&hb->lock)
2195{
2196	struct futex_hash_bucket *hb;
2197
2198	hb = hash_futex(&q->key);
2199
2200	/*
2201	 * Increment the counter before taking the lock so that
2202	 * a potential waker won't miss a to-be-slept task that is
2203	 * waiting for the spinlock. This is safe as all queue_lock()
2204	 * users end up calling queue_me(). Similarly, for housekeeping,
2205	 * decrement the counter at queue_unlock() when some error has
2206	 * occurred and we don't end up adding the task to the list.
2207	 */
2208	hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2209
2210	q->lock_ptr = &hb->lock;
2211
2212	spin_lock(&hb->lock);
2213	return hb;
2214}
2215
2216static inline void
2217queue_unlock(struct futex_hash_bucket *hb)
2218	__releases(&hb->lock)
2219{
2220	spin_unlock(&hb->lock);
2221	hb_waiters_dec(hb);
2222}
2223
2224static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
 
 
 
 
 
 
 
 
 
 
 
 
 
2225{
2226	int prio;
2227
2228	/*
2229	 * The priority used to register this element is
2230	 * - either the real thread-priority for the real-time threads
2231	 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2232	 * - or MAX_RT_PRIO for non-RT threads.
2233	 * Thus, all RT-threads are woken first in priority order, and
2234	 * the others are woken last, in FIFO order.
2235	 */
2236	prio = min(current->normal_prio, MAX_RT_PRIO);
2237
2238	plist_node_init(&q->list, prio);
2239	plist_add(&q->list, &hb->chain);
2240	q->task = current;
2241}
2242
2243/**
2244 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2245 * @q:	The futex_q to enqueue
2246 * @hb:	The destination hash bucket
2247 *
2248 * The hb->lock must be held by the caller, and is released here. A call to
2249 * queue_me() is typically paired with exactly one call to unqueue_me().  The
2250 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2251 * or nothing if the unqueue is done as part of the wake process and the unqueue
2252 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2253 * an example).
2254 */
2255static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2256	__releases(&hb->lock)
2257{
2258	__queue_me(q, hb);
2259	spin_unlock(&hb->lock);
2260}
2261
2262/**
2263 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2264 * @q:	The futex_q to unqueue
2265 *
2266 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2267 * be paired with exactly one earlier call to queue_me().
2268 *
2269 * Return:
2270 *  - 1 - if the futex_q was still queued (and we removed unqueued it);
2271 *  - 0 - if the futex_q was already removed by the waking thread
2272 */
2273static int unqueue_me(struct futex_q *q)
2274{
2275	spinlock_t *lock_ptr;
2276	int ret = 0;
2277
2278	/* In the common case we don't take the spinlock, which is nice. */
2279retry:
2280	/*
2281	 * q->lock_ptr can change between this read and the following spin_lock.
2282	 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2283	 * optimizing lock_ptr out of the logic below.
2284	 */
2285	lock_ptr = READ_ONCE(q->lock_ptr);
2286	if (lock_ptr != NULL) {
2287		spin_lock(lock_ptr);
2288		/*
2289		 * q->lock_ptr can change between reading it and
2290		 * spin_lock(), causing us to take the wrong lock.  This
2291		 * corrects the race condition.
2292		 *
2293		 * Reasoning goes like this: if we have the wrong lock,
2294		 * q->lock_ptr must have changed (maybe several times)
2295		 * between reading it and the spin_lock().  It can
2296		 * change again after the spin_lock() but only if it was
2297		 * already changed before the spin_lock().  It cannot,
2298		 * however, change back to the original value.  Therefore
2299		 * we can detect whether we acquired the correct lock.
2300		 */
2301		if (unlikely(lock_ptr != q->lock_ptr)) {
2302			spin_unlock(lock_ptr);
2303			goto retry;
2304		}
2305		__unqueue_futex(q);
2306
2307		BUG_ON(q->pi_state);
2308
2309		spin_unlock(lock_ptr);
2310		ret = 1;
2311	}
2312
 
2313	return ret;
2314}
2315
2316/*
2317 * PI futexes can not be requeued and must remove themselves from the
2318 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held.
 
2319 */
2320static void unqueue_me_pi(struct futex_q *q)
 
2321{
2322	__unqueue_futex(q);
2323
2324	BUG_ON(!q->pi_state);
2325	put_pi_state(q->pi_state);
2326	q->pi_state = NULL;
 
 
2327}
2328
2329static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2330				  struct task_struct *argowner)
 
 
 
 
 
 
2331{
 
2332	struct futex_pi_state *pi_state = q->pi_state;
2333	struct task_struct *oldowner, *newowner;
2334	u32 uval, curval, newval, newtid;
2335	int err = 0;
2336
2337	oldowner = pi_state->owner;
 
 
2338
2339	/*
2340	 * We are here because either:
2341	 *
2342	 *  - we stole the lock and pi_state->owner needs updating to reflect
2343	 *    that (@argowner == current),
2344	 *
2345	 * or:
2346	 *
2347	 *  - someone stole our lock and we need to fix things to point to the
2348	 *    new owner (@argowner == NULL).
2349	 *
2350	 * Either way, we have to replace the TID in the user space variable.
2351	 * This must be atomic as we have to preserve the owner died bit here.
2352	 *
2353	 * Note: We write the user space value _before_ changing the pi_state
2354	 * because we can fault here. Imagine swapped out pages or a fork
2355	 * that marked all the anonymous memory readonly for cow.
2356	 *
2357	 * Modifying pi_state _before_ the user space value would leave the
2358	 * pi_state in an inconsistent state when we fault here, because we
2359	 * need to drop the locks to handle the fault. This might be observed
2360	 * in the PID check in lookup_pi_state.
 
2361	 */
2362retry:
2363	if (!argowner) {
2364		if (oldowner != current) {
2365			/*
2366			 * We raced against a concurrent self; things are
2367			 * already fixed up. Nothing to do.
2368			 */
2369			return 0;
2370		}
2371
2372		if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2373			/* We got the lock. pi_state is correct. Tell caller. */
2374			return 1;
2375		}
2376
2377		/*
2378		 * The trylock just failed, so either there is an owner or
2379		 * there is a higher priority waiter than this one.
2380		 */
2381		newowner = rt_mutex_owner(&pi_state->pi_mutex);
2382		/*
2383		 * If the higher priority waiter has not yet taken over the
2384		 * rtmutex then newowner is NULL. We can't return here with
2385		 * that state because it's inconsistent vs. the user space
2386		 * state. So drop the locks and try again. It's a valid
2387		 * situation and not any different from the other retry
2388		 * conditions.
2389		 */
2390		if (unlikely(!newowner)) {
2391			err = -EAGAIN;
2392			goto handle_err;
2393		}
2394	} else {
2395		WARN_ON_ONCE(argowner != current);
2396		if (oldowner == current) {
2397			/*
2398			 * We raced against a concurrent self; things are
2399			 * already fixed up. Nothing to do.
2400			 */
2401			return 1;
2402		}
2403		newowner = argowner;
2404	}
2405
2406	newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2407	/* Owner died? */
2408	if (!pi_state->owner)
2409		newtid |= FUTEX_OWNER_DIED;
2410
2411	err = get_futex_value_locked(&uval, uaddr);
2412	if (err)
2413		goto handle_err;
2414
2415	for (;;) {
2416		newval = (uval & FUTEX_OWNER_DIED) | newtid;
2417
2418		err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2419		if (err)
2420			goto handle_err;
2421
2422		if (curval == uval)
2423			break;
2424		uval = curval;
2425	}
2426
2427	/*
2428	 * We fixed up user space. Now we need to fix the pi_state
2429	 * itself.
2430	 */
2431	pi_state_update_owner(pi_state, newowner);
 
 
 
 
 
2432
2433	return argowner == current;
 
 
 
 
 
 
2434
2435	/*
2436	 * In order to reschedule or handle a page fault, we need to drop the
2437	 * locks here. In the case of a fault, this gives the other task
2438	 * (either the highest priority waiter itself or the task which stole
2439	 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2440	 * are back from handling the fault we need to check the pi_state after
2441	 * reacquiring the locks and before trying to do another fixup. When
2442	 * the fixup has been done already we simply return.
2443	 *
2444	 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2445	 * drop hb->lock since the caller owns the hb -> futex_q relation.
2446	 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2447	 */
2448handle_err:
2449	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2450	spin_unlock(q->lock_ptr);
2451
2452	switch (err) {
2453	case -EFAULT:
2454		err = fault_in_user_writeable(uaddr);
2455		break;
2456
2457	case -EAGAIN:
2458		cond_resched();
2459		err = 0;
2460		break;
2461
2462	default:
2463		WARN_ON_ONCE(1);
2464		break;
2465	}
2466
2467	spin_lock(q->lock_ptr);
2468	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2469
2470	/*
2471	 * Check if someone else fixed it for us:
2472	 */
2473	if (pi_state->owner != oldowner)
2474		return argowner == current;
2475
2476	/* Retry if err was -EAGAIN or the fault in succeeded */
2477	if (!err)
2478		goto retry;
2479
2480	/*
2481	 * fault_in_user_writeable() failed so user state is immutable. At
2482	 * best we can make the kernel state consistent but user state will
2483	 * be most likely hosed and any subsequent unlock operation will be
2484	 * rejected due to PI futex rule [10].
2485	 *
2486	 * Ensure that the rtmutex owner is also the pi_state owner despite
2487	 * the user space value claiming something different. There is no
2488	 * point in unlocking the rtmutex if current is the owner as it
2489	 * would need to wait until the next waiter has taken the rtmutex
2490	 * to guarantee consistent state. Keep it simple. Userspace asked
2491	 * for this wreckaged state.
2492	 *
2493	 * The rtmutex has an owner - either current or some other
2494	 * task. See the EAGAIN loop above.
2495	 */
2496	pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2497
2498	return err;
2499}
2500
2501static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2502				struct task_struct *argowner)
2503{
2504	struct futex_pi_state *pi_state = q->pi_state;
2505	int ret;
2506
2507	lockdep_assert_held(q->lock_ptr);
2508
2509	raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2510	ret = __fixup_pi_state_owner(uaddr, q, argowner);
2511	raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2512	return ret;
2513}
2514
2515static long futex_wait_restart(struct restart_block *restart);
2516
2517/**
2518 * fixup_owner() - Post lock pi_state and corner case management
2519 * @uaddr:	user address of the futex
2520 * @q:		futex_q (contains pi_state and access to the rt_mutex)
2521 * @locked:	if the attempt to take the rt_mutex succeeded (1) or not (0)
2522 *
2523 * After attempting to lock an rt_mutex, this function is called to cleanup
2524 * the pi_state owner as well as handle race conditions that may allow us to
2525 * acquire the lock. Must be called with the hb lock held.
2526 *
2527 * Return:
2528 *  -  1 - success, lock taken;
2529 *  -  0 - success, lock not taken;
2530 *  - <0 - on error (-EFAULT)
2531 */
2532static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2533{
 
 
 
2534	if (locked) {
2535		/*
2536		 * Got the lock. We might not be the anticipated owner if we
2537		 * did a lock-steal - fix up the PI-state in that case:
2538		 *
2539		 * Speculative pi_state->owner read (we don't hold wait_lock);
2540		 * since we own the lock pi_state->owner == current is the
2541		 * stable state, anything else needs more attention.
2542		 */
2543		if (q->pi_state->owner != current)
2544			return fixup_pi_state_owner(uaddr, q, current);
2545		return 1;
2546	}
2547
2548	/*
2549	 * If we didn't get the lock; check if anybody stole it from us. In
2550	 * that case, we need to fix up the uval to point to them instead of
2551	 * us, otherwise bad things happen. [10]
2552	 *
2553	 * Another speculative read; pi_state->owner == current is unstable
2554	 * but needs our attention.
2555	 */
2556	if (q->pi_state->owner == current)
2557		return fixup_pi_state_owner(uaddr, q, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2558
2559	/*
2560	 * Paranoia check. If we did not take the lock, then we should not be
2561	 * the owner of the rt_mutex. Warn and establish consistent state.
2562	 */
2563	if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2564		return fixup_pi_state_owner(uaddr, q, current);
 
 
 
2565
2566	return 0;
 
2567}
2568
2569/**
2570 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2571 * @hb:		the futex hash bucket, must be locked by the caller
2572 * @q:		the futex_q to queue up on
2573 * @timeout:	the prepared hrtimer_sleeper, or null for no timeout
2574 */
2575static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2576				struct hrtimer_sleeper *timeout)
2577{
2578	/*
2579	 * The task state is guaranteed to be set before another task can
2580	 * wake it. set_current_state() is implemented using smp_store_mb() and
2581	 * queue_me() calls spin_unlock() upon completion, both serializing
2582	 * access to the hash list and forcing another memory barrier.
2583	 */
2584	set_current_state(TASK_INTERRUPTIBLE);
2585	queue_me(q, hb);
2586
2587	/* Arm the timer */
2588	if (timeout)
2589		hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
 
 
 
2590
2591	/*
2592	 * If we have been removed from the hash list, then another task
2593	 * has tried to wake us, and we can skip the call to schedule().
2594	 */
2595	if (likely(!plist_node_empty(&q->list))) {
2596		/*
2597		 * If the timer has already expired, current will already be
2598		 * flagged for rescheduling. Only call schedule if there
2599		 * is no timeout, or if it has yet to expire.
2600		 */
2601		if (!timeout || timeout->task)
2602			freezable_schedule();
2603	}
2604	__set_current_state(TASK_RUNNING);
2605}
2606
2607/**
2608 * futex_wait_setup() - Prepare to wait on a futex
2609 * @uaddr:	the futex userspace address
2610 * @val:	the expected value
2611 * @flags:	futex flags (FLAGS_SHARED, etc.)
2612 * @q:		the associated futex_q
2613 * @hb:		storage for hash_bucket pointer to be returned to caller
2614 *
2615 * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2616 * compare it with the expected value.  Handle atomic faults internally.
2617 * Return with the hb lock held and a q.key reference on success, and unlocked
2618 * with no q.key reference on failure.
2619 *
2620 * Return:
2621 *  -  0 - uaddr contains val and hb has been locked;
2622 *  - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2623 */
2624static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2625			   struct futex_q *q, struct futex_hash_bucket **hb)
2626{
2627	u32 uval;
2628	int ret;
2629
2630	/*
2631	 * Access the page AFTER the hash-bucket is locked.
2632	 * Order is important:
2633	 *
2634	 *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2635	 *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2636	 *
2637	 * The basic logical guarantee of a futex is that it blocks ONLY
2638	 * if cond(var) is known to be true at the time of blocking, for
2639	 * any cond.  If we locked the hash-bucket after testing *uaddr, that
2640	 * would open a race condition where we could block indefinitely with
2641	 * cond(var) false, which would violate the guarantee.
2642	 *
2643	 * On the other hand, we insert q and release the hash-bucket only
2644	 * after testing *uaddr.  This guarantees that futex_wait() will NOT
2645	 * absorb a wakeup if *uaddr does not match the desired values
2646	 * while the syscall executes.
2647	 */
2648retry:
2649	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2650	if (unlikely(ret != 0))
2651		return ret;
2652
2653retry_private:
2654	*hb = queue_lock(q);
2655
2656	ret = get_futex_value_locked(&uval, uaddr);
2657
2658	if (ret) {
2659		queue_unlock(*hb);
2660
2661		ret = get_user(uval, uaddr);
2662		if (ret)
2663			return ret;
2664
2665		if (!(flags & FLAGS_SHARED))
2666			goto retry_private;
2667
 
2668		goto retry;
2669	}
2670
2671	if (uval != val) {
2672		queue_unlock(*hb);
2673		ret = -EWOULDBLOCK;
2674	}
2675
 
 
 
2676	return ret;
2677}
2678
2679static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2680		      ktime_t *abs_time, u32 bitset)
2681{
2682	struct hrtimer_sleeper timeout, *to;
2683	struct restart_block *restart;
2684	struct futex_hash_bucket *hb;
2685	struct futex_q q = futex_q_init;
2686	int ret;
2687
2688	if (!bitset)
2689		return -EINVAL;
2690	q.bitset = bitset;
2691
2692	to = futex_setup_timer(abs_time, &timeout, flags,
2693			       current->timer_slack_ns);
 
 
 
 
 
 
 
 
 
2694retry:
2695	/*
2696	 * Prepare to wait on uaddr. On success, holds hb lock and increments
2697	 * q.key refs.
2698	 */
2699	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2700	if (ret)
2701		goto out;
2702
2703	/* queue_me and wait for wakeup, timeout, or a signal. */
2704	futex_wait_queue_me(hb, &q, to);
2705
2706	/* If we were woken (and unqueued), we succeeded, whatever. */
2707	ret = 0;
2708	/* unqueue_me() drops q.key ref */
2709	if (!unqueue_me(&q))
2710		goto out;
2711	ret = -ETIMEDOUT;
2712	if (to && !to->task)
2713		goto out;
2714
2715	/*
2716	 * We expect signal_pending(current), but we might be the
2717	 * victim of a spurious wakeup as well.
2718	 */
2719	if (!signal_pending(current))
2720		goto retry;
2721
2722	ret = -ERESTARTSYS;
2723	if (!abs_time)
2724		goto out;
2725
2726	restart = &current->restart_block;
 
2727	restart->futex.uaddr = uaddr;
2728	restart->futex.val = val;
2729	restart->futex.time = *abs_time;
2730	restart->futex.bitset = bitset;
2731	restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2732
2733	ret = set_restart_fn(restart, futex_wait_restart);
2734
2735out:
2736	if (to) {
2737		hrtimer_cancel(&to->timer);
2738		destroy_hrtimer_on_stack(&to->timer);
2739	}
2740	return ret;
2741}
2742
2743
2744static long futex_wait_restart(struct restart_block *restart)
2745{
2746	u32 __user *uaddr = restart->futex.uaddr;
2747	ktime_t t, *tp = NULL;
2748
2749	if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2750		t = restart->futex.time;
2751		tp = &t;
2752	}
2753	restart->fn = do_no_restart_syscall;
2754
2755	return (long)futex_wait(uaddr, restart->futex.flags,
2756				restart->futex.val, tp, restart->futex.bitset);
2757}
2758
2759
2760/*
2761 * Userspace tried a 0 -> TID atomic transition of the futex value
2762 * and failed. The kernel side here does the whole locking operation:
2763 * if there are waiters then it will block as a consequence of relying
2764 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2765 * a 0 value of the futex too.).
2766 *
2767 * Also serves as futex trylock_pi()'ing, and due semantics.
2768 */
2769static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2770			 ktime_t *time, int trylock)
2771{
2772	struct hrtimer_sleeper timeout, *to;
2773	struct task_struct *exiting = NULL;
2774	struct rt_mutex_waiter rt_waiter;
2775	struct futex_hash_bucket *hb;
2776	struct futex_q q = futex_q_init;
2777	int res, ret;
2778
2779	if (!IS_ENABLED(CONFIG_FUTEX_PI))
2780		return -ENOSYS;
2781
2782	if (refill_pi_state_cache())
2783		return -ENOMEM;
2784
2785	to = futex_setup_timer(time, &timeout, flags, 0);
 
 
 
 
 
 
2786
2787retry:
2788	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2789	if (unlikely(ret != 0))
2790		goto out;
2791
2792retry_private:
2793	hb = queue_lock(&q);
2794
2795	ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2796				   &exiting, 0);
2797	if (unlikely(ret)) {
2798		/*
2799		 * Atomic work succeeded and we got the lock,
2800		 * or failed. Either way, we do _not_ block.
2801		 */
2802		switch (ret) {
2803		case 1:
2804			/* We got the lock. */
2805			ret = 0;
2806			goto out_unlock_put_key;
2807		case -EFAULT:
2808			goto uaddr_faulted;
2809		case -EBUSY:
2810		case -EAGAIN:
2811			/*
2812			 * Two reasons for this:
2813			 * - EBUSY: Task is exiting and we just wait for the
2814			 *   exit to complete.
2815			 * - EAGAIN: The user space value changed.
2816			 */
2817			queue_unlock(hb);
2818			/*
2819			 * Handle the case where the owner is in the middle of
2820			 * exiting. Wait for the exit to complete otherwise
2821			 * this task might loop forever, aka. live lock.
2822			 */
2823			wait_for_owner_exiting(ret, exiting);
2824			cond_resched();
2825			goto retry;
2826		default:
2827			goto out_unlock_put_key;
2828		}
2829	}
2830
2831	WARN_ON(!q.pi_state);
2832
2833	/*
2834	 * Only actually queue now that the atomic ops are done:
2835	 */
2836	__queue_me(&q, hb);
2837
2838	if (trylock) {
2839		ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
 
 
 
 
 
 
2840		/* Fixup the trylock return value: */
2841		ret = ret ? 0 : -EWOULDBLOCK;
2842		goto no_block;
2843	}
2844
2845	rt_mutex_init_waiter(&rt_waiter);
2846
2847	/*
2848	 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2849	 * hold it while doing rt_mutex_start_proxy(), because then it will
2850	 * include hb->lock in the blocking chain, even through we'll not in
2851	 * fact hold it while blocking. This will lead it to report -EDEADLK
2852	 * and BUG when futex_unlock_pi() interleaves with this.
2853	 *
2854	 * Therefore acquire wait_lock while holding hb->lock, but drop the
2855	 * latter before calling __rt_mutex_start_proxy_lock(). This
2856	 * interleaves with futex_unlock_pi() -- which does a similar lock
2857	 * handoff -- such that the latter can observe the futex_q::pi_state
2858	 * before __rt_mutex_start_proxy_lock() is done.
2859	 */
2860	raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2861	spin_unlock(q.lock_ptr);
2862	/*
2863	 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2864	 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2865	 * it sees the futex_q::pi_state.
2866	 */
2867	ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2868	raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2869
2870	if (ret) {
2871		if (ret == 1)
2872			ret = 0;
2873		goto cleanup;
2874	}
2875
2876	if (unlikely(to))
2877		hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2878
2879	ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2880
2881cleanup:
2882	spin_lock(q.lock_ptr);
2883	/*
2884	 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2885	 * first acquire the hb->lock before removing the lock from the
2886	 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2887	 * lists consistent.
2888	 *
2889	 * In particular; it is important that futex_unlock_pi() can not
2890	 * observe this inconsistency.
2891	 */
2892	if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2893		ret = 0;
2894
2895no_block:
2896	/*
2897	 * Fixup the pi_state owner and possibly acquire the lock if we
2898	 * haven't already.
2899	 */
2900	res = fixup_owner(uaddr, &q, !ret);
2901	/*
2902	 * If fixup_owner() returned an error, propagate that.  If it acquired
2903	 * the lock, clear our -ETIMEDOUT or -EINTR.
2904	 */
2905	if (res)
2906		ret = (res < 0) ? res : 0;
2907
 
 
 
 
 
 
 
 
2908	unqueue_me_pi(&q);
2909	spin_unlock(q.lock_ptr);
2910	goto out;
2911
2912out_unlock_put_key:
2913	queue_unlock(hb);
2914
 
 
2915out:
2916	if (to) {
2917		hrtimer_cancel(&to->timer);
2918		destroy_hrtimer_on_stack(&to->timer);
2919	}
2920	return ret != -EINTR ? ret : -ERESTARTNOINTR;
2921
2922uaddr_faulted:
2923	queue_unlock(hb);
2924
2925	ret = fault_in_user_writeable(uaddr);
2926	if (ret)
2927		goto out;
2928
2929	if (!(flags & FLAGS_SHARED))
2930		goto retry_private;
2931
 
2932	goto retry;
2933}
2934
2935/*
2936 * Userspace attempted a TID -> 0 atomic transition, and failed.
2937 * This is the in-kernel slowpath: we look up the PI state (if any),
2938 * and do the rt-mutex unlock.
2939 */
2940static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2941{
2942	u32 curval, uval, vpid = task_pid_vnr(current);
 
2943	union futex_key key = FUTEX_KEY_INIT;
2944	struct futex_hash_bucket *hb;
2945	struct futex_q *top_waiter;
2946	int ret;
2947
2948	if (!IS_ENABLED(CONFIG_FUTEX_PI))
2949		return -ENOSYS;
2950
2951retry:
2952	if (get_user(uval, uaddr))
2953		return -EFAULT;
2954	/*
2955	 * We release only a lock we actually own:
2956	 */
2957	if ((uval & FUTEX_TID_MASK) != vpid)
2958		return -EPERM;
2959
2960	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2961	if (ret)
2962		return ret;
2963
2964	hb = hash_futex(&key);
2965	spin_lock(&hb->lock);
2966
2967	/*
2968	 * Check waiters first. We do not trust user space values at
2969	 * all and we at least want to know if user space fiddled
2970	 * with the futex value instead of blindly unlocking.
2971	 */
2972	top_waiter = futex_top_waiter(hb, &key);
2973	if (top_waiter) {
2974		struct futex_pi_state *pi_state = top_waiter->pi_state;
 
 
 
 
 
 
 
2975
2976		ret = -EINVAL;
2977		if (!pi_state)
2978			goto out_unlock;
2979
2980		/*
2981		 * If current does not own the pi_state then the futex is
2982		 * inconsistent and user space fiddled with the futex value.
2983		 */
2984		if (pi_state->owner != current)
2985			goto out_unlock;
2986
2987		get_pi_state(pi_state);
2988		/*
2989		 * By taking wait_lock while still holding hb->lock, we ensure
2990		 * there is no point where we hold neither; and therefore
2991		 * wake_futex_pi() must observe a state consistent with what we
2992		 * observed.
2993		 *
2994		 * In particular; this forces __rt_mutex_start_proxy() to
2995		 * complete such that we're guaranteed to observe the
2996		 * rt_waiter. Also see the WARN in wake_futex_pi().
2997		 */
2998		raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2999		spin_unlock(&hb->lock);
3000
3001		/* drops pi_state->pi_mutex.wait_lock */
3002		ret = wake_futex_pi(uaddr, uval, pi_state);
3003
3004		put_pi_state(pi_state);
3005
3006		/*
3007		 * Success, we're done! No tricky corner cases.
3008		 */
3009		if (!ret)
3010			return ret;
3011		/*
3012		 * The atomic access to the futex value generated a
3013		 * pagefault, so retry the user-access and the wakeup:
 
3014		 */
3015		if (ret == -EFAULT)
3016			goto pi_faulted;
3017		/*
3018		 * A unconditional UNLOCK_PI op raced against a waiter
3019		 * setting the FUTEX_WAITERS bit. Try again.
3020		 */
3021		if (ret == -EAGAIN)
3022			goto pi_retry;
3023		/*
3024		 * wake_futex_pi has detected invalid state. Tell user
3025		 * space.
3026		 */
3027		return ret;
3028	}
3029
3030	/*
3031	 * We have no kernel internal state, i.e. no waiters in the
3032	 * kernel. Waiters which are about to queue themselves are stuck
3033	 * on hb->lock. So we can safely ignore them. We do neither
3034	 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3035	 * owner.
3036	 */
3037	if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3038		spin_unlock(&hb->lock);
3039		switch (ret) {
3040		case -EFAULT:
3041			goto pi_faulted;
3042
3043		case -EAGAIN:
3044			goto pi_retry;
3045
3046		default:
3047			WARN_ON_ONCE(1);
3048			return ret;
3049		}
3050	}
3051
3052	/*
3053	 * If uval has changed, let user space handle it.
3054	 */
3055	ret = (curval == uval) ? 0 : -EAGAIN;
3056
3057out_unlock:
3058	spin_unlock(&hb->lock);
 
 
 
3059	return ret;
3060
3061pi_retry:
3062	cond_resched();
3063	goto retry;
3064
3065pi_faulted:
 
 
3066
3067	ret = fault_in_user_writeable(uaddr);
3068	if (!ret)
3069		goto retry;
3070
3071	return ret;
3072}
3073
3074/**
3075 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3076 * @hb:		the hash_bucket futex_q was original enqueued on
3077 * @q:		the futex_q woken while waiting to be requeued
3078 * @key2:	the futex_key of the requeue target futex
3079 * @timeout:	the timeout associated with the wait (NULL if none)
3080 *
3081 * Detect if the task was woken on the initial futex as opposed to the requeue
3082 * target futex.  If so, determine if it was a timeout or a signal that caused
3083 * the wakeup and return the appropriate error code to the caller.  Must be
3084 * called with the hb lock held.
3085 *
3086 * Return:
3087 *  -  0 = no early wakeup detected;
3088 *  - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3089 */
3090static inline
3091int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3092				   struct futex_q *q, union futex_key *key2,
3093				   struct hrtimer_sleeper *timeout)
3094{
3095	int ret = 0;
3096
3097	/*
3098	 * With the hb lock held, we avoid races while we process the wakeup.
3099	 * We only need to hold hb (and not hb2) to ensure atomicity as the
3100	 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3101	 * It can't be requeued from uaddr2 to something else since we don't
3102	 * support a PI aware source futex for requeue.
3103	 */
3104	if (!match_futex(&q->key, key2)) {
3105		WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3106		/*
3107		 * We were woken prior to requeue by a timeout or a signal.
3108		 * Unqueue the futex_q and determine which it was.
3109		 */
3110		plist_del(&q->list, &hb->chain);
3111		hb_waiters_dec(hb);
3112
3113		/* Handle spurious wakeups gracefully */
3114		ret = -EWOULDBLOCK;
3115		if (timeout && !timeout->task)
3116			ret = -ETIMEDOUT;
3117		else if (signal_pending(current))
3118			ret = -ERESTARTNOINTR;
3119	}
3120	return ret;
3121}
3122
3123/**
3124 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3125 * @uaddr:	the futex we initially wait on (non-pi)
3126 * @flags:	futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3127 *		the same type, no requeueing from private to shared, etc.
3128 * @val:	the expected value of uaddr
3129 * @abs_time:	absolute timeout
3130 * @bitset:	32 bit wakeup bitset set by userspace, defaults to all
3131 * @uaddr2:	the pi futex we will take prior to returning to user-space
3132 *
3133 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3134 * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
3135 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3136 * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
3137 * without one, the pi logic would not know which task to boost/deboost, if
3138 * there was a need to.
3139 *
3140 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3141 * via the following--
3142 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3143 * 2) wakeup on uaddr2 after a requeue
3144 * 3) signal
3145 * 4) timeout
3146 *
3147 * If 3, cleanup and return -ERESTARTNOINTR.
3148 *
3149 * If 2, we may then block on trying to take the rt_mutex and return via:
3150 * 5) successful lock
3151 * 6) signal
3152 * 7) timeout
3153 * 8) other lock acquisition failure
3154 *
3155 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3156 *
3157 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3158 *
3159 * Return:
3160 *  -  0 - On success;
3161 *  - <0 - On error
3162 */
3163static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3164				 u32 val, ktime_t *abs_time, u32 bitset,
3165				 u32 __user *uaddr2)
3166{
3167	struct hrtimer_sleeper timeout, *to;
3168	struct rt_mutex_waiter rt_waiter;
 
3169	struct futex_hash_bucket *hb;
3170	union futex_key key2 = FUTEX_KEY_INIT;
3171	struct futex_q q = futex_q_init;
3172	int res, ret;
3173
3174	if (!IS_ENABLED(CONFIG_FUTEX_PI))
3175		return -ENOSYS;
3176
3177	if (uaddr == uaddr2)
3178		return -EINVAL;
3179
3180	if (!bitset)
3181		return -EINVAL;
3182
3183	to = futex_setup_timer(abs_time, &timeout, flags,
3184			       current->timer_slack_ns);
 
 
 
 
 
 
 
3185
3186	/*
3187	 * The waiter is allocated on our stack, manipulated by the requeue
3188	 * code while we sleep on uaddr.
3189	 */
3190	rt_mutex_init_waiter(&rt_waiter);
 
 
 
3191
3192	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3193	if (unlikely(ret != 0))
3194		goto out;
3195
3196	q.bitset = bitset;
3197	q.rt_waiter = &rt_waiter;
3198	q.requeue_pi_key = &key2;
3199
3200	/*
3201	 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3202	 * count.
3203	 */
3204	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3205	if (ret)
3206		goto out;
3207
3208	/*
3209	 * The check above which compares uaddrs is not sufficient for
3210	 * shared futexes. We need to compare the keys:
3211	 */
3212	if (match_futex(&q.key, &key2)) {
3213		queue_unlock(hb);
3214		ret = -EINVAL;
3215		goto out;
3216	}
3217
3218	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
3219	futex_wait_queue_me(hb, &q, to);
3220
3221	spin_lock(&hb->lock);
3222	ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3223	spin_unlock(&hb->lock);
3224	if (ret)
3225		goto out;
3226
3227	/*
3228	 * In order for us to be here, we know our q.key == key2, and since
3229	 * we took the hb->lock above, we also know that futex_requeue() has
3230	 * completed and we no longer have to concern ourselves with a wakeup
3231	 * race with the atomic proxy lock acquisition by the requeue code. The
3232	 * futex_requeue dropped our key1 reference and incremented our key2
3233	 * reference count.
3234	 */
3235
3236	/*
3237	 * Check if the requeue code acquired the second futex for us and do
3238	 * any pertinent fixup.
3239	 */
3240	if (!q.rt_waiter) {
 
 
 
 
3241		if (q.pi_state && (q.pi_state->owner != current)) {
3242			spin_lock(q.lock_ptr);
3243			ret = fixup_owner(uaddr2, &q, true);
3244			/*
3245			 * Drop the reference to the pi state which
3246			 * the requeue_pi() code acquired for us.
3247			 */
3248			put_pi_state(q.pi_state);
3249			spin_unlock(q.lock_ptr);
3250			/*
3251			 * Adjust the return value. It's either -EFAULT or
3252			 * success (1) but the caller expects 0 for success.
3253			 */
3254			ret = ret < 0 ? ret : 0;
3255		}
3256	} else {
3257		struct rt_mutex *pi_mutex;
3258
3259		/*
3260		 * We have been woken up by futex_unlock_pi(), a timeout, or a
3261		 * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
3262		 * the pi_state.
3263		 */
3264		WARN_ON(!q.pi_state);
3265		pi_mutex = &q.pi_state->pi_mutex;
3266		ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
 
3267
3268		spin_lock(q.lock_ptr);
3269		if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3270			ret = 0;
3271
3272		debug_rt_mutex_free_waiter(&rt_waiter);
3273		/*
3274		 * Fixup the pi_state owner and possibly acquire the lock if we
3275		 * haven't already.
3276		 */
3277		res = fixup_owner(uaddr2, &q, !ret);
3278		/*
3279		 * If fixup_owner() returned an error, propagate that.  If it
3280		 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3281		 */
3282		if (res)
3283			ret = (res < 0) ? res : 0;
3284
 
3285		unqueue_me_pi(&q);
3286		spin_unlock(q.lock_ptr);
3287	}
3288
3289	if (ret == -EINTR) {
 
 
 
 
 
 
 
3290		/*
3291		 * We've already been requeued, but cannot restart by calling
3292		 * futex_lock_pi() directly. We could restart this syscall, but
3293		 * it would detect that the user space "val" changed and return
3294		 * -EWOULDBLOCK.  Save the overhead of the restart and return
3295		 * -EWOULDBLOCK directly.
3296		 */
3297		ret = -EWOULDBLOCK;
3298	}
3299
 
 
 
 
 
3300out:
3301	if (to) {
3302		hrtimer_cancel(&to->timer);
3303		destroy_hrtimer_on_stack(&to->timer);
3304	}
3305	return ret;
3306}
3307
3308/*
3309 * Support for robust futexes: the kernel cleans up held futexes at
3310 * thread exit time.
3311 *
3312 * Implementation: user-space maintains a per-thread list of locks it
3313 * is holding. Upon do_exit(), the kernel carefully walks this list,
3314 * and marks all locks that are owned by this thread with the
3315 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3316 * always manipulated with the lock held, so the list is private and
3317 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3318 * field, to allow the kernel to clean up if the thread dies after
3319 * acquiring the lock, but just before it could have added itself to
3320 * the list. There can only be one such pending lock.
3321 */
3322
3323/**
3324 * sys_set_robust_list() - Set the robust-futex list head of a task
3325 * @head:	pointer to the list-head
3326 * @len:	length of the list-head, as userspace expects
3327 */
3328SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3329		size_t, len)
3330{
3331	if (!futex_cmpxchg_enabled)
3332		return -ENOSYS;
3333	/*
3334	 * The kernel knows only one size for now:
3335	 */
3336	if (unlikely(len != sizeof(*head)))
3337		return -EINVAL;
3338
3339	current->robust_list = head;
3340
3341	return 0;
3342}
3343
3344/**
3345 * sys_get_robust_list() - Get the robust-futex list head of a task
3346 * @pid:	pid of the process [zero for current task]
3347 * @head_ptr:	pointer to a list-head pointer, the kernel fills it in
3348 * @len_ptr:	pointer to a length field, the kernel fills in the header size
3349 */
3350SYSCALL_DEFINE3(get_robust_list, int, pid,
3351		struct robust_list_head __user * __user *, head_ptr,
3352		size_t __user *, len_ptr)
3353{
3354	struct robust_list_head __user *head;
3355	unsigned long ret;
3356	struct task_struct *p;
3357
3358	if (!futex_cmpxchg_enabled)
3359		return -ENOSYS;
3360
3361	rcu_read_lock();
3362
3363	ret = -ESRCH;
3364	if (!pid)
3365		p = current;
3366	else {
3367		p = find_task_by_vpid(pid);
3368		if (!p)
3369			goto err_unlock;
3370	}
3371
3372	ret = -EPERM;
3373	if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3374		goto err_unlock;
3375
3376	head = p->robust_list;
3377	rcu_read_unlock();
3378
3379	if (put_user(sizeof(*head), len_ptr))
3380		return -EFAULT;
3381	return put_user(head, head_ptr);
3382
3383err_unlock:
3384	rcu_read_unlock();
3385
3386	return ret;
3387}
3388
3389/* Constants for the pending_op argument of handle_futex_death */
3390#define HANDLE_DEATH_PENDING	true
3391#define HANDLE_DEATH_LIST	false
3392
3393/*
3394 * Process a futex-list entry, check whether it's owned by the
3395 * dying task, and do notification if so:
3396 */
3397static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3398			      bool pi, bool pending_op)
3399{
3400	u32 uval, nval, mval;
3401	int err;
3402
3403	/* Futex address must be 32bit aligned */
3404	if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3405		return -1;
3406
3407retry:
3408	if (get_user(uval, uaddr))
3409		return -1;
3410
3411	/*
3412	 * Special case for regular (non PI) futexes. The unlock path in
3413	 * user space has two race scenarios:
3414	 *
3415	 * 1. The unlock path releases the user space futex value and
3416	 *    before it can execute the futex() syscall to wake up
3417	 *    waiters it is killed.
3418	 *
3419	 * 2. A woken up waiter is killed before it can acquire the
3420	 *    futex in user space.
3421	 *
3422	 * In both cases the TID validation below prevents a wakeup of
3423	 * potential waiters which can cause these waiters to block
3424	 * forever.
3425	 *
3426	 * In both cases the following conditions are met:
3427	 *
3428	 *	1) task->robust_list->list_op_pending != NULL
3429	 *	   @pending_op == true
3430	 *	2) User space futex value == 0
3431	 *	3) Regular futex: @pi == false
3432	 *
3433	 * If these conditions are met, it is safe to attempt waking up a
3434	 * potential waiter without touching the user space futex value and
3435	 * trying to set the OWNER_DIED bit. The user space futex value is
3436	 * uncontended and the rest of the user space mutex state is
3437	 * consistent, so a woken waiter will just take over the
3438	 * uncontended futex. Setting the OWNER_DIED bit would create
3439	 * inconsistent state and malfunction of the user space owner died
3440	 * handling.
3441	 */
3442	if (pending_op && !pi && !uval) {
3443		futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3444		return 0;
3445	}
3446
3447	if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr))
3448		return 0;
3449
3450	/*
3451	 * Ok, this dying thread is truly holding a futex
3452	 * of interest. Set the OWNER_DIED bit atomically
3453	 * via cmpxchg, and if the value had FUTEX_WAITERS
3454	 * set, wake up a waiter (if any). (We have to do a
3455	 * futex_wake() even if OWNER_DIED is already set -
3456	 * to handle the rare but possible case of recursive
3457	 * thread-death.) The rest of the cleanup is done in
3458	 * userspace.
3459	 */
3460	mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3461
3462	/*
3463	 * We are not holding a lock here, but we want to have
3464	 * the pagefault_disable/enable() protection because
3465	 * we want to handle the fault gracefully. If the
3466	 * access fails we try to fault in the futex with R/W
3467	 * verification via get_user_pages. get_user() above
3468	 * does not guarantee R/W access. If that fails we
3469	 * give up and leave the futex locked.
3470	 */
3471	if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3472		switch (err) {
3473		case -EFAULT:
3474			if (fault_in_user_writeable(uaddr))
3475				return -1;
3476			goto retry;
3477
3478		case -EAGAIN:
3479			cond_resched();
3480			goto retry;
3481
3482		default:
3483			WARN_ON_ONCE(1);
3484			return err;
3485		}
 
 
3486	}
3487
3488	if (nval != uval)
3489		goto retry;
3490
3491	/*
3492	 * Wake robust non-PI futexes here. The wakeup of
3493	 * PI futexes happens in exit_pi_state():
3494	 */
3495	if (!pi && (uval & FUTEX_WAITERS))
3496		futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3497
3498	return 0;
3499}
3500
3501/*
3502 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3503 */
3504static inline int fetch_robust_entry(struct robust_list __user **entry,
3505				     struct robust_list __user * __user *head,
3506				     unsigned int *pi)
3507{
3508	unsigned long uentry;
3509
3510	if (get_user(uentry, (unsigned long __user *)head))
3511		return -EFAULT;
3512
3513	*entry = (void __user *)(uentry & ~1UL);
3514	*pi = uentry & 1;
3515
3516	return 0;
3517}
3518
3519/*
3520 * Walk curr->robust_list (very carefully, it's a userspace list!)
3521 * and mark any locks found there dead, and notify any waiters.
3522 *
3523 * We silently return on any sign of list-walking problem.
3524 */
3525static void exit_robust_list(struct task_struct *curr)
3526{
3527	struct robust_list_head __user *head = curr->robust_list;
3528	struct robust_list __user *entry, *next_entry, *pending;
3529	unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3530	unsigned int next_pi;
3531	unsigned long futex_offset;
3532	int rc;
3533
3534	if (!futex_cmpxchg_enabled)
3535		return;
3536
3537	/*
3538	 * Fetch the list head (which was registered earlier, via
3539	 * sys_set_robust_list()):
3540	 */
3541	if (fetch_robust_entry(&entry, &head->list.next, &pi))
3542		return;
3543	/*
3544	 * Fetch the relative futex offset:
3545	 */
3546	if (get_user(futex_offset, &head->futex_offset))
3547		return;
3548	/*
3549	 * Fetch any possibly pending lock-add first, and handle it
3550	 * if it exists:
3551	 */
3552	if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3553		return;
3554
3555	next_entry = NULL;	/* avoid warning with gcc */
3556	while (entry != &head->list) {
3557		/*
3558		 * Fetch the next entry in the list before calling
3559		 * handle_futex_death:
3560		 */
3561		rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3562		/*
3563		 * A pending lock might already be on the list, so
3564		 * don't process it twice:
3565		 */
3566		if (entry != pending) {
3567			if (handle_futex_death((void __user *)entry + futex_offset,
3568						curr, pi, HANDLE_DEATH_LIST))
3569				return;
3570		}
3571		if (rc)
3572			return;
3573		entry = next_entry;
3574		pi = next_pi;
3575		/*
3576		 * Avoid excessively long or circular lists:
3577		 */
3578		if (!--limit)
3579			break;
3580
3581		cond_resched();
3582	}
3583
3584	if (pending) {
3585		handle_futex_death((void __user *)pending + futex_offset,
3586				   curr, pip, HANDLE_DEATH_PENDING);
3587	}
3588}
3589
3590static void futex_cleanup(struct task_struct *tsk)
3591{
3592	if (unlikely(tsk->robust_list)) {
3593		exit_robust_list(tsk);
3594		tsk->robust_list = NULL;
3595	}
3596
3597#ifdef CONFIG_COMPAT
3598	if (unlikely(tsk->compat_robust_list)) {
3599		compat_exit_robust_list(tsk);
3600		tsk->compat_robust_list = NULL;
3601	}
3602#endif
3603
3604	if (unlikely(!list_empty(&tsk->pi_state_list)))
3605		exit_pi_state_list(tsk);
3606}
3607
3608/**
3609 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3610 * @tsk:	task to set the state on
3611 *
3612 * Set the futex exit state of the task lockless. The futex waiter code
3613 * observes that state when a task is exiting and loops until the task has
3614 * actually finished the futex cleanup. The worst case for this is that the
3615 * waiter runs through the wait loop until the state becomes visible.
3616 *
3617 * This is called from the recursive fault handling path in do_exit().
3618 *
3619 * This is best effort. Either the futex exit code has run already or
3620 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3621 * take it over. If not, the problem is pushed back to user space. If the
3622 * futex exit code did not run yet, then an already queued waiter might
3623 * block forever, but there is nothing which can be done about that.
3624 */
3625void futex_exit_recursive(struct task_struct *tsk)
3626{
3627	/* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3628	if (tsk->futex_state == FUTEX_STATE_EXITING)
3629		mutex_unlock(&tsk->futex_exit_mutex);
3630	tsk->futex_state = FUTEX_STATE_DEAD;
3631}
3632
3633static void futex_cleanup_begin(struct task_struct *tsk)
3634{
3635	/*
3636	 * Prevent various race issues against a concurrent incoming waiter
3637	 * including live locks by forcing the waiter to block on
3638	 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3639	 * attach_to_pi_owner().
3640	 */
3641	mutex_lock(&tsk->futex_exit_mutex);
3642
3643	/*
3644	 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3645	 *
3646	 * This ensures that all subsequent checks of tsk->futex_state in
3647	 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3648	 * tsk->pi_lock held.
3649	 *
3650	 * It guarantees also that a pi_state which was queued right before
3651	 * the state change under tsk->pi_lock by a concurrent waiter must
3652	 * be observed in exit_pi_state_list().
3653	 */
3654	raw_spin_lock_irq(&tsk->pi_lock);
3655	tsk->futex_state = FUTEX_STATE_EXITING;
3656	raw_spin_unlock_irq(&tsk->pi_lock);
3657}
3658
3659static void futex_cleanup_end(struct task_struct *tsk, int state)
3660{
3661	/*
3662	 * Lockless store. The only side effect is that an observer might
3663	 * take another loop until it becomes visible.
3664	 */
3665	tsk->futex_state = state;
3666	/*
3667	 * Drop the exit protection. This unblocks waiters which observed
3668	 * FUTEX_STATE_EXITING to reevaluate the state.
3669	 */
3670	mutex_unlock(&tsk->futex_exit_mutex);
3671}
3672
3673void futex_exec_release(struct task_struct *tsk)
3674{
3675	/*
3676	 * The state handling is done for consistency, but in the case of
3677	 * exec() there is no way to prevent further damage as the PID stays
3678	 * the same. But for the unlikely and arguably buggy case that a
3679	 * futex is held on exec(), this provides at least as much state
3680	 * consistency protection which is possible.
3681	 */
3682	futex_cleanup_begin(tsk);
3683	futex_cleanup(tsk);
3684	/*
3685	 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3686	 * exec a new binary.
3687	 */
3688	futex_cleanup_end(tsk, FUTEX_STATE_OK);
3689}
3690
3691void futex_exit_release(struct task_struct *tsk)
3692{
3693	futex_cleanup_begin(tsk);
3694	futex_cleanup(tsk);
3695	futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3696}
3697
3698long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3699		u32 __user *uaddr2, u32 val2, u32 val3)
3700{
3701	int cmd = op & FUTEX_CMD_MASK;
3702	unsigned int flags = 0;
3703
3704	if (!(op & FUTEX_PRIVATE_FLAG))
3705		flags |= FLAGS_SHARED;
3706
3707	if (op & FUTEX_CLOCK_REALTIME) {
3708		flags |= FLAGS_CLOCKRT;
3709		if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI &&
3710		    cmd != FUTEX_LOCK_PI2)
3711			return -ENOSYS;
3712	}
3713
3714	switch (cmd) {
3715	case FUTEX_LOCK_PI:
3716	case FUTEX_LOCK_PI2:
3717	case FUTEX_UNLOCK_PI:
3718	case FUTEX_TRYLOCK_PI:
3719	case FUTEX_WAIT_REQUEUE_PI:
3720	case FUTEX_CMP_REQUEUE_PI:
3721		if (!futex_cmpxchg_enabled)
3722			return -ENOSYS;
3723	}
3724
3725	switch (cmd) {
3726	case FUTEX_WAIT:
3727		val3 = FUTEX_BITSET_MATCH_ANY;
3728		fallthrough;
3729	case FUTEX_WAIT_BITSET:
3730		return futex_wait(uaddr, flags, val, timeout, val3);
3731	case FUTEX_WAKE:
3732		val3 = FUTEX_BITSET_MATCH_ANY;
3733		fallthrough;
3734	case FUTEX_WAKE_BITSET:
3735		return futex_wake(uaddr, flags, val, val3);
3736	case FUTEX_REQUEUE:
3737		return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3738	case FUTEX_CMP_REQUEUE:
3739		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3740	case FUTEX_WAKE_OP:
3741		return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3742	case FUTEX_LOCK_PI:
3743		flags |= FLAGS_CLOCKRT;
3744		fallthrough;
3745	case FUTEX_LOCK_PI2:
3746		return futex_lock_pi(uaddr, flags, timeout, 0);
3747	case FUTEX_UNLOCK_PI:
3748		return futex_unlock_pi(uaddr, flags);
3749	case FUTEX_TRYLOCK_PI:
3750		return futex_lock_pi(uaddr, flags, NULL, 1);
3751	case FUTEX_WAIT_REQUEUE_PI:
3752		val3 = FUTEX_BITSET_MATCH_ANY;
3753		return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3754					     uaddr2);
3755	case FUTEX_CMP_REQUEUE_PI:
3756		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3757	}
3758	return -ENOSYS;
3759}
3760
3761static __always_inline bool futex_cmd_has_timeout(u32 cmd)
3762{
3763	switch (cmd) {
3764	case FUTEX_WAIT:
3765	case FUTEX_LOCK_PI:
3766	case FUTEX_LOCK_PI2:
3767	case FUTEX_WAIT_BITSET:
3768	case FUTEX_WAIT_REQUEUE_PI:
3769		return true;
3770	}
3771	return false;
3772}
3773
3774static __always_inline int
3775futex_init_timeout(u32 cmd, u32 op, struct timespec64 *ts, ktime_t *t)
3776{
3777	if (!timespec64_valid(ts))
3778		return -EINVAL;
3779
3780	*t = timespec64_to_ktime(*ts);
3781	if (cmd == FUTEX_WAIT)
3782		*t = ktime_add_safe(ktime_get(), *t);
3783	else if (cmd != FUTEX_LOCK_PI && !(op & FUTEX_CLOCK_REALTIME))
3784		*t = timens_ktime_to_host(CLOCK_MONOTONIC, *t);
3785	return 0;
3786}
3787
3788SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3789		const struct __kernel_timespec __user *, utime,
3790		u32 __user *, uaddr2, u32, val3)
3791{
3792	int ret, cmd = op & FUTEX_CMD_MASK;
3793	ktime_t t, *tp = NULL;
3794	struct timespec64 ts;
 
3795
3796	if (utime && futex_cmd_has_timeout(cmd)) {
3797		if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
 
 
3798			return -EFAULT;
3799		if (get_timespec64(&ts, utime))
3800			return -EFAULT;
3801		ret = futex_init_timeout(cmd, op, &ts, &t);
3802		if (ret)
3803			return ret;
 
3804		tp = &t;
3805	}
3806
3807	return do_futex(uaddr, op, val, tp, uaddr2, (unsigned long)utime, val3);
3808}
3809
3810#ifdef CONFIG_COMPAT
3811/*
3812 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3813 */
3814static inline int
3815compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3816		   compat_uptr_t __user *head, unsigned int *pi)
3817{
3818	if (get_user(*uentry, head))
3819		return -EFAULT;
3820
3821	*entry = compat_ptr((*uentry) & ~1);
3822	*pi = (unsigned int)(*uentry) & 1;
3823
3824	return 0;
3825}
3826
3827static void __user *futex_uaddr(struct robust_list __user *entry,
3828				compat_long_t futex_offset)
3829{
3830	compat_uptr_t base = ptr_to_compat(entry);
3831	void __user *uaddr = compat_ptr(base + futex_offset);
3832
3833	return uaddr;
3834}
3835
3836/*
3837 * Walk curr->robust_list (very carefully, it's a userspace list!)
3838 * and mark any locks found there dead, and notify any waiters.
3839 *
3840 * We silently return on any sign of list-walking problem.
3841 */
3842static void compat_exit_robust_list(struct task_struct *curr)
3843{
3844	struct compat_robust_list_head __user *head = curr->compat_robust_list;
3845	struct robust_list __user *entry, *next_entry, *pending;
3846	unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3847	unsigned int next_pi;
3848	compat_uptr_t uentry, next_uentry, upending;
3849	compat_long_t futex_offset;
3850	int rc;
3851
3852	if (!futex_cmpxchg_enabled)
3853		return;
3854
3855	/*
3856	 * Fetch the list head (which was registered earlier, via
3857	 * sys_set_robust_list()):
3858	 */
3859	if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3860		return;
3861	/*
3862	 * Fetch the relative futex offset:
 
3863	 */
3864	if (get_user(futex_offset, &head->futex_offset))
3865		return;
3866	/*
3867	 * Fetch any possibly pending lock-add first, and handle it
3868	 * if it exists:
3869	 */
3870	if (compat_fetch_robust_entry(&upending, &pending,
3871			       &head->list_op_pending, &pip))
3872		return;
3873
3874	next_entry = NULL;	/* avoid warning with gcc */
3875	while (entry != (struct robust_list __user *) &head->list) {
3876		/*
3877		 * Fetch the next entry in the list before calling
3878		 * handle_futex_death:
3879		 */
3880		rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3881			(compat_uptr_t __user *)&entry->next, &next_pi);
3882		/*
3883		 * A pending lock might already be on the list, so
3884		 * dont process it twice:
3885		 */
3886		if (entry != pending) {
3887			void __user *uaddr = futex_uaddr(entry, futex_offset);
3888
3889			if (handle_futex_death(uaddr, curr, pi,
3890					       HANDLE_DEATH_LIST))
3891				return;
3892		}
3893		if (rc)
3894			return;
3895		uentry = next_uentry;
3896		entry = next_entry;
3897		pi = next_pi;
3898		/*
3899		 * Avoid excessively long or circular lists:
3900		 */
3901		if (!--limit)
3902			break;
3903
3904		cond_resched();
3905	}
3906	if (pending) {
3907		void __user *uaddr = futex_uaddr(pending, futex_offset);
3908
3909		handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3910	}
3911}
3912
3913COMPAT_SYSCALL_DEFINE2(set_robust_list,
3914		struct compat_robust_list_head __user *, head,
3915		compat_size_t, len)
3916{
3917	if (!futex_cmpxchg_enabled)
3918		return -ENOSYS;
3919
3920	if (unlikely(len != sizeof(*head)))
3921		return -EINVAL;
3922
3923	current->compat_robust_list = head;
3924
3925	return 0;
3926}
3927
3928COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3929			compat_uptr_t __user *, head_ptr,
3930			compat_size_t __user *, len_ptr)
3931{
3932	struct compat_robust_list_head __user *head;
3933	unsigned long ret;
3934	struct task_struct *p;
3935
3936	if (!futex_cmpxchg_enabled)
3937		return -ENOSYS;
3938
3939	rcu_read_lock();
3940
3941	ret = -ESRCH;
3942	if (!pid)
3943		p = current;
3944	else {
3945		p = find_task_by_vpid(pid);
3946		if (!p)
3947			goto err_unlock;
3948	}
3949
3950	ret = -EPERM;
3951	if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3952		goto err_unlock;
3953
3954	head = p->compat_robust_list;
3955	rcu_read_unlock();
3956
3957	if (put_user(sizeof(*head), len_ptr))
3958		return -EFAULT;
3959	return put_user(ptr_to_compat(head), head_ptr);
3960
3961err_unlock:
3962	rcu_read_unlock();
3963
3964	return ret;
3965}
3966#endif /* CONFIG_COMPAT */
3967
3968#ifdef CONFIG_COMPAT_32BIT_TIME
3969SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3970		const struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3971		u32, val3)
3972{
3973	int ret, cmd = op & FUTEX_CMD_MASK;
3974	ktime_t t, *tp = NULL;
3975	struct timespec64 ts;
3976
3977	if (utime && futex_cmd_has_timeout(cmd)) {
3978		if (get_old_timespec32(&ts, utime))
3979			return -EFAULT;
3980		ret = futex_init_timeout(cmd, op, &ts, &t);
3981		if (ret)
3982			return ret;
3983		tp = &t;
3984	}
3985
3986	return do_futex(uaddr, op, val, tp, uaddr2, (unsigned long)utime, val3);
3987}
3988#endif /* CONFIG_COMPAT_32BIT_TIME */
3989
3990static void __init futex_detect_cmpxchg(void)
3991{
3992#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3993	u32 curval;
3994
3995	/*
3996	 * This will fail and we want it. Some arch implementations do
3997	 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3998	 * functionality. We want to know that before we call in any
3999	 * of the complex code paths. Also we want to prevent
4000	 * registration of robust lists in that case. NULL is
4001	 * guaranteed to fault and we get -EFAULT on functional
4002	 * implementation, the non-functional ones will return
4003	 * -ENOSYS.
4004	 */
4005	if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4006		futex_cmpxchg_enabled = 1;
4007#endif
4008}
4009
4010static int __init futex_init(void)
4011{
4012	unsigned int futex_shift;
4013	unsigned long i;
4014
4015#if CONFIG_BASE_SMALL
4016	futex_hashsize = 16;
4017#else
4018	futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4019#endif
4020
4021	futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4022					       futex_hashsize, 0,
4023					       futex_hashsize < 256 ? HASH_SMALL : 0,
4024					       &futex_shift, NULL,
4025					       futex_hashsize, futex_hashsize);
4026	futex_hashsize = 1UL << futex_shift;
4027
4028	futex_detect_cmpxchg();
4029
4030	for (i = 0; i < futex_hashsize; i++) {
4031		atomic_set(&futex_queues[i].waiters, 0);
4032		plist_head_init(&futex_queues[i].chain);
4033		spin_lock_init(&futex_queues[i].lock);
4034	}
4035
4036	return 0;
4037}
4038core_initcall(futex_init);