Root/kernel/futex.c

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/module.h>
59#include <linux/magic.h>
60#include <linux/pid.h>
61#include <linux/nsproxy.h>
62
63#include <asm/futex.h>
64
65#include "rtmutex_common.h"
66
67int __read_mostly futex_cmpxchg_enabled;
68
69#define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8)
70
71/*
72 * Priority Inheritance state:
73 */
74struct futex_pi_state {
75    /*
76     * list of 'owned' pi_state instances - these have to be
77     * cleaned up in do_exit() if the task exits prematurely:
78     */
79    struct list_head list;
80
81    /*
82     * The PI object:
83     */
84    struct rt_mutex pi_mutex;
85
86    struct task_struct *owner;
87    atomic_t refcount;
88
89    union futex_key key;
90};
91
92/**
93 * struct futex_q - The hashed futex queue entry, one per waiting task
94 * @task: the task waiting on the futex
95 * @lock_ptr: the hash bucket lock
96 * @key: the key the futex is hashed on
97 * @pi_state: optional priority inheritance state
98 * @rt_waiter: rt_waiter storage for use with requeue_pi
99 * @requeue_pi_key: the requeue_pi target futex key
100 * @bitset: bitset for the optional bitmasked wakeup
101 *
102 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
103 * we can wake only the relevant ones (hashed queues may be shared).
104 *
105 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
106 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
107 * The order of wakup is always to make the first condition true, then
108 * the second.
109 *
110 * PI futexes are typically woken before they are removed from the hash list via
111 * the rt_mutex code. See unqueue_me_pi().
112 */
113struct futex_q {
114    struct plist_node list;
115
116    struct task_struct *task;
117    spinlock_t *lock_ptr;
118    union futex_key key;
119    struct futex_pi_state *pi_state;
120    struct rt_mutex_waiter *rt_waiter;
121    union futex_key *requeue_pi_key;
122    u32 bitset;
123};
124
125/*
126 * Hash buckets are shared by all the futex_keys that hash to the same
127 * location. Each key may have multiple futex_q structures, one for each task
128 * waiting on a futex.
129 */
130struct futex_hash_bucket {
131    spinlock_t lock;
132    struct plist_head chain;
133};
134
135static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS];
136
137/*
138 * We hash on the keys returned from get_futex_key (see below).
139 */
140static struct futex_hash_bucket *hash_futex(union futex_key *key)
141{
142    u32 hash = jhash2((u32*)&key->both.word,
143              (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
144              key->both.offset);
145    return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)];
146}
147
148/*
149 * Return 1 if two futex_keys are equal, 0 otherwise.
150 */
151static inline int match_futex(union futex_key *key1, union futex_key *key2)
152{
153    return (key1 && key2
154        && key1->both.word == key2->both.word
155        && key1->both.ptr == key2->both.ptr
156        && key1->both.offset == key2->both.offset);
157}
158
159/*
160 * Take a reference to the resource addressed by a key.
161 * Can be called while holding spinlocks.
162 *
163 */
164static void get_futex_key_refs(union futex_key *key)
165{
166    if (!key->both.ptr)
167        return;
168
169    switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
170    case FUT_OFF_INODE:
171        atomic_inc(&key->shared.inode->i_count);
172        break;
173    case FUT_OFF_MMSHARED:
174        atomic_inc(&key->private.mm->mm_count);
175        break;
176    }
177}
178
179/*
180 * Drop a reference to the resource addressed by a key.
181 * The hash bucket spinlock must not be held.
182 */
183static void drop_futex_key_refs(union futex_key *key)
184{
185    if (!key->both.ptr) {
186        /* If we're here then we tried to put a key we failed to get */
187        WARN_ON_ONCE(1);
188        return;
189    }
190
191    switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
192    case FUT_OFF_INODE:
193        iput(key->shared.inode);
194        break;
195    case FUT_OFF_MMSHARED:
196        mmdrop(key->private.mm);
197        break;
198    }
199}
200
201/**
202 * get_futex_key() - Get parameters which are the keys for a futex
203 * @uaddr: virtual address of the futex
204 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
205 * @key: address where result is stored.
206 *
207 * Returns a negative error code or 0
208 * The key words are stored in *key on success.
209 *
210 * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
211 * offset_within_page). For private mappings, it's (uaddr, current->mm).
212 * We can usually work out the index without swapping in the page.
213 *
214 * lock_page() might sleep, the caller should not hold a spinlock.
215 */
216static int
217get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key)
218{
219    unsigned long address = (unsigned long)uaddr;
220    struct mm_struct *mm = current->mm;
221    struct page *page;
222    int err;
223
224    /*
225     * The futex address must be "naturally" aligned.
226     */
227    key->both.offset = address % PAGE_SIZE;
228    if (unlikely((address % sizeof(u32)) != 0))
229        return -EINVAL;
230    address -= key->both.offset;
231
232    /*
233     * PROCESS_PRIVATE futexes are fast.
234     * As the mm cannot disappear under us and the 'key' only needs
235     * virtual address, we dont even have to find the underlying vma.
236     * Note : We do have to check 'uaddr' is a valid user address,
237     * but access_ok() should be faster than find_vma()
238     */
239    if (!fshared) {
240        if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
241            return -EFAULT;
242        key->private.mm = mm;
243        key->private.address = address;
244        get_futex_key_refs(key);
245        return 0;
246    }
247
248again:
249    err = get_user_pages_fast(address, 1, 1, &page);
250    if (err < 0)
251        return err;
252
253    page = compound_head(page);
254    lock_page(page);
255    if (!page->mapping) {
256        unlock_page(page);
257        put_page(page);
258        goto again;
259    }
260
261    /*
262     * Private mappings are handled in a simple way.
263     *
264     * NOTE: When userspace waits on a MAP_SHARED mapping, even if
265     * it's a read-only handle, it's expected that futexes attach to
266     * the object not the particular process.
267     */
268    if (PageAnon(page)) {
269        key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
270        key->private.mm = mm;
271        key->private.address = address;
272    } else {
273        key->both.offset |= FUT_OFF_INODE; /* inode-based key */
274        key->shared.inode = page->mapping->host;
275        key->shared.pgoff = page->index;
276    }
277
278    get_futex_key_refs(key);
279
280    unlock_page(page);
281    put_page(page);
282    return 0;
283}
284
285static inline
286void put_futex_key(int fshared, union futex_key *key)
287{
288    drop_futex_key_refs(key);
289}
290
291/**
292 * fault_in_user_writeable() - Fault in user address and verify RW access
293 * @uaddr: pointer to faulting user space address
294 *
295 * Slow path to fixup the fault we just took in the atomic write
296 * access to @uaddr.
297 *
298 * We have no generic implementation of a non destructive write to the
299 * user address. We know that we faulted in the atomic pagefault
300 * disabled section so we can as well avoid the #PF overhead by
301 * calling get_user_pages() right away.
302 */
303static int fault_in_user_writeable(u32 __user *uaddr)
304{
305    struct mm_struct *mm = current->mm;
306    int ret;
307
308    down_read(&mm->mmap_sem);
309    ret = get_user_pages(current, mm, (unsigned long)uaddr,
310                 1, 1, 0, NULL, NULL);
311    up_read(&mm->mmap_sem);
312
313    return ret < 0 ? ret : 0;
314}
315
316/**
317 * futex_top_waiter() - Return the highest priority waiter on a futex
318 * @hb: the hash bucket the futex_q's reside in
319 * @key: the futex key (to distinguish it from other futex futex_q's)
320 *
321 * Must be called with the hb lock held.
322 */
323static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
324                    union futex_key *key)
325{
326    struct futex_q *this;
327
328    plist_for_each_entry(this, &hb->chain, list) {
329        if (match_futex(&this->key, key))
330            return this;
331    }
332    return NULL;
333}
334
335static u32 cmpxchg_futex_value_locked(u32 __user *uaddr, u32 uval, u32 newval)
336{
337    u32 curval;
338
339    pagefault_disable();
340    curval = futex_atomic_cmpxchg_inatomic(uaddr, uval, newval);
341    pagefault_enable();
342
343    return curval;
344}
345
346static int get_futex_value_locked(u32 *dest, u32 __user *from)
347{
348    int ret;
349
350    pagefault_disable();
351    ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
352    pagefault_enable();
353
354    return ret ? -EFAULT : 0;
355}
356
357
358/*
359 * PI code:
360 */
361static int refill_pi_state_cache(void)
362{
363    struct futex_pi_state *pi_state;
364
365    if (likely(current->pi_state_cache))
366        return 0;
367
368    pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
369
370    if (!pi_state)
371        return -ENOMEM;
372
373    INIT_LIST_HEAD(&pi_state->list);
374    /* pi_mutex gets initialized later */
375    pi_state->owner = NULL;
376    atomic_set(&pi_state->refcount, 1);
377    pi_state->key = FUTEX_KEY_INIT;
378
379    current->pi_state_cache = pi_state;
380
381    return 0;
382}
383
384static struct futex_pi_state * alloc_pi_state(void)
385{
386    struct futex_pi_state *pi_state = current->pi_state_cache;
387
388    WARN_ON(!pi_state);
389    current->pi_state_cache = NULL;
390
391    return pi_state;
392}
393
394static void free_pi_state(struct futex_pi_state *pi_state)
395{
396    if (!atomic_dec_and_test(&pi_state->refcount))
397        return;
398
399    /*
400     * If pi_state->owner is NULL, the owner is most probably dying
401     * and has cleaned up the pi_state already
402     */
403    if (pi_state->owner) {
404        raw_spin_lock_irq(&pi_state->owner->pi_lock);
405        list_del_init(&pi_state->list);
406        raw_spin_unlock_irq(&pi_state->owner->pi_lock);
407
408        rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
409    }
410
411    if (current->pi_state_cache)
412        kfree(pi_state);
413    else {
414        /*
415         * pi_state->list is already empty.
416         * clear pi_state->owner.
417         * refcount is at 0 - put it back to 1.
418         */
419        pi_state->owner = NULL;
420        atomic_set(&pi_state->refcount, 1);
421        current->pi_state_cache = pi_state;
422    }
423}
424
425/*
426 * Look up the task based on what TID userspace gave us.
427 * We dont trust it.
428 */
429static struct task_struct * futex_find_get_task(pid_t pid)
430{
431    struct task_struct *p;
432
433    rcu_read_lock();
434    p = find_task_by_vpid(pid);
435    if (p)
436        get_task_struct(p);
437
438    rcu_read_unlock();
439
440    return p;
441}
442
443/*
444 * This task is holding PI mutexes at exit time => bad.
445 * Kernel cleans up PI-state, but userspace is likely hosed.
446 * (Robust-futex cleanup is separate and might save the day for userspace.)
447 */
448void exit_pi_state_list(struct task_struct *curr)
449{
450    struct list_head *next, *head = &curr->pi_state_list;
451    struct futex_pi_state *pi_state;
452    struct futex_hash_bucket *hb;
453    union futex_key key = FUTEX_KEY_INIT;
454
455    if (!futex_cmpxchg_enabled)
456        return;
457    /*
458     * We are a ZOMBIE and nobody can enqueue itself on
459     * pi_state_list anymore, but we have to be careful
460     * versus waiters unqueueing themselves:
461     */
462    raw_spin_lock_irq(&curr->pi_lock);
463    while (!list_empty(head)) {
464
465        next = head->next;
466        pi_state = list_entry(next, struct futex_pi_state, list);
467        key = pi_state->key;
468        hb = hash_futex(&key);
469        raw_spin_unlock_irq(&curr->pi_lock);
470
471        spin_lock(&hb->lock);
472
473        raw_spin_lock_irq(&curr->pi_lock);
474        /*
475         * We dropped the pi-lock, so re-check whether this
476         * task still owns the PI-state:
477         */
478        if (head->next != next) {
479            spin_unlock(&hb->lock);
480            continue;
481        }
482
483        WARN_ON(pi_state->owner != curr);
484        WARN_ON(list_empty(&pi_state->list));
485        list_del_init(&pi_state->list);
486        pi_state->owner = NULL;
487        raw_spin_unlock_irq(&curr->pi_lock);
488
489        rt_mutex_unlock(&pi_state->pi_mutex);
490
491        spin_unlock(&hb->lock);
492
493        raw_spin_lock_irq(&curr->pi_lock);
494    }
495    raw_spin_unlock_irq(&curr->pi_lock);
496}
497
498static int
499lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
500        union futex_key *key, struct futex_pi_state **ps)
501{
502    struct futex_pi_state *pi_state = NULL;
503    struct futex_q *this, *next;
504    struct plist_head *head;
505    struct task_struct *p;
506    pid_t pid = uval & FUTEX_TID_MASK;
507
508    head = &hb->chain;
509
510    plist_for_each_entry_safe(this, next, head, list) {
511        if (match_futex(&this->key, key)) {
512            /*
513             * Another waiter already exists - bump up
514             * the refcount and return its pi_state:
515             */
516            pi_state = this->pi_state;
517            /*
518             * Userspace might have messed up non PI and PI futexes
519             */
520            if (unlikely(!pi_state))
521                return -EINVAL;
522
523            WARN_ON(!atomic_read(&pi_state->refcount));
524
525            /*
526             * When pi_state->owner is NULL then the owner died
527             * and another waiter is on the fly. pi_state->owner
528             * is fixed up by the task which acquires
529             * pi_state->rt_mutex.
530             *
531             * We do not check for pid == 0 which can happen when
532             * the owner died and robust_list_exit() cleared the
533             * TID.
534             */
535            if (pid && pi_state->owner) {
536                /*
537                 * Bail out if user space manipulated the
538                 * futex value.
539                 */
540                if (pid != task_pid_vnr(pi_state->owner))
541                    return -EINVAL;
542            }
543
544            atomic_inc(&pi_state->refcount);
545            *ps = pi_state;
546
547            return 0;
548        }
549    }
550
551    /*
552     * We are the first waiter - try to look up the real owner and attach
553     * the new pi_state to it, but bail out when TID = 0
554     */
555    if (!pid)
556        return -ESRCH;
557    p = futex_find_get_task(pid);
558    if (!p)
559        return -ESRCH;
560
561    /*
562     * We need to look at the task state flags to figure out,
563     * whether the task is exiting. To protect against the do_exit
564     * change of the task flags, we do this protected by
565     * p->pi_lock:
566     */
567    raw_spin_lock_irq(&p->pi_lock);
568    if (unlikely(p->flags & PF_EXITING)) {
569        /*
570         * The task is on the way out. When PF_EXITPIDONE is
571         * set, we know that the task has finished the
572         * cleanup:
573         */
574        int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
575
576        raw_spin_unlock_irq(&p->pi_lock);
577        put_task_struct(p);
578        return ret;
579    }
580
581    pi_state = alloc_pi_state();
582
583    /*
584     * Initialize the pi_mutex in locked state and make 'p'
585     * the owner of it:
586     */
587    rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
588
589    /* Store the key for possible exit cleanups: */
590    pi_state->key = *key;
591
592    WARN_ON(!list_empty(&pi_state->list));
593    list_add(&pi_state->list, &p->pi_state_list);
594    pi_state->owner = p;
595    raw_spin_unlock_irq(&p->pi_lock);
596
597    put_task_struct(p);
598
599    *ps = pi_state;
600
601    return 0;
602}
603
604/**
605 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
606 * @uaddr: the pi futex user address
607 * @hb: the pi futex hash bucket
608 * @key: the futex key associated with uaddr and hb
609 * @ps: the pi_state pointer where we store the result of the
610 * lookup
611 * @task: the task to perform the atomic lock work for. This will
612 * be "current" except in the case of requeue pi.
613 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
614 *
615 * Returns:
616 * 0 - ready to wait
617 * 1 - acquired the lock
618 * <0 - error
619 *
620 * The hb->lock and futex_key refs shall be held by the caller.
621 */
622static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
623                union futex_key *key,
624                struct futex_pi_state **ps,
625                struct task_struct *task, int set_waiters)
626{
627    int lock_taken, ret, ownerdied = 0;
628    u32 uval, newval, curval;
629
630retry:
631    ret = lock_taken = 0;
632
633    /*
634     * To avoid races, we attempt to take the lock here again
635     * (by doing a 0 -> TID atomic cmpxchg), while holding all
636     * the locks. It will most likely not succeed.
637     */
638    newval = task_pid_vnr(task);
639    if (set_waiters)
640        newval |= FUTEX_WAITERS;
641
642    curval = cmpxchg_futex_value_locked(uaddr, 0, newval);
643
644    if (unlikely(curval == -EFAULT))
645        return -EFAULT;
646
647    /*
648     * Detect deadlocks.
649     */
650    if ((unlikely((curval & FUTEX_TID_MASK) == task_pid_vnr(task))))
651        return -EDEADLK;
652
653    /*
654     * Surprise - we got the lock. Just return to userspace:
655     */
656    if (unlikely(!curval))
657        return 1;
658
659    uval = curval;
660
661    /*
662     * Set the FUTEX_WAITERS flag, so the owner will know it has someone
663     * to wake at the next unlock.
664     */
665    newval = curval | FUTEX_WAITERS;
666
667    /*
668     * There are two cases, where a futex might have no owner (the
669     * owner TID is 0): OWNER_DIED. We take over the futex in this
670     * case. We also do an unconditional take over, when the owner
671     * of the futex died.
672     *
673     * This is safe as we are protected by the hash bucket lock !
674     */
675    if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
676        /* Keep the OWNER_DIED bit */
677        newval = (curval & ~FUTEX_TID_MASK) | task_pid_vnr(task);
678        ownerdied = 0;
679        lock_taken = 1;
680    }
681
682    curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
683
684    if (unlikely(curval == -EFAULT))
685        return -EFAULT;
686    if (unlikely(curval != uval))
687        goto retry;
688
689    /*
690     * We took the lock due to owner died take over.
691     */
692    if (unlikely(lock_taken))
693        return 1;
694
695    /*
696     * We dont have the lock. Look up the PI state (or create it if
697     * we are the first waiter):
698     */
699    ret = lookup_pi_state(uval, hb, key, ps);
700
701    if (unlikely(ret)) {
702        switch (ret) {
703        case -ESRCH:
704            /*
705             * No owner found for this futex. Check if the
706             * OWNER_DIED bit is set to figure out whether
707             * this is a robust futex or not.
708             */
709            if (get_futex_value_locked(&curval, uaddr))
710                return -EFAULT;
711
712            /*
713             * We simply start over in case of a robust
714             * futex. The code above will take the futex
715             * and return happy.
716             */
717            if (curval & FUTEX_OWNER_DIED) {
718                ownerdied = 1;
719                goto retry;
720            }
721        default:
722            break;
723        }
724    }
725
726    return ret;
727}
728
729/*
730 * The hash bucket lock must be held when this is called.
731 * Afterwards, the futex_q must not be accessed.
732 */
733static void wake_futex(struct futex_q *q)
734{
735    struct task_struct *p = q->task;
736
737    /*
738     * We set q->lock_ptr = NULL _before_ we wake up the task. If
739     * a non futex wake up happens on another CPU then the task
740     * might exit and p would dereference a non existing task
741     * struct. Prevent this by holding a reference on p across the
742     * wake up.
743     */
744    get_task_struct(p);
745
746    plist_del(&q->list, &q->list.plist);
747    /*
748     * The waiting task can free the futex_q as soon as
749     * q->lock_ptr = NULL is written, without taking any locks. A
750     * memory barrier is required here to prevent the following
751     * store to lock_ptr from getting ahead of the plist_del.
752     */
753    smp_wmb();
754    q->lock_ptr = NULL;
755
756    wake_up_state(p, TASK_NORMAL);
757    put_task_struct(p);
758}
759
760static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
761{
762    struct task_struct *new_owner;
763    struct futex_pi_state *pi_state = this->pi_state;
764    u32 curval, newval;
765
766    if (!pi_state)
767        return -EINVAL;
768
769    /*
770     * If current does not own the pi_state then the futex is
771     * inconsistent and user space fiddled with the futex value.
772     */
773    if (pi_state->owner != current)
774        return -EINVAL;
775
776    raw_spin_lock(&pi_state->pi_mutex.wait_lock);
777    new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
778
779    /*
780     * This happens when we have stolen the lock and the original
781     * pending owner did not enqueue itself back on the rt_mutex.
782     * Thats not a tragedy. We know that way, that a lock waiter
783     * is on the fly. We make the futex_q waiter the pending owner.
784     */
785    if (!new_owner)
786        new_owner = this->task;
787
788    /*
789     * We pass it to the next owner. (The WAITERS bit is always
790     * kept enabled while there is PI state around. We must also
791     * preserve the owner died bit.)
792     */
793    if (!(uval & FUTEX_OWNER_DIED)) {
794        int ret = 0;
795
796        newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
797
798        curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
799
800        if (curval == -EFAULT)
801            ret = -EFAULT;
802        else if (curval != uval)
803            ret = -EINVAL;
804        if (ret) {
805            raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
806            return ret;
807        }
808    }
809
810    raw_spin_lock_irq(&pi_state->owner->pi_lock);
811    WARN_ON(list_empty(&pi_state->list));
812    list_del_init(&pi_state->list);
813    raw_spin_unlock_irq(&pi_state->owner->pi_lock);
814
815    raw_spin_lock_irq(&new_owner->pi_lock);
816    WARN_ON(!list_empty(&pi_state->list));
817    list_add(&pi_state->list, &new_owner->pi_state_list);
818    pi_state->owner = new_owner;
819    raw_spin_unlock_irq(&new_owner->pi_lock);
820
821    raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
822    rt_mutex_unlock(&pi_state->pi_mutex);
823
824    return 0;
825}
826
827static int unlock_futex_pi(u32 __user *uaddr, u32 uval)
828{
829    u32 oldval;
830
831    /*
832     * There is no waiter, so we unlock the futex. The owner died
833     * bit has not to be preserved here. We are the owner:
834     */
835    oldval = cmpxchg_futex_value_locked(uaddr, uval, 0);
836
837    if (oldval == -EFAULT)
838        return oldval;
839    if (oldval != uval)
840        return -EAGAIN;
841
842    return 0;
843}
844
845/*
846 * Express the locking dependencies for lockdep:
847 */
848static inline void
849double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
850{
851    if (hb1 <= hb2) {
852        spin_lock(&hb1->lock);
853        if (hb1 < hb2)
854            spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
855    } else { /* hb1 > hb2 */
856        spin_lock(&hb2->lock);
857        spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
858    }
859}
860
861static inline void
862double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
863{
864    spin_unlock(&hb1->lock);
865    if (hb1 != hb2)
866        spin_unlock(&hb2->lock);
867}
868
869/*
870 * Wake up waiters matching bitset queued on this futex (uaddr).
871 */
872static int futex_wake(u32 __user *uaddr, int fshared, int nr_wake, u32 bitset)
873{
874    struct futex_hash_bucket *hb;
875    struct futex_q *this, *next;
876    struct plist_head *head;
877    union futex_key key = FUTEX_KEY_INIT;
878    int ret;
879
880    if (!bitset)
881        return -EINVAL;
882
883    ret = get_futex_key(uaddr, fshared, &key);
884    if (unlikely(ret != 0))
885        goto out;
886
887    hb = hash_futex(&key);
888    spin_lock(&hb->lock);
889    head = &hb->chain;
890
891    plist_for_each_entry_safe(this, next, head, list) {
892        if (match_futex (&this->key, &key)) {
893            if (this->pi_state || this->rt_waiter) {
894                ret = -EINVAL;
895                break;
896            }
897
898            /* Check if one of the bits is set in both bitsets */
899            if (!(this->bitset & bitset))
900                continue;
901
902            wake_futex(this);
903            if (++ret >= nr_wake)
904                break;
905        }
906    }
907
908    spin_unlock(&hb->lock);
909    put_futex_key(fshared, &key);
910out:
911    return ret;
912}
913
914/*
915 * Wake up all waiters hashed on the physical page that is mapped
916 * to this virtual address:
917 */
918static int
919futex_wake_op(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
920          int nr_wake, int nr_wake2, int op)
921{
922    union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
923    struct futex_hash_bucket *hb1, *hb2;
924    struct plist_head *head;
925    struct futex_q *this, *next;
926    int ret, op_ret;
927
928retry:
929    ret = get_futex_key(uaddr1, fshared, &key1);
930    if (unlikely(ret != 0))
931        goto out;
932    ret = get_futex_key(uaddr2, fshared, &key2);
933    if (unlikely(ret != 0))
934        goto out_put_key1;
935
936    hb1 = hash_futex(&key1);
937    hb2 = hash_futex(&key2);
938
939retry_private:
940    double_lock_hb(hb1, hb2);
941    op_ret = futex_atomic_op_inuser(op, uaddr2);
942    if (unlikely(op_ret < 0)) {
943
944        double_unlock_hb(hb1, hb2);
945
946#ifndef CONFIG_MMU
947        /*
948         * we don't get EFAULT from MMU faults if we don't have an MMU,
949         * but we might get them from range checking
950         */
951        ret = op_ret;
952        goto out_put_keys;
953#endif
954
955        if (unlikely(op_ret != -EFAULT)) {
956            ret = op_ret;
957            goto out_put_keys;
958        }
959
960        ret = fault_in_user_writeable(uaddr2);
961        if (ret)
962            goto out_put_keys;
963
964        if (!fshared)
965            goto retry_private;
966
967        put_futex_key(fshared, &key2);
968        put_futex_key(fshared, &key1);
969        goto retry;
970    }
971
972    head = &hb1->chain;
973
974    plist_for_each_entry_safe(this, next, head, list) {
975        if (match_futex (&this->key, &key1)) {
976            wake_futex(this);
977            if (++ret >= nr_wake)
978                break;
979        }
980    }
981
982    if (op_ret > 0) {
983        head = &hb2->chain;
984
985        op_ret = 0;
986        plist_for_each_entry_safe(this, next, head, list) {
987            if (match_futex (&this->key, &key2)) {
988                wake_futex(this);
989                if (++op_ret >= nr_wake2)
990                    break;
991            }
992        }
993        ret += op_ret;
994    }
995
996    double_unlock_hb(hb1, hb2);
997out_put_keys:
998    put_futex_key(fshared, &key2);
999out_put_key1:
1000    put_futex_key(fshared, &key1);
1001out:
1002    return ret;
1003}
1004
1005/**
1006 * requeue_futex() - Requeue a futex_q from one hb to another
1007 * @q: the futex_q to requeue
1008 * @hb1: the source hash_bucket
1009 * @hb2: the target hash_bucket
1010 * @key2: the new key for the requeued futex_q
1011 */
1012static inline
1013void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1014           struct futex_hash_bucket *hb2, union futex_key *key2)
1015{
1016
1017    /*
1018     * If key1 and key2 hash to the same bucket, no need to
1019     * requeue.
1020     */
1021    if (likely(&hb1->chain != &hb2->chain)) {
1022        plist_del(&q->list, &hb1->chain);
1023        plist_add(&q->list, &hb2->chain);
1024        q->lock_ptr = &hb2->lock;
1025#ifdef CONFIG_DEBUG_PI_LIST
1026        q->list.plist.spinlock = &hb2->lock;
1027#endif
1028    }
1029    get_futex_key_refs(key2);
1030    q->key = *key2;
1031}
1032
1033/**
1034 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1035 * @q: the futex_q
1036 * @key: the key of the requeue target futex
1037 * @hb: the hash_bucket of the requeue target futex
1038 *
1039 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1040 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1041 * to the requeue target futex so the waiter can detect the wakeup on the right
1042 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1043 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1044 * to protect access to the pi_state to fixup the owner later. Must be called
1045 * with both q->lock_ptr and hb->lock held.
1046 */
1047static inline
1048void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1049               struct futex_hash_bucket *hb)
1050{
1051    get_futex_key_refs(key);
1052    q->key = *key;
1053
1054    WARN_ON(plist_node_empty(&q->list));
1055    plist_del(&q->list, &q->list.plist);
1056
1057    WARN_ON(!q->rt_waiter);
1058    q->rt_waiter = NULL;
1059
1060    q->lock_ptr = &hb->lock;
1061#ifdef CONFIG_DEBUG_PI_LIST
1062    q->list.plist.spinlock = &hb->lock;
1063#endif
1064
1065    wake_up_state(q->task, TASK_NORMAL);
1066}
1067
1068/**
1069 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1070 * @pifutex: the user address of the to futex
1071 * @hb1: the from futex hash bucket, must be locked by the caller
1072 * @hb2: the to futex hash bucket, must be locked by the caller
1073 * @key1: the from futex key
1074 * @key2: the to futex key
1075 * @ps: address to store the pi_state pointer
1076 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1077 *
1078 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1079 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1080 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1081 * hb1 and hb2 must be held by the caller.
1082 *
1083 * Returns:
1084 * 0 - failed to acquire the lock atomicly
1085 * 1 - acquired the lock
1086 * <0 - error
1087 */
1088static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1089                 struct futex_hash_bucket *hb1,
1090                 struct futex_hash_bucket *hb2,
1091                 union futex_key *key1, union futex_key *key2,
1092                 struct futex_pi_state **ps, int set_waiters)
1093{
1094    struct futex_q *top_waiter = NULL;
1095    u32 curval;
1096    int ret;
1097
1098    if (get_futex_value_locked(&curval, pifutex))
1099        return -EFAULT;
1100
1101    /*
1102     * Find the top_waiter and determine if there are additional waiters.
1103     * If the caller intends to requeue more than 1 waiter to pifutex,
1104     * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1105     * as we have means to handle the possible fault. If not, don't set
1106     * the bit unecessarily as it will force the subsequent unlock to enter
1107     * the kernel.
1108     */
1109    top_waiter = futex_top_waiter(hb1, key1);
1110
1111    /* There are no waiters, nothing for us to do. */
1112    if (!top_waiter)
1113        return 0;
1114
1115    /* Ensure we requeue to the expected futex. */
1116    if (!match_futex(top_waiter->requeue_pi_key, key2))
1117        return -EINVAL;
1118
1119    /*
1120     * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1121     * the contended case or if set_waiters is 1. The pi_state is returned
1122     * in ps in contended cases.
1123     */
1124    ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1125                   set_waiters);
1126    if (ret == 1)
1127        requeue_pi_wake_futex(top_waiter, key2, hb2);
1128
1129    return ret;
1130}
1131
1132/**
1133 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1134 * uaddr1: source futex user address
1135 * uaddr2: target futex user address
1136 * nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1137 * nr_requeue: number of waiters to requeue (0-INT_MAX)
1138 * requeue_pi: if we are attempting to requeue from a non-pi futex to a
1139 * pi futex (pi to pi requeue is not supported)
1140 *
1141 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1142 * uaddr2 atomically on behalf of the top waiter.
1143 *
1144 * Returns:
1145 * >=0 - on success, the number of tasks requeued or woken
1146 * <0 - on error
1147 */
1148static int futex_requeue(u32 __user *uaddr1, int fshared, u32 __user *uaddr2,
1149             int nr_wake, int nr_requeue, u32 *cmpval,
1150             int requeue_pi)
1151{
1152    union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1153    int drop_count = 0, task_count = 0, ret;
1154    struct futex_pi_state *pi_state = NULL;
1155    struct futex_hash_bucket *hb1, *hb2;
1156    struct plist_head *head1;
1157    struct futex_q *this, *next;
1158    u32 curval2;
1159
1160    if (requeue_pi) {
1161        /*
1162         * requeue_pi requires a pi_state, try to allocate it now
1163         * without any locks in case it fails.
1164         */
1165        if (refill_pi_state_cache())
1166            return -ENOMEM;
1167        /*
1168         * requeue_pi must wake as many tasks as it can, up to nr_wake
1169         * + nr_requeue, since it acquires the rt_mutex prior to
1170         * returning to userspace, so as to not leave the rt_mutex with
1171         * waiters and no owner. However, second and third wake-ups
1172         * cannot be predicted as they involve race conditions with the
1173         * first wake and a fault while looking up the pi_state. Both
1174         * pthread_cond_signal() and pthread_cond_broadcast() should
1175         * use nr_wake=1.
1176         */
1177        if (nr_wake != 1)
1178            return -EINVAL;
1179    }
1180
1181retry:
1182    if (pi_state != NULL) {
1183        /*
1184         * We will have to lookup the pi_state again, so free this one
1185         * to keep the accounting correct.
1186         */
1187        free_pi_state(pi_state);
1188        pi_state = NULL;
1189    }
1190
1191    ret = get_futex_key(uaddr1, fshared, &key1);
1192    if (unlikely(ret != 0))
1193        goto out;
1194    ret = get_futex_key(uaddr2, fshared, &key2);
1195    if (unlikely(ret != 0))
1196        goto out_put_key1;
1197
1198    hb1 = hash_futex(&key1);
1199    hb2 = hash_futex(&key2);
1200
1201retry_private:
1202    double_lock_hb(hb1, hb2);
1203
1204    if (likely(cmpval != NULL)) {
1205        u32 curval;
1206
1207        ret = get_futex_value_locked(&curval, uaddr1);
1208
1209        if (unlikely(ret)) {
1210            double_unlock_hb(hb1, hb2);
1211
1212            ret = get_user(curval, uaddr1);
1213            if (ret)
1214                goto out_put_keys;
1215
1216            if (!fshared)
1217                goto retry_private;
1218
1219            put_futex_key(fshared, &key2);
1220            put_futex_key(fshared, &key1);
1221            goto retry;
1222        }
1223        if (curval != *cmpval) {
1224            ret = -EAGAIN;
1225            goto out_unlock;
1226        }
1227    }
1228
1229    if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1230        /*
1231         * Attempt to acquire uaddr2 and wake the top waiter. If we
1232         * intend to requeue waiters, force setting the FUTEX_WAITERS
1233         * bit. We force this here where we are able to easily handle
1234         * faults rather in the requeue loop below.
1235         */
1236        ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1237                         &key2, &pi_state, nr_requeue);
1238
1239        /*
1240         * At this point the top_waiter has either taken uaddr2 or is
1241         * waiting on it. If the former, then the pi_state will not
1242         * exist yet, look it up one more time to ensure we have a
1243         * reference to it.
1244         */
1245        if (ret == 1) {
1246            WARN_ON(pi_state);
1247            drop_count++;
1248            task_count++;
1249            ret = get_futex_value_locked(&curval2, uaddr2);
1250            if (!ret)
1251                ret = lookup_pi_state(curval2, hb2, &key2,
1252                              &pi_state);
1253        }
1254
1255        switch (ret) {
1256        case 0:
1257            break;
1258        case -EFAULT:
1259            double_unlock_hb(hb1, hb2);
1260            put_futex_key(fshared, &key2);
1261            put_futex_key(fshared, &key1);
1262            ret = fault_in_user_writeable(uaddr2);
1263            if (!ret)
1264                goto retry;
1265            goto out;
1266        case -EAGAIN:
1267            /* The owner was exiting, try again. */
1268            double_unlock_hb(hb1, hb2);
1269            put_futex_key(fshared, &key2);
1270            put_futex_key(fshared, &key1);
1271            cond_resched();
1272            goto retry;
1273        default:
1274            goto out_unlock;
1275        }
1276    }
1277
1278    head1 = &hb1->chain;
1279    plist_for_each_entry_safe(this, next, head1, list) {
1280        if (task_count - nr_wake >= nr_requeue)
1281            break;
1282
1283        if (!match_futex(&this->key, &key1))
1284            continue;
1285
1286        /*
1287         * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1288         * be paired with each other and no other futex ops.
1289         */
1290        if ((requeue_pi && !this->rt_waiter) ||
1291            (!requeue_pi && this->rt_waiter)) {
1292            ret = -EINVAL;
1293            break;
1294        }
1295
1296        /*
1297         * Wake nr_wake waiters. For requeue_pi, if we acquired the
1298         * lock, we already woke the top_waiter. If not, it will be
1299         * woken by futex_unlock_pi().
1300         */
1301        if (++task_count <= nr_wake && !requeue_pi) {
1302            wake_futex(this);
1303            continue;
1304        }
1305
1306        /* Ensure we requeue to the expected futex for requeue_pi. */
1307        if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1308            ret = -EINVAL;
1309            break;
1310        }
1311
1312        /*
1313         * Requeue nr_requeue waiters and possibly one more in the case
1314         * of requeue_pi if we couldn't acquire the lock atomically.
1315         */
1316        if (requeue_pi) {
1317            /* Prepare the waiter to take the rt_mutex. */
1318            atomic_inc(&pi_state->refcount);
1319            this->pi_state = pi_state;
1320            ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1321                            this->rt_waiter,
1322                            this->task, 1);
1323            if (ret == 1) {
1324                /* We got the lock. */
1325                requeue_pi_wake_futex(this, &key2, hb2);
1326                drop_count++;
1327                continue;
1328            } else if (ret) {
1329                /* -EDEADLK */
1330                this->pi_state = NULL;
1331                free_pi_state(pi_state);
1332                goto out_unlock;
1333            }
1334        }
1335        requeue_futex(this, hb1, hb2, &key2);
1336        drop_count++;
1337    }
1338
1339out_unlock:
1340    double_unlock_hb(hb1, hb2);
1341
1342    /*
1343     * drop_futex_key_refs() must be called outside the spinlocks. During
1344     * the requeue we moved futex_q's from the hash bucket at key1 to the
1345     * one at key2 and updated their key pointer. We no longer need to
1346     * hold the references to key1.
1347     */
1348    while (--drop_count >= 0)
1349        drop_futex_key_refs(&key1);
1350
1351out_put_keys:
1352    put_futex_key(fshared, &key2);
1353out_put_key1:
1354    put_futex_key(fshared, &key1);
1355out:
1356    if (pi_state != NULL)
1357        free_pi_state(pi_state);
1358    return ret ? ret : task_count;
1359}
1360
1361/* The key must be already stored in q->key. */
1362static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1363{
1364    struct futex_hash_bucket *hb;
1365
1366    get_futex_key_refs(&q->key);
1367    hb = hash_futex(&q->key);
1368    q->lock_ptr = &hb->lock;
1369
1370    spin_lock(&hb->lock);
1371    return hb;
1372}
1373
1374static inline void
1375queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb)
1376{
1377    spin_unlock(&hb->lock);
1378    drop_futex_key_refs(&q->key);
1379}
1380
1381/**
1382 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1383 * @q: The futex_q to enqueue
1384 * @hb: The destination hash bucket
1385 *
1386 * The hb->lock must be held by the caller, and is released here. A call to
1387 * queue_me() is typically paired with exactly one call to unqueue_me(). The
1388 * exceptions involve the PI related operations, which may use unqueue_me_pi()
1389 * or nothing if the unqueue is done as part of the wake process and the unqueue
1390 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1391 * an example).
1392 */
1393static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1394{
1395    int prio;
1396
1397    /*
1398     * The priority used to register this element is
1399     * - either the real thread-priority for the real-time threads
1400     * (i.e. threads with a priority lower than MAX_RT_PRIO)
1401     * - or MAX_RT_PRIO for non-RT threads.
1402     * Thus, all RT-threads are woken first in priority order, and
1403     * the others are woken last, in FIFO order.
1404     */
1405    prio = min(current->normal_prio, MAX_RT_PRIO);
1406
1407    plist_node_init(&q->list, prio);
1408#ifdef CONFIG_DEBUG_PI_LIST
1409    q->list.plist.spinlock = &hb->lock;
1410#endif
1411    plist_add(&q->list, &hb->chain);
1412    q->task = current;
1413    spin_unlock(&hb->lock);
1414}
1415
1416/**
1417 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1418 * @q: The futex_q to unqueue
1419 *
1420 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1421 * be paired with exactly one earlier call to queue_me().
1422 *
1423 * Returns:
1424 * 1 - if the futex_q was still queued (and we removed unqueued it)
1425 * 0 - if the futex_q was already removed by the waking thread
1426 */
1427static int unqueue_me(struct futex_q *q)
1428{
1429    spinlock_t *lock_ptr;
1430    int ret = 0;
1431
1432    /* In the common case we don't take the spinlock, which is nice. */
1433retry:
1434    lock_ptr = q->lock_ptr;
1435    barrier();
1436    if (lock_ptr != NULL) {
1437        spin_lock(lock_ptr);
1438        /*
1439         * q->lock_ptr can change between reading it and
1440         * spin_lock(), causing us to take the wrong lock. This
1441         * corrects the race condition.
1442         *
1443         * Reasoning goes like this: if we have the wrong lock,
1444         * q->lock_ptr must have changed (maybe several times)
1445         * between reading it and the spin_lock(). It can
1446         * change again after the spin_lock() but only if it was
1447         * already changed before the spin_lock(). It cannot,
1448         * however, change back to the original value. Therefore
1449         * we can detect whether we acquired the correct lock.
1450         */
1451        if (unlikely(lock_ptr != q->lock_ptr)) {
1452            spin_unlock(lock_ptr);
1453            goto retry;
1454        }
1455        WARN_ON(plist_node_empty(&q->list));
1456        plist_del(&q->list, &q->list.plist);
1457
1458        BUG_ON(q->pi_state);
1459
1460        spin_unlock(lock_ptr);
1461        ret = 1;
1462    }
1463
1464    drop_futex_key_refs(&q->key);
1465    return ret;
1466}
1467
1468/*
1469 * PI futexes can not be requeued and must remove themself from the
1470 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1471 * and dropped here.
1472 */
1473static void unqueue_me_pi(struct futex_q *q)
1474{
1475    WARN_ON(plist_node_empty(&q->list));
1476    plist_del(&q->list, &q->list.plist);
1477
1478    BUG_ON(!q->pi_state);
1479    free_pi_state(q->pi_state);
1480    q->pi_state = NULL;
1481
1482    spin_unlock(q->lock_ptr);
1483
1484    drop_futex_key_refs(&q->key);
1485}
1486
1487/*
1488 * Fixup the pi_state owner with the new owner.
1489 *
1490 * Must be called with hash bucket lock held and mm->sem held for non
1491 * private futexes.
1492 */
1493static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1494                struct task_struct *newowner, int fshared)
1495{
1496    u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1497    struct futex_pi_state *pi_state = q->pi_state;
1498    struct task_struct *oldowner = pi_state->owner;
1499    u32 uval, curval, newval;
1500    int ret;
1501
1502    /* Owner died? */
1503    if (!pi_state->owner)
1504        newtid |= FUTEX_OWNER_DIED;
1505
1506    /*
1507     * We are here either because we stole the rtmutex from the
1508     * pending owner or we are the pending owner which failed to
1509     * get the rtmutex. We have to replace the pending owner TID
1510     * in the user space variable. This must be atomic as we have
1511     * to preserve the owner died bit here.
1512     *
1513     * Note: We write the user space value _before_ changing the pi_state
1514     * because we can fault here. Imagine swapped out pages or a fork
1515     * that marked all the anonymous memory readonly for cow.
1516     *
1517     * Modifying pi_state _before_ the user space value would
1518     * leave the pi_state in an inconsistent state when we fault
1519     * here, because we need to drop the hash bucket lock to
1520     * handle the fault. This might be observed in the PID check
1521     * in lookup_pi_state.
1522     */
1523retry:
1524    if (get_futex_value_locked(&uval, uaddr))
1525        goto handle_fault;
1526
1527    while (1) {
1528        newval = (uval & FUTEX_OWNER_DIED) | newtid;
1529
1530        curval = cmpxchg_futex_value_locked(uaddr, uval, newval);
1531
1532        if (curval == -EFAULT)
1533            goto handle_fault;
1534        if (curval == uval)
1535            break;
1536        uval = curval;
1537    }
1538
1539    /*
1540     * We fixed up user space. Now we need to fix the pi_state
1541     * itself.
1542     */
1543    if (pi_state->owner != NULL) {
1544        raw_spin_lock_irq(&pi_state->owner->pi_lock);
1545        WARN_ON(list_empty(&pi_state->list));
1546        list_del_init(&pi_state->list);
1547        raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1548    }
1549
1550    pi_state->owner = newowner;
1551
1552    raw_spin_lock_irq(&newowner->pi_lock);
1553    WARN_ON(!list_empty(&pi_state->list));
1554    list_add(&pi_state->list, &newowner->pi_state_list);
1555    raw_spin_unlock_irq(&newowner->pi_lock);
1556    return 0;
1557
1558    /*
1559     * To handle the page fault we need to drop the hash bucket
1560     * lock here. That gives the other task (either the pending
1561     * owner itself or the task which stole the rtmutex) the
1562     * chance to try the fixup of the pi_state. So once we are
1563     * back from handling the fault we need to check the pi_state
1564     * after reacquiring the hash bucket lock and before trying to
1565     * do another fixup. When the fixup has been done already we
1566     * simply return.
1567     */
1568handle_fault:
1569    spin_unlock(q->lock_ptr);
1570
1571    ret = fault_in_user_writeable(uaddr);
1572
1573    spin_lock(q->lock_ptr);
1574
1575    /*
1576     * Check if someone else fixed it for us:
1577     */
1578    if (pi_state->owner != oldowner)
1579        return 0;
1580
1581    if (ret)
1582        return ret;
1583
1584    goto retry;
1585}
1586
1587/*
1588 * In case we must use restart_block to restart a futex_wait,
1589 * we encode in the 'flags' shared capability
1590 */
1591#define FLAGS_SHARED 0x01
1592#define FLAGS_CLOCKRT 0x02
1593#define FLAGS_HAS_TIMEOUT 0x04
1594
1595static long futex_wait_restart(struct restart_block *restart);
1596
1597/**
1598 * fixup_owner() - Post lock pi_state and corner case management
1599 * @uaddr: user address of the futex
1600 * @fshared: whether the futex is shared (1) or not (0)
1601 * @q: futex_q (contains pi_state and access to the rt_mutex)
1602 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
1603 *
1604 * After attempting to lock an rt_mutex, this function is called to cleanup
1605 * the pi_state owner as well as handle race conditions that may allow us to
1606 * acquire the lock. Must be called with the hb lock held.
1607 *
1608 * Returns:
1609 * 1 - success, lock taken
1610 * 0 - success, lock not taken
1611 * <0 - on error (-EFAULT)
1612 */
1613static int fixup_owner(u32 __user *uaddr, int fshared, struct futex_q *q,
1614               int locked)
1615{
1616    struct task_struct *owner;
1617    int ret = 0;
1618
1619    if (locked) {
1620        /*
1621         * Got the lock. We might not be the anticipated owner if we
1622         * did a lock-steal - fix up the PI-state in that case:
1623         */
1624        if (q->pi_state->owner != current)
1625            ret = fixup_pi_state_owner(uaddr, q, current, fshared);
1626        goto out;
1627    }
1628
1629    /*
1630     * Catch the rare case, where the lock was released when we were on the
1631     * way back before we locked the hash bucket.
1632     */
1633    if (q->pi_state->owner == current) {
1634        /*
1635         * Try to get the rt_mutex now. This might fail as some other
1636         * task acquired the rt_mutex after we removed ourself from the
1637         * rt_mutex waiters list.
1638         */
1639        if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
1640            locked = 1;
1641            goto out;
1642        }
1643
1644        /*
1645         * pi_state is incorrect, some other task did a lock steal and
1646         * we returned due to timeout or signal without taking the
1647         * rt_mutex. Too late. We can access the rt_mutex_owner without
1648         * locking, as the other task is now blocked on the hash bucket
1649         * lock. Fix the state up.
1650         */
1651        owner = rt_mutex_owner(&q->pi_state->pi_mutex);
1652        ret = fixup_pi_state_owner(uaddr, q, owner, fshared);
1653        goto out;
1654    }
1655
1656    /*
1657     * Paranoia check. If we did not take the lock, then we should not be
1658     * the owner, nor the pending owner, of the rt_mutex.
1659     */
1660    if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
1661        printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
1662                "pi-state %p\n", ret,
1663                q->pi_state->pi_mutex.owner,
1664                q->pi_state->owner);
1665
1666out:
1667    return ret ? ret : locked;
1668}
1669
1670/**
1671 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
1672 * @hb: the futex hash bucket, must be locked by the caller
1673 * @q: the futex_q to queue up on
1674 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
1675 */
1676static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
1677                struct hrtimer_sleeper *timeout)
1678{
1679    /*
1680     * The task state is guaranteed to be set before another task can
1681     * wake it. set_current_state() is implemented using set_mb() and
1682     * queue_me() calls spin_unlock() upon completion, both serializing
1683     * access to the hash list and forcing another memory barrier.
1684     */
1685    set_current_state(TASK_INTERRUPTIBLE);
1686    queue_me(q, hb);
1687
1688    /* Arm the timer */
1689    if (timeout) {
1690        hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
1691        if (!hrtimer_active(&timeout->timer))
1692            timeout->task = NULL;
1693    }
1694
1695    /*
1696     * If we have been removed from the hash list, then another task
1697     * has tried to wake us, and we can skip the call to schedule().
1698     */
1699    if (likely(!plist_node_empty(&q->list))) {
1700        /*
1701         * If the timer has already expired, current will already be
1702         * flagged for rescheduling. Only call schedule if there
1703         * is no timeout, or if it has yet to expire.
1704         */
1705        if (!timeout || timeout->task)
1706            schedule();
1707    }
1708    __set_current_state(TASK_RUNNING);
1709}
1710
1711/**
1712 * futex_wait_setup() - Prepare to wait on a futex
1713 * @uaddr: the futex userspace address
1714 * @val: the expected value
1715 * @fshared: whether the futex is shared (1) or not (0)
1716 * @q: the associated futex_q
1717 * @hb: storage for hash_bucket pointer to be returned to caller
1718 *
1719 * Setup the futex_q and locate the hash_bucket. Get the futex value and
1720 * compare it with the expected value. Handle atomic faults internally.
1721 * Return with the hb lock held and a q.key reference on success, and unlocked
1722 * with no q.key reference on failure.
1723 *
1724 * Returns:
1725 * 0 - uaddr contains val and hb has been locked
1726 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlcoked
1727 */
1728static int futex_wait_setup(u32 __user *uaddr, u32 val, int fshared,
1729               struct futex_q *q, struct futex_hash_bucket **hb)
1730{
1731    u32 uval;
1732    int ret;
1733
1734    /*
1735     * Access the page AFTER the hash-bucket is locked.
1736     * Order is important:
1737     *
1738     * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
1739     * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
1740     *
1741     * The basic logical guarantee of a futex is that it blocks ONLY
1742     * if cond(var) is known to be true at the time of blocking, for
1743     * any cond. If we queued after testing *uaddr, that would open
1744     * a race condition where we could block indefinitely with
1745     * cond(var) false, which would violate the guarantee.
1746     *
1747     * A consequence is that futex_wait() can return zero and absorb
1748     * a wakeup when *uaddr != val on entry to the syscall. This is
1749     * rare, but normal.
1750     */
1751retry:
1752    q->key = FUTEX_KEY_INIT;
1753    ret = get_futex_key(uaddr, fshared, &q->key);
1754    if (unlikely(ret != 0))
1755        return ret;
1756
1757retry_private:
1758    *hb = queue_lock(q);
1759
1760    ret = get_futex_value_locked(&uval, uaddr);
1761
1762    if (ret) {
1763        queue_unlock(q, *hb);
1764
1765        ret = get_user(uval, uaddr);
1766        if (ret)
1767            goto out;
1768
1769        if (!fshared)
1770            goto retry_private;
1771
1772        put_futex_key(fshared, &q->key);
1773        goto retry;
1774    }
1775
1776    if (uval != val) {
1777        queue_unlock(q, *hb);
1778        ret = -EWOULDBLOCK;
1779    }
1780
1781out:
1782    if (ret)
1783        put_futex_key(fshared, &q->key);
1784    return ret;
1785}
1786
1787static int futex_wait(u32 __user *uaddr, int fshared,
1788              u32 val, ktime_t *abs_time, u32 bitset, int clockrt)
1789{
1790    struct hrtimer_sleeper timeout, *to = NULL;
1791    struct restart_block *restart;
1792    struct futex_hash_bucket *hb;
1793    struct futex_q q;
1794    int ret;
1795
1796    if (!bitset)
1797        return -EINVAL;
1798
1799    q.pi_state = NULL;
1800    q.bitset = bitset;
1801    q.rt_waiter = NULL;
1802    q.requeue_pi_key = NULL;
1803
1804    if (abs_time) {
1805        to = &timeout;
1806
1807        hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
1808                      CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1809        hrtimer_init_sleeper(to, current);
1810        hrtimer_set_expires_range_ns(&to->timer, *abs_time,
1811                         current->timer_slack_ns);
1812    }
1813
1814retry:
1815    /* Prepare to wait on uaddr. */
1816    ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
1817    if (ret)
1818        goto out;
1819
1820    /* queue_me and wait for wakeup, timeout, or a signal. */
1821    futex_wait_queue_me(hb, &q, to);
1822
1823    /* If we were woken (and unqueued), we succeeded, whatever. */
1824    ret = 0;
1825    if (!unqueue_me(&q))
1826        goto out_put_key;
1827    ret = -ETIMEDOUT;
1828    if (to && !to->task)
1829        goto out_put_key;
1830
1831    /*
1832     * We expect signal_pending(current), but we might be the
1833     * victim of a spurious wakeup as well.
1834     */
1835    if (!signal_pending(current)) {
1836        put_futex_key(fshared, &q.key);
1837        goto retry;
1838    }
1839
1840    ret = -ERESTARTSYS;
1841    if (!abs_time)
1842        goto out_put_key;
1843
1844    restart = &current_thread_info()->restart_block;
1845    restart->fn = futex_wait_restart;
1846    restart->futex.uaddr = (u32 *)uaddr;
1847    restart->futex.val = val;
1848    restart->futex.time = abs_time->tv64;
1849    restart->futex.bitset = bitset;
1850    restart->futex.flags = FLAGS_HAS_TIMEOUT;
1851
1852    if (fshared)
1853        restart->futex.flags |= FLAGS_SHARED;
1854    if (clockrt)
1855        restart->futex.flags |= FLAGS_CLOCKRT;
1856
1857    ret = -ERESTART_RESTARTBLOCK;
1858
1859out_put_key:
1860    put_futex_key(fshared, &q.key);
1861out:
1862    if (to) {
1863        hrtimer_cancel(&to->timer);
1864        destroy_hrtimer_on_stack(&to->timer);
1865    }
1866    return ret;
1867}
1868
1869
1870static long futex_wait_restart(struct restart_block *restart)
1871{
1872    u32 __user *uaddr = (u32 __user *)restart->futex.uaddr;
1873    int fshared = 0;
1874    ktime_t t, *tp = NULL;
1875
1876    if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
1877        t.tv64 = restart->futex.time;
1878        tp = &t;
1879    }
1880    restart->fn = do_no_restart_syscall;
1881    if (restart->futex.flags & FLAGS_SHARED)
1882        fshared = 1;
1883    return (long)futex_wait(uaddr, fshared, restart->futex.val, tp,
1884                restart->futex.bitset,
1885                restart->futex.flags & FLAGS_CLOCKRT);
1886}
1887
1888
1889/*
1890 * Userspace tried a 0 -> TID atomic transition of the futex value
1891 * and failed. The kernel side here does the whole locking operation:
1892 * if there are waiters then it will block, it does PI, etc. (Due to
1893 * races the kernel might see a 0 value of the futex too.)
1894 */
1895static int futex_lock_pi(u32 __user *uaddr, int fshared,
1896             int detect, ktime_t *time, int trylock)
1897{
1898    struct hrtimer_sleeper timeout, *to = NULL;
1899    struct futex_hash_bucket *hb;
1900    struct futex_q q;
1901    int res, ret;
1902
1903    if (refill_pi_state_cache())
1904        return -ENOMEM;
1905
1906    if (time) {
1907        to = &timeout;
1908        hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
1909                      HRTIMER_MODE_ABS);
1910        hrtimer_init_sleeper(to, current);
1911        hrtimer_set_expires(&to->timer, *time);
1912    }
1913
1914    q.pi_state = NULL;
1915    q.rt_waiter = NULL;
1916    q.requeue_pi_key = NULL;
1917retry:
1918    q.key = FUTEX_KEY_INIT;
1919    ret = get_futex_key(uaddr, fshared, &q.key);
1920    if (unlikely(ret != 0))
1921        goto out;
1922
1923retry_private:
1924    hb = queue_lock(&q);
1925
1926    ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
1927    if (unlikely(ret)) {
1928        switch (ret) {
1929        case 1:
1930            /* We got the lock. */
1931            ret = 0;
1932            goto out_unlock_put_key;
1933        case -EFAULT:
1934            goto uaddr_faulted;
1935        case -EAGAIN:
1936            /*
1937             * Task is exiting and we just wait for the
1938             * exit to complete.
1939             */
1940            queue_unlock(&q, hb);
1941            put_futex_key(fshared, &q.key);
1942            cond_resched();
1943            goto retry;
1944        default:
1945            goto out_unlock_put_key;
1946        }
1947    }
1948
1949    /*
1950     * Only actually queue now that the atomic ops are done:
1951     */
1952    queue_me(&q, hb);
1953
1954    WARN_ON(!q.pi_state);
1955    /*
1956     * Block on the PI mutex:
1957     */
1958    if (!trylock)
1959        ret = rt_mutex_timed_lock(&q.pi_state->pi_mutex, to, 1);
1960    else {
1961        ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
1962        /* Fixup the trylock return value: */
1963        ret = ret ? 0 : -EWOULDBLOCK;
1964    }
1965
1966    spin_lock(q.lock_ptr);
1967    /*
1968     * Fixup the pi_state owner and possibly acquire the lock if we
1969     * haven't already.
1970     */
1971    res = fixup_owner(uaddr, fshared, &q, !ret);
1972    /*
1973     * If fixup_owner() returned an error, proprogate that. If it acquired
1974     * the lock, clear our -ETIMEDOUT or -EINTR.
1975     */
1976    if (res)
1977        ret = (res < 0) ? res : 0;
1978
1979    /*
1980     * If fixup_owner() faulted and was unable to handle the fault, unlock
1981     * it and return the fault to userspace.
1982     */
1983    if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
1984        rt_mutex_unlock(&q.pi_state->pi_mutex);
1985
1986    /* Unqueue and drop the lock */
1987    unqueue_me_pi(&q);
1988
1989    goto out_put_key;
1990
1991out_unlock_put_key:
1992    queue_unlock(&q, hb);
1993
1994out_put_key:
1995    put_futex_key(fshared, &q.key);
1996out:
1997    if (to)
1998        destroy_hrtimer_on_stack(&to->timer);
1999    return ret != -EINTR ? ret : -ERESTARTNOINTR;
2000
2001uaddr_faulted:
2002    queue_unlock(&q, hb);
2003
2004    ret = fault_in_user_writeable(uaddr);
2005    if (ret)
2006        goto out_put_key;
2007
2008    if (!fshared)
2009        goto retry_private;
2010
2011    put_futex_key(fshared, &q.key);
2012    goto retry;
2013}
2014
2015/*
2016 * Userspace attempted a TID -> 0 atomic transition, and failed.
2017 * This is the in-kernel slowpath: we look up the PI state (if any),
2018 * and do the rt-mutex unlock.
2019 */
2020static int futex_unlock_pi(u32 __user *uaddr, int fshared)
2021{
2022    struct futex_hash_bucket *hb;
2023    struct futex_q *this, *next;
2024    u32 uval;
2025    struct plist_head *head;
2026    union futex_key key = FUTEX_KEY_INIT;
2027    int ret;
2028
2029retry:
2030    if (get_user(uval, uaddr))
2031        return -EFAULT;
2032    /*
2033     * We release only a lock we actually own:
2034     */
2035    if ((uval & FUTEX_TID_MASK) != task_pid_vnr(current))
2036        return -EPERM;
2037
2038    ret = get_futex_key(uaddr, fshared, &key);
2039    if (unlikely(ret != 0))
2040        goto out;
2041
2042    hb = hash_futex(&key);
2043    spin_lock(&hb->lock);
2044
2045    /*
2046     * To avoid races, try to do the TID -> 0 atomic transition
2047     * again. If it succeeds then we can return without waking
2048     * anyone else up:
2049     */
2050    if (!(uval & FUTEX_OWNER_DIED))
2051        uval = cmpxchg_futex_value_locked(uaddr, task_pid_vnr(current), 0);
2052
2053
2054    if (unlikely(uval == -EFAULT))
2055        goto pi_faulted;
2056    /*
2057     * Rare case: we managed to release the lock atomically,
2058     * no need to wake anyone else up:
2059     */
2060    if (unlikely(uval == task_pid_vnr(current)))
2061        goto out_unlock;
2062
2063    /*
2064     * Ok, other tasks may need to be woken up - check waiters
2065     * and do the wakeup if necessary:
2066     */
2067    head = &hb->chain;
2068
2069    plist_for_each_entry_safe(this, next, head, list) {
2070        if (!match_futex (&this->key, &key))
2071            continue;
2072        ret = wake_futex_pi(uaddr, uval, this);
2073        /*
2074         * The atomic access to the futex value
2075         * generated a pagefault, so retry the
2076         * user-access and the wakeup:
2077         */
2078        if (ret == -EFAULT)
2079            goto pi_faulted;
2080        goto out_unlock;
2081    }
2082    /*
2083     * No waiters - kernel unlocks the futex:
2084     */
2085    if (!(uval & FUTEX_OWNER_DIED)) {
2086        ret = unlock_futex_pi(uaddr, uval);
2087        if (ret == -EFAULT)
2088            goto pi_faulted;
2089    }
2090
2091out_unlock:
2092    spin_unlock(&hb->lock);
2093    put_futex_key(fshared, &key);
2094
2095out:
2096    return ret;
2097
2098pi_faulted:
2099    spin_unlock(&hb->lock);
2100    put_futex_key(fshared, &key);
2101
2102    ret = fault_in_user_writeable(uaddr);
2103    if (!ret)
2104        goto retry;
2105
2106    return ret;
2107}
2108
2109/**
2110 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2111 * @hb: the hash_bucket futex_q was original enqueued on
2112 * @q: the futex_q woken while waiting to be requeued
2113 * @key2: the futex_key of the requeue target futex
2114 * @timeout: the timeout associated with the wait (NULL if none)
2115 *
2116 * Detect if the task was woken on the initial futex as opposed to the requeue
2117 * target futex. If so, determine if it was a timeout or a signal that caused
2118 * the wakeup and return the appropriate error code to the caller. Must be
2119 * called with the hb lock held.
2120 *
2121 * Returns
2122 * 0 - no early wakeup detected
2123 * <0 - -ETIMEDOUT or -ERESTARTNOINTR
2124 */
2125static inline
2126int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2127                   struct futex_q *q, union futex_key *key2,
2128                   struct hrtimer_sleeper *timeout)
2129{
2130    int ret = 0;
2131
2132    /*
2133     * With the hb lock held, we avoid races while we process the wakeup.
2134     * We only need to hold hb (and not hb2) to ensure atomicity as the
2135     * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2136     * It can't be requeued from uaddr2 to something else since we don't
2137     * support a PI aware source futex for requeue.
2138     */
2139    if (!match_futex(&q->key, key2)) {
2140        WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2141        /*
2142         * We were woken prior to requeue by a timeout or a signal.
2143         * Unqueue the futex_q and determine which it was.
2144         */
2145        plist_del(&q->list, &q->list.plist);
2146
2147        /* Handle spurious wakeups gracefully */
2148        ret = -EWOULDBLOCK;
2149        if (timeout && !timeout->task)
2150            ret = -ETIMEDOUT;
2151        else if (signal_pending(current))
2152            ret = -ERESTARTNOINTR;
2153    }
2154    return ret;
2155}
2156
2157/**
2158 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2159 * @uaddr: the futex we initially wait on (non-pi)
2160 * @fshared: whether the futexes are shared (1) or not (0). They must be
2161 * the same type, no requeueing from private to shared, etc.
2162 * @val: the expected value of uaddr
2163 * @abs_time: absolute timeout
2164 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
2165 * @clockrt: whether to use CLOCK_REALTIME (1) or CLOCK_MONOTONIC (0)
2166 * @uaddr2: the pi futex we will take prior to returning to user-space
2167 *
2168 * The caller will wait on uaddr and will be requeued by futex_requeue() to
2169 * uaddr2 which must be PI aware. Normal wakeup will wake on uaddr2 and
2170 * complete the acquisition of the rt_mutex prior to returning to userspace.
2171 * This ensures the rt_mutex maintains an owner when it has waiters; without
2172 * one, the pi logic wouldn't know which task to boost/deboost, if there was a
2173 * need to.
2174 *
2175 * We call schedule in futex_wait_queue_me() when we enqueue and return there
2176 * via the following:
2177 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2178 * 2) wakeup on uaddr2 after a requeue
2179 * 3) signal
2180 * 4) timeout
2181 *
2182 * If 3, cleanup and return -ERESTARTNOINTR.
2183 *
2184 * If 2, we may then block on trying to take the rt_mutex and return via:
2185 * 5) successful lock
2186 * 6) signal
2187 * 7) timeout
2188 * 8) other lock acquisition failure
2189 *
2190 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2191 *
2192 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2193 *
2194 * Returns:
2195 * 0 - On success
2196 * <0 - On error
2197 */
2198static int futex_wait_requeue_pi(u32 __user *uaddr, int fshared,
2199                 u32 val, ktime_t *abs_time, u32 bitset,
2200                 int clockrt, u32 __user *uaddr2)
2201{
2202    struct hrtimer_sleeper timeout, *to = NULL;
2203    struct rt_mutex_waiter rt_waiter;
2204    struct rt_mutex *pi_mutex = NULL;
2205    struct futex_hash_bucket *hb;
2206    union futex_key key2;
2207    struct futex_q q;
2208    int res, ret;
2209
2210    if (!bitset)
2211        return -EINVAL;
2212
2213    if (abs_time) {
2214        to = &timeout;
2215        hrtimer_init_on_stack(&to->timer, clockrt ? CLOCK_REALTIME :
2216                      CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
2217        hrtimer_init_sleeper(to, current);
2218        hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2219                         current->timer_slack_ns);
2220    }
2221
2222    /*
2223     * The waiter is allocated on our stack, manipulated by the requeue
2224     * code while we sleep on uaddr.
2225     */
2226    debug_rt_mutex_init_waiter(&rt_waiter);
2227    rt_waiter.task = NULL;
2228
2229    key2 = FUTEX_KEY_INIT;
2230    ret = get_futex_key(uaddr2, fshared, &key2);
2231    if (unlikely(ret != 0))
2232        goto out;
2233
2234    q.pi_state = NULL;
2235    q.bitset = bitset;
2236    q.rt_waiter = &rt_waiter;
2237    q.requeue_pi_key = &key2;
2238
2239    /* Prepare to wait on uaddr. */
2240    ret = futex_wait_setup(uaddr, val, fshared, &q, &hb);
2241    if (ret)
2242        goto out_key2;
2243
2244    /* Queue the futex_q, drop the hb lock, wait for wakeup. */
2245    futex_wait_queue_me(hb, &q, to);
2246
2247    spin_lock(&hb->lock);
2248    ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2249    spin_unlock(&hb->lock);
2250    if (ret)
2251        goto out_put_keys;
2252
2253    /*
2254     * In order for us to be here, we know our q.key == key2, and since
2255     * we took the hb->lock above, we also know that futex_requeue() has
2256     * completed and we no longer have to concern ourselves with a wakeup
2257     * race with the atomic proxy lock acquition by the requeue code.
2258     */
2259
2260    /* Check if the requeue code acquired the second futex for us. */
2261    if (!q.rt_waiter) {
2262        /*
2263         * Got the lock. We might not be the anticipated owner if we
2264         * did a lock-steal - fix up the PI-state in that case.
2265         */
2266        if (q.pi_state && (q.pi_state->owner != current)) {
2267            spin_lock(q.lock_ptr);
2268            ret = fixup_pi_state_owner(uaddr2, &q, current,
2269                           fshared);
2270            spin_unlock(q.lock_ptr);
2271        }
2272    } else {
2273        /*
2274         * We have been woken up by futex_unlock_pi(), a timeout, or a
2275         * signal. futex_unlock_pi() will not destroy the lock_ptr nor
2276         * the pi_state.
2277         */
2278        WARN_ON(!&q.pi_state);
2279        pi_mutex = &q.pi_state->pi_mutex;
2280        ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter, 1);
2281        debug_rt_mutex_free_waiter(&rt_waiter);
2282
2283        spin_lock(q.lock_ptr);
2284        /*
2285         * Fixup the pi_state owner and possibly acquire the lock if we
2286         * haven't already.
2287         */
2288        res = fixup_owner(uaddr2, fshared, &q, !ret);
2289        /*
2290         * If fixup_owner() returned an error, proprogate that. If it
2291         * acquired the lock, clear -ETIMEDOUT or -EINTR.
2292         */
2293        if (res)
2294            ret = (res < 0) ? res : 0;
2295
2296        /* Unqueue and drop the lock. */
2297        unqueue_me_pi(&q);
2298    }
2299
2300    /*
2301     * If fixup_pi_state_owner() faulted and was unable to handle the
2302     * fault, unlock the rt_mutex and return the fault to userspace.
2303     */
2304    if (ret == -EFAULT) {
2305        if (rt_mutex_owner(pi_mutex) == current)
2306            rt_mutex_unlock(pi_mutex);
2307    } else if (ret == -EINTR) {
2308        /*
2309         * We've already been requeued, but cannot restart by calling
2310         * futex_lock_pi() directly. We could restart this syscall, but
2311         * it would detect that the user space "val" changed and return
2312         * -EWOULDBLOCK. Save the overhead of the restart and return
2313         * -EWOULDBLOCK directly.
2314         */
2315        ret = -EWOULDBLOCK;
2316    }
2317
2318out_put_keys:
2319    put_futex_key(fshared, &q.key);
2320out_key2:
2321    put_futex_key(fshared, &key2);
2322
2323out:
2324    if (to) {
2325        hrtimer_cancel(&to->timer);
2326        destroy_hrtimer_on_stack(&to->timer);
2327    }
2328    return ret;
2329}
2330
2331/*
2332 * Support for robust futexes: the kernel cleans up held futexes at
2333 * thread exit time.
2334 *
2335 * Implementation: user-space maintains a per-thread list of locks it
2336 * is holding. Upon do_exit(), the kernel carefully walks this list,
2337 * and marks all locks that are owned by this thread with the
2338 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2339 * always manipulated with the lock held, so the list is private and
2340 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2341 * field, to allow the kernel to clean up if the thread dies after
2342 * acquiring the lock, but just before it could have added itself to
2343 * the list. There can only be one such pending lock.
2344 */
2345
2346/**
2347 * sys_set_robust_list() - Set the robust-futex list head of a task
2348 * @head: pointer to the list-head
2349 * @len: length of the list-head, as userspace expects
2350 */
2351SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2352        size_t, len)
2353{
2354    if (!futex_cmpxchg_enabled)
2355        return -ENOSYS;
2356    /*
2357     * The kernel knows only one size for now:
2358     */
2359    if (unlikely(len != sizeof(*head)))
2360        return -EINVAL;
2361
2362    current->robust_list = head;
2363
2364    return 0;
2365}
2366
2367/**
2368 * sys_get_robust_list() - Get the robust-futex list head of a task
2369 * @pid: pid of the process [zero for current task]
2370 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
2371 * @len_ptr: pointer to a length field, the kernel fills in the header size
2372 */
2373SYSCALL_DEFINE3(get_robust_list, int, pid,
2374        struct robust_list_head __user * __user *, head_ptr,
2375        size_t __user *, len_ptr)
2376{
2377    struct robust_list_head __user *head;
2378    unsigned long ret;
2379    const struct cred *cred = current_cred(), *pcred;
2380
2381    if (!futex_cmpxchg_enabled)
2382        return -ENOSYS;
2383
2384    if (!pid)
2385        head = current->robust_list;
2386    else {
2387        struct task_struct *p;
2388
2389        ret = -ESRCH;
2390        rcu_read_lock();
2391        p = find_task_by_vpid(pid);
2392        if (!p)
2393            goto err_unlock;
2394        ret = -EPERM;
2395        pcred = __task_cred(p);
2396        if (cred->euid != pcred->euid &&
2397            cred->euid != pcred->uid &&
2398            !capable(CAP_SYS_PTRACE))
2399            goto err_unlock;
2400        head = p->robust_list;
2401        rcu_read_unlock();
2402    }
2403
2404    if (put_user(sizeof(*head), len_ptr))
2405        return -EFAULT;
2406    return put_user(head, head_ptr);
2407
2408err_unlock:
2409    rcu_read_unlock();
2410
2411    return ret;
2412}
2413
2414/*
2415 * Process a futex-list entry, check whether it's owned by the
2416 * dying task, and do notification if so:
2417 */
2418int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2419{
2420    u32 uval, nval, mval;
2421
2422retry:
2423    if (get_user(uval, uaddr))
2424        return -1;
2425
2426    if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2427        /*
2428         * Ok, this dying thread is truly holding a futex
2429         * of interest. Set the OWNER_DIED bit atomically
2430         * via cmpxchg, and if the value had FUTEX_WAITERS
2431         * set, wake up a waiter (if any). (We have to do a
2432         * futex_wake() even if OWNER_DIED is already set -
2433         * to handle the rare but possible case of recursive
2434         * thread-death.) The rest of the cleanup is done in
2435         * userspace.
2436         */
2437        mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2438        nval = futex_atomic_cmpxchg_inatomic(uaddr, uval, mval);
2439
2440        if (nval == -EFAULT)
2441            return -1;
2442
2443        if (nval != uval)
2444            goto retry;
2445
2446        /*
2447         * Wake robust non-PI futexes here. The wakeup of
2448         * PI futexes happens in exit_pi_state():
2449         */
2450        if (!pi && (uval & FUTEX_WAITERS))
2451            futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2452    }
2453    return 0;
2454}
2455
2456/*
2457 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2458 */
2459static inline int fetch_robust_entry(struct robust_list __user **entry,
2460                     struct robust_list __user * __user *head,
2461                     int *pi)
2462{
2463    unsigned long uentry;
2464
2465    if (get_user(uentry, (unsigned long __user *)head))
2466        return -EFAULT;
2467
2468    *entry = (void __user *)(uentry & ~1UL);
2469    *pi = uentry & 1;
2470
2471    return 0;
2472}
2473
2474/*
2475 * Walk curr->robust_list (very carefully, it's a userspace list!)
2476 * and mark any locks found there dead, and notify any waiters.
2477 *
2478 * We silently return on any sign of list-walking problem.
2479 */
2480void exit_robust_list(struct task_struct *curr)
2481{
2482    struct robust_list_head __user *head = curr->robust_list;
2483    struct robust_list __user *entry, *next_entry, *pending;
2484    unsigned int limit = ROBUST_LIST_LIMIT, pi, next_pi, pip;
2485    unsigned long futex_offset;
2486    int rc;
2487
2488    if (!futex_cmpxchg_enabled)
2489        return;
2490
2491    /*
2492     * Fetch the list head (which was registered earlier, via
2493     * sys_set_robust_list()):
2494     */
2495    if (fetch_robust_entry(&entry, &head->list.next, &pi))
2496        return;
2497    /*
2498     * Fetch the relative futex offset:
2499     */
2500    if (get_user(futex_offset, &head->futex_offset))
2501        return;
2502    /*
2503     * Fetch any possibly pending lock-add first, and handle it
2504     * if it exists:
2505     */
2506    if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2507        return;
2508
2509    next_entry = NULL; /* avoid warning with gcc */
2510    while (entry != &head->list) {
2511        /*
2512         * Fetch the next entry in the list before calling
2513         * handle_futex_death:
2514         */
2515        rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2516        /*
2517         * A pending lock might already be on the list, so
2518         * don't process it twice:
2519         */
2520        if (entry != pending)
2521            if (handle_futex_death((void __user *)entry + futex_offset,
2522                        curr, pi))
2523                return;
2524        if (rc)
2525            return;
2526        entry = next_entry;
2527        pi = next_pi;
2528        /*
2529         * Avoid excessively long or circular lists:
2530         */
2531        if (!--limit)
2532            break;
2533
2534        cond_resched();
2535    }
2536
2537    if (pending)
2538        handle_futex_death((void __user *)pending + futex_offset,
2539                   curr, pip);
2540}
2541
2542long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2543        u32 __user *uaddr2, u32 val2, u32 val3)
2544{
2545    int clockrt, ret = -ENOSYS;
2546    int cmd = op & FUTEX_CMD_MASK;
2547    int fshared = 0;
2548
2549    if (!(op & FUTEX_PRIVATE_FLAG))
2550        fshared = 1;
2551
2552    clockrt = op & FUTEX_CLOCK_REALTIME;
2553    if (clockrt && cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2554        return -ENOSYS;
2555
2556    switch (cmd) {
2557    case FUTEX_WAIT:
2558        val3 = FUTEX_BITSET_MATCH_ANY;
2559    case FUTEX_WAIT_BITSET:
2560        ret = futex_wait(uaddr, fshared, val, timeout, val3, clockrt);
2561        break;
2562    case FUTEX_WAKE:
2563        val3 = FUTEX_BITSET_MATCH_ANY;
2564    case FUTEX_WAKE_BITSET:
2565        ret = futex_wake(uaddr, fshared, val, val3);
2566        break;
2567    case FUTEX_REQUEUE:
2568        ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL, 0);
2569        break;
2570    case FUTEX_CMP_REQUEUE:
2571        ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2572                    0);
2573        break;
2574    case FUTEX_WAKE_OP:
2575        ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
2576        break;
2577    case FUTEX_LOCK_PI:
2578        if (futex_cmpxchg_enabled)
2579            ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
2580        break;
2581    case FUTEX_UNLOCK_PI:
2582        if (futex_cmpxchg_enabled)
2583            ret = futex_unlock_pi(uaddr, fshared);
2584        break;
2585    case FUTEX_TRYLOCK_PI:
2586        if (futex_cmpxchg_enabled)
2587            ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
2588        break;
2589    case FUTEX_WAIT_REQUEUE_PI:
2590        val3 = FUTEX_BITSET_MATCH_ANY;
2591        ret = futex_wait_requeue_pi(uaddr, fshared, val, timeout, val3,
2592                        clockrt, uaddr2);
2593        break;
2594    case FUTEX_CMP_REQUEUE_PI:
2595        ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3,
2596                    1);
2597        break;
2598    default:
2599        ret = -ENOSYS;
2600    }
2601    return ret;
2602}
2603
2604
2605SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
2606        struct timespec __user *, utime, u32 __user *, uaddr2,
2607        u32, val3)
2608{
2609    struct timespec ts;
2610    ktime_t t, *tp = NULL;
2611    u32 val2 = 0;
2612    int cmd = op & FUTEX_CMD_MASK;
2613
2614    if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
2615              cmd == FUTEX_WAIT_BITSET ||
2616              cmd == FUTEX_WAIT_REQUEUE_PI)) {
2617        if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
2618            return -EFAULT;
2619        if (!timespec_valid(&ts))
2620            return -EINVAL;
2621
2622        t = timespec_to_ktime(ts);
2623        if (cmd == FUTEX_WAIT)
2624            t = ktime_add_safe(ktime_get(), t);
2625        tp = &t;
2626    }
2627    /*
2628     * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
2629     * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
2630     */
2631    if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
2632        cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
2633        val2 = (u32) (unsigned long) utime;
2634
2635    return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
2636}
2637
2638static int __init futex_init(void)
2639{
2640    u32 curval;
2641    int i;
2642
2643    /*
2644     * This will fail and we want it. Some arch implementations do
2645     * runtime detection of the futex_atomic_cmpxchg_inatomic()
2646     * functionality. We want to know that before we call in any
2647     * of the complex code paths. Also we want to prevent
2648     * registration of robust lists in that case. NULL is
2649     * guaranteed to fault and we get -EFAULT on functional
2650     * implementation, the non functional ones will return
2651     * -ENOSYS.
2652     */
2653    curval = cmpxchg_futex_value_locked(NULL, 0, 0);
2654    if (curval == -EFAULT)
2655        futex_cmpxchg_enabled = 1;
2656
2657    for (i = 0; i < ARRAY_SIZE(futex_queues); i++) {
2658        plist_head_init(&futex_queues[i].chain, &futex_queues[i].lock);
2659        spin_lock_init(&futex_queues[i].lock);
2660    }
2661
2662    return 0;
2663}
2664__initcall(futex_init);
2665

Archive Download this file



interactive