Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

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

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us