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

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