~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/futex.c

Version: ~ [ linux-6.1-rc7 ] ~ [ linux-6.0.10 ] ~ [ linux-5.19.17 ] ~ [ linux-5.18.19 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.80 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.156 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.225 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.267 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.300 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.334 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp