• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Fast Userspace Mutexes (which I call "Futexes!").
3  *  (C) Rusty Russell, IBM 2002
4  *
5  *  Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6  *  (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7  *
8  *  Removed page pinning, fix privately mapped COW pages and other cleanups
9  *  (C) Copyright 2003, 2004 Jamie Lokier
10  *
11  *  Robust futex support started by Ingo Molnar
12  *  (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13  *  Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14  *
15  *  PI-futex support started by Ingo Molnar and Thomas Gleixner
16  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18  *
19  *  PRIVATE futexes by Eric Dumazet
20  *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21  *
22  *  Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23  *  Copyright (C) IBM Corporation, 2009
24  *  Thanks to Thomas Gleixner for conceptual design and careful reviews.
25  *
26  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27  *  enough at me, Linus for the original (flawed) idea, Matthew
28  *  Kirkwood for proof-of-concept implementation.
29  *
30  *  "The futexes are also cursed."
31  *  "But they come in a choice of three flavours!"
32  *
33  *  This program is free software; you can redistribute it and/or modify
34  *  it under the terms of the GNU General Public License as published by
35  *  the Free Software Foundation; either version 2 of the License, or
36  *  (at your option) any later version.
37  *
38  *  This program is distributed in the hope that it will be useful,
39  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
40  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
41  *  GNU General Public License for more details.
42  *
43  *  You should have received a copy of the GNU General Public License
44  *  along with this program; if not, write to the Free Software
45  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
46  */
47 #include <linux/slab.h>
48 #include <linux/poll.h>
49 #include <linux/fs.h>
50 #include <linux/file.h>
51 #include <linux/jhash.h>
52 #include <linux/init.h>
53 #include <linux/futex.h>
54 #include <linux/mount.h>
55 #include <linux/pagemap.h>
56 #include <linux/syscalls.h>
57 #include <linux/signal.h>
58 #include <linux/export.h>
59 #include <linux/magic.h>
60 #include <linux/pid.h>
61 #include <linux/nsproxy.h>
62 #include <linux/ptrace.h>
63 #include <linux/sched/rt.h>
64 #include <linux/hugetlb.h>
65 #include <linux/freezer.h>
66 #include <linux/bootmem.h>
67 
68 #include <asm/futex.h>
69 
70 #include "locking/rtmutex_common.h"
71 
72 /*
73  * READ this before attempting to hack on futexes!
74  *
75  * Basic futex operation and ordering guarantees
76  * =============================================
77  *
78  * The waiter reads the futex value in user space and calls
79  * futex_wait(). This function computes the hash bucket and acquires
80  * the hash bucket lock. After that it reads the futex user space value
81  * again and verifies that the data has not changed. If it has not changed
82  * it enqueues itself into the hash bucket, releases the hash bucket lock
83  * and schedules.
84  *
85  * The waker side modifies the user space value of the futex and calls
86  * futex_wake(). This function computes the hash bucket and acquires the
87  * hash bucket lock. Then it looks for waiters on that futex in the hash
88  * bucket and wakes them.
89  *
90  * In futex wake up scenarios where no tasks are blocked on a futex, taking
91  * the hb spinlock can be avoided and simply return. In order for this
92  * optimization to work, ordering guarantees must exist so that the waiter
93  * being added to the list is acknowledged when the list is concurrently being
94  * checked by the waker, avoiding scenarios like the following:
95  *
96  * CPU 0                               CPU 1
97  * val = *futex;
98  * sys_futex(WAIT, futex, val);
99  *   futex_wait(futex, val);
100  *   uval = *futex;
101  *                                     *futex = newval;
102  *                                     sys_futex(WAKE, futex);
103  *                                       futex_wake(futex);
104  *                                       if (queue_empty())
105  *                                         return;
106  *   if (uval == val)
107  *      lock(hash_bucket(futex));
108  *      queue();
109  *     unlock(hash_bucket(futex));
110  *     schedule();
111  *
112  * This would cause the waiter on CPU 0 to wait forever because it
113  * missed the transition of the user space value from val to newval
114  * and the waker did not find the waiter in the hash bucket queue.
115  *
116  * The correct serialization ensures that a waiter either observes
117  * the changed user space value before blocking or is woken by a
118  * concurrent waker:
119  *
120  * CPU 0                                 CPU 1
121  * val = *futex;
122  * sys_futex(WAIT, futex, val);
123  *   futex_wait(futex, val);
124  *
125  *   waiters++; (a)
126  *   mb(); (A) <-- paired with -.
127  *                              |
128  *   lock(hash_bucket(futex));  |
129  *                              |
130  *   uval = *futex;             |
131  *                              |        *futex = newval;
132  *                              |        sys_futex(WAKE, futex);
133  *                              |          futex_wake(futex);
134  *                              |
135  *                              `------->  mb(); (B)
136  *   if (uval == val)
137  *     queue();
138  *     unlock(hash_bucket(futex));
139  *     schedule();                         if (waiters)
140  *                                           lock(hash_bucket(futex));
141  *   else                                    wake_waiters(futex);
142  *     waiters--; (b)                        unlock(hash_bucket(futex));
143  *
144  * Where (A) orders the waiters increment and the futex value read through
145  * atomic operations (see hb_waiters_inc) and where (B) orders the write
146  * to futex and the waiters read -- this is done by the barriers for both
147  * shared and private futexes in get_futex_key_refs().
148  *
149  * This yields the following case (where X:=waiters, Y:=futex):
150  *
151  *	X = Y = 0
152  *
153  *	w[X]=1		w[Y]=1
154  *	MB		MB
155  *	r[Y]=y		r[X]=x
156  *
157  * Which guarantees that x==0 && y==0 is impossible; which translates back into
158  * the guarantee that we cannot both miss the futex variable change and the
159  * enqueue.
160  *
161  * Note that a new waiter is accounted for in (a) even when it is possible that
162  * the wait call can return error, in which case we backtrack from it in (b).
163  * Refer to the comment in queue_lock().
164  *
165  * Similarly, in order to account for waiters being requeued on another
166  * address we always increment the waiters for the destination bucket before
167  * acquiring the lock. It then decrements them again  after releasing it -
168  * the code that actually moves the futex(es) between hash buckets (requeue_futex)
169  * will do the additional required waiter count housekeeping. This is done for
170  * double_lock_hb() and double_unlock_hb(), respectively.
171  */
172 
173 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
174 int __read_mostly futex_cmpxchg_enabled;
175 #endif
176 
177 /*
178  * Futex flags used to encode options to functions and preserve them across
179  * restarts.
180  */
181 #define FLAGS_SHARED		0x01
182 #define FLAGS_CLOCKRT		0x02
183 #define FLAGS_HAS_TIMEOUT	0x04
184 
185 /*
186  * Priority Inheritance state:
187  */
188 struct futex_pi_state {
189 	/*
190 	 * list of 'owned' pi_state instances - these have to be
191 	 * cleaned up in do_exit() if the task exits prematurely:
192 	 */
193 	struct list_head list;
194 
195 	/*
196 	 * The PI object:
197 	 */
198 	struct rt_mutex pi_mutex;
199 
200 	struct task_struct *owner;
201 	atomic_t refcount;
202 
203 	union futex_key key;
204 };
205 
206 /**
207  * struct futex_q - The hashed futex queue entry, one per waiting task
208  * @list:		priority-sorted list of tasks waiting on this futex
209  * @task:		the task waiting on the futex
210  * @lock_ptr:		the hash bucket lock
211  * @key:		the key the futex is hashed on
212  * @pi_state:		optional priority inheritance state
213  * @rt_waiter:		rt_waiter storage for use with requeue_pi
214  * @requeue_pi_key:	the requeue_pi target futex key
215  * @bitset:		bitset for the optional bitmasked wakeup
216  *
217  * We use this hashed waitqueue, instead of a normal wait_queue_t, so
218  * we can wake only the relevant ones (hashed queues may be shared).
219  *
220  * A futex_q has a woken state, just like tasks have TASK_RUNNING.
221  * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
222  * The order of wakeup is always to make the first condition true, then
223  * the second.
224  *
225  * PI futexes are typically woken before they are removed from the hash list via
226  * the rt_mutex code. See unqueue_me_pi().
227  */
228 struct futex_q {
229 	struct plist_node list;
230 
231 	struct task_struct *task;
232 	spinlock_t *lock_ptr;
233 	union futex_key key;
234 	struct futex_pi_state *pi_state;
235 	struct rt_mutex_waiter *rt_waiter;
236 	union futex_key *requeue_pi_key;
237 	u32 bitset;
238 };
239 
240 static const struct futex_q futex_q_init = {
241 	/* list gets initialized in queue_me()*/
242 	.key = FUTEX_KEY_INIT,
243 	.bitset = FUTEX_BITSET_MATCH_ANY
244 };
245 
246 /*
247  * Hash buckets are shared by all the futex_keys that hash to the same
248  * location.  Each key may have multiple futex_q structures, one for each task
249  * waiting on a futex.
250  */
251 struct futex_hash_bucket {
252 	atomic_t waiters;
253 	spinlock_t lock;
254 	struct plist_head chain;
255 } ____cacheline_aligned_in_smp;
256 
257 static unsigned long __read_mostly futex_hashsize;
258 
259 static struct futex_hash_bucket *futex_queues;
260 
futex_get_mm(union futex_key * key)261 static inline void futex_get_mm(union futex_key *key)
262 {
263 	atomic_inc(&key->private.mm->mm_count);
264 	/*
265 	 * Ensure futex_get_mm() implies a full barrier such that
266 	 * get_futex_key() implies a full barrier. This is relied upon
267 	 * as full barrier (B), see the ordering comment above.
268 	 */
269 	smp_mb__after_atomic();
270 }
271 
272 /*
273  * Reflects a new waiter being added to the waitqueue.
274  */
hb_waiters_inc(struct futex_hash_bucket * hb)275 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
276 {
277 #ifdef CONFIG_SMP
278 	atomic_inc(&hb->waiters);
279 	/*
280 	 * Full barrier (A), see the ordering comment above.
281 	 */
282 	smp_mb__after_atomic();
283 #endif
284 }
285 
286 /*
287  * Reflects a waiter being removed from the waitqueue by wakeup
288  * paths.
289  */
hb_waiters_dec(struct futex_hash_bucket * hb)290 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
291 {
292 #ifdef CONFIG_SMP
293 	atomic_dec(&hb->waiters);
294 #endif
295 }
296 
hb_waiters_pending(struct futex_hash_bucket * hb)297 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
298 {
299 #ifdef CONFIG_SMP
300 	return atomic_read(&hb->waiters);
301 #else
302 	return 1;
303 #endif
304 }
305 
306 /*
307  * We hash on the keys returned from get_futex_key (see below).
308  */
hash_futex(union futex_key * key)309 static struct futex_hash_bucket *hash_futex(union futex_key *key)
310 {
311 	u32 hash = jhash2((u32*)&key->both.word,
312 			  (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
313 			  key->both.offset);
314 	return &futex_queues[hash & (futex_hashsize - 1)];
315 }
316 
317 /*
318  * Return 1 if two futex_keys are equal, 0 otherwise.
319  */
match_futex(union futex_key * key1,union futex_key * key2)320 static inline int match_futex(union futex_key *key1, union futex_key *key2)
321 {
322 	return (key1 && key2
323 		&& key1->both.word == key2->both.word
324 		&& key1->both.ptr == key2->both.ptr
325 		&& key1->both.offset == key2->both.offset);
326 }
327 
328 /*
329  * Take a reference to the resource addressed by a key.
330  * Can be called while holding spinlocks.
331  *
332  */
get_futex_key_refs(union futex_key * key)333 static void get_futex_key_refs(union futex_key *key)
334 {
335 	if (!key->both.ptr)
336 		return;
337 
338 	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
339 	case FUT_OFF_INODE:
340 		ihold(key->shared.inode); /* implies MB (B) */
341 		break;
342 	case FUT_OFF_MMSHARED:
343 		futex_get_mm(key); /* implies MB (B) */
344 		break;
345 	default:
346 		/*
347 		 * Private futexes do not hold reference on an inode or
348 		 * mm, therefore the only purpose of calling get_futex_key_refs
349 		 * is because we need the barrier for the lockless waiter check.
350 		 */
351 		smp_mb(); /* explicit MB (B) */
352 	}
353 }
354 
355 /*
356  * Drop a reference to the resource addressed by a key.
357  * The hash bucket spinlock must not be held. This is
358  * a no-op for private futexes, see comment in the get
359  * counterpart.
360  */
drop_futex_key_refs(union futex_key * key)361 static void drop_futex_key_refs(union futex_key *key)
362 {
363 	if (!key->both.ptr) {
364 		/* If we're here then we tried to put a key we failed to get */
365 		WARN_ON_ONCE(1);
366 		return;
367 	}
368 
369 	switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
370 	case FUT_OFF_INODE:
371 		iput(key->shared.inode);
372 		break;
373 	case FUT_OFF_MMSHARED:
374 		mmdrop(key->private.mm);
375 		break;
376 	}
377 }
378 
379 /**
380  * get_futex_key() - Get parameters which are the keys for a futex
381  * @uaddr:	virtual address of the futex
382  * @fshared:	0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
383  * @key:	address where result is stored.
384  * @rw:		mapping needs to be read/write (values: VERIFY_READ,
385  *              VERIFY_WRITE)
386  *
387  * Return: a negative error code or 0
388  *
389  * The key words are stored in *key on success.
390  *
391  * For shared mappings, it's (page->index, file_inode(vma->vm_file),
392  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
393  * We can usually work out the index without swapping in the page.
394  *
395  * lock_page() might sleep, the caller should not hold a spinlock.
396  */
397 static int
get_futex_key(u32 __user * uaddr,int fshared,union futex_key * key,int rw)398 get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
399 {
400 	unsigned long address = (unsigned long)uaddr;
401 	struct mm_struct *mm = current->mm;
402 	struct page *page, *page_head;
403 	int err, ro = 0;
404 
405 	/*
406 	 * The futex address must be "naturally" aligned.
407 	 */
408 	key->both.offset = address % PAGE_SIZE;
409 	if (unlikely((address % sizeof(u32)) != 0))
410 		return -EINVAL;
411 	address -= key->both.offset;
412 
413 	if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
414 		return -EFAULT;
415 
416 	/*
417 	 * PROCESS_PRIVATE futexes are fast.
418 	 * As the mm cannot disappear under us and the 'key' only needs
419 	 * virtual address, we dont even have to find the underlying vma.
420 	 * Note : We do have to check 'uaddr' is a valid user address,
421 	 *        but access_ok() should be faster than find_vma()
422 	 */
423 	if (!fshared) {
424 		key->private.mm = mm;
425 		key->private.address = address;
426 		get_futex_key_refs(key);  /* implies MB (B) */
427 		return 0;
428 	}
429 
430 again:
431 	err = get_user_pages_fast(address, 1, 1, &page);
432 	/*
433 	 * If write access is not required (eg. FUTEX_WAIT), try
434 	 * and get read-only access.
435 	 */
436 	if (err == -EFAULT && rw == VERIFY_READ) {
437 		err = get_user_pages_fast(address, 1, 0, &page);
438 		ro = 1;
439 	}
440 	if (err < 0)
441 		return err;
442 	else
443 		err = 0;
444 
445 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
446 	page_head = page;
447 	if (unlikely(PageTail(page))) {
448 		put_page(page);
449 		/* serialize against __split_huge_page_splitting() */
450 		local_irq_disable();
451 		if (likely(__get_user_pages_fast(address, 1, !ro, &page) == 1)) {
452 			page_head = compound_head(page);
453 			/*
454 			 * page_head is valid pointer but we must pin
455 			 * it before taking the PG_lock and/or
456 			 * PG_compound_lock. The moment we re-enable
457 			 * irqs __split_huge_page_splitting() can
458 			 * return and the head page can be freed from
459 			 * under us. We can't take the PG_lock and/or
460 			 * PG_compound_lock on a page that could be
461 			 * freed from under us.
462 			 */
463 			if (page != page_head) {
464 				get_page(page_head);
465 				put_page(page);
466 			}
467 			local_irq_enable();
468 		} else {
469 			local_irq_enable();
470 			goto again;
471 		}
472 	}
473 #else
474 	page_head = compound_head(page);
475 	if (page != page_head) {
476 		get_page(page_head);
477 		put_page(page);
478 	}
479 #endif
480 
481 	lock_page(page_head);
482 
483 	/*
484 	 * If page_head->mapping is NULL, then it cannot be a PageAnon
485 	 * page; but it might be the ZERO_PAGE or in the gate area or
486 	 * in a special mapping (all cases which we are happy to fail);
487 	 * or it may have been a good file page when get_user_pages_fast
488 	 * found it, but truncated or holepunched or subjected to
489 	 * invalidate_complete_page2 before we got the page lock (also
490 	 * cases which we are happy to fail).  And we hold a reference,
491 	 * so refcount care in invalidate_complete_page's remove_mapping
492 	 * prevents drop_caches from setting mapping to NULL beneath us.
493 	 *
494 	 * The case we do have to guard against is when memory pressure made
495 	 * shmem_writepage move it from filecache to swapcache beneath us:
496 	 * an unlikely race, but we do need to retry for page_head->mapping.
497 	 */
498 	if (!page_head->mapping) {
499 		int shmem_swizzled = PageSwapCache(page_head);
500 		unlock_page(page_head);
501 		put_page(page_head);
502 		if (shmem_swizzled)
503 			goto again;
504 		return -EFAULT;
505 	}
506 
507 	/*
508 	 * Private mappings are handled in a simple way.
509 	 *
510 	 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
511 	 * it's a read-only handle, it's expected that futexes attach to
512 	 * the object not the particular process.
513 	 */
514 	if (PageAnon(page_head)) {
515 		/*
516 		 * A RO anonymous page will never change and thus doesn't make
517 		 * sense for futex operations.
518 		 */
519 		if (ro) {
520 			err = -EFAULT;
521 			goto out;
522 		}
523 
524 		key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
525 		key->private.mm = mm;
526 		key->private.address = address;
527 	} else {
528 		key->both.offset |= FUT_OFF_INODE; /* inode-based key */
529 		key->shared.inode = page_head->mapping->host;
530 		key->shared.pgoff = basepage_index(page);
531 	}
532 
533 	get_futex_key_refs(key); /* implies MB (B) */
534 
535 out:
536 	unlock_page(page_head);
537 	put_page(page_head);
538 	return err;
539 }
540 
put_futex_key(union futex_key * key)541 static inline void put_futex_key(union futex_key *key)
542 {
543 	drop_futex_key_refs(key);
544 }
545 
546 /**
547  * fault_in_user_writeable() - Fault in user address and verify RW access
548  * @uaddr:	pointer to faulting user space address
549  *
550  * Slow path to fixup the fault we just took in the atomic write
551  * access to @uaddr.
552  *
553  * We have no generic implementation of a non-destructive write to the
554  * user address. We know that we faulted in the atomic pagefault
555  * disabled section so we can as well avoid the #PF overhead by
556  * calling get_user_pages() right away.
557  */
fault_in_user_writeable(u32 __user * uaddr)558 static int fault_in_user_writeable(u32 __user *uaddr)
559 {
560 	struct mm_struct *mm = current->mm;
561 	int ret;
562 
563 	down_read(&mm->mmap_sem);
564 	ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
565 			       FAULT_FLAG_WRITE);
566 	up_read(&mm->mmap_sem);
567 
568 	return ret < 0 ? ret : 0;
569 }
570 
571 /**
572  * futex_top_waiter() - Return the highest priority waiter on a futex
573  * @hb:		the hash bucket the futex_q's reside in
574  * @key:	the futex key (to distinguish it from other futex futex_q's)
575  *
576  * Must be called with the hb lock held.
577  */
futex_top_waiter(struct futex_hash_bucket * hb,union futex_key * key)578 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
579 					union futex_key *key)
580 {
581 	struct futex_q *this;
582 
583 	plist_for_each_entry(this, &hb->chain, list) {
584 		if (match_futex(&this->key, key))
585 			return this;
586 	}
587 	return NULL;
588 }
589 
cmpxchg_futex_value_locked(u32 * curval,u32 __user * uaddr,u32 uval,u32 newval)590 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
591 				      u32 uval, u32 newval)
592 {
593 	int ret;
594 
595 	pagefault_disable();
596 	ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
597 	pagefault_enable();
598 
599 	return ret;
600 }
601 
get_futex_value_locked(u32 * dest,u32 __user * from)602 static int get_futex_value_locked(u32 *dest, u32 __user *from)
603 {
604 	int ret;
605 
606 	pagefault_disable();
607 	ret = __copy_from_user_inatomic(dest, from, sizeof(u32));
608 	pagefault_enable();
609 
610 	return ret ? -EFAULT : 0;
611 }
612 
613 
614 /*
615  * PI code:
616  */
refill_pi_state_cache(void)617 static int refill_pi_state_cache(void)
618 {
619 	struct futex_pi_state *pi_state;
620 
621 	if (likely(current->pi_state_cache))
622 		return 0;
623 
624 	pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
625 
626 	if (!pi_state)
627 		return -ENOMEM;
628 
629 	INIT_LIST_HEAD(&pi_state->list);
630 	/* pi_mutex gets initialized later */
631 	pi_state->owner = NULL;
632 	atomic_set(&pi_state->refcount, 1);
633 	pi_state->key = FUTEX_KEY_INIT;
634 
635 	current->pi_state_cache = pi_state;
636 
637 	return 0;
638 }
639 
alloc_pi_state(void)640 static struct futex_pi_state * alloc_pi_state(void)
641 {
642 	struct futex_pi_state *pi_state = current->pi_state_cache;
643 
644 	WARN_ON(!pi_state);
645 	current->pi_state_cache = NULL;
646 
647 	return pi_state;
648 }
649 
650 /*
651  * Must be called with the hb lock held.
652  */
free_pi_state(struct futex_pi_state * pi_state)653 static void free_pi_state(struct futex_pi_state *pi_state)
654 {
655 	if (!pi_state)
656 		return;
657 
658 	if (!atomic_dec_and_test(&pi_state->refcount))
659 		return;
660 
661 	/*
662 	 * If pi_state->owner is NULL, the owner is most probably dying
663 	 * and has cleaned up the pi_state already
664 	 */
665 	if (pi_state->owner) {
666 		raw_spin_lock_irq(&pi_state->owner->pi_lock);
667 		list_del_init(&pi_state->list);
668 		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
669 
670 		rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
671 	}
672 
673 	if (current->pi_state_cache)
674 		kfree(pi_state);
675 	else {
676 		/*
677 		 * pi_state->list is already empty.
678 		 * clear pi_state->owner.
679 		 * refcount is at 0 - put it back to 1.
680 		 */
681 		pi_state->owner = NULL;
682 		atomic_set(&pi_state->refcount, 1);
683 		current->pi_state_cache = pi_state;
684 	}
685 }
686 
687 /*
688  * Look up the task based on what TID userspace gave us.
689  * We dont trust it.
690  */
futex_find_get_task(pid_t pid)691 static struct task_struct * futex_find_get_task(pid_t pid)
692 {
693 	struct task_struct *p;
694 
695 	rcu_read_lock();
696 	p = find_task_by_vpid(pid);
697 	if (p)
698 		get_task_struct(p);
699 
700 	rcu_read_unlock();
701 
702 	return p;
703 }
704 
705 /*
706  * This task is holding PI mutexes at exit time => bad.
707  * Kernel cleans up PI-state, but userspace is likely hosed.
708  * (Robust-futex cleanup is separate and might save the day for userspace.)
709  */
exit_pi_state_list(struct task_struct * curr)710 void exit_pi_state_list(struct task_struct *curr)
711 {
712 	struct list_head *next, *head = &curr->pi_state_list;
713 	struct futex_pi_state *pi_state;
714 	struct futex_hash_bucket *hb;
715 	union futex_key key = FUTEX_KEY_INIT;
716 
717 	if (!futex_cmpxchg_enabled)
718 		return;
719 	/*
720 	 * We are a ZOMBIE and nobody can enqueue itself on
721 	 * pi_state_list anymore, but we have to be careful
722 	 * versus waiters unqueueing themselves:
723 	 */
724 	raw_spin_lock_irq(&curr->pi_lock);
725 	while (!list_empty(head)) {
726 
727 		next = head->next;
728 		pi_state = list_entry(next, struct futex_pi_state, list);
729 		key = pi_state->key;
730 		hb = hash_futex(&key);
731 		raw_spin_unlock_irq(&curr->pi_lock);
732 
733 		spin_lock(&hb->lock);
734 
735 		raw_spin_lock_irq(&curr->pi_lock);
736 		/*
737 		 * We dropped the pi-lock, so re-check whether this
738 		 * task still owns the PI-state:
739 		 */
740 		if (head->next != next) {
741 			spin_unlock(&hb->lock);
742 			continue;
743 		}
744 
745 		WARN_ON(pi_state->owner != curr);
746 		WARN_ON(list_empty(&pi_state->list));
747 		list_del_init(&pi_state->list);
748 		pi_state->owner = NULL;
749 		raw_spin_unlock_irq(&curr->pi_lock);
750 
751 		rt_mutex_unlock(&pi_state->pi_mutex);
752 
753 		spin_unlock(&hb->lock);
754 
755 		raw_spin_lock_irq(&curr->pi_lock);
756 	}
757 	raw_spin_unlock_irq(&curr->pi_lock);
758 }
759 
760 /*
761  * We need to check the following states:
762  *
763  *      Waiter | pi_state | pi->owner | uTID      | uODIED | ?
764  *
765  * [1]  NULL   | ---      | ---       | 0         | 0/1    | Valid
766  * [2]  NULL   | ---      | ---       | >0        | 0/1    | Valid
767  *
768  * [3]  Found  | NULL     | --        | Any       | 0/1    | Invalid
769  *
770  * [4]  Found  | Found    | NULL      | 0         | 1      | Valid
771  * [5]  Found  | Found    | NULL      | >0        | 1      | Invalid
772  *
773  * [6]  Found  | Found    | task      | 0         | 1      | Valid
774  *
775  * [7]  Found  | Found    | NULL      | Any       | 0      | Invalid
776  *
777  * [8]  Found  | Found    | task      | ==taskTID | 0/1    | Valid
778  * [9]  Found  | Found    | task      | 0         | 0      | Invalid
779  * [10] Found  | Found    | task      | !=taskTID | 0/1    | Invalid
780  *
781  * [1]	Indicates that the kernel can acquire the futex atomically. We
782  *	came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
783  *
784  * [2]	Valid, if TID does not belong to a kernel thread. If no matching
785  *      thread is found then it indicates that the owner TID has died.
786  *
787  * [3]	Invalid. The waiter is queued on a non PI futex
788  *
789  * [4]	Valid state after exit_robust_list(), which sets the user space
790  *	value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
791  *
792  * [5]	The user space value got manipulated between exit_robust_list()
793  *	and exit_pi_state_list()
794  *
795  * [6]	Valid state after exit_pi_state_list() which sets the new owner in
796  *	the pi_state but cannot access the user space value.
797  *
798  * [7]	pi_state->owner can only be NULL when the OWNER_DIED bit is set.
799  *
800  * [8]	Owner and user space value match
801  *
802  * [9]	There is no transient state which sets the user space TID to 0
803  *	except exit_robust_list(), but this is indicated by the
804  *	FUTEX_OWNER_DIED bit. See [4]
805  *
806  * [10] There is no transient state which leaves owner and user space
807  *	TID out of sync.
808  */
809 
810 /*
811  * Validate that the existing waiter has a pi_state and sanity check
812  * the pi_state against the user space value. If correct, attach to
813  * it.
814  */
attach_to_pi_state(u32 uval,struct futex_pi_state * pi_state,struct futex_pi_state ** ps)815 static int attach_to_pi_state(u32 uval, struct futex_pi_state *pi_state,
816 			      struct futex_pi_state **ps)
817 {
818 	pid_t pid = uval & FUTEX_TID_MASK;
819 
820 	/*
821 	 * Userspace might have messed up non-PI and PI futexes [3]
822 	 */
823 	if (unlikely(!pi_state))
824 		return -EINVAL;
825 
826 	WARN_ON(!atomic_read(&pi_state->refcount));
827 
828 	/*
829 	 * Handle the owner died case:
830 	 */
831 	if (uval & FUTEX_OWNER_DIED) {
832 		/*
833 		 * exit_pi_state_list sets owner to NULL and wakes the
834 		 * topmost waiter. The task which acquires the
835 		 * pi_state->rt_mutex will fixup owner.
836 		 */
837 		if (!pi_state->owner) {
838 			/*
839 			 * No pi state owner, but the user space TID
840 			 * is not 0. Inconsistent state. [5]
841 			 */
842 			if (pid)
843 				return -EINVAL;
844 			/*
845 			 * Take a ref on the state and return success. [4]
846 			 */
847 			goto out_state;
848 		}
849 
850 		/*
851 		 * If TID is 0, then either the dying owner has not
852 		 * yet executed exit_pi_state_list() or some waiter
853 		 * acquired the rtmutex in the pi state, but did not
854 		 * yet fixup the TID in user space.
855 		 *
856 		 * Take a ref on the state and return success. [6]
857 		 */
858 		if (!pid)
859 			goto out_state;
860 	} else {
861 		/*
862 		 * If the owner died bit is not set, then the pi_state
863 		 * must have an owner. [7]
864 		 */
865 		if (!pi_state->owner)
866 			return -EINVAL;
867 	}
868 
869 	/*
870 	 * Bail out if user space manipulated the futex value. If pi
871 	 * state exists then the owner TID must be the same as the
872 	 * user space TID. [9/10]
873 	 */
874 	if (pid != task_pid_vnr(pi_state->owner))
875 		return -EINVAL;
876 out_state:
877 	atomic_inc(&pi_state->refcount);
878 	*ps = pi_state;
879 	return 0;
880 }
881 
882 /*
883  * Lookup the task for the TID provided from user space and attach to
884  * it after doing proper sanity checks.
885  */
attach_to_pi_owner(u32 uval,union futex_key * key,struct futex_pi_state ** ps)886 static int attach_to_pi_owner(u32 uval, union futex_key *key,
887 			      struct futex_pi_state **ps)
888 {
889 	pid_t pid = uval & FUTEX_TID_MASK;
890 	struct futex_pi_state *pi_state;
891 	struct task_struct *p;
892 
893 	/*
894 	 * We are the first waiter - try to look up the real owner and attach
895 	 * the new pi_state to it, but bail out when TID = 0 [1]
896 	 */
897 	if (!pid)
898 		return -ESRCH;
899 	p = futex_find_get_task(pid);
900 	if (!p)
901 		return -ESRCH;
902 
903 	if (!p->mm) {
904 		put_task_struct(p);
905 		return -EPERM;
906 	}
907 
908 	/*
909 	 * We need to look at the task state flags to figure out,
910 	 * whether the task is exiting. To protect against the do_exit
911 	 * change of the task flags, we do this protected by
912 	 * p->pi_lock:
913 	 */
914 	raw_spin_lock_irq(&p->pi_lock);
915 	if (unlikely(p->flags & PF_EXITING)) {
916 		/*
917 		 * The task is on the way out. When PF_EXITPIDONE is
918 		 * set, we know that the task has finished the
919 		 * cleanup:
920 		 */
921 		int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
922 
923 		raw_spin_unlock_irq(&p->pi_lock);
924 		put_task_struct(p);
925 		return ret;
926 	}
927 
928 	/*
929 	 * No existing pi state. First waiter. [2]
930 	 */
931 	pi_state = alloc_pi_state();
932 
933 	/*
934 	 * Initialize the pi_mutex in locked state and make @p
935 	 * the owner of it:
936 	 */
937 	rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
938 
939 	/* Store the key for possible exit cleanups: */
940 	pi_state->key = *key;
941 
942 	WARN_ON(!list_empty(&pi_state->list));
943 	list_add(&pi_state->list, &p->pi_state_list);
944 	pi_state->owner = p;
945 	raw_spin_unlock_irq(&p->pi_lock);
946 
947 	put_task_struct(p);
948 
949 	*ps = pi_state;
950 
951 	return 0;
952 }
953 
lookup_pi_state(u32 uval,struct futex_hash_bucket * hb,union futex_key * key,struct futex_pi_state ** ps)954 static int lookup_pi_state(u32 uval, struct futex_hash_bucket *hb,
955 			   union futex_key *key, struct futex_pi_state **ps)
956 {
957 	struct futex_q *match = futex_top_waiter(hb, key);
958 
959 	/*
960 	 * If there is a waiter on that futex, validate it and
961 	 * attach to the pi_state when the validation succeeds.
962 	 */
963 	if (match)
964 		return attach_to_pi_state(uval, match->pi_state, ps);
965 
966 	/*
967 	 * We are the first waiter - try to look up the owner based on
968 	 * @uval and attach to it.
969 	 */
970 	return attach_to_pi_owner(uval, key, ps);
971 }
972 
lock_pi_update_atomic(u32 __user * uaddr,u32 uval,u32 newval)973 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
974 {
975 	u32 uninitialized_var(curval);
976 
977 	if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
978 		return -EFAULT;
979 
980 	/*If user space value changed, let the caller retry */
981 	return curval != uval ? -EAGAIN : 0;
982 }
983 
984 /**
985  * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
986  * @uaddr:		the pi futex user address
987  * @hb:			the pi futex hash bucket
988  * @key:		the futex key associated with uaddr and hb
989  * @ps:			the pi_state pointer where we store the result of the
990  *			lookup
991  * @task:		the task to perform the atomic lock work for.  This will
992  *			be "current" except in the case of requeue pi.
993  * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
994  *
995  * Return:
996  *  0 - ready to wait;
997  *  1 - acquired the lock;
998  * <0 - error
999  *
1000  * The hb->lock and futex_key refs shall be held by the caller.
1001  */
futex_lock_pi_atomic(u32 __user * uaddr,struct futex_hash_bucket * hb,union futex_key * key,struct futex_pi_state ** ps,struct task_struct * task,int set_waiters)1002 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1003 				union futex_key *key,
1004 				struct futex_pi_state **ps,
1005 				struct task_struct *task, int set_waiters)
1006 {
1007 	u32 uval, newval, vpid = task_pid_vnr(task);
1008 	struct futex_q *match;
1009 	int ret;
1010 
1011 	/*
1012 	 * Read the user space value first so we can validate a few
1013 	 * things before proceeding further.
1014 	 */
1015 	if (get_futex_value_locked(&uval, uaddr))
1016 		return -EFAULT;
1017 
1018 	/*
1019 	 * Detect deadlocks.
1020 	 */
1021 	if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1022 		return -EDEADLK;
1023 
1024 	/*
1025 	 * Lookup existing state first. If it exists, try to attach to
1026 	 * its pi_state.
1027 	 */
1028 	match = futex_top_waiter(hb, key);
1029 	if (match)
1030 		return attach_to_pi_state(uval, match->pi_state, ps);
1031 
1032 	/*
1033 	 * No waiter and user TID is 0. We are here because the
1034 	 * waiters or the owner died bit is set or called from
1035 	 * requeue_cmp_pi or for whatever reason something took the
1036 	 * syscall.
1037 	 */
1038 	if (!(uval & FUTEX_TID_MASK)) {
1039 		/*
1040 		 * We take over the futex. No other waiters and the user space
1041 		 * TID is 0. We preserve the owner died bit.
1042 		 */
1043 		newval = uval & FUTEX_OWNER_DIED;
1044 		newval |= vpid;
1045 
1046 		/* The futex requeue_pi code can enforce the waiters bit */
1047 		if (set_waiters)
1048 			newval |= FUTEX_WAITERS;
1049 
1050 		ret = lock_pi_update_atomic(uaddr, uval, newval);
1051 		/* If the take over worked, return 1 */
1052 		return ret < 0 ? ret : 1;
1053 	}
1054 
1055 	/*
1056 	 * First waiter. Set the waiters bit before attaching ourself to
1057 	 * the owner. If owner tries to unlock, it will be forced into
1058 	 * the kernel and blocked on hb->lock.
1059 	 */
1060 	newval = uval | FUTEX_WAITERS;
1061 	ret = lock_pi_update_atomic(uaddr, uval, newval);
1062 	if (ret)
1063 		return ret;
1064 	/*
1065 	 * If the update of the user space value succeeded, we try to
1066 	 * attach to the owner. If that fails, no harm done, we only
1067 	 * set the FUTEX_WAITERS bit in the user space variable.
1068 	 */
1069 	return attach_to_pi_owner(uval, key, ps);
1070 }
1071 
1072 /**
1073  * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1074  * @q:	The futex_q to unqueue
1075  *
1076  * The q->lock_ptr must not be NULL and must be held by the caller.
1077  */
__unqueue_futex(struct futex_q * q)1078 static void __unqueue_futex(struct futex_q *q)
1079 {
1080 	struct futex_hash_bucket *hb;
1081 
1082 	if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1083 	    || WARN_ON(plist_node_empty(&q->list)))
1084 		return;
1085 
1086 	hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1087 	plist_del(&q->list, &hb->chain);
1088 	hb_waiters_dec(hb);
1089 }
1090 
1091 /*
1092  * The hash bucket lock must be held when this is called.
1093  * Afterwards, the futex_q must not be accessed.
1094  */
wake_futex(struct futex_q * q)1095 static void wake_futex(struct futex_q *q)
1096 {
1097 	struct task_struct *p = q->task;
1098 
1099 	if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1100 		return;
1101 
1102 	/*
1103 	 * We set q->lock_ptr = NULL _before_ we wake up the task. If
1104 	 * a non-futex wake up happens on another CPU then the task
1105 	 * might exit and p would dereference a non-existing task
1106 	 * struct. Prevent this by holding a reference on p across the
1107 	 * wake up.
1108 	 */
1109 	get_task_struct(p);
1110 
1111 	__unqueue_futex(q);
1112 	/*
1113 	 * The waiting task can free the futex_q as soon as
1114 	 * q->lock_ptr = NULL is written, without taking any locks. A
1115 	 * memory barrier is required here to prevent the following
1116 	 * store to lock_ptr from getting ahead of the plist_del.
1117 	 */
1118 	smp_wmb();
1119 	q->lock_ptr = NULL;
1120 
1121 	wake_up_state(p, TASK_NORMAL);
1122 	put_task_struct(p);
1123 }
1124 
wake_futex_pi(u32 __user * uaddr,u32 uval,struct futex_q * this)1125 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this)
1126 {
1127 	struct task_struct *new_owner;
1128 	struct futex_pi_state *pi_state = this->pi_state;
1129 	u32 uninitialized_var(curval), newval;
1130 	int ret = 0;
1131 
1132 	if (!pi_state)
1133 		return -EINVAL;
1134 
1135 	/*
1136 	 * If current does not own the pi_state then the futex is
1137 	 * inconsistent and user space fiddled with the futex value.
1138 	 */
1139 	if (pi_state->owner != current)
1140 		return -EINVAL;
1141 
1142 	raw_spin_lock(&pi_state->pi_mutex.wait_lock);
1143 	new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1144 
1145 	/*
1146 	 * It is possible that the next waiter (the one that brought
1147 	 * this owner to the kernel) timed out and is no longer
1148 	 * waiting on the lock.
1149 	 */
1150 	if (!new_owner)
1151 		new_owner = this->task;
1152 
1153 	/*
1154 	 * We pass it to the next owner. The WAITERS bit is always
1155 	 * kept enabled while there is PI state around. We cleanup the
1156 	 * owner died bit, because we are the owner.
1157 	 */
1158 	newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1159 
1160 	if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
1161 		ret = -EFAULT;
1162 	} else if (curval != uval) {
1163 		/*
1164 		 * If a unconditional UNLOCK_PI operation (user space did not
1165 		 * try the TID->0 transition) raced with a waiter setting the
1166 		 * FUTEX_WAITERS flag between get_user() and locking the hash
1167 		 * bucket lock, retry the operation.
1168 		 */
1169 		if ((FUTEX_TID_MASK & curval) == uval)
1170 			ret = -EAGAIN;
1171 		else
1172 			ret = -EINVAL;
1173 	}
1174 	if (ret) {
1175 		raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1176 		return ret;
1177 	}
1178 
1179 	raw_spin_lock_irq(&pi_state->owner->pi_lock);
1180 	WARN_ON(list_empty(&pi_state->list));
1181 	list_del_init(&pi_state->list);
1182 	raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1183 
1184 	raw_spin_lock_irq(&new_owner->pi_lock);
1185 	WARN_ON(!list_empty(&pi_state->list));
1186 	list_add(&pi_state->list, &new_owner->pi_state_list);
1187 	pi_state->owner = new_owner;
1188 	raw_spin_unlock_irq(&new_owner->pi_lock);
1189 
1190 	raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
1191 	rt_mutex_unlock(&pi_state->pi_mutex);
1192 
1193 	return 0;
1194 }
1195 
1196 /*
1197  * Express the locking dependencies for lockdep:
1198  */
1199 static inline void
double_lock_hb(struct futex_hash_bucket * hb1,struct futex_hash_bucket * hb2)1200 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1201 {
1202 	if (hb1 <= hb2) {
1203 		spin_lock(&hb1->lock);
1204 		if (hb1 < hb2)
1205 			spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1206 	} else { /* hb1 > hb2 */
1207 		spin_lock(&hb2->lock);
1208 		spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1209 	}
1210 }
1211 
1212 static inline void
double_unlock_hb(struct futex_hash_bucket * hb1,struct futex_hash_bucket * hb2)1213 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1214 {
1215 	spin_unlock(&hb1->lock);
1216 	if (hb1 != hb2)
1217 		spin_unlock(&hb2->lock);
1218 }
1219 
1220 /*
1221  * Wake up waiters matching bitset queued on this futex (uaddr).
1222  */
1223 static int
futex_wake(u32 __user * uaddr,unsigned int flags,int nr_wake,u32 bitset)1224 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1225 {
1226 	struct futex_hash_bucket *hb;
1227 	struct futex_q *this, *next;
1228 	union futex_key key = FUTEX_KEY_INIT;
1229 	int ret;
1230 
1231 	if (!bitset)
1232 		return -EINVAL;
1233 
1234 	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
1235 	if (unlikely(ret != 0))
1236 		goto out;
1237 
1238 	hb = hash_futex(&key);
1239 
1240 	/* Make sure we really have tasks to wakeup */
1241 	if (!hb_waiters_pending(hb))
1242 		goto out_put_key;
1243 
1244 	spin_lock(&hb->lock);
1245 
1246 	plist_for_each_entry_safe(this, next, &hb->chain, list) {
1247 		if (match_futex (&this->key, &key)) {
1248 			if (this->pi_state || this->rt_waiter) {
1249 				ret = -EINVAL;
1250 				break;
1251 			}
1252 
1253 			/* Check if one of the bits is set in both bitsets */
1254 			if (!(this->bitset & bitset))
1255 				continue;
1256 
1257 			wake_futex(this);
1258 			if (++ret >= nr_wake)
1259 				break;
1260 		}
1261 	}
1262 
1263 	spin_unlock(&hb->lock);
1264 out_put_key:
1265 	put_futex_key(&key);
1266 out:
1267 	return ret;
1268 }
1269 
1270 /*
1271  * Wake up all waiters hashed on the physical page that is mapped
1272  * to this virtual address:
1273  */
1274 static int
futex_wake_op(u32 __user * uaddr1,unsigned int flags,u32 __user * uaddr2,int nr_wake,int nr_wake2,int op)1275 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1276 	      int nr_wake, int nr_wake2, int op)
1277 {
1278 	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1279 	struct futex_hash_bucket *hb1, *hb2;
1280 	struct futex_q *this, *next;
1281 	int ret, op_ret;
1282 
1283 retry:
1284 	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1285 	if (unlikely(ret != 0))
1286 		goto out;
1287 	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
1288 	if (unlikely(ret != 0))
1289 		goto out_put_key1;
1290 
1291 	hb1 = hash_futex(&key1);
1292 	hb2 = hash_futex(&key2);
1293 
1294 retry_private:
1295 	double_lock_hb(hb1, hb2);
1296 	op_ret = futex_atomic_op_inuser(op, uaddr2);
1297 	if (unlikely(op_ret < 0)) {
1298 
1299 		double_unlock_hb(hb1, hb2);
1300 
1301 #ifndef CONFIG_MMU
1302 		/*
1303 		 * we don't get EFAULT from MMU faults if we don't have an MMU,
1304 		 * but we might get them from range checking
1305 		 */
1306 		ret = op_ret;
1307 		goto out_put_keys;
1308 #endif
1309 
1310 		if (unlikely(op_ret != -EFAULT)) {
1311 			ret = op_ret;
1312 			goto out_put_keys;
1313 		}
1314 
1315 		ret = fault_in_user_writeable(uaddr2);
1316 		if (ret)
1317 			goto out_put_keys;
1318 
1319 		if (!(flags & FLAGS_SHARED))
1320 			goto retry_private;
1321 
1322 		put_futex_key(&key2);
1323 		put_futex_key(&key1);
1324 		goto retry;
1325 	}
1326 
1327 	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1328 		if (match_futex (&this->key, &key1)) {
1329 			if (this->pi_state || this->rt_waiter) {
1330 				ret = -EINVAL;
1331 				goto out_unlock;
1332 			}
1333 			wake_futex(this);
1334 			if (++ret >= nr_wake)
1335 				break;
1336 		}
1337 	}
1338 
1339 	if (op_ret > 0) {
1340 		op_ret = 0;
1341 		plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1342 			if (match_futex (&this->key, &key2)) {
1343 				if (this->pi_state || this->rt_waiter) {
1344 					ret = -EINVAL;
1345 					goto out_unlock;
1346 				}
1347 				wake_futex(this);
1348 				if (++op_ret >= nr_wake2)
1349 					break;
1350 			}
1351 		}
1352 		ret += op_ret;
1353 	}
1354 
1355 out_unlock:
1356 	double_unlock_hb(hb1, hb2);
1357 out_put_keys:
1358 	put_futex_key(&key2);
1359 out_put_key1:
1360 	put_futex_key(&key1);
1361 out:
1362 	return ret;
1363 }
1364 
1365 /**
1366  * requeue_futex() - Requeue a futex_q from one hb to another
1367  * @q:		the futex_q to requeue
1368  * @hb1:	the source hash_bucket
1369  * @hb2:	the target hash_bucket
1370  * @key2:	the new key for the requeued futex_q
1371  */
1372 static inline
requeue_futex(struct futex_q * q,struct futex_hash_bucket * hb1,struct futex_hash_bucket * hb2,union futex_key * key2)1373 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1374 		   struct futex_hash_bucket *hb2, union futex_key *key2)
1375 {
1376 
1377 	/*
1378 	 * If key1 and key2 hash to the same bucket, no need to
1379 	 * requeue.
1380 	 */
1381 	if (likely(&hb1->chain != &hb2->chain)) {
1382 		plist_del(&q->list, &hb1->chain);
1383 		hb_waiters_dec(hb1);
1384 		hb_waiters_inc(hb2);
1385 		plist_add(&q->list, &hb2->chain);
1386 		q->lock_ptr = &hb2->lock;
1387 	}
1388 	get_futex_key_refs(key2);
1389 	q->key = *key2;
1390 }
1391 
1392 /**
1393  * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1394  * @q:		the futex_q
1395  * @key:	the key of the requeue target futex
1396  * @hb:		the hash_bucket of the requeue target futex
1397  *
1398  * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1399  * target futex if it is uncontended or via a lock steal.  Set the futex_q key
1400  * to the requeue target futex so the waiter can detect the wakeup on the right
1401  * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1402  * atomic lock acquisition.  Set the q->lock_ptr to the requeue target hb->lock
1403  * to protect access to the pi_state to fixup the owner later.  Must be called
1404  * with both q->lock_ptr and hb->lock held.
1405  */
1406 static inline
requeue_pi_wake_futex(struct futex_q * q,union futex_key * key,struct futex_hash_bucket * hb)1407 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1408 			   struct futex_hash_bucket *hb)
1409 {
1410 	get_futex_key_refs(key);
1411 	q->key = *key;
1412 
1413 	__unqueue_futex(q);
1414 
1415 	WARN_ON(!q->rt_waiter);
1416 	q->rt_waiter = NULL;
1417 
1418 	q->lock_ptr = &hb->lock;
1419 
1420 	wake_up_state(q->task, TASK_NORMAL);
1421 }
1422 
1423 /**
1424  * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1425  * @pifutex:		the user address of the to futex
1426  * @hb1:		the from futex hash bucket, must be locked by the caller
1427  * @hb2:		the to futex hash bucket, must be locked by the caller
1428  * @key1:		the from futex key
1429  * @key2:		the to futex key
1430  * @ps:			address to store the pi_state pointer
1431  * @set_waiters:	force setting the FUTEX_WAITERS bit (1) or not (0)
1432  *
1433  * Try and get the lock on behalf of the top waiter if we can do it atomically.
1434  * Wake the top waiter if we succeed.  If the caller specified set_waiters,
1435  * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1436  * hb1 and hb2 must be held by the caller.
1437  *
1438  * Return:
1439  *  0 - failed to acquire the lock atomically;
1440  * >0 - acquired the lock, return value is vpid of the top_waiter
1441  * <0 - error
1442  */
futex_proxy_trylock_atomic(u32 __user * pifutex,struct futex_hash_bucket * hb1,struct futex_hash_bucket * hb2,union futex_key * key1,union futex_key * key2,struct futex_pi_state ** ps,int set_waiters)1443 static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1444 				 struct futex_hash_bucket *hb1,
1445 				 struct futex_hash_bucket *hb2,
1446 				 union futex_key *key1, union futex_key *key2,
1447 				 struct futex_pi_state **ps, int set_waiters)
1448 {
1449 	struct futex_q *top_waiter = NULL;
1450 	u32 curval;
1451 	int ret, vpid;
1452 
1453 	if (get_futex_value_locked(&curval, pifutex))
1454 		return -EFAULT;
1455 
1456 	/*
1457 	 * Find the top_waiter and determine if there are additional waiters.
1458 	 * If the caller intends to requeue more than 1 waiter to pifutex,
1459 	 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1460 	 * as we have means to handle the possible fault.  If not, don't set
1461 	 * the bit unecessarily as it will force the subsequent unlock to enter
1462 	 * the kernel.
1463 	 */
1464 	top_waiter = futex_top_waiter(hb1, key1);
1465 
1466 	/* There are no waiters, nothing for us to do. */
1467 	if (!top_waiter)
1468 		return 0;
1469 
1470 	/* Ensure we requeue to the expected futex. */
1471 	if (!match_futex(top_waiter->requeue_pi_key, key2))
1472 		return -EINVAL;
1473 
1474 	/*
1475 	 * Try to take the lock for top_waiter.  Set the FUTEX_WAITERS bit in
1476 	 * the contended case or if set_waiters is 1.  The pi_state is returned
1477 	 * in ps in contended cases.
1478 	 */
1479 	vpid = task_pid_vnr(top_waiter->task);
1480 	ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1481 				   set_waiters);
1482 	if (ret == 1) {
1483 		requeue_pi_wake_futex(top_waiter, key2, hb2);
1484 		return vpid;
1485 	}
1486 	return ret;
1487 }
1488 
1489 /**
1490  * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1491  * @uaddr1:	source futex user address
1492  * @flags:	futex flags (FLAGS_SHARED, etc.)
1493  * @uaddr2:	target futex user address
1494  * @nr_wake:	number of waiters to wake (must be 1 for requeue_pi)
1495  * @nr_requeue:	number of waiters to requeue (0-INT_MAX)
1496  * @cmpval:	@uaddr1 expected value (or %NULL)
1497  * @requeue_pi:	if we are attempting to requeue from a non-pi futex to a
1498  *		pi futex (pi to pi requeue is not supported)
1499  *
1500  * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1501  * uaddr2 atomically on behalf of the top waiter.
1502  *
1503  * Return:
1504  * >=0 - on success, the number of tasks requeued or woken;
1505  *  <0 - on error
1506  */
futex_requeue(u32 __user * uaddr1,unsigned int flags,u32 __user * uaddr2,int nr_wake,int nr_requeue,u32 * cmpval,int requeue_pi)1507 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1508 			 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1509 			 u32 *cmpval, int requeue_pi)
1510 {
1511 	union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1512 	int drop_count = 0, task_count = 0, ret;
1513 	struct futex_pi_state *pi_state = NULL;
1514 	struct futex_hash_bucket *hb1, *hb2;
1515 	struct futex_q *this, *next;
1516 
1517 	if (nr_wake < 0 || nr_requeue < 0)
1518 		return -EINVAL;
1519 
1520 	if (requeue_pi) {
1521 		/*
1522 		 * Requeue PI only works on two distinct uaddrs. This
1523 		 * check is only valid for private futexes. See below.
1524 		 */
1525 		if (uaddr1 == uaddr2)
1526 			return -EINVAL;
1527 
1528 		/*
1529 		 * requeue_pi requires a pi_state, try to allocate it now
1530 		 * without any locks in case it fails.
1531 		 */
1532 		if (refill_pi_state_cache())
1533 			return -ENOMEM;
1534 		/*
1535 		 * requeue_pi must wake as many tasks as it can, up to nr_wake
1536 		 * + nr_requeue, since it acquires the rt_mutex prior to
1537 		 * returning to userspace, so as to not leave the rt_mutex with
1538 		 * waiters and no owner.  However, second and third wake-ups
1539 		 * cannot be predicted as they involve race conditions with the
1540 		 * first wake and a fault while looking up the pi_state.  Both
1541 		 * pthread_cond_signal() and pthread_cond_broadcast() should
1542 		 * use nr_wake=1.
1543 		 */
1544 		if (nr_wake != 1)
1545 			return -EINVAL;
1546 	}
1547 
1548 retry:
1549 	ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
1550 	if (unlikely(ret != 0))
1551 		goto out;
1552 	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1553 			    requeue_pi ? VERIFY_WRITE : VERIFY_READ);
1554 	if (unlikely(ret != 0))
1555 		goto out_put_key1;
1556 
1557 	/*
1558 	 * The check above which compares uaddrs is not sufficient for
1559 	 * shared futexes. We need to compare the keys:
1560 	 */
1561 	if (requeue_pi && match_futex(&key1, &key2)) {
1562 		ret = -EINVAL;
1563 		goto out_put_keys;
1564 	}
1565 
1566 	hb1 = hash_futex(&key1);
1567 	hb2 = hash_futex(&key2);
1568 
1569 retry_private:
1570 	hb_waiters_inc(hb2);
1571 	double_lock_hb(hb1, hb2);
1572 
1573 	if (likely(cmpval != NULL)) {
1574 		u32 curval;
1575 
1576 		ret = get_futex_value_locked(&curval, uaddr1);
1577 
1578 		if (unlikely(ret)) {
1579 			double_unlock_hb(hb1, hb2);
1580 			hb_waiters_dec(hb2);
1581 
1582 			ret = get_user(curval, uaddr1);
1583 			if (ret)
1584 				goto out_put_keys;
1585 
1586 			if (!(flags & FLAGS_SHARED))
1587 				goto retry_private;
1588 
1589 			put_futex_key(&key2);
1590 			put_futex_key(&key1);
1591 			goto retry;
1592 		}
1593 		if (curval != *cmpval) {
1594 			ret = -EAGAIN;
1595 			goto out_unlock;
1596 		}
1597 	}
1598 
1599 	if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
1600 		/*
1601 		 * Attempt to acquire uaddr2 and wake the top waiter. If we
1602 		 * intend to requeue waiters, force setting the FUTEX_WAITERS
1603 		 * bit.  We force this here where we are able to easily handle
1604 		 * faults rather in the requeue loop below.
1605 		 */
1606 		ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
1607 						 &key2, &pi_state, nr_requeue);
1608 
1609 		/*
1610 		 * At this point the top_waiter has either taken uaddr2 or is
1611 		 * waiting on it.  If the former, then the pi_state will not
1612 		 * exist yet, look it up one more time to ensure we have a
1613 		 * reference to it. If the lock was taken, ret contains the
1614 		 * vpid of the top waiter task.
1615 		 */
1616 		if (ret > 0) {
1617 			WARN_ON(pi_state);
1618 			drop_count++;
1619 			task_count++;
1620 			/*
1621 			 * If we acquired the lock, then the user
1622 			 * space value of uaddr2 should be vpid. It
1623 			 * cannot be changed by the top waiter as it
1624 			 * is blocked on hb2 lock if it tries to do
1625 			 * so. If something fiddled with it behind our
1626 			 * back the pi state lookup might unearth
1627 			 * it. So we rather use the known value than
1628 			 * rereading and handing potential crap to
1629 			 * lookup_pi_state.
1630 			 */
1631 			ret = lookup_pi_state(ret, hb2, &key2, &pi_state);
1632 		}
1633 
1634 		switch (ret) {
1635 		case 0:
1636 			break;
1637 		case -EFAULT:
1638 			free_pi_state(pi_state);
1639 			pi_state = NULL;
1640 			double_unlock_hb(hb1, hb2);
1641 			hb_waiters_dec(hb2);
1642 			put_futex_key(&key2);
1643 			put_futex_key(&key1);
1644 			ret = fault_in_user_writeable(uaddr2);
1645 			if (!ret)
1646 				goto retry;
1647 			goto out;
1648 		case -EAGAIN:
1649 			/*
1650 			 * Two reasons for this:
1651 			 * - Owner is exiting and we just wait for the
1652 			 *   exit to complete.
1653 			 * - The user space value changed.
1654 			 */
1655 			free_pi_state(pi_state);
1656 			pi_state = NULL;
1657 			double_unlock_hb(hb1, hb2);
1658 			hb_waiters_dec(hb2);
1659 			put_futex_key(&key2);
1660 			put_futex_key(&key1);
1661 			cond_resched();
1662 			goto retry;
1663 		default:
1664 			goto out_unlock;
1665 		}
1666 	}
1667 
1668 	plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1669 		if (task_count - nr_wake >= nr_requeue)
1670 			break;
1671 
1672 		if (!match_futex(&this->key, &key1))
1673 			continue;
1674 
1675 		/*
1676 		 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1677 		 * be paired with each other and no other futex ops.
1678 		 *
1679 		 * We should never be requeueing a futex_q with a pi_state,
1680 		 * which is awaiting a futex_unlock_pi().
1681 		 */
1682 		if ((requeue_pi && !this->rt_waiter) ||
1683 		    (!requeue_pi && this->rt_waiter) ||
1684 		    this->pi_state) {
1685 			ret = -EINVAL;
1686 			break;
1687 		}
1688 
1689 		/*
1690 		 * Wake nr_wake waiters.  For requeue_pi, if we acquired the
1691 		 * lock, we already woke the top_waiter.  If not, it will be
1692 		 * woken by futex_unlock_pi().
1693 		 */
1694 		if (++task_count <= nr_wake && !requeue_pi) {
1695 			wake_futex(this);
1696 			continue;
1697 		}
1698 
1699 		/* Ensure we requeue to the expected futex for requeue_pi. */
1700 		if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1701 			ret = -EINVAL;
1702 			break;
1703 		}
1704 
1705 		/*
1706 		 * Requeue nr_requeue waiters and possibly one more in the case
1707 		 * of requeue_pi if we couldn't acquire the lock atomically.
1708 		 */
1709 		if (requeue_pi) {
1710 			/* Prepare the waiter to take the rt_mutex. */
1711 			atomic_inc(&pi_state->refcount);
1712 			this->pi_state = pi_state;
1713 			ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1714 							this->rt_waiter,
1715 							this->task);
1716 			if (ret == 1) {
1717 				/* We got the lock. */
1718 				requeue_pi_wake_futex(this, &key2, hb2);
1719 				drop_count++;
1720 				continue;
1721 			} else if (ret) {
1722 				/* -EDEADLK */
1723 				this->pi_state = NULL;
1724 				free_pi_state(pi_state);
1725 				goto out_unlock;
1726 			}
1727 		}
1728 		requeue_futex(this, hb1, hb2, &key2);
1729 		drop_count++;
1730 	}
1731 
1732 out_unlock:
1733 	free_pi_state(pi_state);
1734 	double_unlock_hb(hb1, hb2);
1735 	hb_waiters_dec(hb2);
1736 
1737 	/*
1738 	 * drop_futex_key_refs() must be called outside the spinlocks. During
1739 	 * the requeue we moved futex_q's from the hash bucket at key1 to the
1740 	 * one at key2 and updated their key pointer.  We no longer need to
1741 	 * hold the references to key1.
1742 	 */
1743 	while (--drop_count >= 0)
1744 		drop_futex_key_refs(&key1);
1745 
1746 out_put_keys:
1747 	put_futex_key(&key2);
1748 out_put_key1:
1749 	put_futex_key(&key1);
1750 out:
1751 	return ret ? ret : task_count;
1752 }
1753 
1754 /* The key must be already stored in q->key. */
queue_lock(struct futex_q * q)1755 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
1756 	__acquires(&hb->lock)
1757 {
1758 	struct futex_hash_bucket *hb;
1759 
1760 	hb = hash_futex(&q->key);
1761 
1762 	/*
1763 	 * Increment the counter before taking the lock so that
1764 	 * a potential waker won't miss a to-be-slept task that is
1765 	 * waiting for the spinlock. This is safe as all queue_lock()
1766 	 * users end up calling queue_me(). Similarly, for housekeeping,
1767 	 * decrement the counter at queue_unlock() when some error has
1768 	 * occurred and we don't end up adding the task to the list.
1769 	 */
1770 	hb_waiters_inc(hb);
1771 
1772 	q->lock_ptr = &hb->lock;
1773 
1774 	spin_lock(&hb->lock); /* implies MB (A) */
1775 	return hb;
1776 }
1777 
1778 static inline void
queue_unlock(struct futex_hash_bucket * hb)1779 queue_unlock(struct futex_hash_bucket *hb)
1780 	__releases(&hb->lock)
1781 {
1782 	spin_unlock(&hb->lock);
1783 	hb_waiters_dec(hb);
1784 }
1785 
1786 /**
1787  * queue_me() - Enqueue the futex_q on the futex_hash_bucket
1788  * @q:	The futex_q to enqueue
1789  * @hb:	The destination hash bucket
1790  *
1791  * The hb->lock must be held by the caller, and is released here. A call to
1792  * queue_me() is typically paired with exactly one call to unqueue_me().  The
1793  * exceptions involve the PI related operations, which may use unqueue_me_pi()
1794  * or nothing if the unqueue is done as part of the wake process and the unqueue
1795  * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
1796  * an example).
1797  */
queue_me(struct futex_q * q,struct futex_hash_bucket * hb)1798 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
1799 	__releases(&hb->lock)
1800 {
1801 	int prio;
1802 
1803 	/*
1804 	 * The priority used to register this element is
1805 	 * - either the real thread-priority for the real-time threads
1806 	 * (i.e. threads with a priority lower than MAX_RT_PRIO)
1807 	 * - or MAX_RT_PRIO for non-RT threads.
1808 	 * Thus, all RT-threads are woken first in priority order, and
1809 	 * the others are woken last, in FIFO order.
1810 	 */
1811 	prio = min(current->normal_prio, MAX_RT_PRIO);
1812 
1813 	plist_node_init(&q->list, prio);
1814 	plist_add(&q->list, &hb->chain);
1815 	q->task = current;
1816 	spin_unlock(&hb->lock);
1817 }
1818 
1819 /**
1820  * unqueue_me() - Remove the futex_q from its futex_hash_bucket
1821  * @q:	The futex_q to unqueue
1822  *
1823  * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
1824  * be paired with exactly one earlier call to queue_me().
1825  *
1826  * Return:
1827  *   1 - if the futex_q was still queued (and we removed unqueued it);
1828  *   0 - if the futex_q was already removed by the waking thread
1829  */
unqueue_me(struct futex_q * q)1830 static int unqueue_me(struct futex_q *q)
1831 {
1832 	spinlock_t *lock_ptr;
1833 	int ret = 0;
1834 
1835 	/* In the common case we don't take the spinlock, which is nice. */
1836 retry:
1837 	lock_ptr = q->lock_ptr;
1838 	barrier();
1839 	if (lock_ptr != NULL) {
1840 		spin_lock(lock_ptr);
1841 		/*
1842 		 * q->lock_ptr can change between reading it and
1843 		 * spin_lock(), causing us to take the wrong lock.  This
1844 		 * corrects the race condition.
1845 		 *
1846 		 * Reasoning goes like this: if we have the wrong lock,
1847 		 * q->lock_ptr must have changed (maybe several times)
1848 		 * between reading it and the spin_lock().  It can
1849 		 * change again after the spin_lock() but only if it was
1850 		 * already changed before the spin_lock().  It cannot,
1851 		 * however, change back to the original value.  Therefore
1852 		 * we can detect whether we acquired the correct lock.
1853 		 */
1854 		if (unlikely(lock_ptr != q->lock_ptr)) {
1855 			spin_unlock(lock_ptr);
1856 			goto retry;
1857 		}
1858 		__unqueue_futex(q);
1859 
1860 		BUG_ON(q->pi_state);
1861 
1862 		spin_unlock(lock_ptr);
1863 		ret = 1;
1864 	}
1865 
1866 	drop_futex_key_refs(&q->key);
1867 	return ret;
1868 }
1869 
1870 /*
1871  * PI futexes can not be requeued and must remove themself from the
1872  * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
1873  * and dropped here.
1874  */
unqueue_me_pi(struct futex_q * q)1875 static void unqueue_me_pi(struct futex_q *q)
1876 	__releases(q->lock_ptr)
1877 {
1878 	__unqueue_futex(q);
1879 
1880 	BUG_ON(!q->pi_state);
1881 	free_pi_state(q->pi_state);
1882 	q->pi_state = NULL;
1883 
1884 	spin_unlock(q->lock_ptr);
1885 }
1886 
1887 /*
1888  * Fixup the pi_state owner with the new owner.
1889  *
1890  * Must be called with hash bucket lock held and mm->sem held for non
1891  * private futexes.
1892  */
fixup_pi_state_owner(u32 __user * uaddr,struct futex_q * q,struct task_struct * newowner)1893 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
1894 				struct task_struct *newowner)
1895 {
1896 	u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
1897 	struct futex_pi_state *pi_state = q->pi_state;
1898 	struct task_struct *oldowner = pi_state->owner;
1899 	u32 uval, uninitialized_var(curval), newval;
1900 	int ret;
1901 
1902 	/* Owner died? */
1903 	if (!pi_state->owner)
1904 		newtid |= FUTEX_OWNER_DIED;
1905 
1906 	/*
1907 	 * We are here either because we stole the rtmutex from the
1908 	 * previous highest priority waiter or we are the highest priority
1909 	 * waiter but failed to get the rtmutex the first time.
1910 	 * We have to replace the newowner TID in the user space variable.
1911 	 * This must be atomic as we have to preserve the owner died bit here.
1912 	 *
1913 	 * Note: We write the user space value _before_ changing the pi_state
1914 	 * because we can fault here. Imagine swapped out pages or a fork
1915 	 * that marked all the anonymous memory readonly for cow.
1916 	 *
1917 	 * Modifying pi_state _before_ the user space value would
1918 	 * leave the pi_state in an inconsistent state when we fault
1919 	 * here, because we need to drop the hash bucket lock to
1920 	 * handle the fault. This might be observed in the PID check
1921 	 * in lookup_pi_state.
1922 	 */
1923 retry:
1924 	if (get_futex_value_locked(&uval, uaddr))
1925 		goto handle_fault;
1926 
1927 	while (1) {
1928 		newval = (uval & FUTEX_OWNER_DIED) | newtid;
1929 
1930 		if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
1931 			goto handle_fault;
1932 		if (curval == uval)
1933 			break;
1934 		uval = curval;
1935 	}
1936 
1937 	/*
1938 	 * We fixed up user space. Now we need to fix the pi_state
1939 	 * itself.
1940 	 */
1941 	if (pi_state->owner != NULL) {
1942 		raw_spin_lock_irq(&pi_state->owner->pi_lock);
1943 		WARN_ON(list_empty(&pi_state->list));
1944 		list_del_init(&pi_state->list);
1945 		raw_spin_unlock_irq(&pi_state->owner->pi_lock);
1946 	}
1947 
1948 	pi_state->owner = newowner;
1949 
1950 	raw_spin_lock_irq(&newowner->pi_lock);
1951 	WARN_ON(!list_empty(&pi_state->list));
1952 	list_add(&pi_state->list, &newowner->pi_state_list);
1953 	raw_spin_unlock_irq(&newowner->pi_lock);
1954 	return 0;
1955 
1956 	/*
1957 	 * To handle the page fault we need to drop the hash bucket
1958 	 * lock here. That gives the other task (either the highest priority
1959 	 * waiter itself or the task which stole the rtmutex) the
1960 	 * chance to try the fixup of the pi_state. So once we are
1961 	 * back from handling the fault we need to check the pi_state
1962 	 * after reacquiring the hash bucket lock and before trying to
1963 	 * do another fixup. When the fixup has been done already we
1964 	 * simply return.
1965 	 */
1966 handle_fault:
1967 	spin_unlock(q->lock_ptr);
1968 
1969 	ret = fault_in_user_writeable(uaddr);
1970 
1971 	spin_lock(q->lock_ptr);
1972 
1973 	/*
1974 	 * Check if someone else fixed it for us:
1975 	 */
1976 	if (pi_state->owner != oldowner)
1977 		return 0;
1978 
1979 	if (ret)
1980 		return ret;
1981 
1982 	goto retry;
1983 }
1984 
1985 static long futex_wait_restart(struct restart_block *restart);
1986 
1987 /**
1988  * fixup_owner() - Post lock pi_state and corner case management
1989  * @uaddr:	user address of the futex
1990  * @q:		futex_q (contains pi_state and access to the rt_mutex)
1991  * @locked:	if the attempt to take the rt_mutex succeeded (1) or not (0)
1992  *
1993  * After attempting to lock an rt_mutex, this function is called to cleanup
1994  * the pi_state owner as well as handle race conditions that may allow us to
1995  * acquire the lock. Must be called with the hb lock held.
1996  *
1997  * Return:
1998  *  1 - success, lock taken;
1999  *  0 - success, lock not taken;
2000  * <0 - on error (-EFAULT)
2001  */
fixup_owner(u32 __user * uaddr,struct futex_q * q,int locked)2002 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2003 {
2004 	struct task_struct *owner;
2005 	int ret = 0;
2006 
2007 	if (locked) {
2008 		/*
2009 		 * Got the lock. We might not be the anticipated owner if we
2010 		 * did a lock-steal - fix up the PI-state in that case:
2011 		 */
2012 		if (q->pi_state->owner != current)
2013 			ret = fixup_pi_state_owner(uaddr, q, current);
2014 		goto out;
2015 	}
2016 
2017 	/*
2018 	 * Catch the rare case, where the lock was released when we were on the
2019 	 * way back before we locked the hash bucket.
2020 	 */
2021 	if (q->pi_state->owner == current) {
2022 		/*
2023 		 * Try to get the rt_mutex now. This might fail as some other
2024 		 * task acquired the rt_mutex after we removed ourself from the
2025 		 * rt_mutex waiters list.
2026 		 */
2027 		if (rt_mutex_trylock(&q->pi_state->pi_mutex)) {
2028 			locked = 1;
2029 			goto out;
2030 		}
2031 
2032 		/*
2033 		 * pi_state is incorrect, some other task did a lock steal and
2034 		 * we returned due to timeout or signal without taking the
2035 		 * rt_mutex. Too late.
2036 		 */
2037 		raw_spin_lock(&q->pi_state->pi_mutex.wait_lock);
2038 		owner = rt_mutex_owner(&q->pi_state->pi_mutex);
2039 		if (!owner)
2040 			owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
2041 		raw_spin_unlock(&q->pi_state->pi_mutex.wait_lock);
2042 		ret = fixup_pi_state_owner(uaddr, q, owner);
2043 		goto out;
2044 	}
2045 
2046 	/*
2047 	 * Paranoia check. If we did not take the lock, then we should not be
2048 	 * the owner of the rt_mutex.
2049 	 */
2050 	if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2051 		printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2052 				"pi-state %p\n", ret,
2053 				q->pi_state->pi_mutex.owner,
2054 				q->pi_state->owner);
2055 
2056 out:
2057 	return ret ? ret : locked;
2058 }
2059 
2060 /**
2061  * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2062  * @hb:		the futex hash bucket, must be locked by the caller
2063  * @q:		the futex_q to queue up on
2064  * @timeout:	the prepared hrtimer_sleeper, or null for no timeout
2065  */
futex_wait_queue_me(struct futex_hash_bucket * hb,struct futex_q * q,struct hrtimer_sleeper * timeout)2066 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2067 				struct hrtimer_sleeper *timeout)
2068 {
2069 	/*
2070 	 * The task state is guaranteed to be set before another task can
2071 	 * wake it. set_current_state() is implemented using set_mb() and
2072 	 * queue_me() calls spin_unlock() upon completion, both serializing
2073 	 * access to the hash list and forcing another memory barrier.
2074 	 */
2075 	set_current_state(TASK_INTERRUPTIBLE);
2076 	queue_me(q, hb);
2077 
2078 	/* Arm the timer */
2079 	if (timeout) {
2080 		hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
2081 		if (!hrtimer_active(&timeout->timer))
2082 			timeout->task = NULL;
2083 	}
2084 
2085 	/*
2086 	 * If we have been removed from the hash list, then another task
2087 	 * has tried to wake us, and we can skip the call to schedule().
2088 	 */
2089 	if (likely(!plist_node_empty(&q->list))) {
2090 		/*
2091 		 * If the timer has already expired, current will already be
2092 		 * flagged for rescheduling. Only call schedule if there
2093 		 * is no timeout, or if it has yet to expire.
2094 		 */
2095 		if (!timeout || timeout->task)
2096 			freezable_schedule();
2097 	}
2098 	__set_current_state(TASK_RUNNING);
2099 }
2100 
2101 /**
2102  * futex_wait_setup() - Prepare to wait on a futex
2103  * @uaddr:	the futex userspace address
2104  * @val:	the expected value
2105  * @flags:	futex flags (FLAGS_SHARED, etc.)
2106  * @q:		the associated futex_q
2107  * @hb:		storage for hash_bucket pointer to be returned to caller
2108  *
2109  * Setup the futex_q and locate the hash_bucket.  Get the futex value and
2110  * compare it with the expected value.  Handle atomic faults internally.
2111  * Return with the hb lock held and a q.key reference on success, and unlocked
2112  * with no q.key reference on failure.
2113  *
2114  * Return:
2115  *  0 - uaddr contains val and hb has been locked;
2116  * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2117  */
futex_wait_setup(u32 __user * uaddr,u32 val,unsigned int flags,struct futex_q * q,struct futex_hash_bucket ** hb)2118 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2119 			   struct futex_q *q, struct futex_hash_bucket **hb)
2120 {
2121 	u32 uval;
2122 	int ret;
2123 
2124 	/*
2125 	 * Access the page AFTER the hash-bucket is locked.
2126 	 * Order is important:
2127 	 *
2128 	 *   Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2129 	 *   Userspace waker:  if (cond(var)) { var = new; futex_wake(&var); }
2130 	 *
2131 	 * The basic logical guarantee of a futex is that it blocks ONLY
2132 	 * if cond(var) is known to be true at the time of blocking, for
2133 	 * any cond.  If we locked the hash-bucket after testing *uaddr, that
2134 	 * would open a race condition where we could block indefinitely with
2135 	 * cond(var) false, which would violate the guarantee.
2136 	 *
2137 	 * On the other hand, we insert q and release the hash-bucket only
2138 	 * after testing *uaddr.  This guarantees that futex_wait() will NOT
2139 	 * absorb a wakeup if *uaddr does not match the desired values
2140 	 * while the syscall executes.
2141 	 */
2142 retry:
2143 	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
2144 	if (unlikely(ret != 0))
2145 		return ret;
2146 
2147 retry_private:
2148 	*hb = queue_lock(q);
2149 
2150 	ret = get_futex_value_locked(&uval, uaddr);
2151 
2152 	if (ret) {
2153 		queue_unlock(*hb);
2154 
2155 		ret = get_user(uval, uaddr);
2156 		if (ret)
2157 			goto out;
2158 
2159 		if (!(flags & FLAGS_SHARED))
2160 			goto retry_private;
2161 
2162 		put_futex_key(&q->key);
2163 		goto retry;
2164 	}
2165 
2166 	if (uval != val) {
2167 		queue_unlock(*hb);
2168 		ret = -EWOULDBLOCK;
2169 	}
2170 
2171 out:
2172 	if (ret)
2173 		put_futex_key(&q->key);
2174 	return ret;
2175 }
2176 
futex_wait(u32 __user * uaddr,unsigned int flags,u32 val,ktime_t * abs_time,u32 bitset)2177 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2178 		      ktime_t *abs_time, u32 bitset)
2179 {
2180 	struct hrtimer_sleeper timeout, *to = NULL;
2181 	struct restart_block *restart;
2182 	struct futex_hash_bucket *hb;
2183 	struct futex_q q = futex_q_init;
2184 	int ret;
2185 
2186 	if (!bitset)
2187 		return -EINVAL;
2188 	q.bitset = bitset;
2189 
2190 	if (abs_time) {
2191 		to = &timeout;
2192 
2193 		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2194 				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2195 				      HRTIMER_MODE_ABS);
2196 		hrtimer_init_sleeper(to, current);
2197 		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2198 					     current->timer_slack_ns);
2199 	}
2200 
2201 retry:
2202 	/*
2203 	 * Prepare to wait on uaddr. On success, holds hb lock and increments
2204 	 * q.key refs.
2205 	 */
2206 	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2207 	if (ret)
2208 		goto out;
2209 
2210 	/* queue_me and wait for wakeup, timeout, or a signal. */
2211 	futex_wait_queue_me(hb, &q, to);
2212 
2213 	/* If we were woken (and unqueued), we succeeded, whatever. */
2214 	ret = 0;
2215 	/* unqueue_me() drops q.key ref */
2216 	if (!unqueue_me(&q))
2217 		goto out;
2218 	ret = -ETIMEDOUT;
2219 	if (to && !to->task)
2220 		goto out;
2221 
2222 	/*
2223 	 * We expect signal_pending(current), but we might be the
2224 	 * victim of a spurious wakeup as well.
2225 	 */
2226 	if (!signal_pending(current))
2227 		goto retry;
2228 
2229 	ret = -ERESTARTSYS;
2230 	if (!abs_time)
2231 		goto out;
2232 
2233 	restart = &current->restart_block;
2234 	restart->fn = futex_wait_restart;
2235 	restart->futex.uaddr = uaddr;
2236 	restart->futex.val = val;
2237 	restart->futex.time = abs_time->tv64;
2238 	restart->futex.bitset = bitset;
2239 	restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2240 
2241 	ret = -ERESTART_RESTARTBLOCK;
2242 
2243 out:
2244 	if (to) {
2245 		hrtimer_cancel(&to->timer);
2246 		destroy_hrtimer_on_stack(&to->timer);
2247 	}
2248 	return ret;
2249 }
2250 
2251 
futex_wait_restart(struct restart_block * restart)2252 static long futex_wait_restart(struct restart_block *restart)
2253 {
2254 	u32 __user *uaddr = restart->futex.uaddr;
2255 	ktime_t t, *tp = NULL;
2256 
2257 	if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2258 		t.tv64 = restart->futex.time;
2259 		tp = &t;
2260 	}
2261 	restart->fn = do_no_restart_syscall;
2262 
2263 	return (long)futex_wait(uaddr, restart->futex.flags,
2264 				restart->futex.val, tp, restart->futex.bitset);
2265 }
2266 
2267 
2268 /*
2269  * Userspace tried a 0 -> TID atomic transition of the futex value
2270  * and failed. The kernel side here does the whole locking operation:
2271  * if there are waiters then it will block, it does PI, etc. (Due to
2272  * races the kernel might see a 0 value of the futex too.)
2273  */
futex_lock_pi(u32 __user * uaddr,unsigned int flags,int detect,ktime_t * time,int trylock)2274 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, int detect,
2275 			 ktime_t *time, int trylock)
2276 {
2277 	struct hrtimer_sleeper timeout, *to = NULL;
2278 	struct futex_hash_bucket *hb;
2279 	struct futex_q q = futex_q_init;
2280 	int res, ret;
2281 
2282 	if (refill_pi_state_cache())
2283 		return -ENOMEM;
2284 
2285 	if (time) {
2286 		to = &timeout;
2287 		hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2288 				      HRTIMER_MODE_ABS);
2289 		hrtimer_init_sleeper(to, current);
2290 		hrtimer_set_expires(&to->timer, *time);
2291 	}
2292 
2293 retry:
2294 	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
2295 	if (unlikely(ret != 0))
2296 		goto out;
2297 
2298 retry_private:
2299 	hb = queue_lock(&q);
2300 
2301 	ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
2302 	if (unlikely(ret)) {
2303 		switch (ret) {
2304 		case 1:
2305 			/* We got the lock. */
2306 			ret = 0;
2307 			goto out_unlock_put_key;
2308 		case -EFAULT:
2309 			goto uaddr_faulted;
2310 		case -EAGAIN:
2311 			/*
2312 			 * Two reasons for this:
2313 			 * - Task is exiting and we just wait for the
2314 			 *   exit to complete.
2315 			 * - The user space value changed.
2316 			 */
2317 			queue_unlock(hb);
2318 			put_futex_key(&q.key);
2319 			cond_resched();
2320 			goto retry;
2321 		default:
2322 			goto out_unlock_put_key;
2323 		}
2324 	}
2325 
2326 	/*
2327 	 * Only actually queue now that the atomic ops are done:
2328 	 */
2329 	queue_me(&q, hb);
2330 
2331 	WARN_ON(!q.pi_state);
2332 	/*
2333 	 * Block on the PI mutex:
2334 	 */
2335 	if (!trylock) {
2336 		ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2337 	} else {
2338 		ret = rt_mutex_trylock(&q.pi_state->pi_mutex);
2339 		/* Fixup the trylock return value: */
2340 		ret = ret ? 0 : -EWOULDBLOCK;
2341 	}
2342 
2343 	spin_lock(q.lock_ptr);
2344 	/*
2345 	 * Fixup the pi_state owner and possibly acquire the lock if we
2346 	 * haven't already.
2347 	 */
2348 	res = fixup_owner(uaddr, &q, !ret);
2349 	/*
2350 	 * If fixup_owner() returned an error, proprogate that.  If it acquired
2351 	 * the lock, clear our -ETIMEDOUT or -EINTR.
2352 	 */
2353 	if (res)
2354 		ret = (res < 0) ? res : 0;
2355 
2356 	/*
2357 	 * If fixup_owner() faulted and was unable to handle the fault, unlock
2358 	 * it and return the fault to userspace.
2359 	 */
2360 	if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
2361 		rt_mutex_unlock(&q.pi_state->pi_mutex);
2362 
2363 	/* Unqueue and drop the lock */
2364 	unqueue_me_pi(&q);
2365 
2366 	goto out_put_key;
2367 
2368 out_unlock_put_key:
2369 	queue_unlock(hb);
2370 
2371 out_put_key:
2372 	put_futex_key(&q.key);
2373 out:
2374 	if (to)
2375 		destroy_hrtimer_on_stack(&to->timer);
2376 	return ret != -EINTR ? ret : -ERESTARTNOINTR;
2377 
2378 uaddr_faulted:
2379 	queue_unlock(hb);
2380 
2381 	ret = fault_in_user_writeable(uaddr);
2382 	if (ret)
2383 		goto out_put_key;
2384 
2385 	if (!(flags & FLAGS_SHARED))
2386 		goto retry_private;
2387 
2388 	put_futex_key(&q.key);
2389 	goto retry;
2390 }
2391 
2392 /*
2393  * Userspace attempted a TID -> 0 atomic transition, and failed.
2394  * This is the in-kernel slowpath: we look up the PI state (if any),
2395  * and do the rt-mutex unlock.
2396  */
futex_unlock_pi(u32 __user * uaddr,unsigned int flags)2397 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2398 {
2399 	u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
2400 	union futex_key key = FUTEX_KEY_INIT;
2401 	struct futex_hash_bucket *hb;
2402 	struct futex_q *match;
2403 	int ret;
2404 
2405 retry:
2406 	if (get_user(uval, uaddr))
2407 		return -EFAULT;
2408 	/*
2409 	 * We release only a lock we actually own:
2410 	 */
2411 	if ((uval & FUTEX_TID_MASK) != vpid)
2412 		return -EPERM;
2413 
2414 	ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
2415 	if (ret)
2416 		return ret;
2417 
2418 	hb = hash_futex(&key);
2419 	spin_lock(&hb->lock);
2420 
2421 	/*
2422 	 * Check waiters first. We do not trust user space values at
2423 	 * all and we at least want to know if user space fiddled
2424 	 * with the futex value instead of blindly unlocking.
2425 	 */
2426 	match = futex_top_waiter(hb, &key);
2427 	if (match) {
2428 		ret = wake_futex_pi(uaddr, uval, match);
2429 		/*
2430 		 * The atomic access to the futex value generated a
2431 		 * pagefault, so retry the user-access and the wakeup:
2432 		 */
2433 		if (ret == -EFAULT)
2434 			goto pi_faulted;
2435 		/*
2436 		 * A unconditional UNLOCK_PI op raced against a waiter
2437 		 * setting the FUTEX_WAITERS bit. Try again.
2438 		 */
2439 		if (ret == -EAGAIN) {
2440 			spin_unlock(&hb->lock);
2441 			put_futex_key(&key);
2442 			goto retry;
2443 		}
2444 		goto out_unlock;
2445 	}
2446 
2447 	/*
2448 	 * We have no kernel internal state, i.e. no waiters in the
2449 	 * kernel. Waiters which are about to queue themselves are stuck
2450 	 * on hb->lock. So we can safely ignore them. We do neither
2451 	 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2452 	 * owner.
2453 	 */
2454 	if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
2455 		goto pi_faulted;
2456 
2457 	/*
2458 	 * If uval has changed, let user space handle it.
2459 	 */
2460 	ret = (curval == uval) ? 0 : -EAGAIN;
2461 
2462 out_unlock:
2463 	spin_unlock(&hb->lock);
2464 	put_futex_key(&key);
2465 	return ret;
2466 
2467 pi_faulted:
2468 	spin_unlock(&hb->lock);
2469 	put_futex_key(&key);
2470 
2471 	ret = fault_in_user_writeable(uaddr);
2472 	if (!ret)
2473 		goto retry;
2474 
2475 	return ret;
2476 }
2477 
2478 /**
2479  * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2480  * @hb:		the hash_bucket futex_q was original enqueued on
2481  * @q:		the futex_q woken while waiting to be requeued
2482  * @key2:	the futex_key of the requeue target futex
2483  * @timeout:	the timeout associated with the wait (NULL if none)
2484  *
2485  * Detect if the task was woken on the initial futex as opposed to the requeue
2486  * target futex.  If so, determine if it was a timeout or a signal that caused
2487  * the wakeup and return the appropriate error code to the caller.  Must be
2488  * called with the hb lock held.
2489  *
2490  * Return:
2491  *  0 = no early wakeup detected;
2492  * <0 = -ETIMEDOUT or -ERESTARTNOINTR
2493  */
2494 static inline
handle_early_requeue_pi_wakeup(struct futex_hash_bucket * hb,struct futex_q * q,union futex_key * key2,struct hrtimer_sleeper * timeout)2495 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2496 				   struct futex_q *q, union futex_key *key2,
2497 				   struct hrtimer_sleeper *timeout)
2498 {
2499 	int ret = 0;
2500 
2501 	/*
2502 	 * With the hb lock held, we avoid races while we process the wakeup.
2503 	 * We only need to hold hb (and not hb2) to ensure atomicity as the
2504 	 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2505 	 * It can't be requeued from uaddr2 to something else since we don't
2506 	 * support a PI aware source futex for requeue.
2507 	 */
2508 	if (!match_futex(&q->key, key2)) {
2509 		WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2510 		/*
2511 		 * We were woken prior to requeue by a timeout or a signal.
2512 		 * Unqueue the futex_q and determine which it was.
2513 		 */
2514 		plist_del(&q->list, &hb->chain);
2515 		hb_waiters_dec(hb);
2516 
2517 		/* Handle spurious wakeups gracefully */
2518 		ret = -EWOULDBLOCK;
2519 		if (timeout && !timeout->task)
2520 			ret = -ETIMEDOUT;
2521 		else if (signal_pending(current))
2522 			ret = -ERESTARTNOINTR;
2523 	}
2524 	return ret;
2525 }
2526 
2527 /**
2528  * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
2529  * @uaddr:	the futex we initially wait on (non-pi)
2530  * @flags:	futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
2531  * 		the same type, no requeueing from private to shared, etc.
2532  * @val:	the expected value of uaddr
2533  * @abs_time:	absolute timeout
2534  * @bitset:	32 bit wakeup bitset set by userspace, defaults to all
2535  * @uaddr2:	the pi futex we will take prior to returning to user-space
2536  *
2537  * The caller will wait on uaddr and will be requeued by futex_requeue() to
2538  * uaddr2 which must be PI aware and unique from uaddr.  Normal wakeup will wake
2539  * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2540  * userspace.  This ensures the rt_mutex maintains an owner when it has waiters;
2541  * without one, the pi logic would not know which task to boost/deboost, if
2542  * there was a need to.
2543  *
2544  * We call schedule in futex_wait_queue_me() when we enqueue and return there
2545  * via the following--
2546  * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
2547  * 2) wakeup on uaddr2 after a requeue
2548  * 3) signal
2549  * 4) timeout
2550  *
2551  * If 3, cleanup and return -ERESTARTNOINTR.
2552  *
2553  * If 2, we may then block on trying to take the rt_mutex and return via:
2554  * 5) successful lock
2555  * 6) signal
2556  * 7) timeout
2557  * 8) other lock acquisition failure
2558  *
2559  * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
2560  *
2561  * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2562  *
2563  * Return:
2564  *  0 - On success;
2565  * <0 - On error
2566  */
futex_wait_requeue_pi(u32 __user * uaddr,unsigned int flags,u32 val,ktime_t * abs_time,u32 bitset,u32 __user * uaddr2)2567 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
2568 				 u32 val, ktime_t *abs_time, u32 bitset,
2569 				 u32 __user *uaddr2)
2570 {
2571 	struct hrtimer_sleeper timeout, *to = NULL;
2572 	struct rt_mutex_waiter rt_waiter;
2573 	struct futex_hash_bucket *hb;
2574 	union futex_key key2 = FUTEX_KEY_INIT;
2575 	struct futex_q q = futex_q_init;
2576 	int res, ret;
2577 
2578 	if (uaddr == uaddr2)
2579 		return -EINVAL;
2580 
2581 	if (!bitset)
2582 		return -EINVAL;
2583 
2584 	if (abs_time) {
2585 		to = &timeout;
2586 		hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2587 				      CLOCK_REALTIME : CLOCK_MONOTONIC,
2588 				      HRTIMER_MODE_ABS);
2589 		hrtimer_init_sleeper(to, current);
2590 		hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2591 					     current->timer_slack_ns);
2592 	}
2593 
2594 	/*
2595 	 * The waiter is allocated on our stack, manipulated by the requeue
2596 	 * code while we sleep on uaddr.
2597 	 */
2598 	debug_rt_mutex_init_waiter(&rt_waiter);
2599 	RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2600 	RB_CLEAR_NODE(&rt_waiter.tree_entry);
2601 	rt_waiter.task = NULL;
2602 
2603 	ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
2604 	if (unlikely(ret != 0))
2605 		goto out;
2606 
2607 	q.bitset = bitset;
2608 	q.rt_waiter = &rt_waiter;
2609 	q.requeue_pi_key = &key2;
2610 
2611 	/*
2612 	 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2613 	 * count.
2614 	 */
2615 	ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2616 	if (ret)
2617 		goto out_key2;
2618 
2619 	/*
2620 	 * The check above which compares uaddrs is not sufficient for
2621 	 * shared futexes. We need to compare the keys:
2622 	 */
2623 	if (match_futex(&q.key, &key2)) {
2624 		queue_unlock(hb);
2625 		ret = -EINVAL;
2626 		goto out_put_keys;
2627 	}
2628 
2629 	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
2630 	futex_wait_queue_me(hb, &q, to);
2631 
2632 	spin_lock(&hb->lock);
2633 	ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2634 	spin_unlock(&hb->lock);
2635 	if (ret)
2636 		goto out_put_keys;
2637 
2638 	/*
2639 	 * In order for us to be here, we know our q.key == key2, and since
2640 	 * we took the hb->lock above, we also know that futex_requeue() has
2641 	 * completed and we no longer have to concern ourselves with a wakeup
2642 	 * race with the atomic proxy lock acquisition by the requeue code. The
2643 	 * futex_requeue dropped our key1 reference and incremented our key2
2644 	 * reference count.
2645 	 */
2646 
2647 	/* Check if the requeue code acquired the second futex for us. */
2648 	if (!q.rt_waiter) {
2649 		/*
2650 		 * Got the lock. We might not be the anticipated owner if we
2651 		 * did a lock-steal - fix up the PI-state in that case.
2652 		 */
2653 		if (q.pi_state && (q.pi_state->owner != current)) {
2654 			spin_lock(q.lock_ptr);
2655 			ret = fixup_pi_state_owner(uaddr2, &q, current);
2656 			if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current)
2657 				rt_mutex_unlock(&q.pi_state->pi_mutex);
2658 			/*
2659 			 * Drop the reference to the pi state which
2660 			 * the requeue_pi() code acquired for us.
2661 			 */
2662 			free_pi_state(q.pi_state);
2663 			spin_unlock(q.lock_ptr);
2664 		}
2665 	} else {
2666 		struct rt_mutex *pi_mutex;
2667 
2668 		/*
2669 		 * We have been woken up by futex_unlock_pi(), a timeout, or a
2670 		 * signal.  futex_unlock_pi() will not destroy the lock_ptr nor
2671 		 * the pi_state.
2672 		 */
2673 		WARN_ON(!q.pi_state);
2674 		pi_mutex = &q.pi_state->pi_mutex;
2675 		ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
2676 		debug_rt_mutex_free_waiter(&rt_waiter);
2677 
2678 		spin_lock(q.lock_ptr);
2679 		/*
2680 		 * Fixup the pi_state owner and possibly acquire the lock if we
2681 		 * haven't already.
2682 		 */
2683 		res = fixup_owner(uaddr2, &q, !ret);
2684 		/*
2685 		 * If fixup_owner() returned an error, proprogate that.  If it
2686 		 * acquired the lock, clear -ETIMEDOUT or -EINTR.
2687 		 */
2688 		if (res)
2689 			ret = (res < 0) ? res : 0;
2690 
2691 		/*
2692 		 * If fixup_pi_state_owner() faulted and was unable to handle
2693 		 * the fault, unlock the rt_mutex and return the fault to
2694 		 * userspace.
2695 		 */
2696 		if (ret && rt_mutex_owner(pi_mutex) == current)
2697 			rt_mutex_unlock(pi_mutex);
2698 
2699 		/* Unqueue and drop the lock. */
2700 		unqueue_me_pi(&q);
2701 	}
2702 
2703 	if (ret == -EINTR) {
2704 		/*
2705 		 * We've already been requeued, but cannot restart by calling
2706 		 * futex_lock_pi() directly. We could restart this syscall, but
2707 		 * it would detect that the user space "val" changed and return
2708 		 * -EWOULDBLOCK.  Save the overhead of the restart and return
2709 		 * -EWOULDBLOCK directly.
2710 		 */
2711 		ret = -EWOULDBLOCK;
2712 	}
2713 
2714 out_put_keys:
2715 	put_futex_key(&q.key);
2716 out_key2:
2717 	put_futex_key(&key2);
2718 
2719 out:
2720 	if (to) {
2721 		hrtimer_cancel(&to->timer);
2722 		destroy_hrtimer_on_stack(&to->timer);
2723 	}
2724 	return ret;
2725 }
2726 
2727 /*
2728  * Support for robust futexes: the kernel cleans up held futexes at
2729  * thread exit time.
2730  *
2731  * Implementation: user-space maintains a per-thread list of locks it
2732  * is holding. Upon do_exit(), the kernel carefully walks this list,
2733  * and marks all locks that are owned by this thread with the
2734  * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
2735  * always manipulated with the lock held, so the list is private and
2736  * per-thread. Userspace also maintains a per-thread 'list_op_pending'
2737  * field, to allow the kernel to clean up if the thread dies after
2738  * acquiring the lock, but just before it could have added itself to
2739  * the list. There can only be one such pending lock.
2740  */
2741 
2742 /**
2743  * sys_set_robust_list() - Set the robust-futex list head of a task
2744  * @head:	pointer to the list-head
2745  * @len:	length of the list-head, as userspace expects
2746  */
SYSCALL_DEFINE2(set_robust_list,struct robust_list_head __user *,head,size_t,len)2747 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
2748 		size_t, len)
2749 {
2750 	if (!futex_cmpxchg_enabled)
2751 		return -ENOSYS;
2752 	/*
2753 	 * The kernel knows only one size for now:
2754 	 */
2755 	if (unlikely(len != sizeof(*head)))
2756 		return -EINVAL;
2757 
2758 	current->robust_list = head;
2759 
2760 	return 0;
2761 }
2762 
2763 /**
2764  * sys_get_robust_list() - Get the robust-futex list head of a task
2765  * @pid:	pid of the process [zero for current task]
2766  * @head_ptr:	pointer to a list-head pointer, the kernel fills it in
2767  * @len_ptr:	pointer to a length field, the kernel fills in the header size
2768  */
SYSCALL_DEFINE3(get_robust_list,int,pid,struct robust_list_head __user * __user *,head_ptr,size_t __user *,len_ptr)2769 SYSCALL_DEFINE3(get_robust_list, int, pid,
2770 		struct robust_list_head __user * __user *, head_ptr,
2771 		size_t __user *, len_ptr)
2772 {
2773 	struct robust_list_head __user *head;
2774 	unsigned long ret;
2775 	struct task_struct *p;
2776 
2777 	if (!futex_cmpxchg_enabled)
2778 		return -ENOSYS;
2779 
2780 	rcu_read_lock();
2781 
2782 	ret = -ESRCH;
2783 	if (!pid)
2784 		p = current;
2785 	else {
2786 		p = find_task_by_vpid(pid);
2787 		if (!p)
2788 			goto err_unlock;
2789 	}
2790 
2791 	ret = -EPERM;
2792 	if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
2793 		goto err_unlock;
2794 
2795 	head = p->robust_list;
2796 	rcu_read_unlock();
2797 
2798 	if (put_user(sizeof(*head), len_ptr))
2799 		return -EFAULT;
2800 	return put_user(head, head_ptr);
2801 
2802 err_unlock:
2803 	rcu_read_unlock();
2804 
2805 	return ret;
2806 }
2807 
2808 /*
2809  * Process a futex-list entry, check whether it's owned by the
2810  * dying task, and do notification if so:
2811  */
handle_futex_death(u32 __user * uaddr,struct task_struct * curr,int pi)2812 int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
2813 {
2814 	u32 uval, uninitialized_var(nval), mval;
2815 
2816 retry:
2817 	if (get_user(uval, uaddr))
2818 		return -1;
2819 
2820 	if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
2821 		/*
2822 		 * Ok, this dying thread is truly holding a futex
2823 		 * of interest. Set the OWNER_DIED bit atomically
2824 		 * via cmpxchg, and if the value had FUTEX_WAITERS
2825 		 * set, wake up a waiter (if any). (We have to do a
2826 		 * futex_wake() even if OWNER_DIED is already set -
2827 		 * to handle the rare but possible case of recursive
2828 		 * thread-death.) The rest of the cleanup is done in
2829 		 * userspace.
2830 		 */
2831 		mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
2832 		/*
2833 		 * We are not holding a lock here, but we want to have
2834 		 * the pagefault_disable/enable() protection because
2835 		 * we want to handle the fault gracefully. If the
2836 		 * access fails we try to fault in the futex with R/W
2837 		 * verification via get_user_pages. get_user() above
2838 		 * does not guarantee R/W access. If that fails we
2839 		 * give up and leave the futex locked.
2840 		 */
2841 		if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
2842 			if (fault_in_user_writeable(uaddr))
2843 				return -1;
2844 			goto retry;
2845 		}
2846 		if (nval != uval)
2847 			goto retry;
2848 
2849 		/*
2850 		 * Wake robust non-PI futexes here. The wakeup of
2851 		 * PI futexes happens in exit_pi_state():
2852 		 */
2853 		if (!pi && (uval & FUTEX_WAITERS))
2854 			futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
2855 	}
2856 	return 0;
2857 }
2858 
2859 /*
2860  * Fetch a robust-list pointer. Bit 0 signals PI futexes:
2861  */
fetch_robust_entry(struct robust_list __user ** entry,struct robust_list __user * __user * head,unsigned int * pi)2862 static inline int fetch_robust_entry(struct robust_list __user **entry,
2863 				     struct robust_list __user * __user *head,
2864 				     unsigned int *pi)
2865 {
2866 	unsigned long uentry;
2867 
2868 	if (get_user(uentry, (unsigned long __user *)head))
2869 		return -EFAULT;
2870 
2871 	*entry = (void __user *)(uentry & ~1UL);
2872 	*pi = uentry & 1;
2873 
2874 	return 0;
2875 }
2876 
2877 /*
2878  * Walk curr->robust_list (very carefully, it's a userspace list!)
2879  * and mark any locks found there dead, and notify any waiters.
2880  *
2881  * We silently return on any sign of list-walking problem.
2882  */
exit_robust_list(struct task_struct * curr)2883 void exit_robust_list(struct task_struct *curr)
2884 {
2885 	struct robust_list_head __user *head = curr->robust_list;
2886 	struct robust_list __user *entry, *next_entry, *pending;
2887 	unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
2888 	unsigned int uninitialized_var(next_pi);
2889 	unsigned long futex_offset;
2890 	int rc;
2891 
2892 	if (!futex_cmpxchg_enabled)
2893 		return;
2894 
2895 	/*
2896 	 * Fetch the list head (which was registered earlier, via
2897 	 * sys_set_robust_list()):
2898 	 */
2899 	if (fetch_robust_entry(&entry, &head->list.next, &pi))
2900 		return;
2901 	/*
2902 	 * Fetch the relative futex offset:
2903 	 */
2904 	if (get_user(futex_offset, &head->futex_offset))
2905 		return;
2906 	/*
2907 	 * Fetch any possibly pending lock-add first, and handle it
2908 	 * if it exists:
2909 	 */
2910 	if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
2911 		return;
2912 
2913 	next_entry = NULL;	/* avoid warning with gcc */
2914 	while (entry != &head->list) {
2915 		/*
2916 		 * Fetch the next entry in the list before calling
2917 		 * handle_futex_death:
2918 		 */
2919 		rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
2920 		/*
2921 		 * A pending lock might already be on the list, so
2922 		 * don't process it twice:
2923 		 */
2924 		if (entry != pending)
2925 			if (handle_futex_death((void __user *)entry + futex_offset,
2926 						curr, pi))
2927 				return;
2928 		if (rc)
2929 			return;
2930 		entry = next_entry;
2931 		pi = next_pi;
2932 		/*
2933 		 * Avoid excessively long or circular lists:
2934 		 */
2935 		if (!--limit)
2936 			break;
2937 
2938 		cond_resched();
2939 	}
2940 
2941 	if (pending)
2942 		handle_futex_death((void __user *)pending + futex_offset,
2943 				   curr, pip);
2944 }
2945 
do_futex(u32 __user * uaddr,int op,u32 val,ktime_t * timeout,u32 __user * uaddr2,u32 val2,u32 val3)2946 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
2947 		u32 __user *uaddr2, u32 val2, u32 val3)
2948 {
2949 	int cmd = op & FUTEX_CMD_MASK;
2950 	unsigned int flags = 0;
2951 
2952 	if (!(op & FUTEX_PRIVATE_FLAG))
2953 		flags |= FLAGS_SHARED;
2954 
2955 	if (op & FUTEX_CLOCK_REALTIME) {
2956 		flags |= FLAGS_CLOCKRT;
2957 		if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
2958 			return -ENOSYS;
2959 	}
2960 
2961 	switch (cmd) {
2962 	case FUTEX_LOCK_PI:
2963 	case FUTEX_UNLOCK_PI:
2964 	case FUTEX_TRYLOCK_PI:
2965 	case FUTEX_WAIT_REQUEUE_PI:
2966 	case FUTEX_CMP_REQUEUE_PI:
2967 		if (!futex_cmpxchg_enabled)
2968 			return -ENOSYS;
2969 	}
2970 
2971 	switch (cmd) {
2972 	case FUTEX_WAIT:
2973 		val3 = FUTEX_BITSET_MATCH_ANY;
2974 	case FUTEX_WAIT_BITSET:
2975 		return futex_wait(uaddr, flags, val, timeout, val3);
2976 	case FUTEX_WAKE:
2977 		val3 = FUTEX_BITSET_MATCH_ANY;
2978 	case FUTEX_WAKE_BITSET:
2979 		return futex_wake(uaddr, flags, val, val3);
2980 	case FUTEX_REQUEUE:
2981 		return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
2982 	case FUTEX_CMP_REQUEUE:
2983 		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
2984 	case FUTEX_WAKE_OP:
2985 		return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
2986 	case FUTEX_LOCK_PI:
2987 		return futex_lock_pi(uaddr, flags, val, timeout, 0);
2988 	case FUTEX_UNLOCK_PI:
2989 		return futex_unlock_pi(uaddr, flags);
2990 	case FUTEX_TRYLOCK_PI:
2991 		return futex_lock_pi(uaddr, flags, 0, timeout, 1);
2992 	case FUTEX_WAIT_REQUEUE_PI:
2993 		val3 = FUTEX_BITSET_MATCH_ANY;
2994 		return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
2995 					     uaddr2);
2996 	case FUTEX_CMP_REQUEUE_PI:
2997 		return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
2998 	}
2999 	return -ENOSYS;
3000 }
3001 
3002 
SYSCALL_DEFINE6(futex,u32 __user *,uaddr,int,op,u32,val,struct timespec __user *,utime,u32 __user *,uaddr2,u32,val3)3003 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3004 		struct timespec __user *, utime, u32 __user *, uaddr2,
3005 		u32, val3)
3006 {
3007 	struct timespec ts;
3008 	ktime_t t, *tp = NULL;
3009 	u32 val2 = 0;
3010 	int cmd = op & FUTEX_CMD_MASK;
3011 
3012 	if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3013 		      cmd == FUTEX_WAIT_BITSET ||
3014 		      cmd == FUTEX_WAIT_REQUEUE_PI)) {
3015 		if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
3016 			return -EFAULT;
3017 		if (!timespec_valid(&ts))
3018 			return -EINVAL;
3019 
3020 		t = timespec_to_ktime(ts);
3021 		if (cmd == FUTEX_WAIT)
3022 			t = ktime_add_safe(ktime_get(), t);
3023 		tp = &t;
3024 	}
3025 	/*
3026 	 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3027 	 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3028 	 */
3029 	if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3030 	    cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3031 		val2 = (u32) (unsigned long) utime;
3032 
3033 	return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3034 }
3035 
futex_detect_cmpxchg(void)3036 static void __init futex_detect_cmpxchg(void)
3037 {
3038 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3039 	u32 curval;
3040 
3041 	/*
3042 	 * This will fail and we want it. Some arch implementations do
3043 	 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3044 	 * functionality. We want to know that before we call in any
3045 	 * of the complex code paths. Also we want to prevent
3046 	 * registration of robust lists in that case. NULL is
3047 	 * guaranteed to fault and we get -EFAULT on functional
3048 	 * implementation, the non-functional ones will return
3049 	 * -ENOSYS.
3050 	 */
3051 	if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3052 		futex_cmpxchg_enabled = 1;
3053 #endif
3054 }
3055 
futex_init(void)3056 static int __init futex_init(void)
3057 {
3058 	unsigned int futex_shift;
3059 	unsigned long i;
3060 
3061 #if CONFIG_BASE_SMALL
3062 	futex_hashsize = 16;
3063 #else
3064 	futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3065 #endif
3066 
3067 	futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3068 					       futex_hashsize, 0,
3069 					       futex_hashsize < 256 ? HASH_SMALL : 0,
3070 					       &futex_shift, NULL,
3071 					       futex_hashsize, futex_hashsize);
3072 	futex_hashsize = 1UL << futex_shift;
3073 
3074 	futex_detect_cmpxchg();
3075 
3076 	for (i = 0; i < futex_hashsize; i++) {
3077 		atomic_set(&futex_queues[i].waiters, 0);
3078 		plist_head_init(&futex_queues[i].chain);
3079 		spin_lock_init(&futex_queues[i].lock);
3080 	}
3081 
3082 	return 0;
3083 }
3084 __initcall(futex_init);
3085