]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Fast Userspace Mutexes (which I call "Futexes!"). | |
3 | * (C) Rusty Russell, IBM 2002 | |
4 | * | |
5 | * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar | |
6 | * (C) Copyright 2003 Red Hat Inc, All Rights Reserved | |
7 | * | |
8 | * Removed page pinning, fix privately mapped COW pages and other cleanups | |
9 | * (C) Copyright 2003, 2004 Jamie Lokier | |
10 | * | |
11 | * Robust futex support started by Ingo Molnar | |
12 | * (C) Copyright 2006 Red Hat Inc, All Rights Reserved | |
13 | * Thanks to Thomas Gleixner for suggestions, analysis and fixes. | |
14 | * | |
15 | * PI-futex support started by Ingo Molnar and Thomas Gleixner | |
16 | * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
17 | * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com> | |
18 | * | |
19 | * PRIVATE futexes by Eric Dumazet | |
20 | * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com> | |
21 | * | |
22 | * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com> | |
23 | * Copyright (C) IBM Corporation, 2009 | |
24 | * Thanks to Thomas Gleixner for conceptual design and careful reviews. | |
25 | * | |
26 | * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly | |
27 | * enough at me, Linus for the original (flawed) idea, Matthew | |
28 | * Kirkwood for proof-of-concept implementation. | |
29 | * | |
30 | * "The futexes are also cursed." | |
31 | * "But they come in a choice of three flavours!" | |
32 | * | |
33 | * This program is free software; you can redistribute it and/or modify | |
34 | * it under the terms of the GNU General Public License as published by | |
35 | * the Free Software Foundation; either version 2 of the License, or | |
36 | * (at your option) any later version. | |
37 | * | |
38 | * This program is distributed in the hope that it will be useful, | |
39 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
40 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
41 | * GNU General Public License for more details. | |
42 | * | |
43 | * You should have received a copy of the GNU General Public License | |
44 | * along with this program; if not, write to the Free Software | |
45 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
46 | */ | |
47 | #include <linux/slab.h> | |
48 | #include <linux/poll.h> | |
49 | #include <linux/fs.h> | |
50 | #include <linux/file.h> | |
51 | #include <linux/jhash.h> | |
52 | #include <linux/init.h> | |
53 | #include <linux/futex.h> | |
54 | #include <linux/mount.h> | |
55 | #include <linux/pagemap.h> | |
56 | #include <linux/syscalls.h> | |
57 | #include <linux/signal.h> | |
58 | #include <linux/export.h> | |
59 | #include <linux/magic.h> | |
60 | #include <linux/pid.h> | |
61 | #include <linux/nsproxy.h> | |
62 | #include <linux/ptrace.h> | |
63 | #include <linux/sched/rt.h> | |
64 | #include <linux/sched/wake_q.h> | |
65 | #include <linux/sched/mm.h> | |
66 | #include <linux/hugetlb.h> | |
67 | #include <linux/freezer.h> | |
68 | #include <linux/bootmem.h> | |
69 | #include <linux/fault-inject.h> | |
70 | ||
71 | #include <asm/futex.h> | |
72 | ||
73 | #include "locking/rtmutex_common.h" | |
74 | ||
75 | /* | |
76 | * READ this before attempting to hack on futexes! | |
77 | * | |
78 | * Basic futex operation and ordering guarantees | |
79 | * ============================================= | |
80 | * | |
81 | * The waiter reads the futex value in user space and calls | |
82 | * futex_wait(). This function computes the hash bucket and acquires | |
83 | * the hash bucket lock. After that it reads the futex user space value | |
84 | * again and verifies that the data has not changed. If it has not changed | |
85 | * it enqueues itself into the hash bucket, releases the hash bucket lock | |
86 | * and schedules. | |
87 | * | |
88 | * The waker side modifies the user space value of the futex and calls | |
89 | * futex_wake(). This function computes the hash bucket and acquires the | |
90 | * hash bucket lock. Then it looks for waiters on that futex in the hash | |
91 | * bucket and wakes them. | |
92 | * | |
93 | * In futex wake up scenarios where no tasks are blocked on a futex, taking | |
94 | * the hb spinlock can be avoided and simply return. In order for this | |
95 | * optimization to work, ordering guarantees must exist so that the waiter | |
96 | * being added to the list is acknowledged when the list is concurrently being | |
97 | * checked by the waker, avoiding scenarios like the following: | |
98 | * | |
99 | * CPU 0 CPU 1 | |
100 | * val = *futex; | |
101 | * sys_futex(WAIT, futex, val); | |
102 | * futex_wait(futex, val); | |
103 | * uval = *futex; | |
104 | * *futex = newval; | |
105 | * sys_futex(WAKE, futex); | |
106 | * futex_wake(futex); | |
107 | * if (queue_empty()) | |
108 | * return; | |
109 | * if (uval == val) | |
110 | * lock(hash_bucket(futex)); | |
111 | * queue(); | |
112 | * unlock(hash_bucket(futex)); | |
113 | * schedule(); | |
114 | * | |
115 | * This would cause the waiter on CPU 0 to wait forever because it | |
116 | * missed the transition of the user space value from val to newval | |
117 | * and the waker did not find the waiter in the hash bucket queue. | |
118 | * | |
119 | * The correct serialization ensures that a waiter either observes | |
120 | * the changed user space value before blocking or is woken by a | |
121 | * concurrent waker: | |
122 | * | |
123 | * CPU 0 CPU 1 | |
124 | * val = *futex; | |
125 | * sys_futex(WAIT, futex, val); | |
126 | * futex_wait(futex, val); | |
127 | * | |
128 | * waiters++; (a) | |
129 | * smp_mb(); (A) <-- paired with -. | |
130 | * | | |
131 | * lock(hash_bucket(futex)); | | |
132 | * | | |
133 | * uval = *futex; | | |
134 | * | *futex = newval; | |
135 | * | sys_futex(WAKE, futex); | |
136 | * | futex_wake(futex); | |
137 | * | | |
138 | * `--------> smp_mb(); (B) | |
139 | * if (uval == val) | |
140 | * queue(); | |
141 | * unlock(hash_bucket(futex)); | |
142 | * schedule(); if (waiters) | |
143 | * lock(hash_bucket(futex)); | |
144 | * else wake_waiters(futex); | |
145 | * waiters--; (b) unlock(hash_bucket(futex)); | |
146 | * | |
147 | * Where (A) orders the waiters increment and the futex value read through | |
148 | * atomic operations (see hb_waiters_inc) and where (B) orders the write | |
149 | * to futex and the waiters read -- this is done by the barriers for both | |
150 | * shared and private futexes in get_futex_key_refs(). | |
151 | * | |
152 | * This yields the following case (where X:=waiters, Y:=futex): | |
153 | * | |
154 | * X = Y = 0 | |
155 | * | |
156 | * w[X]=1 w[Y]=1 | |
157 | * MB MB | |
158 | * r[Y]=y r[X]=x | |
159 | * | |
160 | * Which guarantees that x==0 && y==0 is impossible; which translates back into | |
161 | * the guarantee that we cannot both miss the futex variable change and the | |
162 | * enqueue. | |
163 | * | |
164 | * Note that a new waiter is accounted for in (a) even when it is possible that | |
165 | * the wait call can return error, in which case we backtrack from it in (b). | |
166 | * Refer to the comment in queue_lock(). | |
167 | * | |
168 | * Similarly, in order to account for waiters being requeued on another | |
169 | * address we always increment the waiters for the destination bucket before | |
170 | * acquiring the lock. It then decrements them again after releasing it - | |
171 | * the code that actually moves the futex(es) between hash buckets (requeue_futex) | |
172 | * will do the additional required waiter count housekeeping. This is done for | |
173 | * double_lock_hb() and double_unlock_hb(), respectively. | |
174 | */ | |
175 | ||
176 | #ifndef CONFIG_HAVE_FUTEX_CMPXCHG | |
177 | int __read_mostly futex_cmpxchg_enabled; | |
178 | #endif | |
179 | ||
180 | /* | |
181 | * Futex flags used to encode options to functions and preserve them across | |
182 | * restarts. | |
183 | */ | |
184 | #ifdef CONFIG_MMU | |
185 | # define FLAGS_SHARED 0x01 | |
186 | #else | |
187 | /* | |
188 | * NOMMU does not have per process address space. Let the compiler optimize | |
189 | * code away. | |
190 | */ | |
191 | # define FLAGS_SHARED 0x00 | |
192 | #endif | |
193 | #define FLAGS_CLOCKRT 0x02 | |
194 | #define FLAGS_HAS_TIMEOUT 0x04 | |
195 | ||
196 | /* | |
197 | * Priority Inheritance state: | |
198 | */ | |
199 | struct futex_pi_state { | |
200 | /* | |
201 | * list of 'owned' pi_state instances - these have to be | |
202 | * cleaned up in do_exit() if the task exits prematurely: | |
203 | */ | |
204 | struct list_head list; | |
205 | ||
206 | /* | |
207 | * The PI object: | |
208 | */ | |
209 | struct rt_mutex pi_mutex; | |
210 | ||
211 | struct task_struct *owner; | |
212 | atomic_t refcount; | |
213 | ||
214 | union futex_key key; | |
215 | }; | |
216 | ||
217 | /** | |
218 | * struct futex_q - The hashed futex queue entry, one per waiting task | |
219 | * @list: priority-sorted list of tasks waiting on this futex | |
220 | * @task: the task waiting on the futex | |
221 | * @lock_ptr: the hash bucket lock | |
222 | * @key: the key the futex is hashed on | |
223 | * @pi_state: optional priority inheritance state | |
224 | * @rt_waiter: rt_waiter storage for use with requeue_pi | |
225 | * @requeue_pi_key: the requeue_pi target futex key | |
226 | * @bitset: bitset for the optional bitmasked wakeup | |
227 | * | |
228 | * We use this hashed waitqueue, instead of a normal wait_queue_t, so | |
229 | * we can wake only the relevant ones (hashed queues may be shared). | |
230 | * | |
231 | * A futex_q has a woken state, just like tasks have TASK_RUNNING. | |
232 | * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0. | |
233 | * The order of wakeup is always to make the first condition true, then | |
234 | * the second. | |
235 | * | |
236 | * PI futexes are typically woken before they are removed from the hash list via | |
237 | * the rt_mutex code. See unqueue_me_pi(). | |
238 | */ | |
239 | struct futex_q { | |
240 | struct plist_node list; | |
241 | ||
242 | struct task_struct *task; | |
243 | spinlock_t *lock_ptr; | |
244 | union futex_key key; | |
245 | struct futex_pi_state *pi_state; | |
246 | struct rt_mutex_waiter *rt_waiter; | |
247 | union futex_key *requeue_pi_key; | |
248 | u32 bitset; | |
249 | }; | |
250 | ||
251 | static const struct futex_q futex_q_init = { | |
252 | /* list gets initialized in queue_me()*/ | |
253 | .key = FUTEX_KEY_INIT, | |
254 | .bitset = FUTEX_BITSET_MATCH_ANY | |
255 | }; | |
256 | ||
257 | /* | |
258 | * Hash buckets are shared by all the futex_keys that hash to the same | |
259 | * location. Each key may have multiple futex_q structures, one for each task | |
260 | * waiting on a futex. | |
261 | */ | |
262 | struct futex_hash_bucket { | |
263 | atomic_t waiters; | |
264 | spinlock_t lock; | |
265 | struct plist_head chain; | |
266 | } ____cacheline_aligned_in_smp; | |
267 | ||
268 | /* | |
269 | * The base of the bucket array and its size are always used together | |
270 | * (after initialization only in hash_futex()), so ensure that they | |
271 | * reside in the same cacheline. | |
272 | */ | |
273 | static struct { | |
274 | struct futex_hash_bucket *queues; | |
275 | unsigned long hashsize; | |
276 | } __futex_data __read_mostly __aligned(2*sizeof(long)); | |
277 | #define futex_queues (__futex_data.queues) | |
278 | #define futex_hashsize (__futex_data.hashsize) | |
279 | ||
280 | ||
281 | /* | |
282 | * Fault injections for futexes. | |
283 | */ | |
284 | #ifdef CONFIG_FAIL_FUTEX | |
285 | ||
286 | static struct { | |
287 | struct fault_attr attr; | |
288 | ||
289 | bool ignore_private; | |
290 | } fail_futex = { | |
291 | .attr = FAULT_ATTR_INITIALIZER, | |
292 | .ignore_private = false, | |
293 | }; | |
294 | ||
295 | static int __init setup_fail_futex(char *str) | |
296 | { | |
297 | return setup_fault_attr(&fail_futex.attr, str); | |
298 | } | |
299 | __setup("fail_futex=", setup_fail_futex); | |
300 | ||
301 | static bool should_fail_futex(bool fshared) | |
302 | { | |
303 | if (fail_futex.ignore_private && !fshared) | |
304 | return false; | |
305 | ||
306 | return should_fail(&fail_futex.attr, 1); | |
307 | } | |
308 | ||
309 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
310 | ||
311 | static int __init fail_futex_debugfs(void) | |
312 | { | |
313 | umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
314 | struct dentry *dir; | |
315 | ||
316 | dir = fault_create_debugfs_attr("fail_futex", NULL, | |
317 | &fail_futex.attr); | |
318 | if (IS_ERR(dir)) | |
319 | return PTR_ERR(dir); | |
320 | ||
321 | if (!debugfs_create_bool("ignore-private", mode, dir, | |
322 | &fail_futex.ignore_private)) { | |
323 | debugfs_remove_recursive(dir); | |
324 | return -ENOMEM; | |
325 | } | |
326 | ||
327 | return 0; | |
328 | } | |
329 | ||
330 | late_initcall(fail_futex_debugfs); | |
331 | ||
332 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
333 | ||
334 | #else | |
335 | static inline bool should_fail_futex(bool fshared) | |
336 | { | |
337 | return false; | |
338 | } | |
339 | #endif /* CONFIG_FAIL_FUTEX */ | |
340 | ||
341 | static inline void futex_get_mm(union futex_key *key) | |
342 | { | |
343 | mmgrab(key->private.mm); | |
344 | /* | |
345 | * Ensure futex_get_mm() implies a full barrier such that | |
346 | * get_futex_key() implies a full barrier. This is relied upon | |
347 | * as smp_mb(); (B), see the ordering comment above. | |
348 | */ | |
349 | smp_mb__after_atomic(); | |
350 | } | |
351 | ||
352 | /* | |
353 | * Reflects a new waiter being added to the waitqueue. | |
354 | */ | |
355 | static inline void hb_waiters_inc(struct futex_hash_bucket *hb) | |
356 | { | |
357 | #ifdef CONFIG_SMP | |
358 | atomic_inc(&hb->waiters); | |
359 | /* | |
360 | * Full barrier (A), see the ordering comment above. | |
361 | */ | |
362 | smp_mb__after_atomic(); | |
363 | #endif | |
364 | } | |
365 | ||
366 | /* | |
367 | * Reflects a waiter being removed from the waitqueue by wakeup | |
368 | * paths. | |
369 | */ | |
370 | static inline void hb_waiters_dec(struct futex_hash_bucket *hb) | |
371 | { | |
372 | #ifdef CONFIG_SMP | |
373 | atomic_dec(&hb->waiters); | |
374 | #endif | |
375 | } | |
376 | ||
377 | static inline int hb_waiters_pending(struct futex_hash_bucket *hb) | |
378 | { | |
379 | #ifdef CONFIG_SMP | |
380 | return atomic_read(&hb->waiters); | |
381 | #else | |
382 | return 1; | |
383 | #endif | |
384 | } | |
385 | ||
386 | /** | |
387 | * hash_futex - Return the hash bucket in the global hash | |
388 | * @key: Pointer to the futex key for which the hash is calculated | |
389 | * | |
390 | * We hash on the keys returned from get_futex_key (see below) and return the | |
391 | * corresponding hash bucket in the global hash. | |
392 | */ | |
393 | static struct futex_hash_bucket *hash_futex(union futex_key *key) | |
394 | { | |
395 | u32 hash = jhash2((u32*)&key->both.word, | |
396 | (sizeof(key->both.word)+sizeof(key->both.ptr))/4, | |
397 | key->both.offset); | |
398 | return &futex_queues[hash & (futex_hashsize - 1)]; | |
399 | } | |
400 | ||
401 | ||
402 | /** | |
403 | * match_futex - Check whether two futex keys are equal | |
404 | * @key1: Pointer to key1 | |
405 | * @key2: Pointer to key2 | |
406 | * | |
407 | * Return 1 if two futex_keys are equal, 0 otherwise. | |
408 | */ | |
409 | static inline int match_futex(union futex_key *key1, union futex_key *key2) | |
410 | { | |
411 | return (key1 && key2 | |
412 | && key1->both.word == key2->both.word | |
413 | && key1->both.ptr == key2->both.ptr | |
414 | && key1->both.offset == key2->both.offset); | |
415 | } | |
416 | ||
417 | /* | |
418 | * Take a reference to the resource addressed by a key. | |
419 | * Can be called while holding spinlocks. | |
420 | * | |
421 | */ | |
422 | static void get_futex_key_refs(union futex_key *key) | |
423 | { | |
424 | if (!key->both.ptr) | |
425 | return; | |
426 | ||
427 | /* | |
428 | * On MMU less systems futexes are always "private" as there is no per | |
429 | * process address space. We need the smp wmb nevertheless - yes, | |
430 | * arch/blackfin has MMU less SMP ... | |
431 | */ | |
432 | if (!IS_ENABLED(CONFIG_MMU)) { | |
433 | smp_mb(); /* explicit smp_mb(); (B) */ | |
434 | return; | |
435 | } | |
436 | ||
437 | switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) { | |
438 | case FUT_OFF_INODE: | |
439 | ihold(key->shared.inode); /* implies smp_mb(); (B) */ | |
440 | break; | |
441 | case FUT_OFF_MMSHARED: | |
442 | futex_get_mm(key); /* implies smp_mb(); (B) */ | |
443 | break; | |
444 | default: | |
445 | /* | |
446 | * Private futexes do not hold reference on an inode or | |
447 | * mm, therefore the only purpose of calling get_futex_key_refs | |
448 | * is because we need the barrier for the lockless waiter check. | |
449 | */ | |
450 | smp_mb(); /* explicit smp_mb(); (B) */ | |
451 | } | |
452 | } | |
453 | ||
454 | /* | |
455 | * Drop a reference to the resource addressed by a key. | |
456 | * The hash bucket spinlock must not be held. This is | |
457 | * a no-op for private futexes, see comment in the get | |
458 | * counterpart. | |
459 | */ | |
460 | static void drop_futex_key_refs(union futex_key *key) | |
461 | { | |
462 | if (!key->both.ptr) { | |
463 | /* If we're here then we tried to put a key we failed to get */ | |
464 | WARN_ON_ONCE(1); | |
465 | return; | |
466 | } | |
467 | ||
468 | if (!IS_ENABLED(CONFIG_MMU)) | |
469 | return; | |
470 | ||
471 | switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) { | |
472 | case FUT_OFF_INODE: | |
473 | iput(key->shared.inode); | |
474 | break; | |
475 | case FUT_OFF_MMSHARED: | |
476 | mmdrop(key->private.mm); | |
477 | break; | |
478 | } | |
479 | } | |
480 | ||
481 | /** | |
482 | * get_futex_key() - Get parameters which are the keys for a futex | |
483 | * @uaddr: virtual address of the futex | |
484 | * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED | |
485 | * @key: address where result is stored. | |
486 | * @rw: mapping needs to be read/write (values: VERIFY_READ, | |
487 | * VERIFY_WRITE) | |
488 | * | |
489 | * Return: a negative error code or 0 | |
490 | * | |
491 | * The key words are stored in *key on success. | |
492 | * | |
493 | * For shared mappings, it's (page->index, file_inode(vma->vm_file), | |
494 | * offset_within_page). For private mappings, it's (uaddr, current->mm). | |
495 | * We can usually work out the index without swapping in the page. | |
496 | * | |
497 | * lock_page() might sleep, the caller should not hold a spinlock. | |
498 | */ | |
499 | static int | |
500 | get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw) | |
501 | { | |
502 | unsigned long address = (unsigned long)uaddr; | |
503 | struct mm_struct *mm = current->mm; | |
504 | struct page *page, *tail; | |
505 | struct address_space *mapping; | |
506 | int err, ro = 0; | |
507 | ||
508 | /* | |
509 | * The futex address must be "naturally" aligned. | |
510 | */ | |
511 | key->both.offset = address % PAGE_SIZE; | |
512 | if (unlikely((address % sizeof(u32)) != 0)) | |
513 | return -EINVAL; | |
514 | address -= key->both.offset; | |
515 | ||
516 | if (unlikely(!access_ok(rw, uaddr, sizeof(u32)))) | |
517 | return -EFAULT; | |
518 | ||
519 | if (unlikely(should_fail_futex(fshared))) | |
520 | return -EFAULT; | |
521 | ||
522 | /* | |
523 | * PROCESS_PRIVATE futexes are fast. | |
524 | * As the mm cannot disappear under us and the 'key' only needs | |
525 | * virtual address, we dont even have to find the underlying vma. | |
526 | * Note : We do have to check 'uaddr' is a valid user address, | |
527 | * but access_ok() should be faster than find_vma() | |
528 | */ | |
529 | if (!fshared) { | |
530 | key->private.mm = mm; | |
531 | key->private.address = address; | |
532 | get_futex_key_refs(key); /* implies smp_mb(); (B) */ | |
533 | return 0; | |
534 | } | |
535 | ||
536 | again: | |
537 | /* Ignore any VERIFY_READ mapping (futex common case) */ | |
538 | if (unlikely(should_fail_futex(fshared))) | |
539 | return -EFAULT; | |
540 | ||
541 | err = get_user_pages_fast(address, 1, 1, &page); | |
542 | /* | |
543 | * If write access is not required (eg. FUTEX_WAIT), try | |
544 | * and get read-only access. | |
545 | */ | |
546 | if (err == -EFAULT && rw == VERIFY_READ) { | |
547 | err = get_user_pages_fast(address, 1, 0, &page); | |
548 | ro = 1; | |
549 | } | |
550 | if (err < 0) | |
551 | return err; | |
552 | else | |
553 | err = 0; | |
554 | ||
555 | /* | |
556 | * The treatment of mapping from this point on is critical. The page | |
557 | * lock protects many things but in this context the page lock | |
558 | * stabilizes mapping, prevents inode freeing in the shared | |
559 | * file-backed region case and guards against movement to swap cache. | |
560 | * | |
561 | * Strictly speaking the page lock is not needed in all cases being | |
562 | * considered here and page lock forces unnecessarily serialization | |
563 | * From this point on, mapping will be re-verified if necessary and | |
564 | * page lock will be acquired only if it is unavoidable | |
565 | * | |
566 | * Mapping checks require the head page for any compound page so the | |
567 | * head page and mapping is looked up now. For anonymous pages, it | |
568 | * does not matter if the page splits in the future as the key is | |
569 | * based on the address. For filesystem-backed pages, the tail is | |
570 | * required as the index of the page determines the key. For | |
571 | * base pages, there is no tail page and tail == page. | |
572 | */ | |
573 | tail = page; | |
574 | page = compound_head(page); | |
575 | mapping = READ_ONCE(page->mapping); | |
576 | ||
577 | /* | |
578 | * If page->mapping is NULL, then it cannot be a PageAnon | |
579 | * page; but it might be the ZERO_PAGE or in the gate area or | |
580 | * in a special mapping (all cases which we are happy to fail); | |
581 | * or it may have been a good file page when get_user_pages_fast | |
582 | * found it, but truncated or holepunched or subjected to | |
583 | * invalidate_complete_page2 before we got the page lock (also | |
584 | * cases which we are happy to fail). And we hold a reference, | |
585 | * so refcount care in invalidate_complete_page's remove_mapping | |
586 | * prevents drop_caches from setting mapping to NULL beneath us. | |
587 | * | |
588 | * The case we do have to guard against is when memory pressure made | |
589 | * shmem_writepage move it from filecache to swapcache beneath us: | |
590 | * an unlikely race, but we do need to retry for page->mapping. | |
591 | */ | |
592 | if (unlikely(!mapping)) { | |
593 | int shmem_swizzled; | |
594 | ||
595 | /* | |
596 | * Page lock is required to identify which special case above | |
597 | * applies. If this is really a shmem page then the page lock | |
598 | * will prevent unexpected transitions. | |
599 | */ | |
600 | lock_page(page); | |
601 | shmem_swizzled = PageSwapCache(page) || page->mapping; | |
602 | unlock_page(page); | |
603 | put_page(page); | |
604 | ||
605 | if (shmem_swizzled) | |
606 | goto again; | |
607 | ||
608 | return -EFAULT; | |
609 | } | |
610 | ||
611 | /* | |
612 | * Private mappings are handled in a simple way. | |
613 | * | |
614 | * If the futex key is stored on an anonymous page, then the associated | |
615 | * object is the mm which is implicitly pinned by the calling process. | |
616 | * | |
617 | * NOTE: When userspace waits on a MAP_SHARED mapping, even if | |
618 | * it's a read-only handle, it's expected that futexes attach to | |
619 | * the object not the particular process. | |
620 | */ | |
621 | if (PageAnon(page)) { | |
622 | /* | |
623 | * A RO anonymous page will never change and thus doesn't make | |
624 | * sense for futex operations. | |
625 | */ | |
626 | if (unlikely(should_fail_futex(fshared)) || ro) { | |
627 | err = -EFAULT; | |
628 | goto out; | |
629 | } | |
630 | ||
631 | key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */ | |
632 | key->private.mm = mm; | |
633 | key->private.address = address; | |
634 | ||
635 | get_futex_key_refs(key); /* implies smp_mb(); (B) */ | |
636 | ||
637 | } else { | |
638 | struct inode *inode; | |
639 | ||
640 | /* | |
641 | * The associated futex object in this case is the inode and | |
642 | * the page->mapping must be traversed. Ordinarily this should | |
643 | * be stabilised under page lock but it's not strictly | |
644 | * necessary in this case as we just want to pin the inode, not | |
645 | * update the radix tree or anything like that. | |
646 | * | |
647 | * The RCU read lock is taken as the inode is finally freed | |
648 | * under RCU. If the mapping still matches expectations then the | |
649 | * mapping->host can be safely accessed as being a valid inode. | |
650 | */ | |
651 | rcu_read_lock(); | |
652 | ||
653 | if (READ_ONCE(page->mapping) != mapping) { | |
654 | rcu_read_unlock(); | |
655 | put_page(page); | |
656 | ||
657 | goto again; | |
658 | } | |
659 | ||
660 | inode = READ_ONCE(mapping->host); | |
661 | if (!inode) { | |
662 | rcu_read_unlock(); | |
663 | put_page(page); | |
664 | ||
665 | goto again; | |
666 | } | |
667 | ||
668 | /* | |
669 | * Take a reference unless it is about to be freed. Previously | |
670 | * this reference was taken by ihold under the page lock | |
671 | * pinning the inode in place so i_lock was unnecessary. The | |
672 | * only way for this check to fail is if the inode was | |
673 | * truncated in parallel so warn for now if this happens. | |
674 | * | |
675 | * We are not calling into get_futex_key_refs() in file-backed | |
676 | * cases, therefore a successful atomic_inc return below will | |
677 | * guarantee that get_futex_key() will still imply smp_mb(); (B). | |
678 | */ | |
679 | if (WARN_ON_ONCE(!atomic_inc_not_zero(&inode->i_count))) { | |
680 | rcu_read_unlock(); | |
681 | put_page(page); | |
682 | ||
683 | goto again; | |
684 | } | |
685 | ||
686 | /* Should be impossible but lets be paranoid for now */ | |
687 | if (WARN_ON_ONCE(inode->i_mapping != mapping)) { | |
688 | err = -EFAULT; | |
689 | rcu_read_unlock(); | |
690 | iput(inode); | |
691 | ||
692 | goto out; | |
693 | } | |
694 | ||
695 | key->both.offset |= FUT_OFF_INODE; /* inode-based key */ | |
696 | key->shared.inode = inode; | |
697 | key->shared.pgoff = basepage_index(tail); | |
698 | rcu_read_unlock(); | |
699 | } | |
700 | ||
701 | out: | |
702 | put_page(page); | |
703 | return err; | |
704 | } | |
705 | ||
706 | static inline void put_futex_key(union futex_key *key) | |
707 | { | |
708 | drop_futex_key_refs(key); | |
709 | } | |
710 | ||
711 | /** | |
712 | * fault_in_user_writeable() - Fault in user address and verify RW access | |
713 | * @uaddr: pointer to faulting user space address | |
714 | * | |
715 | * Slow path to fixup the fault we just took in the atomic write | |
716 | * access to @uaddr. | |
717 | * | |
718 | * We have no generic implementation of a non-destructive write to the | |
719 | * user address. We know that we faulted in the atomic pagefault | |
720 | * disabled section so we can as well avoid the #PF overhead by | |
721 | * calling get_user_pages() right away. | |
722 | */ | |
723 | static int fault_in_user_writeable(u32 __user *uaddr) | |
724 | { | |
725 | struct mm_struct *mm = current->mm; | |
726 | int ret; | |
727 | ||
728 | down_read(&mm->mmap_sem); | |
729 | ret = fixup_user_fault(current, mm, (unsigned long)uaddr, | |
730 | FAULT_FLAG_WRITE, NULL); | |
731 | up_read(&mm->mmap_sem); | |
732 | ||
733 | return ret < 0 ? ret : 0; | |
734 | } | |
735 | ||
736 | /** | |
737 | * futex_top_waiter() - Return the highest priority waiter on a futex | |
738 | * @hb: the hash bucket the futex_q's reside in | |
739 | * @key: the futex key (to distinguish it from other futex futex_q's) | |
740 | * | |
741 | * Must be called with the hb lock held. | |
742 | */ | |
743 | static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb, | |
744 | union futex_key *key) | |
745 | { | |
746 | struct futex_q *this; | |
747 | ||
748 | plist_for_each_entry(this, &hb->chain, list) { | |
749 | if (match_futex(&this->key, key)) | |
750 | return this; | |
751 | } | |
752 | return NULL; | |
753 | } | |
754 | ||
755 | static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr, | |
756 | u32 uval, u32 newval) | |
757 | { | |
758 | int ret; | |
759 | ||
760 | pagefault_disable(); | |
761 | ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval); | |
762 | pagefault_enable(); | |
763 | ||
764 | return ret; | |
765 | } | |
766 | ||
767 | static int get_futex_value_locked(u32 *dest, u32 __user *from) | |
768 | { | |
769 | int ret; | |
770 | ||
771 | pagefault_disable(); | |
772 | ret = __get_user(*dest, from); | |
773 | pagefault_enable(); | |
774 | ||
775 | return ret ? -EFAULT : 0; | |
776 | } | |
777 | ||
778 | ||
779 | /* | |
780 | * PI code: | |
781 | */ | |
782 | static int refill_pi_state_cache(void) | |
783 | { | |
784 | struct futex_pi_state *pi_state; | |
785 | ||
786 | if (likely(current->pi_state_cache)) | |
787 | return 0; | |
788 | ||
789 | pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL); | |
790 | ||
791 | if (!pi_state) | |
792 | return -ENOMEM; | |
793 | ||
794 | INIT_LIST_HEAD(&pi_state->list); | |
795 | /* pi_mutex gets initialized later */ | |
796 | pi_state->owner = NULL; | |
797 | atomic_set(&pi_state->refcount, 1); | |
798 | pi_state->key = FUTEX_KEY_INIT; | |
799 | ||
800 | current->pi_state_cache = pi_state; | |
801 | ||
802 | return 0; | |
803 | } | |
804 | ||
805 | static struct futex_pi_state *alloc_pi_state(void) | |
806 | { | |
807 | struct futex_pi_state *pi_state = current->pi_state_cache; | |
808 | ||
809 | WARN_ON(!pi_state); | |
810 | current->pi_state_cache = NULL; | |
811 | ||
812 | return pi_state; | |
813 | } | |
814 | ||
815 | static void get_pi_state(struct futex_pi_state *pi_state) | |
816 | { | |
817 | WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount)); | |
818 | } | |
819 | ||
820 | /* | |
821 | * Drops a reference to the pi_state object and frees or caches it | |
822 | * when the last reference is gone. | |
823 | * | |
824 | * Must be called with the hb lock held. | |
825 | */ | |
826 | static void put_pi_state(struct futex_pi_state *pi_state) | |
827 | { | |
828 | if (!pi_state) | |
829 | return; | |
830 | ||
831 | if (!atomic_dec_and_test(&pi_state->refcount)) | |
832 | return; | |
833 | ||
834 | /* | |
835 | * If pi_state->owner is NULL, the owner is most probably dying | |
836 | * and has cleaned up the pi_state already | |
837 | */ | |
838 | if (pi_state->owner) { | |
839 | raw_spin_lock_irq(&pi_state->owner->pi_lock); | |
840 | list_del_init(&pi_state->list); | |
841 | raw_spin_unlock_irq(&pi_state->owner->pi_lock); | |
842 | ||
843 | rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner); | |
844 | } | |
845 | ||
846 | if (current->pi_state_cache) | |
847 | kfree(pi_state); | |
848 | else { | |
849 | /* | |
850 | * pi_state->list is already empty. | |
851 | * clear pi_state->owner. | |
852 | * refcount is at 0 - put it back to 1. | |
853 | */ | |
854 | pi_state->owner = NULL; | |
855 | atomic_set(&pi_state->refcount, 1); | |
856 | current->pi_state_cache = pi_state; | |
857 | } | |
858 | } | |
859 | ||
860 | /* | |
861 | * Look up the task based on what TID userspace gave us. | |
862 | * We dont trust it. | |
863 | */ | |
864 | static struct task_struct *futex_find_get_task(pid_t pid) | |
865 | { | |
866 | struct task_struct *p; | |
867 | ||
868 | rcu_read_lock(); | |
869 | p = find_task_by_vpid(pid); | |
870 | if (p) | |
871 | get_task_struct(p); | |
872 | ||
873 | rcu_read_unlock(); | |
874 | ||
875 | return p; | |
876 | } | |
877 | ||
878 | /* | |
879 | * This task is holding PI mutexes at exit time => bad. | |
880 | * Kernel cleans up PI-state, but userspace is likely hosed. | |
881 | * (Robust-futex cleanup is separate and might save the day for userspace.) | |
882 | */ | |
883 | void exit_pi_state_list(struct task_struct *curr) | |
884 | { | |
885 | struct list_head *next, *head = &curr->pi_state_list; | |
886 | struct futex_pi_state *pi_state; | |
887 | struct futex_hash_bucket *hb; | |
888 | union futex_key key = FUTEX_KEY_INIT; | |
889 | ||
890 | if (!futex_cmpxchg_enabled) | |
891 | return; | |
892 | /* | |
893 | * We are a ZOMBIE and nobody can enqueue itself on | |
894 | * pi_state_list anymore, but we have to be careful | |
895 | * versus waiters unqueueing themselves: | |
896 | */ | |
897 | raw_spin_lock_irq(&curr->pi_lock); | |
898 | while (!list_empty(head)) { | |
899 | ||
900 | next = head->next; | |
901 | pi_state = list_entry(next, struct futex_pi_state, list); | |
902 | key = pi_state->key; | |
903 | hb = hash_futex(&key); | |
904 | raw_spin_unlock_irq(&curr->pi_lock); | |
905 | ||
906 | spin_lock(&hb->lock); | |
907 | ||
908 | raw_spin_lock_irq(&curr->pi_lock); | |
909 | /* | |
910 | * We dropped the pi-lock, so re-check whether this | |
911 | * task still owns the PI-state: | |
912 | */ | |
913 | if (head->next != next) { | |
914 | spin_unlock(&hb->lock); | |
915 | continue; | |
916 | } | |
917 | ||
918 | WARN_ON(pi_state->owner != curr); | |
919 | WARN_ON(list_empty(&pi_state->list)); | |
920 | list_del_init(&pi_state->list); | |
921 | pi_state->owner = NULL; | |
922 | raw_spin_unlock_irq(&curr->pi_lock); | |
923 | ||
924 | get_pi_state(pi_state); | |
925 | spin_unlock(&hb->lock); | |
926 | ||
927 | rt_mutex_futex_unlock(&pi_state->pi_mutex); | |
928 | put_pi_state(pi_state); | |
929 | ||
930 | raw_spin_lock_irq(&curr->pi_lock); | |
931 | } | |
932 | raw_spin_unlock_irq(&curr->pi_lock); | |
933 | } | |
934 | ||
935 | /* | |
936 | * We need to check the following states: | |
937 | * | |
938 | * Waiter | pi_state | pi->owner | uTID | uODIED | ? | |
939 | * | |
940 | * [1] NULL | --- | --- | 0 | 0/1 | Valid | |
941 | * [2] NULL | --- | --- | >0 | 0/1 | Valid | |
942 | * | |
943 | * [3] Found | NULL | -- | Any | 0/1 | Invalid | |
944 | * | |
945 | * [4] Found | Found | NULL | 0 | 1 | Valid | |
946 | * [5] Found | Found | NULL | >0 | 1 | Invalid | |
947 | * | |
948 | * [6] Found | Found | task | 0 | 1 | Valid | |
949 | * | |
950 | * [7] Found | Found | NULL | Any | 0 | Invalid | |
951 | * | |
952 | * [8] Found | Found | task | ==taskTID | 0/1 | Valid | |
953 | * [9] Found | Found | task | 0 | 0 | Invalid | |
954 | * [10] Found | Found | task | !=taskTID | 0/1 | Invalid | |
955 | * | |
956 | * [1] Indicates that the kernel can acquire the futex atomically. We | |
957 | * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit. | |
958 | * | |
959 | * [2] Valid, if TID does not belong to a kernel thread. If no matching | |
960 | * thread is found then it indicates that the owner TID has died. | |
961 | * | |
962 | * [3] Invalid. The waiter is queued on a non PI futex | |
963 | * | |
964 | * [4] Valid state after exit_robust_list(), which sets the user space | |
965 | * value to FUTEX_WAITERS | FUTEX_OWNER_DIED. | |
966 | * | |
967 | * [5] The user space value got manipulated between exit_robust_list() | |
968 | * and exit_pi_state_list() | |
969 | * | |
970 | * [6] Valid state after exit_pi_state_list() which sets the new owner in | |
971 | * the pi_state but cannot access the user space value. | |
972 | * | |
973 | * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set. | |
974 | * | |
975 | * [8] Owner and user space value match | |
976 | * | |
977 | * [9] There is no transient state which sets the user space TID to 0 | |
978 | * except exit_robust_list(), but this is indicated by the | |
979 | * FUTEX_OWNER_DIED bit. See [4] | |
980 | * | |
981 | * [10] There is no transient state which leaves owner and user space | |
982 | * TID out of sync. | |
983 | * | |
984 | * | |
985 | * Serialization and lifetime rules: | |
986 | * | |
987 | * hb->lock: | |
988 | * | |
989 | * hb -> futex_q, relation | |
990 | * futex_q -> pi_state, relation | |
991 | * | |
992 | * (cannot be raw because hb can contain arbitrary amount | |
993 | * of futex_q's) | |
994 | * | |
995 | * pi_mutex->wait_lock: | |
996 | * | |
997 | * {uval, pi_state} | |
998 | * | |
999 | * (and pi_mutex 'obviously') | |
1000 | * | |
1001 | * p->pi_lock: | |
1002 | * | |
1003 | * p->pi_state_list -> pi_state->list, relation | |
1004 | * | |
1005 | * pi_state->refcount: | |
1006 | * | |
1007 | * pi_state lifetime | |
1008 | * | |
1009 | * | |
1010 | * Lock order: | |
1011 | * | |
1012 | * hb->lock | |
1013 | * pi_mutex->wait_lock | |
1014 | * p->pi_lock | |
1015 | * | |
1016 | */ | |
1017 | ||
1018 | /* | |
1019 | * Validate that the existing waiter has a pi_state and sanity check | |
1020 | * the pi_state against the user space value. If correct, attach to | |
1021 | * it. | |
1022 | */ | |
1023 | static int attach_to_pi_state(u32 __user *uaddr, u32 uval, | |
1024 | struct futex_pi_state *pi_state, | |
1025 | struct futex_pi_state **ps) | |
1026 | { | |
1027 | pid_t pid = uval & FUTEX_TID_MASK; | |
1028 | u32 uval2; | |
1029 | int ret; | |
1030 | ||
1031 | /* | |
1032 | * Userspace might have messed up non-PI and PI futexes [3] | |
1033 | */ | |
1034 | if (unlikely(!pi_state)) | |
1035 | return -EINVAL; | |
1036 | ||
1037 | /* | |
1038 | * We get here with hb->lock held, and having found a | |
1039 | * futex_top_waiter(). This means that futex_lock_pi() of said futex_q | |
1040 | * has dropped the hb->lock in between queue_me() and unqueue_me_pi(), | |
1041 | * which in turn means that futex_lock_pi() still has a reference on | |
1042 | * our pi_state. | |
1043 | * | |
1044 | * The waiter holding a reference on @pi_state also protects against | |
1045 | * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi() | |
1046 | * and futex_wait_requeue_pi() as it cannot go to 0 and consequently | |
1047 | * free pi_state before we can take a reference ourselves. | |
1048 | */ | |
1049 | WARN_ON(!atomic_read(&pi_state->refcount)); | |
1050 | ||
1051 | /* | |
1052 | * Now that we have a pi_state, we can acquire wait_lock | |
1053 | * and do the state validation. | |
1054 | */ | |
1055 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); | |
1056 | ||
1057 | /* | |
1058 | * Since {uval, pi_state} is serialized by wait_lock, and our current | |
1059 | * uval was read without holding it, it can have changed. Verify it | |
1060 | * still is what we expect it to be, otherwise retry the entire | |
1061 | * operation. | |
1062 | */ | |
1063 | if (get_futex_value_locked(&uval2, uaddr)) | |
1064 | goto out_efault; | |
1065 | ||
1066 | if (uval != uval2) | |
1067 | goto out_eagain; | |
1068 | ||
1069 | /* | |
1070 | * Handle the owner died case: | |
1071 | */ | |
1072 | if (uval & FUTEX_OWNER_DIED) { | |
1073 | /* | |
1074 | * exit_pi_state_list sets owner to NULL and wakes the | |
1075 | * topmost waiter. The task which acquires the | |
1076 | * pi_state->rt_mutex will fixup owner. | |
1077 | */ | |
1078 | if (!pi_state->owner) { | |
1079 | /* | |
1080 | * No pi state owner, but the user space TID | |
1081 | * is not 0. Inconsistent state. [5] | |
1082 | */ | |
1083 | if (pid) | |
1084 | goto out_einval; | |
1085 | /* | |
1086 | * Take a ref on the state and return success. [4] | |
1087 | */ | |
1088 | goto out_attach; | |
1089 | } | |
1090 | ||
1091 | /* | |
1092 | * If TID is 0, then either the dying owner has not | |
1093 | * yet executed exit_pi_state_list() or some waiter | |
1094 | * acquired the rtmutex in the pi state, but did not | |
1095 | * yet fixup the TID in user space. | |
1096 | * | |
1097 | * Take a ref on the state and return success. [6] | |
1098 | */ | |
1099 | if (!pid) | |
1100 | goto out_attach; | |
1101 | } else { | |
1102 | /* | |
1103 | * If the owner died bit is not set, then the pi_state | |
1104 | * must have an owner. [7] | |
1105 | */ | |
1106 | if (!pi_state->owner) | |
1107 | goto out_einval; | |
1108 | } | |
1109 | ||
1110 | /* | |
1111 | * Bail out if user space manipulated the futex value. If pi | |
1112 | * state exists then the owner TID must be the same as the | |
1113 | * user space TID. [9/10] | |
1114 | */ | |
1115 | if (pid != task_pid_vnr(pi_state->owner)) | |
1116 | goto out_einval; | |
1117 | ||
1118 | out_attach: | |
1119 | get_pi_state(pi_state); | |
1120 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
1121 | *ps = pi_state; | |
1122 | return 0; | |
1123 | ||
1124 | out_einval: | |
1125 | ret = -EINVAL; | |
1126 | goto out_error; | |
1127 | ||
1128 | out_eagain: | |
1129 | ret = -EAGAIN; | |
1130 | goto out_error; | |
1131 | ||
1132 | out_efault: | |
1133 | ret = -EFAULT; | |
1134 | goto out_error; | |
1135 | ||
1136 | out_error: | |
1137 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
1138 | return ret; | |
1139 | } | |
1140 | ||
1141 | /* | |
1142 | * Lookup the task for the TID provided from user space and attach to | |
1143 | * it after doing proper sanity checks. | |
1144 | */ | |
1145 | static int attach_to_pi_owner(u32 uval, union futex_key *key, | |
1146 | struct futex_pi_state **ps) | |
1147 | { | |
1148 | pid_t pid = uval & FUTEX_TID_MASK; | |
1149 | struct futex_pi_state *pi_state; | |
1150 | struct task_struct *p; | |
1151 | ||
1152 | /* | |
1153 | * We are the first waiter - try to look up the real owner and attach | |
1154 | * the new pi_state to it, but bail out when TID = 0 [1] | |
1155 | */ | |
1156 | if (!pid) | |
1157 | return -ESRCH; | |
1158 | p = futex_find_get_task(pid); | |
1159 | if (!p) | |
1160 | return -ESRCH; | |
1161 | ||
1162 | if (unlikely(p->flags & PF_KTHREAD)) { | |
1163 | put_task_struct(p); | |
1164 | return -EPERM; | |
1165 | } | |
1166 | ||
1167 | /* | |
1168 | * We need to look at the task state flags to figure out, | |
1169 | * whether the task is exiting. To protect against the do_exit | |
1170 | * change of the task flags, we do this protected by | |
1171 | * p->pi_lock: | |
1172 | */ | |
1173 | raw_spin_lock_irq(&p->pi_lock); | |
1174 | if (unlikely(p->flags & PF_EXITING)) { | |
1175 | /* | |
1176 | * The task is on the way out. When PF_EXITPIDONE is | |
1177 | * set, we know that the task has finished the | |
1178 | * cleanup: | |
1179 | */ | |
1180 | int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN; | |
1181 | ||
1182 | raw_spin_unlock_irq(&p->pi_lock); | |
1183 | put_task_struct(p); | |
1184 | return ret; | |
1185 | } | |
1186 | ||
1187 | /* | |
1188 | * No existing pi state. First waiter. [2] | |
1189 | * | |
1190 | * This creates pi_state, we have hb->lock held, this means nothing can | |
1191 | * observe this state, wait_lock is irrelevant. | |
1192 | */ | |
1193 | pi_state = alloc_pi_state(); | |
1194 | ||
1195 | /* | |
1196 | * Initialize the pi_mutex in locked state and make @p | |
1197 | * the owner of it: | |
1198 | */ | |
1199 | rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p); | |
1200 | ||
1201 | /* Store the key for possible exit cleanups: */ | |
1202 | pi_state->key = *key; | |
1203 | ||
1204 | WARN_ON(!list_empty(&pi_state->list)); | |
1205 | list_add(&pi_state->list, &p->pi_state_list); | |
1206 | pi_state->owner = p; | |
1207 | raw_spin_unlock_irq(&p->pi_lock); | |
1208 | ||
1209 | put_task_struct(p); | |
1210 | ||
1211 | *ps = pi_state; | |
1212 | ||
1213 | return 0; | |
1214 | } | |
1215 | ||
1216 | static int lookup_pi_state(u32 __user *uaddr, u32 uval, | |
1217 | struct futex_hash_bucket *hb, | |
1218 | union futex_key *key, struct futex_pi_state **ps) | |
1219 | { | |
1220 | struct futex_q *top_waiter = futex_top_waiter(hb, key); | |
1221 | ||
1222 | /* | |
1223 | * If there is a waiter on that futex, validate it and | |
1224 | * attach to the pi_state when the validation succeeds. | |
1225 | */ | |
1226 | if (top_waiter) | |
1227 | return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps); | |
1228 | ||
1229 | /* | |
1230 | * We are the first waiter - try to look up the owner based on | |
1231 | * @uval and attach to it. | |
1232 | */ | |
1233 | return attach_to_pi_owner(uval, key, ps); | |
1234 | } | |
1235 | ||
1236 | static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval) | |
1237 | { | |
1238 | u32 uninitialized_var(curval); | |
1239 | ||
1240 | if (unlikely(should_fail_futex(true))) | |
1241 | return -EFAULT; | |
1242 | ||
1243 | if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))) | |
1244 | return -EFAULT; | |
1245 | ||
1246 | /* If user space value changed, let the caller retry */ | |
1247 | return curval != uval ? -EAGAIN : 0; | |
1248 | } | |
1249 | ||
1250 | /** | |
1251 | * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex | |
1252 | * @uaddr: the pi futex user address | |
1253 | * @hb: the pi futex hash bucket | |
1254 | * @key: the futex key associated with uaddr and hb | |
1255 | * @ps: the pi_state pointer where we store the result of the | |
1256 | * lookup | |
1257 | * @task: the task to perform the atomic lock work for. This will | |
1258 | * be "current" except in the case of requeue pi. | |
1259 | * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0) | |
1260 | * | |
1261 | * Return: | |
1262 | * 0 - ready to wait; | |
1263 | * 1 - acquired the lock; | |
1264 | * <0 - error | |
1265 | * | |
1266 | * The hb->lock and futex_key refs shall be held by the caller. | |
1267 | */ | |
1268 | static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb, | |
1269 | union futex_key *key, | |
1270 | struct futex_pi_state **ps, | |
1271 | struct task_struct *task, int set_waiters) | |
1272 | { | |
1273 | u32 uval, newval, vpid = task_pid_vnr(task); | |
1274 | struct futex_q *top_waiter; | |
1275 | int ret; | |
1276 | ||
1277 | /* | |
1278 | * Read the user space value first so we can validate a few | |
1279 | * things before proceeding further. | |
1280 | */ | |
1281 | if (get_futex_value_locked(&uval, uaddr)) | |
1282 | return -EFAULT; | |
1283 | ||
1284 | if (unlikely(should_fail_futex(true))) | |
1285 | return -EFAULT; | |
1286 | ||
1287 | /* | |
1288 | * Detect deadlocks. | |
1289 | */ | |
1290 | if ((unlikely((uval & FUTEX_TID_MASK) == vpid))) | |
1291 | return -EDEADLK; | |
1292 | ||
1293 | if ((unlikely(should_fail_futex(true)))) | |
1294 | return -EDEADLK; | |
1295 | ||
1296 | /* | |
1297 | * Lookup existing state first. If it exists, try to attach to | |
1298 | * its pi_state. | |
1299 | */ | |
1300 | top_waiter = futex_top_waiter(hb, key); | |
1301 | if (top_waiter) | |
1302 | return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps); | |
1303 | ||
1304 | /* | |
1305 | * No waiter and user TID is 0. We are here because the | |
1306 | * waiters or the owner died bit is set or called from | |
1307 | * requeue_cmp_pi or for whatever reason something took the | |
1308 | * syscall. | |
1309 | */ | |
1310 | if (!(uval & FUTEX_TID_MASK)) { | |
1311 | /* | |
1312 | * We take over the futex. No other waiters and the user space | |
1313 | * TID is 0. We preserve the owner died bit. | |
1314 | */ | |
1315 | newval = uval & FUTEX_OWNER_DIED; | |
1316 | newval |= vpid; | |
1317 | ||
1318 | /* The futex requeue_pi code can enforce the waiters bit */ | |
1319 | if (set_waiters) | |
1320 | newval |= FUTEX_WAITERS; | |
1321 | ||
1322 | ret = lock_pi_update_atomic(uaddr, uval, newval); | |
1323 | /* If the take over worked, return 1 */ | |
1324 | return ret < 0 ? ret : 1; | |
1325 | } | |
1326 | ||
1327 | /* | |
1328 | * First waiter. Set the waiters bit before attaching ourself to | |
1329 | * the owner. If owner tries to unlock, it will be forced into | |
1330 | * the kernel and blocked on hb->lock. | |
1331 | */ | |
1332 | newval = uval | FUTEX_WAITERS; | |
1333 | ret = lock_pi_update_atomic(uaddr, uval, newval); | |
1334 | if (ret) | |
1335 | return ret; | |
1336 | /* | |
1337 | * If the update of the user space value succeeded, we try to | |
1338 | * attach to the owner. If that fails, no harm done, we only | |
1339 | * set the FUTEX_WAITERS bit in the user space variable. | |
1340 | */ | |
1341 | return attach_to_pi_owner(uval, key, ps); | |
1342 | } | |
1343 | ||
1344 | /** | |
1345 | * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket | |
1346 | * @q: The futex_q to unqueue | |
1347 | * | |
1348 | * The q->lock_ptr must not be NULL and must be held by the caller. | |
1349 | */ | |
1350 | static void __unqueue_futex(struct futex_q *q) | |
1351 | { | |
1352 | struct futex_hash_bucket *hb; | |
1353 | ||
1354 | if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr)) | |
1355 | || WARN_ON(plist_node_empty(&q->list))) | |
1356 | return; | |
1357 | ||
1358 | hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock); | |
1359 | plist_del(&q->list, &hb->chain); | |
1360 | hb_waiters_dec(hb); | |
1361 | } | |
1362 | ||
1363 | /* | |
1364 | * The hash bucket lock must be held when this is called. | |
1365 | * Afterwards, the futex_q must not be accessed. Callers | |
1366 | * must ensure to later call wake_up_q() for the actual | |
1367 | * wakeups to occur. | |
1368 | */ | |
1369 | static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q) | |
1370 | { | |
1371 | struct task_struct *p = q->task; | |
1372 | ||
1373 | if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n")) | |
1374 | return; | |
1375 | ||
1376 | /* | |
1377 | * Queue the task for later wakeup for after we've released | |
1378 | * the hb->lock. wake_q_add() grabs reference to p. | |
1379 | */ | |
1380 | wake_q_add(wake_q, p); | |
1381 | __unqueue_futex(q); | |
1382 | /* | |
1383 | * The waiting task can free the futex_q as soon as q->lock_ptr = NULL | |
1384 | * is written, without taking any locks. This is possible in the event | |
1385 | * of a spurious wakeup, for example. A memory barrier is required here | |
1386 | * to prevent the following store to lock_ptr from getting ahead of the | |
1387 | * plist_del in __unqueue_futex(). | |
1388 | */ | |
1389 | smp_store_release(&q->lock_ptr, NULL); | |
1390 | } | |
1391 | ||
1392 | /* | |
1393 | * Caller must hold a reference on @pi_state. | |
1394 | */ | |
1395 | static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state) | |
1396 | { | |
1397 | u32 uninitialized_var(curval), newval; | |
1398 | struct task_struct *new_owner; | |
1399 | bool postunlock = false; | |
1400 | DEFINE_WAKE_Q(wake_q); | |
1401 | int ret = 0; | |
1402 | ||
1403 | new_owner = rt_mutex_next_owner(&pi_state->pi_mutex); | |
1404 | if (WARN_ON_ONCE(!new_owner)) { | |
1405 | /* | |
1406 | * As per the comment in futex_unlock_pi() this should not happen. | |
1407 | * | |
1408 | * When this happens, give up our locks and try again, giving | |
1409 | * the futex_lock_pi() instance time to complete, either by | |
1410 | * waiting on the rtmutex or removing itself from the futex | |
1411 | * queue. | |
1412 | */ | |
1413 | ret = -EAGAIN; | |
1414 | goto out_unlock; | |
1415 | } | |
1416 | ||
1417 | /* | |
1418 | * We pass it to the next owner. The WAITERS bit is always kept | |
1419 | * enabled while there is PI state around. We cleanup the owner | |
1420 | * died bit, because we are the owner. | |
1421 | */ | |
1422 | newval = FUTEX_WAITERS | task_pid_vnr(new_owner); | |
1423 | ||
1424 | if (unlikely(should_fail_futex(true))) | |
1425 | ret = -EFAULT; | |
1426 | ||
1427 | if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) { | |
1428 | ret = -EFAULT; | |
1429 | ||
1430 | } else if (curval != uval) { | |
1431 | /* | |
1432 | * If a unconditional UNLOCK_PI operation (user space did not | |
1433 | * try the TID->0 transition) raced with a waiter setting the | |
1434 | * FUTEX_WAITERS flag between get_user() and locking the hash | |
1435 | * bucket lock, retry the operation. | |
1436 | */ | |
1437 | if ((FUTEX_TID_MASK & curval) == uval) | |
1438 | ret = -EAGAIN; | |
1439 | else | |
1440 | ret = -EINVAL; | |
1441 | } | |
1442 | ||
1443 | if (ret) | |
1444 | goto out_unlock; | |
1445 | ||
1446 | /* | |
1447 | * This is a point of no return; once we modify the uval there is no | |
1448 | * going back and subsequent operations must not fail. | |
1449 | */ | |
1450 | ||
1451 | raw_spin_lock(&pi_state->owner->pi_lock); | |
1452 | WARN_ON(list_empty(&pi_state->list)); | |
1453 | list_del_init(&pi_state->list); | |
1454 | raw_spin_unlock(&pi_state->owner->pi_lock); | |
1455 | ||
1456 | raw_spin_lock(&new_owner->pi_lock); | |
1457 | WARN_ON(!list_empty(&pi_state->list)); | |
1458 | list_add(&pi_state->list, &new_owner->pi_state_list); | |
1459 | pi_state->owner = new_owner; | |
1460 | raw_spin_unlock(&new_owner->pi_lock); | |
1461 | ||
1462 | postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q); | |
1463 | ||
1464 | out_unlock: | |
1465 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
1466 | ||
1467 | if (postunlock) | |
1468 | rt_mutex_postunlock(&wake_q); | |
1469 | ||
1470 | return ret; | |
1471 | } | |
1472 | ||
1473 | /* | |
1474 | * Express the locking dependencies for lockdep: | |
1475 | */ | |
1476 | static inline void | |
1477 | double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2) | |
1478 | { | |
1479 | if (hb1 <= hb2) { | |
1480 | spin_lock(&hb1->lock); | |
1481 | if (hb1 < hb2) | |
1482 | spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING); | |
1483 | } else { /* hb1 > hb2 */ | |
1484 | spin_lock(&hb2->lock); | |
1485 | spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING); | |
1486 | } | |
1487 | } | |
1488 | ||
1489 | static inline void | |
1490 | double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2) | |
1491 | { | |
1492 | spin_unlock(&hb1->lock); | |
1493 | if (hb1 != hb2) | |
1494 | spin_unlock(&hb2->lock); | |
1495 | } | |
1496 | ||
1497 | /* | |
1498 | * Wake up waiters matching bitset queued on this futex (uaddr). | |
1499 | */ | |
1500 | static int | |
1501 | futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset) | |
1502 | { | |
1503 | struct futex_hash_bucket *hb; | |
1504 | struct futex_q *this, *next; | |
1505 | union futex_key key = FUTEX_KEY_INIT; | |
1506 | int ret; | |
1507 | DEFINE_WAKE_Q(wake_q); | |
1508 | ||
1509 | if (!bitset) | |
1510 | return -EINVAL; | |
1511 | ||
1512 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ); | |
1513 | if (unlikely(ret != 0)) | |
1514 | goto out; | |
1515 | ||
1516 | hb = hash_futex(&key); | |
1517 | ||
1518 | /* Make sure we really have tasks to wakeup */ | |
1519 | if (!hb_waiters_pending(hb)) | |
1520 | goto out_put_key; | |
1521 | ||
1522 | spin_lock(&hb->lock); | |
1523 | ||
1524 | plist_for_each_entry_safe(this, next, &hb->chain, list) { | |
1525 | if (match_futex (&this->key, &key)) { | |
1526 | if (this->pi_state || this->rt_waiter) { | |
1527 | ret = -EINVAL; | |
1528 | break; | |
1529 | } | |
1530 | ||
1531 | /* Check if one of the bits is set in both bitsets */ | |
1532 | if (!(this->bitset & bitset)) | |
1533 | continue; | |
1534 | ||
1535 | mark_wake_futex(&wake_q, this); | |
1536 | if (++ret >= nr_wake) | |
1537 | break; | |
1538 | } | |
1539 | } | |
1540 | ||
1541 | spin_unlock(&hb->lock); | |
1542 | wake_up_q(&wake_q); | |
1543 | out_put_key: | |
1544 | put_futex_key(&key); | |
1545 | out: | |
1546 | return ret; | |
1547 | } | |
1548 | ||
1549 | /* | |
1550 | * Wake up all waiters hashed on the physical page that is mapped | |
1551 | * to this virtual address: | |
1552 | */ | |
1553 | static int | |
1554 | futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2, | |
1555 | int nr_wake, int nr_wake2, int op) | |
1556 | { | |
1557 | union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT; | |
1558 | struct futex_hash_bucket *hb1, *hb2; | |
1559 | struct futex_q *this, *next; | |
1560 | int ret, op_ret; | |
1561 | DEFINE_WAKE_Q(wake_q); | |
1562 | ||
1563 | retry: | |
1564 | ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ); | |
1565 | if (unlikely(ret != 0)) | |
1566 | goto out; | |
1567 | ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE); | |
1568 | if (unlikely(ret != 0)) | |
1569 | goto out_put_key1; | |
1570 | ||
1571 | hb1 = hash_futex(&key1); | |
1572 | hb2 = hash_futex(&key2); | |
1573 | ||
1574 | retry_private: | |
1575 | double_lock_hb(hb1, hb2); | |
1576 | op_ret = futex_atomic_op_inuser(op, uaddr2); | |
1577 | if (unlikely(op_ret < 0)) { | |
1578 | ||
1579 | double_unlock_hb(hb1, hb2); | |
1580 | ||
1581 | #ifndef CONFIG_MMU | |
1582 | /* | |
1583 | * we don't get EFAULT from MMU faults if we don't have an MMU, | |
1584 | * but we might get them from range checking | |
1585 | */ | |
1586 | ret = op_ret; | |
1587 | goto out_put_keys; | |
1588 | #endif | |
1589 | ||
1590 | if (unlikely(op_ret != -EFAULT)) { | |
1591 | ret = op_ret; | |
1592 | goto out_put_keys; | |
1593 | } | |
1594 | ||
1595 | ret = fault_in_user_writeable(uaddr2); | |
1596 | if (ret) | |
1597 | goto out_put_keys; | |
1598 | ||
1599 | if (!(flags & FLAGS_SHARED)) | |
1600 | goto retry_private; | |
1601 | ||
1602 | put_futex_key(&key2); | |
1603 | put_futex_key(&key1); | |
1604 | goto retry; | |
1605 | } | |
1606 | ||
1607 | plist_for_each_entry_safe(this, next, &hb1->chain, list) { | |
1608 | if (match_futex (&this->key, &key1)) { | |
1609 | if (this->pi_state || this->rt_waiter) { | |
1610 | ret = -EINVAL; | |
1611 | goto out_unlock; | |
1612 | } | |
1613 | mark_wake_futex(&wake_q, this); | |
1614 | if (++ret >= nr_wake) | |
1615 | break; | |
1616 | } | |
1617 | } | |
1618 | ||
1619 | if (op_ret > 0) { | |
1620 | op_ret = 0; | |
1621 | plist_for_each_entry_safe(this, next, &hb2->chain, list) { | |
1622 | if (match_futex (&this->key, &key2)) { | |
1623 | if (this->pi_state || this->rt_waiter) { | |
1624 | ret = -EINVAL; | |
1625 | goto out_unlock; | |
1626 | } | |
1627 | mark_wake_futex(&wake_q, this); | |
1628 | if (++op_ret >= nr_wake2) | |
1629 | break; | |
1630 | } | |
1631 | } | |
1632 | ret += op_ret; | |
1633 | } | |
1634 | ||
1635 | out_unlock: | |
1636 | double_unlock_hb(hb1, hb2); | |
1637 | wake_up_q(&wake_q); | |
1638 | out_put_keys: | |
1639 | put_futex_key(&key2); | |
1640 | out_put_key1: | |
1641 | put_futex_key(&key1); | |
1642 | out: | |
1643 | return ret; | |
1644 | } | |
1645 | ||
1646 | /** | |
1647 | * requeue_futex() - Requeue a futex_q from one hb to another | |
1648 | * @q: the futex_q to requeue | |
1649 | * @hb1: the source hash_bucket | |
1650 | * @hb2: the target hash_bucket | |
1651 | * @key2: the new key for the requeued futex_q | |
1652 | */ | |
1653 | static inline | |
1654 | void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1, | |
1655 | struct futex_hash_bucket *hb2, union futex_key *key2) | |
1656 | { | |
1657 | ||
1658 | /* | |
1659 | * If key1 and key2 hash to the same bucket, no need to | |
1660 | * requeue. | |
1661 | */ | |
1662 | if (likely(&hb1->chain != &hb2->chain)) { | |
1663 | plist_del(&q->list, &hb1->chain); | |
1664 | hb_waiters_dec(hb1); | |
1665 | hb_waiters_inc(hb2); | |
1666 | plist_add(&q->list, &hb2->chain); | |
1667 | q->lock_ptr = &hb2->lock; | |
1668 | } | |
1669 | get_futex_key_refs(key2); | |
1670 | q->key = *key2; | |
1671 | } | |
1672 | ||
1673 | /** | |
1674 | * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue | |
1675 | * @q: the futex_q | |
1676 | * @key: the key of the requeue target futex | |
1677 | * @hb: the hash_bucket of the requeue target futex | |
1678 | * | |
1679 | * During futex_requeue, with requeue_pi=1, it is possible to acquire the | |
1680 | * target futex if it is uncontended or via a lock steal. Set the futex_q key | |
1681 | * to the requeue target futex so the waiter can detect the wakeup on the right | |
1682 | * futex, but remove it from the hb and NULL the rt_waiter so it can detect | |
1683 | * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock | |
1684 | * to protect access to the pi_state to fixup the owner later. Must be called | |
1685 | * with both q->lock_ptr and hb->lock held. | |
1686 | */ | |
1687 | static inline | |
1688 | void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key, | |
1689 | struct futex_hash_bucket *hb) | |
1690 | { | |
1691 | get_futex_key_refs(key); | |
1692 | q->key = *key; | |
1693 | ||
1694 | __unqueue_futex(q); | |
1695 | ||
1696 | WARN_ON(!q->rt_waiter); | |
1697 | q->rt_waiter = NULL; | |
1698 | ||
1699 | q->lock_ptr = &hb->lock; | |
1700 | ||
1701 | wake_up_state(q->task, TASK_NORMAL); | |
1702 | } | |
1703 | ||
1704 | /** | |
1705 | * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter | |
1706 | * @pifutex: the user address of the to futex | |
1707 | * @hb1: the from futex hash bucket, must be locked by the caller | |
1708 | * @hb2: the to futex hash bucket, must be locked by the caller | |
1709 | * @key1: the from futex key | |
1710 | * @key2: the to futex key | |
1711 | * @ps: address to store the pi_state pointer | |
1712 | * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0) | |
1713 | * | |
1714 | * Try and get the lock on behalf of the top waiter if we can do it atomically. | |
1715 | * Wake the top waiter if we succeed. If the caller specified set_waiters, | |
1716 | * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit. | |
1717 | * hb1 and hb2 must be held by the caller. | |
1718 | * | |
1719 | * Return: | |
1720 | * 0 - failed to acquire the lock atomically; | |
1721 | * >0 - acquired the lock, return value is vpid of the top_waiter | |
1722 | * <0 - error | |
1723 | */ | |
1724 | static int futex_proxy_trylock_atomic(u32 __user *pifutex, | |
1725 | struct futex_hash_bucket *hb1, | |
1726 | struct futex_hash_bucket *hb2, | |
1727 | union futex_key *key1, union futex_key *key2, | |
1728 | struct futex_pi_state **ps, int set_waiters) | |
1729 | { | |
1730 | struct futex_q *top_waiter = NULL; | |
1731 | u32 curval; | |
1732 | int ret, vpid; | |
1733 | ||
1734 | if (get_futex_value_locked(&curval, pifutex)) | |
1735 | return -EFAULT; | |
1736 | ||
1737 | if (unlikely(should_fail_futex(true))) | |
1738 | return -EFAULT; | |
1739 | ||
1740 | /* | |
1741 | * Find the top_waiter and determine if there are additional waiters. | |
1742 | * If the caller intends to requeue more than 1 waiter to pifutex, | |
1743 | * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now, | |
1744 | * as we have means to handle the possible fault. If not, don't set | |
1745 | * the bit unecessarily as it will force the subsequent unlock to enter | |
1746 | * the kernel. | |
1747 | */ | |
1748 | top_waiter = futex_top_waiter(hb1, key1); | |
1749 | ||
1750 | /* There are no waiters, nothing for us to do. */ | |
1751 | if (!top_waiter) | |
1752 | return 0; | |
1753 | ||
1754 | /* Ensure we requeue to the expected futex. */ | |
1755 | if (!match_futex(top_waiter->requeue_pi_key, key2)) | |
1756 | return -EINVAL; | |
1757 | ||
1758 | /* | |
1759 | * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in | |
1760 | * the contended case or if set_waiters is 1. The pi_state is returned | |
1761 | * in ps in contended cases. | |
1762 | */ | |
1763 | vpid = task_pid_vnr(top_waiter->task); | |
1764 | ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task, | |
1765 | set_waiters); | |
1766 | if (ret == 1) { | |
1767 | requeue_pi_wake_futex(top_waiter, key2, hb2); | |
1768 | return vpid; | |
1769 | } | |
1770 | return ret; | |
1771 | } | |
1772 | ||
1773 | /** | |
1774 | * futex_requeue() - Requeue waiters from uaddr1 to uaddr2 | |
1775 | * @uaddr1: source futex user address | |
1776 | * @flags: futex flags (FLAGS_SHARED, etc.) | |
1777 | * @uaddr2: target futex user address | |
1778 | * @nr_wake: number of waiters to wake (must be 1 for requeue_pi) | |
1779 | * @nr_requeue: number of waiters to requeue (0-INT_MAX) | |
1780 | * @cmpval: @uaddr1 expected value (or %NULL) | |
1781 | * @requeue_pi: if we are attempting to requeue from a non-pi futex to a | |
1782 | * pi futex (pi to pi requeue is not supported) | |
1783 | * | |
1784 | * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire | |
1785 | * uaddr2 atomically on behalf of the top waiter. | |
1786 | * | |
1787 | * Return: | |
1788 | * >=0 - on success, the number of tasks requeued or woken; | |
1789 | * <0 - on error | |
1790 | */ | |
1791 | static int futex_requeue(u32 __user *uaddr1, unsigned int flags, | |
1792 | u32 __user *uaddr2, int nr_wake, int nr_requeue, | |
1793 | u32 *cmpval, int requeue_pi) | |
1794 | { | |
1795 | union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT; | |
1796 | int drop_count = 0, task_count = 0, ret; | |
1797 | struct futex_pi_state *pi_state = NULL; | |
1798 | struct futex_hash_bucket *hb1, *hb2; | |
1799 | struct futex_q *this, *next; | |
1800 | DEFINE_WAKE_Q(wake_q); | |
1801 | ||
1802 | if (requeue_pi) { | |
1803 | /* | |
1804 | * Requeue PI only works on two distinct uaddrs. This | |
1805 | * check is only valid for private futexes. See below. | |
1806 | */ | |
1807 | if (uaddr1 == uaddr2) | |
1808 | return -EINVAL; | |
1809 | ||
1810 | /* | |
1811 | * requeue_pi requires a pi_state, try to allocate it now | |
1812 | * without any locks in case it fails. | |
1813 | */ | |
1814 | if (refill_pi_state_cache()) | |
1815 | return -ENOMEM; | |
1816 | /* | |
1817 | * requeue_pi must wake as many tasks as it can, up to nr_wake | |
1818 | * + nr_requeue, since it acquires the rt_mutex prior to | |
1819 | * returning to userspace, so as to not leave the rt_mutex with | |
1820 | * waiters and no owner. However, second and third wake-ups | |
1821 | * cannot be predicted as they involve race conditions with the | |
1822 | * first wake and a fault while looking up the pi_state. Both | |
1823 | * pthread_cond_signal() and pthread_cond_broadcast() should | |
1824 | * use nr_wake=1. | |
1825 | */ | |
1826 | if (nr_wake != 1) | |
1827 | return -EINVAL; | |
1828 | } | |
1829 | ||
1830 | retry: | |
1831 | ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ); | |
1832 | if (unlikely(ret != 0)) | |
1833 | goto out; | |
1834 | ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, | |
1835 | requeue_pi ? VERIFY_WRITE : VERIFY_READ); | |
1836 | if (unlikely(ret != 0)) | |
1837 | goto out_put_key1; | |
1838 | ||
1839 | /* | |
1840 | * The check above which compares uaddrs is not sufficient for | |
1841 | * shared futexes. We need to compare the keys: | |
1842 | */ | |
1843 | if (requeue_pi && match_futex(&key1, &key2)) { | |
1844 | ret = -EINVAL; | |
1845 | goto out_put_keys; | |
1846 | } | |
1847 | ||
1848 | hb1 = hash_futex(&key1); | |
1849 | hb2 = hash_futex(&key2); | |
1850 | ||
1851 | retry_private: | |
1852 | hb_waiters_inc(hb2); | |
1853 | double_lock_hb(hb1, hb2); | |
1854 | ||
1855 | if (likely(cmpval != NULL)) { | |
1856 | u32 curval; | |
1857 | ||
1858 | ret = get_futex_value_locked(&curval, uaddr1); | |
1859 | ||
1860 | if (unlikely(ret)) { | |
1861 | double_unlock_hb(hb1, hb2); | |
1862 | hb_waiters_dec(hb2); | |
1863 | ||
1864 | ret = get_user(curval, uaddr1); | |
1865 | if (ret) | |
1866 | goto out_put_keys; | |
1867 | ||
1868 | if (!(flags & FLAGS_SHARED)) | |
1869 | goto retry_private; | |
1870 | ||
1871 | put_futex_key(&key2); | |
1872 | put_futex_key(&key1); | |
1873 | goto retry; | |
1874 | } | |
1875 | if (curval != *cmpval) { | |
1876 | ret = -EAGAIN; | |
1877 | goto out_unlock; | |
1878 | } | |
1879 | } | |
1880 | ||
1881 | if (requeue_pi && (task_count - nr_wake < nr_requeue)) { | |
1882 | /* | |
1883 | * Attempt to acquire uaddr2 and wake the top waiter. If we | |
1884 | * intend to requeue waiters, force setting the FUTEX_WAITERS | |
1885 | * bit. We force this here where we are able to easily handle | |
1886 | * faults rather in the requeue loop below. | |
1887 | */ | |
1888 | ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1, | |
1889 | &key2, &pi_state, nr_requeue); | |
1890 | ||
1891 | /* | |
1892 | * At this point the top_waiter has either taken uaddr2 or is | |
1893 | * waiting on it. If the former, then the pi_state will not | |
1894 | * exist yet, look it up one more time to ensure we have a | |
1895 | * reference to it. If the lock was taken, ret contains the | |
1896 | * vpid of the top waiter task. | |
1897 | * If the lock was not taken, we have pi_state and an initial | |
1898 | * refcount on it. In case of an error we have nothing. | |
1899 | */ | |
1900 | if (ret > 0) { | |
1901 | WARN_ON(pi_state); | |
1902 | drop_count++; | |
1903 | task_count++; | |
1904 | /* | |
1905 | * If we acquired the lock, then the user space value | |
1906 | * of uaddr2 should be vpid. It cannot be changed by | |
1907 | * the top waiter as it is blocked on hb2 lock if it | |
1908 | * tries to do so. If something fiddled with it behind | |
1909 | * our back the pi state lookup might unearth it. So | |
1910 | * we rather use the known value than rereading and | |
1911 | * handing potential crap to lookup_pi_state. | |
1912 | * | |
1913 | * If that call succeeds then we have pi_state and an | |
1914 | * initial refcount on it. | |
1915 | */ | |
1916 | ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state); | |
1917 | } | |
1918 | ||
1919 | switch (ret) { | |
1920 | case 0: | |
1921 | /* We hold a reference on the pi state. */ | |
1922 | break; | |
1923 | ||
1924 | /* If the above failed, then pi_state is NULL */ | |
1925 | case -EFAULT: | |
1926 | double_unlock_hb(hb1, hb2); | |
1927 | hb_waiters_dec(hb2); | |
1928 | put_futex_key(&key2); | |
1929 | put_futex_key(&key1); | |
1930 | ret = fault_in_user_writeable(uaddr2); | |
1931 | if (!ret) | |
1932 | goto retry; | |
1933 | goto out; | |
1934 | case -EAGAIN: | |
1935 | /* | |
1936 | * Two reasons for this: | |
1937 | * - Owner is exiting and we just wait for the | |
1938 | * exit to complete. | |
1939 | * - The user space value changed. | |
1940 | */ | |
1941 | double_unlock_hb(hb1, hb2); | |
1942 | hb_waiters_dec(hb2); | |
1943 | put_futex_key(&key2); | |
1944 | put_futex_key(&key1); | |
1945 | cond_resched(); | |
1946 | goto retry; | |
1947 | default: | |
1948 | goto out_unlock; | |
1949 | } | |
1950 | } | |
1951 | ||
1952 | plist_for_each_entry_safe(this, next, &hb1->chain, list) { | |
1953 | if (task_count - nr_wake >= nr_requeue) | |
1954 | break; | |
1955 | ||
1956 | if (!match_futex(&this->key, &key1)) | |
1957 | continue; | |
1958 | ||
1959 | /* | |
1960 | * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always | |
1961 | * be paired with each other and no other futex ops. | |
1962 | * | |
1963 | * We should never be requeueing a futex_q with a pi_state, | |
1964 | * which is awaiting a futex_unlock_pi(). | |
1965 | */ | |
1966 | if ((requeue_pi && !this->rt_waiter) || | |
1967 | (!requeue_pi && this->rt_waiter) || | |
1968 | this->pi_state) { | |
1969 | ret = -EINVAL; | |
1970 | break; | |
1971 | } | |
1972 | ||
1973 | /* | |
1974 | * Wake nr_wake waiters. For requeue_pi, if we acquired the | |
1975 | * lock, we already woke the top_waiter. If not, it will be | |
1976 | * woken by futex_unlock_pi(). | |
1977 | */ | |
1978 | if (++task_count <= nr_wake && !requeue_pi) { | |
1979 | mark_wake_futex(&wake_q, this); | |
1980 | continue; | |
1981 | } | |
1982 | ||
1983 | /* Ensure we requeue to the expected futex for requeue_pi. */ | |
1984 | if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) { | |
1985 | ret = -EINVAL; | |
1986 | break; | |
1987 | } | |
1988 | ||
1989 | /* | |
1990 | * Requeue nr_requeue waiters and possibly one more in the case | |
1991 | * of requeue_pi if we couldn't acquire the lock atomically. | |
1992 | */ | |
1993 | if (requeue_pi) { | |
1994 | /* | |
1995 | * Prepare the waiter to take the rt_mutex. Take a | |
1996 | * refcount on the pi_state and store the pointer in | |
1997 | * the futex_q object of the waiter. | |
1998 | */ | |
1999 | get_pi_state(pi_state); | |
2000 | this->pi_state = pi_state; | |
2001 | ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex, | |
2002 | this->rt_waiter, | |
2003 | this->task); | |
2004 | if (ret == 1) { | |
2005 | /* | |
2006 | * We got the lock. We do neither drop the | |
2007 | * refcount on pi_state nor clear | |
2008 | * this->pi_state because the waiter needs the | |
2009 | * pi_state for cleaning up the user space | |
2010 | * value. It will drop the refcount after | |
2011 | * doing so. | |
2012 | */ | |
2013 | requeue_pi_wake_futex(this, &key2, hb2); | |
2014 | drop_count++; | |
2015 | continue; | |
2016 | } else if (ret) { | |
2017 | /* | |
2018 | * rt_mutex_start_proxy_lock() detected a | |
2019 | * potential deadlock when we tried to queue | |
2020 | * that waiter. Drop the pi_state reference | |
2021 | * which we took above and remove the pointer | |
2022 | * to the state from the waiters futex_q | |
2023 | * object. | |
2024 | */ | |
2025 | this->pi_state = NULL; | |
2026 | put_pi_state(pi_state); | |
2027 | /* | |
2028 | * We stop queueing more waiters and let user | |
2029 | * space deal with the mess. | |
2030 | */ | |
2031 | break; | |
2032 | } | |
2033 | } | |
2034 | requeue_futex(this, hb1, hb2, &key2); | |
2035 | drop_count++; | |
2036 | } | |
2037 | ||
2038 | /* | |
2039 | * We took an extra initial reference to the pi_state either | |
2040 | * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We | |
2041 | * need to drop it here again. | |
2042 | */ | |
2043 | put_pi_state(pi_state); | |
2044 | ||
2045 | out_unlock: | |
2046 | double_unlock_hb(hb1, hb2); | |
2047 | wake_up_q(&wake_q); | |
2048 | hb_waiters_dec(hb2); | |
2049 | ||
2050 | /* | |
2051 | * drop_futex_key_refs() must be called outside the spinlocks. During | |
2052 | * the requeue we moved futex_q's from the hash bucket at key1 to the | |
2053 | * one at key2 and updated their key pointer. We no longer need to | |
2054 | * hold the references to key1. | |
2055 | */ | |
2056 | while (--drop_count >= 0) | |
2057 | drop_futex_key_refs(&key1); | |
2058 | ||
2059 | out_put_keys: | |
2060 | put_futex_key(&key2); | |
2061 | out_put_key1: | |
2062 | put_futex_key(&key1); | |
2063 | out: | |
2064 | return ret ? ret : task_count; | |
2065 | } | |
2066 | ||
2067 | /* The key must be already stored in q->key. */ | |
2068 | static inline struct futex_hash_bucket *queue_lock(struct futex_q *q) | |
2069 | __acquires(&hb->lock) | |
2070 | { | |
2071 | struct futex_hash_bucket *hb; | |
2072 | ||
2073 | hb = hash_futex(&q->key); | |
2074 | ||
2075 | /* | |
2076 | * Increment the counter before taking the lock so that | |
2077 | * a potential waker won't miss a to-be-slept task that is | |
2078 | * waiting for the spinlock. This is safe as all queue_lock() | |
2079 | * users end up calling queue_me(). Similarly, for housekeeping, | |
2080 | * decrement the counter at queue_unlock() when some error has | |
2081 | * occurred and we don't end up adding the task to the list. | |
2082 | */ | |
2083 | hb_waiters_inc(hb); | |
2084 | ||
2085 | q->lock_ptr = &hb->lock; | |
2086 | ||
2087 | spin_lock(&hb->lock); /* implies smp_mb(); (A) */ | |
2088 | return hb; | |
2089 | } | |
2090 | ||
2091 | static inline void | |
2092 | queue_unlock(struct futex_hash_bucket *hb) | |
2093 | __releases(&hb->lock) | |
2094 | { | |
2095 | spin_unlock(&hb->lock); | |
2096 | hb_waiters_dec(hb); | |
2097 | } | |
2098 | ||
2099 | static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb) | |
2100 | { | |
2101 | int prio; | |
2102 | ||
2103 | /* | |
2104 | * The priority used to register this element is | |
2105 | * - either the real thread-priority for the real-time threads | |
2106 | * (i.e. threads with a priority lower than MAX_RT_PRIO) | |
2107 | * - or MAX_RT_PRIO for non-RT threads. | |
2108 | * Thus, all RT-threads are woken first in priority order, and | |
2109 | * the others are woken last, in FIFO order. | |
2110 | */ | |
2111 | prio = min(current->normal_prio, MAX_RT_PRIO); | |
2112 | ||
2113 | plist_node_init(&q->list, prio); | |
2114 | plist_add(&q->list, &hb->chain); | |
2115 | q->task = current; | |
2116 | } | |
2117 | ||
2118 | /** | |
2119 | * queue_me() - Enqueue the futex_q on the futex_hash_bucket | |
2120 | * @q: The futex_q to enqueue | |
2121 | * @hb: The destination hash bucket | |
2122 | * | |
2123 | * The hb->lock must be held by the caller, and is released here. A call to | |
2124 | * queue_me() is typically paired with exactly one call to unqueue_me(). The | |
2125 | * exceptions involve the PI related operations, which may use unqueue_me_pi() | |
2126 | * or nothing if the unqueue is done as part of the wake process and the unqueue | |
2127 | * state is implicit in the state of woken task (see futex_wait_requeue_pi() for | |
2128 | * an example). | |
2129 | */ | |
2130 | static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb) | |
2131 | __releases(&hb->lock) | |
2132 | { | |
2133 | __queue_me(q, hb); | |
2134 | spin_unlock(&hb->lock); | |
2135 | } | |
2136 | ||
2137 | /** | |
2138 | * unqueue_me() - Remove the futex_q from its futex_hash_bucket | |
2139 | * @q: The futex_q to unqueue | |
2140 | * | |
2141 | * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must | |
2142 | * be paired with exactly one earlier call to queue_me(). | |
2143 | * | |
2144 | * Return: | |
2145 | * 1 - if the futex_q was still queued (and we removed unqueued it); | |
2146 | * 0 - if the futex_q was already removed by the waking thread | |
2147 | */ | |
2148 | static int unqueue_me(struct futex_q *q) | |
2149 | { | |
2150 | spinlock_t *lock_ptr; | |
2151 | int ret = 0; | |
2152 | ||
2153 | /* In the common case we don't take the spinlock, which is nice. */ | |
2154 | retry: | |
2155 | /* | |
2156 | * q->lock_ptr can change between this read and the following spin_lock. | |
2157 | * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and | |
2158 | * optimizing lock_ptr out of the logic below. | |
2159 | */ | |
2160 | lock_ptr = READ_ONCE(q->lock_ptr); | |
2161 | if (lock_ptr != NULL) { | |
2162 | spin_lock(lock_ptr); | |
2163 | /* | |
2164 | * q->lock_ptr can change between reading it and | |
2165 | * spin_lock(), causing us to take the wrong lock. This | |
2166 | * corrects the race condition. | |
2167 | * | |
2168 | * Reasoning goes like this: if we have the wrong lock, | |
2169 | * q->lock_ptr must have changed (maybe several times) | |
2170 | * between reading it and the spin_lock(). It can | |
2171 | * change again after the spin_lock() but only if it was | |
2172 | * already changed before the spin_lock(). It cannot, | |
2173 | * however, change back to the original value. Therefore | |
2174 | * we can detect whether we acquired the correct lock. | |
2175 | */ | |
2176 | if (unlikely(lock_ptr != q->lock_ptr)) { | |
2177 | spin_unlock(lock_ptr); | |
2178 | goto retry; | |
2179 | } | |
2180 | __unqueue_futex(q); | |
2181 | ||
2182 | BUG_ON(q->pi_state); | |
2183 | ||
2184 | spin_unlock(lock_ptr); | |
2185 | ret = 1; | |
2186 | } | |
2187 | ||
2188 | drop_futex_key_refs(&q->key); | |
2189 | return ret; | |
2190 | } | |
2191 | ||
2192 | /* | |
2193 | * PI futexes can not be requeued and must remove themself from the | |
2194 | * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry | |
2195 | * and dropped here. | |
2196 | */ | |
2197 | static void unqueue_me_pi(struct futex_q *q) | |
2198 | __releases(q->lock_ptr) | |
2199 | { | |
2200 | __unqueue_futex(q); | |
2201 | ||
2202 | BUG_ON(!q->pi_state); | |
2203 | put_pi_state(q->pi_state); | |
2204 | q->pi_state = NULL; | |
2205 | ||
2206 | spin_unlock(q->lock_ptr); | |
2207 | } | |
2208 | ||
2209 | /* | |
2210 | * Fixup the pi_state owner with the new owner. | |
2211 | * | |
2212 | * Must be called with hash bucket lock held and mm->sem held for non | |
2213 | * private futexes. | |
2214 | */ | |
2215 | static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q, | |
2216 | struct task_struct *newowner) | |
2217 | { | |
2218 | u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS; | |
2219 | struct futex_pi_state *pi_state = q->pi_state; | |
2220 | u32 uval, uninitialized_var(curval), newval; | |
2221 | struct task_struct *oldowner; | |
2222 | int ret; | |
2223 | ||
2224 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); | |
2225 | ||
2226 | oldowner = pi_state->owner; | |
2227 | /* Owner died? */ | |
2228 | if (!pi_state->owner) | |
2229 | newtid |= FUTEX_OWNER_DIED; | |
2230 | ||
2231 | /* | |
2232 | * We are here either because we stole the rtmutex from the | |
2233 | * previous highest priority waiter or we are the highest priority | |
2234 | * waiter but have failed to get the rtmutex the first time. | |
2235 | * | |
2236 | * We have to replace the newowner TID in the user space variable. | |
2237 | * This must be atomic as we have to preserve the owner died bit here. | |
2238 | * | |
2239 | * Note: We write the user space value _before_ changing the pi_state | |
2240 | * because we can fault here. Imagine swapped out pages or a fork | |
2241 | * that marked all the anonymous memory readonly for cow. | |
2242 | * | |
2243 | * Modifying pi_state _before_ the user space value would leave the | |
2244 | * pi_state in an inconsistent state when we fault here, because we | |
2245 | * need to drop the locks to handle the fault. This might be observed | |
2246 | * in the PID check in lookup_pi_state. | |
2247 | */ | |
2248 | retry: | |
2249 | if (get_futex_value_locked(&uval, uaddr)) | |
2250 | goto handle_fault; | |
2251 | ||
2252 | for (;;) { | |
2253 | newval = (uval & FUTEX_OWNER_DIED) | newtid; | |
2254 | ||
2255 | if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) | |
2256 | goto handle_fault; | |
2257 | if (curval == uval) | |
2258 | break; | |
2259 | uval = curval; | |
2260 | } | |
2261 | ||
2262 | /* | |
2263 | * We fixed up user space. Now we need to fix the pi_state | |
2264 | * itself. | |
2265 | */ | |
2266 | if (pi_state->owner != NULL) { | |
2267 | raw_spin_lock(&pi_state->owner->pi_lock); | |
2268 | WARN_ON(list_empty(&pi_state->list)); | |
2269 | list_del_init(&pi_state->list); | |
2270 | raw_spin_unlock(&pi_state->owner->pi_lock); | |
2271 | } | |
2272 | ||
2273 | pi_state->owner = newowner; | |
2274 | ||
2275 | raw_spin_lock(&newowner->pi_lock); | |
2276 | WARN_ON(!list_empty(&pi_state->list)); | |
2277 | list_add(&pi_state->list, &newowner->pi_state_list); | |
2278 | raw_spin_unlock(&newowner->pi_lock); | |
2279 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
2280 | ||
2281 | return 0; | |
2282 | ||
2283 | /* | |
2284 | * To handle the page fault we need to drop the locks here. That gives | |
2285 | * the other task (either the highest priority waiter itself or the | |
2286 | * task which stole the rtmutex) the chance to try the fixup of the | |
2287 | * pi_state. So once we are back from handling the fault we need to | |
2288 | * check the pi_state after reacquiring the locks and before trying to | |
2289 | * do another fixup. When the fixup has been done already we simply | |
2290 | * return. | |
2291 | * | |
2292 | * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely | |
2293 | * drop hb->lock since the caller owns the hb -> futex_q relation. | |
2294 | * Dropping the pi_mutex->wait_lock requires the state revalidate. | |
2295 | */ | |
2296 | handle_fault: | |
2297 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
2298 | spin_unlock(q->lock_ptr); | |
2299 | ||
2300 | ret = fault_in_user_writeable(uaddr); | |
2301 | ||
2302 | spin_lock(q->lock_ptr); | |
2303 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); | |
2304 | ||
2305 | /* | |
2306 | * Check if someone else fixed it for us: | |
2307 | */ | |
2308 | if (pi_state->owner != oldowner) { | |
2309 | ret = 0; | |
2310 | goto out_unlock; | |
2311 | } | |
2312 | ||
2313 | if (ret) | |
2314 | goto out_unlock; | |
2315 | ||
2316 | goto retry; | |
2317 | ||
2318 | out_unlock: | |
2319 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
2320 | return ret; | |
2321 | } | |
2322 | ||
2323 | static long futex_wait_restart(struct restart_block *restart); | |
2324 | ||
2325 | /** | |
2326 | * fixup_owner() - Post lock pi_state and corner case management | |
2327 | * @uaddr: user address of the futex | |
2328 | * @q: futex_q (contains pi_state and access to the rt_mutex) | |
2329 | * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0) | |
2330 | * | |
2331 | * After attempting to lock an rt_mutex, this function is called to cleanup | |
2332 | * the pi_state owner as well as handle race conditions that may allow us to | |
2333 | * acquire the lock. Must be called with the hb lock held. | |
2334 | * | |
2335 | * Return: | |
2336 | * 1 - success, lock taken; | |
2337 | * 0 - success, lock not taken; | |
2338 | * <0 - on error (-EFAULT) | |
2339 | */ | |
2340 | static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked) | |
2341 | { | |
2342 | int ret = 0; | |
2343 | ||
2344 | if (locked) { | |
2345 | /* | |
2346 | * Got the lock. We might not be the anticipated owner if we | |
2347 | * did a lock-steal - fix up the PI-state in that case: | |
2348 | * | |
2349 | * We can safely read pi_state->owner without holding wait_lock | |
2350 | * because we now own the rt_mutex, only the owner will attempt | |
2351 | * to change it. | |
2352 | */ | |
2353 | if (q->pi_state->owner != current) | |
2354 | ret = fixup_pi_state_owner(uaddr, q, current); | |
2355 | goto out; | |
2356 | } | |
2357 | ||
2358 | /* | |
2359 | * Paranoia check. If we did not take the lock, then we should not be | |
2360 | * the owner of the rt_mutex. | |
2361 | */ | |
2362 | if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) { | |
2363 | printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p " | |
2364 | "pi-state %p\n", ret, | |
2365 | q->pi_state->pi_mutex.owner, | |
2366 | q->pi_state->owner); | |
2367 | } | |
2368 | ||
2369 | out: | |
2370 | return ret ? ret : locked; | |
2371 | } | |
2372 | ||
2373 | /** | |
2374 | * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal | |
2375 | * @hb: the futex hash bucket, must be locked by the caller | |
2376 | * @q: the futex_q to queue up on | |
2377 | * @timeout: the prepared hrtimer_sleeper, or null for no timeout | |
2378 | */ | |
2379 | static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q, | |
2380 | struct hrtimer_sleeper *timeout) | |
2381 | { | |
2382 | /* | |
2383 | * The task state is guaranteed to be set before another task can | |
2384 | * wake it. set_current_state() is implemented using smp_store_mb() and | |
2385 | * queue_me() calls spin_unlock() upon completion, both serializing | |
2386 | * access to the hash list and forcing another memory barrier. | |
2387 | */ | |
2388 | set_current_state(TASK_INTERRUPTIBLE); | |
2389 | queue_me(q, hb); | |
2390 | ||
2391 | /* Arm the timer */ | |
2392 | if (timeout) | |
2393 | hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS); | |
2394 | ||
2395 | /* | |
2396 | * If we have been removed from the hash list, then another task | |
2397 | * has tried to wake us, and we can skip the call to schedule(). | |
2398 | */ | |
2399 | if (likely(!plist_node_empty(&q->list))) { | |
2400 | /* | |
2401 | * If the timer has already expired, current will already be | |
2402 | * flagged for rescheduling. Only call schedule if there | |
2403 | * is no timeout, or if it has yet to expire. | |
2404 | */ | |
2405 | if (!timeout || timeout->task) | |
2406 | freezable_schedule(); | |
2407 | } | |
2408 | __set_current_state(TASK_RUNNING); | |
2409 | } | |
2410 | ||
2411 | /** | |
2412 | * futex_wait_setup() - Prepare to wait on a futex | |
2413 | * @uaddr: the futex userspace address | |
2414 | * @val: the expected value | |
2415 | * @flags: futex flags (FLAGS_SHARED, etc.) | |
2416 | * @q: the associated futex_q | |
2417 | * @hb: storage for hash_bucket pointer to be returned to caller | |
2418 | * | |
2419 | * Setup the futex_q and locate the hash_bucket. Get the futex value and | |
2420 | * compare it with the expected value. Handle atomic faults internally. | |
2421 | * Return with the hb lock held and a q.key reference on success, and unlocked | |
2422 | * with no q.key reference on failure. | |
2423 | * | |
2424 | * Return: | |
2425 | * 0 - uaddr contains val and hb has been locked; | |
2426 | * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked | |
2427 | */ | |
2428 | static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags, | |
2429 | struct futex_q *q, struct futex_hash_bucket **hb) | |
2430 | { | |
2431 | u32 uval; | |
2432 | int ret; | |
2433 | ||
2434 | /* | |
2435 | * Access the page AFTER the hash-bucket is locked. | |
2436 | * Order is important: | |
2437 | * | |
2438 | * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val); | |
2439 | * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); } | |
2440 | * | |
2441 | * The basic logical guarantee of a futex is that it blocks ONLY | |
2442 | * if cond(var) is known to be true at the time of blocking, for | |
2443 | * any cond. If we locked the hash-bucket after testing *uaddr, that | |
2444 | * would open a race condition where we could block indefinitely with | |
2445 | * cond(var) false, which would violate the guarantee. | |
2446 | * | |
2447 | * On the other hand, we insert q and release the hash-bucket only | |
2448 | * after testing *uaddr. This guarantees that futex_wait() will NOT | |
2449 | * absorb a wakeup if *uaddr does not match the desired values | |
2450 | * while the syscall executes. | |
2451 | */ | |
2452 | retry: | |
2453 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ); | |
2454 | if (unlikely(ret != 0)) | |
2455 | return ret; | |
2456 | ||
2457 | retry_private: | |
2458 | *hb = queue_lock(q); | |
2459 | ||
2460 | ret = get_futex_value_locked(&uval, uaddr); | |
2461 | ||
2462 | if (ret) { | |
2463 | queue_unlock(*hb); | |
2464 | ||
2465 | ret = get_user(uval, uaddr); | |
2466 | if (ret) | |
2467 | goto out; | |
2468 | ||
2469 | if (!(flags & FLAGS_SHARED)) | |
2470 | goto retry_private; | |
2471 | ||
2472 | put_futex_key(&q->key); | |
2473 | goto retry; | |
2474 | } | |
2475 | ||
2476 | if (uval != val) { | |
2477 | queue_unlock(*hb); | |
2478 | ret = -EWOULDBLOCK; | |
2479 | } | |
2480 | ||
2481 | out: | |
2482 | if (ret) | |
2483 | put_futex_key(&q->key); | |
2484 | return ret; | |
2485 | } | |
2486 | ||
2487 | static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val, | |
2488 | ktime_t *abs_time, u32 bitset) | |
2489 | { | |
2490 | struct hrtimer_sleeper timeout, *to = NULL; | |
2491 | struct restart_block *restart; | |
2492 | struct futex_hash_bucket *hb; | |
2493 | struct futex_q q = futex_q_init; | |
2494 | int ret; | |
2495 | ||
2496 | if (!bitset) | |
2497 | return -EINVAL; | |
2498 | q.bitset = bitset; | |
2499 | ||
2500 | if (abs_time) { | |
2501 | to = &timeout; | |
2502 | ||
2503 | hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ? | |
2504 | CLOCK_REALTIME : CLOCK_MONOTONIC, | |
2505 | HRTIMER_MODE_ABS); | |
2506 | hrtimer_init_sleeper(to, current); | |
2507 | hrtimer_set_expires_range_ns(&to->timer, *abs_time, | |
2508 | current->timer_slack_ns); | |
2509 | } | |
2510 | ||
2511 | retry: | |
2512 | /* | |
2513 | * Prepare to wait on uaddr. On success, holds hb lock and increments | |
2514 | * q.key refs. | |
2515 | */ | |
2516 | ret = futex_wait_setup(uaddr, val, flags, &q, &hb); | |
2517 | if (ret) | |
2518 | goto out; | |
2519 | ||
2520 | /* queue_me and wait for wakeup, timeout, or a signal. */ | |
2521 | futex_wait_queue_me(hb, &q, to); | |
2522 | ||
2523 | /* If we were woken (and unqueued), we succeeded, whatever. */ | |
2524 | ret = 0; | |
2525 | /* unqueue_me() drops q.key ref */ | |
2526 | if (!unqueue_me(&q)) | |
2527 | goto out; | |
2528 | ret = -ETIMEDOUT; | |
2529 | if (to && !to->task) | |
2530 | goto out; | |
2531 | ||
2532 | /* | |
2533 | * We expect signal_pending(current), but we might be the | |
2534 | * victim of a spurious wakeup as well. | |
2535 | */ | |
2536 | if (!signal_pending(current)) | |
2537 | goto retry; | |
2538 | ||
2539 | ret = -ERESTARTSYS; | |
2540 | if (!abs_time) | |
2541 | goto out; | |
2542 | ||
2543 | restart = ¤t->restart_block; | |
2544 | restart->fn = futex_wait_restart; | |
2545 | restart->futex.uaddr = uaddr; | |
2546 | restart->futex.val = val; | |
2547 | restart->futex.time = *abs_time; | |
2548 | restart->futex.bitset = bitset; | |
2549 | restart->futex.flags = flags | FLAGS_HAS_TIMEOUT; | |
2550 | ||
2551 | ret = -ERESTART_RESTARTBLOCK; | |
2552 | ||
2553 | out: | |
2554 | if (to) { | |
2555 | hrtimer_cancel(&to->timer); | |
2556 | destroy_hrtimer_on_stack(&to->timer); | |
2557 | } | |
2558 | return ret; | |
2559 | } | |
2560 | ||
2561 | ||
2562 | static long futex_wait_restart(struct restart_block *restart) | |
2563 | { | |
2564 | u32 __user *uaddr = restart->futex.uaddr; | |
2565 | ktime_t t, *tp = NULL; | |
2566 | ||
2567 | if (restart->futex.flags & FLAGS_HAS_TIMEOUT) { | |
2568 | t = restart->futex.time; | |
2569 | tp = &t; | |
2570 | } | |
2571 | restart->fn = do_no_restart_syscall; | |
2572 | ||
2573 | return (long)futex_wait(uaddr, restart->futex.flags, | |
2574 | restart->futex.val, tp, restart->futex.bitset); | |
2575 | } | |
2576 | ||
2577 | ||
2578 | /* | |
2579 | * Userspace tried a 0 -> TID atomic transition of the futex value | |
2580 | * and failed. The kernel side here does the whole locking operation: | |
2581 | * if there are waiters then it will block as a consequence of relying | |
2582 | * on rt-mutexes, it does PI, etc. (Due to races the kernel might see | |
2583 | * a 0 value of the futex too.). | |
2584 | * | |
2585 | * Also serves as futex trylock_pi()'ing, and due semantics. | |
2586 | */ | |
2587 | static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, | |
2588 | ktime_t *time, int trylock) | |
2589 | { | |
2590 | struct hrtimer_sleeper timeout, *to = NULL; | |
2591 | struct futex_pi_state *pi_state = NULL; | |
2592 | struct rt_mutex_waiter rt_waiter; | |
2593 | struct futex_hash_bucket *hb; | |
2594 | struct futex_q q = futex_q_init; | |
2595 | int res, ret; | |
2596 | ||
2597 | if (refill_pi_state_cache()) | |
2598 | return -ENOMEM; | |
2599 | ||
2600 | if (time) { | |
2601 | to = &timeout; | |
2602 | hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME, | |
2603 | HRTIMER_MODE_ABS); | |
2604 | hrtimer_init_sleeper(to, current); | |
2605 | hrtimer_set_expires(&to->timer, *time); | |
2606 | } | |
2607 | ||
2608 | retry: | |
2609 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE); | |
2610 | if (unlikely(ret != 0)) | |
2611 | goto out; | |
2612 | ||
2613 | retry_private: | |
2614 | hb = queue_lock(&q); | |
2615 | ||
2616 | ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0); | |
2617 | if (unlikely(ret)) { | |
2618 | /* | |
2619 | * Atomic work succeeded and we got the lock, | |
2620 | * or failed. Either way, we do _not_ block. | |
2621 | */ | |
2622 | switch (ret) { | |
2623 | case 1: | |
2624 | /* We got the lock. */ | |
2625 | ret = 0; | |
2626 | goto out_unlock_put_key; | |
2627 | case -EFAULT: | |
2628 | goto uaddr_faulted; | |
2629 | case -EAGAIN: | |
2630 | /* | |
2631 | * Two reasons for this: | |
2632 | * - Task is exiting and we just wait for the | |
2633 | * exit to complete. | |
2634 | * - The user space value changed. | |
2635 | */ | |
2636 | queue_unlock(hb); | |
2637 | put_futex_key(&q.key); | |
2638 | cond_resched(); | |
2639 | goto retry; | |
2640 | default: | |
2641 | goto out_unlock_put_key; | |
2642 | } | |
2643 | } | |
2644 | ||
2645 | WARN_ON(!q.pi_state); | |
2646 | ||
2647 | /* | |
2648 | * Only actually queue now that the atomic ops are done: | |
2649 | */ | |
2650 | __queue_me(&q, hb); | |
2651 | ||
2652 | if (trylock) { | |
2653 | ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex); | |
2654 | /* Fixup the trylock return value: */ | |
2655 | ret = ret ? 0 : -EWOULDBLOCK; | |
2656 | goto no_block; | |
2657 | } | |
2658 | ||
2659 | rt_mutex_init_waiter(&rt_waiter); | |
2660 | ||
2661 | /* | |
2662 | * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not | |
2663 | * hold it while doing rt_mutex_start_proxy(), because then it will | |
2664 | * include hb->lock in the blocking chain, even through we'll not in | |
2665 | * fact hold it while blocking. This will lead it to report -EDEADLK | |
2666 | * and BUG when futex_unlock_pi() interleaves with this. | |
2667 | * | |
2668 | * Therefore acquire wait_lock while holding hb->lock, but drop the | |
2669 | * latter before calling rt_mutex_start_proxy_lock(). This still fully | |
2670 | * serializes against futex_unlock_pi() as that does the exact same | |
2671 | * lock handoff sequence. | |
2672 | */ | |
2673 | raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock); | |
2674 | spin_unlock(q.lock_ptr); | |
2675 | ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current); | |
2676 | raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock); | |
2677 | ||
2678 | if (ret) { | |
2679 | if (ret == 1) | |
2680 | ret = 0; | |
2681 | ||
2682 | spin_lock(q.lock_ptr); | |
2683 | goto no_block; | |
2684 | } | |
2685 | ||
2686 | ||
2687 | if (unlikely(to)) | |
2688 | hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS); | |
2689 | ||
2690 | ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter); | |
2691 | ||
2692 | spin_lock(q.lock_ptr); | |
2693 | /* | |
2694 | * If we failed to acquire the lock (signal/timeout), we must | |
2695 | * first acquire the hb->lock before removing the lock from the | |
2696 | * rt_mutex waitqueue, such that we can keep the hb and rt_mutex | |
2697 | * wait lists consistent. | |
2698 | * | |
2699 | * In particular; it is important that futex_unlock_pi() can not | |
2700 | * observe this inconsistency. | |
2701 | */ | |
2702 | if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter)) | |
2703 | ret = 0; | |
2704 | ||
2705 | no_block: | |
2706 | /* | |
2707 | * Fixup the pi_state owner and possibly acquire the lock if we | |
2708 | * haven't already. | |
2709 | */ | |
2710 | res = fixup_owner(uaddr, &q, !ret); | |
2711 | /* | |
2712 | * If fixup_owner() returned an error, proprogate that. If it acquired | |
2713 | * the lock, clear our -ETIMEDOUT or -EINTR. | |
2714 | */ | |
2715 | if (res) | |
2716 | ret = (res < 0) ? res : 0; | |
2717 | ||
2718 | /* | |
2719 | * If fixup_owner() faulted and was unable to handle the fault, unlock | |
2720 | * it and return the fault to userspace. | |
2721 | */ | |
2722 | if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) { | |
2723 | pi_state = q.pi_state; | |
2724 | get_pi_state(pi_state); | |
2725 | } | |
2726 | ||
2727 | /* Unqueue and drop the lock */ | |
2728 | unqueue_me_pi(&q); | |
2729 | ||
2730 | if (pi_state) { | |
2731 | rt_mutex_futex_unlock(&pi_state->pi_mutex); | |
2732 | put_pi_state(pi_state); | |
2733 | } | |
2734 | ||
2735 | goto out_put_key; | |
2736 | ||
2737 | out_unlock_put_key: | |
2738 | queue_unlock(hb); | |
2739 | ||
2740 | out_put_key: | |
2741 | put_futex_key(&q.key); | |
2742 | out: | |
2743 | if (to) { | |
2744 | hrtimer_cancel(&to->timer); | |
2745 | destroy_hrtimer_on_stack(&to->timer); | |
2746 | } | |
2747 | return ret != -EINTR ? ret : -ERESTARTNOINTR; | |
2748 | ||
2749 | uaddr_faulted: | |
2750 | queue_unlock(hb); | |
2751 | ||
2752 | ret = fault_in_user_writeable(uaddr); | |
2753 | if (ret) | |
2754 | goto out_put_key; | |
2755 | ||
2756 | if (!(flags & FLAGS_SHARED)) | |
2757 | goto retry_private; | |
2758 | ||
2759 | put_futex_key(&q.key); | |
2760 | goto retry; | |
2761 | } | |
2762 | ||
2763 | /* | |
2764 | * Userspace attempted a TID -> 0 atomic transition, and failed. | |
2765 | * This is the in-kernel slowpath: we look up the PI state (if any), | |
2766 | * and do the rt-mutex unlock. | |
2767 | */ | |
2768 | static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags) | |
2769 | { | |
2770 | u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current); | |
2771 | union futex_key key = FUTEX_KEY_INIT; | |
2772 | struct futex_hash_bucket *hb; | |
2773 | struct futex_q *top_waiter; | |
2774 | int ret; | |
2775 | ||
2776 | retry: | |
2777 | if (get_user(uval, uaddr)) | |
2778 | return -EFAULT; | |
2779 | /* | |
2780 | * We release only a lock we actually own: | |
2781 | */ | |
2782 | if ((uval & FUTEX_TID_MASK) != vpid) | |
2783 | return -EPERM; | |
2784 | ||
2785 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE); | |
2786 | if (ret) | |
2787 | return ret; | |
2788 | ||
2789 | hb = hash_futex(&key); | |
2790 | spin_lock(&hb->lock); | |
2791 | ||
2792 | /* | |
2793 | * Check waiters first. We do not trust user space values at | |
2794 | * all and we at least want to know if user space fiddled | |
2795 | * with the futex value instead of blindly unlocking. | |
2796 | */ | |
2797 | top_waiter = futex_top_waiter(hb, &key); | |
2798 | if (top_waiter) { | |
2799 | struct futex_pi_state *pi_state = top_waiter->pi_state; | |
2800 | ||
2801 | ret = -EINVAL; | |
2802 | if (!pi_state) | |
2803 | goto out_unlock; | |
2804 | ||
2805 | /* | |
2806 | * If current does not own the pi_state then the futex is | |
2807 | * inconsistent and user space fiddled with the futex value. | |
2808 | */ | |
2809 | if (pi_state->owner != current) | |
2810 | goto out_unlock; | |
2811 | ||
2812 | get_pi_state(pi_state); | |
2813 | /* | |
2814 | * By taking wait_lock while still holding hb->lock, we ensure | |
2815 | * there is no point where we hold neither; and therefore | |
2816 | * wake_futex_pi() must observe a state consistent with what we | |
2817 | * observed. | |
2818 | */ | |
2819 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); | |
2820 | spin_unlock(&hb->lock); | |
2821 | ||
2822 | ret = wake_futex_pi(uaddr, uval, pi_state); | |
2823 | ||
2824 | put_pi_state(pi_state); | |
2825 | ||
2826 | /* | |
2827 | * Success, we're done! No tricky corner cases. | |
2828 | */ | |
2829 | if (!ret) | |
2830 | goto out_putkey; | |
2831 | /* | |
2832 | * The atomic access to the futex value generated a | |
2833 | * pagefault, so retry the user-access and the wakeup: | |
2834 | */ | |
2835 | if (ret == -EFAULT) | |
2836 | goto pi_faulted; | |
2837 | /* | |
2838 | * A unconditional UNLOCK_PI op raced against a waiter | |
2839 | * setting the FUTEX_WAITERS bit. Try again. | |
2840 | */ | |
2841 | if (ret == -EAGAIN) { | |
2842 | put_futex_key(&key); | |
2843 | goto retry; | |
2844 | } | |
2845 | /* | |
2846 | * wake_futex_pi has detected invalid state. Tell user | |
2847 | * space. | |
2848 | */ | |
2849 | goto out_putkey; | |
2850 | } | |
2851 | ||
2852 | /* | |
2853 | * We have no kernel internal state, i.e. no waiters in the | |
2854 | * kernel. Waiters which are about to queue themselves are stuck | |
2855 | * on hb->lock. So we can safely ignore them. We do neither | |
2856 | * preserve the WAITERS bit not the OWNER_DIED one. We are the | |
2857 | * owner. | |
2858 | */ | |
2859 | if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) { | |
2860 | spin_unlock(&hb->lock); | |
2861 | goto pi_faulted; | |
2862 | } | |
2863 | ||
2864 | /* | |
2865 | * If uval has changed, let user space handle it. | |
2866 | */ | |
2867 | ret = (curval == uval) ? 0 : -EAGAIN; | |
2868 | ||
2869 | out_unlock: | |
2870 | spin_unlock(&hb->lock); | |
2871 | out_putkey: | |
2872 | put_futex_key(&key); | |
2873 | return ret; | |
2874 | ||
2875 | pi_faulted: | |
2876 | put_futex_key(&key); | |
2877 | ||
2878 | ret = fault_in_user_writeable(uaddr); | |
2879 | if (!ret) | |
2880 | goto retry; | |
2881 | ||
2882 | return ret; | |
2883 | } | |
2884 | ||
2885 | /** | |
2886 | * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex | |
2887 | * @hb: the hash_bucket futex_q was original enqueued on | |
2888 | * @q: the futex_q woken while waiting to be requeued | |
2889 | * @key2: the futex_key of the requeue target futex | |
2890 | * @timeout: the timeout associated with the wait (NULL if none) | |
2891 | * | |
2892 | * Detect if the task was woken on the initial futex as opposed to the requeue | |
2893 | * target futex. If so, determine if it was a timeout or a signal that caused | |
2894 | * the wakeup and return the appropriate error code to the caller. Must be | |
2895 | * called with the hb lock held. | |
2896 | * | |
2897 | * Return: | |
2898 | * 0 = no early wakeup detected; | |
2899 | * <0 = -ETIMEDOUT or -ERESTARTNOINTR | |
2900 | */ | |
2901 | static inline | |
2902 | int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb, | |
2903 | struct futex_q *q, union futex_key *key2, | |
2904 | struct hrtimer_sleeper *timeout) | |
2905 | { | |
2906 | int ret = 0; | |
2907 | ||
2908 | /* | |
2909 | * With the hb lock held, we avoid races while we process the wakeup. | |
2910 | * We only need to hold hb (and not hb2) to ensure atomicity as the | |
2911 | * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb. | |
2912 | * It can't be requeued from uaddr2 to something else since we don't | |
2913 | * support a PI aware source futex for requeue. | |
2914 | */ | |
2915 | if (!match_futex(&q->key, key2)) { | |
2916 | WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr)); | |
2917 | /* | |
2918 | * We were woken prior to requeue by a timeout or a signal. | |
2919 | * Unqueue the futex_q and determine which it was. | |
2920 | */ | |
2921 | plist_del(&q->list, &hb->chain); | |
2922 | hb_waiters_dec(hb); | |
2923 | ||
2924 | /* Handle spurious wakeups gracefully */ | |
2925 | ret = -EWOULDBLOCK; | |
2926 | if (timeout && !timeout->task) | |
2927 | ret = -ETIMEDOUT; | |
2928 | else if (signal_pending(current)) | |
2929 | ret = -ERESTARTNOINTR; | |
2930 | } | |
2931 | return ret; | |
2932 | } | |
2933 | ||
2934 | /** | |
2935 | * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2 | |
2936 | * @uaddr: the futex we initially wait on (non-pi) | |
2937 | * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be | |
2938 | * the same type, no requeueing from private to shared, etc. | |
2939 | * @val: the expected value of uaddr | |
2940 | * @abs_time: absolute timeout | |
2941 | * @bitset: 32 bit wakeup bitset set by userspace, defaults to all | |
2942 | * @uaddr2: the pi futex we will take prior to returning to user-space | |
2943 | * | |
2944 | * The caller will wait on uaddr and will be requeued by futex_requeue() to | |
2945 | * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake | |
2946 | * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to | |
2947 | * userspace. This ensures the rt_mutex maintains an owner when it has waiters; | |
2948 | * without one, the pi logic would not know which task to boost/deboost, if | |
2949 | * there was a need to. | |
2950 | * | |
2951 | * We call schedule in futex_wait_queue_me() when we enqueue and return there | |
2952 | * via the following-- | |
2953 | * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue() | |
2954 | * 2) wakeup on uaddr2 after a requeue | |
2955 | * 3) signal | |
2956 | * 4) timeout | |
2957 | * | |
2958 | * If 3, cleanup and return -ERESTARTNOINTR. | |
2959 | * | |
2960 | * If 2, we may then block on trying to take the rt_mutex and return via: | |
2961 | * 5) successful lock | |
2962 | * 6) signal | |
2963 | * 7) timeout | |
2964 | * 8) other lock acquisition failure | |
2965 | * | |
2966 | * If 6, return -EWOULDBLOCK (restarting the syscall would do the same). | |
2967 | * | |
2968 | * If 4 or 7, we cleanup and return with -ETIMEDOUT. | |
2969 | * | |
2970 | * Return: | |
2971 | * 0 - On success; | |
2972 | * <0 - On error | |
2973 | */ | |
2974 | static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags, | |
2975 | u32 val, ktime_t *abs_time, u32 bitset, | |
2976 | u32 __user *uaddr2) | |
2977 | { | |
2978 | struct hrtimer_sleeper timeout, *to = NULL; | |
2979 | struct futex_pi_state *pi_state = NULL; | |
2980 | struct rt_mutex_waiter rt_waiter; | |
2981 | struct futex_hash_bucket *hb; | |
2982 | union futex_key key2 = FUTEX_KEY_INIT; | |
2983 | struct futex_q q = futex_q_init; | |
2984 | int res, ret; | |
2985 | ||
2986 | if (uaddr == uaddr2) | |
2987 | return -EINVAL; | |
2988 | ||
2989 | if (!bitset) | |
2990 | return -EINVAL; | |
2991 | ||
2992 | if (abs_time) { | |
2993 | to = &timeout; | |
2994 | hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ? | |
2995 | CLOCK_REALTIME : CLOCK_MONOTONIC, | |
2996 | HRTIMER_MODE_ABS); | |
2997 | hrtimer_init_sleeper(to, current); | |
2998 | hrtimer_set_expires_range_ns(&to->timer, *abs_time, | |
2999 | current->timer_slack_ns); | |
3000 | } | |
3001 | ||
3002 | /* | |
3003 | * The waiter is allocated on our stack, manipulated by the requeue | |
3004 | * code while we sleep on uaddr. | |
3005 | */ | |
3006 | rt_mutex_init_waiter(&rt_waiter); | |
3007 | ||
3008 | ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE); | |
3009 | if (unlikely(ret != 0)) | |
3010 | goto out; | |
3011 | ||
3012 | q.bitset = bitset; | |
3013 | q.rt_waiter = &rt_waiter; | |
3014 | q.requeue_pi_key = &key2; | |
3015 | ||
3016 | /* | |
3017 | * Prepare to wait on uaddr. On success, increments q.key (key1) ref | |
3018 | * count. | |
3019 | */ | |
3020 | ret = futex_wait_setup(uaddr, val, flags, &q, &hb); | |
3021 | if (ret) | |
3022 | goto out_key2; | |
3023 | ||
3024 | /* | |
3025 | * The check above which compares uaddrs is not sufficient for | |
3026 | * shared futexes. We need to compare the keys: | |
3027 | */ | |
3028 | if (match_futex(&q.key, &key2)) { | |
3029 | queue_unlock(hb); | |
3030 | ret = -EINVAL; | |
3031 | goto out_put_keys; | |
3032 | } | |
3033 | ||
3034 | /* Queue the futex_q, drop the hb lock, wait for wakeup. */ | |
3035 | futex_wait_queue_me(hb, &q, to); | |
3036 | ||
3037 | spin_lock(&hb->lock); | |
3038 | ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to); | |
3039 | spin_unlock(&hb->lock); | |
3040 | if (ret) | |
3041 | goto out_put_keys; | |
3042 | ||
3043 | /* | |
3044 | * In order for us to be here, we know our q.key == key2, and since | |
3045 | * we took the hb->lock above, we also know that futex_requeue() has | |
3046 | * completed and we no longer have to concern ourselves with a wakeup | |
3047 | * race with the atomic proxy lock acquisition by the requeue code. The | |
3048 | * futex_requeue dropped our key1 reference and incremented our key2 | |
3049 | * reference count. | |
3050 | */ | |
3051 | ||
3052 | /* Check if the requeue code acquired the second futex for us. */ | |
3053 | if (!q.rt_waiter) { | |
3054 | /* | |
3055 | * Got the lock. We might not be the anticipated owner if we | |
3056 | * did a lock-steal - fix up the PI-state in that case. | |
3057 | */ | |
3058 | if (q.pi_state && (q.pi_state->owner != current)) { | |
3059 | spin_lock(q.lock_ptr); | |
3060 | ret = fixup_pi_state_owner(uaddr2, &q, current); | |
3061 | if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) { | |
3062 | pi_state = q.pi_state; | |
3063 | get_pi_state(pi_state); | |
3064 | } | |
3065 | /* | |
3066 | * Drop the reference to the pi state which | |
3067 | * the requeue_pi() code acquired for us. | |
3068 | */ | |
3069 | put_pi_state(q.pi_state); | |
3070 | spin_unlock(q.lock_ptr); | |
3071 | } | |
3072 | } else { | |
3073 | struct rt_mutex *pi_mutex; | |
3074 | ||
3075 | /* | |
3076 | * We have been woken up by futex_unlock_pi(), a timeout, or a | |
3077 | * signal. futex_unlock_pi() will not destroy the lock_ptr nor | |
3078 | * the pi_state. | |
3079 | */ | |
3080 | WARN_ON(!q.pi_state); | |
3081 | pi_mutex = &q.pi_state->pi_mutex; | |
3082 | ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter); | |
3083 | ||
3084 | spin_lock(q.lock_ptr); | |
3085 | if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter)) | |
3086 | ret = 0; | |
3087 | ||
3088 | debug_rt_mutex_free_waiter(&rt_waiter); | |
3089 | /* | |
3090 | * Fixup the pi_state owner and possibly acquire the lock if we | |
3091 | * haven't already. | |
3092 | */ | |
3093 | res = fixup_owner(uaddr2, &q, !ret); | |
3094 | /* | |
3095 | * If fixup_owner() returned an error, proprogate that. If it | |
3096 | * acquired the lock, clear -ETIMEDOUT or -EINTR. | |
3097 | */ | |
3098 | if (res) | |
3099 | ret = (res < 0) ? res : 0; | |
3100 | ||
3101 | /* | |
3102 | * If fixup_pi_state_owner() faulted and was unable to handle | |
3103 | * the fault, unlock the rt_mutex and return the fault to | |
3104 | * userspace. | |
3105 | */ | |
3106 | if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) { | |
3107 | pi_state = q.pi_state; | |
3108 | get_pi_state(pi_state); | |
3109 | } | |
3110 | ||
3111 | /* Unqueue and drop the lock. */ | |
3112 | unqueue_me_pi(&q); | |
3113 | } | |
3114 | ||
3115 | if (pi_state) { | |
3116 | rt_mutex_futex_unlock(&pi_state->pi_mutex); | |
3117 | put_pi_state(pi_state); | |
3118 | } | |
3119 | ||
3120 | if (ret == -EINTR) { | |
3121 | /* | |
3122 | * We've already been requeued, but cannot restart by calling | |
3123 | * futex_lock_pi() directly. We could restart this syscall, but | |
3124 | * it would detect that the user space "val" changed and return | |
3125 | * -EWOULDBLOCK. Save the overhead of the restart and return | |
3126 | * -EWOULDBLOCK directly. | |
3127 | */ | |
3128 | ret = -EWOULDBLOCK; | |
3129 | } | |
3130 | ||
3131 | out_put_keys: | |
3132 | put_futex_key(&q.key); | |
3133 | out_key2: | |
3134 | put_futex_key(&key2); | |
3135 | ||
3136 | out: | |
3137 | if (to) { | |
3138 | hrtimer_cancel(&to->timer); | |
3139 | destroy_hrtimer_on_stack(&to->timer); | |
3140 | } | |
3141 | return ret; | |
3142 | } | |
3143 | ||
3144 | /* | |
3145 | * Support for robust futexes: the kernel cleans up held futexes at | |
3146 | * thread exit time. | |
3147 | * | |
3148 | * Implementation: user-space maintains a per-thread list of locks it | |
3149 | * is holding. Upon do_exit(), the kernel carefully walks this list, | |
3150 | * and marks all locks that are owned by this thread with the | |
3151 | * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is | |
3152 | * always manipulated with the lock held, so the list is private and | |
3153 | * per-thread. Userspace also maintains a per-thread 'list_op_pending' | |
3154 | * field, to allow the kernel to clean up if the thread dies after | |
3155 | * acquiring the lock, but just before it could have added itself to | |
3156 | * the list. There can only be one such pending lock. | |
3157 | */ | |
3158 | ||
3159 | /** | |
3160 | * sys_set_robust_list() - Set the robust-futex list head of a task | |
3161 | * @head: pointer to the list-head | |
3162 | * @len: length of the list-head, as userspace expects | |
3163 | */ | |
3164 | SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head, | |
3165 | size_t, len) | |
3166 | { | |
3167 | if (!futex_cmpxchg_enabled) | |
3168 | return -ENOSYS; | |
3169 | /* | |
3170 | * The kernel knows only one size for now: | |
3171 | */ | |
3172 | if (unlikely(len != sizeof(*head))) | |
3173 | return -EINVAL; | |
3174 | ||
3175 | current->robust_list = head; | |
3176 | ||
3177 | return 0; | |
3178 | } | |
3179 | ||
3180 | /** | |
3181 | * sys_get_robust_list() - Get the robust-futex list head of a task | |
3182 | * @pid: pid of the process [zero for current task] | |
3183 | * @head_ptr: pointer to a list-head pointer, the kernel fills it in | |
3184 | * @len_ptr: pointer to a length field, the kernel fills in the header size | |
3185 | */ | |
3186 | SYSCALL_DEFINE3(get_robust_list, int, pid, | |
3187 | struct robust_list_head __user * __user *, head_ptr, | |
3188 | size_t __user *, len_ptr) | |
3189 | { | |
3190 | struct robust_list_head __user *head; | |
3191 | unsigned long ret; | |
3192 | struct task_struct *p; | |
3193 | ||
3194 | if (!futex_cmpxchg_enabled) | |
3195 | return -ENOSYS; | |
3196 | ||
3197 | rcu_read_lock(); | |
3198 | ||
3199 | ret = -ESRCH; | |
3200 | if (!pid) | |
3201 | p = current; | |
3202 | else { | |
3203 | p = find_task_by_vpid(pid); | |
3204 | if (!p) | |
3205 | goto err_unlock; | |
3206 | } | |
3207 | ||
3208 | ret = -EPERM; | |
3209 | if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS)) | |
3210 | goto err_unlock; | |
3211 | ||
3212 | head = p->robust_list; | |
3213 | rcu_read_unlock(); | |
3214 | ||
3215 | if (put_user(sizeof(*head), len_ptr)) | |
3216 | return -EFAULT; | |
3217 | return put_user(head, head_ptr); | |
3218 | ||
3219 | err_unlock: | |
3220 | rcu_read_unlock(); | |
3221 | ||
3222 | return ret; | |
3223 | } | |
3224 | ||
3225 | /* | |
3226 | * Process a futex-list entry, check whether it's owned by the | |
3227 | * dying task, and do notification if so: | |
3228 | */ | |
3229 | int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi) | |
3230 | { | |
3231 | u32 uval, uninitialized_var(nval), mval; | |
3232 | ||
3233 | retry: | |
3234 | if (get_user(uval, uaddr)) | |
3235 | return -1; | |
3236 | ||
3237 | if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) { | |
3238 | /* | |
3239 | * Ok, this dying thread is truly holding a futex | |
3240 | * of interest. Set the OWNER_DIED bit atomically | |
3241 | * via cmpxchg, and if the value had FUTEX_WAITERS | |
3242 | * set, wake up a waiter (if any). (We have to do a | |
3243 | * futex_wake() even if OWNER_DIED is already set - | |
3244 | * to handle the rare but possible case of recursive | |
3245 | * thread-death.) The rest of the cleanup is done in | |
3246 | * userspace. | |
3247 | */ | |
3248 | mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED; | |
3249 | /* | |
3250 | * We are not holding a lock here, but we want to have | |
3251 | * the pagefault_disable/enable() protection because | |
3252 | * we want to handle the fault gracefully. If the | |
3253 | * access fails we try to fault in the futex with R/W | |
3254 | * verification via get_user_pages. get_user() above | |
3255 | * does not guarantee R/W access. If that fails we | |
3256 | * give up and leave the futex locked. | |
3257 | */ | |
3258 | if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) { | |
3259 | if (fault_in_user_writeable(uaddr)) | |
3260 | return -1; | |
3261 | goto retry; | |
3262 | } | |
3263 | if (nval != uval) | |
3264 | goto retry; | |
3265 | ||
3266 | /* | |
3267 | * Wake robust non-PI futexes here. The wakeup of | |
3268 | * PI futexes happens in exit_pi_state(): | |
3269 | */ | |
3270 | if (!pi && (uval & FUTEX_WAITERS)) | |
3271 | futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY); | |
3272 | } | |
3273 | return 0; | |
3274 | } | |
3275 | ||
3276 | /* | |
3277 | * Fetch a robust-list pointer. Bit 0 signals PI futexes: | |
3278 | */ | |
3279 | static inline int fetch_robust_entry(struct robust_list __user **entry, | |
3280 | struct robust_list __user * __user *head, | |
3281 | unsigned int *pi) | |
3282 | { | |
3283 | unsigned long uentry; | |
3284 | ||
3285 | if (get_user(uentry, (unsigned long __user *)head)) | |
3286 | return -EFAULT; | |
3287 | ||
3288 | *entry = (void __user *)(uentry & ~1UL); | |
3289 | *pi = uentry & 1; | |
3290 | ||
3291 | return 0; | |
3292 | } | |
3293 | ||
3294 | /* | |
3295 | * Walk curr->robust_list (very carefully, it's a userspace list!) | |
3296 | * and mark any locks found there dead, and notify any waiters. | |
3297 | * | |
3298 | * We silently return on any sign of list-walking problem. | |
3299 | */ | |
3300 | void exit_robust_list(struct task_struct *curr) | |
3301 | { | |
3302 | struct robust_list_head __user *head = curr->robust_list; | |
3303 | struct robust_list __user *entry, *next_entry, *pending; | |
3304 | unsigned int limit = ROBUST_LIST_LIMIT, pi, pip; | |
3305 | unsigned int uninitialized_var(next_pi); | |
3306 | unsigned long futex_offset; | |
3307 | int rc; | |
3308 | ||
3309 | if (!futex_cmpxchg_enabled) | |
3310 | return; | |
3311 | ||
3312 | /* | |
3313 | * Fetch the list head (which was registered earlier, via | |
3314 | * sys_set_robust_list()): | |
3315 | */ | |
3316 | if (fetch_robust_entry(&entry, &head->list.next, &pi)) | |
3317 | return; | |
3318 | /* | |
3319 | * Fetch the relative futex offset: | |
3320 | */ | |
3321 | if (get_user(futex_offset, &head->futex_offset)) | |
3322 | return; | |
3323 | /* | |
3324 | * Fetch any possibly pending lock-add first, and handle it | |
3325 | * if it exists: | |
3326 | */ | |
3327 | if (fetch_robust_entry(&pending, &head->list_op_pending, &pip)) | |
3328 | return; | |
3329 | ||
3330 | next_entry = NULL; /* avoid warning with gcc */ | |
3331 | while (entry != &head->list) { | |
3332 | /* | |
3333 | * Fetch the next entry in the list before calling | |
3334 | * handle_futex_death: | |
3335 | */ | |
3336 | rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi); | |
3337 | /* | |
3338 | * A pending lock might already be on the list, so | |
3339 | * don't process it twice: | |
3340 | */ | |
3341 | if (entry != pending) | |
3342 | if (handle_futex_death((void __user *)entry + futex_offset, | |
3343 | curr, pi)) | |
3344 | return; | |
3345 | if (rc) | |
3346 | return; | |
3347 | entry = next_entry; | |
3348 | pi = next_pi; | |
3349 | /* | |
3350 | * Avoid excessively long or circular lists: | |
3351 | */ | |
3352 | if (!--limit) | |
3353 | break; | |
3354 | ||
3355 | cond_resched(); | |
3356 | } | |
3357 | ||
3358 | if (pending) | |
3359 | handle_futex_death((void __user *)pending + futex_offset, | |
3360 | curr, pip); | |
3361 | } | |
3362 | ||
3363 | long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout, | |
3364 | u32 __user *uaddr2, u32 val2, u32 val3) | |
3365 | { | |
3366 | int cmd = op & FUTEX_CMD_MASK; | |
3367 | unsigned int flags = 0; | |
3368 | ||
3369 | if (!(op & FUTEX_PRIVATE_FLAG)) | |
3370 | flags |= FLAGS_SHARED; | |
3371 | ||
3372 | if (op & FUTEX_CLOCK_REALTIME) { | |
3373 | flags |= FLAGS_CLOCKRT; | |
3374 | if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \ | |
3375 | cmd != FUTEX_WAIT_REQUEUE_PI) | |
3376 | return -ENOSYS; | |
3377 | } | |
3378 | ||
3379 | switch (cmd) { | |
3380 | case FUTEX_LOCK_PI: | |
3381 | case FUTEX_UNLOCK_PI: | |
3382 | case FUTEX_TRYLOCK_PI: | |
3383 | case FUTEX_WAIT_REQUEUE_PI: | |
3384 | case FUTEX_CMP_REQUEUE_PI: | |
3385 | if (!futex_cmpxchg_enabled) | |
3386 | return -ENOSYS; | |
3387 | } | |
3388 | ||
3389 | switch (cmd) { | |
3390 | case FUTEX_WAIT: | |
3391 | val3 = FUTEX_BITSET_MATCH_ANY; | |
3392 | case FUTEX_WAIT_BITSET: | |
3393 | return futex_wait(uaddr, flags, val, timeout, val3); | |
3394 | case FUTEX_WAKE: | |
3395 | val3 = FUTEX_BITSET_MATCH_ANY; | |
3396 | case FUTEX_WAKE_BITSET: | |
3397 | return futex_wake(uaddr, flags, val, val3); | |
3398 | case FUTEX_REQUEUE: | |
3399 | return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0); | |
3400 | case FUTEX_CMP_REQUEUE: | |
3401 | return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0); | |
3402 | case FUTEX_WAKE_OP: | |
3403 | return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3); | |
3404 | case FUTEX_LOCK_PI: | |
3405 | return futex_lock_pi(uaddr, flags, timeout, 0); | |
3406 | case FUTEX_UNLOCK_PI: | |
3407 | return futex_unlock_pi(uaddr, flags); | |
3408 | case FUTEX_TRYLOCK_PI: | |
3409 | return futex_lock_pi(uaddr, flags, NULL, 1); | |
3410 | case FUTEX_WAIT_REQUEUE_PI: | |
3411 | val3 = FUTEX_BITSET_MATCH_ANY; | |
3412 | return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3, | |
3413 | uaddr2); | |
3414 | case FUTEX_CMP_REQUEUE_PI: | |
3415 | return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1); | |
3416 | } | |
3417 | return -ENOSYS; | |
3418 | } | |
3419 | ||
3420 | ||
3421 | SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val, | |
3422 | struct timespec __user *, utime, u32 __user *, uaddr2, | |
3423 | u32, val3) | |
3424 | { | |
3425 | struct timespec ts; | |
3426 | ktime_t t, *tp = NULL; | |
3427 | u32 val2 = 0; | |
3428 | int cmd = op & FUTEX_CMD_MASK; | |
3429 | ||
3430 | if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI || | |
3431 | cmd == FUTEX_WAIT_BITSET || | |
3432 | cmd == FUTEX_WAIT_REQUEUE_PI)) { | |
3433 | if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG)))) | |
3434 | return -EFAULT; | |
3435 | if (copy_from_user(&ts, utime, sizeof(ts)) != 0) | |
3436 | return -EFAULT; | |
3437 | if (!timespec_valid(&ts)) | |
3438 | return -EINVAL; | |
3439 | ||
3440 | t = timespec_to_ktime(ts); | |
3441 | if (cmd == FUTEX_WAIT) | |
3442 | t = ktime_add_safe(ktime_get(), t); | |
3443 | tp = &t; | |
3444 | } | |
3445 | /* | |
3446 | * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*. | |
3447 | * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP. | |
3448 | */ | |
3449 | if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE || | |
3450 | cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP) | |
3451 | val2 = (u32) (unsigned long) utime; | |
3452 | ||
3453 | return do_futex(uaddr, op, val, tp, uaddr2, val2, val3); | |
3454 | } | |
3455 | ||
3456 | static void __init futex_detect_cmpxchg(void) | |
3457 | { | |
3458 | #ifndef CONFIG_HAVE_FUTEX_CMPXCHG | |
3459 | u32 curval; | |
3460 | ||
3461 | /* | |
3462 | * This will fail and we want it. Some arch implementations do | |
3463 | * runtime detection of the futex_atomic_cmpxchg_inatomic() | |
3464 | * functionality. We want to know that before we call in any | |
3465 | * of the complex code paths. Also we want to prevent | |
3466 | * registration of robust lists in that case. NULL is | |
3467 | * guaranteed to fault and we get -EFAULT on functional | |
3468 | * implementation, the non-functional ones will return | |
3469 | * -ENOSYS. | |
3470 | */ | |
3471 | if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT) | |
3472 | futex_cmpxchg_enabled = 1; | |
3473 | #endif | |
3474 | } | |
3475 | ||
3476 | static int __init futex_init(void) | |
3477 | { | |
3478 | unsigned int futex_shift; | |
3479 | unsigned long i; | |
3480 | ||
3481 | #if CONFIG_BASE_SMALL | |
3482 | futex_hashsize = 16; | |
3483 | #else | |
3484 | futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus()); | |
3485 | #endif | |
3486 | ||
3487 | futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues), | |
3488 | futex_hashsize, 0, | |
3489 | futex_hashsize < 256 ? HASH_SMALL : 0, | |
3490 | &futex_shift, NULL, | |
3491 | futex_hashsize, futex_hashsize); | |
3492 | futex_hashsize = 1UL << futex_shift; | |
3493 | ||
3494 | futex_detect_cmpxchg(); | |
3495 | ||
3496 | for (i = 0; i < futex_hashsize; i++) { | |
3497 | atomic_set(&futex_queues[i].waiters, 0); | |
3498 | plist_head_init(&futex_queues[i].chain); | |
3499 | spin_lock_init(&futex_queues[i].lock); | |
3500 | } | |
3501 | ||
3502 | return 0; | |
3503 | } | |
3504 | core_initcall(futex_init); |