Commit | Line | Data |
---|---|---|
1da177e4 LT |
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 | * | |
0771dfef IM |
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 | * | |
1da177e4 LT |
15 | * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly |
16 | * enough at me, Linus for the original (flawed) idea, Matthew | |
17 | * Kirkwood for proof-of-concept implementation. | |
18 | * | |
19 | * "The futexes are also cursed." | |
20 | * "But they come in a choice of three flavours!" | |
21 | * | |
22 | * This program is free software; you can redistribute it and/or modify | |
23 | * it under the terms of the GNU General Public License as published by | |
24 | * the Free Software Foundation; either version 2 of the License, or | |
25 | * (at your option) any later version. | |
26 | * | |
27 | * This program is distributed in the hope that it will be useful, | |
28 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
29 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
30 | * GNU General Public License for more details. | |
31 | * | |
32 | * You should have received a copy of the GNU General Public License | |
33 | * along with this program; if not, write to the Free Software | |
34 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
35 | */ | |
36 | #include <linux/slab.h> | |
37 | #include <linux/poll.h> | |
38 | #include <linux/fs.h> | |
39 | #include <linux/file.h> | |
40 | #include <linux/jhash.h> | |
41 | #include <linux/init.h> | |
42 | #include <linux/futex.h> | |
43 | #include <linux/mount.h> | |
44 | #include <linux/pagemap.h> | |
45 | #include <linux/syscalls.h> | |
7ed20e1a | 46 | #include <linux/signal.h> |
4732efbe | 47 | #include <asm/futex.h> |
1da177e4 LT |
48 | |
49 | #define FUTEX_HASHBITS (CONFIG_BASE_SMALL ? 4 : 8) | |
50 | ||
51 | /* | |
52 | * Futexes are matched on equal values of this key. | |
53 | * The key type depends on whether it's a shared or private mapping. | |
54 | * Don't rearrange members without looking at hash_futex(). | |
55 | * | |
56 | * offset is aligned to a multiple of sizeof(u32) (== 4) by definition. | |
57 | * We set bit 0 to indicate if it's an inode-based key. | |
58 | */ | |
59 | union futex_key { | |
60 | struct { | |
61 | unsigned long pgoff; | |
62 | struct inode *inode; | |
63 | int offset; | |
64 | } shared; | |
65 | struct { | |
e2970f2f | 66 | unsigned long address; |
1da177e4 LT |
67 | struct mm_struct *mm; |
68 | int offset; | |
69 | } private; | |
70 | struct { | |
71 | unsigned long word; | |
72 | void *ptr; | |
73 | int offset; | |
74 | } both; | |
75 | }; | |
76 | ||
77 | /* | |
78 | * We use this hashed waitqueue instead of a normal wait_queue_t, so | |
79 | * we can wake only the relevant ones (hashed queues may be shared). | |
80 | * | |
81 | * A futex_q has a woken state, just like tasks have TASK_RUNNING. | |
82 | * It is considered woken when list_empty(&q->list) || q->lock_ptr == 0. | |
83 | * The order of wakup is always to make the first condition true, then | |
84 | * wake up q->waiters, then make the second condition true. | |
85 | */ | |
86 | struct futex_q { | |
87 | struct list_head list; | |
88 | wait_queue_head_t waiters; | |
89 | ||
e2970f2f | 90 | /* Which hash list lock to use: */ |
1da177e4 LT |
91 | spinlock_t *lock_ptr; |
92 | ||
e2970f2f | 93 | /* Key which the futex is hashed on: */ |
1da177e4 LT |
94 | union futex_key key; |
95 | ||
e2970f2f | 96 | /* For fd, sigio sent using these: */ |
1da177e4 LT |
97 | int fd; |
98 | struct file *filp; | |
99 | }; | |
100 | ||
101 | /* | |
102 | * Split the global futex_lock into every hash list lock. | |
103 | */ | |
104 | struct futex_hash_bucket { | |
105 | spinlock_t lock; | |
106 | struct list_head chain; | |
107 | }; | |
108 | ||
109 | static struct futex_hash_bucket futex_queues[1<<FUTEX_HASHBITS]; | |
110 | ||
111 | /* Futex-fs vfsmount entry: */ | |
112 | static struct vfsmount *futex_mnt; | |
113 | ||
114 | /* | |
115 | * We hash on the keys returned from get_futex_key (see below). | |
116 | */ | |
117 | static struct futex_hash_bucket *hash_futex(union futex_key *key) | |
118 | { | |
119 | u32 hash = jhash2((u32*)&key->both.word, | |
120 | (sizeof(key->both.word)+sizeof(key->both.ptr))/4, | |
121 | key->both.offset); | |
122 | return &futex_queues[hash & ((1 << FUTEX_HASHBITS)-1)]; | |
123 | } | |
124 | ||
125 | /* | |
126 | * Return 1 if two futex_keys are equal, 0 otherwise. | |
127 | */ | |
128 | static inline int match_futex(union futex_key *key1, union futex_key *key2) | |
129 | { | |
130 | return (key1->both.word == key2->both.word | |
131 | && key1->both.ptr == key2->both.ptr | |
132 | && key1->both.offset == key2->both.offset); | |
133 | } | |
134 | ||
135 | /* | |
136 | * Get parameters which are the keys for a futex. | |
137 | * | |
138 | * For shared mappings, it's (page->index, vma->vm_file->f_dentry->d_inode, | |
139 | * offset_within_page). For private mappings, it's (uaddr, current->mm). | |
140 | * We can usually work out the index without swapping in the page. | |
141 | * | |
142 | * Returns: 0, or negative error code. | |
143 | * The key words are stored in *key on success. | |
144 | * | |
145 | * Should be called with ¤t->mm->mmap_sem but NOT any spinlocks. | |
146 | */ | |
e2970f2f | 147 | static int get_futex_key(u32 __user *uaddr, union futex_key *key) |
1da177e4 | 148 | { |
e2970f2f | 149 | unsigned long address = (unsigned long)uaddr; |
1da177e4 LT |
150 | struct mm_struct *mm = current->mm; |
151 | struct vm_area_struct *vma; | |
152 | struct page *page; | |
153 | int err; | |
154 | ||
155 | /* | |
156 | * The futex address must be "naturally" aligned. | |
157 | */ | |
e2970f2f | 158 | key->both.offset = address % PAGE_SIZE; |
1da177e4 LT |
159 | if (unlikely((key->both.offset % sizeof(u32)) != 0)) |
160 | return -EINVAL; | |
e2970f2f | 161 | address -= key->both.offset; |
1da177e4 LT |
162 | |
163 | /* | |
164 | * The futex is hashed differently depending on whether | |
165 | * it's in a shared or private mapping. So check vma first. | |
166 | */ | |
e2970f2f | 167 | vma = find_extend_vma(mm, address); |
1da177e4 LT |
168 | if (unlikely(!vma)) |
169 | return -EFAULT; | |
170 | ||
171 | /* | |
172 | * Permissions. | |
173 | */ | |
174 | if (unlikely((vma->vm_flags & (VM_IO|VM_READ)) != VM_READ)) | |
175 | return (vma->vm_flags & VM_IO) ? -EPERM : -EACCES; | |
176 | ||
177 | /* | |
178 | * Private mappings are handled in a simple way. | |
179 | * | |
180 | * NOTE: When userspace waits on a MAP_SHARED mapping, even if | |
181 | * it's a read-only handle, it's expected that futexes attach to | |
182 | * the object not the particular process. Therefore we use | |
183 | * VM_MAYSHARE here, not VM_SHARED which is restricted to shared | |
184 | * mappings of _writable_ handles. | |
185 | */ | |
186 | if (likely(!(vma->vm_flags & VM_MAYSHARE))) { | |
187 | key->private.mm = mm; | |
e2970f2f | 188 | key->private.address = address; |
1da177e4 LT |
189 | return 0; |
190 | } | |
191 | ||
192 | /* | |
193 | * Linear file mappings are also simple. | |
194 | */ | |
195 | key->shared.inode = vma->vm_file->f_dentry->d_inode; | |
196 | key->both.offset++; /* Bit 0 of offset indicates inode-based key. */ | |
197 | if (likely(!(vma->vm_flags & VM_NONLINEAR))) { | |
e2970f2f | 198 | key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT) |
1da177e4 LT |
199 | + vma->vm_pgoff); |
200 | return 0; | |
201 | } | |
202 | ||
203 | /* | |
204 | * We could walk the page table to read the non-linear | |
205 | * pte, and get the page index without fetching the page | |
206 | * from swap. But that's a lot of code to duplicate here | |
207 | * for a rare case, so we simply fetch the page. | |
208 | */ | |
e2970f2f | 209 | err = get_user_pages(current, mm, address, 1, 0, 0, &page, NULL); |
1da177e4 LT |
210 | if (err >= 0) { |
211 | key->shared.pgoff = | |
212 | page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
213 | put_page(page); | |
214 | return 0; | |
215 | } | |
216 | return err; | |
217 | } | |
218 | ||
219 | /* | |
220 | * Take a reference to the resource addressed by a key. | |
221 | * Can be called while holding spinlocks. | |
222 | * | |
223 | * NOTE: mmap_sem MUST be held between get_futex_key() and calling this | |
224 | * function, if it is called at all. mmap_sem keeps key->shared.inode valid. | |
225 | */ | |
226 | static inline void get_key_refs(union futex_key *key) | |
227 | { | |
228 | if (key->both.ptr != 0) { | |
229 | if (key->both.offset & 1) | |
230 | atomic_inc(&key->shared.inode->i_count); | |
231 | else | |
232 | atomic_inc(&key->private.mm->mm_count); | |
233 | } | |
234 | } | |
235 | ||
236 | /* | |
237 | * Drop a reference to the resource addressed by a key. | |
238 | * The hash bucket spinlock must not be held. | |
239 | */ | |
240 | static void drop_key_refs(union futex_key *key) | |
241 | { | |
242 | if (key->both.ptr != 0) { | |
243 | if (key->both.offset & 1) | |
244 | iput(key->shared.inode); | |
245 | else | |
246 | mmdrop(key->private.mm); | |
247 | } | |
248 | } | |
249 | ||
e2970f2f | 250 | static inline int get_futex_value_locked(u32 *dest, u32 __user *from) |
1da177e4 LT |
251 | { |
252 | int ret; | |
253 | ||
254 | inc_preempt_count(); | |
e2970f2f | 255 | ret = __copy_from_user_inatomic(dest, from, sizeof(u32)); |
1da177e4 LT |
256 | dec_preempt_count(); |
257 | ||
258 | return ret ? -EFAULT : 0; | |
259 | } | |
260 | ||
261 | /* | |
262 | * The hash bucket lock must be held when this is called. | |
263 | * Afterwards, the futex_q must not be accessed. | |
264 | */ | |
265 | static void wake_futex(struct futex_q *q) | |
266 | { | |
267 | list_del_init(&q->list); | |
268 | if (q->filp) | |
269 | send_sigio(&q->filp->f_owner, q->fd, POLL_IN); | |
270 | /* | |
271 | * The lock in wake_up_all() is a crucial memory barrier after the | |
272 | * list_del_init() and also before assigning to q->lock_ptr. | |
273 | */ | |
274 | wake_up_all(&q->waiters); | |
275 | /* | |
276 | * The waiting task can free the futex_q as soon as this is written, | |
277 | * without taking any locks. This must come last. | |
8e31108b AM |
278 | * |
279 | * A memory barrier is required here to prevent the following store | |
280 | * to lock_ptr from getting ahead of the wakeup. Clearing the lock | |
281 | * at the end of wake_up_all() does not prevent this store from | |
282 | * moving. | |
1da177e4 | 283 | */ |
8e31108b | 284 | wmb(); |
1da177e4 LT |
285 | q->lock_ptr = NULL; |
286 | } | |
287 | ||
288 | /* | |
289 | * Wake up all waiters hashed on the physical page that is mapped | |
290 | * to this virtual address: | |
291 | */ | |
e2970f2f | 292 | static int futex_wake(u32 __user *uaddr, int nr_wake) |
1da177e4 | 293 | { |
e2970f2f | 294 | struct futex_hash_bucket *hb; |
1da177e4 | 295 | struct futex_q *this, *next; |
e2970f2f IM |
296 | struct list_head *head; |
297 | union futex_key key; | |
1da177e4 LT |
298 | int ret; |
299 | ||
300 | down_read(¤t->mm->mmap_sem); | |
301 | ||
302 | ret = get_futex_key(uaddr, &key); | |
303 | if (unlikely(ret != 0)) | |
304 | goto out; | |
305 | ||
e2970f2f IM |
306 | hb = hash_futex(&key); |
307 | spin_lock(&hb->lock); | |
308 | head = &hb->chain; | |
1da177e4 LT |
309 | |
310 | list_for_each_entry_safe(this, next, head, list) { | |
311 | if (match_futex (&this->key, &key)) { | |
312 | wake_futex(this); | |
313 | if (++ret >= nr_wake) | |
314 | break; | |
315 | } | |
316 | } | |
317 | ||
e2970f2f | 318 | spin_unlock(&hb->lock); |
1da177e4 LT |
319 | out: |
320 | up_read(¤t->mm->mmap_sem); | |
321 | return ret; | |
322 | } | |
323 | ||
4732efbe JJ |
324 | /* |
325 | * Wake up all waiters hashed on the physical page that is mapped | |
326 | * to this virtual address: | |
327 | */ | |
e2970f2f IM |
328 | static int |
329 | futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2, | |
330 | int nr_wake, int nr_wake2, int op) | |
4732efbe JJ |
331 | { |
332 | union futex_key key1, key2; | |
e2970f2f | 333 | struct futex_hash_bucket *hb1, *hb2; |
4732efbe JJ |
334 | struct list_head *head; |
335 | struct futex_q *this, *next; | |
336 | int ret, op_ret, attempt = 0; | |
337 | ||
338 | retryfull: | |
339 | down_read(¤t->mm->mmap_sem); | |
340 | ||
341 | ret = get_futex_key(uaddr1, &key1); | |
342 | if (unlikely(ret != 0)) | |
343 | goto out; | |
344 | ret = get_futex_key(uaddr2, &key2); | |
345 | if (unlikely(ret != 0)) | |
346 | goto out; | |
347 | ||
e2970f2f IM |
348 | hb1 = hash_futex(&key1); |
349 | hb2 = hash_futex(&key2); | |
4732efbe JJ |
350 | |
351 | retry: | |
e2970f2f IM |
352 | if (hb1 < hb2) |
353 | spin_lock(&hb1->lock); | |
354 | spin_lock(&hb2->lock); | |
355 | if (hb1 > hb2) | |
356 | spin_lock(&hb1->lock); | |
4732efbe | 357 | |
e2970f2f | 358 | op_ret = futex_atomic_op_inuser(op, uaddr2); |
4732efbe | 359 | if (unlikely(op_ret < 0)) { |
e2970f2f | 360 | u32 dummy; |
4732efbe | 361 | |
e2970f2f IM |
362 | spin_unlock(&hb1->lock); |
363 | if (hb1 != hb2) | |
364 | spin_unlock(&hb2->lock); | |
4732efbe | 365 | |
7ee1dd3f | 366 | #ifndef CONFIG_MMU |
e2970f2f IM |
367 | /* |
368 | * we don't get EFAULT from MMU faults if we don't have an MMU, | |
369 | * but we might get them from range checking | |
370 | */ | |
7ee1dd3f DH |
371 | ret = op_ret; |
372 | goto out; | |
373 | #endif | |
374 | ||
796f8d9b DG |
375 | if (unlikely(op_ret != -EFAULT)) { |
376 | ret = op_ret; | |
377 | goto out; | |
378 | } | |
379 | ||
e2970f2f IM |
380 | /* |
381 | * futex_atomic_op_inuser needs to both read and write | |
4732efbe JJ |
382 | * *(int __user *)uaddr2, but we can't modify it |
383 | * non-atomically. Therefore, if get_user below is not | |
384 | * enough, we need to handle the fault ourselves, while | |
e2970f2f IM |
385 | * still holding the mmap_sem. |
386 | */ | |
4732efbe JJ |
387 | if (attempt++) { |
388 | struct vm_area_struct * vma; | |
389 | struct mm_struct *mm = current->mm; | |
e2970f2f | 390 | unsigned long address = (unsigned long)uaddr2; |
4732efbe JJ |
391 | |
392 | ret = -EFAULT; | |
393 | if (attempt >= 2 || | |
e2970f2f IM |
394 | !(vma = find_vma(mm, address)) || |
395 | vma->vm_start > address || | |
4732efbe JJ |
396 | !(vma->vm_flags & VM_WRITE)) |
397 | goto out; | |
398 | ||
e2970f2f | 399 | switch (handle_mm_fault(mm, vma, address, 1)) { |
4732efbe JJ |
400 | case VM_FAULT_MINOR: |
401 | current->min_flt++; | |
402 | break; | |
403 | case VM_FAULT_MAJOR: | |
404 | current->maj_flt++; | |
405 | break; | |
406 | default: | |
407 | goto out; | |
408 | } | |
409 | goto retry; | |
410 | } | |
411 | ||
e2970f2f IM |
412 | /* |
413 | * If we would have faulted, release mmap_sem, | |
414 | * fault it in and start all over again. | |
415 | */ | |
4732efbe JJ |
416 | up_read(¤t->mm->mmap_sem); |
417 | ||
e2970f2f | 418 | ret = get_user(dummy, uaddr2); |
4732efbe JJ |
419 | if (ret) |
420 | return ret; | |
421 | ||
422 | goto retryfull; | |
423 | } | |
424 | ||
e2970f2f | 425 | head = &hb1->chain; |
4732efbe JJ |
426 | |
427 | list_for_each_entry_safe(this, next, head, list) { | |
428 | if (match_futex (&this->key, &key1)) { | |
429 | wake_futex(this); | |
430 | if (++ret >= nr_wake) | |
431 | break; | |
432 | } | |
433 | } | |
434 | ||
435 | if (op_ret > 0) { | |
e2970f2f | 436 | head = &hb2->chain; |
4732efbe JJ |
437 | |
438 | op_ret = 0; | |
439 | list_for_each_entry_safe(this, next, head, list) { | |
440 | if (match_futex (&this->key, &key2)) { | |
441 | wake_futex(this); | |
442 | if (++op_ret >= nr_wake2) | |
443 | break; | |
444 | } | |
445 | } | |
446 | ret += op_ret; | |
447 | } | |
448 | ||
e2970f2f IM |
449 | spin_unlock(&hb1->lock); |
450 | if (hb1 != hb2) | |
451 | spin_unlock(&hb2->lock); | |
4732efbe JJ |
452 | out: |
453 | up_read(¤t->mm->mmap_sem); | |
454 | return ret; | |
455 | } | |
456 | ||
1da177e4 LT |
457 | /* |
458 | * Requeue all waiters hashed on one physical page to another | |
459 | * physical page. | |
460 | */ | |
e2970f2f IM |
461 | static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2, |
462 | int nr_wake, int nr_requeue, u32 *cmpval) | |
1da177e4 LT |
463 | { |
464 | union futex_key key1, key2; | |
e2970f2f | 465 | struct futex_hash_bucket *hb1, *hb2; |
1da177e4 LT |
466 | struct list_head *head1; |
467 | struct futex_q *this, *next; | |
468 | int ret, drop_count = 0; | |
469 | ||
470 | retry: | |
471 | down_read(¤t->mm->mmap_sem); | |
472 | ||
473 | ret = get_futex_key(uaddr1, &key1); | |
474 | if (unlikely(ret != 0)) | |
475 | goto out; | |
476 | ret = get_futex_key(uaddr2, &key2); | |
477 | if (unlikely(ret != 0)) | |
478 | goto out; | |
479 | ||
e2970f2f IM |
480 | hb1 = hash_futex(&key1); |
481 | hb2 = hash_futex(&key2); | |
1da177e4 | 482 | |
e2970f2f IM |
483 | if (hb1 < hb2) |
484 | spin_lock(&hb1->lock); | |
485 | spin_lock(&hb2->lock); | |
486 | if (hb1 > hb2) | |
487 | spin_lock(&hb1->lock); | |
1da177e4 | 488 | |
e2970f2f IM |
489 | if (likely(cmpval != NULL)) { |
490 | u32 curval; | |
1da177e4 | 491 | |
e2970f2f | 492 | ret = get_futex_value_locked(&curval, uaddr1); |
1da177e4 LT |
493 | |
494 | if (unlikely(ret)) { | |
e2970f2f IM |
495 | spin_unlock(&hb1->lock); |
496 | if (hb1 != hb2) | |
497 | spin_unlock(&hb2->lock); | |
1da177e4 | 498 | |
e2970f2f IM |
499 | /* |
500 | * If we would have faulted, release mmap_sem, fault | |
1da177e4 LT |
501 | * it in and start all over again. |
502 | */ | |
503 | up_read(¤t->mm->mmap_sem); | |
504 | ||
e2970f2f | 505 | ret = get_user(curval, uaddr1); |
1da177e4 LT |
506 | |
507 | if (!ret) | |
508 | goto retry; | |
509 | ||
510 | return ret; | |
511 | } | |
e2970f2f | 512 | if (curval != *cmpval) { |
1da177e4 LT |
513 | ret = -EAGAIN; |
514 | goto out_unlock; | |
515 | } | |
516 | } | |
517 | ||
e2970f2f | 518 | head1 = &hb1->chain; |
1da177e4 LT |
519 | list_for_each_entry_safe(this, next, head1, list) { |
520 | if (!match_futex (&this->key, &key1)) | |
521 | continue; | |
522 | if (++ret <= nr_wake) { | |
523 | wake_futex(this); | |
524 | } else { | |
e2970f2f IM |
525 | list_move_tail(&this->list, &hb2->chain); |
526 | this->lock_ptr = &hb2->lock; | |
1da177e4 LT |
527 | this->key = key2; |
528 | get_key_refs(&key2); | |
529 | drop_count++; | |
530 | ||
531 | if (ret - nr_wake >= nr_requeue) | |
532 | break; | |
e2970f2f IM |
533 | /* Make sure to stop if key1 == key2: */ |
534 | if (head1 == &hb2->chain && head1 != &next->list) | |
1da177e4 LT |
535 | head1 = &this->list; |
536 | } | |
537 | } | |
538 | ||
539 | out_unlock: | |
e2970f2f IM |
540 | spin_unlock(&hb1->lock); |
541 | if (hb1 != hb2) | |
542 | spin_unlock(&hb2->lock); | |
1da177e4 LT |
543 | |
544 | /* drop_key_refs() must be called outside the spinlocks. */ | |
545 | while (--drop_count >= 0) | |
546 | drop_key_refs(&key1); | |
547 | ||
548 | out: | |
549 | up_read(¤t->mm->mmap_sem); | |
550 | return ret; | |
551 | } | |
552 | ||
553 | /* The key must be already stored in q->key. */ | |
554 | static inline struct futex_hash_bucket * | |
555 | queue_lock(struct futex_q *q, int fd, struct file *filp) | |
556 | { | |
e2970f2f | 557 | struct futex_hash_bucket *hb; |
1da177e4 LT |
558 | |
559 | q->fd = fd; | |
560 | q->filp = filp; | |
561 | ||
562 | init_waitqueue_head(&q->waiters); | |
563 | ||
564 | get_key_refs(&q->key); | |
e2970f2f IM |
565 | hb = hash_futex(&q->key); |
566 | q->lock_ptr = &hb->lock; | |
1da177e4 | 567 | |
e2970f2f IM |
568 | spin_lock(&hb->lock); |
569 | return hb; | |
1da177e4 LT |
570 | } |
571 | ||
e2970f2f | 572 | static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb) |
1da177e4 | 573 | { |
e2970f2f IM |
574 | list_add_tail(&q->list, &hb->chain); |
575 | spin_unlock(&hb->lock); | |
1da177e4 LT |
576 | } |
577 | ||
578 | static inline void | |
e2970f2f | 579 | queue_unlock(struct futex_q *q, struct futex_hash_bucket *hb) |
1da177e4 | 580 | { |
e2970f2f | 581 | spin_unlock(&hb->lock); |
1da177e4 LT |
582 | drop_key_refs(&q->key); |
583 | } | |
584 | ||
585 | /* | |
586 | * queue_me and unqueue_me must be called as a pair, each | |
587 | * exactly once. They are called with the hashed spinlock held. | |
588 | */ | |
589 | ||
590 | /* The key must be already stored in q->key. */ | |
591 | static void queue_me(struct futex_q *q, int fd, struct file *filp) | |
592 | { | |
e2970f2f IM |
593 | struct futex_hash_bucket *hb; |
594 | ||
595 | hb = queue_lock(q, fd, filp); | |
596 | __queue_me(q, hb); | |
1da177e4 LT |
597 | } |
598 | ||
599 | /* Return 1 if we were still queued (ie. 0 means we were woken) */ | |
600 | static int unqueue_me(struct futex_q *q) | |
601 | { | |
1da177e4 | 602 | spinlock_t *lock_ptr; |
e2970f2f | 603 | int ret = 0; |
1da177e4 LT |
604 | |
605 | /* In the common case we don't take the spinlock, which is nice. */ | |
606 | retry: | |
607 | lock_ptr = q->lock_ptr; | |
608 | if (lock_ptr != 0) { | |
609 | spin_lock(lock_ptr); | |
610 | /* | |
611 | * q->lock_ptr can change between reading it and | |
612 | * spin_lock(), causing us to take the wrong lock. This | |
613 | * corrects the race condition. | |
614 | * | |
615 | * Reasoning goes like this: if we have the wrong lock, | |
616 | * q->lock_ptr must have changed (maybe several times) | |
617 | * between reading it and the spin_lock(). It can | |
618 | * change again after the spin_lock() but only if it was | |
619 | * already changed before the spin_lock(). It cannot, | |
620 | * however, change back to the original value. Therefore | |
621 | * we can detect whether we acquired the correct lock. | |
622 | */ | |
623 | if (unlikely(lock_ptr != q->lock_ptr)) { | |
624 | spin_unlock(lock_ptr); | |
625 | goto retry; | |
626 | } | |
627 | WARN_ON(list_empty(&q->list)); | |
628 | list_del(&q->list); | |
629 | spin_unlock(lock_ptr); | |
630 | ret = 1; | |
631 | } | |
632 | ||
633 | drop_key_refs(&q->key); | |
634 | return ret; | |
635 | } | |
636 | ||
e2970f2f | 637 | static int futex_wait(u32 __user *uaddr, u32 val, unsigned long time) |
1da177e4 LT |
638 | { |
639 | DECLARE_WAITQUEUE(wait, current); | |
e2970f2f | 640 | struct futex_hash_bucket *hb; |
1da177e4 | 641 | struct futex_q q; |
e2970f2f IM |
642 | u32 uval; |
643 | int ret; | |
1da177e4 LT |
644 | |
645 | retry: | |
646 | down_read(¤t->mm->mmap_sem); | |
647 | ||
648 | ret = get_futex_key(uaddr, &q.key); | |
649 | if (unlikely(ret != 0)) | |
650 | goto out_release_sem; | |
651 | ||
e2970f2f | 652 | hb = queue_lock(&q, -1, NULL); |
1da177e4 LT |
653 | |
654 | /* | |
655 | * Access the page AFTER the futex is queued. | |
656 | * Order is important: | |
657 | * | |
658 | * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val); | |
659 | * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); } | |
660 | * | |
661 | * The basic logical guarantee of a futex is that it blocks ONLY | |
662 | * if cond(var) is known to be true at the time of blocking, for | |
663 | * any cond. If we queued after testing *uaddr, that would open | |
664 | * a race condition where we could block indefinitely with | |
665 | * cond(var) false, which would violate the guarantee. | |
666 | * | |
667 | * A consequence is that futex_wait() can return zero and absorb | |
668 | * a wakeup when *uaddr != val on entry to the syscall. This is | |
669 | * rare, but normal. | |
670 | * | |
671 | * We hold the mmap semaphore, so the mapping cannot have changed | |
672 | * since we looked it up in get_futex_key. | |
673 | */ | |
e2970f2f | 674 | ret = get_futex_value_locked(&uval, uaddr); |
1da177e4 LT |
675 | |
676 | if (unlikely(ret)) { | |
e2970f2f | 677 | queue_unlock(&q, hb); |
1da177e4 | 678 | |
e2970f2f IM |
679 | /* |
680 | * If we would have faulted, release mmap_sem, fault it in and | |
1da177e4 LT |
681 | * start all over again. |
682 | */ | |
683 | up_read(¤t->mm->mmap_sem); | |
684 | ||
e2970f2f | 685 | ret = get_user(uval, uaddr); |
1da177e4 LT |
686 | |
687 | if (!ret) | |
688 | goto retry; | |
689 | return ret; | |
690 | } | |
e2970f2f | 691 | if (uval != val) { |
1da177e4 | 692 | ret = -EWOULDBLOCK; |
e2970f2f | 693 | queue_unlock(&q, hb); |
1da177e4 LT |
694 | goto out_release_sem; |
695 | } | |
696 | ||
697 | /* Only actually queue if *uaddr contained val. */ | |
e2970f2f | 698 | __queue_me(&q, hb); |
1da177e4 LT |
699 | |
700 | /* | |
701 | * Now the futex is queued and we have checked the data, we | |
702 | * don't want to hold mmap_sem while we sleep. | |
703 | */ | |
704 | up_read(¤t->mm->mmap_sem); | |
705 | ||
706 | /* | |
707 | * There might have been scheduling since the queue_me(), as we | |
708 | * cannot hold a spinlock across the get_user() in case it | |
709 | * faults, and we cannot just set TASK_INTERRUPTIBLE state when | |
710 | * queueing ourselves into the futex hash. This code thus has to | |
711 | * rely on the futex_wake() code removing us from hash when it | |
712 | * wakes us up. | |
713 | */ | |
714 | ||
715 | /* add_wait_queue is the barrier after __set_current_state. */ | |
716 | __set_current_state(TASK_INTERRUPTIBLE); | |
717 | add_wait_queue(&q.waiters, &wait); | |
718 | /* | |
719 | * !list_empty() is safe here without any lock. | |
720 | * q.lock_ptr != 0 is not safe, because of ordering against wakeup. | |
721 | */ | |
722 | if (likely(!list_empty(&q.list))) | |
723 | time = schedule_timeout(time); | |
724 | __set_current_state(TASK_RUNNING); | |
725 | ||
726 | /* | |
727 | * NOTE: we don't remove ourselves from the waitqueue because | |
728 | * we are the only user of it. | |
729 | */ | |
730 | ||
731 | /* If we were woken (and unqueued), we succeeded, whatever. */ | |
732 | if (!unqueue_me(&q)) | |
733 | return 0; | |
734 | if (time == 0) | |
735 | return -ETIMEDOUT; | |
e2970f2f IM |
736 | /* |
737 | * We expect signal_pending(current), but another thread may | |
738 | * have handled it for us already. | |
739 | */ | |
1da177e4 LT |
740 | return -EINTR; |
741 | ||
742 | out_release_sem: | |
743 | up_read(¤t->mm->mmap_sem); | |
744 | return ret; | |
745 | } | |
746 | ||
747 | static int futex_close(struct inode *inode, struct file *filp) | |
748 | { | |
749 | struct futex_q *q = filp->private_data; | |
750 | ||
751 | unqueue_me(q); | |
752 | kfree(q); | |
e2970f2f | 753 | |
1da177e4 LT |
754 | return 0; |
755 | } | |
756 | ||
757 | /* This is one-shot: once it's gone off you need a new fd */ | |
758 | static unsigned int futex_poll(struct file *filp, | |
759 | struct poll_table_struct *wait) | |
760 | { | |
761 | struct futex_q *q = filp->private_data; | |
762 | int ret = 0; | |
763 | ||
764 | poll_wait(filp, &q->waiters, wait); | |
765 | ||
766 | /* | |
767 | * list_empty() is safe here without any lock. | |
768 | * q->lock_ptr != 0 is not safe, because of ordering against wakeup. | |
769 | */ | |
770 | if (list_empty(&q->list)) | |
771 | ret = POLLIN | POLLRDNORM; | |
772 | ||
773 | return ret; | |
774 | } | |
775 | ||
776 | static struct file_operations futex_fops = { | |
777 | .release = futex_close, | |
778 | .poll = futex_poll, | |
779 | }; | |
780 | ||
781 | /* | |
782 | * Signal allows caller to avoid the race which would occur if they | |
783 | * set the sigio stuff up afterwards. | |
784 | */ | |
e2970f2f | 785 | static int futex_fd(u32 __user *uaddr, int signal) |
1da177e4 LT |
786 | { |
787 | struct futex_q *q; | |
788 | struct file *filp; | |
789 | int ret, err; | |
790 | ||
791 | ret = -EINVAL; | |
7ed20e1a | 792 | if (!valid_signal(signal)) |
1da177e4 LT |
793 | goto out; |
794 | ||
795 | ret = get_unused_fd(); | |
796 | if (ret < 0) | |
797 | goto out; | |
798 | filp = get_empty_filp(); | |
799 | if (!filp) { | |
800 | put_unused_fd(ret); | |
801 | ret = -ENFILE; | |
802 | goto out; | |
803 | } | |
804 | filp->f_op = &futex_fops; | |
805 | filp->f_vfsmnt = mntget(futex_mnt); | |
806 | filp->f_dentry = dget(futex_mnt->mnt_root); | |
807 | filp->f_mapping = filp->f_dentry->d_inode->i_mapping; | |
808 | ||
809 | if (signal) { | |
1da177e4 LT |
810 | err = f_setown(filp, current->pid, 1); |
811 | if (err < 0) { | |
39ed3fde | 812 | goto error; |
1da177e4 LT |
813 | } |
814 | filp->f_owner.signum = signal; | |
815 | } | |
816 | ||
817 | q = kmalloc(sizeof(*q), GFP_KERNEL); | |
818 | if (!q) { | |
39ed3fde PE |
819 | err = -ENOMEM; |
820 | goto error; | |
1da177e4 LT |
821 | } |
822 | ||
823 | down_read(¤t->mm->mmap_sem); | |
824 | err = get_futex_key(uaddr, &q->key); | |
825 | ||
826 | if (unlikely(err != 0)) { | |
827 | up_read(¤t->mm->mmap_sem); | |
1da177e4 | 828 | kfree(q); |
39ed3fde | 829 | goto error; |
1da177e4 LT |
830 | } |
831 | ||
832 | /* | |
833 | * queue_me() must be called before releasing mmap_sem, because | |
834 | * key->shared.inode needs to be referenced while holding it. | |
835 | */ | |
836 | filp->private_data = q; | |
837 | ||
838 | queue_me(q, ret, filp); | |
839 | up_read(¤t->mm->mmap_sem); | |
840 | ||
841 | /* Now we map fd to filp, so userspace can access it */ | |
842 | fd_install(ret, filp); | |
843 | out: | |
844 | return ret; | |
39ed3fde PE |
845 | error: |
846 | put_unused_fd(ret); | |
847 | put_filp(filp); | |
848 | ret = err; | |
849 | goto out; | |
1da177e4 LT |
850 | } |
851 | ||
0771dfef IM |
852 | /* |
853 | * Support for robust futexes: the kernel cleans up held futexes at | |
854 | * thread exit time. | |
855 | * | |
856 | * Implementation: user-space maintains a per-thread list of locks it | |
857 | * is holding. Upon do_exit(), the kernel carefully walks this list, | |
858 | * and marks all locks that are owned by this thread with the | |
859 | * FUTEX_OWNER_DEAD bit, and wakes up a waiter (if any). The list is | |
860 | * always manipulated with the lock held, so the list is private and | |
861 | * per-thread. Userspace also maintains a per-thread 'list_op_pending' | |
862 | * field, to allow the kernel to clean up if the thread dies after | |
863 | * acquiring the lock, but just before it could have added itself to | |
864 | * the list. There can only be one such pending lock. | |
865 | */ | |
866 | ||
867 | /** | |
868 | * sys_set_robust_list - set the robust-futex list head of a task | |
869 | * @head: pointer to the list-head | |
870 | * @len: length of the list-head, as userspace expects | |
871 | */ | |
872 | asmlinkage long | |
873 | sys_set_robust_list(struct robust_list_head __user *head, | |
874 | size_t len) | |
875 | { | |
876 | /* | |
877 | * The kernel knows only one size for now: | |
878 | */ | |
879 | if (unlikely(len != sizeof(*head))) | |
880 | return -EINVAL; | |
881 | ||
882 | current->robust_list = head; | |
883 | ||
884 | return 0; | |
885 | } | |
886 | ||
887 | /** | |
888 | * sys_get_robust_list - get the robust-futex list head of a task | |
889 | * @pid: pid of the process [zero for current task] | |
890 | * @head_ptr: pointer to a list-head pointer, the kernel fills it in | |
891 | * @len_ptr: pointer to a length field, the kernel fills in the header size | |
892 | */ | |
893 | asmlinkage long | |
894 | sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr, | |
895 | size_t __user *len_ptr) | |
896 | { | |
897 | struct robust_list_head *head; | |
898 | unsigned long ret; | |
899 | ||
900 | if (!pid) | |
901 | head = current->robust_list; | |
902 | else { | |
903 | struct task_struct *p; | |
904 | ||
905 | ret = -ESRCH; | |
906 | read_lock(&tasklist_lock); | |
907 | p = find_task_by_pid(pid); | |
908 | if (!p) | |
909 | goto err_unlock; | |
910 | ret = -EPERM; | |
911 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
912 | !capable(CAP_SYS_PTRACE)) | |
913 | goto err_unlock; | |
914 | head = p->robust_list; | |
915 | read_unlock(&tasklist_lock); | |
916 | } | |
917 | ||
918 | if (put_user(sizeof(*head), len_ptr)) | |
919 | return -EFAULT; | |
920 | return put_user(head, head_ptr); | |
921 | ||
922 | err_unlock: | |
923 | read_unlock(&tasklist_lock); | |
924 | ||
925 | return ret; | |
926 | } | |
927 | ||
928 | /* | |
929 | * Process a futex-list entry, check whether it's owned by the | |
930 | * dying task, and do notification if so: | |
931 | */ | |
8f17d3a5 | 932 | int handle_futex_death(u32 __user *uaddr, struct task_struct *curr) |
0771dfef | 933 | { |
8f17d3a5 | 934 | u32 uval; |
0771dfef | 935 | |
8f17d3a5 IM |
936 | retry: |
937 | if (get_user(uval, uaddr)) | |
0771dfef IM |
938 | return -1; |
939 | ||
8f17d3a5 | 940 | if ((uval & FUTEX_TID_MASK) == curr->pid) { |
0771dfef IM |
941 | /* |
942 | * Ok, this dying thread is truly holding a futex | |
943 | * of interest. Set the OWNER_DIED bit atomically | |
944 | * via cmpxchg, and if the value had FUTEX_WAITERS | |
945 | * set, wake up a waiter (if any). (We have to do a | |
946 | * futex_wake() even if OWNER_DIED is already set - | |
947 | * to handle the rare but possible case of recursive | |
948 | * thread-death.) The rest of the cleanup is done in | |
949 | * userspace. | |
950 | */ | |
8f17d3a5 IM |
951 | if (futex_atomic_cmpxchg_inatomic(uaddr, uval, |
952 | uval | FUTEX_OWNER_DIED) != uval) | |
953 | goto retry; | |
0771dfef | 954 | |
8f17d3a5 | 955 | if (uval & FUTEX_WAITERS) |
e2970f2f | 956 | futex_wake(uaddr, 1); |
0771dfef IM |
957 | } |
958 | return 0; | |
959 | } | |
960 | ||
961 | /* | |
962 | * Walk curr->robust_list (very carefully, it's a userspace list!) | |
963 | * and mark any locks found there dead, and notify any waiters. | |
964 | * | |
965 | * We silently return on any sign of list-walking problem. | |
966 | */ | |
967 | void exit_robust_list(struct task_struct *curr) | |
968 | { | |
969 | struct robust_list_head __user *head = curr->robust_list; | |
970 | struct robust_list __user *entry, *pending; | |
971 | unsigned int limit = ROBUST_LIST_LIMIT; | |
972 | unsigned long futex_offset; | |
973 | ||
974 | /* | |
975 | * Fetch the list head (which was registered earlier, via | |
976 | * sys_set_robust_list()): | |
977 | */ | |
978 | if (get_user(entry, &head->list.next)) | |
979 | return; | |
980 | /* | |
981 | * Fetch the relative futex offset: | |
982 | */ | |
983 | if (get_user(futex_offset, &head->futex_offset)) | |
984 | return; | |
985 | /* | |
986 | * Fetch any possibly pending lock-add first, and handle it | |
987 | * if it exists: | |
988 | */ | |
989 | if (get_user(pending, &head->list_op_pending)) | |
990 | return; | |
991 | if (pending) | |
992 | handle_futex_death((void *)pending + futex_offset, curr); | |
993 | ||
994 | while (entry != &head->list) { | |
995 | /* | |
996 | * A pending lock might already be on the list, so | |
997 | * dont process it twice: | |
998 | */ | |
999 | if (entry != pending) | |
1000 | if (handle_futex_death((void *)entry + futex_offset, | |
1001 | curr)) | |
1002 | return; | |
0771dfef IM |
1003 | /* |
1004 | * Fetch the next entry in the list: | |
1005 | */ | |
1006 | if (get_user(entry, &entry->next)) | |
1007 | return; | |
1008 | /* | |
1009 | * Avoid excessively long or circular lists: | |
1010 | */ | |
1011 | if (!--limit) | |
1012 | break; | |
1013 | ||
1014 | cond_resched(); | |
1015 | } | |
1016 | } | |
1017 | ||
e2970f2f IM |
1018 | long do_futex(u32 __user *uaddr, int op, u32 val, unsigned long timeout, |
1019 | u32 __user *uaddr2, u32 val2, u32 val3) | |
1da177e4 LT |
1020 | { |
1021 | int ret; | |
1022 | ||
1023 | switch (op) { | |
1024 | case FUTEX_WAIT: | |
1025 | ret = futex_wait(uaddr, val, timeout); | |
1026 | break; | |
1027 | case FUTEX_WAKE: | |
1028 | ret = futex_wake(uaddr, val); | |
1029 | break; | |
1030 | case FUTEX_FD: | |
1031 | /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */ | |
1032 | ret = futex_fd(uaddr, val); | |
1033 | break; | |
1034 | case FUTEX_REQUEUE: | |
1035 | ret = futex_requeue(uaddr, uaddr2, val, val2, NULL); | |
1036 | break; | |
1037 | case FUTEX_CMP_REQUEUE: | |
1038 | ret = futex_requeue(uaddr, uaddr2, val, val2, &val3); | |
1039 | break; | |
4732efbe JJ |
1040 | case FUTEX_WAKE_OP: |
1041 | ret = futex_wake_op(uaddr, uaddr2, val, val2, val3); | |
1042 | break; | |
1da177e4 LT |
1043 | default: |
1044 | ret = -ENOSYS; | |
1045 | } | |
1046 | return ret; | |
1047 | } | |
1048 | ||
1049 | ||
e2970f2f | 1050 | asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val, |
1da177e4 | 1051 | struct timespec __user *utime, u32 __user *uaddr2, |
e2970f2f | 1052 | u32 val3) |
1da177e4 LT |
1053 | { |
1054 | struct timespec t; | |
1055 | unsigned long timeout = MAX_SCHEDULE_TIMEOUT; | |
e2970f2f | 1056 | u32 val2 = 0; |
1da177e4 | 1057 | |
9741ef96 | 1058 | if (utime && (op == FUTEX_WAIT)) { |
1da177e4 LT |
1059 | if (copy_from_user(&t, utime, sizeof(t)) != 0) |
1060 | return -EFAULT; | |
9741ef96 TG |
1061 | if (!timespec_valid(&t)) |
1062 | return -EINVAL; | |
1da177e4 LT |
1063 | timeout = timespec_to_jiffies(&t) + 1; |
1064 | } | |
1065 | /* | |
1066 | * requeue parameter in 'utime' if op == FUTEX_REQUEUE. | |
1067 | */ | |
1068 | if (op >= FUTEX_REQUEUE) | |
e2970f2f | 1069 | val2 = (u32) (unsigned long) utime; |
1da177e4 | 1070 | |
e2970f2f | 1071 | return do_futex(uaddr, op, val, timeout, uaddr2, val2, val3); |
1da177e4 LT |
1072 | } |
1073 | ||
454e2398 DH |
1074 | static int futexfs_get_sb(struct file_system_type *fs_type, |
1075 | int flags, const char *dev_name, void *data, | |
1076 | struct vfsmount *mnt) | |
1da177e4 | 1077 | { |
454e2398 | 1078 | return get_sb_pseudo(fs_type, "futex", NULL, 0xBAD1DEA, mnt); |
1da177e4 LT |
1079 | } |
1080 | ||
1081 | static struct file_system_type futex_fs_type = { | |
1082 | .name = "futexfs", | |
1083 | .get_sb = futexfs_get_sb, | |
1084 | .kill_sb = kill_anon_super, | |
1085 | }; | |
1086 | ||
1087 | static int __init init(void) | |
1088 | { | |
1089 | unsigned int i; | |
1090 | ||
1091 | register_filesystem(&futex_fs_type); | |
1092 | futex_mnt = kern_mount(&futex_fs_type); | |
1093 | ||
1094 | for (i = 0; i < ARRAY_SIZE(futex_queues); i++) { | |
1095 | INIT_LIST_HEAD(&futex_queues[i].chain); | |
1096 | spin_lock_init(&futex_queues[i].lock); | |
1097 | } | |
1098 | return 0; | |
1099 | } | |
1100 | __initcall(init); |