]>
Commit | Line | Data |
---|---|---|
1a59d1b8 | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
1da177e4 LT |
2 | /* |
3 | * Fast Userspace Mutexes (which I call "Futexes!"). | |
4 | * (C) Rusty Russell, IBM 2002 | |
5 | * | |
6 | * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar | |
7 | * (C) Copyright 2003 Red Hat Inc, All Rights Reserved | |
8 | * | |
9 | * Removed page pinning, fix privately mapped COW pages and other cleanups | |
10 | * (C) Copyright 2003, 2004 Jamie Lokier | |
11 | * | |
0771dfef IM |
12 | * Robust futex support started by Ingo Molnar |
13 | * (C) Copyright 2006 Red Hat Inc, All Rights Reserved | |
14 | * Thanks to Thomas Gleixner for suggestions, analysis and fixes. | |
15 | * | |
c87e2837 IM |
16 | * PI-futex support started by Ingo Molnar and Thomas Gleixner |
17 | * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
18 | * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com> | |
19 | * | |
34f01cc1 ED |
20 | * PRIVATE futexes by Eric Dumazet |
21 | * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com> | |
22 | * | |
52400ba9 DH |
23 | * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com> |
24 | * Copyright (C) IBM Corporation, 2009 | |
25 | * Thanks to Thomas Gleixner for conceptual design and careful reviews. | |
26 | * | |
1da177e4 LT |
27 | * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly |
28 | * enough at me, Linus for the original (flawed) idea, Matthew | |
29 | * Kirkwood for proof-of-concept implementation. | |
30 | * | |
31 | * "The futexes are also cursed." | |
32 | * "But they come in a choice of three flavours!" | |
1da177e4 | 33 | */ |
04e7712f | 34 | #include <linux/compat.h> |
1da177e4 LT |
35 | #include <linux/slab.h> |
36 | #include <linux/poll.h> | |
37 | #include <linux/fs.h> | |
38 | #include <linux/file.h> | |
39 | #include <linux/jhash.h> | |
40 | #include <linux/init.h> | |
41 | #include <linux/futex.h> | |
42 | #include <linux/mount.h> | |
43 | #include <linux/pagemap.h> | |
44 | #include <linux/syscalls.h> | |
7ed20e1a | 45 | #include <linux/signal.h> |
9984de1a | 46 | #include <linux/export.h> |
fd5eea42 | 47 | #include <linux/magic.h> |
b488893a PE |
48 | #include <linux/pid.h> |
49 | #include <linux/nsproxy.h> | |
bdbb776f | 50 | #include <linux/ptrace.h> |
8bd75c77 | 51 | #include <linux/sched/rt.h> |
84f001e1 | 52 | #include <linux/sched/wake_q.h> |
6e84f315 | 53 | #include <linux/sched/mm.h> |
13d60f4b | 54 | #include <linux/hugetlb.h> |
88c8004f | 55 | #include <linux/freezer.h> |
57c8a661 | 56 | #include <linux/memblock.h> |
ab51fbab | 57 | #include <linux/fault-inject.h> |
49262de2 | 58 | #include <linux/refcount.h> |
b488893a | 59 | |
4732efbe | 60 | #include <asm/futex.h> |
1da177e4 | 61 | |
1696a8be | 62 | #include "locking/rtmutex_common.h" |
c87e2837 | 63 | |
99b60ce6 | 64 | /* |
d7e8af1a DB |
65 | * READ this before attempting to hack on futexes! |
66 | * | |
67 | * Basic futex operation and ordering guarantees | |
68 | * ============================================= | |
99b60ce6 TG |
69 | * |
70 | * The waiter reads the futex value in user space and calls | |
71 | * futex_wait(). This function computes the hash bucket and acquires | |
72 | * the hash bucket lock. After that it reads the futex user space value | |
b0c29f79 DB |
73 | * again and verifies that the data has not changed. If it has not changed |
74 | * it enqueues itself into the hash bucket, releases the hash bucket lock | |
75 | * and schedules. | |
99b60ce6 TG |
76 | * |
77 | * The waker side modifies the user space value of the futex and calls | |
b0c29f79 DB |
78 | * futex_wake(). This function computes the hash bucket and acquires the |
79 | * hash bucket lock. Then it looks for waiters on that futex in the hash | |
80 | * bucket and wakes them. | |
99b60ce6 | 81 | * |
b0c29f79 DB |
82 | * In futex wake up scenarios where no tasks are blocked on a futex, taking |
83 | * the hb spinlock can be avoided and simply return. In order for this | |
84 | * optimization to work, ordering guarantees must exist so that the waiter | |
85 | * being added to the list is acknowledged when the list is concurrently being | |
86 | * checked by the waker, avoiding scenarios like the following: | |
99b60ce6 TG |
87 | * |
88 | * CPU 0 CPU 1 | |
89 | * val = *futex; | |
90 | * sys_futex(WAIT, futex, val); | |
91 | * futex_wait(futex, val); | |
92 | * uval = *futex; | |
93 | * *futex = newval; | |
94 | * sys_futex(WAKE, futex); | |
95 | * futex_wake(futex); | |
96 | * if (queue_empty()) | |
97 | * return; | |
98 | * if (uval == val) | |
99 | * lock(hash_bucket(futex)); | |
100 | * queue(); | |
101 | * unlock(hash_bucket(futex)); | |
102 | * schedule(); | |
103 | * | |
104 | * This would cause the waiter on CPU 0 to wait forever because it | |
105 | * missed the transition of the user space value from val to newval | |
106 | * and the waker did not find the waiter in the hash bucket queue. | |
99b60ce6 | 107 | * |
b0c29f79 DB |
108 | * The correct serialization ensures that a waiter either observes |
109 | * the changed user space value before blocking or is woken by a | |
110 | * concurrent waker: | |
111 | * | |
112 | * CPU 0 CPU 1 | |
99b60ce6 TG |
113 | * val = *futex; |
114 | * sys_futex(WAIT, futex, val); | |
115 | * futex_wait(futex, val); | |
b0c29f79 | 116 | * |
d7e8af1a | 117 | * waiters++; (a) |
8ad7b378 DB |
118 | * smp_mb(); (A) <-- paired with -. |
119 | * | | |
120 | * lock(hash_bucket(futex)); | | |
121 | * | | |
122 | * uval = *futex; | | |
123 | * | *futex = newval; | |
124 | * | sys_futex(WAKE, futex); | |
125 | * | futex_wake(futex); | |
126 | * | | |
127 | * `--------> smp_mb(); (B) | |
99b60ce6 | 128 | * if (uval == val) |
b0c29f79 | 129 | * queue(); |
99b60ce6 | 130 | * unlock(hash_bucket(futex)); |
b0c29f79 DB |
131 | * schedule(); if (waiters) |
132 | * lock(hash_bucket(futex)); | |
d7e8af1a DB |
133 | * else wake_waiters(futex); |
134 | * waiters--; (b) unlock(hash_bucket(futex)); | |
b0c29f79 | 135 | * |
d7e8af1a DB |
136 | * Where (A) orders the waiters increment and the futex value read through |
137 | * atomic operations (see hb_waiters_inc) and where (B) orders the write | |
993b2ff2 DB |
138 | * to futex and the waiters read -- this is done by the barriers for both |
139 | * shared and private futexes in get_futex_key_refs(). | |
b0c29f79 DB |
140 | * |
141 | * This yields the following case (where X:=waiters, Y:=futex): | |
142 | * | |
143 | * X = Y = 0 | |
144 | * | |
145 | * w[X]=1 w[Y]=1 | |
146 | * MB MB | |
147 | * r[Y]=y r[X]=x | |
148 | * | |
149 | * Which guarantees that x==0 && y==0 is impossible; which translates back into | |
150 | * the guarantee that we cannot both miss the futex variable change and the | |
151 | * enqueue. | |
d7e8af1a DB |
152 | * |
153 | * Note that a new waiter is accounted for in (a) even when it is possible that | |
154 | * the wait call can return error, in which case we backtrack from it in (b). | |
155 | * Refer to the comment in queue_lock(). | |
156 | * | |
157 | * Similarly, in order to account for waiters being requeued on another | |
158 | * address we always increment the waiters for the destination bucket before | |
159 | * acquiring the lock. It then decrements them again after releasing it - | |
160 | * the code that actually moves the futex(es) between hash buckets (requeue_futex) | |
161 | * will do the additional required waiter count housekeeping. This is done for | |
162 | * double_lock_hb() and double_unlock_hb(), respectively. | |
99b60ce6 TG |
163 | */ |
164 | ||
04e7712f AB |
165 | #ifdef CONFIG_HAVE_FUTEX_CMPXCHG |
166 | #define futex_cmpxchg_enabled 1 | |
167 | #else | |
168 | static int __read_mostly futex_cmpxchg_enabled; | |
03b8c7b6 | 169 | #endif |
a0c1e907 | 170 | |
b41277dc DH |
171 | /* |
172 | * Futex flags used to encode options to functions and preserve them across | |
173 | * restarts. | |
174 | */ | |
784bdf3b TG |
175 | #ifdef CONFIG_MMU |
176 | # define FLAGS_SHARED 0x01 | |
177 | #else | |
178 | /* | |
179 | * NOMMU does not have per process address space. Let the compiler optimize | |
180 | * code away. | |
181 | */ | |
182 | # define FLAGS_SHARED 0x00 | |
183 | #endif | |
b41277dc DH |
184 | #define FLAGS_CLOCKRT 0x02 |
185 | #define FLAGS_HAS_TIMEOUT 0x04 | |
186 | ||
c87e2837 IM |
187 | /* |
188 | * Priority Inheritance state: | |
189 | */ | |
190 | struct futex_pi_state { | |
191 | /* | |
192 | * list of 'owned' pi_state instances - these have to be | |
193 | * cleaned up in do_exit() if the task exits prematurely: | |
194 | */ | |
195 | struct list_head list; | |
196 | ||
197 | /* | |
198 | * The PI object: | |
199 | */ | |
200 | struct rt_mutex pi_mutex; | |
201 | ||
202 | struct task_struct *owner; | |
49262de2 | 203 | refcount_t refcount; |
c87e2837 IM |
204 | |
205 | union futex_key key; | |
3859a271 | 206 | } __randomize_layout; |
c87e2837 | 207 | |
d8d88fbb DH |
208 | /** |
209 | * struct futex_q - The hashed futex queue entry, one per waiting task | |
fb62db2b | 210 | * @list: priority-sorted list of tasks waiting on this futex |
d8d88fbb DH |
211 | * @task: the task waiting on the futex |
212 | * @lock_ptr: the hash bucket lock | |
213 | * @key: the key the futex is hashed on | |
214 | * @pi_state: optional priority inheritance state | |
215 | * @rt_waiter: rt_waiter storage for use with requeue_pi | |
216 | * @requeue_pi_key: the requeue_pi target futex key | |
217 | * @bitset: bitset for the optional bitmasked wakeup | |
218 | * | |
ac6424b9 | 219 | * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so |
1da177e4 LT |
220 | * we can wake only the relevant ones (hashed queues may be shared). |
221 | * | |
222 | * A futex_q has a woken state, just like tasks have TASK_RUNNING. | |
ec92d082 | 223 | * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0. |
fb62db2b | 224 | * The order of wakeup is always to make the first condition true, then |
d8d88fbb DH |
225 | * the second. |
226 | * | |
227 | * PI futexes are typically woken before they are removed from the hash list via | |
228 | * the rt_mutex code. See unqueue_me_pi(). | |
1da177e4 LT |
229 | */ |
230 | struct futex_q { | |
ec92d082 | 231 | struct plist_node list; |
1da177e4 | 232 | |
d8d88fbb | 233 | struct task_struct *task; |
1da177e4 | 234 | spinlock_t *lock_ptr; |
1da177e4 | 235 | union futex_key key; |
c87e2837 | 236 | struct futex_pi_state *pi_state; |
52400ba9 | 237 | struct rt_mutex_waiter *rt_waiter; |
84bc4af5 | 238 | union futex_key *requeue_pi_key; |
cd689985 | 239 | u32 bitset; |
3859a271 | 240 | } __randomize_layout; |
1da177e4 | 241 | |
5bdb05f9 DH |
242 | static const struct futex_q futex_q_init = { |
243 | /* list gets initialized in queue_me()*/ | |
244 | .key = FUTEX_KEY_INIT, | |
245 | .bitset = FUTEX_BITSET_MATCH_ANY | |
246 | }; | |
247 | ||
1da177e4 | 248 | /* |
b2d0994b DH |
249 | * Hash buckets are shared by all the futex_keys that hash to the same |
250 | * location. Each key may have multiple futex_q structures, one for each task | |
251 | * waiting on a futex. | |
1da177e4 LT |
252 | */ |
253 | struct futex_hash_bucket { | |
11d4616b | 254 | atomic_t waiters; |
ec92d082 PP |
255 | spinlock_t lock; |
256 | struct plist_head chain; | |
a52b89eb | 257 | } ____cacheline_aligned_in_smp; |
1da177e4 | 258 | |
ac742d37 RV |
259 | /* |
260 | * The base of the bucket array and its size are always used together | |
261 | * (after initialization only in hash_futex()), so ensure that they | |
262 | * reside in the same cacheline. | |
263 | */ | |
264 | static struct { | |
265 | struct futex_hash_bucket *queues; | |
266 | unsigned long hashsize; | |
267 | } __futex_data __read_mostly __aligned(2*sizeof(long)); | |
268 | #define futex_queues (__futex_data.queues) | |
269 | #define futex_hashsize (__futex_data.hashsize) | |
a52b89eb | 270 | |
1da177e4 | 271 | |
ab51fbab DB |
272 | /* |
273 | * Fault injections for futexes. | |
274 | */ | |
275 | #ifdef CONFIG_FAIL_FUTEX | |
276 | ||
277 | static struct { | |
278 | struct fault_attr attr; | |
279 | ||
621a5f7a | 280 | bool ignore_private; |
ab51fbab DB |
281 | } fail_futex = { |
282 | .attr = FAULT_ATTR_INITIALIZER, | |
621a5f7a | 283 | .ignore_private = false, |
ab51fbab DB |
284 | }; |
285 | ||
286 | static int __init setup_fail_futex(char *str) | |
287 | { | |
288 | return setup_fault_attr(&fail_futex.attr, str); | |
289 | } | |
290 | __setup("fail_futex=", setup_fail_futex); | |
291 | ||
5d285a7f | 292 | static bool should_fail_futex(bool fshared) |
ab51fbab DB |
293 | { |
294 | if (fail_futex.ignore_private && !fshared) | |
295 | return false; | |
296 | ||
297 | return should_fail(&fail_futex.attr, 1); | |
298 | } | |
299 | ||
300 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
301 | ||
302 | static int __init fail_futex_debugfs(void) | |
303 | { | |
304 | umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
305 | struct dentry *dir; | |
306 | ||
307 | dir = fault_create_debugfs_attr("fail_futex", NULL, | |
308 | &fail_futex.attr); | |
309 | if (IS_ERR(dir)) | |
310 | return PTR_ERR(dir); | |
311 | ||
0365aeba GKH |
312 | debugfs_create_bool("ignore-private", mode, dir, |
313 | &fail_futex.ignore_private); | |
ab51fbab DB |
314 | return 0; |
315 | } | |
316 | ||
317 | late_initcall(fail_futex_debugfs); | |
318 | ||
319 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
320 | ||
321 | #else | |
322 | static inline bool should_fail_futex(bool fshared) | |
323 | { | |
324 | return false; | |
325 | } | |
326 | #endif /* CONFIG_FAIL_FUTEX */ | |
327 | ||
ba31c1a4 TG |
328 | #ifdef CONFIG_COMPAT |
329 | static void compat_exit_robust_list(struct task_struct *curr); | |
330 | #else | |
331 | static inline void compat_exit_robust_list(struct task_struct *curr) { } | |
332 | #endif | |
333 | ||
b0c29f79 DB |
334 | static inline void futex_get_mm(union futex_key *key) |
335 | { | |
f1f10076 | 336 | mmgrab(key->private.mm); |
b0c29f79 DB |
337 | /* |
338 | * Ensure futex_get_mm() implies a full barrier such that | |
339 | * get_futex_key() implies a full barrier. This is relied upon | |
8ad7b378 | 340 | * as smp_mb(); (B), see the ordering comment above. |
b0c29f79 | 341 | */ |
4e857c58 | 342 | smp_mb__after_atomic(); |
b0c29f79 DB |
343 | } |
344 | ||
11d4616b LT |
345 | /* |
346 | * Reflects a new waiter being added to the waitqueue. | |
347 | */ | |
348 | static inline void hb_waiters_inc(struct futex_hash_bucket *hb) | |
b0c29f79 DB |
349 | { |
350 | #ifdef CONFIG_SMP | |
11d4616b | 351 | atomic_inc(&hb->waiters); |
b0c29f79 | 352 | /* |
11d4616b | 353 | * Full barrier (A), see the ordering comment above. |
b0c29f79 | 354 | */ |
4e857c58 | 355 | smp_mb__after_atomic(); |
11d4616b LT |
356 | #endif |
357 | } | |
358 | ||
359 | /* | |
360 | * Reflects a waiter being removed from the waitqueue by wakeup | |
361 | * paths. | |
362 | */ | |
363 | static inline void hb_waiters_dec(struct futex_hash_bucket *hb) | |
364 | { | |
365 | #ifdef CONFIG_SMP | |
366 | atomic_dec(&hb->waiters); | |
367 | #endif | |
368 | } | |
b0c29f79 | 369 | |
11d4616b LT |
370 | static inline int hb_waiters_pending(struct futex_hash_bucket *hb) |
371 | { | |
372 | #ifdef CONFIG_SMP | |
373 | return atomic_read(&hb->waiters); | |
b0c29f79 | 374 | #else |
11d4616b | 375 | return 1; |
b0c29f79 DB |
376 | #endif |
377 | } | |
378 | ||
e8b61b3f TG |
379 | /** |
380 | * hash_futex - Return the hash bucket in the global hash | |
381 | * @key: Pointer to the futex key for which the hash is calculated | |
382 | * | |
383 | * We hash on the keys returned from get_futex_key (see below) and return the | |
384 | * corresponding hash bucket in the global hash. | |
1da177e4 LT |
385 | */ |
386 | static struct futex_hash_bucket *hash_futex(union futex_key *key) | |
387 | { | |
388 | u32 hash = jhash2((u32*)&key->both.word, | |
389 | (sizeof(key->both.word)+sizeof(key->both.ptr))/4, | |
390 | key->both.offset); | |
a52b89eb | 391 | return &futex_queues[hash & (futex_hashsize - 1)]; |
1da177e4 LT |
392 | } |
393 | ||
e8b61b3f TG |
394 | |
395 | /** | |
396 | * match_futex - Check whether two futex keys are equal | |
397 | * @key1: Pointer to key1 | |
398 | * @key2: Pointer to key2 | |
399 | * | |
1da177e4 LT |
400 | * Return 1 if two futex_keys are equal, 0 otherwise. |
401 | */ | |
402 | static inline int match_futex(union futex_key *key1, union futex_key *key2) | |
403 | { | |
2bc87203 DH |
404 | return (key1 && key2 |
405 | && key1->both.word == key2->both.word | |
1da177e4 LT |
406 | && key1->both.ptr == key2->both.ptr |
407 | && key1->both.offset == key2->both.offset); | |
408 | } | |
409 | ||
38d47c1b PZ |
410 | /* |
411 | * Take a reference to the resource addressed by a key. | |
412 | * Can be called while holding spinlocks. | |
413 | * | |
414 | */ | |
415 | static void get_futex_key_refs(union futex_key *key) | |
416 | { | |
417 | if (!key->both.ptr) | |
418 | return; | |
419 | ||
784bdf3b TG |
420 | /* |
421 | * On MMU less systems futexes are always "private" as there is no per | |
422 | * process address space. We need the smp wmb nevertheless - yes, | |
423 | * arch/blackfin has MMU less SMP ... | |
424 | */ | |
425 | if (!IS_ENABLED(CONFIG_MMU)) { | |
426 | smp_mb(); /* explicit smp_mb(); (B) */ | |
427 | return; | |
428 | } | |
429 | ||
38d47c1b PZ |
430 | switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) { |
431 | case FUT_OFF_INODE: | |
8019ad13 | 432 | smp_mb(); /* explicit smp_mb(); (B) */ |
38d47c1b PZ |
433 | break; |
434 | case FUT_OFF_MMSHARED: | |
8ad7b378 | 435 | futex_get_mm(key); /* implies smp_mb(); (B) */ |
38d47c1b | 436 | break; |
76835b0e | 437 | default: |
993b2ff2 DB |
438 | /* |
439 | * Private futexes do not hold reference on an inode or | |
440 | * mm, therefore the only purpose of calling get_futex_key_refs | |
441 | * is because we need the barrier for the lockless waiter check. | |
442 | */ | |
8ad7b378 | 443 | smp_mb(); /* explicit smp_mb(); (B) */ |
38d47c1b PZ |
444 | } |
445 | } | |
446 | ||
447 | /* | |
448 | * Drop a reference to the resource addressed by a key. | |
993b2ff2 DB |
449 | * The hash bucket spinlock must not be held. This is |
450 | * a no-op for private futexes, see comment in the get | |
451 | * counterpart. | |
38d47c1b PZ |
452 | */ |
453 | static void drop_futex_key_refs(union futex_key *key) | |
454 | { | |
90621c40 DH |
455 | if (!key->both.ptr) { |
456 | /* If we're here then we tried to put a key we failed to get */ | |
457 | WARN_ON_ONCE(1); | |
38d47c1b | 458 | return; |
90621c40 | 459 | } |
38d47c1b | 460 | |
784bdf3b TG |
461 | if (!IS_ENABLED(CONFIG_MMU)) |
462 | return; | |
463 | ||
38d47c1b PZ |
464 | switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) { |
465 | case FUT_OFF_INODE: | |
38d47c1b PZ |
466 | break; |
467 | case FUT_OFF_MMSHARED: | |
468 | mmdrop(key->private.mm); | |
469 | break; | |
470 | } | |
471 | } | |
472 | ||
96d4f267 LT |
473 | enum futex_access { |
474 | FUTEX_READ, | |
475 | FUTEX_WRITE | |
476 | }; | |
477 | ||
5ca584d9 WL |
478 | /** |
479 | * futex_setup_timer - set up the sleeping hrtimer. | |
480 | * @time: ptr to the given timeout value | |
481 | * @timeout: the hrtimer_sleeper structure to be set up | |
482 | * @flags: futex flags | |
483 | * @range_ns: optional range in ns | |
484 | * | |
485 | * Return: Initialized hrtimer_sleeper structure or NULL if no timeout | |
486 | * value given | |
487 | */ | |
488 | static inline struct hrtimer_sleeper * | |
489 | futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout, | |
490 | int flags, u64 range_ns) | |
491 | { | |
492 | if (!time) | |
493 | return NULL; | |
494 | ||
dbc1625f SAS |
495 | hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ? |
496 | CLOCK_REALTIME : CLOCK_MONOTONIC, | |
497 | HRTIMER_MODE_ABS); | |
5ca584d9 WL |
498 | /* |
499 | * If range_ns is 0, calling hrtimer_set_expires_range_ns() is | |
500 | * effectively the same as calling hrtimer_set_expires(). | |
501 | */ | |
502 | hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns); | |
503 | ||
504 | return timeout; | |
505 | } | |
506 | ||
8019ad13 PZ |
507 | /* |
508 | * Generate a machine wide unique identifier for this inode. | |
509 | * | |
510 | * This relies on u64 not wrapping in the life-time of the machine; which with | |
511 | * 1ns resolution means almost 585 years. | |
512 | * | |
513 | * This further relies on the fact that a well formed program will not unmap | |
514 | * the file while it has a (shared) futex waiting on it. This mapping will have | |
515 | * a file reference which pins the mount and inode. | |
516 | * | |
517 | * If for some reason an inode gets evicted and read back in again, it will get | |
518 | * a new sequence number and will _NOT_ match, even though it is the exact same | |
519 | * file. | |
520 | * | |
521 | * It is important that match_futex() will never have a false-positive, esp. | |
522 | * for PI futexes that can mess up the state. The above argues that false-negatives | |
523 | * are only possible for malformed programs. | |
524 | */ | |
525 | static u64 get_inode_sequence_number(struct inode *inode) | |
526 | { | |
527 | static atomic64_t i_seq; | |
528 | u64 old; | |
529 | ||
530 | /* Does the inode already have a sequence number? */ | |
531 | old = atomic64_read(&inode->i_sequence); | |
532 | if (likely(old)) | |
533 | return old; | |
534 | ||
535 | for (;;) { | |
536 | u64 new = atomic64_add_return(1, &i_seq); | |
537 | if (WARN_ON_ONCE(!new)) | |
538 | continue; | |
539 | ||
540 | old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new); | |
541 | if (old) | |
542 | return old; | |
543 | return new; | |
544 | } | |
545 | } | |
546 | ||
34f01cc1 | 547 | /** |
d96ee56c DH |
548 | * get_futex_key() - Get parameters which are the keys for a futex |
549 | * @uaddr: virtual address of the futex | |
550 | * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED | |
551 | * @key: address where result is stored. | |
96d4f267 LT |
552 | * @rw: mapping needs to be read/write (values: FUTEX_READ, |
553 | * FUTEX_WRITE) | |
34f01cc1 | 554 | * |
6c23cbbd RD |
555 | * Return: a negative error code or 0 |
556 | * | |
7b4ff1ad | 557 | * The key words are stored in @key on success. |
1da177e4 | 558 | * |
8019ad13 PZ |
559 | * For shared mappings (when @fshared), the key is: |
560 | * ( inode->i_sequence, page->index, offset_within_page ) | |
561 | * [ also see get_inode_sequence_number() ] | |
562 | * | |
563 | * For private mappings (or when !@fshared), the key is: | |
564 | * ( current->mm, address, 0 ) | |
565 | * | |
566 | * This allows (cross process, where applicable) identification of the futex | |
567 | * without keeping the page pinned for the duration of the FUTEX_WAIT. | |
1da177e4 | 568 | * |
b2d0994b | 569 | * lock_page() might sleep, the caller should not hold a spinlock. |
1da177e4 | 570 | */ |
64d1304a | 571 | static int |
96d4f267 | 572 | get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, enum futex_access rw) |
1da177e4 | 573 | { |
e2970f2f | 574 | unsigned long address = (unsigned long)uaddr; |
1da177e4 | 575 | struct mm_struct *mm = current->mm; |
077fa7ae | 576 | struct page *page, *tail; |
14d27abd | 577 | struct address_space *mapping; |
9ea71503 | 578 | int err, ro = 0; |
1da177e4 LT |
579 | |
580 | /* | |
581 | * The futex address must be "naturally" aligned. | |
582 | */ | |
e2970f2f | 583 | key->both.offset = address % PAGE_SIZE; |
34f01cc1 | 584 | if (unlikely((address % sizeof(u32)) != 0)) |
1da177e4 | 585 | return -EINVAL; |
e2970f2f | 586 | address -= key->both.offset; |
1da177e4 | 587 | |
96d4f267 | 588 | if (unlikely(!access_ok(uaddr, sizeof(u32)))) |
5cdec2d8 LT |
589 | return -EFAULT; |
590 | ||
ab51fbab DB |
591 | if (unlikely(should_fail_futex(fshared))) |
592 | return -EFAULT; | |
593 | ||
34f01cc1 ED |
594 | /* |
595 | * PROCESS_PRIVATE futexes are fast. | |
596 | * As the mm cannot disappear under us and the 'key' only needs | |
597 | * virtual address, we dont even have to find the underlying vma. | |
598 | * Note : We do have to check 'uaddr' is a valid user address, | |
599 | * but access_ok() should be faster than find_vma() | |
600 | */ | |
601 | if (!fshared) { | |
34f01cc1 ED |
602 | key->private.mm = mm; |
603 | key->private.address = address; | |
8ad7b378 | 604 | get_futex_key_refs(key); /* implies smp_mb(); (B) */ |
34f01cc1 ED |
605 | return 0; |
606 | } | |
1da177e4 | 607 | |
38d47c1b | 608 | again: |
ab51fbab DB |
609 | /* Ignore any VERIFY_READ mapping (futex common case) */ |
610 | if (unlikely(should_fail_futex(fshared))) | |
611 | return -EFAULT; | |
612 | ||
73b0140b | 613 | err = get_user_pages_fast(address, 1, FOLL_WRITE, &page); |
9ea71503 SB |
614 | /* |
615 | * If write access is not required (eg. FUTEX_WAIT), try | |
616 | * and get read-only access. | |
617 | */ | |
96d4f267 | 618 | if (err == -EFAULT && rw == FUTEX_READ) { |
9ea71503 SB |
619 | err = get_user_pages_fast(address, 1, 0, &page); |
620 | ro = 1; | |
621 | } | |
38d47c1b PZ |
622 | if (err < 0) |
623 | return err; | |
9ea71503 SB |
624 | else |
625 | err = 0; | |
38d47c1b | 626 | |
65d8fc77 MG |
627 | /* |
628 | * The treatment of mapping from this point on is critical. The page | |
629 | * lock protects many things but in this context the page lock | |
630 | * stabilizes mapping, prevents inode freeing in the shared | |
631 | * file-backed region case and guards against movement to swap cache. | |
632 | * | |
633 | * Strictly speaking the page lock is not needed in all cases being | |
634 | * considered here and page lock forces unnecessarily serialization | |
635 | * From this point on, mapping will be re-verified if necessary and | |
636 | * page lock will be acquired only if it is unavoidable | |
077fa7ae MG |
637 | * |
638 | * Mapping checks require the head page for any compound page so the | |
639 | * head page and mapping is looked up now. For anonymous pages, it | |
640 | * does not matter if the page splits in the future as the key is | |
641 | * based on the address. For filesystem-backed pages, the tail is | |
642 | * required as the index of the page determines the key. For | |
643 | * base pages, there is no tail page and tail == page. | |
65d8fc77 | 644 | */ |
077fa7ae | 645 | tail = page; |
65d8fc77 MG |
646 | page = compound_head(page); |
647 | mapping = READ_ONCE(page->mapping); | |
648 | ||
e6780f72 | 649 | /* |
14d27abd | 650 | * If page->mapping is NULL, then it cannot be a PageAnon |
e6780f72 HD |
651 | * page; but it might be the ZERO_PAGE or in the gate area or |
652 | * in a special mapping (all cases which we are happy to fail); | |
653 | * or it may have been a good file page when get_user_pages_fast | |
654 | * found it, but truncated or holepunched or subjected to | |
655 | * invalidate_complete_page2 before we got the page lock (also | |
656 | * cases which we are happy to fail). And we hold a reference, | |
657 | * so refcount care in invalidate_complete_page's remove_mapping | |
658 | * prevents drop_caches from setting mapping to NULL beneath us. | |
659 | * | |
660 | * The case we do have to guard against is when memory pressure made | |
661 | * shmem_writepage move it from filecache to swapcache beneath us: | |
14d27abd | 662 | * an unlikely race, but we do need to retry for page->mapping. |
e6780f72 | 663 | */ |
65d8fc77 MG |
664 | if (unlikely(!mapping)) { |
665 | int shmem_swizzled; | |
666 | ||
667 | /* | |
668 | * Page lock is required to identify which special case above | |
669 | * applies. If this is really a shmem page then the page lock | |
670 | * will prevent unexpected transitions. | |
671 | */ | |
672 | lock_page(page); | |
673 | shmem_swizzled = PageSwapCache(page) || page->mapping; | |
14d27abd KS |
674 | unlock_page(page); |
675 | put_page(page); | |
65d8fc77 | 676 | |
e6780f72 HD |
677 | if (shmem_swizzled) |
678 | goto again; | |
65d8fc77 | 679 | |
e6780f72 | 680 | return -EFAULT; |
38d47c1b | 681 | } |
1da177e4 LT |
682 | |
683 | /* | |
684 | * Private mappings are handled in a simple way. | |
685 | * | |
65d8fc77 MG |
686 | * If the futex key is stored on an anonymous page, then the associated |
687 | * object is the mm which is implicitly pinned by the calling process. | |
688 | * | |
1da177e4 LT |
689 | * NOTE: When userspace waits on a MAP_SHARED mapping, even if |
690 | * it's a read-only handle, it's expected that futexes attach to | |
38d47c1b | 691 | * the object not the particular process. |
1da177e4 | 692 | */ |
14d27abd | 693 | if (PageAnon(page)) { |
9ea71503 SB |
694 | /* |
695 | * A RO anonymous page will never change and thus doesn't make | |
696 | * sense for futex operations. | |
697 | */ | |
ab51fbab | 698 | if (unlikely(should_fail_futex(fshared)) || ro) { |
9ea71503 SB |
699 | err = -EFAULT; |
700 | goto out; | |
701 | } | |
702 | ||
38d47c1b | 703 | key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */ |
1da177e4 | 704 | key->private.mm = mm; |
e2970f2f | 705 | key->private.address = address; |
65d8fc77 | 706 | |
38d47c1b | 707 | } else { |
65d8fc77 MG |
708 | struct inode *inode; |
709 | ||
710 | /* | |
711 | * The associated futex object in this case is the inode and | |
712 | * the page->mapping must be traversed. Ordinarily this should | |
713 | * be stabilised under page lock but it's not strictly | |
714 | * necessary in this case as we just want to pin the inode, not | |
715 | * update the radix tree or anything like that. | |
716 | * | |
717 | * The RCU read lock is taken as the inode is finally freed | |
718 | * under RCU. If the mapping still matches expectations then the | |
719 | * mapping->host can be safely accessed as being a valid inode. | |
720 | */ | |
721 | rcu_read_lock(); | |
722 | ||
723 | if (READ_ONCE(page->mapping) != mapping) { | |
724 | rcu_read_unlock(); | |
725 | put_page(page); | |
726 | ||
727 | goto again; | |
728 | } | |
729 | ||
730 | inode = READ_ONCE(mapping->host); | |
731 | if (!inode) { | |
732 | rcu_read_unlock(); | |
733 | put_page(page); | |
734 | ||
735 | goto again; | |
736 | } | |
737 | ||
38d47c1b | 738 | key->both.offset |= FUT_OFF_INODE; /* inode-based key */ |
8019ad13 | 739 | key->shared.i_seq = get_inode_sequence_number(inode); |
077fa7ae | 740 | key->shared.pgoff = basepage_index(tail); |
65d8fc77 | 741 | rcu_read_unlock(); |
1da177e4 LT |
742 | } |
743 | ||
8019ad13 PZ |
744 | get_futex_key_refs(key); /* implies smp_mb(); (B) */ |
745 | ||
9ea71503 | 746 | out: |
14d27abd | 747 | put_page(page); |
9ea71503 | 748 | return err; |
1da177e4 LT |
749 | } |
750 | ||
ae791a2d | 751 | static inline void put_futex_key(union futex_key *key) |
1da177e4 | 752 | { |
38d47c1b | 753 | drop_futex_key_refs(key); |
1da177e4 LT |
754 | } |
755 | ||
d96ee56c DH |
756 | /** |
757 | * fault_in_user_writeable() - Fault in user address and verify RW access | |
d0725992 TG |
758 | * @uaddr: pointer to faulting user space address |
759 | * | |
760 | * Slow path to fixup the fault we just took in the atomic write | |
761 | * access to @uaddr. | |
762 | * | |
fb62db2b | 763 | * We have no generic implementation of a non-destructive write to the |
d0725992 TG |
764 | * user address. We know that we faulted in the atomic pagefault |
765 | * disabled section so we can as well avoid the #PF overhead by | |
766 | * calling get_user_pages() right away. | |
767 | */ | |
768 | static int fault_in_user_writeable(u32 __user *uaddr) | |
769 | { | |
722d0172 AK |
770 | struct mm_struct *mm = current->mm; |
771 | int ret; | |
772 | ||
773 | down_read(&mm->mmap_sem); | |
2efaca92 | 774 | ret = fixup_user_fault(current, mm, (unsigned long)uaddr, |
4a9e1cda | 775 | FAULT_FLAG_WRITE, NULL); |
722d0172 AK |
776 | up_read(&mm->mmap_sem); |
777 | ||
d0725992 TG |
778 | return ret < 0 ? ret : 0; |
779 | } | |
780 | ||
4b1c486b DH |
781 | /** |
782 | * futex_top_waiter() - Return the highest priority waiter on a futex | |
d96ee56c DH |
783 | * @hb: the hash bucket the futex_q's reside in |
784 | * @key: the futex key (to distinguish it from other futex futex_q's) | |
4b1c486b DH |
785 | * |
786 | * Must be called with the hb lock held. | |
787 | */ | |
788 | static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb, | |
789 | union futex_key *key) | |
790 | { | |
791 | struct futex_q *this; | |
792 | ||
793 | plist_for_each_entry(this, &hb->chain, list) { | |
794 | if (match_futex(&this->key, key)) | |
795 | return this; | |
796 | } | |
797 | return NULL; | |
798 | } | |
799 | ||
37a9d912 ML |
800 | static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr, |
801 | u32 uval, u32 newval) | |
36cf3b5c | 802 | { |
37a9d912 | 803 | int ret; |
36cf3b5c TG |
804 | |
805 | pagefault_disable(); | |
37a9d912 | 806 | ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval); |
36cf3b5c TG |
807 | pagefault_enable(); |
808 | ||
37a9d912 | 809 | return ret; |
36cf3b5c TG |
810 | } |
811 | ||
812 | static int get_futex_value_locked(u32 *dest, u32 __user *from) | |
1da177e4 LT |
813 | { |
814 | int ret; | |
815 | ||
a866374a | 816 | pagefault_disable(); |
bd28b145 | 817 | ret = __get_user(*dest, from); |
a866374a | 818 | pagefault_enable(); |
1da177e4 LT |
819 | |
820 | return ret ? -EFAULT : 0; | |
821 | } | |
822 | ||
c87e2837 IM |
823 | |
824 | /* | |
825 | * PI code: | |
826 | */ | |
827 | static int refill_pi_state_cache(void) | |
828 | { | |
829 | struct futex_pi_state *pi_state; | |
830 | ||
831 | if (likely(current->pi_state_cache)) | |
832 | return 0; | |
833 | ||
4668edc3 | 834 | pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL); |
c87e2837 IM |
835 | |
836 | if (!pi_state) | |
837 | return -ENOMEM; | |
838 | ||
c87e2837 IM |
839 | INIT_LIST_HEAD(&pi_state->list); |
840 | /* pi_mutex gets initialized later */ | |
841 | pi_state->owner = NULL; | |
49262de2 | 842 | refcount_set(&pi_state->refcount, 1); |
38d47c1b | 843 | pi_state->key = FUTEX_KEY_INIT; |
c87e2837 IM |
844 | |
845 | current->pi_state_cache = pi_state; | |
846 | ||
847 | return 0; | |
848 | } | |
849 | ||
bf92cf3a | 850 | static struct futex_pi_state *alloc_pi_state(void) |
c87e2837 IM |
851 | { |
852 | struct futex_pi_state *pi_state = current->pi_state_cache; | |
853 | ||
854 | WARN_ON(!pi_state); | |
855 | current->pi_state_cache = NULL; | |
856 | ||
857 | return pi_state; | |
858 | } | |
859 | ||
bf92cf3a PZ |
860 | static void get_pi_state(struct futex_pi_state *pi_state) |
861 | { | |
49262de2 | 862 | WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount)); |
bf92cf3a PZ |
863 | } |
864 | ||
30a6b803 | 865 | /* |
29e9ee5d TG |
866 | * Drops a reference to the pi_state object and frees or caches it |
867 | * when the last reference is gone. | |
30a6b803 | 868 | */ |
29e9ee5d | 869 | static void put_pi_state(struct futex_pi_state *pi_state) |
c87e2837 | 870 | { |
30a6b803 BS |
871 | if (!pi_state) |
872 | return; | |
873 | ||
49262de2 | 874 | if (!refcount_dec_and_test(&pi_state->refcount)) |
c87e2837 IM |
875 | return; |
876 | ||
877 | /* | |
878 | * If pi_state->owner is NULL, the owner is most probably dying | |
879 | * and has cleaned up the pi_state already | |
880 | */ | |
881 | if (pi_state->owner) { | |
c74aef2d | 882 | struct task_struct *owner; |
c87e2837 | 883 | |
c74aef2d PZ |
884 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); |
885 | owner = pi_state->owner; | |
886 | if (owner) { | |
887 | raw_spin_lock(&owner->pi_lock); | |
888 | list_del_init(&pi_state->list); | |
889 | raw_spin_unlock(&owner->pi_lock); | |
890 | } | |
891 | rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner); | |
892 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
c87e2837 IM |
893 | } |
894 | ||
c74aef2d | 895 | if (current->pi_state_cache) { |
c87e2837 | 896 | kfree(pi_state); |
c74aef2d | 897 | } else { |
c87e2837 IM |
898 | /* |
899 | * pi_state->list is already empty. | |
900 | * clear pi_state->owner. | |
901 | * refcount is at 0 - put it back to 1. | |
902 | */ | |
903 | pi_state->owner = NULL; | |
49262de2 | 904 | refcount_set(&pi_state->refcount, 1); |
c87e2837 IM |
905 | current->pi_state_cache = pi_state; |
906 | } | |
907 | } | |
908 | ||
bc2eecd7 NP |
909 | #ifdef CONFIG_FUTEX_PI |
910 | ||
c87e2837 IM |
911 | /* |
912 | * This task is holding PI mutexes at exit time => bad. | |
913 | * Kernel cleans up PI-state, but userspace is likely hosed. | |
914 | * (Robust-futex cleanup is separate and might save the day for userspace.) | |
915 | */ | |
ba31c1a4 | 916 | static void exit_pi_state_list(struct task_struct *curr) |
c87e2837 | 917 | { |
c87e2837 IM |
918 | struct list_head *next, *head = &curr->pi_state_list; |
919 | struct futex_pi_state *pi_state; | |
627371d7 | 920 | struct futex_hash_bucket *hb; |
38d47c1b | 921 | union futex_key key = FUTEX_KEY_INIT; |
c87e2837 | 922 | |
a0c1e907 TG |
923 | if (!futex_cmpxchg_enabled) |
924 | return; | |
c87e2837 IM |
925 | /* |
926 | * We are a ZOMBIE and nobody can enqueue itself on | |
927 | * pi_state_list anymore, but we have to be careful | |
627371d7 | 928 | * versus waiters unqueueing themselves: |
c87e2837 | 929 | */ |
1d615482 | 930 | raw_spin_lock_irq(&curr->pi_lock); |
c87e2837 | 931 | while (!list_empty(head)) { |
c87e2837 IM |
932 | next = head->next; |
933 | pi_state = list_entry(next, struct futex_pi_state, list); | |
934 | key = pi_state->key; | |
627371d7 | 935 | hb = hash_futex(&key); |
153fbd12 PZ |
936 | |
937 | /* | |
938 | * We can race against put_pi_state() removing itself from the | |
939 | * list (a waiter going away). put_pi_state() will first | |
940 | * decrement the reference count and then modify the list, so | |
941 | * its possible to see the list entry but fail this reference | |
942 | * acquire. | |
943 | * | |
944 | * In that case; drop the locks to let put_pi_state() make | |
945 | * progress and retry the loop. | |
946 | */ | |
49262de2 | 947 | if (!refcount_inc_not_zero(&pi_state->refcount)) { |
153fbd12 PZ |
948 | raw_spin_unlock_irq(&curr->pi_lock); |
949 | cpu_relax(); | |
950 | raw_spin_lock_irq(&curr->pi_lock); | |
951 | continue; | |
952 | } | |
1d615482 | 953 | raw_spin_unlock_irq(&curr->pi_lock); |
c87e2837 | 954 | |
c87e2837 | 955 | spin_lock(&hb->lock); |
c74aef2d PZ |
956 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); |
957 | raw_spin_lock(&curr->pi_lock); | |
627371d7 IM |
958 | /* |
959 | * We dropped the pi-lock, so re-check whether this | |
960 | * task still owns the PI-state: | |
961 | */ | |
c87e2837 | 962 | if (head->next != next) { |
153fbd12 | 963 | /* retain curr->pi_lock for the loop invariant */ |
c74aef2d | 964 | raw_spin_unlock(&pi_state->pi_mutex.wait_lock); |
c87e2837 | 965 | spin_unlock(&hb->lock); |
153fbd12 | 966 | put_pi_state(pi_state); |
c87e2837 IM |
967 | continue; |
968 | } | |
969 | ||
c87e2837 | 970 | WARN_ON(pi_state->owner != curr); |
627371d7 IM |
971 | WARN_ON(list_empty(&pi_state->list)); |
972 | list_del_init(&pi_state->list); | |
c87e2837 | 973 | pi_state->owner = NULL; |
c87e2837 | 974 | |
153fbd12 | 975 | raw_spin_unlock(&curr->pi_lock); |
c74aef2d | 976 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); |
c87e2837 IM |
977 | spin_unlock(&hb->lock); |
978 | ||
16ffa12d PZ |
979 | rt_mutex_futex_unlock(&pi_state->pi_mutex); |
980 | put_pi_state(pi_state); | |
981 | ||
1d615482 | 982 | raw_spin_lock_irq(&curr->pi_lock); |
c87e2837 | 983 | } |
1d615482 | 984 | raw_spin_unlock_irq(&curr->pi_lock); |
c87e2837 | 985 | } |
ba31c1a4 TG |
986 | #else |
987 | static inline void exit_pi_state_list(struct task_struct *curr) { } | |
bc2eecd7 NP |
988 | #endif |
989 | ||
54a21788 TG |
990 | /* |
991 | * We need to check the following states: | |
992 | * | |
993 | * Waiter | pi_state | pi->owner | uTID | uODIED | ? | |
994 | * | |
995 | * [1] NULL | --- | --- | 0 | 0/1 | Valid | |
996 | * [2] NULL | --- | --- | >0 | 0/1 | Valid | |
997 | * | |
998 | * [3] Found | NULL | -- | Any | 0/1 | Invalid | |
999 | * | |
1000 | * [4] Found | Found | NULL | 0 | 1 | Valid | |
1001 | * [5] Found | Found | NULL | >0 | 1 | Invalid | |
1002 | * | |
1003 | * [6] Found | Found | task | 0 | 1 | Valid | |
1004 | * | |
1005 | * [7] Found | Found | NULL | Any | 0 | Invalid | |
1006 | * | |
1007 | * [8] Found | Found | task | ==taskTID | 0/1 | Valid | |
1008 | * [9] Found | Found | task | 0 | 0 | Invalid | |
1009 | * [10] Found | Found | task | !=taskTID | 0/1 | Invalid | |
1010 | * | |
1011 | * [1] Indicates that the kernel can acquire the futex atomically. We | |
1012 | * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit. | |
1013 | * | |
1014 | * [2] Valid, if TID does not belong to a kernel thread. If no matching | |
1015 | * thread is found then it indicates that the owner TID has died. | |
1016 | * | |
1017 | * [3] Invalid. The waiter is queued on a non PI futex | |
1018 | * | |
1019 | * [4] Valid state after exit_robust_list(), which sets the user space | |
1020 | * value to FUTEX_WAITERS | FUTEX_OWNER_DIED. | |
1021 | * | |
1022 | * [5] The user space value got manipulated between exit_robust_list() | |
1023 | * and exit_pi_state_list() | |
1024 | * | |
1025 | * [6] Valid state after exit_pi_state_list() which sets the new owner in | |
1026 | * the pi_state but cannot access the user space value. | |
1027 | * | |
1028 | * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set. | |
1029 | * | |
1030 | * [8] Owner and user space value match | |
1031 | * | |
1032 | * [9] There is no transient state which sets the user space TID to 0 | |
1033 | * except exit_robust_list(), but this is indicated by the | |
1034 | * FUTEX_OWNER_DIED bit. See [4] | |
1035 | * | |
1036 | * [10] There is no transient state which leaves owner and user space | |
1037 | * TID out of sync. | |
734009e9 PZ |
1038 | * |
1039 | * | |
1040 | * Serialization and lifetime rules: | |
1041 | * | |
1042 | * hb->lock: | |
1043 | * | |
1044 | * hb -> futex_q, relation | |
1045 | * futex_q -> pi_state, relation | |
1046 | * | |
1047 | * (cannot be raw because hb can contain arbitrary amount | |
1048 | * of futex_q's) | |
1049 | * | |
1050 | * pi_mutex->wait_lock: | |
1051 | * | |
1052 | * {uval, pi_state} | |
1053 | * | |
1054 | * (and pi_mutex 'obviously') | |
1055 | * | |
1056 | * p->pi_lock: | |
1057 | * | |
1058 | * p->pi_state_list -> pi_state->list, relation | |
1059 | * | |
1060 | * pi_state->refcount: | |
1061 | * | |
1062 | * pi_state lifetime | |
1063 | * | |
1064 | * | |
1065 | * Lock order: | |
1066 | * | |
1067 | * hb->lock | |
1068 | * pi_mutex->wait_lock | |
1069 | * p->pi_lock | |
1070 | * | |
54a21788 | 1071 | */ |
e60cbc5c TG |
1072 | |
1073 | /* | |
1074 | * Validate that the existing waiter has a pi_state and sanity check | |
1075 | * the pi_state against the user space value. If correct, attach to | |
1076 | * it. | |
1077 | */ | |
734009e9 PZ |
1078 | static int attach_to_pi_state(u32 __user *uaddr, u32 uval, |
1079 | struct futex_pi_state *pi_state, | |
e60cbc5c | 1080 | struct futex_pi_state **ps) |
c87e2837 | 1081 | { |
778e9a9c | 1082 | pid_t pid = uval & FUTEX_TID_MASK; |
94ffac5d PZ |
1083 | u32 uval2; |
1084 | int ret; | |
c87e2837 | 1085 | |
e60cbc5c TG |
1086 | /* |
1087 | * Userspace might have messed up non-PI and PI futexes [3] | |
1088 | */ | |
1089 | if (unlikely(!pi_state)) | |
1090 | return -EINVAL; | |
06a9ec29 | 1091 | |
734009e9 PZ |
1092 | /* |
1093 | * We get here with hb->lock held, and having found a | |
1094 | * futex_top_waiter(). This means that futex_lock_pi() of said futex_q | |
1095 | * has dropped the hb->lock in between queue_me() and unqueue_me_pi(), | |
1096 | * which in turn means that futex_lock_pi() still has a reference on | |
1097 | * our pi_state. | |
16ffa12d PZ |
1098 | * |
1099 | * The waiter holding a reference on @pi_state also protects against | |
1100 | * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi() | |
1101 | * and futex_wait_requeue_pi() as it cannot go to 0 and consequently | |
1102 | * free pi_state before we can take a reference ourselves. | |
734009e9 | 1103 | */ |
49262de2 | 1104 | WARN_ON(!refcount_read(&pi_state->refcount)); |
59647b6a | 1105 | |
734009e9 PZ |
1106 | /* |
1107 | * Now that we have a pi_state, we can acquire wait_lock | |
1108 | * and do the state validation. | |
1109 | */ | |
1110 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); | |
1111 | ||
1112 | /* | |
1113 | * Since {uval, pi_state} is serialized by wait_lock, and our current | |
1114 | * uval was read without holding it, it can have changed. Verify it | |
1115 | * still is what we expect it to be, otherwise retry the entire | |
1116 | * operation. | |
1117 | */ | |
1118 | if (get_futex_value_locked(&uval2, uaddr)) | |
1119 | goto out_efault; | |
1120 | ||
1121 | if (uval != uval2) | |
1122 | goto out_eagain; | |
1123 | ||
e60cbc5c TG |
1124 | /* |
1125 | * Handle the owner died case: | |
1126 | */ | |
1127 | if (uval & FUTEX_OWNER_DIED) { | |
bd1dbcc6 | 1128 | /* |
e60cbc5c TG |
1129 | * exit_pi_state_list sets owner to NULL and wakes the |
1130 | * topmost waiter. The task which acquires the | |
1131 | * pi_state->rt_mutex will fixup owner. | |
bd1dbcc6 | 1132 | */ |
e60cbc5c | 1133 | if (!pi_state->owner) { |
59647b6a | 1134 | /* |
e60cbc5c TG |
1135 | * No pi state owner, but the user space TID |
1136 | * is not 0. Inconsistent state. [5] | |
59647b6a | 1137 | */ |
e60cbc5c | 1138 | if (pid) |
734009e9 | 1139 | goto out_einval; |
bd1dbcc6 | 1140 | /* |
e60cbc5c | 1141 | * Take a ref on the state and return success. [4] |
866293ee | 1142 | */ |
734009e9 | 1143 | goto out_attach; |
c87e2837 | 1144 | } |
bd1dbcc6 TG |
1145 | |
1146 | /* | |
e60cbc5c TG |
1147 | * If TID is 0, then either the dying owner has not |
1148 | * yet executed exit_pi_state_list() or some waiter | |
1149 | * acquired the rtmutex in the pi state, but did not | |
1150 | * yet fixup the TID in user space. | |
1151 | * | |
1152 | * Take a ref on the state and return success. [6] | |
1153 | */ | |
1154 | if (!pid) | |
734009e9 | 1155 | goto out_attach; |
e60cbc5c TG |
1156 | } else { |
1157 | /* | |
1158 | * If the owner died bit is not set, then the pi_state | |
1159 | * must have an owner. [7] | |
bd1dbcc6 | 1160 | */ |
e60cbc5c | 1161 | if (!pi_state->owner) |
734009e9 | 1162 | goto out_einval; |
c87e2837 IM |
1163 | } |
1164 | ||
e60cbc5c TG |
1165 | /* |
1166 | * Bail out if user space manipulated the futex value. If pi | |
1167 | * state exists then the owner TID must be the same as the | |
1168 | * user space TID. [9/10] | |
1169 | */ | |
1170 | if (pid != task_pid_vnr(pi_state->owner)) | |
734009e9 PZ |
1171 | goto out_einval; |
1172 | ||
1173 | out_attach: | |
bf92cf3a | 1174 | get_pi_state(pi_state); |
734009e9 | 1175 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); |
e60cbc5c TG |
1176 | *ps = pi_state; |
1177 | return 0; | |
734009e9 PZ |
1178 | |
1179 | out_einval: | |
1180 | ret = -EINVAL; | |
1181 | goto out_error; | |
1182 | ||
1183 | out_eagain: | |
1184 | ret = -EAGAIN; | |
1185 | goto out_error; | |
1186 | ||
1187 | out_efault: | |
1188 | ret = -EFAULT; | |
1189 | goto out_error; | |
1190 | ||
1191 | out_error: | |
1192 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
1193 | return ret; | |
e60cbc5c TG |
1194 | } |
1195 | ||
3ef240ea TG |
1196 | /** |
1197 | * wait_for_owner_exiting - Block until the owner has exited | |
51bfb1d1 | 1198 | * @ret: owner's current futex lock status |
3ef240ea TG |
1199 | * @exiting: Pointer to the exiting task |
1200 | * | |
1201 | * Caller must hold a refcount on @exiting. | |
1202 | */ | |
1203 | static void wait_for_owner_exiting(int ret, struct task_struct *exiting) | |
1204 | { | |
1205 | if (ret != -EBUSY) { | |
1206 | WARN_ON_ONCE(exiting); | |
1207 | return; | |
1208 | } | |
1209 | ||
1210 | if (WARN_ON_ONCE(ret == -EBUSY && !exiting)) | |
1211 | return; | |
1212 | ||
1213 | mutex_lock(&exiting->futex_exit_mutex); | |
1214 | /* | |
1215 | * No point in doing state checking here. If the waiter got here | |
1216 | * while the task was in exec()->exec_futex_release() then it can | |
1217 | * have any FUTEX_STATE_* value when the waiter has acquired the | |
1218 | * mutex. OK, if running, EXITING or DEAD if it reached exit() | |
1219 | * already. Highly unlikely and not a problem. Just one more round | |
1220 | * through the futex maze. | |
1221 | */ | |
1222 | mutex_unlock(&exiting->futex_exit_mutex); | |
1223 | ||
1224 | put_task_struct(exiting); | |
1225 | } | |
1226 | ||
da791a66 TG |
1227 | static int handle_exit_race(u32 __user *uaddr, u32 uval, |
1228 | struct task_struct *tsk) | |
1229 | { | |
1230 | u32 uval2; | |
1231 | ||
1232 | /* | |
ac31c7ff TG |
1233 | * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the |
1234 | * caller that the alleged owner is busy. | |
da791a66 | 1235 | */ |
3d4775df | 1236 | if (tsk && tsk->futex_state != FUTEX_STATE_DEAD) |
ac31c7ff | 1237 | return -EBUSY; |
da791a66 TG |
1238 | |
1239 | /* | |
1240 | * Reread the user space value to handle the following situation: | |
1241 | * | |
1242 | * CPU0 CPU1 | |
1243 | * | |
1244 | * sys_exit() sys_futex() | |
1245 | * do_exit() futex_lock_pi() | |
1246 | * futex_lock_pi_atomic() | |
1247 | * exit_signals(tsk) No waiters: | |
1248 | * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID | |
1249 | * mm_release(tsk) Set waiter bit | |
1250 | * exit_robust_list(tsk) { *uaddr = 0x80000PID; | |
1251 | * Set owner died attach_to_pi_owner() { | |
1252 | * *uaddr = 0xC0000000; tsk = get_task(PID); | |
1253 | * } if (!tsk->flags & PF_EXITING) { | |
1254 | * ... attach(); | |
3d4775df TG |
1255 | * tsk->futex_state = } else { |
1256 | * FUTEX_STATE_DEAD; if (tsk->futex_state != | |
1257 | * FUTEX_STATE_DEAD) | |
da791a66 TG |
1258 | * return -EAGAIN; |
1259 | * return -ESRCH; <--- FAIL | |
1260 | * } | |
1261 | * | |
1262 | * Returning ESRCH unconditionally is wrong here because the | |
1263 | * user space value has been changed by the exiting task. | |
1264 | * | |
1265 | * The same logic applies to the case where the exiting task is | |
1266 | * already gone. | |
1267 | */ | |
1268 | if (get_futex_value_locked(&uval2, uaddr)) | |
1269 | return -EFAULT; | |
1270 | ||
1271 | /* If the user space value has changed, try again. */ | |
1272 | if (uval2 != uval) | |
1273 | return -EAGAIN; | |
1274 | ||
1275 | /* | |
1276 | * The exiting task did not have a robust list, the robust list was | |
1277 | * corrupted or the user space value in *uaddr is simply bogus. | |
1278 | * Give up and tell user space. | |
1279 | */ | |
1280 | return -ESRCH; | |
1281 | } | |
1282 | ||
04e1b2e5 TG |
1283 | /* |
1284 | * Lookup the task for the TID provided from user space and attach to | |
1285 | * it after doing proper sanity checks. | |
1286 | */ | |
da791a66 | 1287 | static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key, |
3ef240ea TG |
1288 | struct futex_pi_state **ps, |
1289 | struct task_struct **exiting) | |
e60cbc5c | 1290 | { |
e60cbc5c | 1291 | pid_t pid = uval & FUTEX_TID_MASK; |
04e1b2e5 TG |
1292 | struct futex_pi_state *pi_state; |
1293 | struct task_struct *p; | |
e60cbc5c | 1294 | |
c87e2837 | 1295 | /* |
e3f2ddea | 1296 | * We are the first waiter - try to look up the real owner and attach |
54a21788 | 1297 | * the new pi_state to it, but bail out when TID = 0 [1] |
da791a66 TG |
1298 | * |
1299 | * The !pid check is paranoid. None of the call sites should end up | |
1300 | * with pid == 0, but better safe than sorry. Let the caller retry | |
c87e2837 | 1301 | */ |
778e9a9c | 1302 | if (!pid) |
da791a66 | 1303 | return -EAGAIN; |
2ee08260 | 1304 | p = find_get_task_by_vpid(pid); |
7a0ea09a | 1305 | if (!p) |
da791a66 | 1306 | return handle_exit_race(uaddr, uval, NULL); |
778e9a9c | 1307 | |
a2129464 | 1308 | if (unlikely(p->flags & PF_KTHREAD)) { |
f0d71b3d TG |
1309 | put_task_struct(p); |
1310 | return -EPERM; | |
1311 | } | |
1312 | ||
778e9a9c | 1313 | /* |
3d4775df TG |
1314 | * We need to look at the task state to figure out, whether the |
1315 | * task is exiting. To protect against the change of the task state | |
1316 | * in futex_exit_release(), we do this protected by p->pi_lock: | |
778e9a9c | 1317 | */ |
1d615482 | 1318 | raw_spin_lock_irq(&p->pi_lock); |
3d4775df | 1319 | if (unlikely(p->futex_state != FUTEX_STATE_OK)) { |
778e9a9c | 1320 | /* |
3d4775df TG |
1321 | * The task is on the way out. When the futex state is |
1322 | * FUTEX_STATE_DEAD, we know that the task has finished | |
1323 | * the cleanup: | |
778e9a9c | 1324 | */ |
da791a66 | 1325 | int ret = handle_exit_race(uaddr, uval, p); |
778e9a9c | 1326 | |
1d615482 | 1327 | raw_spin_unlock_irq(&p->pi_lock); |
3ef240ea TG |
1328 | /* |
1329 | * If the owner task is between FUTEX_STATE_EXITING and | |
1330 | * FUTEX_STATE_DEAD then store the task pointer and keep | |
1331 | * the reference on the task struct. The calling code will | |
1332 | * drop all locks, wait for the task to reach | |
1333 | * FUTEX_STATE_DEAD and then drop the refcount. This is | |
1334 | * required to prevent a live lock when the current task | |
1335 | * preempted the exiting task between the two states. | |
1336 | */ | |
1337 | if (ret == -EBUSY) | |
1338 | *exiting = p; | |
1339 | else | |
1340 | put_task_struct(p); | |
778e9a9c AK |
1341 | return ret; |
1342 | } | |
c87e2837 | 1343 | |
54a21788 TG |
1344 | /* |
1345 | * No existing pi state. First waiter. [2] | |
734009e9 PZ |
1346 | * |
1347 | * This creates pi_state, we have hb->lock held, this means nothing can | |
1348 | * observe this state, wait_lock is irrelevant. | |
54a21788 | 1349 | */ |
c87e2837 IM |
1350 | pi_state = alloc_pi_state(); |
1351 | ||
1352 | /* | |
04e1b2e5 | 1353 | * Initialize the pi_mutex in locked state and make @p |
c87e2837 IM |
1354 | * the owner of it: |
1355 | */ | |
1356 | rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p); | |
1357 | ||
1358 | /* Store the key for possible exit cleanups: */ | |
d0aa7a70 | 1359 | pi_state->key = *key; |
c87e2837 | 1360 | |
627371d7 | 1361 | WARN_ON(!list_empty(&pi_state->list)); |
c87e2837 | 1362 | list_add(&pi_state->list, &p->pi_state_list); |
c74aef2d PZ |
1363 | /* |
1364 | * Assignment without holding pi_state->pi_mutex.wait_lock is safe | |
1365 | * because there is no concurrency as the object is not published yet. | |
1366 | */ | |
c87e2837 | 1367 | pi_state->owner = p; |
1d615482 | 1368 | raw_spin_unlock_irq(&p->pi_lock); |
c87e2837 IM |
1369 | |
1370 | put_task_struct(p); | |
1371 | ||
d0aa7a70 | 1372 | *ps = pi_state; |
c87e2837 IM |
1373 | |
1374 | return 0; | |
1375 | } | |
1376 | ||
734009e9 PZ |
1377 | static int lookup_pi_state(u32 __user *uaddr, u32 uval, |
1378 | struct futex_hash_bucket *hb, | |
3ef240ea TG |
1379 | union futex_key *key, struct futex_pi_state **ps, |
1380 | struct task_struct **exiting) | |
04e1b2e5 | 1381 | { |
499f5aca | 1382 | struct futex_q *top_waiter = futex_top_waiter(hb, key); |
04e1b2e5 TG |
1383 | |
1384 | /* | |
1385 | * If there is a waiter on that futex, validate it and | |
1386 | * attach to the pi_state when the validation succeeds. | |
1387 | */ | |
499f5aca | 1388 | if (top_waiter) |
734009e9 | 1389 | return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps); |
04e1b2e5 TG |
1390 | |
1391 | /* | |
1392 | * We are the first waiter - try to look up the owner based on | |
1393 | * @uval and attach to it. | |
1394 | */ | |
3ef240ea | 1395 | return attach_to_pi_owner(uaddr, uval, key, ps, exiting); |
04e1b2e5 TG |
1396 | } |
1397 | ||
af54d6a1 TG |
1398 | static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval) |
1399 | { | |
6b4f4bc9 | 1400 | int err; |
af54d6a1 TG |
1401 | u32 uninitialized_var(curval); |
1402 | ||
ab51fbab DB |
1403 | if (unlikely(should_fail_futex(true))) |
1404 | return -EFAULT; | |
1405 | ||
6b4f4bc9 WD |
1406 | err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval); |
1407 | if (unlikely(err)) | |
1408 | return err; | |
af54d6a1 | 1409 | |
734009e9 | 1410 | /* If user space value changed, let the caller retry */ |
af54d6a1 TG |
1411 | return curval != uval ? -EAGAIN : 0; |
1412 | } | |
1413 | ||
1a52084d | 1414 | /** |
d96ee56c | 1415 | * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex |
bab5bc9e DH |
1416 | * @uaddr: the pi futex user address |
1417 | * @hb: the pi futex hash bucket | |
1418 | * @key: the futex key associated with uaddr and hb | |
1419 | * @ps: the pi_state pointer where we store the result of the | |
1420 | * lookup | |
1421 | * @task: the task to perform the atomic lock work for. This will | |
1422 | * be "current" except in the case of requeue pi. | |
3ef240ea TG |
1423 | * @exiting: Pointer to store the task pointer of the owner task |
1424 | * which is in the middle of exiting | |
bab5bc9e | 1425 | * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0) |
1a52084d | 1426 | * |
6c23cbbd | 1427 | * Return: |
7b4ff1ad MCC |
1428 | * - 0 - ready to wait; |
1429 | * - 1 - acquired the lock; | |
1430 | * - <0 - error | |
1a52084d DH |
1431 | * |
1432 | * The hb->lock and futex_key refs shall be held by the caller. | |
3ef240ea TG |
1433 | * |
1434 | * @exiting is only set when the return value is -EBUSY. If so, this holds | |
1435 | * a refcount on the exiting task on return and the caller needs to drop it | |
1436 | * after waiting for the exit to complete. | |
1a52084d DH |
1437 | */ |
1438 | static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb, | |
1439 | union futex_key *key, | |
1440 | struct futex_pi_state **ps, | |
3ef240ea TG |
1441 | struct task_struct *task, |
1442 | struct task_struct **exiting, | |
1443 | int set_waiters) | |
1a52084d | 1444 | { |
af54d6a1 | 1445 | u32 uval, newval, vpid = task_pid_vnr(task); |
499f5aca | 1446 | struct futex_q *top_waiter; |
af54d6a1 | 1447 | int ret; |
1a52084d DH |
1448 | |
1449 | /* | |
af54d6a1 TG |
1450 | * Read the user space value first so we can validate a few |
1451 | * things before proceeding further. | |
1a52084d | 1452 | */ |
af54d6a1 | 1453 | if (get_futex_value_locked(&uval, uaddr)) |
1a52084d DH |
1454 | return -EFAULT; |
1455 | ||
ab51fbab DB |
1456 | if (unlikely(should_fail_futex(true))) |
1457 | return -EFAULT; | |
1458 | ||
1a52084d DH |
1459 | /* |
1460 | * Detect deadlocks. | |
1461 | */ | |
af54d6a1 | 1462 | if ((unlikely((uval & FUTEX_TID_MASK) == vpid))) |
1a52084d DH |
1463 | return -EDEADLK; |
1464 | ||
ab51fbab DB |
1465 | if ((unlikely(should_fail_futex(true)))) |
1466 | return -EDEADLK; | |
1467 | ||
1a52084d | 1468 | /* |
af54d6a1 TG |
1469 | * Lookup existing state first. If it exists, try to attach to |
1470 | * its pi_state. | |
1a52084d | 1471 | */ |
499f5aca PZ |
1472 | top_waiter = futex_top_waiter(hb, key); |
1473 | if (top_waiter) | |
734009e9 | 1474 | return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps); |
1a52084d DH |
1475 | |
1476 | /* | |
af54d6a1 TG |
1477 | * No waiter and user TID is 0. We are here because the |
1478 | * waiters or the owner died bit is set or called from | |
1479 | * requeue_cmp_pi or for whatever reason something took the | |
1480 | * syscall. | |
1a52084d | 1481 | */ |
af54d6a1 | 1482 | if (!(uval & FUTEX_TID_MASK)) { |
59fa6245 | 1483 | /* |
af54d6a1 TG |
1484 | * We take over the futex. No other waiters and the user space |
1485 | * TID is 0. We preserve the owner died bit. | |
59fa6245 | 1486 | */ |
af54d6a1 TG |
1487 | newval = uval & FUTEX_OWNER_DIED; |
1488 | newval |= vpid; | |
1a52084d | 1489 | |
af54d6a1 TG |
1490 | /* The futex requeue_pi code can enforce the waiters bit */ |
1491 | if (set_waiters) | |
1492 | newval |= FUTEX_WAITERS; | |
1493 | ||
1494 | ret = lock_pi_update_atomic(uaddr, uval, newval); | |
1495 | /* If the take over worked, return 1 */ | |
1496 | return ret < 0 ? ret : 1; | |
1497 | } | |
1a52084d DH |
1498 | |
1499 | /* | |
af54d6a1 TG |
1500 | * First waiter. Set the waiters bit before attaching ourself to |
1501 | * the owner. If owner tries to unlock, it will be forced into | |
1502 | * the kernel and blocked on hb->lock. | |
1a52084d | 1503 | */ |
af54d6a1 TG |
1504 | newval = uval | FUTEX_WAITERS; |
1505 | ret = lock_pi_update_atomic(uaddr, uval, newval); | |
1506 | if (ret) | |
1507 | return ret; | |
1a52084d | 1508 | /* |
af54d6a1 TG |
1509 | * If the update of the user space value succeeded, we try to |
1510 | * attach to the owner. If that fails, no harm done, we only | |
1511 | * set the FUTEX_WAITERS bit in the user space variable. | |
1a52084d | 1512 | */ |
3ef240ea | 1513 | return attach_to_pi_owner(uaddr, newval, key, ps, exiting); |
1a52084d DH |
1514 | } |
1515 | ||
2e12978a LJ |
1516 | /** |
1517 | * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket | |
1518 | * @q: The futex_q to unqueue | |
1519 | * | |
1520 | * The q->lock_ptr must not be NULL and must be held by the caller. | |
1521 | */ | |
1522 | static void __unqueue_futex(struct futex_q *q) | |
1523 | { | |
1524 | struct futex_hash_bucket *hb; | |
1525 | ||
4de1a293 | 1526 | if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list))) |
2e12978a | 1527 | return; |
4de1a293 | 1528 | lockdep_assert_held(q->lock_ptr); |
2e12978a LJ |
1529 | |
1530 | hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock); | |
1531 | plist_del(&q->list, &hb->chain); | |
11d4616b | 1532 | hb_waiters_dec(hb); |
2e12978a LJ |
1533 | } |
1534 | ||
1da177e4 LT |
1535 | /* |
1536 | * The hash bucket lock must be held when this is called. | |
1d0dcb3a DB |
1537 | * Afterwards, the futex_q must not be accessed. Callers |
1538 | * must ensure to later call wake_up_q() for the actual | |
1539 | * wakeups to occur. | |
1da177e4 | 1540 | */ |
1d0dcb3a | 1541 | static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q) |
1da177e4 | 1542 | { |
f1a11e05 TG |
1543 | struct task_struct *p = q->task; |
1544 | ||
aa10990e DH |
1545 | if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n")) |
1546 | return; | |
1547 | ||
b061c38b | 1548 | get_task_struct(p); |
2e12978a | 1549 | __unqueue_futex(q); |
1da177e4 | 1550 | /* |
38fcd06e DHV |
1551 | * The waiting task can free the futex_q as soon as q->lock_ptr = NULL |
1552 | * is written, without taking any locks. This is possible in the event | |
1553 | * of a spurious wakeup, for example. A memory barrier is required here | |
1554 | * to prevent the following store to lock_ptr from getting ahead of the | |
1555 | * plist_del in __unqueue_futex(). | |
1da177e4 | 1556 | */ |
1b367ece | 1557 | smp_store_release(&q->lock_ptr, NULL); |
b061c38b PZ |
1558 | |
1559 | /* | |
1560 | * Queue the task for later wakeup for after we've released | |
75145904 | 1561 | * the hb->lock. |
b061c38b | 1562 | */ |
07879c6a | 1563 | wake_q_add_safe(wake_q, p); |
1da177e4 LT |
1564 | } |
1565 | ||
16ffa12d PZ |
1566 | /* |
1567 | * Caller must hold a reference on @pi_state. | |
1568 | */ | |
1569 | static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state) | |
c87e2837 | 1570 | { |
7cfdaf38 | 1571 | u32 uninitialized_var(curval), newval; |
16ffa12d | 1572 | struct task_struct *new_owner; |
aa2bfe55 | 1573 | bool postunlock = false; |
194a6b5b | 1574 | DEFINE_WAKE_Q(wake_q); |
13fbca4c | 1575 | int ret = 0; |
c87e2837 | 1576 | |
c87e2837 | 1577 | new_owner = rt_mutex_next_owner(&pi_state->pi_mutex); |
bebe5b51 | 1578 | if (WARN_ON_ONCE(!new_owner)) { |
16ffa12d | 1579 | /* |
bebe5b51 | 1580 | * As per the comment in futex_unlock_pi() this should not happen. |
16ffa12d PZ |
1581 | * |
1582 | * When this happens, give up our locks and try again, giving | |
1583 | * the futex_lock_pi() instance time to complete, either by | |
1584 | * waiting on the rtmutex or removing itself from the futex | |
1585 | * queue. | |
1586 | */ | |
1587 | ret = -EAGAIN; | |
1588 | goto out_unlock; | |
73d786bd | 1589 | } |
c87e2837 IM |
1590 | |
1591 | /* | |
16ffa12d PZ |
1592 | * We pass it to the next owner. The WAITERS bit is always kept |
1593 | * enabled while there is PI state around. We cleanup the owner | |
1594 | * died bit, because we are the owner. | |
c87e2837 | 1595 | */ |
13fbca4c | 1596 | newval = FUTEX_WAITERS | task_pid_vnr(new_owner); |
e3f2ddea | 1597 | |
ab51fbab DB |
1598 | if (unlikely(should_fail_futex(true))) |
1599 | ret = -EFAULT; | |
1600 | ||
6b4f4bc9 WD |
1601 | ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval); |
1602 | if (!ret && (curval != uval)) { | |
89e9e66b SAS |
1603 | /* |
1604 | * If a unconditional UNLOCK_PI operation (user space did not | |
1605 | * try the TID->0 transition) raced with a waiter setting the | |
1606 | * FUTEX_WAITERS flag between get_user() and locking the hash | |
1607 | * bucket lock, retry the operation. | |
1608 | */ | |
1609 | if ((FUTEX_TID_MASK & curval) == uval) | |
1610 | ret = -EAGAIN; | |
1611 | else | |
1612 | ret = -EINVAL; | |
1613 | } | |
734009e9 | 1614 | |
16ffa12d PZ |
1615 | if (ret) |
1616 | goto out_unlock; | |
c87e2837 | 1617 | |
94ffac5d PZ |
1618 | /* |
1619 | * This is a point of no return; once we modify the uval there is no | |
1620 | * going back and subsequent operations must not fail. | |
1621 | */ | |
1622 | ||
b4abf910 | 1623 | raw_spin_lock(&pi_state->owner->pi_lock); |
627371d7 IM |
1624 | WARN_ON(list_empty(&pi_state->list)); |
1625 | list_del_init(&pi_state->list); | |
b4abf910 | 1626 | raw_spin_unlock(&pi_state->owner->pi_lock); |
627371d7 | 1627 | |
b4abf910 | 1628 | raw_spin_lock(&new_owner->pi_lock); |
627371d7 | 1629 | WARN_ON(!list_empty(&pi_state->list)); |
c87e2837 IM |
1630 | list_add(&pi_state->list, &new_owner->pi_state_list); |
1631 | pi_state->owner = new_owner; | |
b4abf910 | 1632 | raw_spin_unlock(&new_owner->pi_lock); |
627371d7 | 1633 | |
aa2bfe55 | 1634 | postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q); |
5293c2ef | 1635 | |
16ffa12d | 1636 | out_unlock: |
5293c2ef | 1637 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); |
5293c2ef | 1638 | |
aa2bfe55 PZ |
1639 | if (postunlock) |
1640 | rt_mutex_postunlock(&wake_q); | |
c87e2837 | 1641 | |
16ffa12d | 1642 | return ret; |
c87e2837 IM |
1643 | } |
1644 | ||
8b8f319f IM |
1645 | /* |
1646 | * Express the locking dependencies for lockdep: | |
1647 | */ | |
1648 | static inline void | |
1649 | double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2) | |
1650 | { | |
1651 | if (hb1 <= hb2) { | |
1652 | spin_lock(&hb1->lock); | |
1653 | if (hb1 < hb2) | |
1654 | spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING); | |
1655 | } else { /* hb1 > hb2 */ | |
1656 | spin_lock(&hb2->lock); | |
1657 | spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING); | |
1658 | } | |
1659 | } | |
1660 | ||
5eb3dc62 DH |
1661 | static inline void |
1662 | double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2) | |
1663 | { | |
f061d351 | 1664 | spin_unlock(&hb1->lock); |
88f502fe IM |
1665 | if (hb1 != hb2) |
1666 | spin_unlock(&hb2->lock); | |
5eb3dc62 DH |
1667 | } |
1668 | ||
1da177e4 | 1669 | /* |
b2d0994b | 1670 | * Wake up waiters matching bitset queued on this futex (uaddr). |
1da177e4 | 1671 | */ |
b41277dc DH |
1672 | static int |
1673 | futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset) | |
1da177e4 | 1674 | { |
e2970f2f | 1675 | struct futex_hash_bucket *hb; |
1da177e4 | 1676 | struct futex_q *this, *next; |
38d47c1b | 1677 | union futex_key key = FUTEX_KEY_INIT; |
1da177e4 | 1678 | int ret; |
194a6b5b | 1679 | DEFINE_WAKE_Q(wake_q); |
1da177e4 | 1680 | |
cd689985 TG |
1681 | if (!bitset) |
1682 | return -EINVAL; | |
1683 | ||
96d4f267 | 1684 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ); |
1da177e4 LT |
1685 | if (unlikely(ret != 0)) |
1686 | goto out; | |
1687 | ||
e2970f2f | 1688 | hb = hash_futex(&key); |
b0c29f79 DB |
1689 | |
1690 | /* Make sure we really have tasks to wakeup */ | |
1691 | if (!hb_waiters_pending(hb)) | |
1692 | goto out_put_key; | |
1693 | ||
e2970f2f | 1694 | spin_lock(&hb->lock); |
1da177e4 | 1695 | |
0d00c7b2 | 1696 | plist_for_each_entry_safe(this, next, &hb->chain, list) { |
1da177e4 | 1697 | if (match_futex (&this->key, &key)) { |
52400ba9 | 1698 | if (this->pi_state || this->rt_waiter) { |
ed6f7b10 IM |
1699 | ret = -EINVAL; |
1700 | break; | |
1701 | } | |
cd689985 TG |
1702 | |
1703 | /* Check if one of the bits is set in both bitsets */ | |
1704 | if (!(this->bitset & bitset)) | |
1705 | continue; | |
1706 | ||
1d0dcb3a | 1707 | mark_wake_futex(&wake_q, this); |
1da177e4 LT |
1708 | if (++ret >= nr_wake) |
1709 | break; | |
1710 | } | |
1711 | } | |
1712 | ||
e2970f2f | 1713 | spin_unlock(&hb->lock); |
1d0dcb3a | 1714 | wake_up_q(&wake_q); |
b0c29f79 | 1715 | out_put_key: |
ae791a2d | 1716 | put_futex_key(&key); |
42d35d48 | 1717 | out: |
1da177e4 LT |
1718 | return ret; |
1719 | } | |
1720 | ||
30d6e0a4 JS |
1721 | static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr) |
1722 | { | |
1723 | unsigned int op = (encoded_op & 0x70000000) >> 28; | |
1724 | unsigned int cmp = (encoded_op & 0x0f000000) >> 24; | |
d70ef228 JS |
1725 | int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11); |
1726 | int cmparg = sign_extend32(encoded_op & 0x00000fff, 11); | |
30d6e0a4 JS |
1727 | int oldval, ret; |
1728 | ||
1729 | if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) { | |
e78c38f6 JS |
1730 | if (oparg < 0 || oparg > 31) { |
1731 | char comm[sizeof(current->comm)]; | |
1732 | /* | |
1733 | * kill this print and return -EINVAL when userspace | |
1734 | * is sane again | |
1735 | */ | |
1736 | pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n", | |
1737 | get_task_comm(comm, current), oparg); | |
1738 | oparg &= 31; | |
1739 | } | |
30d6e0a4 JS |
1740 | oparg = 1 << oparg; |
1741 | } | |
1742 | ||
96d4f267 | 1743 | if (!access_ok(uaddr, sizeof(u32))) |
30d6e0a4 JS |
1744 | return -EFAULT; |
1745 | ||
1746 | ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr); | |
1747 | if (ret) | |
1748 | return ret; | |
1749 | ||
1750 | switch (cmp) { | |
1751 | case FUTEX_OP_CMP_EQ: | |
1752 | return oldval == cmparg; | |
1753 | case FUTEX_OP_CMP_NE: | |
1754 | return oldval != cmparg; | |
1755 | case FUTEX_OP_CMP_LT: | |
1756 | return oldval < cmparg; | |
1757 | case FUTEX_OP_CMP_GE: | |
1758 | return oldval >= cmparg; | |
1759 | case FUTEX_OP_CMP_LE: | |
1760 | return oldval <= cmparg; | |
1761 | case FUTEX_OP_CMP_GT: | |
1762 | return oldval > cmparg; | |
1763 | default: | |
1764 | return -ENOSYS; | |
1765 | } | |
1766 | } | |
1767 | ||
4732efbe JJ |
1768 | /* |
1769 | * Wake up all waiters hashed on the physical page that is mapped | |
1770 | * to this virtual address: | |
1771 | */ | |
e2970f2f | 1772 | static int |
b41277dc | 1773 | futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2, |
e2970f2f | 1774 | int nr_wake, int nr_wake2, int op) |
4732efbe | 1775 | { |
38d47c1b | 1776 | union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT; |
e2970f2f | 1777 | struct futex_hash_bucket *hb1, *hb2; |
4732efbe | 1778 | struct futex_q *this, *next; |
e4dc5b7a | 1779 | int ret, op_ret; |
194a6b5b | 1780 | DEFINE_WAKE_Q(wake_q); |
4732efbe | 1781 | |
e4dc5b7a | 1782 | retry: |
96d4f267 | 1783 | ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ); |
4732efbe JJ |
1784 | if (unlikely(ret != 0)) |
1785 | goto out; | |
96d4f267 | 1786 | ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE); |
4732efbe | 1787 | if (unlikely(ret != 0)) |
42d35d48 | 1788 | goto out_put_key1; |
4732efbe | 1789 | |
e2970f2f IM |
1790 | hb1 = hash_futex(&key1); |
1791 | hb2 = hash_futex(&key2); | |
4732efbe | 1792 | |
e4dc5b7a | 1793 | retry_private: |
eaaea803 | 1794 | double_lock_hb(hb1, hb2); |
e2970f2f | 1795 | op_ret = futex_atomic_op_inuser(op, uaddr2); |
4732efbe | 1796 | if (unlikely(op_ret < 0)) { |
5eb3dc62 | 1797 | double_unlock_hb(hb1, hb2); |
4732efbe | 1798 | |
6b4f4bc9 WD |
1799 | if (!IS_ENABLED(CONFIG_MMU) || |
1800 | unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) { | |
1801 | /* | |
1802 | * we don't get EFAULT from MMU faults if we don't have | |
1803 | * an MMU, but we might get them from range checking | |
1804 | */ | |
796f8d9b | 1805 | ret = op_ret; |
42d35d48 | 1806 | goto out_put_keys; |
796f8d9b DG |
1807 | } |
1808 | ||
6b4f4bc9 WD |
1809 | if (op_ret == -EFAULT) { |
1810 | ret = fault_in_user_writeable(uaddr2); | |
1811 | if (ret) | |
1812 | goto out_put_keys; | |
1813 | } | |
4732efbe | 1814 | |
6b4f4bc9 WD |
1815 | if (!(flags & FLAGS_SHARED)) { |
1816 | cond_resched(); | |
e4dc5b7a | 1817 | goto retry_private; |
6b4f4bc9 | 1818 | } |
e4dc5b7a | 1819 | |
ae791a2d TG |
1820 | put_futex_key(&key2); |
1821 | put_futex_key(&key1); | |
6b4f4bc9 | 1822 | cond_resched(); |
e4dc5b7a | 1823 | goto retry; |
4732efbe JJ |
1824 | } |
1825 | ||
0d00c7b2 | 1826 | plist_for_each_entry_safe(this, next, &hb1->chain, list) { |
4732efbe | 1827 | if (match_futex (&this->key, &key1)) { |
aa10990e DH |
1828 | if (this->pi_state || this->rt_waiter) { |
1829 | ret = -EINVAL; | |
1830 | goto out_unlock; | |
1831 | } | |
1d0dcb3a | 1832 | mark_wake_futex(&wake_q, this); |
4732efbe JJ |
1833 | if (++ret >= nr_wake) |
1834 | break; | |
1835 | } | |
1836 | } | |
1837 | ||
1838 | if (op_ret > 0) { | |
4732efbe | 1839 | op_ret = 0; |
0d00c7b2 | 1840 | plist_for_each_entry_safe(this, next, &hb2->chain, list) { |
4732efbe | 1841 | if (match_futex (&this->key, &key2)) { |
aa10990e DH |
1842 | if (this->pi_state || this->rt_waiter) { |
1843 | ret = -EINVAL; | |
1844 | goto out_unlock; | |
1845 | } | |
1d0dcb3a | 1846 | mark_wake_futex(&wake_q, this); |
4732efbe JJ |
1847 | if (++op_ret >= nr_wake2) |
1848 | break; | |
1849 | } | |
1850 | } | |
1851 | ret += op_ret; | |
1852 | } | |
1853 | ||
aa10990e | 1854 | out_unlock: |
5eb3dc62 | 1855 | double_unlock_hb(hb1, hb2); |
1d0dcb3a | 1856 | wake_up_q(&wake_q); |
42d35d48 | 1857 | out_put_keys: |
ae791a2d | 1858 | put_futex_key(&key2); |
42d35d48 | 1859 | out_put_key1: |
ae791a2d | 1860 | put_futex_key(&key1); |
42d35d48 | 1861 | out: |
4732efbe JJ |
1862 | return ret; |
1863 | } | |
1864 | ||
9121e478 DH |
1865 | /** |
1866 | * requeue_futex() - Requeue a futex_q from one hb to another | |
1867 | * @q: the futex_q to requeue | |
1868 | * @hb1: the source hash_bucket | |
1869 | * @hb2: the target hash_bucket | |
1870 | * @key2: the new key for the requeued futex_q | |
1871 | */ | |
1872 | static inline | |
1873 | void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1, | |
1874 | struct futex_hash_bucket *hb2, union futex_key *key2) | |
1875 | { | |
1876 | ||
1877 | /* | |
1878 | * If key1 and key2 hash to the same bucket, no need to | |
1879 | * requeue. | |
1880 | */ | |
1881 | if (likely(&hb1->chain != &hb2->chain)) { | |
1882 | plist_del(&q->list, &hb1->chain); | |
11d4616b | 1883 | hb_waiters_dec(hb1); |
11d4616b | 1884 | hb_waiters_inc(hb2); |
fe1bce9e | 1885 | plist_add(&q->list, &hb2->chain); |
9121e478 | 1886 | q->lock_ptr = &hb2->lock; |
9121e478 DH |
1887 | } |
1888 | get_futex_key_refs(key2); | |
1889 | q->key = *key2; | |
1890 | } | |
1891 | ||
52400ba9 DH |
1892 | /** |
1893 | * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue | |
d96ee56c DH |
1894 | * @q: the futex_q |
1895 | * @key: the key of the requeue target futex | |
1896 | * @hb: the hash_bucket of the requeue target futex | |
52400ba9 DH |
1897 | * |
1898 | * During futex_requeue, with requeue_pi=1, it is possible to acquire the | |
1899 | * target futex if it is uncontended or via a lock steal. Set the futex_q key | |
1900 | * to the requeue target futex so the waiter can detect the wakeup on the right | |
1901 | * futex, but remove it from the hb and NULL the rt_waiter so it can detect | |
beda2c7e DH |
1902 | * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock |
1903 | * to protect access to the pi_state to fixup the owner later. Must be called | |
1904 | * with both q->lock_ptr and hb->lock held. | |
52400ba9 DH |
1905 | */ |
1906 | static inline | |
beda2c7e DH |
1907 | void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key, |
1908 | struct futex_hash_bucket *hb) | |
52400ba9 | 1909 | { |
52400ba9 DH |
1910 | get_futex_key_refs(key); |
1911 | q->key = *key; | |
1912 | ||
2e12978a | 1913 | __unqueue_futex(q); |
52400ba9 DH |
1914 | |
1915 | WARN_ON(!q->rt_waiter); | |
1916 | q->rt_waiter = NULL; | |
1917 | ||
beda2c7e | 1918 | q->lock_ptr = &hb->lock; |
beda2c7e | 1919 | |
f1a11e05 | 1920 | wake_up_state(q->task, TASK_NORMAL); |
52400ba9 DH |
1921 | } |
1922 | ||
1923 | /** | |
1924 | * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter | |
bab5bc9e DH |
1925 | * @pifutex: the user address of the to futex |
1926 | * @hb1: the from futex hash bucket, must be locked by the caller | |
1927 | * @hb2: the to futex hash bucket, must be locked by the caller | |
1928 | * @key1: the from futex key | |
1929 | * @key2: the to futex key | |
1930 | * @ps: address to store the pi_state pointer | |
3ef240ea TG |
1931 | * @exiting: Pointer to store the task pointer of the owner task |
1932 | * which is in the middle of exiting | |
bab5bc9e | 1933 | * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0) |
52400ba9 DH |
1934 | * |
1935 | * Try and get the lock on behalf of the top waiter if we can do it atomically. | |
bab5bc9e DH |
1936 | * Wake the top waiter if we succeed. If the caller specified set_waiters, |
1937 | * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit. | |
1938 | * hb1 and hb2 must be held by the caller. | |
52400ba9 | 1939 | * |
3ef240ea TG |
1940 | * @exiting is only set when the return value is -EBUSY. If so, this holds |
1941 | * a refcount on the exiting task on return and the caller needs to drop it | |
1942 | * after waiting for the exit to complete. | |
1943 | * | |
6c23cbbd | 1944 | * Return: |
7b4ff1ad MCC |
1945 | * - 0 - failed to acquire the lock atomically; |
1946 | * - >0 - acquired the lock, return value is vpid of the top_waiter | |
1947 | * - <0 - error | |
52400ba9 | 1948 | */ |
3ef240ea TG |
1949 | static int |
1950 | futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1, | |
1951 | struct futex_hash_bucket *hb2, union futex_key *key1, | |
1952 | union futex_key *key2, struct futex_pi_state **ps, | |
1953 | struct task_struct **exiting, int set_waiters) | |
52400ba9 | 1954 | { |
bab5bc9e | 1955 | struct futex_q *top_waiter = NULL; |
52400ba9 | 1956 | u32 curval; |
866293ee | 1957 | int ret, vpid; |
52400ba9 DH |
1958 | |
1959 | if (get_futex_value_locked(&curval, pifutex)) | |
1960 | return -EFAULT; | |
1961 | ||
ab51fbab DB |
1962 | if (unlikely(should_fail_futex(true))) |
1963 | return -EFAULT; | |
1964 | ||
bab5bc9e DH |
1965 | /* |
1966 | * Find the top_waiter and determine if there are additional waiters. | |
1967 | * If the caller intends to requeue more than 1 waiter to pifutex, | |
1968 | * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now, | |
1969 | * as we have means to handle the possible fault. If not, don't set | |
1970 | * the bit unecessarily as it will force the subsequent unlock to enter | |
1971 | * the kernel. | |
1972 | */ | |
52400ba9 DH |
1973 | top_waiter = futex_top_waiter(hb1, key1); |
1974 | ||
1975 | /* There are no waiters, nothing for us to do. */ | |
1976 | if (!top_waiter) | |
1977 | return 0; | |
1978 | ||
84bc4af5 DH |
1979 | /* Ensure we requeue to the expected futex. */ |
1980 | if (!match_futex(top_waiter->requeue_pi_key, key2)) | |
1981 | return -EINVAL; | |
1982 | ||
52400ba9 | 1983 | /* |
bab5bc9e DH |
1984 | * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in |
1985 | * the contended case or if set_waiters is 1. The pi_state is returned | |
1986 | * in ps in contended cases. | |
52400ba9 | 1987 | */ |
866293ee | 1988 | vpid = task_pid_vnr(top_waiter->task); |
bab5bc9e | 1989 | ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task, |
3ef240ea | 1990 | exiting, set_waiters); |
866293ee | 1991 | if (ret == 1) { |
beda2c7e | 1992 | requeue_pi_wake_futex(top_waiter, key2, hb2); |
866293ee TG |
1993 | return vpid; |
1994 | } | |
52400ba9 DH |
1995 | return ret; |
1996 | } | |
1997 | ||
1998 | /** | |
1999 | * futex_requeue() - Requeue waiters from uaddr1 to uaddr2 | |
fb62db2b | 2000 | * @uaddr1: source futex user address |
b41277dc | 2001 | * @flags: futex flags (FLAGS_SHARED, etc.) |
fb62db2b RD |
2002 | * @uaddr2: target futex user address |
2003 | * @nr_wake: number of waiters to wake (must be 1 for requeue_pi) | |
2004 | * @nr_requeue: number of waiters to requeue (0-INT_MAX) | |
2005 | * @cmpval: @uaddr1 expected value (or %NULL) | |
2006 | * @requeue_pi: if we are attempting to requeue from a non-pi futex to a | |
b41277dc | 2007 | * pi futex (pi to pi requeue is not supported) |
52400ba9 DH |
2008 | * |
2009 | * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire | |
2010 | * uaddr2 atomically on behalf of the top waiter. | |
2011 | * | |
6c23cbbd | 2012 | * Return: |
7b4ff1ad MCC |
2013 | * - >=0 - on success, the number of tasks requeued or woken; |
2014 | * - <0 - on error | |
1da177e4 | 2015 | */ |
b41277dc DH |
2016 | static int futex_requeue(u32 __user *uaddr1, unsigned int flags, |
2017 | u32 __user *uaddr2, int nr_wake, int nr_requeue, | |
2018 | u32 *cmpval, int requeue_pi) | |
1da177e4 | 2019 | { |
38d47c1b | 2020 | union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT; |
52400ba9 DH |
2021 | int drop_count = 0, task_count = 0, ret; |
2022 | struct futex_pi_state *pi_state = NULL; | |
e2970f2f | 2023 | struct futex_hash_bucket *hb1, *hb2; |
1da177e4 | 2024 | struct futex_q *this, *next; |
194a6b5b | 2025 | DEFINE_WAKE_Q(wake_q); |
52400ba9 | 2026 | |
fbe0e839 LJ |
2027 | if (nr_wake < 0 || nr_requeue < 0) |
2028 | return -EINVAL; | |
2029 | ||
bc2eecd7 NP |
2030 | /* |
2031 | * When PI not supported: return -ENOSYS if requeue_pi is true, | |
2032 | * consequently the compiler knows requeue_pi is always false past | |
2033 | * this point which will optimize away all the conditional code | |
2034 | * further down. | |
2035 | */ | |
2036 | if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi) | |
2037 | return -ENOSYS; | |
2038 | ||
52400ba9 | 2039 | if (requeue_pi) { |
e9c243a5 TG |
2040 | /* |
2041 | * Requeue PI only works on two distinct uaddrs. This | |
2042 | * check is only valid for private futexes. See below. | |
2043 | */ | |
2044 | if (uaddr1 == uaddr2) | |
2045 | return -EINVAL; | |
2046 | ||
52400ba9 DH |
2047 | /* |
2048 | * requeue_pi requires a pi_state, try to allocate it now | |
2049 | * without any locks in case it fails. | |
2050 | */ | |
2051 | if (refill_pi_state_cache()) | |
2052 | return -ENOMEM; | |
2053 | /* | |
2054 | * requeue_pi must wake as many tasks as it can, up to nr_wake | |
2055 | * + nr_requeue, since it acquires the rt_mutex prior to | |
2056 | * returning to userspace, so as to not leave the rt_mutex with | |
2057 | * waiters and no owner. However, second and third wake-ups | |
2058 | * cannot be predicted as they involve race conditions with the | |
2059 | * first wake and a fault while looking up the pi_state. Both | |
2060 | * pthread_cond_signal() and pthread_cond_broadcast() should | |
2061 | * use nr_wake=1. | |
2062 | */ | |
2063 | if (nr_wake != 1) | |
2064 | return -EINVAL; | |
2065 | } | |
1da177e4 | 2066 | |
42d35d48 | 2067 | retry: |
96d4f267 | 2068 | ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ); |
1da177e4 LT |
2069 | if (unlikely(ret != 0)) |
2070 | goto out; | |
9ea71503 | 2071 | ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, |
96d4f267 | 2072 | requeue_pi ? FUTEX_WRITE : FUTEX_READ); |
1da177e4 | 2073 | if (unlikely(ret != 0)) |
42d35d48 | 2074 | goto out_put_key1; |
1da177e4 | 2075 | |
e9c243a5 TG |
2076 | /* |
2077 | * The check above which compares uaddrs is not sufficient for | |
2078 | * shared futexes. We need to compare the keys: | |
2079 | */ | |
2080 | if (requeue_pi && match_futex(&key1, &key2)) { | |
2081 | ret = -EINVAL; | |
2082 | goto out_put_keys; | |
2083 | } | |
2084 | ||
e2970f2f IM |
2085 | hb1 = hash_futex(&key1); |
2086 | hb2 = hash_futex(&key2); | |
1da177e4 | 2087 | |
e4dc5b7a | 2088 | retry_private: |
69cd9eba | 2089 | hb_waiters_inc(hb2); |
8b8f319f | 2090 | double_lock_hb(hb1, hb2); |
1da177e4 | 2091 | |
e2970f2f IM |
2092 | if (likely(cmpval != NULL)) { |
2093 | u32 curval; | |
1da177e4 | 2094 | |
e2970f2f | 2095 | ret = get_futex_value_locked(&curval, uaddr1); |
1da177e4 LT |
2096 | |
2097 | if (unlikely(ret)) { | |
5eb3dc62 | 2098 | double_unlock_hb(hb1, hb2); |
69cd9eba | 2099 | hb_waiters_dec(hb2); |
1da177e4 | 2100 | |
e2970f2f | 2101 | ret = get_user(curval, uaddr1); |
e4dc5b7a DH |
2102 | if (ret) |
2103 | goto out_put_keys; | |
1da177e4 | 2104 | |
b41277dc | 2105 | if (!(flags & FLAGS_SHARED)) |
e4dc5b7a | 2106 | goto retry_private; |
1da177e4 | 2107 | |
ae791a2d TG |
2108 | put_futex_key(&key2); |
2109 | put_futex_key(&key1); | |
e4dc5b7a | 2110 | goto retry; |
1da177e4 | 2111 | } |
e2970f2f | 2112 | if (curval != *cmpval) { |
1da177e4 LT |
2113 | ret = -EAGAIN; |
2114 | goto out_unlock; | |
2115 | } | |
2116 | } | |
2117 | ||
52400ba9 | 2118 | if (requeue_pi && (task_count - nr_wake < nr_requeue)) { |
3ef240ea TG |
2119 | struct task_struct *exiting = NULL; |
2120 | ||
bab5bc9e DH |
2121 | /* |
2122 | * Attempt to acquire uaddr2 and wake the top waiter. If we | |
2123 | * intend to requeue waiters, force setting the FUTEX_WAITERS | |
2124 | * bit. We force this here where we are able to easily handle | |
2125 | * faults rather in the requeue loop below. | |
2126 | */ | |
52400ba9 | 2127 | ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1, |
3ef240ea TG |
2128 | &key2, &pi_state, |
2129 | &exiting, nr_requeue); | |
52400ba9 DH |
2130 | |
2131 | /* | |
2132 | * At this point the top_waiter has either taken uaddr2 or is | |
2133 | * waiting on it. If the former, then the pi_state will not | |
2134 | * exist yet, look it up one more time to ensure we have a | |
866293ee TG |
2135 | * reference to it. If the lock was taken, ret contains the |
2136 | * vpid of the top waiter task. | |
ecb38b78 TG |
2137 | * If the lock was not taken, we have pi_state and an initial |
2138 | * refcount on it. In case of an error we have nothing. | |
52400ba9 | 2139 | */ |
866293ee | 2140 | if (ret > 0) { |
52400ba9 | 2141 | WARN_ON(pi_state); |
89061d3d | 2142 | drop_count++; |
52400ba9 | 2143 | task_count++; |
866293ee | 2144 | /* |
ecb38b78 TG |
2145 | * If we acquired the lock, then the user space value |
2146 | * of uaddr2 should be vpid. It cannot be changed by | |
2147 | * the top waiter as it is blocked on hb2 lock if it | |
2148 | * tries to do so. If something fiddled with it behind | |
2149 | * our back the pi state lookup might unearth it. So | |
2150 | * we rather use the known value than rereading and | |
2151 | * handing potential crap to lookup_pi_state. | |
2152 | * | |
2153 | * If that call succeeds then we have pi_state and an | |
2154 | * initial refcount on it. | |
866293ee | 2155 | */ |
3ef240ea TG |
2156 | ret = lookup_pi_state(uaddr2, ret, hb2, &key2, |
2157 | &pi_state, &exiting); | |
52400ba9 DH |
2158 | } |
2159 | ||
2160 | switch (ret) { | |
2161 | case 0: | |
ecb38b78 | 2162 | /* We hold a reference on the pi state. */ |
52400ba9 | 2163 | break; |
4959f2de TG |
2164 | |
2165 | /* If the above failed, then pi_state is NULL */ | |
52400ba9 DH |
2166 | case -EFAULT: |
2167 | double_unlock_hb(hb1, hb2); | |
69cd9eba | 2168 | hb_waiters_dec(hb2); |
ae791a2d TG |
2169 | put_futex_key(&key2); |
2170 | put_futex_key(&key1); | |
d0725992 | 2171 | ret = fault_in_user_writeable(uaddr2); |
52400ba9 DH |
2172 | if (!ret) |
2173 | goto retry; | |
2174 | goto out; | |
ac31c7ff | 2175 | case -EBUSY: |
52400ba9 | 2176 | case -EAGAIN: |
af54d6a1 TG |
2177 | /* |
2178 | * Two reasons for this: | |
ac31c7ff | 2179 | * - EBUSY: Owner is exiting and we just wait for the |
af54d6a1 | 2180 | * exit to complete. |
ac31c7ff | 2181 | * - EAGAIN: The user space value changed. |
af54d6a1 | 2182 | */ |
52400ba9 | 2183 | double_unlock_hb(hb1, hb2); |
69cd9eba | 2184 | hb_waiters_dec(hb2); |
ae791a2d TG |
2185 | put_futex_key(&key2); |
2186 | put_futex_key(&key1); | |
3ef240ea TG |
2187 | /* |
2188 | * Handle the case where the owner is in the middle of | |
2189 | * exiting. Wait for the exit to complete otherwise | |
2190 | * this task might loop forever, aka. live lock. | |
2191 | */ | |
2192 | wait_for_owner_exiting(ret, exiting); | |
52400ba9 DH |
2193 | cond_resched(); |
2194 | goto retry; | |
2195 | default: | |
2196 | goto out_unlock; | |
2197 | } | |
2198 | } | |
2199 | ||
0d00c7b2 | 2200 | plist_for_each_entry_safe(this, next, &hb1->chain, list) { |
52400ba9 DH |
2201 | if (task_count - nr_wake >= nr_requeue) |
2202 | break; | |
2203 | ||
2204 | if (!match_futex(&this->key, &key1)) | |
1da177e4 | 2205 | continue; |
52400ba9 | 2206 | |
392741e0 DH |
2207 | /* |
2208 | * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always | |
2209 | * be paired with each other and no other futex ops. | |
aa10990e DH |
2210 | * |
2211 | * We should never be requeueing a futex_q with a pi_state, | |
2212 | * which is awaiting a futex_unlock_pi(). | |
392741e0 DH |
2213 | */ |
2214 | if ((requeue_pi && !this->rt_waiter) || | |
aa10990e DH |
2215 | (!requeue_pi && this->rt_waiter) || |
2216 | this->pi_state) { | |
392741e0 DH |
2217 | ret = -EINVAL; |
2218 | break; | |
2219 | } | |
52400ba9 DH |
2220 | |
2221 | /* | |
2222 | * Wake nr_wake waiters. For requeue_pi, if we acquired the | |
2223 | * lock, we already woke the top_waiter. If not, it will be | |
2224 | * woken by futex_unlock_pi(). | |
2225 | */ | |
2226 | if (++task_count <= nr_wake && !requeue_pi) { | |
1d0dcb3a | 2227 | mark_wake_futex(&wake_q, this); |
52400ba9 DH |
2228 | continue; |
2229 | } | |
1da177e4 | 2230 | |
84bc4af5 DH |
2231 | /* Ensure we requeue to the expected futex for requeue_pi. */ |
2232 | if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) { | |
2233 | ret = -EINVAL; | |
2234 | break; | |
2235 | } | |
2236 | ||
52400ba9 DH |
2237 | /* |
2238 | * Requeue nr_requeue waiters and possibly one more in the case | |
2239 | * of requeue_pi if we couldn't acquire the lock atomically. | |
2240 | */ | |
2241 | if (requeue_pi) { | |
ecb38b78 TG |
2242 | /* |
2243 | * Prepare the waiter to take the rt_mutex. Take a | |
2244 | * refcount on the pi_state and store the pointer in | |
2245 | * the futex_q object of the waiter. | |
2246 | */ | |
bf92cf3a | 2247 | get_pi_state(pi_state); |
52400ba9 DH |
2248 | this->pi_state = pi_state; |
2249 | ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex, | |
2250 | this->rt_waiter, | |
c051b21f | 2251 | this->task); |
52400ba9 | 2252 | if (ret == 1) { |
ecb38b78 TG |
2253 | /* |
2254 | * We got the lock. We do neither drop the | |
2255 | * refcount on pi_state nor clear | |
2256 | * this->pi_state because the waiter needs the | |
2257 | * pi_state for cleaning up the user space | |
2258 | * value. It will drop the refcount after | |
2259 | * doing so. | |
2260 | */ | |
beda2c7e | 2261 | requeue_pi_wake_futex(this, &key2, hb2); |
89061d3d | 2262 | drop_count++; |
52400ba9 DH |
2263 | continue; |
2264 | } else if (ret) { | |
ecb38b78 TG |
2265 | /* |
2266 | * rt_mutex_start_proxy_lock() detected a | |
2267 | * potential deadlock when we tried to queue | |
2268 | * that waiter. Drop the pi_state reference | |
2269 | * which we took above and remove the pointer | |
2270 | * to the state from the waiters futex_q | |
2271 | * object. | |
2272 | */ | |
52400ba9 | 2273 | this->pi_state = NULL; |
29e9ee5d | 2274 | put_pi_state(pi_state); |
885c2cb7 TG |
2275 | /* |
2276 | * We stop queueing more waiters and let user | |
2277 | * space deal with the mess. | |
2278 | */ | |
2279 | break; | |
52400ba9 | 2280 | } |
1da177e4 | 2281 | } |
52400ba9 DH |
2282 | requeue_futex(this, hb1, hb2, &key2); |
2283 | drop_count++; | |
1da177e4 LT |
2284 | } |
2285 | ||
ecb38b78 TG |
2286 | /* |
2287 | * We took an extra initial reference to the pi_state either | |
2288 | * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We | |
2289 | * need to drop it here again. | |
2290 | */ | |
29e9ee5d | 2291 | put_pi_state(pi_state); |
885c2cb7 TG |
2292 | |
2293 | out_unlock: | |
5eb3dc62 | 2294 | double_unlock_hb(hb1, hb2); |
1d0dcb3a | 2295 | wake_up_q(&wake_q); |
69cd9eba | 2296 | hb_waiters_dec(hb2); |
1da177e4 | 2297 | |
cd84a42f DH |
2298 | /* |
2299 | * drop_futex_key_refs() must be called outside the spinlocks. During | |
2300 | * the requeue we moved futex_q's from the hash bucket at key1 to the | |
2301 | * one at key2 and updated their key pointer. We no longer need to | |
2302 | * hold the references to key1. | |
2303 | */ | |
1da177e4 | 2304 | while (--drop_count >= 0) |
9adef58b | 2305 | drop_futex_key_refs(&key1); |
1da177e4 | 2306 | |
42d35d48 | 2307 | out_put_keys: |
ae791a2d | 2308 | put_futex_key(&key2); |
42d35d48 | 2309 | out_put_key1: |
ae791a2d | 2310 | put_futex_key(&key1); |
42d35d48 | 2311 | out: |
52400ba9 | 2312 | return ret ? ret : task_count; |
1da177e4 LT |
2313 | } |
2314 | ||
2315 | /* The key must be already stored in q->key. */ | |
82af7aca | 2316 | static inline struct futex_hash_bucket *queue_lock(struct futex_q *q) |
15e408cd | 2317 | __acquires(&hb->lock) |
1da177e4 | 2318 | { |
e2970f2f | 2319 | struct futex_hash_bucket *hb; |
1da177e4 | 2320 | |
e2970f2f | 2321 | hb = hash_futex(&q->key); |
11d4616b LT |
2322 | |
2323 | /* | |
2324 | * Increment the counter before taking the lock so that | |
2325 | * a potential waker won't miss a to-be-slept task that is | |
2326 | * waiting for the spinlock. This is safe as all queue_lock() | |
2327 | * users end up calling queue_me(). Similarly, for housekeeping, | |
2328 | * decrement the counter at queue_unlock() when some error has | |
2329 | * occurred and we don't end up adding the task to the list. | |
2330 | */ | |
6f568ebe | 2331 | hb_waiters_inc(hb); /* implies smp_mb(); (A) */ |
11d4616b | 2332 | |
e2970f2f | 2333 | q->lock_ptr = &hb->lock; |
1da177e4 | 2334 | |
6f568ebe | 2335 | spin_lock(&hb->lock); |
e2970f2f | 2336 | return hb; |
1da177e4 LT |
2337 | } |
2338 | ||
d40d65c8 | 2339 | static inline void |
0d00c7b2 | 2340 | queue_unlock(struct futex_hash_bucket *hb) |
15e408cd | 2341 | __releases(&hb->lock) |
d40d65c8 DH |
2342 | { |
2343 | spin_unlock(&hb->lock); | |
11d4616b | 2344 | hb_waiters_dec(hb); |
d40d65c8 DH |
2345 | } |
2346 | ||
cfafcd11 | 2347 | static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb) |
1da177e4 | 2348 | { |
ec92d082 PP |
2349 | int prio; |
2350 | ||
2351 | /* | |
2352 | * The priority used to register this element is | |
2353 | * - either the real thread-priority for the real-time threads | |
2354 | * (i.e. threads with a priority lower than MAX_RT_PRIO) | |
2355 | * - or MAX_RT_PRIO for non-RT threads. | |
2356 | * Thus, all RT-threads are woken first in priority order, and | |
2357 | * the others are woken last, in FIFO order. | |
2358 | */ | |
2359 | prio = min(current->normal_prio, MAX_RT_PRIO); | |
2360 | ||
2361 | plist_node_init(&q->list, prio); | |
ec92d082 | 2362 | plist_add(&q->list, &hb->chain); |
c87e2837 | 2363 | q->task = current; |
cfafcd11 PZ |
2364 | } |
2365 | ||
2366 | /** | |
2367 | * queue_me() - Enqueue the futex_q on the futex_hash_bucket | |
2368 | * @q: The futex_q to enqueue | |
2369 | * @hb: The destination hash bucket | |
2370 | * | |
2371 | * The hb->lock must be held by the caller, and is released here. A call to | |
2372 | * queue_me() is typically paired with exactly one call to unqueue_me(). The | |
2373 | * exceptions involve the PI related operations, which may use unqueue_me_pi() | |
2374 | * or nothing if the unqueue is done as part of the wake process and the unqueue | |
2375 | * state is implicit in the state of woken task (see futex_wait_requeue_pi() for | |
2376 | * an example). | |
2377 | */ | |
2378 | static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb) | |
2379 | __releases(&hb->lock) | |
2380 | { | |
2381 | __queue_me(q, hb); | |
e2970f2f | 2382 | spin_unlock(&hb->lock); |
1da177e4 LT |
2383 | } |
2384 | ||
d40d65c8 DH |
2385 | /** |
2386 | * unqueue_me() - Remove the futex_q from its futex_hash_bucket | |
2387 | * @q: The futex_q to unqueue | |
2388 | * | |
2389 | * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must | |
2390 | * be paired with exactly one earlier call to queue_me(). | |
2391 | * | |
6c23cbbd | 2392 | * Return: |
7b4ff1ad MCC |
2393 | * - 1 - if the futex_q was still queued (and we removed unqueued it); |
2394 | * - 0 - if the futex_q was already removed by the waking thread | |
1da177e4 | 2395 | */ |
1da177e4 LT |
2396 | static int unqueue_me(struct futex_q *q) |
2397 | { | |
1da177e4 | 2398 | spinlock_t *lock_ptr; |
e2970f2f | 2399 | int ret = 0; |
1da177e4 LT |
2400 | |
2401 | /* In the common case we don't take the spinlock, which is nice. */ | |
42d35d48 | 2402 | retry: |
29b75eb2 JZ |
2403 | /* |
2404 | * q->lock_ptr can change between this read and the following spin_lock. | |
2405 | * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and | |
2406 | * optimizing lock_ptr out of the logic below. | |
2407 | */ | |
2408 | lock_ptr = READ_ONCE(q->lock_ptr); | |
c80544dc | 2409 | if (lock_ptr != NULL) { |
1da177e4 LT |
2410 | spin_lock(lock_ptr); |
2411 | /* | |
2412 | * q->lock_ptr can change between reading it and | |
2413 | * spin_lock(), causing us to take the wrong lock. This | |
2414 | * corrects the race condition. | |
2415 | * | |
2416 | * Reasoning goes like this: if we have the wrong lock, | |
2417 | * q->lock_ptr must have changed (maybe several times) | |
2418 | * between reading it and the spin_lock(). It can | |
2419 | * change again after the spin_lock() but only if it was | |
2420 | * already changed before the spin_lock(). It cannot, | |
2421 | * however, change back to the original value. Therefore | |
2422 | * we can detect whether we acquired the correct lock. | |
2423 | */ | |
2424 | if (unlikely(lock_ptr != q->lock_ptr)) { | |
2425 | spin_unlock(lock_ptr); | |
2426 | goto retry; | |
2427 | } | |
2e12978a | 2428 | __unqueue_futex(q); |
c87e2837 IM |
2429 | |
2430 | BUG_ON(q->pi_state); | |
2431 | ||
1da177e4 LT |
2432 | spin_unlock(lock_ptr); |
2433 | ret = 1; | |
2434 | } | |
2435 | ||
9adef58b | 2436 | drop_futex_key_refs(&q->key); |
1da177e4 LT |
2437 | return ret; |
2438 | } | |
2439 | ||
c87e2837 IM |
2440 | /* |
2441 | * PI futexes can not be requeued and must remove themself from the | |
d0aa7a70 PP |
2442 | * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry |
2443 | * and dropped here. | |
c87e2837 | 2444 | */ |
d0aa7a70 | 2445 | static void unqueue_me_pi(struct futex_q *q) |
15e408cd | 2446 | __releases(q->lock_ptr) |
c87e2837 | 2447 | { |
2e12978a | 2448 | __unqueue_futex(q); |
c87e2837 IM |
2449 | |
2450 | BUG_ON(!q->pi_state); | |
29e9ee5d | 2451 | put_pi_state(q->pi_state); |
c87e2837 IM |
2452 | q->pi_state = NULL; |
2453 | ||
d0aa7a70 | 2454 | spin_unlock(q->lock_ptr); |
c87e2837 IM |
2455 | } |
2456 | ||
778e9a9c | 2457 | static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q, |
c1e2f0ea | 2458 | struct task_struct *argowner) |
d0aa7a70 | 2459 | { |
d0aa7a70 | 2460 | struct futex_pi_state *pi_state = q->pi_state; |
7cfdaf38 | 2461 | u32 uval, uninitialized_var(curval), newval; |
c1e2f0ea PZ |
2462 | struct task_struct *oldowner, *newowner; |
2463 | u32 newtid; | |
6b4f4bc9 | 2464 | int ret, err = 0; |
d0aa7a70 | 2465 | |
c1e2f0ea PZ |
2466 | lockdep_assert_held(q->lock_ptr); |
2467 | ||
734009e9 PZ |
2468 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); |
2469 | ||
2470 | oldowner = pi_state->owner; | |
1b7558e4 TG |
2471 | |
2472 | /* | |
c1e2f0ea | 2473 | * We are here because either: |
16ffa12d | 2474 | * |
c1e2f0ea PZ |
2475 | * - we stole the lock and pi_state->owner needs updating to reflect |
2476 | * that (@argowner == current), | |
2477 | * | |
2478 | * or: | |
2479 | * | |
2480 | * - someone stole our lock and we need to fix things to point to the | |
2481 | * new owner (@argowner == NULL). | |
2482 | * | |
2483 | * Either way, we have to replace the TID in the user space variable. | |
8161239a | 2484 | * This must be atomic as we have to preserve the owner died bit here. |
1b7558e4 | 2485 | * |
b2d0994b DH |
2486 | * Note: We write the user space value _before_ changing the pi_state |
2487 | * because we can fault here. Imagine swapped out pages or a fork | |
2488 | * that marked all the anonymous memory readonly for cow. | |
1b7558e4 | 2489 | * |
734009e9 PZ |
2490 | * Modifying pi_state _before_ the user space value would leave the |
2491 | * pi_state in an inconsistent state when we fault here, because we | |
2492 | * need to drop the locks to handle the fault. This might be observed | |
2493 | * in the PID check in lookup_pi_state. | |
1b7558e4 TG |
2494 | */ |
2495 | retry: | |
c1e2f0ea PZ |
2496 | if (!argowner) { |
2497 | if (oldowner != current) { | |
2498 | /* | |
2499 | * We raced against a concurrent self; things are | |
2500 | * already fixed up. Nothing to do. | |
2501 | */ | |
2502 | ret = 0; | |
2503 | goto out_unlock; | |
2504 | } | |
2505 | ||
2506 | if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) { | |
2507 | /* We got the lock after all, nothing to fix. */ | |
2508 | ret = 0; | |
2509 | goto out_unlock; | |
2510 | } | |
2511 | ||
2512 | /* | |
2513 | * Since we just failed the trylock; there must be an owner. | |
2514 | */ | |
2515 | newowner = rt_mutex_owner(&pi_state->pi_mutex); | |
2516 | BUG_ON(!newowner); | |
2517 | } else { | |
2518 | WARN_ON_ONCE(argowner != current); | |
2519 | if (oldowner == current) { | |
2520 | /* | |
2521 | * We raced against a concurrent self; things are | |
2522 | * already fixed up. Nothing to do. | |
2523 | */ | |
2524 | ret = 0; | |
2525 | goto out_unlock; | |
2526 | } | |
2527 | newowner = argowner; | |
2528 | } | |
2529 | ||
2530 | newtid = task_pid_vnr(newowner) | FUTEX_WAITERS; | |
a97cb0e7 PZ |
2531 | /* Owner died? */ |
2532 | if (!pi_state->owner) | |
2533 | newtid |= FUTEX_OWNER_DIED; | |
c1e2f0ea | 2534 | |
6b4f4bc9 WD |
2535 | err = get_futex_value_locked(&uval, uaddr); |
2536 | if (err) | |
2537 | goto handle_err; | |
1b7558e4 | 2538 | |
16ffa12d | 2539 | for (;;) { |
1b7558e4 TG |
2540 | newval = (uval & FUTEX_OWNER_DIED) | newtid; |
2541 | ||
6b4f4bc9 WD |
2542 | err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval); |
2543 | if (err) | |
2544 | goto handle_err; | |
2545 | ||
1b7558e4 TG |
2546 | if (curval == uval) |
2547 | break; | |
2548 | uval = curval; | |
2549 | } | |
2550 | ||
2551 | /* | |
2552 | * We fixed up user space. Now we need to fix the pi_state | |
2553 | * itself. | |
2554 | */ | |
d0aa7a70 | 2555 | if (pi_state->owner != NULL) { |
734009e9 | 2556 | raw_spin_lock(&pi_state->owner->pi_lock); |
d0aa7a70 PP |
2557 | WARN_ON(list_empty(&pi_state->list)); |
2558 | list_del_init(&pi_state->list); | |
734009e9 | 2559 | raw_spin_unlock(&pi_state->owner->pi_lock); |
1b7558e4 | 2560 | } |
d0aa7a70 | 2561 | |
cdf71a10 | 2562 | pi_state->owner = newowner; |
d0aa7a70 | 2563 | |
734009e9 | 2564 | raw_spin_lock(&newowner->pi_lock); |
d0aa7a70 | 2565 | WARN_ON(!list_empty(&pi_state->list)); |
cdf71a10 | 2566 | list_add(&pi_state->list, &newowner->pi_state_list); |
734009e9 PZ |
2567 | raw_spin_unlock(&newowner->pi_lock); |
2568 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
2569 | ||
1b7558e4 | 2570 | return 0; |
d0aa7a70 | 2571 | |
d0aa7a70 | 2572 | /* |
6b4f4bc9 WD |
2573 | * In order to reschedule or handle a page fault, we need to drop the |
2574 | * locks here. In the case of a fault, this gives the other task | |
2575 | * (either the highest priority waiter itself or the task which stole | |
2576 | * the rtmutex) the chance to try the fixup of the pi_state. So once we | |
2577 | * are back from handling the fault we need to check the pi_state after | |
2578 | * reacquiring the locks and before trying to do another fixup. When | |
2579 | * the fixup has been done already we simply return. | |
734009e9 PZ |
2580 | * |
2581 | * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely | |
2582 | * drop hb->lock since the caller owns the hb -> futex_q relation. | |
2583 | * Dropping the pi_mutex->wait_lock requires the state revalidate. | |
d0aa7a70 | 2584 | */ |
6b4f4bc9 | 2585 | handle_err: |
734009e9 | 2586 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); |
1b7558e4 | 2587 | spin_unlock(q->lock_ptr); |
778e9a9c | 2588 | |
6b4f4bc9 WD |
2589 | switch (err) { |
2590 | case -EFAULT: | |
2591 | ret = fault_in_user_writeable(uaddr); | |
2592 | break; | |
2593 | ||
2594 | case -EAGAIN: | |
2595 | cond_resched(); | |
2596 | ret = 0; | |
2597 | break; | |
2598 | ||
2599 | default: | |
2600 | WARN_ON_ONCE(1); | |
2601 | ret = err; | |
2602 | break; | |
2603 | } | |
778e9a9c | 2604 | |
1b7558e4 | 2605 | spin_lock(q->lock_ptr); |
734009e9 | 2606 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); |
778e9a9c | 2607 | |
1b7558e4 TG |
2608 | /* |
2609 | * Check if someone else fixed it for us: | |
2610 | */ | |
734009e9 PZ |
2611 | if (pi_state->owner != oldowner) { |
2612 | ret = 0; | |
2613 | goto out_unlock; | |
2614 | } | |
1b7558e4 TG |
2615 | |
2616 | if (ret) | |
734009e9 | 2617 | goto out_unlock; |
1b7558e4 TG |
2618 | |
2619 | goto retry; | |
734009e9 PZ |
2620 | |
2621 | out_unlock: | |
2622 | raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock); | |
2623 | return ret; | |
d0aa7a70 PP |
2624 | } |
2625 | ||
72c1bbf3 | 2626 | static long futex_wait_restart(struct restart_block *restart); |
36cf3b5c | 2627 | |
dd973998 DH |
2628 | /** |
2629 | * fixup_owner() - Post lock pi_state and corner case management | |
2630 | * @uaddr: user address of the futex | |
dd973998 DH |
2631 | * @q: futex_q (contains pi_state and access to the rt_mutex) |
2632 | * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0) | |
2633 | * | |
2634 | * After attempting to lock an rt_mutex, this function is called to cleanup | |
2635 | * the pi_state owner as well as handle race conditions that may allow us to | |
2636 | * acquire the lock. Must be called with the hb lock held. | |
2637 | * | |
6c23cbbd | 2638 | * Return: |
7b4ff1ad MCC |
2639 | * - 1 - success, lock taken; |
2640 | * - 0 - success, lock not taken; | |
2641 | * - <0 - on error (-EFAULT) | |
dd973998 | 2642 | */ |
ae791a2d | 2643 | static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked) |
dd973998 | 2644 | { |
dd973998 DH |
2645 | int ret = 0; |
2646 | ||
2647 | if (locked) { | |
2648 | /* | |
2649 | * Got the lock. We might not be the anticipated owner if we | |
2650 | * did a lock-steal - fix up the PI-state in that case: | |
16ffa12d | 2651 | * |
c1e2f0ea PZ |
2652 | * Speculative pi_state->owner read (we don't hold wait_lock); |
2653 | * since we own the lock pi_state->owner == current is the | |
2654 | * stable state, anything else needs more attention. | |
dd973998 DH |
2655 | */ |
2656 | if (q->pi_state->owner != current) | |
ae791a2d | 2657 | ret = fixup_pi_state_owner(uaddr, q, current); |
dd973998 DH |
2658 | goto out; |
2659 | } | |
2660 | ||
c1e2f0ea PZ |
2661 | /* |
2662 | * If we didn't get the lock; check if anybody stole it from us. In | |
2663 | * that case, we need to fix up the uval to point to them instead of | |
2664 | * us, otherwise bad things happen. [10] | |
2665 | * | |
2666 | * Another speculative read; pi_state->owner == current is unstable | |
2667 | * but needs our attention. | |
2668 | */ | |
2669 | if (q->pi_state->owner == current) { | |
2670 | ret = fixup_pi_state_owner(uaddr, q, NULL); | |
2671 | goto out; | |
2672 | } | |
2673 | ||
dd973998 DH |
2674 | /* |
2675 | * Paranoia check. If we did not take the lock, then we should not be | |
8161239a | 2676 | * the owner of the rt_mutex. |
dd973998 | 2677 | */ |
73d786bd | 2678 | if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) { |
dd973998 DH |
2679 | printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p " |
2680 | "pi-state %p\n", ret, | |
2681 | q->pi_state->pi_mutex.owner, | |
2682 | q->pi_state->owner); | |
73d786bd | 2683 | } |
dd973998 DH |
2684 | |
2685 | out: | |
2686 | return ret ? ret : locked; | |
2687 | } | |
2688 | ||
ca5f9524 DH |
2689 | /** |
2690 | * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal | |
2691 | * @hb: the futex hash bucket, must be locked by the caller | |
2692 | * @q: the futex_q to queue up on | |
2693 | * @timeout: the prepared hrtimer_sleeper, or null for no timeout | |
ca5f9524 DH |
2694 | */ |
2695 | static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q, | |
f1a11e05 | 2696 | struct hrtimer_sleeper *timeout) |
ca5f9524 | 2697 | { |
9beba3c5 DH |
2698 | /* |
2699 | * The task state is guaranteed to be set before another task can | |
b92b8b35 | 2700 | * wake it. set_current_state() is implemented using smp_store_mb() and |
9beba3c5 DH |
2701 | * queue_me() calls spin_unlock() upon completion, both serializing |
2702 | * access to the hash list and forcing another memory barrier. | |
2703 | */ | |
f1a11e05 | 2704 | set_current_state(TASK_INTERRUPTIBLE); |
0729e196 | 2705 | queue_me(q, hb); |
ca5f9524 DH |
2706 | |
2707 | /* Arm the timer */ | |
2e4b0d3f | 2708 | if (timeout) |
9dd8813e | 2709 | hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS); |
ca5f9524 DH |
2710 | |
2711 | /* | |
0729e196 DH |
2712 | * If we have been removed from the hash list, then another task |
2713 | * has tried to wake us, and we can skip the call to schedule(). | |
ca5f9524 DH |
2714 | */ |
2715 | if (likely(!plist_node_empty(&q->list))) { | |
2716 | /* | |
2717 | * If the timer has already expired, current will already be | |
2718 | * flagged for rescheduling. Only call schedule if there | |
2719 | * is no timeout, or if it has yet to expire. | |
2720 | */ | |
2721 | if (!timeout || timeout->task) | |
88c8004f | 2722 | freezable_schedule(); |
ca5f9524 DH |
2723 | } |
2724 | __set_current_state(TASK_RUNNING); | |
2725 | } | |
2726 | ||
f801073f DH |
2727 | /** |
2728 | * futex_wait_setup() - Prepare to wait on a futex | |
2729 | * @uaddr: the futex userspace address | |
2730 | * @val: the expected value | |
b41277dc | 2731 | * @flags: futex flags (FLAGS_SHARED, etc.) |
f801073f DH |
2732 | * @q: the associated futex_q |
2733 | * @hb: storage for hash_bucket pointer to be returned to caller | |
2734 | * | |
2735 | * Setup the futex_q and locate the hash_bucket. Get the futex value and | |
2736 | * compare it with the expected value. Handle atomic faults internally. | |
2737 | * Return with the hb lock held and a q.key reference on success, and unlocked | |
2738 | * with no q.key reference on failure. | |
2739 | * | |
6c23cbbd | 2740 | * Return: |
7b4ff1ad MCC |
2741 | * - 0 - uaddr contains val and hb has been locked; |
2742 | * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked | |
f801073f | 2743 | */ |
b41277dc | 2744 | static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags, |
f801073f | 2745 | struct futex_q *q, struct futex_hash_bucket **hb) |
1da177e4 | 2746 | { |
e2970f2f IM |
2747 | u32 uval; |
2748 | int ret; | |
1da177e4 | 2749 | |
1da177e4 | 2750 | /* |
b2d0994b | 2751 | * Access the page AFTER the hash-bucket is locked. |
1da177e4 LT |
2752 | * Order is important: |
2753 | * | |
2754 | * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val); | |
2755 | * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); } | |
2756 | * | |
2757 | * The basic logical guarantee of a futex is that it blocks ONLY | |
2758 | * if cond(var) is known to be true at the time of blocking, for | |
8fe8f545 ML |
2759 | * any cond. If we locked the hash-bucket after testing *uaddr, that |
2760 | * would open a race condition where we could block indefinitely with | |
1da177e4 LT |
2761 | * cond(var) false, which would violate the guarantee. |
2762 | * | |
8fe8f545 ML |
2763 | * On the other hand, we insert q and release the hash-bucket only |
2764 | * after testing *uaddr. This guarantees that futex_wait() will NOT | |
2765 | * absorb a wakeup if *uaddr does not match the desired values | |
2766 | * while the syscall executes. | |
1da177e4 | 2767 | */ |
f801073f | 2768 | retry: |
96d4f267 | 2769 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ); |
f801073f | 2770 | if (unlikely(ret != 0)) |
a5a2a0c7 | 2771 | return ret; |
f801073f DH |
2772 | |
2773 | retry_private: | |
2774 | *hb = queue_lock(q); | |
2775 | ||
e2970f2f | 2776 | ret = get_futex_value_locked(&uval, uaddr); |
1da177e4 | 2777 | |
f801073f | 2778 | if (ret) { |
0d00c7b2 | 2779 | queue_unlock(*hb); |
1da177e4 | 2780 | |
e2970f2f | 2781 | ret = get_user(uval, uaddr); |
e4dc5b7a | 2782 | if (ret) |
f801073f | 2783 | goto out; |
1da177e4 | 2784 | |
b41277dc | 2785 | if (!(flags & FLAGS_SHARED)) |
e4dc5b7a DH |
2786 | goto retry_private; |
2787 | ||
ae791a2d | 2788 | put_futex_key(&q->key); |
e4dc5b7a | 2789 | goto retry; |
1da177e4 | 2790 | } |
ca5f9524 | 2791 | |
f801073f | 2792 | if (uval != val) { |
0d00c7b2 | 2793 | queue_unlock(*hb); |
f801073f | 2794 | ret = -EWOULDBLOCK; |
2fff78c7 | 2795 | } |
1da177e4 | 2796 | |
f801073f DH |
2797 | out: |
2798 | if (ret) | |
ae791a2d | 2799 | put_futex_key(&q->key); |
f801073f DH |
2800 | return ret; |
2801 | } | |
2802 | ||
b41277dc DH |
2803 | static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val, |
2804 | ktime_t *abs_time, u32 bitset) | |
f801073f | 2805 | { |
5ca584d9 | 2806 | struct hrtimer_sleeper timeout, *to; |
f801073f DH |
2807 | struct restart_block *restart; |
2808 | struct futex_hash_bucket *hb; | |
5bdb05f9 | 2809 | struct futex_q q = futex_q_init; |
f801073f DH |
2810 | int ret; |
2811 | ||
2812 | if (!bitset) | |
2813 | return -EINVAL; | |
f801073f DH |
2814 | q.bitset = bitset; |
2815 | ||
5ca584d9 WL |
2816 | to = futex_setup_timer(abs_time, &timeout, flags, |
2817 | current->timer_slack_ns); | |
d58e6576 | 2818 | retry: |
7ada876a DH |
2819 | /* |
2820 | * Prepare to wait on uaddr. On success, holds hb lock and increments | |
2821 | * q.key refs. | |
2822 | */ | |
b41277dc | 2823 | ret = futex_wait_setup(uaddr, val, flags, &q, &hb); |
f801073f DH |
2824 | if (ret) |
2825 | goto out; | |
2826 | ||
ca5f9524 | 2827 | /* queue_me and wait for wakeup, timeout, or a signal. */ |
f1a11e05 | 2828 | futex_wait_queue_me(hb, &q, to); |
1da177e4 LT |
2829 | |
2830 | /* If we were woken (and unqueued), we succeeded, whatever. */ | |
2fff78c7 | 2831 | ret = 0; |
7ada876a | 2832 | /* unqueue_me() drops q.key ref */ |
1da177e4 | 2833 | if (!unqueue_me(&q)) |
7ada876a | 2834 | goto out; |
2fff78c7 | 2835 | ret = -ETIMEDOUT; |
ca5f9524 | 2836 | if (to && !to->task) |
7ada876a | 2837 | goto out; |
72c1bbf3 | 2838 | |
e2970f2f | 2839 | /* |
d58e6576 TG |
2840 | * We expect signal_pending(current), but we might be the |
2841 | * victim of a spurious wakeup as well. | |
e2970f2f | 2842 | */ |
7ada876a | 2843 | if (!signal_pending(current)) |
d58e6576 | 2844 | goto retry; |
d58e6576 | 2845 | |
2fff78c7 | 2846 | ret = -ERESTARTSYS; |
c19384b5 | 2847 | if (!abs_time) |
7ada876a | 2848 | goto out; |
1da177e4 | 2849 | |
f56141e3 | 2850 | restart = ¤t->restart_block; |
2fff78c7 | 2851 | restart->fn = futex_wait_restart; |
a3c74c52 | 2852 | restart->futex.uaddr = uaddr; |
2fff78c7 | 2853 | restart->futex.val = val; |
2456e855 | 2854 | restart->futex.time = *abs_time; |
2fff78c7 | 2855 | restart->futex.bitset = bitset; |
0cd9c649 | 2856 | restart->futex.flags = flags | FLAGS_HAS_TIMEOUT; |
42d35d48 | 2857 | |
2fff78c7 PZ |
2858 | ret = -ERESTART_RESTARTBLOCK; |
2859 | ||
42d35d48 | 2860 | out: |
ca5f9524 DH |
2861 | if (to) { |
2862 | hrtimer_cancel(&to->timer); | |
2863 | destroy_hrtimer_on_stack(&to->timer); | |
2864 | } | |
c87e2837 IM |
2865 | return ret; |
2866 | } | |
2867 | ||
72c1bbf3 NP |
2868 | |
2869 | static long futex_wait_restart(struct restart_block *restart) | |
2870 | { | |
a3c74c52 | 2871 | u32 __user *uaddr = restart->futex.uaddr; |
a72188d8 | 2872 | ktime_t t, *tp = NULL; |
72c1bbf3 | 2873 | |
a72188d8 | 2874 | if (restart->futex.flags & FLAGS_HAS_TIMEOUT) { |
2456e855 | 2875 | t = restart->futex.time; |
a72188d8 DH |
2876 | tp = &t; |
2877 | } | |
72c1bbf3 | 2878 | restart->fn = do_no_restart_syscall; |
b41277dc DH |
2879 | |
2880 | return (long)futex_wait(uaddr, restart->futex.flags, | |
2881 | restart->futex.val, tp, restart->futex.bitset); | |
72c1bbf3 NP |
2882 | } |
2883 | ||
2884 | ||
c87e2837 IM |
2885 | /* |
2886 | * Userspace tried a 0 -> TID atomic transition of the futex value | |
2887 | * and failed. The kernel side here does the whole locking operation: | |
767f509c DB |
2888 | * if there are waiters then it will block as a consequence of relying |
2889 | * on rt-mutexes, it does PI, etc. (Due to races the kernel might see | |
2890 | * a 0 value of the futex too.). | |
2891 | * | |
2892 | * Also serves as futex trylock_pi()'ing, and due semantics. | |
c87e2837 | 2893 | */ |
996636dd | 2894 | static int futex_lock_pi(u32 __user *uaddr, unsigned int flags, |
b41277dc | 2895 | ktime_t *time, int trylock) |
c87e2837 | 2896 | { |
5ca584d9 | 2897 | struct hrtimer_sleeper timeout, *to; |
16ffa12d | 2898 | struct futex_pi_state *pi_state = NULL; |
3ef240ea | 2899 | struct task_struct *exiting = NULL; |
cfafcd11 | 2900 | struct rt_mutex_waiter rt_waiter; |
c87e2837 | 2901 | struct futex_hash_bucket *hb; |
5bdb05f9 | 2902 | struct futex_q q = futex_q_init; |
dd973998 | 2903 | int res, ret; |
c87e2837 | 2904 | |
bc2eecd7 NP |
2905 | if (!IS_ENABLED(CONFIG_FUTEX_PI)) |
2906 | return -ENOSYS; | |
2907 | ||
c87e2837 IM |
2908 | if (refill_pi_state_cache()) |
2909 | return -ENOMEM; | |
2910 | ||
5ca584d9 | 2911 | to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0); |
c5780e97 | 2912 | |
42d35d48 | 2913 | retry: |
96d4f267 | 2914 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE); |
c87e2837 | 2915 | if (unlikely(ret != 0)) |
42d35d48 | 2916 | goto out; |
c87e2837 | 2917 | |
e4dc5b7a | 2918 | retry_private: |
82af7aca | 2919 | hb = queue_lock(&q); |
c87e2837 | 2920 | |
3ef240ea TG |
2921 | ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, |
2922 | &exiting, 0); | |
c87e2837 | 2923 | if (unlikely(ret)) { |
767f509c DB |
2924 | /* |
2925 | * Atomic work succeeded and we got the lock, | |
2926 | * or failed. Either way, we do _not_ block. | |
2927 | */ | |
778e9a9c | 2928 | switch (ret) { |
1a52084d DH |
2929 | case 1: |
2930 | /* We got the lock. */ | |
2931 | ret = 0; | |
2932 | goto out_unlock_put_key; | |
2933 | case -EFAULT: | |
2934 | goto uaddr_faulted; | |
ac31c7ff | 2935 | case -EBUSY: |
778e9a9c AK |
2936 | case -EAGAIN: |
2937 | /* | |
af54d6a1 | 2938 | * Two reasons for this: |
ac31c7ff | 2939 | * - EBUSY: Task is exiting and we just wait for the |
af54d6a1 | 2940 | * exit to complete. |
ac31c7ff | 2941 | * - EAGAIN: The user space value changed. |
778e9a9c | 2942 | */ |
0d00c7b2 | 2943 | queue_unlock(hb); |
ae791a2d | 2944 | put_futex_key(&q.key); |
3ef240ea TG |
2945 | /* |
2946 | * Handle the case where the owner is in the middle of | |
2947 | * exiting. Wait for the exit to complete otherwise | |
2948 | * this task might loop forever, aka. live lock. | |
2949 | */ | |
2950 | wait_for_owner_exiting(ret, exiting); | |
778e9a9c AK |
2951 | cond_resched(); |
2952 | goto retry; | |
778e9a9c | 2953 | default: |
42d35d48 | 2954 | goto out_unlock_put_key; |
c87e2837 | 2955 | } |
c87e2837 IM |
2956 | } |
2957 | ||
cfafcd11 PZ |
2958 | WARN_ON(!q.pi_state); |
2959 | ||
c87e2837 IM |
2960 | /* |
2961 | * Only actually queue now that the atomic ops are done: | |
2962 | */ | |
cfafcd11 | 2963 | __queue_me(&q, hb); |
c87e2837 | 2964 | |
cfafcd11 | 2965 | if (trylock) { |
5293c2ef | 2966 | ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex); |
c87e2837 IM |
2967 | /* Fixup the trylock return value: */ |
2968 | ret = ret ? 0 : -EWOULDBLOCK; | |
cfafcd11 | 2969 | goto no_block; |
c87e2837 IM |
2970 | } |
2971 | ||
56222b21 PZ |
2972 | rt_mutex_init_waiter(&rt_waiter); |
2973 | ||
cfafcd11 | 2974 | /* |
56222b21 PZ |
2975 | * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not |
2976 | * hold it while doing rt_mutex_start_proxy(), because then it will | |
2977 | * include hb->lock in the blocking chain, even through we'll not in | |
2978 | * fact hold it while blocking. This will lead it to report -EDEADLK | |
2979 | * and BUG when futex_unlock_pi() interleaves with this. | |
2980 | * | |
2981 | * Therefore acquire wait_lock while holding hb->lock, but drop the | |
1a1fb985 TG |
2982 | * latter before calling __rt_mutex_start_proxy_lock(). This |
2983 | * interleaves with futex_unlock_pi() -- which does a similar lock | |
2984 | * handoff -- such that the latter can observe the futex_q::pi_state | |
2985 | * before __rt_mutex_start_proxy_lock() is done. | |
cfafcd11 | 2986 | */ |
56222b21 PZ |
2987 | raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock); |
2988 | spin_unlock(q.lock_ptr); | |
1a1fb985 TG |
2989 | /* |
2990 | * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter | |
2991 | * such that futex_unlock_pi() is guaranteed to observe the waiter when | |
2992 | * it sees the futex_q::pi_state. | |
2993 | */ | |
56222b21 PZ |
2994 | ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current); |
2995 | raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock); | |
2996 | ||
cfafcd11 PZ |
2997 | if (ret) { |
2998 | if (ret == 1) | |
2999 | ret = 0; | |
1a1fb985 | 3000 | goto cleanup; |
cfafcd11 PZ |
3001 | } |
3002 | ||
cfafcd11 | 3003 | if (unlikely(to)) |
9dd8813e | 3004 | hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS); |
cfafcd11 PZ |
3005 | |
3006 | ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter); | |
3007 | ||
1a1fb985 | 3008 | cleanup: |
a99e4e41 | 3009 | spin_lock(q.lock_ptr); |
cfafcd11 | 3010 | /* |
1a1fb985 | 3011 | * If we failed to acquire the lock (deadlock/signal/timeout), we must |
cfafcd11 | 3012 | * first acquire the hb->lock before removing the lock from the |
1a1fb985 TG |
3013 | * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait |
3014 | * lists consistent. | |
56222b21 PZ |
3015 | * |
3016 | * In particular; it is important that futex_unlock_pi() can not | |
3017 | * observe this inconsistency. | |
cfafcd11 PZ |
3018 | */ |
3019 | if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter)) | |
3020 | ret = 0; | |
3021 | ||
3022 | no_block: | |
dd973998 DH |
3023 | /* |
3024 | * Fixup the pi_state owner and possibly acquire the lock if we | |
3025 | * haven't already. | |
3026 | */ | |
ae791a2d | 3027 | res = fixup_owner(uaddr, &q, !ret); |
dd973998 DH |
3028 | /* |
3029 | * If fixup_owner() returned an error, proprogate that. If it acquired | |
3030 | * the lock, clear our -ETIMEDOUT or -EINTR. | |
3031 | */ | |
3032 | if (res) | |
3033 | ret = (res < 0) ? res : 0; | |
c87e2837 | 3034 | |
e8f6386c | 3035 | /* |
dd973998 DH |
3036 | * If fixup_owner() faulted and was unable to handle the fault, unlock |
3037 | * it and return the fault to userspace. | |
e8f6386c | 3038 | */ |
16ffa12d PZ |
3039 | if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) { |
3040 | pi_state = q.pi_state; | |
3041 | get_pi_state(pi_state); | |
3042 | } | |
e8f6386c | 3043 | |
778e9a9c AK |
3044 | /* Unqueue and drop the lock */ |
3045 | unqueue_me_pi(&q); | |
c87e2837 | 3046 | |
16ffa12d PZ |
3047 | if (pi_state) { |
3048 | rt_mutex_futex_unlock(&pi_state->pi_mutex); | |
3049 | put_pi_state(pi_state); | |
3050 | } | |
3051 | ||
5ecb01cf | 3052 | goto out_put_key; |
c87e2837 | 3053 | |
42d35d48 | 3054 | out_unlock_put_key: |
0d00c7b2 | 3055 | queue_unlock(hb); |
c87e2837 | 3056 | |
42d35d48 | 3057 | out_put_key: |
ae791a2d | 3058 | put_futex_key(&q.key); |
42d35d48 | 3059 | out: |
97181f9b TG |
3060 | if (to) { |
3061 | hrtimer_cancel(&to->timer); | |
237fc6e7 | 3062 | destroy_hrtimer_on_stack(&to->timer); |
97181f9b | 3063 | } |
dd973998 | 3064 | return ret != -EINTR ? ret : -ERESTARTNOINTR; |
c87e2837 | 3065 | |
42d35d48 | 3066 | uaddr_faulted: |
0d00c7b2 | 3067 | queue_unlock(hb); |
778e9a9c | 3068 | |
d0725992 | 3069 | ret = fault_in_user_writeable(uaddr); |
e4dc5b7a DH |
3070 | if (ret) |
3071 | goto out_put_key; | |
c87e2837 | 3072 | |
b41277dc | 3073 | if (!(flags & FLAGS_SHARED)) |
e4dc5b7a DH |
3074 | goto retry_private; |
3075 | ||
ae791a2d | 3076 | put_futex_key(&q.key); |
e4dc5b7a | 3077 | goto retry; |
c87e2837 IM |
3078 | } |
3079 | ||
c87e2837 IM |
3080 | /* |
3081 | * Userspace attempted a TID -> 0 atomic transition, and failed. | |
3082 | * This is the in-kernel slowpath: we look up the PI state (if any), | |
3083 | * and do the rt-mutex unlock. | |
3084 | */ | |
b41277dc | 3085 | static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags) |
c87e2837 | 3086 | { |
ccf9e6a8 | 3087 | u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current); |
38d47c1b | 3088 | union futex_key key = FUTEX_KEY_INIT; |
ccf9e6a8 | 3089 | struct futex_hash_bucket *hb; |
499f5aca | 3090 | struct futex_q *top_waiter; |
e4dc5b7a | 3091 | int ret; |
c87e2837 | 3092 | |
bc2eecd7 NP |
3093 | if (!IS_ENABLED(CONFIG_FUTEX_PI)) |
3094 | return -ENOSYS; | |
3095 | ||
c87e2837 IM |
3096 | retry: |
3097 | if (get_user(uval, uaddr)) | |
3098 | return -EFAULT; | |
3099 | /* | |
3100 | * We release only a lock we actually own: | |
3101 | */ | |
c0c9ed15 | 3102 | if ((uval & FUTEX_TID_MASK) != vpid) |
c87e2837 | 3103 | return -EPERM; |
c87e2837 | 3104 | |
96d4f267 | 3105 | ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE); |
ccf9e6a8 TG |
3106 | if (ret) |
3107 | return ret; | |
c87e2837 IM |
3108 | |
3109 | hb = hash_futex(&key); | |
3110 | spin_lock(&hb->lock); | |
3111 | ||
c87e2837 | 3112 | /* |
ccf9e6a8 TG |
3113 | * Check waiters first. We do not trust user space values at |
3114 | * all and we at least want to know if user space fiddled | |
3115 | * with the futex value instead of blindly unlocking. | |
c87e2837 | 3116 | */ |
499f5aca PZ |
3117 | top_waiter = futex_top_waiter(hb, &key); |
3118 | if (top_waiter) { | |
16ffa12d PZ |
3119 | struct futex_pi_state *pi_state = top_waiter->pi_state; |
3120 | ||
3121 | ret = -EINVAL; | |
3122 | if (!pi_state) | |
3123 | goto out_unlock; | |
3124 | ||
3125 | /* | |
3126 | * If current does not own the pi_state then the futex is | |
3127 | * inconsistent and user space fiddled with the futex value. | |
3128 | */ | |
3129 | if (pi_state->owner != current) | |
3130 | goto out_unlock; | |
3131 | ||
bebe5b51 | 3132 | get_pi_state(pi_state); |
802ab58d | 3133 | /* |
bebe5b51 PZ |
3134 | * By taking wait_lock while still holding hb->lock, we ensure |
3135 | * there is no point where we hold neither; and therefore | |
3136 | * wake_futex_pi() must observe a state consistent with what we | |
3137 | * observed. | |
1a1fb985 TG |
3138 | * |
3139 | * In particular; this forces __rt_mutex_start_proxy() to | |
3140 | * complete such that we're guaranteed to observe the | |
3141 | * rt_waiter. Also see the WARN in wake_futex_pi(). | |
16ffa12d | 3142 | */ |
bebe5b51 | 3143 | raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); |
16ffa12d PZ |
3144 | spin_unlock(&hb->lock); |
3145 | ||
c74aef2d | 3146 | /* drops pi_state->pi_mutex.wait_lock */ |
16ffa12d PZ |
3147 | ret = wake_futex_pi(uaddr, uval, pi_state); |
3148 | ||
3149 | put_pi_state(pi_state); | |
3150 | ||
3151 | /* | |
3152 | * Success, we're done! No tricky corner cases. | |
802ab58d SAS |
3153 | */ |
3154 | if (!ret) | |
3155 | goto out_putkey; | |
c87e2837 | 3156 | /* |
ccf9e6a8 TG |
3157 | * The atomic access to the futex value generated a |
3158 | * pagefault, so retry the user-access and the wakeup: | |
c87e2837 IM |
3159 | */ |
3160 | if (ret == -EFAULT) | |
3161 | goto pi_faulted; | |
89e9e66b SAS |
3162 | /* |
3163 | * A unconditional UNLOCK_PI op raced against a waiter | |
3164 | * setting the FUTEX_WAITERS bit. Try again. | |
3165 | */ | |
6b4f4bc9 WD |
3166 | if (ret == -EAGAIN) |
3167 | goto pi_retry; | |
802ab58d SAS |
3168 | /* |
3169 | * wake_futex_pi has detected invalid state. Tell user | |
3170 | * space. | |
3171 | */ | |
16ffa12d | 3172 | goto out_putkey; |
c87e2837 | 3173 | } |
ccf9e6a8 | 3174 | |
c87e2837 | 3175 | /* |
ccf9e6a8 TG |
3176 | * We have no kernel internal state, i.e. no waiters in the |
3177 | * kernel. Waiters which are about to queue themselves are stuck | |
3178 | * on hb->lock. So we can safely ignore them. We do neither | |
3179 | * preserve the WAITERS bit not the OWNER_DIED one. We are the | |
3180 | * owner. | |
c87e2837 | 3181 | */ |
6b4f4bc9 | 3182 | if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) { |
16ffa12d | 3183 | spin_unlock(&hb->lock); |
6b4f4bc9 WD |
3184 | switch (ret) { |
3185 | case -EFAULT: | |
3186 | goto pi_faulted; | |
3187 | ||
3188 | case -EAGAIN: | |
3189 | goto pi_retry; | |
3190 | ||
3191 | default: | |
3192 | WARN_ON_ONCE(1); | |
3193 | goto out_putkey; | |
3194 | } | |
16ffa12d | 3195 | } |
c87e2837 | 3196 | |
ccf9e6a8 TG |
3197 | /* |
3198 | * If uval has changed, let user space handle it. | |
3199 | */ | |
3200 | ret = (curval == uval) ? 0 : -EAGAIN; | |
3201 | ||
c87e2837 IM |
3202 | out_unlock: |
3203 | spin_unlock(&hb->lock); | |
802ab58d | 3204 | out_putkey: |
ae791a2d | 3205 | put_futex_key(&key); |
c87e2837 IM |
3206 | return ret; |
3207 | ||
6b4f4bc9 WD |
3208 | pi_retry: |
3209 | put_futex_key(&key); | |
3210 | cond_resched(); | |
3211 | goto retry; | |
3212 | ||
c87e2837 | 3213 | pi_faulted: |
ae791a2d | 3214 | put_futex_key(&key); |
c87e2837 | 3215 | |
d0725992 | 3216 | ret = fault_in_user_writeable(uaddr); |
b5686363 | 3217 | if (!ret) |
c87e2837 IM |
3218 | goto retry; |
3219 | ||
1da177e4 LT |
3220 | return ret; |
3221 | } | |
3222 | ||
52400ba9 DH |
3223 | /** |
3224 | * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex | |
3225 | * @hb: the hash_bucket futex_q was original enqueued on | |
3226 | * @q: the futex_q woken while waiting to be requeued | |
3227 | * @key2: the futex_key of the requeue target futex | |
3228 | * @timeout: the timeout associated with the wait (NULL if none) | |
3229 | * | |
3230 | * Detect if the task was woken on the initial futex as opposed to the requeue | |
3231 | * target futex. If so, determine if it was a timeout or a signal that caused | |
3232 | * the wakeup and return the appropriate error code to the caller. Must be | |
3233 | * called with the hb lock held. | |
3234 | * | |
6c23cbbd | 3235 | * Return: |
7b4ff1ad MCC |
3236 | * - 0 = no early wakeup detected; |
3237 | * - <0 = -ETIMEDOUT or -ERESTARTNOINTR | |
52400ba9 DH |
3238 | */ |
3239 | static inline | |
3240 | int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb, | |
3241 | struct futex_q *q, union futex_key *key2, | |
3242 | struct hrtimer_sleeper *timeout) | |
3243 | { | |
3244 | int ret = 0; | |
3245 | ||
3246 | /* | |
3247 | * With the hb lock held, we avoid races while we process the wakeup. | |
3248 | * We only need to hold hb (and not hb2) to ensure atomicity as the | |
3249 | * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb. | |
3250 | * It can't be requeued from uaddr2 to something else since we don't | |
3251 | * support a PI aware source futex for requeue. | |
3252 | */ | |
3253 | if (!match_futex(&q->key, key2)) { | |
3254 | WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr)); | |
3255 | /* | |
3256 | * We were woken prior to requeue by a timeout or a signal. | |
3257 | * Unqueue the futex_q and determine which it was. | |
3258 | */ | |
2e12978a | 3259 | plist_del(&q->list, &hb->chain); |
11d4616b | 3260 | hb_waiters_dec(hb); |
52400ba9 | 3261 | |
d58e6576 | 3262 | /* Handle spurious wakeups gracefully */ |
11df6ddd | 3263 | ret = -EWOULDBLOCK; |
52400ba9 DH |
3264 | if (timeout && !timeout->task) |
3265 | ret = -ETIMEDOUT; | |
d58e6576 | 3266 | else if (signal_pending(current)) |
1c840c14 | 3267 | ret = -ERESTARTNOINTR; |
52400ba9 DH |
3268 | } |
3269 | return ret; | |
3270 | } | |
3271 | ||
3272 | /** | |
3273 | * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2 | |
56ec1607 | 3274 | * @uaddr: the futex we initially wait on (non-pi) |
b41277dc | 3275 | * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be |
ab51fbab | 3276 | * the same type, no requeueing from private to shared, etc. |
52400ba9 DH |
3277 | * @val: the expected value of uaddr |
3278 | * @abs_time: absolute timeout | |
56ec1607 | 3279 | * @bitset: 32 bit wakeup bitset set by userspace, defaults to all |
52400ba9 DH |
3280 | * @uaddr2: the pi futex we will take prior to returning to user-space |
3281 | * | |
3282 | * The caller will wait on uaddr and will be requeued by futex_requeue() to | |
6f7b0a2a DH |
3283 | * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake |
3284 | * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to | |
3285 | * userspace. This ensures the rt_mutex maintains an owner when it has waiters; | |
3286 | * without one, the pi logic would not know which task to boost/deboost, if | |
3287 | * there was a need to. | |
52400ba9 DH |
3288 | * |
3289 | * We call schedule in futex_wait_queue_me() when we enqueue and return there | |
6c23cbbd | 3290 | * via the following-- |
52400ba9 | 3291 | * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue() |
cc6db4e6 DH |
3292 | * 2) wakeup on uaddr2 after a requeue |
3293 | * 3) signal | |
3294 | * 4) timeout | |
52400ba9 | 3295 | * |
cc6db4e6 | 3296 | * If 3, cleanup and return -ERESTARTNOINTR. |
52400ba9 DH |
3297 | * |
3298 | * If 2, we may then block on trying to take the rt_mutex and return via: | |
3299 | * 5) successful lock | |
3300 | * 6) signal | |
3301 | * 7) timeout | |
3302 | * 8) other lock acquisition failure | |
3303 | * | |
cc6db4e6 | 3304 | * If 6, return -EWOULDBLOCK (restarting the syscall would do the same). |
52400ba9 DH |
3305 | * |
3306 | * If 4 or 7, we cleanup and return with -ETIMEDOUT. | |
3307 | * | |
6c23cbbd | 3308 | * Return: |
7b4ff1ad MCC |
3309 | * - 0 - On success; |
3310 | * - <0 - On error | |
52400ba9 | 3311 | */ |
b41277dc | 3312 | static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags, |
52400ba9 | 3313 | u32 val, ktime_t *abs_time, u32 bitset, |
b41277dc | 3314 | u32 __user *uaddr2) |
52400ba9 | 3315 | { |
5ca584d9 | 3316 | struct hrtimer_sleeper timeout, *to; |
16ffa12d | 3317 | struct futex_pi_state *pi_state = NULL; |
52400ba9 | 3318 | struct rt_mutex_waiter rt_waiter; |
52400ba9 | 3319 | struct futex_hash_bucket *hb; |
5bdb05f9 DH |
3320 | union futex_key key2 = FUTEX_KEY_INIT; |
3321 | struct futex_q q = futex_q_init; | |
52400ba9 | 3322 | int res, ret; |
52400ba9 | 3323 | |
bc2eecd7 NP |
3324 | if (!IS_ENABLED(CONFIG_FUTEX_PI)) |
3325 | return -ENOSYS; | |
3326 | ||
6f7b0a2a DH |
3327 | if (uaddr == uaddr2) |
3328 | return -EINVAL; | |
3329 | ||
52400ba9 DH |
3330 | if (!bitset) |
3331 | return -EINVAL; | |
3332 | ||
5ca584d9 WL |
3333 | to = futex_setup_timer(abs_time, &timeout, flags, |
3334 | current->timer_slack_ns); | |
52400ba9 DH |
3335 | |
3336 | /* | |
3337 | * The waiter is allocated on our stack, manipulated by the requeue | |
3338 | * code while we sleep on uaddr. | |
3339 | */ | |
50809358 | 3340 | rt_mutex_init_waiter(&rt_waiter); |
52400ba9 | 3341 | |
96d4f267 | 3342 | ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE); |
52400ba9 DH |
3343 | if (unlikely(ret != 0)) |
3344 | goto out; | |
3345 | ||
84bc4af5 DH |
3346 | q.bitset = bitset; |
3347 | q.rt_waiter = &rt_waiter; | |
3348 | q.requeue_pi_key = &key2; | |
3349 | ||
7ada876a DH |
3350 | /* |
3351 | * Prepare to wait on uaddr. On success, increments q.key (key1) ref | |
3352 | * count. | |
3353 | */ | |
b41277dc | 3354 | ret = futex_wait_setup(uaddr, val, flags, &q, &hb); |
c8b15a70 TG |
3355 | if (ret) |
3356 | goto out_key2; | |
52400ba9 | 3357 | |
e9c243a5 TG |
3358 | /* |
3359 | * The check above which compares uaddrs is not sufficient for | |
3360 | * shared futexes. We need to compare the keys: | |
3361 | */ | |
3362 | if (match_futex(&q.key, &key2)) { | |
13c42c2f | 3363 | queue_unlock(hb); |
e9c243a5 TG |
3364 | ret = -EINVAL; |
3365 | goto out_put_keys; | |
3366 | } | |
3367 | ||
52400ba9 | 3368 | /* Queue the futex_q, drop the hb lock, wait for wakeup. */ |
f1a11e05 | 3369 | futex_wait_queue_me(hb, &q, to); |
52400ba9 DH |
3370 | |
3371 | spin_lock(&hb->lock); | |
3372 | ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to); | |
3373 | spin_unlock(&hb->lock); | |
3374 | if (ret) | |
3375 | goto out_put_keys; | |
3376 | ||
3377 | /* | |
3378 | * In order for us to be here, we know our q.key == key2, and since | |
3379 | * we took the hb->lock above, we also know that futex_requeue() has | |
3380 | * completed and we no longer have to concern ourselves with a wakeup | |
7ada876a DH |
3381 | * race with the atomic proxy lock acquisition by the requeue code. The |
3382 | * futex_requeue dropped our key1 reference and incremented our key2 | |
3383 | * reference count. | |
52400ba9 DH |
3384 | */ |
3385 | ||
3386 | /* Check if the requeue code acquired the second futex for us. */ | |
3387 | if (!q.rt_waiter) { | |
3388 | /* | |
3389 | * Got the lock. We might not be the anticipated owner if we | |
3390 | * did a lock-steal - fix up the PI-state in that case. | |
3391 | */ | |
3392 | if (q.pi_state && (q.pi_state->owner != current)) { | |
3393 | spin_lock(q.lock_ptr); | |
ae791a2d | 3394 | ret = fixup_pi_state_owner(uaddr2, &q, current); |
16ffa12d PZ |
3395 | if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) { |
3396 | pi_state = q.pi_state; | |
3397 | get_pi_state(pi_state); | |
3398 | } | |
fb75a428 TG |
3399 | /* |
3400 | * Drop the reference to the pi state which | |
3401 | * the requeue_pi() code acquired for us. | |
3402 | */ | |
29e9ee5d | 3403 | put_pi_state(q.pi_state); |
52400ba9 DH |
3404 | spin_unlock(q.lock_ptr); |
3405 | } | |
3406 | } else { | |
c236c8e9 PZ |
3407 | struct rt_mutex *pi_mutex; |
3408 | ||
52400ba9 DH |
3409 | /* |
3410 | * We have been woken up by futex_unlock_pi(), a timeout, or a | |
3411 | * signal. futex_unlock_pi() will not destroy the lock_ptr nor | |
3412 | * the pi_state. | |
3413 | */ | |
f27071cb | 3414 | WARN_ON(!q.pi_state); |
52400ba9 | 3415 | pi_mutex = &q.pi_state->pi_mutex; |
38d589f2 | 3416 | ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter); |
52400ba9 DH |
3417 | |
3418 | spin_lock(q.lock_ptr); | |
38d589f2 PZ |
3419 | if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter)) |
3420 | ret = 0; | |
3421 | ||
3422 | debug_rt_mutex_free_waiter(&rt_waiter); | |
52400ba9 DH |
3423 | /* |
3424 | * Fixup the pi_state owner and possibly acquire the lock if we | |
3425 | * haven't already. | |
3426 | */ | |
ae791a2d | 3427 | res = fixup_owner(uaddr2, &q, !ret); |
52400ba9 DH |
3428 | /* |
3429 | * If fixup_owner() returned an error, proprogate that. If it | |
56ec1607 | 3430 | * acquired the lock, clear -ETIMEDOUT or -EINTR. |
52400ba9 DH |
3431 | */ |
3432 | if (res) | |
3433 | ret = (res < 0) ? res : 0; | |
3434 | ||
c236c8e9 PZ |
3435 | /* |
3436 | * If fixup_pi_state_owner() faulted and was unable to handle | |
3437 | * the fault, unlock the rt_mutex and return the fault to | |
3438 | * userspace. | |
3439 | */ | |
16ffa12d PZ |
3440 | if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) { |
3441 | pi_state = q.pi_state; | |
3442 | get_pi_state(pi_state); | |
3443 | } | |
c236c8e9 | 3444 | |
52400ba9 DH |
3445 | /* Unqueue and drop the lock. */ |
3446 | unqueue_me_pi(&q); | |
3447 | } | |
3448 | ||
16ffa12d PZ |
3449 | if (pi_state) { |
3450 | rt_mutex_futex_unlock(&pi_state->pi_mutex); | |
3451 | put_pi_state(pi_state); | |
3452 | } | |
3453 | ||
c236c8e9 | 3454 | if (ret == -EINTR) { |
52400ba9 | 3455 | /* |
cc6db4e6 DH |
3456 | * We've already been requeued, but cannot restart by calling |
3457 | * futex_lock_pi() directly. We could restart this syscall, but | |
3458 | * it would detect that the user space "val" changed and return | |
3459 | * -EWOULDBLOCK. Save the overhead of the restart and return | |
3460 | * -EWOULDBLOCK directly. | |
52400ba9 | 3461 | */ |
2070887f | 3462 | ret = -EWOULDBLOCK; |
52400ba9 DH |
3463 | } |
3464 | ||
3465 | out_put_keys: | |
ae791a2d | 3466 | put_futex_key(&q.key); |
c8b15a70 | 3467 | out_key2: |
ae791a2d | 3468 | put_futex_key(&key2); |
52400ba9 DH |
3469 | |
3470 | out: | |
3471 | if (to) { | |
3472 | hrtimer_cancel(&to->timer); | |
3473 | destroy_hrtimer_on_stack(&to->timer); | |
3474 | } | |
3475 | return ret; | |
3476 | } | |
3477 | ||
0771dfef IM |
3478 | /* |
3479 | * Support for robust futexes: the kernel cleans up held futexes at | |
3480 | * thread exit time. | |
3481 | * | |
3482 | * Implementation: user-space maintains a per-thread list of locks it | |
3483 | * is holding. Upon do_exit(), the kernel carefully walks this list, | |
3484 | * and marks all locks that are owned by this thread with the | |
c87e2837 | 3485 | * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is |
0771dfef IM |
3486 | * always manipulated with the lock held, so the list is private and |
3487 | * per-thread. Userspace also maintains a per-thread 'list_op_pending' | |
3488 | * field, to allow the kernel to clean up if the thread dies after | |
3489 | * acquiring the lock, but just before it could have added itself to | |
3490 | * the list. There can only be one such pending lock. | |
3491 | */ | |
3492 | ||
3493 | /** | |
d96ee56c DH |
3494 | * sys_set_robust_list() - Set the robust-futex list head of a task |
3495 | * @head: pointer to the list-head | |
3496 | * @len: length of the list-head, as userspace expects | |
0771dfef | 3497 | */ |
836f92ad HC |
3498 | SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head, |
3499 | size_t, len) | |
0771dfef | 3500 | { |
a0c1e907 TG |
3501 | if (!futex_cmpxchg_enabled) |
3502 | return -ENOSYS; | |
0771dfef IM |
3503 | /* |
3504 | * The kernel knows only one size for now: | |
3505 | */ | |
3506 | if (unlikely(len != sizeof(*head))) | |
3507 | return -EINVAL; | |
3508 | ||
3509 | current->robust_list = head; | |
3510 | ||
3511 | return 0; | |
3512 | } | |
3513 | ||
3514 | /** | |
d96ee56c DH |
3515 | * sys_get_robust_list() - Get the robust-futex list head of a task |
3516 | * @pid: pid of the process [zero for current task] | |
3517 | * @head_ptr: pointer to a list-head pointer, the kernel fills it in | |
3518 | * @len_ptr: pointer to a length field, the kernel fills in the header size | |
0771dfef | 3519 | */ |
836f92ad HC |
3520 | SYSCALL_DEFINE3(get_robust_list, int, pid, |
3521 | struct robust_list_head __user * __user *, head_ptr, | |
3522 | size_t __user *, len_ptr) | |
0771dfef | 3523 | { |
ba46df98 | 3524 | struct robust_list_head __user *head; |
0771dfef | 3525 | unsigned long ret; |
bdbb776f | 3526 | struct task_struct *p; |
0771dfef | 3527 | |
a0c1e907 TG |
3528 | if (!futex_cmpxchg_enabled) |
3529 | return -ENOSYS; | |
3530 | ||
bdbb776f KC |
3531 | rcu_read_lock(); |
3532 | ||
3533 | ret = -ESRCH; | |
0771dfef | 3534 | if (!pid) |
bdbb776f | 3535 | p = current; |
0771dfef | 3536 | else { |
228ebcbe | 3537 | p = find_task_by_vpid(pid); |
0771dfef IM |
3538 | if (!p) |
3539 | goto err_unlock; | |
0771dfef IM |
3540 | } |
3541 | ||
bdbb776f | 3542 | ret = -EPERM; |
caaee623 | 3543 | if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS)) |
bdbb776f KC |
3544 | goto err_unlock; |
3545 | ||
3546 | head = p->robust_list; | |
3547 | rcu_read_unlock(); | |
3548 | ||
0771dfef IM |
3549 | if (put_user(sizeof(*head), len_ptr)) |
3550 | return -EFAULT; | |
3551 | return put_user(head, head_ptr); | |
3552 | ||
3553 | err_unlock: | |
aaa2a97e | 3554 | rcu_read_unlock(); |
0771dfef IM |
3555 | |
3556 | return ret; | |
3557 | } | |
3558 | ||
ca16d5be YT |
3559 | /* Constants for the pending_op argument of handle_futex_death */ |
3560 | #define HANDLE_DEATH_PENDING true | |
3561 | #define HANDLE_DEATH_LIST false | |
3562 | ||
0771dfef IM |
3563 | /* |
3564 | * Process a futex-list entry, check whether it's owned by the | |
3565 | * dying task, and do notification if so: | |
3566 | */ | |
ca16d5be YT |
3567 | static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, |
3568 | bool pi, bool pending_op) | |
0771dfef | 3569 | { |
7cfdaf38 | 3570 | u32 uval, uninitialized_var(nval), mval; |
6b4f4bc9 | 3571 | int err; |
0771dfef | 3572 | |
5a07168d CJ |
3573 | /* Futex address must be 32bit aligned */ |
3574 | if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0) | |
3575 | return -1; | |
3576 | ||
8f17d3a5 IM |
3577 | retry: |
3578 | if (get_user(uval, uaddr)) | |
0771dfef IM |
3579 | return -1; |
3580 | ||
ca16d5be YT |
3581 | /* |
3582 | * Special case for regular (non PI) futexes. The unlock path in | |
3583 | * user space has two race scenarios: | |
3584 | * | |
3585 | * 1. The unlock path releases the user space futex value and | |
3586 | * before it can execute the futex() syscall to wake up | |
3587 | * waiters it is killed. | |
3588 | * | |
3589 | * 2. A woken up waiter is killed before it can acquire the | |
3590 | * futex in user space. | |
3591 | * | |
3592 | * In both cases the TID validation below prevents a wakeup of | |
3593 | * potential waiters which can cause these waiters to block | |
3594 | * forever. | |
3595 | * | |
3596 | * In both cases the following conditions are met: | |
3597 | * | |
3598 | * 1) task->robust_list->list_op_pending != NULL | |
3599 | * @pending_op == true | |
3600 | * 2) User space futex value == 0 | |
3601 | * 3) Regular futex: @pi == false | |
3602 | * | |
3603 | * If these conditions are met, it is safe to attempt waking up a | |
3604 | * potential waiter without touching the user space futex value and | |
3605 | * trying to set the OWNER_DIED bit. The user space futex value is | |
3606 | * uncontended and the rest of the user space mutex state is | |
3607 | * consistent, so a woken waiter will just take over the | |
3608 | * uncontended futex. Setting the OWNER_DIED bit would create | |
3609 | * inconsistent state and malfunction of the user space owner died | |
3610 | * handling. | |
3611 | */ | |
3612 | if (pending_op && !pi && !uval) { | |
3613 | futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY); | |
3614 | return 0; | |
3615 | } | |
3616 | ||
6b4f4bc9 WD |
3617 | if ((uval & FUTEX_TID_MASK) != task_pid_vnr(curr)) |
3618 | return 0; | |
3619 | ||
3620 | /* | |
3621 | * Ok, this dying thread is truly holding a futex | |
3622 | * of interest. Set the OWNER_DIED bit atomically | |
3623 | * via cmpxchg, and if the value had FUTEX_WAITERS | |
3624 | * set, wake up a waiter (if any). (We have to do a | |
3625 | * futex_wake() even if OWNER_DIED is already set - | |
3626 | * to handle the rare but possible case of recursive | |
3627 | * thread-death.) The rest of the cleanup is done in | |
3628 | * userspace. | |
3629 | */ | |
3630 | mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED; | |
3631 | ||
3632 | /* | |
3633 | * We are not holding a lock here, but we want to have | |
3634 | * the pagefault_disable/enable() protection because | |
3635 | * we want to handle the fault gracefully. If the | |
3636 | * access fails we try to fault in the futex with R/W | |
3637 | * verification via get_user_pages. get_user() above | |
3638 | * does not guarantee R/W access. If that fails we | |
3639 | * give up and leave the futex locked. | |
3640 | */ | |
3641 | if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) { | |
3642 | switch (err) { | |
3643 | case -EFAULT: | |
6e0aa9f8 TG |
3644 | if (fault_in_user_writeable(uaddr)) |
3645 | return -1; | |
3646 | goto retry; | |
6b4f4bc9 WD |
3647 | |
3648 | case -EAGAIN: | |
3649 | cond_resched(); | |
8f17d3a5 | 3650 | goto retry; |
0771dfef | 3651 | |
6b4f4bc9 WD |
3652 | default: |
3653 | WARN_ON_ONCE(1); | |
3654 | return err; | |
3655 | } | |
0771dfef | 3656 | } |
6b4f4bc9 WD |
3657 | |
3658 | if (nval != uval) | |
3659 | goto retry; | |
3660 | ||
3661 | /* | |
3662 | * Wake robust non-PI futexes here. The wakeup of | |
3663 | * PI futexes happens in exit_pi_state(): | |
3664 | */ | |
3665 | if (!pi && (uval & FUTEX_WAITERS)) | |
3666 | futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY); | |
3667 | ||
0771dfef IM |
3668 | return 0; |
3669 | } | |
3670 | ||
e3f2ddea IM |
3671 | /* |
3672 | * Fetch a robust-list pointer. Bit 0 signals PI futexes: | |
3673 | */ | |
3674 | static inline int fetch_robust_entry(struct robust_list __user **entry, | |
ba46df98 | 3675 | struct robust_list __user * __user *head, |
1dcc41bb | 3676 | unsigned int *pi) |
e3f2ddea IM |
3677 | { |
3678 | unsigned long uentry; | |
3679 | ||
ba46df98 | 3680 | if (get_user(uentry, (unsigned long __user *)head)) |
e3f2ddea IM |
3681 | return -EFAULT; |
3682 | ||
ba46df98 | 3683 | *entry = (void __user *)(uentry & ~1UL); |
e3f2ddea IM |
3684 | *pi = uentry & 1; |
3685 | ||
3686 | return 0; | |
3687 | } | |
3688 | ||
0771dfef IM |
3689 | /* |
3690 | * Walk curr->robust_list (very carefully, it's a userspace list!) | |
3691 | * and mark any locks found there dead, and notify any waiters. | |
3692 | * | |
3693 | * We silently return on any sign of list-walking problem. | |
3694 | */ | |
ba31c1a4 | 3695 | static void exit_robust_list(struct task_struct *curr) |
0771dfef IM |
3696 | { |
3697 | struct robust_list_head __user *head = curr->robust_list; | |
9f96cb1e | 3698 | struct robust_list __user *entry, *next_entry, *pending; |
4c115e95 DH |
3699 | unsigned int limit = ROBUST_LIST_LIMIT, pi, pip; |
3700 | unsigned int uninitialized_var(next_pi); | |
0771dfef | 3701 | unsigned long futex_offset; |
9f96cb1e | 3702 | int rc; |
0771dfef | 3703 | |
a0c1e907 TG |
3704 | if (!futex_cmpxchg_enabled) |
3705 | return; | |
3706 | ||
0771dfef IM |
3707 | /* |
3708 | * Fetch the list head (which was registered earlier, via | |
3709 | * sys_set_robust_list()): | |
3710 | */ | |
e3f2ddea | 3711 | if (fetch_robust_entry(&entry, &head->list.next, &pi)) |
0771dfef IM |
3712 | return; |
3713 | /* | |
3714 | * Fetch the relative futex offset: | |
3715 | */ | |
3716 | if (get_user(futex_offset, &head->futex_offset)) | |
3717 | return; | |
3718 | /* | |
3719 | * Fetch any possibly pending lock-add first, and handle it | |
3720 | * if it exists: | |
3721 | */ | |
e3f2ddea | 3722 | if (fetch_robust_entry(&pending, &head->list_op_pending, &pip)) |
0771dfef | 3723 | return; |
e3f2ddea | 3724 | |
9f96cb1e | 3725 | next_entry = NULL; /* avoid warning with gcc */ |
0771dfef | 3726 | while (entry != &head->list) { |
9f96cb1e MS |
3727 | /* |
3728 | * Fetch the next entry in the list before calling | |
3729 | * handle_futex_death: | |
3730 | */ | |
3731 | rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi); | |
0771dfef IM |
3732 | /* |
3733 | * A pending lock might already be on the list, so | |
c87e2837 | 3734 | * don't process it twice: |
0771dfef | 3735 | */ |
ca16d5be | 3736 | if (entry != pending) { |
ba46df98 | 3737 | if (handle_futex_death((void __user *)entry + futex_offset, |
ca16d5be | 3738 | curr, pi, HANDLE_DEATH_LIST)) |
0771dfef | 3739 | return; |
ca16d5be | 3740 | } |
9f96cb1e | 3741 | if (rc) |
0771dfef | 3742 | return; |
9f96cb1e MS |
3743 | entry = next_entry; |
3744 | pi = next_pi; | |
0771dfef IM |
3745 | /* |
3746 | * Avoid excessively long or circular lists: | |
3747 | */ | |
3748 | if (!--limit) | |
3749 | break; | |
3750 | ||
3751 | cond_resched(); | |
3752 | } | |
9f96cb1e | 3753 | |
ca16d5be | 3754 | if (pending) { |
9f96cb1e | 3755 | handle_futex_death((void __user *)pending + futex_offset, |
ca16d5be YT |
3756 | curr, pip, HANDLE_DEATH_PENDING); |
3757 | } | |
0771dfef IM |
3758 | } |
3759 | ||
af8cbda2 | 3760 | static void futex_cleanup(struct task_struct *tsk) |
ba31c1a4 TG |
3761 | { |
3762 | if (unlikely(tsk->robust_list)) { | |
3763 | exit_robust_list(tsk); | |
3764 | tsk->robust_list = NULL; | |
3765 | } | |
3766 | ||
3767 | #ifdef CONFIG_COMPAT | |
3768 | if (unlikely(tsk->compat_robust_list)) { | |
3769 | compat_exit_robust_list(tsk); | |
3770 | tsk->compat_robust_list = NULL; | |
3771 | } | |
3772 | #endif | |
3773 | ||
3774 | if (unlikely(!list_empty(&tsk->pi_state_list))) | |
3775 | exit_pi_state_list(tsk); | |
3776 | } | |
3777 | ||
18f69438 TG |
3778 | /** |
3779 | * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD | |
3780 | * @tsk: task to set the state on | |
3781 | * | |
3782 | * Set the futex exit state of the task lockless. The futex waiter code | |
3783 | * observes that state when a task is exiting and loops until the task has | |
3784 | * actually finished the futex cleanup. The worst case for this is that the | |
3785 | * waiter runs through the wait loop until the state becomes visible. | |
3786 | * | |
3787 | * This is called from the recursive fault handling path in do_exit(). | |
3788 | * | |
3789 | * This is best effort. Either the futex exit code has run already or | |
3790 | * not. If the OWNER_DIED bit has been set on the futex then the waiter can | |
3791 | * take it over. If not, the problem is pushed back to user space. If the | |
3792 | * futex exit code did not run yet, then an already queued waiter might | |
3793 | * block forever, but there is nothing which can be done about that. | |
3794 | */ | |
3795 | void futex_exit_recursive(struct task_struct *tsk) | |
3796 | { | |
3f186d97 TG |
3797 | /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */ |
3798 | if (tsk->futex_state == FUTEX_STATE_EXITING) | |
3799 | mutex_unlock(&tsk->futex_exit_mutex); | |
18f69438 TG |
3800 | tsk->futex_state = FUTEX_STATE_DEAD; |
3801 | } | |
3802 | ||
af8cbda2 | 3803 | static void futex_cleanup_begin(struct task_struct *tsk) |
150d7158 | 3804 | { |
3f186d97 TG |
3805 | /* |
3806 | * Prevent various race issues against a concurrent incoming waiter | |
3807 | * including live locks by forcing the waiter to block on | |
3808 | * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in | |
3809 | * attach_to_pi_owner(). | |
3810 | */ | |
3811 | mutex_lock(&tsk->futex_exit_mutex); | |
3812 | ||
18f69438 | 3813 | /* |
4a8e991b TG |
3814 | * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock. |
3815 | * | |
3816 | * This ensures that all subsequent checks of tsk->futex_state in | |
3817 | * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with | |
3818 | * tsk->pi_lock held. | |
3819 | * | |
3820 | * It guarantees also that a pi_state which was queued right before | |
3821 | * the state change under tsk->pi_lock by a concurrent waiter must | |
3822 | * be observed in exit_pi_state_list(). | |
18f69438 TG |
3823 | */ |
3824 | raw_spin_lock_irq(&tsk->pi_lock); | |
4a8e991b | 3825 | tsk->futex_state = FUTEX_STATE_EXITING; |
18f69438 | 3826 | raw_spin_unlock_irq(&tsk->pi_lock); |
af8cbda2 | 3827 | } |
18f69438 | 3828 | |
af8cbda2 TG |
3829 | static void futex_cleanup_end(struct task_struct *tsk, int state) |
3830 | { | |
3831 | /* | |
3832 | * Lockless store. The only side effect is that an observer might | |
3833 | * take another loop until it becomes visible. | |
3834 | */ | |
3835 | tsk->futex_state = state; | |
3f186d97 TG |
3836 | /* |
3837 | * Drop the exit protection. This unblocks waiters which observed | |
3838 | * FUTEX_STATE_EXITING to reevaluate the state. | |
3839 | */ | |
3840 | mutex_unlock(&tsk->futex_exit_mutex); | |
af8cbda2 | 3841 | } |
18f69438 | 3842 | |
af8cbda2 TG |
3843 | void futex_exec_release(struct task_struct *tsk) |
3844 | { | |
3845 | /* | |
3846 | * The state handling is done for consistency, but in the case of | |
3847 | * exec() there is no way to prevent futher damage as the PID stays | |
3848 | * the same. But for the unlikely and arguably buggy case that a | |
3849 | * futex is held on exec(), this provides at least as much state | |
3850 | * consistency protection which is possible. | |
3851 | */ | |
3852 | futex_cleanup_begin(tsk); | |
3853 | futex_cleanup(tsk); | |
3854 | /* | |
3855 | * Reset the state to FUTEX_STATE_OK. The task is alive and about | |
3856 | * exec a new binary. | |
3857 | */ | |
3858 | futex_cleanup_end(tsk, FUTEX_STATE_OK); | |
3859 | } | |
3860 | ||
3861 | void futex_exit_release(struct task_struct *tsk) | |
3862 | { | |
3863 | futex_cleanup_begin(tsk); | |
3864 | futex_cleanup(tsk); | |
3865 | futex_cleanup_end(tsk, FUTEX_STATE_DEAD); | |
150d7158 TG |
3866 | } |
3867 | ||
c19384b5 | 3868 | long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout, |
e2970f2f | 3869 | u32 __user *uaddr2, u32 val2, u32 val3) |
1da177e4 | 3870 | { |
81b40539 | 3871 | int cmd = op & FUTEX_CMD_MASK; |
b41277dc | 3872 | unsigned int flags = 0; |
34f01cc1 ED |
3873 | |
3874 | if (!(op & FUTEX_PRIVATE_FLAG)) | |
b41277dc | 3875 | flags |= FLAGS_SHARED; |
1da177e4 | 3876 | |
b41277dc DH |
3877 | if (op & FUTEX_CLOCK_REALTIME) { |
3878 | flags |= FLAGS_CLOCKRT; | |
337f1304 DH |
3879 | if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \ |
3880 | cmd != FUTEX_WAIT_REQUEUE_PI) | |
b41277dc DH |
3881 | return -ENOSYS; |
3882 | } | |
1da177e4 | 3883 | |
59263b51 TG |
3884 | switch (cmd) { |
3885 | case FUTEX_LOCK_PI: | |
3886 | case FUTEX_UNLOCK_PI: | |
3887 | case FUTEX_TRYLOCK_PI: | |
3888 | case FUTEX_WAIT_REQUEUE_PI: | |
3889 | case FUTEX_CMP_REQUEUE_PI: | |
3890 | if (!futex_cmpxchg_enabled) | |
3891 | return -ENOSYS; | |
3892 | } | |
3893 | ||
34f01cc1 | 3894 | switch (cmd) { |
1da177e4 | 3895 | case FUTEX_WAIT: |
cd689985 | 3896 | val3 = FUTEX_BITSET_MATCH_ANY; |
b639186f | 3897 | /* fall through */ |
cd689985 | 3898 | case FUTEX_WAIT_BITSET: |
81b40539 | 3899 | return futex_wait(uaddr, flags, val, timeout, val3); |
1da177e4 | 3900 | case FUTEX_WAKE: |
cd689985 | 3901 | val3 = FUTEX_BITSET_MATCH_ANY; |
b639186f | 3902 | /* fall through */ |
cd689985 | 3903 | case FUTEX_WAKE_BITSET: |
81b40539 | 3904 | return futex_wake(uaddr, flags, val, val3); |
1da177e4 | 3905 | case FUTEX_REQUEUE: |
81b40539 | 3906 | return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0); |
1da177e4 | 3907 | case FUTEX_CMP_REQUEUE: |
81b40539 | 3908 | return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0); |
4732efbe | 3909 | case FUTEX_WAKE_OP: |
81b40539 | 3910 | return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3); |
c87e2837 | 3911 | case FUTEX_LOCK_PI: |
996636dd | 3912 | return futex_lock_pi(uaddr, flags, timeout, 0); |
c87e2837 | 3913 | case FUTEX_UNLOCK_PI: |
81b40539 | 3914 | return futex_unlock_pi(uaddr, flags); |
c87e2837 | 3915 | case FUTEX_TRYLOCK_PI: |
996636dd | 3916 | return futex_lock_pi(uaddr, flags, NULL, 1); |
52400ba9 DH |
3917 | case FUTEX_WAIT_REQUEUE_PI: |
3918 | val3 = FUTEX_BITSET_MATCH_ANY; | |
81b40539 TG |
3919 | return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3, |
3920 | uaddr2); | |
52400ba9 | 3921 | case FUTEX_CMP_REQUEUE_PI: |
81b40539 | 3922 | return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1); |
1da177e4 | 3923 | } |
81b40539 | 3924 | return -ENOSYS; |
1da177e4 LT |
3925 | } |
3926 | ||
3927 | ||
17da2bd9 | 3928 | SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val, |
bec2f7cb | 3929 | struct __kernel_timespec __user *, utime, u32 __user *, uaddr2, |
17da2bd9 | 3930 | u32, val3) |
1da177e4 | 3931 | { |
bec2f7cb | 3932 | struct timespec64 ts; |
c19384b5 | 3933 | ktime_t t, *tp = NULL; |
e2970f2f | 3934 | u32 val2 = 0; |
34f01cc1 | 3935 | int cmd = op & FUTEX_CMD_MASK; |
1da177e4 | 3936 | |
cd689985 | 3937 | if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI || |
52400ba9 DH |
3938 | cmd == FUTEX_WAIT_BITSET || |
3939 | cmd == FUTEX_WAIT_REQUEUE_PI)) { | |
ab51fbab DB |
3940 | if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG)))) |
3941 | return -EFAULT; | |
bec2f7cb | 3942 | if (get_timespec64(&ts, utime)) |
1da177e4 | 3943 | return -EFAULT; |
bec2f7cb | 3944 | if (!timespec64_valid(&ts)) |
9741ef96 | 3945 | return -EINVAL; |
c19384b5 | 3946 | |
bec2f7cb | 3947 | t = timespec64_to_ktime(ts); |
34f01cc1 | 3948 | if (cmd == FUTEX_WAIT) |
5a7780e7 | 3949 | t = ktime_add_safe(ktime_get(), t); |
c19384b5 | 3950 | tp = &t; |
1da177e4 LT |
3951 | } |
3952 | /* | |
52400ba9 | 3953 | * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*. |
f54f0986 | 3954 | * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP. |
1da177e4 | 3955 | */ |
f54f0986 | 3956 | if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE || |
ba9c22f2 | 3957 | cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP) |
e2970f2f | 3958 | val2 = (u32) (unsigned long) utime; |
1da177e4 | 3959 | |
c19384b5 | 3960 | return do_futex(uaddr, op, val, tp, uaddr2, val2, val3); |
1da177e4 LT |
3961 | } |
3962 | ||
04e7712f AB |
3963 | #ifdef CONFIG_COMPAT |
3964 | /* | |
3965 | * Fetch a robust-list pointer. Bit 0 signals PI futexes: | |
3966 | */ | |
3967 | static inline int | |
3968 | compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry, | |
3969 | compat_uptr_t __user *head, unsigned int *pi) | |
3970 | { | |
3971 | if (get_user(*uentry, head)) | |
3972 | return -EFAULT; | |
3973 | ||
3974 | *entry = compat_ptr((*uentry) & ~1); | |
3975 | *pi = (unsigned int)(*uentry) & 1; | |
3976 | ||
3977 | return 0; | |
3978 | } | |
3979 | ||
3980 | static void __user *futex_uaddr(struct robust_list __user *entry, | |
3981 | compat_long_t futex_offset) | |
3982 | { | |
3983 | compat_uptr_t base = ptr_to_compat(entry); | |
3984 | void __user *uaddr = compat_ptr(base + futex_offset); | |
3985 | ||
3986 | return uaddr; | |
3987 | } | |
3988 | ||
3989 | /* | |
3990 | * Walk curr->robust_list (very carefully, it's a userspace list!) | |
3991 | * and mark any locks found there dead, and notify any waiters. | |
3992 | * | |
3993 | * We silently return on any sign of list-walking problem. | |
3994 | */ | |
ba31c1a4 | 3995 | static void compat_exit_robust_list(struct task_struct *curr) |
04e7712f AB |
3996 | { |
3997 | struct compat_robust_list_head __user *head = curr->compat_robust_list; | |
3998 | struct robust_list __user *entry, *next_entry, *pending; | |
3999 | unsigned int limit = ROBUST_LIST_LIMIT, pi, pip; | |
4000 | unsigned int uninitialized_var(next_pi); | |
4001 | compat_uptr_t uentry, next_uentry, upending; | |
4002 | compat_long_t futex_offset; | |
4003 | int rc; | |
4004 | ||
4005 | if (!futex_cmpxchg_enabled) | |
4006 | return; | |
4007 | ||
4008 | /* | |
4009 | * Fetch the list head (which was registered earlier, via | |
4010 | * sys_set_robust_list()): | |
4011 | */ | |
4012 | if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi)) | |
4013 | return; | |
4014 | /* | |
4015 | * Fetch the relative futex offset: | |
4016 | */ | |
4017 | if (get_user(futex_offset, &head->futex_offset)) | |
4018 | return; | |
4019 | /* | |
4020 | * Fetch any possibly pending lock-add first, and handle it | |
4021 | * if it exists: | |
4022 | */ | |
4023 | if (compat_fetch_robust_entry(&upending, &pending, | |
4024 | &head->list_op_pending, &pip)) | |
4025 | return; | |
4026 | ||
4027 | next_entry = NULL; /* avoid warning with gcc */ | |
4028 | while (entry != (struct robust_list __user *) &head->list) { | |
4029 | /* | |
4030 | * Fetch the next entry in the list before calling | |
4031 | * handle_futex_death: | |
4032 | */ | |
4033 | rc = compat_fetch_robust_entry(&next_uentry, &next_entry, | |
4034 | (compat_uptr_t __user *)&entry->next, &next_pi); | |
4035 | /* | |
4036 | * A pending lock might already be on the list, so | |
4037 | * dont process it twice: | |
4038 | */ | |
4039 | if (entry != pending) { | |
4040 | void __user *uaddr = futex_uaddr(entry, futex_offset); | |
4041 | ||
ca16d5be YT |
4042 | if (handle_futex_death(uaddr, curr, pi, |
4043 | HANDLE_DEATH_LIST)) | |
04e7712f AB |
4044 | return; |
4045 | } | |
4046 | if (rc) | |
4047 | return; | |
4048 | uentry = next_uentry; | |
4049 | entry = next_entry; | |
4050 | pi = next_pi; | |
4051 | /* | |
4052 | * Avoid excessively long or circular lists: | |
4053 | */ | |
4054 | if (!--limit) | |
4055 | break; | |
4056 | ||
4057 | cond_resched(); | |
4058 | } | |
4059 | if (pending) { | |
4060 | void __user *uaddr = futex_uaddr(pending, futex_offset); | |
4061 | ||
ca16d5be | 4062 | handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING); |
04e7712f AB |
4063 | } |
4064 | } | |
4065 | ||
4066 | COMPAT_SYSCALL_DEFINE2(set_robust_list, | |
4067 | struct compat_robust_list_head __user *, head, | |
4068 | compat_size_t, len) | |
4069 | { | |
4070 | if (!futex_cmpxchg_enabled) | |
4071 | return -ENOSYS; | |
4072 | ||
4073 | if (unlikely(len != sizeof(*head))) | |
4074 | return -EINVAL; | |
4075 | ||
4076 | current->compat_robust_list = head; | |
4077 | ||
4078 | return 0; | |
4079 | } | |
4080 | ||
4081 | COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid, | |
4082 | compat_uptr_t __user *, head_ptr, | |
4083 | compat_size_t __user *, len_ptr) | |
4084 | { | |
4085 | struct compat_robust_list_head __user *head; | |
4086 | unsigned long ret; | |
4087 | struct task_struct *p; | |
4088 | ||
4089 | if (!futex_cmpxchg_enabled) | |
4090 | return -ENOSYS; | |
4091 | ||
4092 | rcu_read_lock(); | |
4093 | ||
4094 | ret = -ESRCH; | |
4095 | if (!pid) | |
4096 | p = current; | |
4097 | else { | |
4098 | p = find_task_by_vpid(pid); | |
4099 | if (!p) | |
4100 | goto err_unlock; | |
4101 | } | |
4102 | ||
4103 | ret = -EPERM; | |
4104 | if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS)) | |
4105 | goto err_unlock; | |
4106 | ||
4107 | head = p->compat_robust_list; | |
4108 | rcu_read_unlock(); | |
4109 | ||
4110 | if (put_user(sizeof(*head), len_ptr)) | |
4111 | return -EFAULT; | |
4112 | return put_user(ptr_to_compat(head), head_ptr); | |
4113 | ||
4114 | err_unlock: | |
4115 | rcu_read_unlock(); | |
4116 | ||
4117 | return ret; | |
4118 | } | |
bec2f7cb | 4119 | #endif /* CONFIG_COMPAT */ |
04e7712f | 4120 | |
bec2f7cb | 4121 | #ifdef CONFIG_COMPAT_32BIT_TIME |
8dabe724 | 4122 | SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val, |
04e7712f AB |
4123 | struct old_timespec32 __user *, utime, u32 __user *, uaddr2, |
4124 | u32, val3) | |
4125 | { | |
bec2f7cb | 4126 | struct timespec64 ts; |
04e7712f AB |
4127 | ktime_t t, *tp = NULL; |
4128 | int val2 = 0; | |
4129 | int cmd = op & FUTEX_CMD_MASK; | |
4130 | ||
4131 | if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI || | |
4132 | cmd == FUTEX_WAIT_BITSET || | |
4133 | cmd == FUTEX_WAIT_REQUEUE_PI)) { | |
bec2f7cb | 4134 | if (get_old_timespec32(&ts, utime)) |
04e7712f | 4135 | return -EFAULT; |
bec2f7cb | 4136 | if (!timespec64_valid(&ts)) |
04e7712f AB |
4137 | return -EINVAL; |
4138 | ||
bec2f7cb | 4139 | t = timespec64_to_ktime(ts); |
04e7712f AB |
4140 | if (cmd == FUTEX_WAIT) |
4141 | t = ktime_add_safe(ktime_get(), t); | |
4142 | tp = &t; | |
4143 | } | |
4144 | if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE || | |
4145 | cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP) | |
4146 | val2 = (int) (unsigned long) utime; | |
4147 | ||
4148 | return do_futex(uaddr, op, val, tp, uaddr2, val2, val3); | |
4149 | } | |
bec2f7cb | 4150 | #endif /* CONFIG_COMPAT_32BIT_TIME */ |
04e7712f | 4151 | |
03b8c7b6 | 4152 | static void __init futex_detect_cmpxchg(void) |
1da177e4 | 4153 | { |
03b8c7b6 | 4154 | #ifndef CONFIG_HAVE_FUTEX_CMPXCHG |
a0c1e907 | 4155 | u32 curval; |
03b8c7b6 HC |
4156 | |
4157 | /* | |
4158 | * This will fail and we want it. Some arch implementations do | |
4159 | * runtime detection of the futex_atomic_cmpxchg_inatomic() | |
4160 | * functionality. We want to know that before we call in any | |
4161 | * of the complex code paths. Also we want to prevent | |
4162 | * registration of robust lists in that case. NULL is | |
4163 | * guaranteed to fault and we get -EFAULT on functional | |
4164 | * implementation, the non-functional ones will return | |
4165 | * -ENOSYS. | |
4166 | */ | |
4167 | if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT) | |
4168 | futex_cmpxchg_enabled = 1; | |
4169 | #endif | |
4170 | } | |
4171 | ||
4172 | static int __init futex_init(void) | |
4173 | { | |
63b1a816 | 4174 | unsigned int futex_shift; |
a52b89eb DB |
4175 | unsigned long i; |
4176 | ||
4177 | #if CONFIG_BASE_SMALL | |
4178 | futex_hashsize = 16; | |
4179 | #else | |
4180 | futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus()); | |
4181 | #endif | |
4182 | ||
4183 | futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues), | |
4184 | futex_hashsize, 0, | |
4185 | futex_hashsize < 256 ? HASH_SMALL : 0, | |
63b1a816 HC |
4186 | &futex_shift, NULL, |
4187 | futex_hashsize, futex_hashsize); | |
4188 | futex_hashsize = 1UL << futex_shift; | |
03b8c7b6 HC |
4189 | |
4190 | futex_detect_cmpxchg(); | |
a0c1e907 | 4191 | |
a52b89eb | 4192 | for (i = 0; i < futex_hashsize; i++) { |
11d4616b | 4193 | atomic_set(&futex_queues[i].waiters, 0); |
732375c6 | 4194 | plist_head_init(&futex_queues[i].chain); |
3e4ab747 TG |
4195 | spin_lock_init(&futex_queues[i].lock); |
4196 | } | |
4197 | ||
1da177e4 LT |
4198 | return 0; |
4199 | } | |
25f71d1c | 4200 | core_initcall(futex_init); |