blob: c3eebcdac206da4fd32a58c05610b1a093917189 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Fast Userspace Mutexes (which I call "Futexes!").
3 * (C) Rusty Russell, IBM 2002
4 *
5 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
6 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
7 *
8 * Removed page pinning, fix privately mapped COW pages and other cleanups
9 * (C) Copyright 2003, 2004 Jamie Lokier
10 *
Ingo Molnar0771dfe2006-03-27 01:16:22 -080011 * Robust futex support started by Ingo Molnar
12 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
13 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
14 *
Ingo Molnarc87e2832006-06-27 02:54:58 -070015 * PI-futex support started by Ingo Molnar and Thomas Gleixner
16 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
17 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
18 *
Eric Dumazet34f01cc2007-05-09 02:35:04 -070019 * PRIVATE futexes by Eric Dumazet
20 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
21 *
Darren Hart52400ba2009-04-03 13:40:49 -070022 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
23 * Copyright (C) IBM Corporation, 2009
24 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
25 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070026 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
27 * enough at me, Linus for the original (flawed) idea, Matthew
28 * Kirkwood for proof-of-concept implementation.
29 *
30 * "The futexes are also cursed."
31 * "But they come in a choice of three flavours!"
32 *
33 * This program is free software; you can redistribute it and/or modify
34 * it under the terms of the GNU General Public License as published by
35 * the Free Software Foundation; either version 2 of the License, or
36 * (at your option) any later version.
37 *
38 * This program is distributed in the hope that it will be useful,
39 * but WITHOUT ANY WARRANTY; without even the implied warranty of
40 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
41 * GNU General Public License for more details.
42 *
43 * You should have received a copy of the GNU General Public License
44 * along with this program; if not, write to the Free Software
45 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
46 */
47#include <linux/slab.h>
48#include <linux/poll.h>
49#include <linux/fs.h>
50#include <linux/file.h>
51#include <linux/jhash.h>
52#include <linux/init.h>
53#include <linux/futex.h>
54#include <linux/mount.h>
55#include <linux/pagemap.h>
56#include <linux/syscalls.h>
Jesper Juhl7ed20e12005-05-01 08:59:14 -070057#include <linux/signal.h>
Paul Gortmaker9984de12011-05-23 14:51:41 -040058#include <linux/export.h>
Andrey Mirkinfd5eea42007-10-16 23:30:13 -070059#include <linux/magic.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070060#include <linux/pid.h>
61#include <linux/nsproxy.h>
Kees Cookbdbb7762012-03-19 16:12:53 -070062#include <linux/ptrace.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060063#include <linux/sched/rt.h>
Ingo Molnar84f001e2017-02-01 16:36:40 +010064#include <linux/sched/wake_q.h>
Ingo Molnar6e84f312017-02-08 18:51:29 +010065#include <linux/sched/mm.h>
Zhang Yi13d60f42013-06-25 21:19:31 +080066#include <linux/hugetlb.h>
Colin Cross88c80042013-05-01 18:35:05 -070067#include <linux/freezer.h>
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -080068#include <linux/bootmem.h>
Davidlohr Buesoab51fba2015-06-29 23:26:02 -070069#include <linux/fault-inject.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070070
Jakub Jelinek4732efbe2005-09-06 15:16:25 -070071#include <asm/futex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070072
Peter Zijlstra1696a8b2013-10-31 18:18:19 +010073#include "locking/rtmutex_common.h"
Ingo Molnarc87e2832006-06-27 02:54:58 -070074
Thomas Gleixner99b60ce2014-01-12 15:31:24 -080075/*
Davidlohr Buesod7e8af12014-04-09 11:55:07 -070076 * READ this before attempting to hack on futexes!
77 *
78 * Basic futex operation and ordering guarantees
79 * =============================================
Thomas Gleixner99b60ce2014-01-12 15:31:24 -080080 *
81 * The waiter reads the futex value in user space and calls
82 * futex_wait(). This function computes the hash bucket and acquires
83 * the hash bucket lock. After that it reads the futex user space value
Davidlohr Buesob0c29f72014-01-12 15:31:25 -080084 * again and verifies that the data has not changed. If it has not changed
85 * it enqueues itself into the hash bucket, releases the hash bucket lock
86 * and schedules.
Thomas Gleixner99b60ce2014-01-12 15:31:24 -080087 *
88 * The waker side modifies the user space value of the futex and calls
Davidlohr Buesob0c29f72014-01-12 15:31:25 -080089 * futex_wake(). This function computes the hash bucket and acquires the
90 * hash bucket lock. Then it looks for waiters on that futex in the hash
91 * bucket and wakes them.
Thomas Gleixner99b60ce2014-01-12 15:31:24 -080092 *
Davidlohr Buesob0c29f72014-01-12 15:31:25 -080093 * In futex wake up scenarios where no tasks are blocked on a futex, taking
94 * the hb spinlock can be avoided and simply return. In order for this
95 * optimization to work, ordering guarantees must exist so that the waiter
96 * being added to the list is acknowledged when the list is concurrently being
97 * checked by the waker, avoiding scenarios like the following:
Thomas Gleixner99b60ce2014-01-12 15:31:24 -080098 *
99 * CPU 0 CPU 1
100 * val = *futex;
101 * sys_futex(WAIT, futex, val);
102 * futex_wait(futex, val);
103 * uval = *futex;
104 * *futex = newval;
105 * sys_futex(WAKE, futex);
106 * futex_wake(futex);
107 * if (queue_empty())
108 * return;
109 * if (uval == val)
110 * lock(hash_bucket(futex));
111 * queue();
112 * unlock(hash_bucket(futex));
113 * schedule();
114 *
115 * This would cause the waiter on CPU 0 to wait forever because it
116 * missed the transition of the user space value from val to newval
117 * and the waker did not find the waiter in the hash bucket queue.
Thomas Gleixner99b60ce2014-01-12 15:31:24 -0800118 *
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800119 * The correct serialization ensures that a waiter either observes
120 * the changed user space value before blocking or is woken by a
121 * concurrent waker:
122 *
123 * CPU 0 CPU 1
Thomas Gleixner99b60ce2014-01-12 15:31:24 -0800124 * val = *futex;
125 * sys_futex(WAIT, futex, val);
126 * futex_wait(futex, val);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800127 *
Davidlohr Buesod7e8af12014-04-09 11:55:07 -0700128 * waiters++; (a)
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -0800129 * smp_mb(); (A) <-- paired with -.
130 * |
131 * lock(hash_bucket(futex)); |
132 * |
133 * uval = *futex; |
134 * | *futex = newval;
135 * | sys_futex(WAKE, futex);
136 * | futex_wake(futex);
137 * |
138 * `--------> smp_mb(); (B)
Thomas Gleixner99b60ce2014-01-12 15:31:24 -0800139 * if (uval == val)
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800140 * queue();
Thomas Gleixner99b60ce2014-01-12 15:31:24 -0800141 * unlock(hash_bucket(futex));
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800142 * schedule(); if (waiters)
143 * lock(hash_bucket(futex));
Davidlohr Buesod7e8af12014-04-09 11:55:07 -0700144 * else wake_waiters(futex);
145 * waiters--; (b) unlock(hash_bucket(futex));
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800146 *
Davidlohr Buesod7e8af12014-04-09 11:55:07 -0700147 * Where (A) orders the waiters increment and the futex value read through
148 * atomic operations (see hb_waiters_inc) and where (B) orders the write
Davidlohr Bueso993b2ff2014-10-23 20:27:00 -0700149 * to futex and the waiters read -- this is done by the barriers for both
150 * shared and private futexes in get_futex_key_refs().
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800151 *
152 * This yields the following case (where X:=waiters, Y:=futex):
153 *
154 * X = Y = 0
155 *
156 * w[X]=1 w[Y]=1
157 * MB MB
158 * r[Y]=y r[X]=x
159 *
160 * Which guarantees that x==0 && y==0 is impossible; which translates back into
161 * the guarantee that we cannot both miss the futex variable change and the
162 * enqueue.
Davidlohr Buesod7e8af12014-04-09 11:55:07 -0700163 *
164 * Note that a new waiter is accounted for in (a) even when it is possible that
165 * the wait call can return error, in which case we backtrack from it in (b).
166 * Refer to the comment in queue_lock().
167 *
168 * Similarly, in order to account for waiters being requeued on another
169 * address we always increment the waiters for the destination bucket before
170 * acquiring the lock. It then decrements them again after releasing it -
171 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
172 * will do the additional required waiter count housekeeping. This is done for
173 * double_lock_hb() and double_unlock_hb(), respectively.
Thomas Gleixner99b60ce2014-01-12 15:31:24 -0800174 */
175
Heiko Carstens03b8c7b2014-03-02 13:09:47 +0100176#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
Thomas Gleixnera0c1e902008-02-23 15:23:57 -0800177int __read_mostly futex_cmpxchg_enabled;
Heiko Carstens03b8c7b2014-03-02 13:09:47 +0100178#endif
Thomas Gleixnera0c1e902008-02-23 15:23:57 -0800179
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180/*
Darren Hartb41277d2010-11-08 13:10:09 -0800181 * Futex flags used to encode options to functions and preserve them across
182 * restarts.
183 */
Thomas Gleixner784bdf32016-07-29 16:32:30 +0200184#ifdef CONFIG_MMU
185# define FLAGS_SHARED 0x01
186#else
187/*
188 * NOMMU does not have per process address space. Let the compiler optimize
189 * code away.
190 */
191# define FLAGS_SHARED 0x00
192#endif
Darren Hartb41277d2010-11-08 13:10:09 -0800193#define FLAGS_CLOCKRT 0x02
194#define FLAGS_HAS_TIMEOUT 0x04
195
196/*
Ingo Molnarc87e2832006-06-27 02:54:58 -0700197 * Priority Inheritance state:
198 */
199struct futex_pi_state {
200 /*
201 * list of 'owned' pi_state instances - these have to be
202 * cleaned up in do_exit() if the task exits prematurely:
203 */
204 struct list_head list;
205
206 /*
207 * The PI object:
208 */
209 struct rt_mutex pi_mutex;
210
211 struct task_struct *owner;
212 atomic_t refcount;
213
214 union futex_key key;
215};
216
Darren Hartd8d88fb2009-09-21 22:30:30 -0700217/**
218 * struct futex_q - The hashed futex queue entry, one per waiting task
Randy Dunlapfb62db22010-10-13 11:02:34 -0700219 * @list: priority-sorted list of tasks waiting on this futex
Darren Hartd8d88fb2009-09-21 22:30:30 -0700220 * @task: the task waiting on the futex
221 * @lock_ptr: the hash bucket lock
222 * @key: the key the futex is hashed on
223 * @pi_state: optional priority inheritance state
224 * @rt_waiter: rt_waiter storage for use with requeue_pi
225 * @requeue_pi_key: the requeue_pi target futex key
226 * @bitset: bitset for the optional bitmasked wakeup
227 *
228 * We use this hashed waitqueue, instead of a normal wait_queue_t, so
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 * we can wake only the relevant ones (hashed queues may be shared).
230 *
231 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
Pierre Peifferec92d082007-05-09 02:35:00 -0700232 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
Randy Dunlapfb62db22010-10-13 11:02:34 -0700233 * The order of wakeup is always to make the first condition true, then
Darren Hartd8d88fb2009-09-21 22:30:30 -0700234 * the second.
235 *
236 * PI futexes are typically woken before they are removed from the hash list via
237 * the rt_mutex code. See unqueue_me_pi().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 */
239struct futex_q {
Pierre Peifferec92d082007-05-09 02:35:00 -0700240 struct plist_node list;
Darren Hartd8d88fb2009-09-21 22:30:30 -0700241
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +0200242 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 spinlock_t *lock_ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 union futex_key key;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700245 struct futex_pi_state *pi_state;
Darren Hart52400ba2009-04-03 13:40:49 -0700246 struct rt_mutex_waiter *rt_waiter;
Darren Hart84bc4af2009-08-13 17:36:53 -0700247 union futex_key *requeue_pi_key;
Thomas Gleixnercd689982008-02-01 17:45:14 +0100248 u32 bitset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249};
250
Darren Hart5bdb05f2010-11-08 13:40:28 -0800251static const struct futex_q futex_q_init = {
252 /* list gets initialized in queue_me()*/
253 .key = FUTEX_KEY_INIT,
254 .bitset = FUTEX_BITSET_MATCH_ANY
255};
256
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257/*
Darren Hartb2d09942009-03-12 00:55:37 -0700258 * Hash buckets are shared by all the futex_keys that hash to the same
259 * location. Each key may have multiple futex_q structures, one for each task
260 * waiting on a futex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 */
262struct futex_hash_bucket {
Linus Torvalds11d46162014-03-20 22:11:17 -0700263 atomic_t waiters;
Pierre Peifferec92d082007-05-09 02:35:00 -0700264 spinlock_t lock;
265 struct plist_head chain;
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -0800266} ____cacheline_aligned_in_smp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
Rasmus Villemoesac742d32015-09-09 23:36:40 +0200268/*
269 * The base of the bucket array and its size are always used together
270 * (after initialization only in hash_futex()), so ensure that they
271 * reside in the same cacheline.
272 */
273static struct {
274 struct futex_hash_bucket *queues;
275 unsigned long hashsize;
276} __futex_data __read_mostly __aligned(2*sizeof(long));
277#define futex_queues (__futex_data.queues)
278#define futex_hashsize (__futex_data.hashsize)
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -0800279
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280
Davidlohr Buesoab51fba2015-06-29 23:26:02 -0700281/*
282 * Fault injections for futexes.
283 */
284#ifdef CONFIG_FAIL_FUTEX
285
286static struct {
287 struct fault_attr attr;
288
Viresh Kumar621a5f72015-09-26 15:04:07 -0700289 bool ignore_private;
Davidlohr Buesoab51fba2015-06-29 23:26:02 -0700290} fail_futex = {
291 .attr = FAULT_ATTR_INITIALIZER,
Viresh Kumar621a5f72015-09-26 15:04:07 -0700292 .ignore_private = false,
Davidlohr Buesoab51fba2015-06-29 23:26:02 -0700293};
294
295static int __init setup_fail_futex(char *str)
296{
297 return setup_fault_attr(&fail_futex.attr, str);
298}
299__setup("fail_futex=", setup_fail_futex);
300
kbuild test robot5d285a72015-07-21 01:40:45 +0800301static bool should_fail_futex(bool fshared)
Davidlohr Buesoab51fba2015-06-29 23:26:02 -0700302{
303 if (fail_futex.ignore_private && !fshared)
304 return false;
305
306 return should_fail(&fail_futex.attr, 1);
307}
308
309#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
310
311static int __init fail_futex_debugfs(void)
312{
313 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
314 struct dentry *dir;
315
316 dir = fault_create_debugfs_attr("fail_futex", NULL,
317 &fail_futex.attr);
318 if (IS_ERR(dir))
319 return PTR_ERR(dir);
320
321 if (!debugfs_create_bool("ignore-private", mode, dir,
322 &fail_futex.ignore_private)) {
323 debugfs_remove_recursive(dir);
324 return -ENOMEM;
325 }
326
327 return 0;
328}
329
330late_initcall(fail_futex_debugfs);
331
332#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
333
334#else
335static inline bool should_fail_futex(bool fshared)
336{
337 return false;
338}
339#endif /* CONFIG_FAIL_FUTEX */
340
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800341static inline void futex_get_mm(union futex_key *key)
342{
Vegard Nossumf1f10072017-02-27 14:30:07 -0800343 mmgrab(key->private.mm);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800344 /*
345 * Ensure futex_get_mm() implies a full barrier such that
346 * get_futex_key() implies a full barrier. This is relied upon
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -0800347 * as smp_mb(); (B), see the ordering comment above.
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800348 */
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100349 smp_mb__after_atomic();
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800350}
351
Linus Torvalds11d46162014-03-20 22:11:17 -0700352/*
353 * Reflects a new waiter being added to the waitqueue.
354 */
355static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800356{
357#ifdef CONFIG_SMP
Linus Torvalds11d46162014-03-20 22:11:17 -0700358 atomic_inc(&hb->waiters);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800359 /*
Linus Torvalds11d46162014-03-20 22:11:17 -0700360 * Full barrier (A), see the ordering comment above.
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800361 */
Peter Zijlstra4e857c52014-03-17 18:06:10 +0100362 smp_mb__after_atomic();
Linus Torvalds11d46162014-03-20 22:11:17 -0700363#endif
364}
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800365
Linus Torvalds11d46162014-03-20 22:11:17 -0700366/*
367 * Reflects a waiter being removed from the waitqueue by wakeup
368 * paths.
369 */
370static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
371{
372#ifdef CONFIG_SMP
373 atomic_dec(&hb->waiters);
374#endif
375}
376
377static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
378{
379#ifdef CONFIG_SMP
380 return atomic_read(&hb->waiters);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800381#else
Linus Torvalds11d46162014-03-20 22:11:17 -0700382 return 1;
Davidlohr Buesob0c29f72014-01-12 15:31:25 -0800383#endif
384}
385
Thomas Gleixnere8b61b32016-06-01 10:43:29 +0200386/**
387 * hash_futex - Return the hash bucket in the global hash
388 * @key: Pointer to the futex key for which the hash is calculated
389 *
390 * We hash on the keys returned from get_futex_key (see below) and return the
391 * corresponding hash bucket in the global hash.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 */
393static struct futex_hash_bucket *hash_futex(union futex_key *key)
394{
395 u32 hash = jhash2((u32*)&key->both.word,
396 (sizeof(key->both.word)+sizeof(key->both.ptr))/4,
397 key->both.offset);
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -0800398 return &futex_queues[hash & (futex_hashsize - 1)];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399}
400
Thomas Gleixnere8b61b32016-06-01 10:43:29 +0200401
402/**
403 * match_futex - Check whether two futex keys are equal
404 * @key1: Pointer to key1
405 * @key2: Pointer to key2
406 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 * Return 1 if two futex_keys are equal, 0 otherwise.
408 */
409static inline int match_futex(union futex_key *key1, union futex_key *key2)
410{
Darren Hart2bc87202009-10-14 10:12:39 -0700411 return (key1 && key2
412 && key1->both.word == key2->both.word
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 && key1->both.ptr == key2->both.ptr
414 && key1->both.offset == key2->both.offset);
415}
416
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200417/*
418 * Take a reference to the resource addressed by a key.
419 * Can be called while holding spinlocks.
420 *
421 */
422static void get_futex_key_refs(union futex_key *key)
423{
424 if (!key->both.ptr)
425 return;
426
Thomas Gleixner784bdf32016-07-29 16:32:30 +0200427 /*
428 * On MMU less systems futexes are always "private" as there is no per
429 * process address space. We need the smp wmb nevertheless - yes,
430 * arch/blackfin has MMU less SMP ...
431 */
432 if (!IS_ENABLED(CONFIG_MMU)) {
433 smp_mb(); /* explicit smp_mb(); (B) */
434 return;
435 }
436
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200437 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
438 case FUT_OFF_INODE:
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -0800439 ihold(key->shared.inode); /* implies smp_mb(); (B) */
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200440 break;
441 case FUT_OFF_MMSHARED:
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -0800442 futex_get_mm(key); /* implies smp_mb(); (B) */
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200443 break;
Catalin Marinas76835b0e2014-10-17 17:38:49 +0100444 default:
Davidlohr Bueso993b2ff2014-10-23 20:27:00 -0700445 /*
446 * Private futexes do not hold reference on an inode or
447 * mm, therefore the only purpose of calling get_futex_key_refs
448 * is because we need the barrier for the lockless waiter check.
449 */
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -0800450 smp_mb(); /* explicit smp_mb(); (B) */
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200451 }
452}
453
454/*
455 * Drop a reference to the resource addressed by a key.
Davidlohr Bueso993b2ff2014-10-23 20:27:00 -0700456 * The hash bucket spinlock must not be held. This is
457 * a no-op for private futexes, see comment in the get
458 * counterpart.
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200459 */
460static void drop_futex_key_refs(union futex_key *key)
461{
Darren Hart90621c42008-12-29 19:43:21 -0800462 if (!key->both.ptr) {
463 /* If we're here then we tried to put a key we failed to get */
464 WARN_ON_ONCE(1);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200465 return;
Darren Hart90621c42008-12-29 19:43:21 -0800466 }
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200467
Thomas Gleixner784bdf32016-07-29 16:32:30 +0200468 if (!IS_ENABLED(CONFIG_MMU))
469 return;
470
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200471 switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
472 case FUT_OFF_INODE:
473 iput(key->shared.inode);
474 break;
475 case FUT_OFF_MMSHARED:
476 mmdrop(key->private.mm);
477 break;
478 }
479}
480
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700481/**
Darren Hartd96ee562009-09-21 22:30:22 -0700482 * get_futex_key() - Get parameters which are the keys for a futex
483 * @uaddr: virtual address of the futex
484 * @fshared: 0 for a PROCESS_PRIVATE futex, 1 for PROCESS_SHARED
485 * @key: address where result is stored.
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500486 * @rw: mapping needs to be read/write (values: VERIFY_READ,
487 * VERIFY_WRITE)
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700488 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -0800489 * Return: a negative error code or 0
490 *
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700491 * The key words are stored in *key on success.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 *
Al Viro6131ffa2013-02-27 16:59:05 -0500493 * For shared mappings, it's (page->index, file_inode(vma->vm_file),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 * offset_within_page). For private mappings, it's (uaddr, current->mm).
495 * We can usually work out the index without swapping in the page.
496 *
Darren Hartb2d09942009-03-12 00:55:37 -0700497 * lock_page() might sleep, the caller should not hold a spinlock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 */
Thomas Gleixner64d13042009-05-18 21:20:10 +0200499static int
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500500get_futex_key(u32 __user *uaddr, int fshared, union futex_key *key, int rw)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501{
Ingo Molnare2970f22006-06-27 02:54:47 -0700502 unsigned long address = (unsigned long)uaddr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 struct mm_struct *mm = current->mm;
Mel Gorman077fa7a2016-06-08 14:25:22 +0100504 struct page *page, *tail;
Kirill A. Shutemov14d27ab2016-01-15 16:53:00 -0800505 struct address_space *mapping;
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500506 int err, ro = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507
508 /*
509 * The futex address must be "naturally" aligned.
510 */
Ingo Molnare2970f22006-06-27 02:54:47 -0700511 key->both.offset = address % PAGE_SIZE;
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700512 if (unlikely((address % sizeof(u32)) != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 return -EINVAL;
Ingo Molnare2970f22006-06-27 02:54:47 -0700514 address -= key->both.offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515
Linus Torvalds5cdec2d2013-12-12 09:53:51 -0800516 if (unlikely(!access_ok(rw, uaddr, sizeof(u32))))
517 return -EFAULT;
518
Davidlohr Buesoab51fba2015-06-29 23:26:02 -0700519 if (unlikely(should_fail_futex(fshared)))
520 return -EFAULT;
521
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 /*
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700523 * PROCESS_PRIVATE futexes are fast.
524 * As the mm cannot disappear under us and the 'key' only needs
525 * virtual address, we dont even have to find the underlying vma.
526 * Note : We do have to check 'uaddr' is a valid user address,
527 * but access_ok() should be faster than find_vma()
528 */
529 if (!fshared) {
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700530 key->private.mm = mm;
531 key->private.address = address;
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -0800532 get_futex_key_refs(key); /* implies smp_mb(); (B) */
Eric Dumazet34f01cc2007-05-09 02:35:04 -0700533 return 0;
534 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200536again:
Davidlohr Buesoab51fba2015-06-29 23:26:02 -0700537 /* Ignore any VERIFY_READ mapping (futex common case) */
538 if (unlikely(should_fail_futex(fshared)))
539 return -EFAULT;
540
KOSAKI Motohiro7485d0d2010-01-05 16:32:43 +0900541 err = get_user_pages_fast(address, 1, 1, &page);
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500542 /*
543 * If write access is not required (eg. FUTEX_WAIT), try
544 * and get read-only access.
545 */
546 if (err == -EFAULT && rw == VERIFY_READ) {
547 err = get_user_pages_fast(address, 1, 0, &page);
548 ro = 1;
549 }
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200550 if (err < 0)
551 return err;
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500552 else
553 err = 0;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200554
Mel Gorman65d8fc72016-02-09 11:15:14 -0800555 /*
556 * The treatment of mapping from this point on is critical. The page
557 * lock protects many things but in this context the page lock
558 * stabilizes mapping, prevents inode freeing in the shared
559 * file-backed region case and guards against movement to swap cache.
560 *
561 * Strictly speaking the page lock is not needed in all cases being
562 * considered here and page lock forces unnecessarily serialization
563 * From this point on, mapping will be re-verified if necessary and
564 * page lock will be acquired only if it is unavoidable
Mel Gorman077fa7a2016-06-08 14:25:22 +0100565 *
566 * Mapping checks require the head page for any compound page so the
567 * head page and mapping is looked up now. For anonymous pages, it
568 * does not matter if the page splits in the future as the key is
569 * based on the address. For filesystem-backed pages, the tail is
570 * required as the index of the page determines the key. For
571 * base pages, there is no tail page and tail == page.
Mel Gorman65d8fc72016-02-09 11:15:14 -0800572 */
Mel Gorman077fa7a2016-06-08 14:25:22 +0100573 tail = page;
Mel Gorman65d8fc72016-02-09 11:15:14 -0800574 page = compound_head(page);
575 mapping = READ_ONCE(page->mapping);
576
Hugh Dickinse6780f72011-12-31 11:44:01 -0800577 /*
Kirill A. Shutemov14d27ab2016-01-15 16:53:00 -0800578 * If page->mapping is NULL, then it cannot be a PageAnon
Hugh Dickinse6780f72011-12-31 11:44:01 -0800579 * page; but it might be the ZERO_PAGE or in the gate area or
580 * in a special mapping (all cases which we are happy to fail);
581 * or it may have been a good file page when get_user_pages_fast
582 * found it, but truncated or holepunched or subjected to
583 * invalidate_complete_page2 before we got the page lock (also
584 * cases which we are happy to fail). And we hold a reference,
585 * so refcount care in invalidate_complete_page's remove_mapping
586 * prevents drop_caches from setting mapping to NULL beneath us.
587 *
588 * The case we do have to guard against is when memory pressure made
589 * shmem_writepage move it from filecache to swapcache beneath us:
Kirill A. Shutemov14d27ab2016-01-15 16:53:00 -0800590 * an unlikely race, but we do need to retry for page->mapping.
Hugh Dickinse6780f72011-12-31 11:44:01 -0800591 */
Mel Gorman65d8fc72016-02-09 11:15:14 -0800592 if (unlikely(!mapping)) {
593 int shmem_swizzled;
594
595 /*
596 * Page lock is required to identify which special case above
597 * applies. If this is really a shmem page then the page lock
598 * will prevent unexpected transitions.
599 */
600 lock_page(page);
601 shmem_swizzled = PageSwapCache(page) || page->mapping;
Kirill A. Shutemov14d27ab2016-01-15 16:53:00 -0800602 unlock_page(page);
603 put_page(page);
Mel Gorman65d8fc72016-02-09 11:15:14 -0800604
Hugh Dickinse6780f72011-12-31 11:44:01 -0800605 if (shmem_swizzled)
606 goto again;
Mel Gorman65d8fc72016-02-09 11:15:14 -0800607
Hugh Dickinse6780f72011-12-31 11:44:01 -0800608 return -EFAULT;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200609 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
611 /*
612 * Private mappings are handled in a simple way.
613 *
Mel Gorman65d8fc72016-02-09 11:15:14 -0800614 * If the futex key is stored on an anonymous page, then the associated
615 * object is the mm which is implicitly pinned by the calling process.
616 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
618 * it's a read-only handle, it's expected that futexes attach to
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200619 * the object not the particular process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 */
Kirill A. Shutemov14d27ab2016-01-15 16:53:00 -0800621 if (PageAnon(page)) {
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500622 /*
623 * A RO anonymous page will never change and thus doesn't make
624 * sense for futex operations.
625 */
Davidlohr Buesoab51fba2015-06-29 23:26:02 -0700626 if (unlikely(should_fail_futex(fshared)) || ro) {
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500627 err = -EFAULT;
628 goto out;
629 }
630
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200631 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 key->private.mm = mm;
Ingo Molnare2970f22006-06-27 02:54:47 -0700633 key->private.address = address;
Mel Gorman65d8fc72016-02-09 11:15:14 -0800634
635 get_futex_key_refs(key); /* implies smp_mb(); (B) */
636
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200637 } else {
Mel Gorman65d8fc72016-02-09 11:15:14 -0800638 struct inode *inode;
639
640 /*
641 * The associated futex object in this case is the inode and
642 * the page->mapping must be traversed. Ordinarily this should
643 * be stabilised under page lock but it's not strictly
644 * necessary in this case as we just want to pin the inode, not
645 * update the radix tree or anything like that.
646 *
647 * The RCU read lock is taken as the inode is finally freed
648 * under RCU. If the mapping still matches expectations then the
649 * mapping->host can be safely accessed as being a valid inode.
650 */
651 rcu_read_lock();
652
653 if (READ_ONCE(page->mapping) != mapping) {
654 rcu_read_unlock();
655 put_page(page);
656
657 goto again;
658 }
659
660 inode = READ_ONCE(mapping->host);
661 if (!inode) {
662 rcu_read_unlock();
663 put_page(page);
664
665 goto again;
666 }
667
668 /*
669 * Take a reference unless it is about to be freed. Previously
670 * this reference was taken by ihold under the page lock
671 * pinning the inode in place so i_lock was unnecessary. The
672 * only way for this check to fail is if the inode was
673 * truncated in parallel so warn for now if this happens.
674 *
675 * We are not calling into get_futex_key_refs() in file-backed
676 * cases, therefore a successful atomic_inc return below will
677 * guarantee that get_futex_key() will still imply smp_mb(); (B).
678 */
679 if (WARN_ON_ONCE(!atomic_inc_not_zero(&inode->i_count))) {
680 rcu_read_unlock();
681 put_page(page);
682
683 goto again;
684 }
685
686 /* Should be impossible but lets be paranoid for now */
687 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
688 err = -EFAULT;
689 rcu_read_unlock();
690 iput(inode);
691
692 goto out;
693 }
694
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200695 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
Mel Gorman65d8fc72016-02-09 11:15:14 -0800696 key->shared.inode = inode;
Mel Gorman077fa7a2016-06-08 14:25:22 +0100697 key->shared.pgoff = basepage_index(tail);
Mel Gorman65d8fc72016-02-09 11:15:14 -0800698 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 }
700
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500701out:
Kirill A. Shutemov14d27ab2016-01-15 16:53:00 -0800702 put_page(page);
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500703 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704}
705
Thomas Gleixnerae791a22010-11-10 13:30:36 +0100706static inline void put_futex_key(union futex_key *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707{
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200708 drop_futex_key_refs(key);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709}
710
Darren Hartd96ee562009-09-21 22:30:22 -0700711/**
712 * fault_in_user_writeable() - Fault in user address and verify RW access
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200713 * @uaddr: pointer to faulting user space address
714 *
715 * Slow path to fixup the fault we just took in the atomic write
716 * access to @uaddr.
717 *
Randy Dunlapfb62db22010-10-13 11:02:34 -0700718 * We have no generic implementation of a non-destructive write to the
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200719 * user address. We know that we faulted in the atomic pagefault
720 * disabled section so we can as well avoid the #PF overhead by
721 * calling get_user_pages() right away.
722 */
723static int fault_in_user_writeable(u32 __user *uaddr)
724{
Andi Kleen722d0172009-12-08 13:19:42 +0100725 struct mm_struct *mm = current->mm;
726 int ret;
727
728 down_read(&mm->mmap_sem);
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -0700729 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800730 FAULT_FLAG_WRITE, NULL);
Andi Kleen722d0172009-12-08 13:19:42 +0100731 up_read(&mm->mmap_sem);
732
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200733 return ret < 0 ? ret : 0;
734}
735
Darren Hart4b1c4862009-04-03 13:39:42 -0700736/**
737 * futex_top_waiter() - Return the highest priority waiter on a futex
Darren Hartd96ee562009-09-21 22:30:22 -0700738 * @hb: the hash bucket the futex_q's reside in
739 * @key: the futex key (to distinguish it from other futex futex_q's)
Darren Hart4b1c4862009-04-03 13:39:42 -0700740 *
741 * Must be called with the hb lock held.
742 */
743static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
744 union futex_key *key)
745{
746 struct futex_q *this;
747
748 plist_for_each_entry(this, &hb->chain, list) {
749 if (match_futex(&this->key, key))
750 return this;
751 }
752 return NULL;
753}
754
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800755static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
756 u32 uval, u32 newval)
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700757{
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800758 int ret;
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700759
760 pagefault_disable();
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800761 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700762 pagefault_enable();
763
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800764 return ret;
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700765}
766
767static int get_futex_value_locked(u32 *dest, u32 __user *from)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
769 int ret;
770
Peter Zijlstraa8663742006-12-06 20:32:20 -0800771 pagefault_disable();
Linus Torvaldsbd28b142016-05-22 17:21:27 -0700772 ret = __get_user(*dest, from);
Peter Zijlstraa8663742006-12-06 20:32:20 -0800773 pagefault_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
775 return ret ? -EFAULT : 0;
776}
777
Ingo Molnarc87e2832006-06-27 02:54:58 -0700778
779/*
780 * PI code:
781 */
782static int refill_pi_state_cache(void)
783{
784 struct futex_pi_state *pi_state;
785
786 if (likely(current->pi_state_cache))
787 return 0;
788
Burman Yan4668edc2006-12-06 20:38:51 -0800789 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700790
791 if (!pi_state)
792 return -ENOMEM;
793
Ingo Molnarc87e2832006-06-27 02:54:58 -0700794 INIT_LIST_HEAD(&pi_state->list);
795 /* pi_mutex gets initialized later */
796 pi_state->owner = NULL;
797 atomic_set(&pi_state->refcount, 1);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200798 pi_state->key = FUTEX_KEY_INIT;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700799
800 current->pi_state_cache = pi_state;
801
802 return 0;
803}
804
Peter Zijlstrabf92cf32017-03-22 11:35:53 +0100805static struct futex_pi_state *alloc_pi_state(void)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700806{
807 struct futex_pi_state *pi_state = current->pi_state_cache;
808
809 WARN_ON(!pi_state);
810 current->pi_state_cache = NULL;
811
812 return pi_state;
813}
814
Peter Zijlstrabf92cf32017-03-22 11:35:53 +0100815static void get_pi_state(struct futex_pi_state *pi_state)
816{
817 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
818}
819
Brian Silverman30a6b802014-10-25 20:20:37 -0400820/*
Thomas Gleixner29e9ee52015-12-19 20:07:39 +0000821 * Drops a reference to the pi_state object and frees or caches it
822 * when the last reference is gone.
823 *
Brian Silverman30a6b802014-10-25 20:20:37 -0400824 * Must be called with the hb lock held.
825 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +0000826static void put_pi_state(struct futex_pi_state *pi_state)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700827{
Brian Silverman30a6b802014-10-25 20:20:37 -0400828 if (!pi_state)
829 return;
830
Ingo Molnarc87e2832006-06-27 02:54:58 -0700831 if (!atomic_dec_and_test(&pi_state->refcount))
832 return;
833
834 /*
835 * If pi_state->owner is NULL, the owner is most probably dying
836 * and has cleaned up the pi_state already
837 */
838 if (pi_state->owner) {
Thomas Gleixner1d615482009-11-17 14:54:03 +0100839 raw_spin_lock_irq(&pi_state->owner->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700840 list_del_init(&pi_state->list);
Thomas Gleixner1d615482009-11-17 14:54:03 +0100841 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700842
843 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
844 }
845
846 if (current->pi_state_cache)
847 kfree(pi_state);
848 else {
849 /*
850 * pi_state->list is already empty.
851 * clear pi_state->owner.
852 * refcount is at 0 - put it back to 1.
853 */
854 pi_state->owner = NULL;
855 atomic_set(&pi_state->refcount, 1);
856 current->pi_state_cache = pi_state;
857 }
858}
859
860/*
861 * Look up the task based on what TID userspace gave us.
862 * We dont trust it.
863 */
Peter Zijlstrabf92cf32017-03-22 11:35:53 +0100864static struct task_struct *futex_find_get_task(pid_t pid)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700865{
866 struct task_struct *p;
867
Oleg Nesterovd359b542006-09-29 02:00:55 -0700868 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -0700869 p = find_task_by_vpid(pid);
Michal Hocko7a0ea092010-06-30 09:51:19 +0200870 if (p)
871 get_task_struct(p);
Thomas Gleixnera06381f2007-06-23 11:48:40 +0200872
Oleg Nesterovd359b542006-09-29 02:00:55 -0700873 rcu_read_unlock();
Ingo Molnarc87e2832006-06-27 02:54:58 -0700874
875 return p;
876}
877
878/*
879 * This task is holding PI mutexes at exit time => bad.
880 * Kernel cleans up PI-state, but userspace is likely hosed.
881 * (Robust-futex cleanup is separate and might save the day for userspace.)
882 */
883void exit_pi_state_list(struct task_struct *curr)
884{
Ingo Molnarc87e2832006-06-27 02:54:58 -0700885 struct list_head *next, *head = &curr->pi_state_list;
886 struct futex_pi_state *pi_state;
Ingo Molnar627371d2006-07-29 05:16:20 +0200887 struct futex_hash_bucket *hb;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200888 union futex_key key = FUTEX_KEY_INIT;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700889
Thomas Gleixnera0c1e902008-02-23 15:23:57 -0800890 if (!futex_cmpxchg_enabled)
891 return;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700892 /*
893 * We are a ZOMBIE and nobody can enqueue itself on
894 * pi_state_list anymore, but we have to be careful
Ingo Molnar627371d2006-07-29 05:16:20 +0200895 * versus waiters unqueueing themselves:
Ingo Molnarc87e2832006-06-27 02:54:58 -0700896 */
Thomas Gleixner1d615482009-11-17 14:54:03 +0100897 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700898 while (!list_empty(head)) {
899
900 next = head->next;
901 pi_state = list_entry(next, struct futex_pi_state, list);
902 key = pi_state->key;
Ingo Molnar627371d2006-07-29 05:16:20 +0200903 hb = hash_futex(&key);
Thomas Gleixner1d615482009-11-17 14:54:03 +0100904 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700905
Ingo Molnarc87e2832006-06-27 02:54:58 -0700906 spin_lock(&hb->lock);
907
Thomas Gleixner1d615482009-11-17 14:54:03 +0100908 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +0200909 /*
910 * We dropped the pi-lock, so re-check whether this
911 * task still owns the PI-state:
912 */
Ingo Molnarc87e2832006-06-27 02:54:58 -0700913 if (head->next != next) {
914 spin_unlock(&hb->lock);
915 continue;
916 }
917
Ingo Molnarc87e2832006-06-27 02:54:58 -0700918 WARN_ON(pi_state->owner != curr);
Ingo Molnar627371d2006-07-29 05:16:20 +0200919 WARN_ON(list_empty(&pi_state->list));
920 list_del_init(&pi_state->list);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700921 pi_state->owner = NULL;
Thomas Gleixner1d615482009-11-17 14:54:03 +0100922 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700923
Peter Zijlstra16ffa122017-03-22 11:35:55 +0100924 get_pi_state(pi_state);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700925 spin_unlock(&hb->lock);
926
Peter Zijlstra16ffa122017-03-22 11:35:55 +0100927 rt_mutex_futex_unlock(&pi_state->pi_mutex);
928 put_pi_state(pi_state);
929
Thomas Gleixner1d615482009-11-17 14:54:03 +0100930 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700931 }
Thomas Gleixner1d615482009-11-17 14:54:03 +0100932 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700933}
934
Thomas Gleixner54a21782014-06-03 12:27:08 +0000935/*
936 * We need to check the following states:
937 *
938 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
939 *
940 * [1] NULL | --- | --- | 0 | 0/1 | Valid
941 * [2] NULL | --- | --- | >0 | 0/1 | Valid
942 *
943 * [3] Found | NULL | -- | Any | 0/1 | Invalid
944 *
945 * [4] Found | Found | NULL | 0 | 1 | Valid
946 * [5] Found | Found | NULL | >0 | 1 | Invalid
947 *
948 * [6] Found | Found | task | 0 | 1 | Valid
949 *
950 * [7] Found | Found | NULL | Any | 0 | Invalid
951 *
952 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
953 * [9] Found | Found | task | 0 | 0 | Invalid
954 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
955 *
956 * [1] Indicates that the kernel can acquire the futex atomically. We
957 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
958 *
959 * [2] Valid, if TID does not belong to a kernel thread. If no matching
960 * thread is found then it indicates that the owner TID has died.
961 *
962 * [3] Invalid. The waiter is queued on a non PI futex
963 *
964 * [4] Valid state after exit_robust_list(), which sets the user space
965 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
966 *
967 * [5] The user space value got manipulated between exit_robust_list()
968 * and exit_pi_state_list()
969 *
970 * [6] Valid state after exit_pi_state_list() which sets the new owner in
971 * the pi_state but cannot access the user space value.
972 *
973 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
974 *
975 * [8] Owner and user space value match
976 *
977 * [9] There is no transient state which sets the user space TID to 0
978 * except exit_robust_list(), but this is indicated by the
979 * FUTEX_OWNER_DIED bit. See [4]
980 *
981 * [10] There is no transient state which leaves owner and user space
982 * TID out of sync.
Peter Zijlstra734009e2017-03-22 11:35:52 +0100983 *
984 *
985 * Serialization and lifetime rules:
986 *
987 * hb->lock:
988 *
989 * hb -> futex_q, relation
990 * futex_q -> pi_state, relation
991 *
992 * (cannot be raw because hb can contain arbitrary amount
993 * of futex_q's)
994 *
995 * pi_mutex->wait_lock:
996 *
997 * {uval, pi_state}
998 *
999 * (and pi_mutex 'obviously')
1000 *
1001 * p->pi_lock:
1002 *
1003 * p->pi_state_list -> pi_state->list, relation
1004 *
1005 * pi_state->refcount:
1006 *
1007 * pi_state lifetime
1008 *
1009 *
1010 * Lock order:
1011 *
1012 * hb->lock
1013 * pi_mutex->wait_lock
1014 * p->pi_lock
1015 *
Thomas Gleixner54a21782014-06-03 12:27:08 +00001016 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001017
1018/*
1019 * Validate that the existing waiter has a pi_state and sanity check
1020 * the pi_state against the user space value. If correct, attach to
1021 * it.
1022 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001023static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1024 struct futex_pi_state *pi_state,
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001025 struct futex_pi_state **ps)
1026{
1027 pid_t pid = uval & FUTEX_TID_MASK;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001028 int ret, uval2;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001029
1030 /*
1031 * Userspace might have messed up non-PI and PI futexes [3]
1032 */
1033 if (unlikely(!pi_state))
1034 return -EINVAL;
1035
Peter Zijlstra734009e2017-03-22 11:35:52 +01001036 /*
1037 * We get here with hb->lock held, and having found a
1038 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1039 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1040 * which in turn means that futex_lock_pi() still has a reference on
1041 * our pi_state.
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001042 *
1043 * The waiter holding a reference on @pi_state also protects against
1044 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1045 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1046 * free pi_state before we can take a reference ourselves.
Peter Zijlstra734009e2017-03-22 11:35:52 +01001047 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001048 WARN_ON(!atomic_read(&pi_state->refcount));
1049
1050 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01001051 * Now that we have a pi_state, we can acquire wait_lock
1052 * and do the state validation.
1053 */
1054 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1055
1056 /*
1057 * Since {uval, pi_state} is serialized by wait_lock, and our current
1058 * uval was read without holding it, it can have changed. Verify it
1059 * still is what we expect it to be, otherwise retry the entire
1060 * operation.
1061 */
1062 if (get_futex_value_locked(&uval2, uaddr))
1063 goto out_efault;
1064
1065 if (uval != uval2)
1066 goto out_eagain;
1067
1068 /*
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001069 * Handle the owner died case:
1070 */
1071 if (uval & FUTEX_OWNER_DIED) {
1072 /*
1073 * exit_pi_state_list sets owner to NULL and wakes the
1074 * topmost waiter. The task which acquires the
1075 * pi_state->rt_mutex will fixup owner.
1076 */
1077 if (!pi_state->owner) {
1078 /*
1079 * No pi state owner, but the user space TID
1080 * is not 0. Inconsistent state. [5]
1081 */
1082 if (pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001083 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001084 /*
1085 * Take a ref on the state and return success. [4]
1086 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001087 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001088 }
1089
1090 /*
1091 * If TID is 0, then either the dying owner has not
1092 * yet executed exit_pi_state_list() or some waiter
1093 * acquired the rtmutex in the pi state, but did not
1094 * yet fixup the TID in user space.
1095 *
1096 * Take a ref on the state and return success. [6]
1097 */
1098 if (!pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001099 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001100 } else {
1101 /*
1102 * If the owner died bit is not set, then the pi_state
1103 * must have an owner. [7]
1104 */
1105 if (!pi_state->owner)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001106 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001107 }
1108
1109 /*
1110 * Bail out if user space manipulated the futex value. If pi
1111 * state exists then the owner TID must be the same as the
1112 * user space TID. [9/10]
1113 */
1114 if (pid != task_pid_vnr(pi_state->owner))
Peter Zijlstra734009e2017-03-22 11:35:52 +01001115 goto out_einval;
1116
1117out_attach:
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01001118 get_pi_state(pi_state);
Peter Zijlstra734009e2017-03-22 11:35:52 +01001119 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001120 *ps = pi_state;
1121 return 0;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001122
1123out_einval:
1124 ret = -EINVAL;
1125 goto out_error;
1126
1127out_eagain:
1128 ret = -EAGAIN;
1129 goto out_error;
1130
1131out_efault:
1132 ret = -EFAULT;
1133 goto out_error;
1134
1135out_error:
1136 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1137 return ret;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001138}
1139
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001140/*
1141 * Lookup the task for the TID provided from user space and attach to
1142 * it after doing proper sanity checks.
1143 */
1144static int attach_to_pi_owner(u32 uval, union futex_key *key,
1145 struct futex_pi_state **ps)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001146{
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001147 pid_t pid = uval & FUTEX_TID_MASK;
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001148 struct futex_pi_state *pi_state;
1149 struct task_struct *p;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001150
1151 /*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001152 * We are the first waiter - try to look up the real owner and attach
Thomas Gleixner54a21782014-06-03 12:27:08 +00001153 * the new pi_state to it, but bail out when TID = 0 [1]
Ingo Molnarc87e2832006-06-27 02:54:58 -07001154 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001155 if (!pid)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001156 return -ESRCH;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001157 p = futex_find_get_task(pid);
Michal Hocko7a0ea092010-06-30 09:51:19 +02001158 if (!p)
1159 return -ESRCH;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001160
Oleg Nesterova2129462015-02-02 15:05:36 +01001161 if (unlikely(p->flags & PF_KTHREAD)) {
Thomas Gleixnerf0d71b32014-05-12 20:45:35 +00001162 put_task_struct(p);
1163 return -EPERM;
1164 }
1165
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001166 /*
1167 * We need to look at the task state flags to figure out,
1168 * whether the task is exiting. To protect against the do_exit
1169 * change of the task flags, we do this protected by
1170 * p->pi_lock:
1171 */
Thomas Gleixner1d615482009-11-17 14:54:03 +01001172 raw_spin_lock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001173 if (unlikely(p->flags & PF_EXITING)) {
1174 /*
1175 * The task is on the way out. When PF_EXITPIDONE is
1176 * set, we know that the task has finished the
1177 * cleanup:
1178 */
1179 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1180
Thomas Gleixner1d615482009-11-17 14:54:03 +01001181 raw_spin_unlock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001182 put_task_struct(p);
1183 return ret;
1184 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001185
Thomas Gleixner54a21782014-06-03 12:27:08 +00001186 /*
1187 * No existing pi state. First waiter. [2]
Peter Zijlstra734009e2017-03-22 11:35:52 +01001188 *
1189 * This creates pi_state, we have hb->lock held, this means nothing can
1190 * observe this state, wait_lock is irrelevant.
Thomas Gleixner54a21782014-06-03 12:27:08 +00001191 */
Ingo Molnarc87e2832006-06-27 02:54:58 -07001192 pi_state = alloc_pi_state();
1193
1194 /*
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001195 * Initialize the pi_mutex in locked state and make @p
Ingo Molnarc87e2832006-06-27 02:54:58 -07001196 * the owner of it:
1197 */
1198 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1199
1200 /* Store the key for possible exit cleanups: */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001201 pi_state->key = *key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001202
Ingo Molnar627371d2006-07-29 05:16:20 +02001203 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001204 list_add(&pi_state->list, &p->pi_state_list);
1205 pi_state->owner = p;
Thomas Gleixner1d615482009-11-17 14:54:03 +01001206 raw_spin_unlock_irq(&p->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001207
1208 put_task_struct(p);
1209
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001210 *ps = pi_state;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001211
1212 return 0;
1213}
1214
Peter Zijlstra734009e2017-03-22 11:35:52 +01001215static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1216 struct futex_hash_bucket *hb,
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001217 union futex_key *key, struct futex_pi_state **ps)
1218{
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001219 struct futex_q *top_waiter = futex_top_waiter(hb, key);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001220
1221 /*
1222 * If there is a waiter on that futex, validate it and
1223 * attach to the pi_state when the validation succeeds.
1224 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001225 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001226 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001227
1228 /*
1229 * We are the first waiter - try to look up the owner based on
1230 * @uval and attach to it.
1231 */
1232 return attach_to_pi_owner(uval, key, ps);
1233}
1234
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001235static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1236{
1237 u32 uninitialized_var(curval);
1238
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001239 if (unlikely(should_fail_futex(true)))
1240 return -EFAULT;
1241
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001242 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1243 return -EFAULT;
1244
Peter Zijlstra734009e2017-03-22 11:35:52 +01001245 /* If user space value changed, let the caller retry */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001246 return curval != uval ? -EAGAIN : 0;
1247}
1248
Darren Hart1a520842009-04-03 13:39:52 -07001249/**
Darren Hartd96ee562009-09-21 22:30:22 -07001250 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
Darren Hartbab5bc92009-04-07 23:23:50 -07001251 * @uaddr: the pi futex user address
1252 * @hb: the pi futex hash bucket
1253 * @key: the futex key associated with uaddr and hb
1254 * @ps: the pi_state pointer where we store the result of the
1255 * lookup
1256 * @task: the task to perform the atomic lock work for. This will
1257 * be "current" except in the case of requeue pi.
1258 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart1a520842009-04-03 13:39:52 -07001259 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001260 * Return:
1261 * 0 - ready to wait;
1262 * 1 - acquired the lock;
Darren Hart1a520842009-04-03 13:39:52 -07001263 * <0 - error
1264 *
1265 * The hb->lock and futex_key refs shall be held by the caller.
1266 */
1267static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1268 union futex_key *key,
1269 struct futex_pi_state **ps,
Darren Hartbab5bc92009-04-07 23:23:50 -07001270 struct task_struct *task, int set_waiters)
Darren Hart1a520842009-04-03 13:39:52 -07001271{
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001272 u32 uval, newval, vpid = task_pid_vnr(task);
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001273 struct futex_q *top_waiter;
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001274 int ret;
Darren Hart1a520842009-04-03 13:39:52 -07001275
1276 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001277 * Read the user space value first so we can validate a few
1278 * things before proceeding further.
Darren Hart1a520842009-04-03 13:39:52 -07001279 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001280 if (get_futex_value_locked(&uval, uaddr))
Darren Hart1a520842009-04-03 13:39:52 -07001281 return -EFAULT;
1282
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001283 if (unlikely(should_fail_futex(true)))
1284 return -EFAULT;
1285
Darren Hart1a520842009-04-03 13:39:52 -07001286 /*
1287 * Detect deadlocks.
1288 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001289 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
Darren Hart1a520842009-04-03 13:39:52 -07001290 return -EDEADLK;
1291
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001292 if ((unlikely(should_fail_futex(true))))
1293 return -EDEADLK;
1294
Darren Hart1a520842009-04-03 13:39:52 -07001295 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001296 * Lookup existing state first. If it exists, try to attach to
1297 * its pi_state.
Darren Hart1a520842009-04-03 13:39:52 -07001298 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001299 top_waiter = futex_top_waiter(hb, key);
1300 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001301 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001302
1303 /*
1304 * No waiter and user TID is 0. We are here because the
1305 * waiters or the owner died bit is set or called from
1306 * requeue_cmp_pi or for whatever reason something took the
1307 * syscall.
1308 */
1309 if (!(uval & FUTEX_TID_MASK)) {
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001310 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001311 * We take over the futex. No other waiters and the user space
1312 * TID is 0. We preserve the owner died bit.
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001313 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001314 newval = uval & FUTEX_OWNER_DIED;
1315 newval |= vpid;
1316
1317 /* The futex requeue_pi code can enforce the waiters bit */
1318 if (set_waiters)
1319 newval |= FUTEX_WAITERS;
1320
1321 ret = lock_pi_update_atomic(uaddr, uval, newval);
1322 /* If the take over worked, return 1 */
1323 return ret < 0 ? ret : 1;
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001324 }
Darren Hart1a520842009-04-03 13:39:52 -07001325
Darren Hart1a520842009-04-03 13:39:52 -07001326 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001327 * First waiter. Set the waiters bit before attaching ourself to
1328 * the owner. If owner tries to unlock, it will be forced into
1329 * the kernel and blocked on hb->lock.
Darren Hart1a520842009-04-03 13:39:52 -07001330 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001331 newval = uval | FUTEX_WAITERS;
1332 ret = lock_pi_update_atomic(uaddr, uval, newval);
1333 if (ret)
1334 return ret;
Darren Hart1a520842009-04-03 13:39:52 -07001335 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001336 * If the update of the user space value succeeded, we try to
1337 * attach to the owner. If that fails, no harm done, we only
1338 * set the FUTEX_WAITERS bit in the user space variable.
Darren Hart1a520842009-04-03 13:39:52 -07001339 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001340 return attach_to_pi_owner(uval, key, ps);
Darren Hart1a520842009-04-03 13:39:52 -07001341}
1342
Lai Jiangshan2e129782010-12-22 14:18:50 +08001343/**
1344 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1345 * @q: The futex_q to unqueue
1346 *
1347 * The q->lock_ptr must not be NULL and must be held by the caller.
1348 */
1349static void __unqueue_futex(struct futex_q *q)
1350{
1351 struct futex_hash_bucket *hb;
1352
Steven Rostedt29096202011-03-17 15:21:07 -04001353 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1354 || WARN_ON(plist_node_empty(&q->list)))
Lai Jiangshan2e129782010-12-22 14:18:50 +08001355 return;
1356
1357 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1358 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001359 hb_waiters_dec(hb);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001360}
1361
Ingo Molnarc87e2832006-06-27 02:54:58 -07001362/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 * The hash bucket lock must be held when this is called.
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001364 * Afterwards, the futex_q must not be accessed. Callers
1365 * must ensure to later call wake_up_q() for the actual
1366 * wakeups to occur.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001368static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369{
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001370 struct task_struct *p = q->task;
1371
Darren Hartaa109902012-11-26 16:29:56 -08001372 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1373 return;
1374
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001375 /*
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001376 * Queue the task for later wakeup for after we've released
1377 * the hb->lock. wake_q_add() grabs reference to p.
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001378 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001379 wake_q_add(wake_q, p);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001380 __unqueue_futex(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 /*
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001382 * The waiting task can free the futex_q as soon as
1383 * q->lock_ptr = NULL is written, without taking any locks. A
1384 * memory barrier is required here to prevent the following
1385 * store to lock_ptr from getting ahead of the plist_del.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 */
Peter Zijlstra1b367ec2017-03-22 11:35:49 +01001387 smp_store_release(&q->lock_ptr, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388}
1389
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001390/*
1391 * Caller must hold a reference on @pi_state.
1392 */
1393static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001394{
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03001395 u32 uninitialized_var(curval), newval;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001396 struct task_struct *new_owner;
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001397 bool postunlock = false;
Waiman Long194a6b52016-11-17 11:46:38 -05001398 DEFINE_WAKE_Q(wake_q);
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001399 int ret = 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001400
Ingo Molnarc87e2832006-06-27 02:54:58 -07001401 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01001402 if (WARN_ON_ONCE(!new_owner)) {
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001403 /*
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01001404 * As per the comment in futex_unlock_pi() this should not happen.
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001405 *
1406 * When this happens, give up our locks and try again, giving
1407 * the futex_lock_pi() instance time to complete, either by
1408 * waiting on the rtmutex or removing itself from the futex
1409 * queue.
1410 */
1411 ret = -EAGAIN;
1412 goto out_unlock;
Peter Zijlstra73d786b2017-03-22 11:35:54 +01001413 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001414
1415 /*
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001416 * We pass it to the next owner. The WAITERS bit is always kept
1417 * enabled while there is PI state around. We cleanup the owner
1418 * died bit, because we are the owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07001419 */
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001420 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001421
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001422 if (unlikely(should_fail_futex(true)))
1423 ret = -EFAULT;
1424
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001425 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001426 ret = -EFAULT;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001427
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001428 } else if (curval != uval) {
1429 /*
1430 * If a unconditional UNLOCK_PI operation (user space did not
1431 * try the TID->0 transition) raced with a waiter setting the
1432 * FUTEX_WAITERS flag between get_user() and locking the hash
1433 * bucket lock, retry the operation.
1434 */
1435 if ((FUTEX_TID_MASK & curval) == uval)
1436 ret = -EAGAIN;
1437 else
1438 ret = -EINVAL;
1439 }
Peter Zijlstra734009e2017-03-22 11:35:52 +01001440
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001441 if (ret)
1442 goto out_unlock;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001443
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001444 raw_spin_lock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001445 WARN_ON(list_empty(&pi_state->list));
1446 list_del_init(&pi_state->list);
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001447 raw_spin_unlock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001448
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001449 raw_spin_lock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001450 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001451 list_add(&pi_state->list, &new_owner->pi_state_list);
1452 pi_state->owner = new_owner;
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001453 raw_spin_unlock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001454
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001455 /*
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001456 * We've updated the uservalue, this unlock cannot fail.
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001457 */
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001458 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001459
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001460out_unlock:
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001461 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001462
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001463 if (postunlock)
1464 rt_mutex_postunlock(&wake_q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001465
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001466 return ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001467}
1468
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469/*
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001470 * Express the locking dependencies for lockdep:
1471 */
1472static inline void
1473double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1474{
1475 if (hb1 <= hb2) {
1476 spin_lock(&hb1->lock);
1477 if (hb1 < hb2)
1478 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1479 } else { /* hb1 > hb2 */
1480 spin_lock(&hb2->lock);
1481 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1482 }
1483}
1484
Darren Hart5eb3dc62009-03-12 00:55:52 -07001485static inline void
1486double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1487{
Darren Hartf061d352009-03-12 15:11:18 -07001488 spin_unlock(&hb1->lock);
Ingo Molnar88f502f2009-03-13 10:32:07 +01001489 if (hb1 != hb2)
1490 spin_unlock(&hb2->lock);
Darren Hart5eb3dc62009-03-12 00:55:52 -07001491}
1492
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001493/*
Darren Hartb2d09942009-03-12 00:55:37 -07001494 * Wake up waiters matching bitset queued on this futex (uaddr).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 */
Darren Hartb41277d2010-11-08 13:10:09 -08001496static int
1497futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498{
Ingo Molnare2970f22006-06-27 02:54:47 -07001499 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 struct futex_q *this, *next;
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001501 union futex_key key = FUTEX_KEY_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 int ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001503 DEFINE_WAKE_Q(wake_q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504
Thomas Gleixnercd689982008-02-01 17:45:14 +01001505 if (!bitset)
1506 return -EINVAL;
1507
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001508 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 if (unlikely(ret != 0))
1510 goto out;
1511
Ingo Molnare2970f22006-06-27 02:54:47 -07001512 hb = hash_futex(&key);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001513
1514 /* Make sure we really have tasks to wakeup */
1515 if (!hb_waiters_pending(hb))
1516 goto out_put_key;
1517
Ingo Molnare2970f22006-06-27 02:54:47 -07001518 spin_lock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519
Jason Low0d00c7b2014-01-12 15:31:22 -08001520 plist_for_each_entry_safe(this, next, &hb->chain, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 if (match_futex (&this->key, &key)) {
Darren Hart52400ba2009-04-03 13:40:49 -07001522 if (this->pi_state || this->rt_waiter) {
Ingo Molnared6f7b12006-07-01 04:35:46 -07001523 ret = -EINVAL;
1524 break;
1525 }
Thomas Gleixnercd689982008-02-01 17:45:14 +01001526
1527 /* Check if one of the bits is set in both bitsets */
1528 if (!(this->bitset & bitset))
1529 continue;
1530
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001531 mark_wake_futex(&wake_q, this);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 if (++ret >= nr_wake)
1533 break;
1534 }
1535 }
1536
Ingo Molnare2970f22006-06-27 02:54:47 -07001537 spin_unlock(&hb->lock);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001538 wake_up_q(&wake_q);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001539out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001540 put_futex_key(&key);
Darren Hart42d35d42008-12-29 15:49:53 -08001541out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 return ret;
1543}
1544
1545/*
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001546 * Wake up all waiters hashed on the physical page that is mapped
1547 * to this virtual address:
1548 */
Ingo Molnare2970f22006-06-27 02:54:47 -07001549static int
Darren Hartb41277d2010-11-08 13:10:09 -08001550futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
Ingo Molnare2970f22006-06-27 02:54:47 -07001551 int nr_wake, int nr_wake2, int op)
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001552{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001553 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Ingo Molnare2970f22006-06-27 02:54:47 -07001554 struct futex_hash_bucket *hb1, *hb2;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001555 struct futex_q *this, *next;
Darren Harte4dc5b72009-03-12 00:56:13 -07001556 int ret, op_ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001557 DEFINE_WAKE_Q(wake_q);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001558
Darren Harte4dc5b72009-03-12 00:56:13 -07001559retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001560 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001561 if (unlikely(ret != 0))
1562 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001563 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001564 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001565 goto out_put_key1;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001566
Ingo Molnare2970f22006-06-27 02:54:47 -07001567 hb1 = hash_futex(&key1);
1568 hb2 = hash_futex(&key2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001569
Darren Harte4dc5b72009-03-12 00:56:13 -07001570retry_private:
Thomas Gleixnereaaea802009-10-04 09:34:17 +02001571 double_lock_hb(hb1, hb2);
Ingo Molnare2970f22006-06-27 02:54:47 -07001572 op_ret = futex_atomic_op_inuser(op, uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001573 if (unlikely(op_ret < 0)) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001574
Darren Hart5eb3dc62009-03-12 00:55:52 -07001575 double_unlock_hb(hb1, hb2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001576
David Howells7ee1dd32006-01-06 00:11:44 -08001577#ifndef CONFIG_MMU
Ingo Molnare2970f22006-06-27 02:54:47 -07001578 /*
1579 * we don't get EFAULT from MMU faults if we don't have an MMU,
1580 * but we might get them from range checking
1581 */
David Howells7ee1dd32006-01-06 00:11:44 -08001582 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001583 goto out_put_keys;
David Howells7ee1dd32006-01-06 00:11:44 -08001584#endif
1585
David Gibson796f8d92005-11-07 00:59:33 -08001586 if (unlikely(op_ret != -EFAULT)) {
1587 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001588 goto out_put_keys;
David Gibson796f8d92005-11-07 00:59:33 -08001589 }
1590
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001591 ret = fault_in_user_writeable(uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001592 if (ret)
Darren Hartde87fcc2009-03-12 00:55:46 -07001593 goto out_put_keys;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001594
Darren Hartb41277d2010-11-08 13:10:09 -08001595 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001596 goto retry_private;
1597
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001598 put_futex_key(&key2);
1599 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001600 goto retry;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001601 }
1602
Jason Low0d00c7b2014-01-12 15:31:22 -08001603 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001604 if (match_futex (&this->key, &key1)) {
Darren Hartaa109902012-11-26 16:29:56 -08001605 if (this->pi_state || this->rt_waiter) {
1606 ret = -EINVAL;
1607 goto out_unlock;
1608 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001609 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001610 if (++ret >= nr_wake)
1611 break;
1612 }
1613 }
1614
1615 if (op_ret > 0) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001616 op_ret = 0;
Jason Low0d00c7b2014-01-12 15:31:22 -08001617 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001618 if (match_futex (&this->key, &key2)) {
Darren Hartaa109902012-11-26 16:29:56 -08001619 if (this->pi_state || this->rt_waiter) {
1620 ret = -EINVAL;
1621 goto out_unlock;
1622 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001623 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001624 if (++op_ret >= nr_wake2)
1625 break;
1626 }
1627 }
1628 ret += op_ret;
1629 }
1630
Darren Hartaa109902012-11-26 16:29:56 -08001631out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07001632 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001633 wake_up_q(&wake_q);
Darren Hart42d35d42008-12-29 15:49:53 -08001634out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001635 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08001636out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001637 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08001638out:
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001639 return ret;
1640}
1641
Darren Hart9121e472009-04-03 13:40:31 -07001642/**
1643 * requeue_futex() - Requeue a futex_q from one hb to another
1644 * @q: the futex_q to requeue
1645 * @hb1: the source hash_bucket
1646 * @hb2: the target hash_bucket
1647 * @key2: the new key for the requeued futex_q
1648 */
1649static inline
1650void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1651 struct futex_hash_bucket *hb2, union futex_key *key2)
1652{
1653
1654 /*
1655 * If key1 and key2 hash to the same bucket, no need to
1656 * requeue.
1657 */
1658 if (likely(&hb1->chain != &hb2->chain)) {
1659 plist_del(&q->list, &hb1->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001660 hb_waiters_dec(hb1);
Linus Torvalds11d46162014-03-20 22:11:17 -07001661 hb_waiters_inc(hb2);
Davidlohr Buesofe1bce92016-04-20 20:09:24 -07001662 plist_add(&q->list, &hb2->chain);
Darren Hart9121e472009-04-03 13:40:31 -07001663 q->lock_ptr = &hb2->lock;
Darren Hart9121e472009-04-03 13:40:31 -07001664 }
1665 get_futex_key_refs(key2);
1666 q->key = *key2;
1667}
1668
Darren Hart52400ba2009-04-03 13:40:49 -07001669/**
1670 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
Darren Hartd96ee562009-09-21 22:30:22 -07001671 * @q: the futex_q
1672 * @key: the key of the requeue target futex
1673 * @hb: the hash_bucket of the requeue target futex
Darren Hart52400ba2009-04-03 13:40:49 -07001674 *
1675 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1676 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1677 * to the requeue target futex so the waiter can detect the wakeup on the right
1678 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
Darren Hartbeda2c72009-08-09 15:34:39 -07001679 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1680 * to protect access to the pi_state to fixup the owner later. Must be called
1681 * with both q->lock_ptr and hb->lock held.
Darren Hart52400ba2009-04-03 13:40:49 -07001682 */
1683static inline
Darren Hartbeda2c72009-08-09 15:34:39 -07001684void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1685 struct futex_hash_bucket *hb)
Darren Hart52400ba2009-04-03 13:40:49 -07001686{
Darren Hart52400ba2009-04-03 13:40:49 -07001687 get_futex_key_refs(key);
1688 q->key = *key;
1689
Lai Jiangshan2e129782010-12-22 14:18:50 +08001690 __unqueue_futex(q);
Darren Hart52400ba2009-04-03 13:40:49 -07001691
1692 WARN_ON(!q->rt_waiter);
1693 q->rt_waiter = NULL;
1694
Darren Hartbeda2c72009-08-09 15:34:39 -07001695 q->lock_ptr = &hb->lock;
Darren Hartbeda2c72009-08-09 15:34:39 -07001696
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001697 wake_up_state(q->task, TASK_NORMAL);
Darren Hart52400ba2009-04-03 13:40:49 -07001698}
1699
1700/**
1701 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
Darren Hartbab5bc92009-04-07 23:23:50 -07001702 * @pifutex: the user address of the to futex
1703 * @hb1: the from futex hash bucket, must be locked by the caller
1704 * @hb2: the to futex hash bucket, must be locked by the caller
1705 * @key1: the from futex key
1706 * @key2: the to futex key
1707 * @ps: address to store the pi_state pointer
1708 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart52400ba2009-04-03 13:40:49 -07001709 *
1710 * Try and get the lock on behalf of the top waiter if we can do it atomically.
Darren Hartbab5bc92009-04-07 23:23:50 -07001711 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1712 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1713 * hb1 and hb2 must be held by the caller.
Darren Hart52400ba2009-04-03 13:40:49 -07001714 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001715 * Return:
1716 * 0 - failed to acquire the lock atomically;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001717 * >0 - acquired the lock, return value is vpid of the top_waiter
Darren Hart52400ba2009-04-03 13:40:49 -07001718 * <0 - error
1719 */
1720static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1721 struct futex_hash_bucket *hb1,
1722 struct futex_hash_bucket *hb2,
1723 union futex_key *key1, union futex_key *key2,
Darren Hartbab5bc92009-04-07 23:23:50 -07001724 struct futex_pi_state **ps, int set_waiters)
Darren Hart52400ba2009-04-03 13:40:49 -07001725{
Darren Hartbab5bc92009-04-07 23:23:50 -07001726 struct futex_q *top_waiter = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07001727 u32 curval;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001728 int ret, vpid;
Darren Hart52400ba2009-04-03 13:40:49 -07001729
1730 if (get_futex_value_locked(&curval, pifutex))
1731 return -EFAULT;
1732
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001733 if (unlikely(should_fail_futex(true)))
1734 return -EFAULT;
1735
Darren Hartbab5bc92009-04-07 23:23:50 -07001736 /*
1737 * Find the top_waiter and determine if there are additional waiters.
1738 * If the caller intends to requeue more than 1 waiter to pifutex,
1739 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1740 * as we have means to handle the possible fault. If not, don't set
1741 * the bit unecessarily as it will force the subsequent unlock to enter
1742 * the kernel.
1743 */
Darren Hart52400ba2009-04-03 13:40:49 -07001744 top_waiter = futex_top_waiter(hb1, key1);
1745
1746 /* There are no waiters, nothing for us to do. */
1747 if (!top_waiter)
1748 return 0;
1749
Darren Hart84bc4af2009-08-13 17:36:53 -07001750 /* Ensure we requeue to the expected futex. */
1751 if (!match_futex(top_waiter->requeue_pi_key, key2))
1752 return -EINVAL;
1753
Darren Hart52400ba2009-04-03 13:40:49 -07001754 /*
Darren Hartbab5bc92009-04-07 23:23:50 -07001755 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1756 * the contended case or if set_waiters is 1. The pi_state is returned
1757 * in ps in contended cases.
Darren Hart52400ba2009-04-03 13:40:49 -07001758 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001759 vpid = task_pid_vnr(top_waiter->task);
Darren Hartbab5bc92009-04-07 23:23:50 -07001760 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1761 set_waiters);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001762 if (ret == 1) {
Darren Hartbeda2c72009-08-09 15:34:39 -07001763 requeue_pi_wake_futex(top_waiter, key2, hb2);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001764 return vpid;
1765 }
Darren Hart52400ba2009-04-03 13:40:49 -07001766 return ret;
1767}
1768
1769/**
1770 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
Randy Dunlapfb62db22010-10-13 11:02:34 -07001771 * @uaddr1: source futex user address
Darren Hartb41277d2010-11-08 13:10:09 -08001772 * @flags: futex flags (FLAGS_SHARED, etc.)
Randy Dunlapfb62db22010-10-13 11:02:34 -07001773 * @uaddr2: target futex user address
1774 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1775 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1776 * @cmpval: @uaddr1 expected value (or %NULL)
1777 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
Darren Hartb41277d2010-11-08 13:10:09 -08001778 * pi futex (pi to pi requeue is not supported)
Darren Hart52400ba2009-04-03 13:40:49 -07001779 *
1780 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1781 * uaddr2 atomically on behalf of the top waiter.
1782 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001783 * Return:
1784 * >=0 - on success, the number of tasks requeued or woken;
Darren Hart52400ba2009-04-03 13:40:49 -07001785 * <0 - on error
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 */
Darren Hartb41277d2010-11-08 13:10:09 -08001787static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1788 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1789 u32 *cmpval, int requeue_pi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001791 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Darren Hart52400ba2009-04-03 13:40:49 -07001792 int drop_count = 0, task_count = 0, ret;
1793 struct futex_pi_state *pi_state = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07001794 struct futex_hash_bucket *hb1, *hb2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 struct futex_q *this, *next;
Waiman Long194a6b52016-11-17 11:46:38 -05001796 DEFINE_WAKE_Q(wake_q);
Darren Hart52400ba2009-04-03 13:40:49 -07001797
1798 if (requeue_pi) {
1799 /*
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001800 * Requeue PI only works on two distinct uaddrs. This
1801 * check is only valid for private futexes. See below.
1802 */
1803 if (uaddr1 == uaddr2)
1804 return -EINVAL;
1805
1806 /*
Darren Hart52400ba2009-04-03 13:40:49 -07001807 * requeue_pi requires a pi_state, try to allocate it now
1808 * without any locks in case it fails.
1809 */
1810 if (refill_pi_state_cache())
1811 return -ENOMEM;
1812 /*
1813 * requeue_pi must wake as many tasks as it can, up to nr_wake
1814 * + nr_requeue, since it acquires the rt_mutex prior to
1815 * returning to userspace, so as to not leave the rt_mutex with
1816 * waiters and no owner. However, second and third wake-ups
1817 * cannot be predicted as they involve race conditions with the
1818 * first wake and a fault while looking up the pi_state. Both
1819 * pthread_cond_signal() and pthread_cond_broadcast() should
1820 * use nr_wake=1.
1821 */
1822 if (nr_wake != 1)
1823 return -EINVAL;
1824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
Darren Hart42d35d42008-12-29 15:49:53 -08001826retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001827 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 if (unlikely(ret != 0))
1829 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001830 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1831 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001833 goto out_put_key1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001835 /*
1836 * The check above which compares uaddrs is not sufficient for
1837 * shared futexes. We need to compare the keys:
1838 */
1839 if (requeue_pi && match_futex(&key1, &key2)) {
1840 ret = -EINVAL;
1841 goto out_put_keys;
1842 }
1843
Ingo Molnare2970f22006-06-27 02:54:47 -07001844 hb1 = hash_futex(&key1);
1845 hb2 = hash_futex(&key2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846
Darren Harte4dc5b72009-03-12 00:56:13 -07001847retry_private:
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001848 hb_waiters_inc(hb2);
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001849 double_lock_hb(hb1, hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850
Ingo Molnare2970f22006-06-27 02:54:47 -07001851 if (likely(cmpval != NULL)) {
1852 u32 curval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853
Ingo Molnare2970f22006-06-27 02:54:47 -07001854 ret = get_futex_value_locked(&curval, uaddr1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855
1856 if (unlikely(ret)) {
Darren Hart5eb3dc62009-03-12 00:55:52 -07001857 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001858 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859
Darren Harte4dc5b72009-03-12 00:56:13 -07001860 ret = get_user(curval, uaddr1);
1861 if (ret)
1862 goto out_put_keys;
1863
Darren Hartb41277d2010-11-08 13:10:09 -08001864 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001865 goto retry_private;
1866
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001867 put_futex_key(&key2);
1868 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001869 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 }
Ingo Molnare2970f22006-06-27 02:54:47 -07001871 if (curval != *cmpval) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 ret = -EAGAIN;
1873 goto out_unlock;
1874 }
1875 }
1876
Darren Hart52400ba2009-04-03 13:40:49 -07001877 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
Darren Hartbab5bc92009-04-07 23:23:50 -07001878 /*
1879 * Attempt to acquire uaddr2 and wake the top waiter. If we
1880 * intend to requeue waiters, force setting the FUTEX_WAITERS
1881 * bit. We force this here where we are able to easily handle
1882 * faults rather in the requeue loop below.
1883 */
Darren Hart52400ba2009-04-03 13:40:49 -07001884 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
Darren Hartbab5bc92009-04-07 23:23:50 -07001885 &key2, &pi_state, nr_requeue);
Darren Hart52400ba2009-04-03 13:40:49 -07001886
1887 /*
1888 * At this point the top_waiter has either taken uaddr2 or is
1889 * waiting on it. If the former, then the pi_state will not
1890 * exist yet, look it up one more time to ensure we have a
Thomas Gleixner866293e2014-05-12 20:45:34 +00001891 * reference to it. If the lock was taken, ret contains the
1892 * vpid of the top waiter task.
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001893 * If the lock was not taken, we have pi_state and an initial
1894 * refcount on it. In case of an error we have nothing.
Darren Hart52400ba2009-04-03 13:40:49 -07001895 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001896 if (ret > 0) {
Darren Hart52400ba2009-04-03 13:40:49 -07001897 WARN_ON(pi_state);
Darren Hart89061d32009-10-15 15:30:48 -07001898 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07001899 task_count++;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001900 /*
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001901 * If we acquired the lock, then the user space value
1902 * of uaddr2 should be vpid. It cannot be changed by
1903 * the top waiter as it is blocked on hb2 lock if it
1904 * tries to do so. If something fiddled with it behind
1905 * our back the pi state lookup might unearth it. So
1906 * we rather use the known value than rereading and
1907 * handing potential crap to lookup_pi_state.
1908 *
1909 * If that call succeeds then we have pi_state and an
1910 * initial refcount on it.
Thomas Gleixner866293e2014-05-12 20:45:34 +00001911 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001912 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07001913 }
1914
1915 switch (ret) {
1916 case 0:
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001917 /* We hold a reference on the pi state. */
Darren Hart52400ba2009-04-03 13:40:49 -07001918 break;
Thomas Gleixner4959f2d2015-12-19 20:07:40 +00001919
1920 /* If the above failed, then pi_state is NULL */
Darren Hart52400ba2009-04-03 13:40:49 -07001921 case -EFAULT:
1922 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001923 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001924 put_futex_key(&key2);
1925 put_futex_key(&key1);
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001926 ret = fault_in_user_writeable(uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07001927 if (!ret)
1928 goto retry;
1929 goto out;
1930 case -EAGAIN:
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001931 /*
1932 * Two reasons for this:
1933 * - Owner is exiting and we just wait for the
1934 * exit to complete.
1935 * - The user space value changed.
1936 */
Darren Hart52400ba2009-04-03 13:40:49 -07001937 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001938 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001939 put_futex_key(&key2);
1940 put_futex_key(&key1);
Darren Hart52400ba2009-04-03 13:40:49 -07001941 cond_resched();
1942 goto retry;
1943 default:
1944 goto out_unlock;
1945 }
1946 }
1947
Jason Low0d00c7b2014-01-12 15:31:22 -08001948 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Darren Hart52400ba2009-04-03 13:40:49 -07001949 if (task_count - nr_wake >= nr_requeue)
1950 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951
Darren Hart52400ba2009-04-03 13:40:49 -07001952 if (!match_futex(&this->key, &key1))
1953 continue;
1954
Darren Hart392741e2009-08-07 15:20:48 -07001955 /*
1956 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1957 * be paired with each other and no other futex ops.
Darren Hartaa109902012-11-26 16:29:56 -08001958 *
1959 * We should never be requeueing a futex_q with a pi_state,
1960 * which is awaiting a futex_unlock_pi().
Darren Hart392741e2009-08-07 15:20:48 -07001961 */
1962 if ((requeue_pi && !this->rt_waiter) ||
Darren Hartaa109902012-11-26 16:29:56 -08001963 (!requeue_pi && this->rt_waiter) ||
1964 this->pi_state) {
Darren Hart392741e2009-08-07 15:20:48 -07001965 ret = -EINVAL;
1966 break;
1967 }
Darren Hart52400ba2009-04-03 13:40:49 -07001968
1969 /*
1970 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1971 * lock, we already woke the top_waiter. If not, it will be
1972 * woken by futex_unlock_pi().
1973 */
1974 if (++task_count <= nr_wake && !requeue_pi) {
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001975 mark_wake_futex(&wake_q, this);
Darren Hart52400ba2009-04-03 13:40:49 -07001976 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 }
Darren Hart52400ba2009-04-03 13:40:49 -07001978
Darren Hart84bc4af2009-08-13 17:36:53 -07001979 /* Ensure we requeue to the expected futex for requeue_pi. */
1980 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1981 ret = -EINVAL;
1982 break;
1983 }
1984
Darren Hart52400ba2009-04-03 13:40:49 -07001985 /*
1986 * Requeue nr_requeue waiters and possibly one more in the case
1987 * of requeue_pi if we couldn't acquire the lock atomically.
1988 */
1989 if (requeue_pi) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001990 /*
1991 * Prepare the waiter to take the rt_mutex. Take a
1992 * refcount on the pi_state and store the pointer in
1993 * the futex_q object of the waiter.
1994 */
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01001995 get_pi_state(pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07001996 this->pi_state = pi_state;
1997 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
1998 this->rt_waiter,
Thomas Gleixnerc051b212014-05-22 03:25:50 +00001999 this->task);
Darren Hart52400ba2009-04-03 13:40:49 -07002000 if (ret == 1) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002001 /*
2002 * We got the lock. We do neither drop the
2003 * refcount on pi_state nor clear
2004 * this->pi_state because the waiter needs the
2005 * pi_state for cleaning up the user space
2006 * value. It will drop the refcount after
2007 * doing so.
2008 */
Darren Hartbeda2c72009-08-09 15:34:39 -07002009 requeue_pi_wake_futex(this, &key2, hb2);
Darren Hart89061d32009-10-15 15:30:48 -07002010 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07002011 continue;
2012 } else if (ret) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002013 /*
2014 * rt_mutex_start_proxy_lock() detected a
2015 * potential deadlock when we tried to queue
2016 * that waiter. Drop the pi_state reference
2017 * which we took above and remove the pointer
2018 * to the state from the waiters futex_q
2019 * object.
2020 */
Darren Hart52400ba2009-04-03 13:40:49 -07002021 this->pi_state = NULL;
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002022 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002023 /*
2024 * We stop queueing more waiters and let user
2025 * space deal with the mess.
2026 */
2027 break;
Darren Hart52400ba2009-04-03 13:40:49 -07002028 }
2029 }
2030 requeue_futex(this, hb1, hb2, &key2);
2031 drop_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 }
2033
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002034 /*
2035 * We took an extra initial reference to the pi_state either
2036 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2037 * need to drop it here again.
2038 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002039 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002040
2041out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07002042 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07002043 wake_up_q(&wake_q);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07002044 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045
Darren Hartcd84a422009-04-02 14:19:38 -07002046 /*
2047 * drop_futex_key_refs() must be called outside the spinlocks. During
2048 * the requeue we moved futex_q's from the hash bucket at key1 to the
2049 * one at key2 and updated their key pointer. We no longer need to
2050 * hold the references to key1.
2051 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 while (--drop_count >= 0)
Rusty Russell9adef582007-05-08 00:26:42 -07002053 drop_futex_key_refs(&key1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054
Darren Hart42d35d42008-12-29 15:49:53 -08002055out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002056 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08002057out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002058 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08002059out:
Darren Hart52400ba2009-04-03 13:40:49 -07002060 return ret ? ret : task_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061}
2062
2063/* The key must be already stored in q->key. */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002064static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002065 __acquires(&hb->lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002066{
Ingo Molnare2970f22006-06-27 02:54:47 -07002067 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068
Ingo Molnare2970f22006-06-27 02:54:47 -07002069 hb = hash_futex(&q->key);
Linus Torvalds11d46162014-03-20 22:11:17 -07002070
2071 /*
2072 * Increment the counter before taking the lock so that
2073 * a potential waker won't miss a to-be-slept task that is
2074 * waiting for the spinlock. This is safe as all queue_lock()
2075 * users end up calling queue_me(). Similarly, for housekeeping,
2076 * decrement the counter at queue_unlock() when some error has
2077 * occurred and we don't end up adding the task to the list.
2078 */
2079 hb_waiters_inc(hb);
2080
Ingo Molnare2970f22006-06-27 02:54:47 -07002081 q->lock_ptr = &hb->lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -08002083 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
Ingo Molnare2970f22006-06-27 02:54:47 -07002084 return hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085}
2086
Darren Hartd40d65c2009-09-21 22:30:15 -07002087static inline void
Jason Low0d00c7b2014-01-12 15:31:22 -08002088queue_unlock(struct futex_hash_bucket *hb)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002089 __releases(&hb->lock)
Darren Hartd40d65c2009-09-21 22:30:15 -07002090{
2091 spin_unlock(&hb->lock);
Linus Torvalds11d46162014-03-20 22:11:17 -07002092 hb_waiters_dec(hb);
Darren Hartd40d65c2009-09-21 22:30:15 -07002093}
2094
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002095static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096{
Pierre Peifferec92d082007-05-09 02:35:00 -07002097 int prio;
2098
2099 /*
2100 * The priority used to register this element is
2101 * - either the real thread-priority for the real-time threads
2102 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2103 * - or MAX_RT_PRIO for non-RT threads.
2104 * Thus, all RT-threads are woken first in priority order, and
2105 * the others are woken last, in FIFO order.
2106 */
2107 prio = min(current->normal_prio, MAX_RT_PRIO);
2108
2109 plist_node_init(&q->list, prio);
Pierre Peifferec92d082007-05-09 02:35:00 -07002110 plist_add(&q->list, &hb->chain);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002111 q->task = current;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002112}
2113
2114/**
2115 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2116 * @q: The futex_q to enqueue
2117 * @hb: The destination hash bucket
2118 *
2119 * The hb->lock must be held by the caller, and is released here. A call to
2120 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2121 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2122 * or nothing if the unqueue is done as part of the wake process and the unqueue
2123 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2124 * an example).
2125 */
2126static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2127 __releases(&hb->lock)
2128{
2129 __queue_me(q, hb);
Ingo Molnare2970f22006-06-27 02:54:47 -07002130 spin_unlock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131}
2132
Darren Hartd40d65c2009-09-21 22:30:15 -07002133/**
2134 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2135 * @q: The futex_q to unqueue
2136 *
2137 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2138 * be paired with exactly one earlier call to queue_me().
2139 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002140 * Return:
2141 * 1 - if the futex_q was still queued (and we removed unqueued it);
Darren Hartd40d65c2009-09-21 22:30:15 -07002142 * 0 - if the futex_q was already removed by the waking thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144static int unqueue_me(struct futex_q *q)
2145{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 spinlock_t *lock_ptr;
Ingo Molnare2970f22006-06-27 02:54:47 -07002147 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148
2149 /* In the common case we don't take the spinlock, which is nice. */
Darren Hart42d35d42008-12-29 15:49:53 -08002150retry:
Jianyu Zhan29b75eb2016-03-07 09:32:24 +08002151 /*
2152 * q->lock_ptr can change between this read and the following spin_lock.
2153 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2154 * optimizing lock_ptr out of the logic below.
2155 */
2156 lock_ptr = READ_ONCE(q->lock_ptr);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07002157 if (lock_ptr != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 spin_lock(lock_ptr);
2159 /*
2160 * q->lock_ptr can change between reading it and
2161 * spin_lock(), causing us to take the wrong lock. This
2162 * corrects the race condition.
2163 *
2164 * Reasoning goes like this: if we have the wrong lock,
2165 * q->lock_ptr must have changed (maybe several times)
2166 * between reading it and the spin_lock(). It can
2167 * change again after the spin_lock() but only if it was
2168 * already changed before the spin_lock(). It cannot,
2169 * however, change back to the original value. Therefore
2170 * we can detect whether we acquired the correct lock.
2171 */
2172 if (unlikely(lock_ptr != q->lock_ptr)) {
2173 spin_unlock(lock_ptr);
2174 goto retry;
2175 }
Lai Jiangshan2e129782010-12-22 14:18:50 +08002176 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002177
2178 BUG_ON(q->pi_state);
2179
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 spin_unlock(lock_ptr);
2181 ret = 1;
2182 }
2183
Rusty Russell9adef582007-05-08 00:26:42 -07002184 drop_futex_key_refs(&q->key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 return ret;
2186}
2187
Ingo Molnarc87e2832006-06-27 02:54:58 -07002188/*
2189 * PI futexes can not be requeued and must remove themself from the
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002190 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2191 * and dropped here.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002192 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002193static void unqueue_me_pi(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002194 __releases(q->lock_ptr)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002195{
Lai Jiangshan2e129782010-12-22 14:18:50 +08002196 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002197
2198 BUG_ON(!q->pi_state);
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002199 put_pi_state(q->pi_state);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002200 q->pi_state = NULL;
2201
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002202 spin_unlock(q->lock_ptr);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002203}
2204
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002205/*
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002206 * Fixup the pi_state owner with the new owner.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002207 *
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002208 * Must be called with hash bucket lock held and mm->sem held for non
2209 * private futexes.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002210 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002211static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002212 struct task_struct *newowner)
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002213{
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002214 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002215 struct futex_pi_state *pi_state = q->pi_state;
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03002216 u32 uval, uninitialized_var(curval), newval;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002217 struct task_struct *oldowner;
Darren Harte4dc5b72009-03-12 00:56:13 -07002218 int ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002219
Peter Zijlstra734009e2017-03-22 11:35:52 +01002220 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2221
2222 oldowner = pi_state->owner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002223 /* Owner died? */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002224 if (!pi_state->owner)
2225 newtid |= FUTEX_OWNER_DIED;
2226
2227 /*
2228 * We are here either because we stole the rtmutex from the
Lai Jiangshan81612392011-01-14 17:09:41 +08002229 * previous highest priority waiter or we are the highest priority
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002230 * waiter but have failed to get the rtmutex the first time.
2231 *
Lai Jiangshan81612392011-01-14 17:09:41 +08002232 * We have to replace the newowner TID in the user space variable.
2233 * This must be atomic as we have to preserve the owner died bit here.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002234 *
Darren Hartb2d09942009-03-12 00:55:37 -07002235 * Note: We write the user space value _before_ changing the pi_state
2236 * because we can fault here. Imagine swapped out pages or a fork
2237 * that marked all the anonymous memory readonly for cow.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002238 *
Peter Zijlstra734009e2017-03-22 11:35:52 +01002239 * Modifying pi_state _before_ the user space value would leave the
2240 * pi_state in an inconsistent state when we fault here, because we
2241 * need to drop the locks to handle the fault. This might be observed
2242 * in the PID check in lookup_pi_state.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002243 */
2244retry:
2245 if (get_futex_value_locked(&uval, uaddr))
2246 goto handle_fault;
2247
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002248 for (;;) {
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002249 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2250
Michel Lespinasse37a9d912011-03-10 18:48:51 -08002251 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002252 goto handle_fault;
2253 if (curval == uval)
2254 break;
2255 uval = curval;
2256 }
2257
2258 /*
2259 * We fixed up user space. Now we need to fix the pi_state
2260 * itself.
2261 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002262 if (pi_state->owner != NULL) {
Peter Zijlstra734009e2017-03-22 11:35:52 +01002263 raw_spin_lock(&pi_state->owner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002264 WARN_ON(list_empty(&pi_state->list));
2265 list_del_init(&pi_state->list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002266 raw_spin_unlock(&pi_state->owner->pi_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002267 }
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002268
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002269 pi_state->owner = newowner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002270
Peter Zijlstra734009e2017-03-22 11:35:52 +01002271 raw_spin_lock(&newowner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002272 WARN_ON(!list_empty(&pi_state->list));
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002273 list_add(&pi_state->list, &newowner->pi_state_list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002274 raw_spin_unlock(&newowner->pi_lock);
2275 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2276
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002277 return 0;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002278
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002279 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01002280 * To handle the page fault we need to drop the locks here. That gives
2281 * the other task (either the highest priority waiter itself or the
2282 * task which stole the rtmutex) the chance to try the fixup of the
2283 * pi_state. So once we are back from handling the fault we need to
2284 * check the pi_state after reacquiring the locks and before trying to
2285 * do another fixup. When the fixup has been done already we simply
2286 * return.
2287 *
2288 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2289 * drop hb->lock since the caller owns the hb -> futex_q relation.
2290 * Dropping the pi_mutex->wait_lock requires the state revalidate.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002291 */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002292handle_fault:
Peter Zijlstra734009e2017-03-22 11:35:52 +01002293 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002294 spin_unlock(q->lock_ptr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002295
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002296 ret = fault_in_user_writeable(uaddr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002297
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002298 spin_lock(q->lock_ptr);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002299 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002300
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002301 /*
2302 * Check if someone else fixed it for us:
2303 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01002304 if (pi_state->owner != oldowner) {
2305 ret = 0;
2306 goto out_unlock;
2307 }
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002308
2309 if (ret)
Peter Zijlstra734009e2017-03-22 11:35:52 +01002310 goto out_unlock;
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002311
2312 goto retry;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002313
2314out_unlock:
2315 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2316 return ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002317}
2318
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002319static long futex_wait_restart(struct restart_block *restart);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07002320
Darren Hartca5f9522009-04-03 13:39:33 -07002321/**
Darren Hartdd973992009-04-03 13:40:02 -07002322 * fixup_owner() - Post lock pi_state and corner case management
2323 * @uaddr: user address of the futex
Darren Hartdd973992009-04-03 13:40:02 -07002324 * @q: futex_q (contains pi_state and access to the rt_mutex)
2325 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2326 *
2327 * After attempting to lock an rt_mutex, this function is called to cleanup
2328 * the pi_state owner as well as handle race conditions that may allow us to
2329 * acquire the lock. Must be called with the hb lock held.
2330 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002331 * Return:
2332 * 1 - success, lock taken;
2333 * 0 - success, lock not taken;
Darren Hartdd973992009-04-03 13:40:02 -07002334 * <0 - on error (-EFAULT)
2335 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002336static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
Darren Hartdd973992009-04-03 13:40:02 -07002337{
Darren Hartdd973992009-04-03 13:40:02 -07002338 int ret = 0;
2339
2340 if (locked) {
2341 /*
2342 * Got the lock. We might not be the anticipated owner if we
2343 * did a lock-steal - fix up the PI-state in that case:
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002344 *
2345 * We can safely read pi_state->owner without holding wait_lock
2346 * because we now own the rt_mutex, only the owner will attempt
2347 * to change it.
Darren Hartdd973992009-04-03 13:40:02 -07002348 */
2349 if (q->pi_state->owner != current)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002350 ret = fixup_pi_state_owner(uaddr, q, current);
Darren Hartdd973992009-04-03 13:40:02 -07002351 goto out;
2352 }
2353
2354 /*
Darren Hartdd973992009-04-03 13:40:02 -07002355 * Paranoia check. If we did not take the lock, then we should not be
Lai Jiangshan81612392011-01-14 17:09:41 +08002356 * the owner of the rt_mutex.
Darren Hartdd973992009-04-03 13:40:02 -07002357 */
Peter Zijlstra73d786b2017-03-22 11:35:54 +01002358 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
Darren Hartdd973992009-04-03 13:40:02 -07002359 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2360 "pi-state %p\n", ret,
2361 q->pi_state->pi_mutex.owner,
2362 q->pi_state->owner);
Peter Zijlstra73d786b2017-03-22 11:35:54 +01002363 }
Darren Hartdd973992009-04-03 13:40:02 -07002364
2365out:
2366 return ret ? ret : locked;
2367}
2368
2369/**
Darren Hartca5f9522009-04-03 13:39:33 -07002370 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2371 * @hb: the futex hash bucket, must be locked by the caller
2372 * @q: the futex_q to queue up on
2373 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
Darren Hartca5f9522009-04-03 13:39:33 -07002374 */
2375static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002376 struct hrtimer_sleeper *timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002377{
Darren Hart9beba3c2009-09-24 11:54:47 -07002378 /*
2379 * The task state is guaranteed to be set before another task can
Peter Zijlstrab92b8b32015-05-12 10:51:55 +02002380 * wake it. set_current_state() is implemented using smp_store_mb() and
Darren Hart9beba3c2009-09-24 11:54:47 -07002381 * queue_me() calls spin_unlock() upon completion, both serializing
2382 * access to the hash list and forcing another memory barrier.
2383 */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002384 set_current_state(TASK_INTERRUPTIBLE);
Darren Hart0729e192009-09-21 22:30:38 -07002385 queue_me(q, hb);
Darren Hartca5f9522009-04-03 13:39:33 -07002386
2387 /* Arm the timer */
Thomas Gleixner2e4b0d32015-04-14 21:09:13 +00002388 if (timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002389 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002390
2391 /*
Darren Hart0729e192009-09-21 22:30:38 -07002392 * If we have been removed from the hash list, then another task
2393 * has tried to wake us, and we can skip the call to schedule().
Darren Hartca5f9522009-04-03 13:39:33 -07002394 */
2395 if (likely(!plist_node_empty(&q->list))) {
2396 /*
2397 * If the timer has already expired, current will already be
2398 * flagged for rescheduling. Only call schedule if there
2399 * is no timeout, or if it has yet to expire.
2400 */
2401 if (!timeout || timeout->task)
Colin Cross88c80042013-05-01 18:35:05 -07002402 freezable_schedule();
Darren Hartca5f9522009-04-03 13:39:33 -07002403 }
2404 __set_current_state(TASK_RUNNING);
2405}
2406
Darren Hartf8010732009-04-03 13:40:40 -07002407/**
2408 * futex_wait_setup() - Prepare to wait on a futex
2409 * @uaddr: the futex userspace address
2410 * @val: the expected value
Darren Hartb41277d2010-11-08 13:10:09 -08002411 * @flags: futex flags (FLAGS_SHARED, etc.)
Darren Hartf8010732009-04-03 13:40:40 -07002412 * @q: the associated futex_q
2413 * @hb: storage for hash_bucket pointer to be returned to caller
2414 *
2415 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2416 * compare it with the expected value. Handle atomic faults internally.
2417 * Return with the hb lock held and a q.key reference on success, and unlocked
2418 * with no q.key reference on failure.
2419 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002420 * Return:
2421 * 0 - uaddr contains val and hb has been locked;
Bart Van Asscheca4a04c2011-07-17 09:01:00 +02002422 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
Darren Hartf8010732009-04-03 13:40:40 -07002423 */
Darren Hartb41277d2010-11-08 13:10:09 -08002424static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
Darren Hartf8010732009-04-03 13:40:40 -07002425 struct futex_q *q, struct futex_hash_bucket **hb)
2426{
2427 u32 uval;
2428 int ret;
2429
2430 /*
2431 * Access the page AFTER the hash-bucket is locked.
2432 * Order is important:
2433 *
2434 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2435 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2436 *
2437 * The basic logical guarantee of a futex is that it blocks ONLY
2438 * if cond(var) is known to be true at the time of blocking, for
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002439 * any cond. If we locked the hash-bucket after testing *uaddr, that
2440 * would open a race condition where we could block indefinitely with
Darren Hartf8010732009-04-03 13:40:40 -07002441 * cond(var) false, which would violate the guarantee.
2442 *
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002443 * On the other hand, we insert q and release the hash-bucket only
2444 * after testing *uaddr. This guarantees that futex_wait() will NOT
2445 * absorb a wakeup if *uaddr does not match the desired values
2446 * while the syscall executes.
Darren Hartf8010732009-04-03 13:40:40 -07002447 */
2448retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002449 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
Darren Hartf8010732009-04-03 13:40:40 -07002450 if (unlikely(ret != 0))
Darren Harta5a2a0c2009-04-10 09:50:05 -07002451 return ret;
Darren Hartf8010732009-04-03 13:40:40 -07002452
2453retry_private:
2454 *hb = queue_lock(q);
2455
2456 ret = get_futex_value_locked(&uval, uaddr);
2457
2458 if (ret) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002459 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002460
2461 ret = get_user(uval, uaddr);
2462 if (ret)
2463 goto out;
2464
Darren Hartb41277d2010-11-08 13:10:09 -08002465 if (!(flags & FLAGS_SHARED))
Darren Hartf8010732009-04-03 13:40:40 -07002466 goto retry_private;
2467
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002468 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002469 goto retry;
2470 }
2471
2472 if (uval != val) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002473 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002474 ret = -EWOULDBLOCK;
2475 }
2476
2477out:
2478 if (ret)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002479 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002480 return ret;
2481}
2482
Darren Hartb41277d2010-11-08 13:10:09 -08002483static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2484 ktime_t *abs_time, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485{
Darren Hartca5f9522009-04-03 13:39:33 -07002486 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002487 struct restart_block *restart;
Ingo Molnare2970f22006-06-27 02:54:47 -07002488 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002489 struct futex_q q = futex_q_init;
Ingo Molnare2970f22006-06-27 02:54:47 -07002490 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491
Thomas Gleixnercd689982008-02-01 17:45:14 +01002492 if (!bitset)
2493 return -EINVAL;
Thomas Gleixnercd689982008-02-01 17:45:14 +01002494 q.bitset = bitset;
Darren Hartca5f9522009-04-03 13:39:33 -07002495
2496 if (abs_time) {
2497 to = &timeout;
2498
Darren Hartb41277d2010-11-08 13:10:09 -08002499 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2500 CLOCK_REALTIME : CLOCK_MONOTONIC,
2501 HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002502 hrtimer_init_sleeper(to, current);
2503 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2504 current->timer_slack_ns);
2505 }
2506
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002507retry:
Darren Hart7ada8762010-10-17 08:35:04 -07002508 /*
2509 * Prepare to wait on uaddr. On success, holds hb lock and increments
2510 * q.key refs.
2511 */
Darren Hartb41277d2010-11-08 13:10:09 -08002512 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Darren Hartf8010732009-04-03 13:40:40 -07002513 if (ret)
Darren Hart42d35d42008-12-29 15:49:53 -08002514 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515
Darren Hartca5f9522009-04-03 13:39:33 -07002516 /* queue_me and wait for wakeup, timeout, or a signal. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002517 futex_wait_queue_me(hb, &q, to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
2519 /* If we were woken (and unqueued), we succeeded, whatever. */
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002520 ret = 0;
Darren Hart7ada8762010-10-17 08:35:04 -07002521 /* unqueue_me() drops q.key ref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 if (!unqueue_me(&q))
Darren Hart7ada8762010-10-17 08:35:04 -07002523 goto out;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002524 ret = -ETIMEDOUT;
Darren Hartca5f9522009-04-03 13:39:33 -07002525 if (to && !to->task)
Darren Hart7ada8762010-10-17 08:35:04 -07002526 goto out;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002527
Ingo Molnare2970f22006-06-27 02:54:47 -07002528 /*
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002529 * We expect signal_pending(current), but we might be the
2530 * victim of a spurious wakeup as well.
Ingo Molnare2970f22006-06-27 02:54:47 -07002531 */
Darren Hart7ada8762010-10-17 08:35:04 -07002532 if (!signal_pending(current))
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002533 goto retry;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002534
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002535 ret = -ERESTARTSYS;
Pierre Peifferc19384b2007-05-09 02:35:02 -07002536 if (!abs_time)
Darren Hart7ada8762010-10-17 08:35:04 -07002537 goto out;
Steven Rostedtce6bd422007-12-05 15:46:09 +01002538
Andy Lutomirskif56141e2015-02-12 15:01:14 -08002539 restart = &current->restart_block;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002540 restart->fn = futex_wait_restart;
Namhyung Kima3c74c52010-09-14 21:43:47 +09002541 restart->futex.uaddr = uaddr;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002542 restart->futex.val = val;
Thomas Gleixner2456e852016-12-25 11:38:40 +01002543 restart->futex.time = *abs_time;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002544 restart->futex.bitset = bitset;
Darren Hart0cd9c642011-04-14 15:41:57 -07002545 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002546
2547 ret = -ERESTART_RESTARTBLOCK;
2548
Darren Hart42d35d42008-12-29 15:49:53 -08002549out:
Darren Hartca5f9522009-04-03 13:39:33 -07002550 if (to) {
2551 hrtimer_cancel(&to->timer);
2552 destroy_hrtimer_on_stack(&to->timer);
2553 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002554 return ret;
2555}
2556
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002557
2558static long futex_wait_restart(struct restart_block *restart)
2559{
Namhyung Kima3c74c52010-09-14 21:43:47 +09002560 u32 __user *uaddr = restart->futex.uaddr;
Darren Harta72188d2009-04-03 13:40:22 -07002561 ktime_t t, *tp = NULL;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002562
Darren Harta72188d2009-04-03 13:40:22 -07002563 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
Thomas Gleixner2456e852016-12-25 11:38:40 +01002564 t = restart->futex.time;
Darren Harta72188d2009-04-03 13:40:22 -07002565 tp = &t;
2566 }
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002567 restart->fn = do_no_restart_syscall;
Darren Hartb41277d2010-11-08 13:10:09 -08002568
2569 return (long)futex_wait(uaddr, restart->futex.flags,
2570 restart->futex.val, tp, restart->futex.bitset);
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002571}
2572
2573
Ingo Molnarc87e2832006-06-27 02:54:58 -07002574/*
2575 * Userspace tried a 0 -> TID atomic transition of the futex value
2576 * and failed. The kernel side here does the whole locking operation:
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002577 * if there are waiters then it will block as a consequence of relying
2578 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2579 * a 0 value of the futex too.).
2580 *
2581 * Also serves as futex trylock_pi()'ing, and due semantics.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002582 */
Michael Kerrisk996636d2015-01-16 20:28:06 +01002583static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
Darren Hartb41277d2010-11-08 13:10:09 -08002584 ktime_t *time, int trylock)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002585{
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002586 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002587 struct futex_pi_state *pi_state = NULL;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002588 struct rt_mutex_waiter rt_waiter;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002589 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002590 struct futex_q q = futex_q_init;
Darren Hartdd973992009-04-03 13:40:02 -07002591 int res, ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002592
2593 if (refill_pi_state_cache())
2594 return -ENOMEM;
2595
Pierre Peifferc19384b2007-05-09 02:35:02 -07002596 if (time) {
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002597 to = &timeout;
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002598 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2599 HRTIMER_MODE_ABS);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002600 hrtimer_init_sleeper(to, current);
Arjan van de Vencc584b22008-09-01 15:02:30 -07002601 hrtimer_set_expires(&to->timer, *time);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002602 }
2603
Darren Hart42d35d42008-12-29 15:49:53 -08002604retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002605 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002606 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08002607 goto out;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002608
Darren Harte4dc5b72009-03-12 00:56:13 -07002609retry_private:
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002610 hb = queue_lock(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002611
Darren Hartbab5bc92009-04-07 23:23:50 -07002612 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002613 if (unlikely(ret)) {
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002614 /*
2615 * Atomic work succeeded and we got the lock,
2616 * or failed. Either way, we do _not_ block.
2617 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002618 switch (ret) {
Darren Hart1a520842009-04-03 13:39:52 -07002619 case 1:
2620 /* We got the lock. */
2621 ret = 0;
2622 goto out_unlock_put_key;
2623 case -EFAULT:
2624 goto uaddr_faulted;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002625 case -EAGAIN:
2626 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00002627 * Two reasons for this:
2628 * - Task is exiting and we just wait for the
2629 * exit to complete.
2630 * - The user space value changed.
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002631 */
Jason Low0d00c7b2014-01-12 15:31:22 -08002632 queue_unlock(hb);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002633 put_futex_key(&q.key);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002634 cond_resched();
2635 goto retry;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002636 default:
Darren Hart42d35d42008-12-29 15:49:53 -08002637 goto out_unlock_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002638 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002639 }
2640
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002641 WARN_ON(!q.pi_state);
2642
Ingo Molnarc87e2832006-06-27 02:54:58 -07002643 /*
2644 * Only actually queue now that the atomic ops are done:
2645 */
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002646 __queue_me(&q, hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002647
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002648 if (trylock) {
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01002649 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002650 /* Fixup the trylock return value: */
2651 ret = ret ? 0 : -EWOULDBLOCK;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002652 goto no_block;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002653 }
2654
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002655 rt_mutex_init_waiter(&rt_waiter);
Peter Zijlstra56222b22017-03-22 11:36:00 +01002656
2657 /*
2658 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2659 * hold it while doing rt_mutex_start_proxy(), because then it will
2660 * include hb->lock in the blocking chain, even through we'll not in
2661 * fact hold it while blocking. This will lead it to report -EDEADLK
2662 * and BUG when futex_unlock_pi() interleaves with this.
2663 *
2664 * Therefore acquire wait_lock while holding hb->lock, but drop the
2665 * latter before calling rt_mutex_start_proxy_lock(). This still fully
2666 * serializes against futex_unlock_pi() as that does the exact same
2667 * lock handoff sequence.
2668 */
2669 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2670 spin_unlock(q.lock_ptr);
2671 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2672 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2673
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002674 if (ret) {
2675 if (ret == 1)
2676 ret = 0;
2677
Peter Zijlstra56222b22017-03-22 11:36:00 +01002678 spin_lock(q.lock_ptr);
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002679 goto no_block;
2680 }
2681
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002682
2683 if (unlikely(to))
2684 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2685
2686 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2687
Vernon Mauerya99e4e42006-07-01 04:35:42 -07002688 spin_lock(q.lock_ptr);
Darren Hartdd973992009-04-03 13:40:02 -07002689 /*
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002690 * If we failed to acquire the lock (signal/timeout), we must
2691 * first acquire the hb->lock before removing the lock from the
2692 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2693 * wait lists consistent.
Peter Zijlstra56222b22017-03-22 11:36:00 +01002694 *
2695 * In particular; it is important that futex_unlock_pi() can not
2696 * observe this inconsistency.
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002697 */
2698 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2699 ret = 0;
2700
2701no_block:
2702 /*
Darren Hartdd973992009-04-03 13:40:02 -07002703 * Fixup the pi_state owner and possibly acquire the lock if we
2704 * haven't already.
2705 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002706 res = fixup_owner(uaddr, &q, !ret);
Darren Hartdd973992009-04-03 13:40:02 -07002707 /*
2708 * If fixup_owner() returned an error, proprogate that. If it acquired
2709 * the lock, clear our -ETIMEDOUT or -EINTR.
2710 */
2711 if (res)
2712 ret = (res < 0) ? res : 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002713
Darren Harte8f63862009-03-12 00:56:06 -07002714 /*
Darren Hartdd973992009-04-03 13:40:02 -07002715 * If fixup_owner() faulted and was unable to handle the fault, unlock
2716 * it and return the fault to userspace.
Darren Harte8f63862009-03-12 00:56:06 -07002717 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002718 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2719 pi_state = q.pi_state;
2720 get_pi_state(pi_state);
2721 }
Darren Harte8f63862009-03-12 00:56:06 -07002722
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002723 /* Unqueue and drop the lock */
2724 unqueue_me_pi(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002725
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002726 if (pi_state) {
2727 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2728 put_pi_state(pi_state);
2729 }
2730
Mikael Pettersson5ecb01c2010-01-23 22:36:29 +01002731 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002732
Darren Hart42d35d42008-12-29 15:49:53 -08002733out_unlock_put_key:
Jason Low0d00c7b2014-01-12 15:31:22 -08002734 queue_unlock(hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002735
Darren Hart42d35d42008-12-29 15:49:53 -08002736out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002737 put_futex_key(&q.key);
Darren Hart42d35d42008-12-29 15:49:53 -08002738out:
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002739 if (to)
2740 destroy_hrtimer_on_stack(&to->timer);
Darren Hartdd973992009-04-03 13:40:02 -07002741 return ret != -EINTR ? ret : -ERESTARTNOINTR;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002742
Darren Hart42d35d42008-12-29 15:49:53 -08002743uaddr_faulted:
Jason Low0d00c7b2014-01-12 15:31:22 -08002744 queue_unlock(hb);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002745
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002746 ret = fault_in_user_writeable(uaddr);
Darren Harte4dc5b72009-03-12 00:56:13 -07002747 if (ret)
2748 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002749
Darren Hartb41277d2010-11-08 13:10:09 -08002750 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07002751 goto retry_private;
2752
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002753 put_futex_key(&q.key);
Darren Harte4dc5b72009-03-12 00:56:13 -07002754 goto retry;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002755}
2756
2757/*
Ingo Molnarc87e2832006-06-27 02:54:58 -07002758 * Userspace attempted a TID -> 0 atomic transition, and failed.
2759 * This is the in-kernel slowpath: we look up the PI state (if any),
2760 * and do the rt-mutex unlock.
2761 */
Darren Hartb41277d2010-11-08 13:10:09 -08002762static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002763{
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002764 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
Peter Zijlstra38d47c12008-09-26 19:32:20 +02002765 union futex_key key = FUTEX_KEY_INIT;
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002766 struct futex_hash_bucket *hb;
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002767 struct futex_q *top_waiter;
Darren Harte4dc5b72009-03-12 00:56:13 -07002768 int ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002769
2770retry:
2771 if (get_user(uval, uaddr))
2772 return -EFAULT;
2773 /*
2774 * We release only a lock we actually own:
2775 */
Thomas Gleixnerc0c9ed12011-03-11 11:51:22 +01002776 if ((uval & FUTEX_TID_MASK) != vpid)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002777 return -EPERM;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002778
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002779 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002780 if (ret)
2781 return ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002782
2783 hb = hash_futex(&key);
2784 spin_lock(&hb->lock);
2785
Ingo Molnarc87e2832006-06-27 02:54:58 -07002786 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002787 * Check waiters first. We do not trust user space values at
2788 * all and we at least want to know if user space fiddled
2789 * with the futex value instead of blindly unlocking.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002790 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002791 top_waiter = futex_top_waiter(hb, &key);
2792 if (top_waiter) {
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002793 struct futex_pi_state *pi_state = top_waiter->pi_state;
2794
2795 ret = -EINVAL;
2796 if (!pi_state)
2797 goto out_unlock;
2798
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002799 /*
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002800 * If current does not own the pi_state then the futex is
2801 * inconsistent and user space fiddled with the futex value.
2802 */
2803 if (pi_state->owner != current)
2804 goto out_unlock;
2805
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002806 get_pi_state(pi_state);
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01002807 /*
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01002808 * By taking wait_lock while still holding hb->lock, we ensure
2809 * there is no point where we hold neither; and therefore
2810 * wake_futex_pi() must observe a state consistent with what we
2811 * observed.
2812 */
2813 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002814 spin_unlock(&hb->lock);
2815
2816 ret = wake_futex_pi(uaddr, uval, pi_state);
2817
2818 put_pi_state(pi_state);
2819
2820 /*
2821 * Success, we're done! No tricky corner cases.
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002822 */
2823 if (!ret)
2824 goto out_putkey;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002825 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002826 * The atomic access to the futex value generated a
2827 * pagefault, so retry the user-access and the wakeup:
Ingo Molnarc87e2832006-06-27 02:54:58 -07002828 */
2829 if (ret == -EFAULT)
2830 goto pi_faulted;
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002831 /*
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02002832 * A unconditional UNLOCK_PI op raced against a waiter
2833 * setting the FUTEX_WAITERS bit. Try again.
2834 */
2835 if (ret == -EAGAIN) {
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02002836 put_futex_key(&key);
2837 goto retry;
2838 }
2839 /*
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002840 * wake_futex_pi has detected invalid state. Tell user
2841 * space.
2842 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002843 goto out_putkey;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002844 }
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002845
Ingo Molnarc87e2832006-06-27 02:54:58 -07002846 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002847 * We have no kernel internal state, i.e. no waiters in the
2848 * kernel. Waiters which are about to queue themselves are stuck
2849 * on hb->lock. So we can safely ignore them. We do neither
2850 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2851 * owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002852 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002853 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
2854 spin_unlock(&hb->lock);
Thomas Gleixner13fbca42014-06-03 12:27:07 +00002855 goto pi_faulted;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002856 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002857
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002858 /*
2859 * If uval has changed, let user space handle it.
2860 */
2861 ret = (curval == uval) ? 0 : -EAGAIN;
2862
Ingo Molnarc87e2832006-06-27 02:54:58 -07002863out_unlock:
2864 spin_unlock(&hb->lock);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002865out_putkey:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002866 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002867 return ret;
2868
2869pi_faulted:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002870 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002871
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002872 ret = fault_in_user_writeable(uaddr);
Darren Hartb5686362008-12-18 15:06:34 -08002873 if (!ret)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002874 goto retry;
2875
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 return ret;
2877}
2878
Darren Hart52400ba2009-04-03 13:40:49 -07002879/**
2880 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2881 * @hb: the hash_bucket futex_q was original enqueued on
2882 * @q: the futex_q woken while waiting to be requeued
2883 * @key2: the futex_key of the requeue target futex
2884 * @timeout: the timeout associated with the wait (NULL if none)
2885 *
2886 * Detect if the task was woken on the initial futex as opposed to the requeue
2887 * target futex. If so, determine if it was a timeout or a signal that caused
2888 * the wakeup and return the appropriate error code to the caller. Must be
2889 * called with the hb lock held.
2890 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002891 * Return:
2892 * 0 = no early wakeup detected;
2893 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
Darren Hart52400ba2009-04-03 13:40:49 -07002894 */
2895static inline
2896int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2897 struct futex_q *q, union futex_key *key2,
2898 struct hrtimer_sleeper *timeout)
2899{
2900 int ret = 0;
2901
2902 /*
2903 * With the hb lock held, we avoid races while we process the wakeup.
2904 * We only need to hold hb (and not hb2) to ensure atomicity as the
2905 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2906 * It can't be requeued from uaddr2 to something else since we don't
2907 * support a PI aware source futex for requeue.
2908 */
2909 if (!match_futex(&q->key, key2)) {
2910 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2911 /*
2912 * We were woken prior to requeue by a timeout or a signal.
2913 * Unqueue the futex_q and determine which it was.
2914 */
Lai Jiangshan2e129782010-12-22 14:18:50 +08002915 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07002916 hb_waiters_dec(hb);
Darren Hart52400ba2009-04-03 13:40:49 -07002917
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002918 /* Handle spurious wakeups gracefully */
Thomas Gleixner11df6dd2009-10-28 20:26:48 +01002919 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07002920 if (timeout && !timeout->task)
2921 ret = -ETIMEDOUT;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002922 else if (signal_pending(current))
Thomas Gleixner1c840c12009-05-20 09:22:40 +02002923 ret = -ERESTARTNOINTR;
Darren Hart52400ba2009-04-03 13:40:49 -07002924 }
2925 return ret;
2926}
2927
2928/**
2929 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
Darren Hart56ec1602009-09-21 22:29:59 -07002930 * @uaddr: the futex we initially wait on (non-pi)
Darren Hartb41277d2010-11-08 13:10:09 -08002931 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07002932 * the same type, no requeueing from private to shared, etc.
Darren Hart52400ba2009-04-03 13:40:49 -07002933 * @val: the expected value of uaddr
2934 * @abs_time: absolute timeout
Darren Hart56ec1602009-09-21 22:29:59 -07002935 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
Darren Hart52400ba2009-04-03 13:40:49 -07002936 * @uaddr2: the pi futex we will take prior to returning to user-space
2937 *
2938 * The caller will wait on uaddr and will be requeued by futex_requeue() to
Darren Hart6f7b0a22012-07-20 11:53:31 -07002939 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2940 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2941 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2942 * without one, the pi logic would not know which task to boost/deboost, if
2943 * there was a need to.
Darren Hart52400ba2009-04-03 13:40:49 -07002944 *
2945 * We call schedule in futex_wait_queue_me() when we enqueue and return there
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002946 * via the following--
Darren Hart52400ba2009-04-03 13:40:49 -07002947 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
Darren Hartcc6db4e2009-07-31 16:20:10 -07002948 * 2) wakeup on uaddr2 after a requeue
2949 * 3) signal
2950 * 4) timeout
Darren Hart52400ba2009-04-03 13:40:49 -07002951 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002952 * If 3, cleanup and return -ERESTARTNOINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07002953 *
2954 * If 2, we may then block on trying to take the rt_mutex and return via:
2955 * 5) successful lock
2956 * 6) signal
2957 * 7) timeout
2958 * 8) other lock acquisition failure
2959 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002960 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
Darren Hart52400ba2009-04-03 13:40:49 -07002961 *
2962 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2963 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002964 * Return:
2965 * 0 - On success;
Darren Hart52400ba2009-04-03 13:40:49 -07002966 * <0 - On error
2967 */
Darren Hartb41277d2010-11-08 13:10:09 -08002968static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
Darren Hart52400ba2009-04-03 13:40:49 -07002969 u32 val, ktime_t *abs_time, u32 bitset,
Darren Hartb41277d2010-11-08 13:10:09 -08002970 u32 __user *uaddr2)
Darren Hart52400ba2009-04-03 13:40:49 -07002971{
2972 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002973 struct futex_pi_state *pi_state = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07002974 struct rt_mutex_waiter rt_waiter;
Darren Hart52400ba2009-04-03 13:40:49 -07002975 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002976 union futex_key key2 = FUTEX_KEY_INIT;
2977 struct futex_q q = futex_q_init;
Darren Hart52400ba2009-04-03 13:40:49 -07002978 int res, ret;
Darren Hart52400ba2009-04-03 13:40:49 -07002979
Darren Hart6f7b0a22012-07-20 11:53:31 -07002980 if (uaddr == uaddr2)
2981 return -EINVAL;
2982
Darren Hart52400ba2009-04-03 13:40:49 -07002983 if (!bitset)
2984 return -EINVAL;
2985
2986 if (abs_time) {
2987 to = &timeout;
Darren Hartb41277d2010-11-08 13:10:09 -08002988 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2989 CLOCK_REALTIME : CLOCK_MONOTONIC,
2990 HRTIMER_MODE_ABS);
Darren Hart52400ba2009-04-03 13:40:49 -07002991 hrtimer_init_sleeper(to, current);
2992 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2993 current->timer_slack_ns);
2994 }
2995
2996 /*
2997 * The waiter is allocated on our stack, manipulated by the requeue
2998 * code while we sleep on uaddr.
2999 */
Peter Zijlstra50809352017-03-22 11:35:56 +01003000 rt_mutex_init_waiter(&rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003001
Shawn Bohrer9ea71502011-06-30 11:21:32 -05003002 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Darren Hart52400ba2009-04-03 13:40:49 -07003003 if (unlikely(ret != 0))
3004 goto out;
3005
Darren Hart84bc4af2009-08-13 17:36:53 -07003006 q.bitset = bitset;
3007 q.rt_waiter = &rt_waiter;
3008 q.requeue_pi_key = &key2;
3009
Darren Hart7ada8762010-10-17 08:35:04 -07003010 /*
3011 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3012 * count.
3013 */
Darren Hartb41277d2010-11-08 13:10:09 -08003014 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02003015 if (ret)
3016 goto out_key2;
Darren Hart52400ba2009-04-03 13:40:49 -07003017
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00003018 /*
3019 * The check above which compares uaddrs is not sufficient for
3020 * shared futexes. We need to compare the keys:
3021 */
3022 if (match_futex(&q.key, &key2)) {
Thomas Gleixner13c42c22014-09-11 23:44:35 +02003023 queue_unlock(hb);
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00003024 ret = -EINVAL;
3025 goto out_put_keys;
3026 }
3027
Darren Hart52400ba2009-04-03 13:40:49 -07003028 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02003029 futex_wait_queue_me(hb, &q, to);
Darren Hart52400ba2009-04-03 13:40:49 -07003030
3031 spin_lock(&hb->lock);
3032 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3033 spin_unlock(&hb->lock);
3034 if (ret)
3035 goto out_put_keys;
3036
3037 /*
3038 * In order for us to be here, we know our q.key == key2, and since
3039 * we took the hb->lock above, we also know that futex_requeue() has
3040 * completed and we no longer have to concern ourselves with a wakeup
Darren Hart7ada8762010-10-17 08:35:04 -07003041 * race with the atomic proxy lock acquisition by the requeue code. The
3042 * futex_requeue dropped our key1 reference and incremented our key2
3043 * reference count.
Darren Hart52400ba2009-04-03 13:40:49 -07003044 */
3045
3046 /* Check if the requeue code acquired the second futex for us. */
3047 if (!q.rt_waiter) {
3048 /*
3049 * Got the lock. We might not be the anticipated owner if we
3050 * did a lock-steal - fix up the PI-state in that case.
3051 */
3052 if (q.pi_state && (q.pi_state->owner != current)) {
3053 spin_lock(q.lock_ptr);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003054 ret = fixup_pi_state_owner(uaddr2, &q, current);
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003055 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3056 pi_state = q.pi_state;
3057 get_pi_state(pi_state);
3058 }
Thomas Gleixnerfb75a422015-12-19 20:07:38 +00003059 /*
3060 * Drop the reference to the pi state which
3061 * the requeue_pi() code acquired for us.
3062 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00003063 put_pi_state(q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003064 spin_unlock(q.lock_ptr);
3065 }
3066 } else {
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003067 struct rt_mutex *pi_mutex;
3068
Darren Hart52400ba2009-04-03 13:40:49 -07003069 /*
3070 * We have been woken up by futex_unlock_pi(), a timeout, or a
3071 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3072 * the pi_state.
3073 */
Darren Hartf27071c2012-07-20 11:53:30 -07003074 WARN_ON(!q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003075 pi_mutex = &q.pi_state->pi_mutex;
Peter Zijlstra38d589f2017-03-22 11:35:57 +01003076 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003077
3078 spin_lock(q.lock_ptr);
Peter Zijlstra38d589f2017-03-22 11:35:57 +01003079 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3080 ret = 0;
3081
3082 debug_rt_mutex_free_waiter(&rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003083 /*
3084 * Fixup the pi_state owner and possibly acquire the lock if we
3085 * haven't already.
3086 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003087 res = fixup_owner(uaddr2, &q, !ret);
Darren Hart52400ba2009-04-03 13:40:49 -07003088 /*
3089 * If fixup_owner() returned an error, proprogate that. If it
Darren Hart56ec1602009-09-21 22:29:59 -07003090 * acquired the lock, clear -ETIMEDOUT or -EINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07003091 */
3092 if (res)
3093 ret = (res < 0) ? res : 0;
3094
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003095 /*
3096 * If fixup_pi_state_owner() faulted and was unable to handle
3097 * the fault, unlock the rt_mutex and return the fault to
3098 * userspace.
3099 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003100 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3101 pi_state = q.pi_state;
3102 get_pi_state(pi_state);
3103 }
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003104
Darren Hart52400ba2009-04-03 13:40:49 -07003105 /* Unqueue and drop the lock. */
3106 unqueue_me_pi(&q);
3107 }
3108
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003109 if (pi_state) {
3110 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3111 put_pi_state(pi_state);
3112 }
3113
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003114 if (ret == -EINTR) {
Darren Hart52400ba2009-04-03 13:40:49 -07003115 /*
Darren Hartcc6db4e2009-07-31 16:20:10 -07003116 * We've already been requeued, but cannot restart by calling
3117 * futex_lock_pi() directly. We could restart this syscall, but
3118 * it would detect that the user space "val" changed and return
3119 * -EWOULDBLOCK. Save the overhead of the restart and return
3120 * -EWOULDBLOCK directly.
Darren Hart52400ba2009-04-03 13:40:49 -07003121 */
Thomas Gleixner20708872009-05-19 23:04:59 +02003122 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07003123 }
3124
3125out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003126 put_futex_key(&q.key);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02003127out_key2:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003128 put_futex_key(&key2);
Darren Hart52400ba2009-04-03 13:40:49 -07003129
3130out:
3131 if (to) {
3132 hrtimer_cancel(&to->timer);
3133 destroy_hrtimer_on_stack(&to->timer);
3134 }
3135 return ret;
3136}
3137
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003138/*
3139 * Support for robust futexes: the kernel cleans up held futexes at
3140 * thread exit time.
3141 *
3142 * Implementation: user-space maintains a per-thread list of locks it
3143 * is holding. Upon do_exit(), the kernel carefully walks this list,
3144 * and marks all locks that are owned by this thread with the
Ingo Molnarc87e2832006-06-27 02:54:58 -07003145 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003146 * always manipulated with the lock held, so the list is private and
3147 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3148 * field, to allow the kernel to clean up if the thread dies after
3149 * acquiring the lock, but just before it could have added itself to
3150 * the list. There can only be one such pending lock.
3151 */
3152
3153/**
Darren Hartd96ee562009-09-21 22:30:22 -07003154 * sys_set_robust_list() - Set the robust-futex list head of a task
3155 * @head: pointer to the list-head
3156 * @len: length of the list-head, as userspace expects
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003157 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003158SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3159 size_t, len)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003160{
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003161 if (!futex_cmpxchg_enabled)
3162 return -ENOSYS;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003163 /*
3164 * The kernel knows only one size for now:
3165 */
3166 if (unlikely(len != sizeof(*head)))
3167 return -EINVAL;
3168
3169 current->robust_list = head;
3170
3171 return 0;
3172}
3173
3174/**
Darren Hartd96ee562009-09-21 22:30:22 -07003175 * sys_get_robust_list() - Get the robust-futex list head of a task
3176 * @pid: pid of the process [zero for current task]
3177 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3178 * @len_ptr: pointer to a length field, the kernel fills in the header size
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003179 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003180SYSCALL_DEFINE3(get_robust_list, int, pid,
3181 struct robust_list_head __user * __user *, head_ptr,
3182 size_t __user *, len_ptr)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003183{
Al Viroba46df92006-10-10 22:46:07 +01003184 struct robust_list_head __user *head;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003185 unsigned long ret;
Kees Cookbdbb7762012-03-19 16:12:53 -07003186 struct task_struct *p;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003187
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003188 if (!futex_cmpxchg_enabled)
3189 return -ENOSYS;
3190
Kees Cookbdbb7762012-03-19 16:12:53 -07003191 rcu_read_lock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003192
Kees Cookbdbb7762012-03-19 16:12:53 -07003193 ret = -ESRCH;
3194 if (!pid)
3195 p = current;
3196 else {
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07003197 p = find_task_by_vpid(pid);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003198 if (!p)
3199 goto err_unlock;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003200 }
3201
Kees Cookbdbb7762012-03-19 16:12:53 -07003202 ret = -EPERM;
Jann Horncaaee622016-01-20 15:00:04 -08003203 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
Kees Cookbdbb7762012-03-19 16:12:53 -07003204 goto err_unlock;
3205
3206 head = p->robust_list;
3207 rcu_read_unlock();
3208
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003209 if (put_user(sizeof(*head), len_ptr))
3210 return -EFAULT;
3211 return put_user(head, head_ptr);
3212
3213err_unlock:
Oleg Nesterovaaa2a972006-09-29 02:00:55 -07003214 rcu_read_unlock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003215
3216 return ret;
3217}
3218
3219/*
3220 * Process a futex-list entry, check whether it's owned by the
3221 * dying task, and do notification if so:
3222 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003223int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003224{
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03003225 u32 uval, uninitialized_var(nval), mval;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003226
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003227retry:
3228 if (get_user(uval, uaddr))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003229 return -1;
3230
Pavel Emelyanovb4888932007-10-18 23:40:14 -07003231 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003232 /*
3233 * Ok, this dying thread is truly holding a futex
3234 * of interest. Set the OWNER_DIED bit atomically
3235 * via cmpxchg, and if the value had FUTEX_WAITERS
3236 * set, wake up a waiter (if any). (We have to do a
3237 * futex_wake() even if OWNER_DIED is already set -
3238 * to handle the rare but possible case of recursive
3239 * thread-death.) The rest of the cleanup is done in
3240 * userspace.
3241 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003242 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
Thomas Gleixner6e0aa9f2011-03-14 10:34:35 +01003243 /*
3244 * We are not holding a lock here, but we want to have
3245 * the pagefault_disable/enable() protection because
3246 * we want to handle the fault gracefully. If the
3247 * access fails we try to fault in the futex with R/W
3248 * verification via get_user_pages. get_user() above
3249 * does not guarantee R/W access. If that fails we
3250 * give up and leave the futex locked.
3251 */
3252 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3253 if (fault_in_user_writeable(uaddr))
3254 return -1;
3255 goto retry;
3256 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07003257 if (nval != uval)
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003258 goto retry;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003259
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003260 /*
3261 * Wake robust non-PI futexes here. The wakeup of
3262 * PI futexes happens in exit_pi_state():
3263 */
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07003264 if (!pi && (uval & FUTEX_WAITERS))
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02003265 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003266 }
3267 return 0;
3268}
3269
3270/*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003271 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3272 */
3273static inline int fetch_robust_entry(struct robust_list __user **entry,
Al Viroba46df92006-10-10 22:46:07 +01003274 struct robust_list __user * __user *head,
Namhyung Kim1dcc41b2010-09-14 21:43:46 +09003275 unsigned int *pi)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003276{
3277 unsigned long uentry;
3278
Al Viroba46df92006-10-10 22:46:07 +01003279 if (get_user(uentry, (unsigned long __user *)head))
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003280 return -EFAULT;
3281
Al Viroba46df92006-10-10 22:46:07 +01003282 *entry = (void __user *)(uentry & ~1UL);
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003283 *pi = uentry & 1;
3284
3285 return 0;
3286}
3287
3288/*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003289 * Walk curr->robust_list (very carefully, it's a userspace list!)
3290 * and mark any locks found there dead, and notify any waiters.
3291 *
3292 * We silently return on any sign of list-walking problem.
3293 */
3294void exit_robust_list(struct task_struct *curr)
3295{
3296 struct robust_list_head __user *head = curr->robust_list;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003297 struct robust_list __user *entry, *next_entry, *pending;
Darren Hart4c115e92010-11-04 15:00:00 -04003298 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3299 unsigned int uninitialized_var(next_pi);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003300 unsigned long futex_offset;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003301 int rc;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003302
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003303 if (!futex_cmpxchg_enabled)
3304 return;
3305
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003306 /*
3307 * Fetch the list head (which was registered earlier, via
3308 * sys_set_robust_list()):
3309 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003310 if (fetch_robust_entry(&entry, &head->list.next, &pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003311 return;
3312 /*
3313 * Fetch the relative futex offset:
3314 */
3315 if (get_user(futex_offset, &head->futex_offset))
3316 return;
3317 /*
3318 * Fetch any possibly pending lock-add first, and handle it
3319 * if it exists:
3320 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003321 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003322 return;
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003323
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003324 next_entry = NULL; /* avoid warning with gcc */
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003325 while (entry != &head->list) {
3326 /*
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003327 * Fetch the next entry in the list before calling
3328 * handle_futex_death:
3329 */
3330 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3331 /*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003332 * A pending lock might already be on the list, so
Ingo Molnarc87e2832006-06-27 02:54:58 -07003333 * don't process it twice:
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003334 */
3335 if (entry != pending)
Al Viroba46df92006-10-10 22:46:07 +01003336 if (handle_futex_death((void __user *)entry + futex_offset,
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003337 curr, pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003338 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003339 if (rc)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003340 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003341 entry = next_entry;
3342 pi = next_pi;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003343 /*
3344 * Avoid excessively long or circular lists:
3345 */
3346 if (!--limit)
3347 break;
3348
3349 cond_resched();
3350 }
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003351
3352 if (pending)
3353 handle_futex_death((void __user *)pending + futex_offset,
3354 curr, pip);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003355}
3356
Pierre Peifferc19384b2007-05-09 02:35:02 -07003357long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
Ingo Molnare2970f22006-06-27 02:54:47 -07003358 u32 __user *uaddr2, u32 val2, u32 val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359{
Thomas Gleixner81b40532012-02-15 12:17:09 +01003360 int cmd = op & FUTEX_CMD_MASK;
Darren Hartb41277d2010-11-08 13:10:09 -08003361 unsigned int flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003363 if (!(op & FUTEX_PRIVATE_FLAG))
Darren Hartb41277d2010-11-08 13:10:09 -08003364 flags |= FLAGS_SHARED;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003365
Darren Hartb41277d2010-11-08 13:10:09 -08003366 if (op & FUTEX_CLOCK_REALTIME) {
3367 flags |= FLAGS_CLOCKRT;
Darren Hart337f1302015-12-18 13:36:37 -08003368 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3369 cmd != FUTEX_WAIT_REQUEUE_PI)
Darren Hartb41277d2010-11-08 13:10:09 -08003370 return -ENOSYS;
3371 }
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003372
3373 switch (cmd) {
Thomas Gleixner59263b52012-02-15 12:08:34 +01003374 case FUTEX_LOCK_PI:
3375 case FUTEX_UNLOCK_PI:
3376 case FUTEX_TRYLOCK_PI:
3377 case FUTEX_WAIT_REQUEUE_PI:
3378 case FUTEX_CMP_REQUEUE_PI:
3379 if (!futex_cmpxchg_enabled)
3380 return -ENOSYS;
3381 }
3382
3383 switch (cmd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 case FUTEX_WAIT:
Thomas Gleixnercd689982008-02-01 17:45:14 +01003385 val3 = FUTEX_BITSET_MATCH_ANY;
3386 case FUTEX_WAIT_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003387 return futex_wait(uaddr, flags, val, timeout, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388 case FUTEX_WAKE:
Thomas Gleixnercd689982008-02-01 17:45:14 +01003389 val3 = FUTEX_BITSET_MATCH_ANY;
3390 case FUTEX_WAKE_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003391 return futex_wake(uaddr, flags, val, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 case FUTEX_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003393 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 case FUTEX_CMP_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003395 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07003396 case FUTEX_WAKE_OP:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003397 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003398 case FUTEX_LOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003399 return futex_lock_pi(uaddr, flags, timeout, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003400 case FUTEX_UNLOCK_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003401 return futex_unlock_pi(uaddr, flags);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003402 case FUTEX_TRYLOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003403 return futex_lock_pi(uaddr, flags, NULL, 1);
Darren Hart52400ba2009-04-03 13:40:49 -07003404 case FUTEX_WAIT_REQUEUE_PI:
3405 val3 = FUTEX_BITSET_MATCH_ANY;
Thomas Gleixner81b40532012-02-15 12:17:09 +01003406 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3407 uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07003408 case FUTEX_CMP_REQUEUE_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003409 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 }
Thomas Gleixner81b40532012-02-15 12:17:09 +01003411 return -ENOSYS;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412}
3413
3414
Heiko Carstens17da2bd2009-01-14 14:14:10 +01003415SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3416 struct timespec __user *, utime, u32 __user *, uaddr2,
3417 u32, val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418{
Pierre Peifferc19384b2007-05-09 02:35:02 -07003419 struct timespec ts;
3420 ktime_t t, *tp = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07003421 u32 val2 = 0;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003422 int cmd = op & FUTEX_CMD_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423
Thomas Gleixnercd689982008-02-01 17:45:14 +01003424 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
Darren Hart52400ba2009-04-03 13:40:49 -07003425 cmd == FUTEX_WAIT_BITSET ||
3426 cmd == FUTEX_WAIT_REQUEUE_PI)) {
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07003427 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3428 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003429 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003431 if (!timespec_valid(&ts))
Thomas Gleixner9741ef962006-03-31 02:31:32 -08003432 return -EINVAL;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003433
3434 t = timespec_to_ktime(ts);
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003435 if (cmd == FUTEX_WAIT)
Thomas Gleixner5a7780e2008-02-13 09:20:43 +01003436 t = ktime_add_safe(ktime_get(), t);
Pierre Peifferc19384b2007-05-09 02:35:02 -07003437 tp = &t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438 }
3439 /*
Darren Hart52400ba2009-04-03 13:40:49 -07003440 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
Andreas Schwabf54f0982007-07-31 00:38:51 -07003441 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442 */
Andreas Schwabf54f0982007-07-31 00:38:51 -07003443 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
Darren Hartba9c22f2009-04-20 22:22:22 -07003444 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
Ingo Molnare2970f22006-06-27 02:54:47 -07003445 val2 = (u32) (unsigned long) utime;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446
Pierre Peifferc19384b2007-05-09 02:35:02 -07003447 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448}
3449
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003450static void __init futex_detect_cmpxchg(void)
3451{
3452#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3453 u32 curval;
3454
3455 /*
3456 * This will fail and we want it. Some arch implementations do
3457 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3458 * functionality. We want to know that before we call in any
3459 * of the complex code paths. Also we want to prevent
3460 * registration of robust lists in that case. NULL is
3461 * guaranteed to fault and we get -EFAULT on functional
3462 * implementation, the non-functional ones will return
3463 * -ENOSYS.
3464 */
3465 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3466 futex_cmpxchg_enabled = 1;
3467#endif
3468}
3469
Benjamin Herrenschmidtf6d107f2008-03-27 14:52:15 +11003470static int __init futex_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471{
Heiko Carstens63b1a812014-01-16 14:54:50 +01003472 unsigned int futex_shift;
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003473 unsigned long i;
3474
3475#if CONFIG_BASE_SMALL
3476 futex_hashsize = 16;
3477#else
3478 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3479#endif
3480
3481 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3482 futex_hashsize, 0,
3483 futex_hashsize < 256 ? HASH_SMALL : 0,
Heiko Carstens63b1a812014-01-16 14:54:50 +01003484 &futex_shift, NULL,
3485 futex_hashsize, futex_hashsize);
3486 futex_hashsize = 1UL << futex_shift;
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003487
3488 futex_detect_cmpxchg();
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003489
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003490 for (i = 0; i < futex_hashsize; i++) {
Linus Torvalds11d46162014-03-20 22:11:17 -07003491 atomic_set(&futex_queues[i].waiters, 0);
Dima Zavin732375c2011-07-07 17:27:59 -07003492 plist_head_init(&futex_queues[i].chain);
Thomas Gleixner3e4ab742008-02-23 15:23:55 -08003493 spin_lock_init(&futex_queues[i].lock);
3494 }
3495
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 return 0;
3497}
Yang Yang25f71d12016-12-30 16:17:55 +08003498core_initcall(futex_init);