blob: 3b6dbeecd91be431eea601ed9d458f00b2d4a68b [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 Zijlstra5293c2e2017-03-22 11:35:51 +0100924 rt_mutex_futex_unlock(&pi_state->pi_mutex);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700925
926 spin_unlock(&hb->lock);
927
Thomas Gleixner1d615482009-11-17 14:54:03 +0100928 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700929 }
Thomas Gleixner1d615482009-11-17 14:54:03 +0100930 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700931}
932
Thomas Gleixner54a21782014-06-03 12:27:08 +0000933/*
934 * We need to check the following states:
935 *
936 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
937 *
938 * [1] NULL | --- | --- | 0 | 0/1 | Valid
939 * [2] NULL | --- | --- | >0 | 0/1 | Valid
940 *
941 * [3] Found | NULL | -- | Any | 0/1 | Invalid
942 *
943 * [4] Found | Found | NULL | 0 | 1 | Valid
944 * [5] Found | Found | NULL | >0 | 1 | Invalid
945 *
946 * [6] Found | Found | task | 0 | 1 | Valid
947 *
948 * [7] Found | Found | NULL | Any | 0 | Invalid
949 *
950 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
951 * [9] Found | Found | task | 0 | 0 | Invalid
952 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
953 *
954 * [1] Indicates that the kernel can acquire the futex atomically. We
955 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
956 *
957 * [2] Valid, if TID does not belong to a kernel thread. If no matching
958 * thread is found then it indicates that the owner TID has died.
959 *
960 * [3] Invalid. The waiter is queued on a non PI futex
961 *
962 * [4] Valid state after exit_robust_list(), which sets the user space
963 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
964 *
965 * [5] The user space value got manipulated between exit_robust_list()
966 * and exit_pi_state_list()
967 *
968 * [6] Valid state after exit_pi_state_list() which sets the new owner in
969 * the pi_state but cannot access the user space value.
970 *
971 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
972 *
973 * [8] Owner and user space value match
974 *
975 * [9] There is no transient state which sets the user space TID to 0
976 * except exit_robust_list(), but this is indicated by the
977 * FUTEX_OWNER_DIED bit. See [4]
978 *
979 * [10] There is no transient state which leaves owner and user space
980 * TID out of sync.
Peter Zijlstra734009e2017-03-22 11:35:52 +0100981 *
982 *
983 * Serialization and lifetime rules:
984 *
985 * hb->lock:
986 *
987 * hb -> futex_q, relation
988 * futex_q -> pi_state, relation
989 *
990 * (cannot be raw because hb can contain arbitrary amount
991 * of futex_q's)
992 *
993 * pi_mutex->wait_lock:
994 *
995 * {uval, pi_state}
996 *
997 * (and pi_mutex 'obviously')
998 *
999 * p->pi_lock:
1000 *
1001 * p->pi_state_list -> pi_state->list, relation
1002 *
1003 * pi_state->refcount:
1004 *
1005 * pi_state lifetime
1006 *
1007 *
1008 * Lock order:
1009 *
1010 * hb->lock
1011 * pi_mutex->wait_lock
1012 * p->pi_lock
1013 *
Thomas Gleixner54a21782014-06-03 12:27:08 +00001014 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001015
1016/*
1017 * Validate that the existing waiter has a pi_state and sanity check
1018 * the pi_state against the user space value. If correct, attach to
1019 * it.
1020 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001021static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1022 struct futex_pi_state *pi_state,
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001023 struct futex_pi_state **ps)
1024{
1025 pid_t pid = uval & FUTEX_TID_MASK;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001026 int ret, uval2;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001027
1028 /*
1029 * Userspace might have messed up non-PI and PI futexes [3]
1030 */
1031 if (unlikely(!pi_state))
1032 return -EINVAL;
1033
Peter Zijlstra734009e2017-03-22 11:35:52 +01001034 /*
1035 * We get here with hb->lock held, and having found a
1036 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1037 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1038 * which in turn means that futex_lock_pi() still has a reference on
1039 * our pi_state.
1040 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001041 WARN_ON(!atomic_read(&pi_state->refcount));
1042
1043 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01001044 * Now that we have a pi_state, we can acquire wait_lock
1045 * and do the state validation.
1046 */
1047 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1048
1049 /*
1050 * Since {uval, pi_state} is serialized by wait_lock, and our current
1051 * uval was read without holding it, it can have changed. Verify it
1052 * still is what we expect it to be, otherwise retry the entire
1053 * operation.
1054 */
1055 if (get_futex_value_locked(&uval2, uaddr))
1056 goto out_efault;
1057
1058 if (uval != uval2)
1059 goto out_eagain;
1060
1061 /*
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001062 * Handle the owner died case:
1063 */
1064 if (uval & FUTEX_OWNER_DIED) {
1065 /*
1066 * exit_pi_state_list sets owner to NULL and wakes the
1067 * topmost waiter. The task which acquires the
1068 * pi_state->rt_mutex will fixup owner.
1069 */
1070 if (!pi_state->owner) {
1071 /*
1072 * No pi state owner, but the user space TID
1073 * is not 0. Inconsistent state. [5]
1074 */
1075 if (pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001076 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001077 /*
1078 * Take a ref on the state and return success. [4]
1079 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001080 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001081 }
1082
1083 /*
1084 * If TID is 0, then either the dying owner has not
1085 * yet executed exit_pi_state_list() or some waiter
1086 * acquired the rtmutex in the pi state, but did not
1087 * yet fixup the TID in user space.
1088 *
1089 * Take a ref on the state and return success. [6]
1090 */
1091 if (!pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001092 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001093 } else {
1094 /*
1095 * If the owner died bit is not set, then the pi_state
1096 * must have an owner. [7]
1097 */
1098 if (!pi_state->owner)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001099 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001100 }
1101
1102 /*
1103 * Bail out if user space manipulated the futex value. If pi
1104 * state exists then the owner TID must be the same as the
1105 * user space TID. [9/10]
1106 */
1107 if (pid != task_pid_vnr(pi_state->owner))
Peter Zijlstra734009e2017-03-22 11:35:52 +01001108 goto out_einval;
1109
1110out_attach:
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01001111 get_pi_state(pi_state);
Peter Zijlstra734009e2017-03-22 11:35:52 +01001112 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001113 *ps = pi_state;
1114 return 0;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001115
1116out_einval:
1117 ret = -EINVAL;
1118 goto out_error;
1119
1120out_eagain:
1121 ret = -EAGAIN;
1122 goto out_error;
1123
1124out_efault:
1125 ret = -EFAULT;
1126 goto out_error;
1127
1128out_error:
1129 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1130 return ret;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001131}
1132
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001133/*
1134 * Lookup the task for the TID provided from user space and attach to
1135 * it after doing proper sanity checks.
1136 */
1137static int attach_to_pi_owner(u32 uval, union futex_key *key,
1138 struct futex_pi_state **ps)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001139{
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001140 pid_t pid = uval & FUTEX_TID_MASK;
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001141 struct futex_pi_state *pi_state;
1142 struct task_struct *p;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001143
1144 /*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001145 * We are the first waiter - try to look up the real owner and attach
Thomas Gleixner54a21782014-06-03 12:27:08 +00001146 * the new pi_state to it, but bail out when TID = 0 [1]
Ingo Molnarc87e2832006-06-27 02:54:58 -07001147 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001148 if (!pid)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001149 return -ESRCH;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001150 p = futex_find_get_task(pid);
Michal Hocko7a0ea092010-06-30 09:51:19 +02001151 if (!p)
1152 return -ESRCH;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001153
Oleg Nesterova2129462015-02-02 15:05:36 +01001154 if (unlikely(p->flags & PF_KTHREAD)) {
Thomas Gleixnerf0d71b32014-05-12 20:45:35 +00001155 put_task_struct(p);
1156 return -EPERM;
1157 }
1158
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001159 /*
1160 * We need to look at the task state flags to figure out,
1161 * whether the task is exiting. To protect against the do_exit
1162 * change of the task flags, we do this protected by
1163 * p->pi_lock:
1164 */
Thomas Gleixner1d615482009-11-17 14:54:03 +01001165 raw_spin_lock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001166 if (unlikely(p->flags & PF_EXITING)) {
1167 /*
1168 * The task is on the way out. When PF_EXITPIDONE is
1169 * set, we know that the task has finished the
1170 * cleanup:
1171 */
1172 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1173
Thomas Gleixner1d615482009-11-17 14:54:03 +01001174 raw_spin_unlock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001175 put_task_struct(p);
1176 return ret;
1177 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001178
Thomas Gleixner54a21782014-06-03 12:27:08 +00001179 /*
1180 * No existing pi state. First waiter. [2]
Peter Zijlstra734009e2017-03-22 11:35:52 +01001181 *
1182 * This creates pi_state, we have hb->lock held, this means nothing can
1183 * observe this state, wait_lock is irrelevant.
Thomas Gleixner54a21782014-06-03 12:27:08 +00001184 */
Ingo Molnarc87e2832006-06-27 02:54:58 -07001185 pi_state = alloc_pi_state();
1186
1187 /*
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001188 * Initialize the pi_mutex in locked state and make @p
Ingo Molnarc87e2832006-06-27 02:54:58 -07001189 * the owner of it:
1190 */
1191 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1192
1193 /* Store the key for possible exit cleanups: */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001194 pi_state->key = *key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001195
Ingo Molnar627371d2006-07-29 05:16:20 +02001196 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001197 list_add(&pi_state->list, &p->pi_state_list);
1198 pi_state->owner = p;
Thomas Gleixner1d615482009-11-17 14:54:03 +01001199 raw_spin_unlock_irq(&p->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001200
1201 put_task_struct(p);
1202
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001203 *ps = pi_state;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001204
1205 return 0;
1206}
1207
Peter Zijlstra734009e2017-03-22 11:35:52 +01001208static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1209 struct futex_hash_bucket *hb,
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001210 union futex_key *key, struct futex_pi_state **ps)
1211{
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001212 struct futex_q *top_waiter = futex_top_waiter(hb, key);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001213
1214 /*
1215 * If there is a waiter on that futex, validate it and
1216 * attach to the pi_state when the validation succeeds.
1217 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001218 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001219 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001220
1221 /*
1222 * We are the first waiter - try to look up the owner based on
1223 * @uval and attach to it.
1224 */
1225 return attach_to_pi_owner(uval, key, ps);
1226}
1227
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001228static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1229{
1230 u32 uninitialized_var(curval);
1231
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001232 if (unlikely(should_fail_futex(true)))
1233 return -EFAULT;
1234
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001235 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1236 return -EFAULT;
1237
Peter Zijlstra734009e2017-03-22 11:35:52 +01001238 /* If user space value changed, let the caller retry */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001239 return curval != uval ? -EAGAIN : 0;
1240}
1241
Darren Hart1a520842009-04-03 13:39:52 -07001242/**
Darren Hartd96ee562009-09-21 22:30:22 -07001243 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
Darren Hartbab5bc92009-04-07 23:23:50 -07001244 * @uaddr: the pi futex user address
1245 * @hb: the pi futex hash bucket
1246 * @key: the futex key associated with uaddr and hb
1247 * @ps: the pi_state pointer where we store the result of the
1248 * lookup
1249 * @task: the task to perform the atomic lock work for. This will
1250 * be "current" except in the case of requeue pi.
1251 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart1a520842009-04-03 13:39:52 -07001252 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001253 * Return:
1254 * 0 - ready to wait;
1255 * 1 - acquired the lock;
Darren Hart1a520842009-04-03 13:39:52 -07001256 * <0 - error
1257 *
1258 * The hb->lock and futex_key refs shall be held by the caller.
1259 */
1260static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1261 union futex_key *key,
1262 struct futex_pi_state **ps,
Darren Hartbab5bc92009-04-07 23:23:50 -07001263 struct task_struct *task, int set_waiters)
Darren Hart1a520842009-04-03 13:39:52 -07001264{
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001265 u32 uval, newval, vpid = task_pid_vnr(task);
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001266 struct futex_q *top_waiter;
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001267 int ret;
Darren Hart1a520842009-04-03 13:39:52 -07001268
1269 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001270 * Read the user space value first so we can validate a few
1271 * things before proceeding further.
Darren Hart1a520842009-04-03 13:39:52 -07001272 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001273 if (get_futex_value_locked(&uval, uaddr))
Darren Hart1a520842009-04-03 13:39:52 -07001274 return -EFAULT;
1275
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001276 if (unlikely(should_fail_futex(true)))
1277 return -EFAULT;
1278
Darren Hart1a520842009-04-03 13:39:52 -07001279 /*
1280 * Detect deadlocks.
1281 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001282 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
Darren Hart1a520842009-04-03 13:39:52 -07001283 return -EDEADLK;
1284
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001285 if ((unlikely(should_fail_futex(true))))
1286 return -EDEADLK;
1287
Darren Hart1a520842009-04-03 13:39:52 -07001288 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001289 * Lookup existing state first. If it exists, try to attach to
1290 * its pi_state.
Darren Hart1a520842009-04-03 13:39:52 -07001291 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001292 top_waiter = futex_top_waiter(hb, key);
1293 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001294 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001295
1296 /*
1297 * No waiter and user TID is 0. We are here because the
1298 * waiters or the owner died bit is set or called from
1299 * requeue_cmp_pi or for whatever reason something took the
1300 * syscall.
1301 */
1302 if (!(uval & FUTEX_TID_MASK)) {
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001303 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001304 * We take over the futex. No other waiters and the user space
1305 * TID is 0. We preserve the owner died bit.
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001306 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001307 newval = uval & FUTEX_OWNER_DIED;
1308 newval |= vpid;
1309
1310 /* The futex requeue_pi code can enforce the waiters bit */
1311 if (set_waiters)
1312 newval |= FUTEX_WAITERS;
1313
1314 ret = lock_pi_update_atomic(uaddr, uval, newval);
1315 /* If the take over worked, return 1 */
1316 return ret < 0 ? ret : 1;
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001317 }
Darren Hart1a520842009-04-03 13:39:52 -07001318
Darren Hart1a520842009-04-03 13:39:52 -07001319 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001320 * First waiter. Set the waiters bit before attaching ourself to
1321 * the owner. If owner tries to unlock, it will be forced into
1322 * the kernel and blocked on hb->lock.
Darren Hart1a520842009-04-03 13:39:52 -07001323 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001324 newval = uval | FUTEX_WAITERS;
1325 ret = lock_pi_update_atomic(uaddr, uval, newval);
1326 if (ret)
1327 return ret;
Darren Hart1a520842009-04-03 13:39:52 -07001328 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001329 * If the update of the user space value succeeded, we try to
1330 * attach to the owner. If that fails, no harm done, we only
1331 * set the FUTEX_WAITERS bit in the user space variable.
Darren Hart1a520842009-04-03 13:39:52 -07001332 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001333 return attach_to_pi_owner(uval, key, ps);
Darren Hart1a520842009-04-03 13:39:52 -07001334}
1335
Lai Jiangshan2e129782010-12-22 14:18:50 +08001336/**
1337 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1338 * @q: The futex_q to unqueue
1339 *
1340 * The q->lock_ptr must not be NULL and must be held by the caller.
1341 */
1342static void __unqueue_futex(struct futex_q *q)
1343{
1344 struct futex_hash_bucket *hb;
1345
Steven Rostedt29096202011-03-17 15:21:07 -04001346 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1347 || WARN_ON(plist_node_empty(&q->list)))
Lai Jiangshan2e129782010-12-22 14:18:50 +08001348 return;
1349
1350 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1351 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001352 hb_waiters_dec(hb);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001353}
1354
Ingo Molnarc87e2832006-06-27 02:54:58 -07001355/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 * The hash bucket lock must be held when this is called.
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001357 * Afterwards, the futex_q must not be accessed. Callers
1358 * must ensure to later call wake_up_q() for the actual
1359 * wakeups to occur.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001361static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362{
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001363 struct task_struct *p = q->task;
1364
Darren Hartaa109902012-11-26 16:29:56 -08001365 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1366 return;
1367
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001368 /*
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001369 * Queue the task for later wakeup for after we've released
1370 * the hb->lock. wake_q_add() grabs reference to p.
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001371 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001372 wake_q_add(wake_q, p);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001373 __unqueue_futex(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 /*
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001375 * The waiting task can free the futex_q as soon as
1376 * q->lock_ptr = NULL is written, without taking any locks. A
1377 * memory barrier is required here to prevent the following
1378 * store to lock_ptr from getting ahead of the plist_del.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 */
Peter Zijlstra1b367ec2017-03-22 11:35:49 +01001380 smp_store_release(&q->lock_ptr, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381}
1382
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001383static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *top_waiter,
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001384 struct futex_hash_bucket *hb)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001385{
1386 struct task_struct *new_owner;
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001387 struct futex_pi_state *pi_state = top_waiter->pi_state;
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03001388 u32 uninitialized_var(curval), newval;
Waiman Long194a6b52016-11-17 11:46:38 -05001389 DEFINE_WAKE_Q(wake_q);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001390 bool deboost;
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001391 int ret = 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001392
1393 if (!pi_state)
1394 return -EINVAL;
1395
Thomas Gleixner51246bf2010-02-02 11:40:27 +01001396 /*
1397 * If current does not own the pi_state then the futex is
1398 * inconsistent and user space fiddled with the futex value.
1399 */
1400 if (pi_state->owner != current)
1401 return -EINVAL;
1402
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001403 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001404 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1405
1406 /*
Steven Rostedtf123c982011-01-06 15:08:29 -05001407 * It is possible that the next waiter (the one that brought
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001408 * top_waiter owner to the kernel) timed out and is no longer
Steven Rostedtf123c982011-01-06 15:08:29 -05001409 * waiting on the lock.
Ingo Molnarc87e2832006-06-27 02:54:58 -07001410 */
1411 if (!new_owner)
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001412 new_owner = top_waiter->task;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001413
1414 /*
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001415 * We pass it to the next owner. The WAITERS bit is always
1416 * kept enabled while there is PI state around. We cleanup the
1417 * owner died bit, because we are the owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07001418 */
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001419 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001420
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001421 if (unlikely(should_fail_futex(true)))
1422 ret = -EFAULT;
1423
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001424 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001425 ret = -EFAULT;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001426
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001427 } else if (curval != uval) {
1428 /*
1429 * If a unconditional UNLOCK_PI operation (user space did not
1430 * try the TID->0 transition) raced with a waiter setting the
1431 * FUTEX_WAITERS flag between get_user() and locking the hash
1432 * bucket lock, retry the operation.
1433 */
1434 if ((FUTEX_TID_MASK & curval) == uval)
1435 ret = -EAGAIN;
1436 else
1437 ret = -EINVAL;
1438 }
Peter Zijlstra734009e2017-03-22 11:35:52 +01001439
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001440 if (ret) {
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001441 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001442 return ret;
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001443 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001444
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001445 raw_spin_lock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001446 WARN_ON(list_empty(&pi_state->list));
1447 list_del_init(&pi_state->list);
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001448 raw_spin_unlock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001449
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001450 raw_spin_lock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001451 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001452 list_add(&pi_state->list, &new_owner->pi_state_list);
1453 pi_state->owner = new_owner;
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001454 raw_spin_unlock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001455
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001456 /*
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001457 * We've updated the uservalue, this unlock cannot fail.
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001458 */
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001459 deboost = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1460
1461 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001462 spin_unlock(&hb->lock);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001463
1464 if (deboost) {
1465 wake_up_q(&wake_q);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02001466 rt_mutex_adjust_prio(current);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001467 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001468
1469 return 0;
1470}
1471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472/*
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001473 * Express the locking dependencies for lockdep:
1474 */
1475static inline void
1476double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1477{
1478 if (hb1 <= hb2) {
1479 spin_lock(&hb1->lock);
1480 if (hb1 < hb2)
1481 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1482 } else { /* hb1 > hb2 */
1483 spin_lock(&hb2->lock);
1484 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1485 }
1486}
1487
Darren Hart5eb3dc62009-03-12 00:55:52 -07001488static inline void
1489double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1490{
Darren Hartf061d352009-03-12 15:11:18 -07001491 spin_unlock(&hb1->lock);
Ingo Molnar88f502f2009-03-13 10:32:07 +01001492 if (hb1 != hb2)
1493 spin_unlock(&hb2->lock);
Darren Hart5eb3dc62009-03-12 00:55:52 -07001494}
1495
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001496/*
Darren Hartb2d09942009-03-12 00:55:37 -07001497 * Wake up waiters matching bitset queued on this futex (uaddr).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 */
Darren Hartb41277d2010-11-08 13:10:09 -08001499static int
1500futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501{
Ingo Molnare2970f22006-06-27 02:54:47 -07001502 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 struct futex_q *this, *next;
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001504 union futex_key key = FUTEX_KEY_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 int ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001506 DEFINE_WAKE_Q(wake_q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507
Thomas Gleixnercd689982008-02-01 17:45:14 +01001508 if (!bitset)
1509 return -EINVAL;
1510
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001511 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 if (unlikely(ret != 0))
1513 goto out;
1514
Ingo Molnare2970f22006-06-27 02:54:47 -07001515 hb = hash_futex(&key);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001516
1517 /* Make sure we really have tasks to wakeup */
1518 if (!hb_waiters_pending(hb))
1519 goto out_put_key;
1520
Ingo Molnare2970f22006-06-27 02:54:47 -07001521 spin_lock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522
Jason Low0d00c7b2014-01-12 15:31:22 -08001523 plist_for_each_entry_safe(this, next, &hb->chain, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 if (match_futex (&this->key, &key)) {
Darren Hart52400ba2009-04-03 13:40:49 -07001525 if (this->pi_state || this->rt_waiter) {
Ingo Molnared6f7b12006-07-01 04:35:46 -07001526 ret = -EINVAL;
1527 break;
1528 }
Thomas Gleixnercd689982008-02-01 17:45:14 +01001529
1530 /* Check if one of the bits is set in both bitsets */
1531 if (!(this->bitset & bitset))
1532 continue;
1533
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001534 mark_wake_futex(&wake_q, this);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535 if (++ret >= nr_wake)
1536 break;
1537 }
1538 }
1539
Ingo Molnare2970f22006-06-27 02:54:47 -07001540 spin_unlock(&hb->lock);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001541 wake_up_q(&wake_q);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001542out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001543 put_futex_key(&key);
Darren Hart42d35d42008-12-29 15:49:53 -08001544out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 return ret;
1546}
1547
1548/*
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001549 * Wake up all waiters hashed on the physical page that is mapped
1550 * to this virtual address:
1551 */
Ingo Molnare2970f22006-06-27 02:54:47 -07001552static int
Darren Hartb41277d2010-11-08 13:10:09 -08001553futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
Ingo Molnare2970f22006-06-27 02:54:47 -07001554 int nr_wake, int nr_wake2, int op)
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001555{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001556 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Ingo Molnare2970f22006-06-27 02:54:47 -07001557 struct futex_hash_bucket *hb1, *hb2;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001558 struct futex_q *this, *next;
Darren Harte4dc5b72009-03-12 00:56:13 -07001559 int ret, op_ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001560 DEFINE_WAKE_Q(wake_q);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001561
Darren Harte4dc5b72009-03-12 00:56:13 -07001562retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001563 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001564 if (unlikely(ret != 0))
1565 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001566 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001567 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001568 goto out_put_key1;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001569
Ingo Molnare2970f22006-06-27 02:54:47 -07001570 hb1 = hash_futex(&key1);
1571 hb2 = hash_futex(&key2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001572
Darren Harte4dc5b72009-03-12 00:56:13 -07001573retry_private:
Thomas Gleixnereaaea802009-10-04 09:34:17 +02001574 double_lock_hb(hb1, hb2);
Ingo Molnare2970f22006-06-27 02:54:47 -07001575 op_ret = futex_atomic_op_inuser(op, uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001576 if (unlikely(op_ret < 0)) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001577
Darren Hart5eb3dc62009-03-12 00:55:52 -07001578 double_unlock_hb(hb1, hb2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001579
David Howells7ee1dd32006-01-06 00:11:44 -08001580#ifndef CONFIG_MMU
Ingo Molnare2970f22006-06-27 02:54:47 -07001581 /*
1582 * we don't get EFAULT from MMU faults if we don't have an MMU,
1583 * but we might get them from range checking
1584 */
David Howells7ee1dd32006-01-06 00:11:44 -08001585 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001586 goto out_put_keys;
David Howells7ee1dd32006-01-06 00:11:44 -08001587#endif
1588
David Gibson796f8d92005-11-07 00:59:33 -08001589 if (unlikely(op_ret != -EFAULT)) {
1590 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001591 goto out_put_keys;
David Gibson796f8d92005-11-07 00:59:33 -08001592 }
1593
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001594 ret = fault_in_user_writeable(uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001595 if (ret)
Darren Hartde87fcc2009-03-12 00:55:46 -07001596 goto out_put_keys;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001597
Darren Hartb41277d2010-11-08 13:10:09 -08001598 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001599 goto retry_private;
1600
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001601 put_futex_key(&key2);
1602 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001603 goto retry;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001604 }
1605
Jason Low0d00c7b2014-01-12 15:31:22 -08001606 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001607 if (match_futex (&this->key, &key1)) {
Darren Hartaa109902012-11-26 16:29:56 -08001608 if (this->pi_state || this->rt_waiter) {
1609 ret = -EINVAL;
1610 goto out_unlock;
1611 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001612 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001613 if (++ret >= nr_wake)
1614 break;
1615 }
1616 }
1617
1618 if (op_ret > 0) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001619 op_ret = 0;
Jason Low0d00c7b2014-01-12 15:31:22 -08001620 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001621 if (match_futex (&this->key, &key2)) {
Darren Hartaa109902012-11-26 16:29:56 -08001622 if (this->pi_state || this->rt_waiter) {
1623 ret = -EINVAL;
1624 goto out_unlock;
1625 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001626 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001627 if (++op_ret >= nr_wake2)
1628 break;
1629 }
1630 }
1631 ret += op_ret;
1632 }
1633
Darren Hartaa109902012-11-26 16:29:56 -08001634out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07001635 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001636 wake_up_q(&wake_q);
Darren Hart42d35d42008-12-29 15:49:53 -08001637out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001638 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08001639out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001640 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08001641out:
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001642 return ret;
1643}
1644
Darren Hart9121e472009-04-03 13:40:31 -07001645/**
1646 * requeue_futex() - Requeue a futex_q from one hb to another
1647 * @q: the futex_q to requeue
1648 * @hb1: the source hash_bucket
1649 * @hb2: the target hash_bucket
1650 * @key2: the new key for the requeued futex_q
1651 */
1652static inline
1653void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1654 struct futex_hash_bucket *hb2, union futex_key *key2)
1655{
1656
1657 /*
1658 * If key1 and key2 hash to the same bucket, no need to
1659 * requeue.
1660 */
1661 if (likely(&hb1->chain != &hb2->chain)) {
1662 plist_del(&q->list, &hb1->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001663 hb_waiters_dec(hb1);
Linus Torvalds11d46162014-03-20 22:11:17 -07001664 hb_waiters_inc(hb2);
Davidlohr Buesofe1bce92016-04-20 20:09:24 -07001665 plist_add(&q->list, &hb2->chain);
Darren Hart9121e472009-04-03 13:40:31 -07001666 q->lock_ptr = &hb2->lock;
Darren Hart9121e472009-04-03 13:40:31 -07001667 }
1668 get_futex_key_refs(key2);
1669 q->key = *key2;
1670}
1671
Darren Hart52400ba2009-04-03 13:40:49 -07001672/**
1673 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
Darren Hartd96ee562009-09-21 22:30:22 -07001674 * @q: the futex_q
1675 * @key: the key of the requeue target futex
1676 * @hb: the hash_bucket of the requeue target futex
Darren Hart52400ba2009-04-03 13:40:49 -07001677 *
1678 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1679 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1680 * to the requeue target futex so the waiter can detect the wakeup on the right
1681 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
Darren Hartbeda2c72009-08-09 15:34:39 -07001682 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1683 * to protect access to the pi_state to fixup the owner later. Must be called
1684 * with both q->lock_ptr and hb->lock held.
Darren Hart52400ba2009-04-03 13:40:49 -07001685 */
1686static inline
Darren Hartbeda2c72009-08-09 15:34:39 -07001687void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1688 struct futex_hash_bucket *hb)
Darren Hart52400ba2009-04-03 13:40:49 -07001689{
Darren Hart52400ba2009-04-03 13:40:49 -07001690 get_futex_key_refs(key);
1691 q->key = *key;
1692
Lai Jiangshan2e129782010-12-22 14:18:50 +08001693 __unqueue_futex(q);
Darren Hart52400ba2009-04-03 13:40:49 -07001694
1695 WARN_ON(!q->rt_waiter);
1696 q->rt_waiter = NULL;
1697
Darren Hartbeda2c72009-08-09 15:34:39 -07001698 q->lock_ptr = &hb->lock;
Darren Hartbeda2c72009-08-09 15:34:39 -07001699
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001700 wake_up_state(q->task, TASK_NORMAL);
Darren Hart52400ba2009-04-03 13:40:49 -07001701}
1702
1703/**
1704 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
Darren Hartbab5bc92009-04-07 23:23:50 -07001705 * @pifutex: the user address of the to futex
1706 * @hb1: the from futex hash bucket, must be locked by the caller
1707 * @hb2: the to futex hash bucket, must be locked by the caller
1708 * @key1: the from futex key
1709 * @key2: the to futex key
1710 * @ps: address to store the pi_state pointer
1711 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart52400ba2009-04-03 13:40:49 -07001712 *
1713 * Try and get the lock on behalf of the top waiter if we can do it atomically.
Darren Hartbab5bc92009-04-07 23:23:50 -07001714 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1715 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1716 * hb1 and hb2 must be held by the caller.
Darren Hart52400ba2009-04-03 13:40:49 -07001717 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001718 * Return:
1719 * 0 - failed to acquire the lock atomically;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001720 * >0 - acquired the lock, return value is vpid of the top_waiter
Darren Hart52400ba2009-04-03 13:40:49 -07001721 * <0 - error
1722 */
1723static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1724 struct futex_hash_bucket *hb1,
1725 struct futex_hash_bucket *hb2,
1726 union futex_key *key1, union futex_key *key2,
Darren Hartbab5bc92009-04-07 23:23:50 -07001727 struct futex_pi_state **ps, int set_waiters)
Darren Hart52400ba2009-04-03 13:40:49 -07001728{
Darren Hartbab5bc92009-04-07 23:23:50 -07001729 struct futex_q *top_waiter = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07001730 u32 curval;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001731 int ret, vpid;
Darren Hart52400ba2009-04-03 13:40:49 -07001732
1733 if (get_futex_value_locked(&curval, pifutex))
1734 return -EFAULT;
1735
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001736 if (unlikely(should_fail_futex(true)))
1737 return -EFAULT;
1738
Darren Hartbab5bc92009-04-07 23:23:50 -07001739 /*
1740 * Find the top_waiter and determine if there are additional waiters.
1741 * If the caller intends to requeue more than 1 waiter to pifutex,
1742 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1743 * as we have means to handle the possible fault. If not, don't set
1744 * the bit unecessarily as it will force the subsequent unlock to enter
1745 * the kernel.
1746 */
Darren Hart52400ba2009-04-03 13:40:49 -07001747 top_waiter = futex_top_waiter(hb1, key1);
1748
1749 /* There are no waiters, nothing for us to do. */
1750 if (!top_waiter)
1751 return 0;
1752
Darren Hart84bc4af2009-08-13 17:36:53 -07001753 /* Ensure we requeue to the expected futex. */
1754 if (!match_futex(top_waiter->requeue_pi_key, key2))
1755 return -EINVAL;
1756
Darren Hart52400ba2009-04-03 13:40:49 -07001757 /*
Darren Hartbab5bc92009-04-07 23:23:50 -07001758 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1759 * the contended case or if set_waiters is 1. The pi_state is returned
1760 * in ps in contended cases.
Darren Hart52400ba2009-04-03 13:40:49 -07001761 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001762 vpid = task_pid_vnr(top_waiter->task);
Darren Hartbab5bc92009-04-07 23:23:50 -07001763 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1764 set_waiters);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001765 if (ret == 1) {
Darren Hartbeda2c72009-08-09 15:34:39 -07001766 requeue_pi_wake_futex(top_waiter, key2, hb2);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001767 return vpid;
1768 }
Darren Hart52400ba2009-04-03 13:40:49 -07001769 return ret;
1770}
1771
1772/**
1773 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
Randy Dunlapfb62db22010-10-13 11:02:34 -07001774 * @uaddr1: source futex user address
Darren Hartb41277d2010-11-08 13:10:09 -08001775 * @flags: futex flags (FLAGS_SHARED, etc.)
Randy Dunlapfb62db22010-10-13 11:02:34 -07001776 * @uaddr2: target futex user address
1777 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1778 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1779 * @cmpval: @uaddr1 expected value (or %NULL)
1780 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
Darren Hartb41277d2010-11-08 13:10:09 -08001781 * pi futex (pi to pi requeue is not supported)
Darren Hart52400ba2009-04-03 13:40:49 -07001782 *
1783 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1784 * uaddr2 atomically on behalf of the top waiter.
1785 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001786 * Return:
1787 * >=0 - on success, the number of tasks requeued or woken;
Darren Hart52400ba2009-04-03 13:40:49 -07001788 * <0 - on error
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 */
Darren Hartb41277d2010-11-08 13:10:09 -08001790static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1791 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1792 u32 *cmpval, int requeue_pi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001794 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Darren Hart52400ba2009-04-03 13:40:49 -07001795 int drop_count = 0, task_count = 0, ret;
1796 struct futex_pi_state *pi_state = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07001797 struct futex_hash_bucket *hb1, *hb2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798 struct futex_q *this, *next;
Waiman Long194a6b52016-11-17 11:46:38 -05001799 DEFINE_WAKE_Q(wake_q);
Darren Hart52400ba2009-04-03 13:40:49 -07001800
1801 if (requeue_pi) {
1802 /*
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001803 * Requeue PI only works on two distinct uaddrs. This
1804 * check is only valid for private futexes. See below.
1805 */
1806 if (uaddr1 == uaddr2)
1807 return -EINVAL;
1808
1809 /*
Darren Hart52400ba2009-04-03 13:40:49 -07001810 * requeue_pi requires a pi_state, try to allocate it now
1811 * without any locks in case it fails.
1812 */
1813 if (refill_pi_state_cache())
1814 return -ENOMEM;
1815 /*
1816 * requeue_pi must wake as many tasks as it can, up to nr_wake
1817 * + nr_requeue, since it acquires the rt_mutex prior to
1818 * returning to userspace, so as to not leave the rt_mutex with
1819 * waiters and no owner. However, second and third wake-ups
1820 * cannot be predicted as they involve race conditions with the
1821 * first wake and a fault while looking up the pi_state. Both
1822 * pthread_cond_signal() and pthread_cond_broadcast() should
1823 * use nr_wake=1.
1824 */
1825 if (nr_wake != 1)
1826 return -EINVAL;
1827 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828
Darren Hart42d35d42008-12-29 15:49:53 -08001829retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001830 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 if (unlikely(ret != 0))
1832 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001833 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1834 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001836 goto out_put_key1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001838 /*
1839 * The check above which compares uaddrs is not sufficient for
1840 * shared futexes. We need to compare the keys:
1841 */
1842 if (requeue_pi && match_futex(&key1, &key2)) {
1843 ret = -EINVAL;
1844 goto out_put_keys;
1845 }
1846
Ingo Molnare2970f22006-06-27 02:54:47 -07001847 hb1 = hash_futex(&key1);
1848 hb2 = hash_futex(&key2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849
Darren Harte4dc5b72009-03-12 00:56:13 -07001850retry_private:
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001851 hb_waiters_inc(hb2);
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001852 double_lock_hb(hb1, hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853
Ingo Molnare2970f22006-06-27 02:54:47 -07001854 if (likely(cmpval != NULL)) {
1855 u32 curval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856
Ingo Molnare2970f22006-06-27 02:54:47 -07001857 ret = get_futex_value_locked(&curval, uaddr1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858
1859 if (unlikely(ret)) {
Darren Hart5eb3dc62009-03-12 00:55:52 -07001860 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001861 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862
Darren Harte4dc5b72009-03-12 00:56:13 -07001863 ret = get_user(curval, uaddr1);
1864 if (ret)
1865 goto out_put_keys;
1866
Darren Hartb41277d2010-11-08 13:10:09 -08001867 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001868 goto retry_private;
1869
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001870 put_futex_key(&key2);
1871 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001872 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873 }
Ingo Molnare2970f22006-06-27 02:54:47 -07001874 if (curval != *cmpval) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 ret = -EAGAIN;
1876 goto out_unlock;
1877 }
1878 }
1879
Darren Hart52400ba2009-04-03 13:40:49 -07001880 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
Darren Hartbab5bc92009-04-07 23:23:50 -07001881 /*
1882 * Attempt to acquire uaddr2 and wake the top waiter. If we
1883 * intend to requeue waiters, force setting the FUTEX_WAITERS
1884 * bit. We force this here where we are able to easily handle
1885 * faults rather in the requeue loop below.
1886 */
Darren Hart52400ba2009-04-03 13:40:49 -07001887 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
Darren Hartbab5bc92009-04-07 23:23:50 -07001888 &key2, &pi_state, nr_requeue);
Darren Hart52400ba2009-04-03 13:40:49 -07001889
1890 /*
1891 * At this point the top_waiter has either taken uaddr2 or is
1892 * waiting on it. If the former, then the pi_state will not
1893 * exist yet, look it up one more time to ensure we have a
Thomas Gleixner866293e2014-05-12 20:45:34 +00001894 * reference to it. If the lock was taken, ret contains the
1895 * vpid of the top waiter task.
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001896 * If the lock was not taken, we have pi_state and an initial
1897 * refcount on it. In case of an error we have nothing.
Darren Hart52400ba2009-04-03 13:40:49 -07001898 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001899 if (ret > 0) {
Darren Hart52400ba2009-04-03 13:40:49 -07001900 WARN_ON(pi_state);
Darren Hart89061d32009-10-15 15:30:48 -07001901 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07001902 task_count++;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001903 /*
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001904 * If we acquired the lock, then the user space value
1905 * of uaddr2 should be vpid. It cannot be changed by
1906 * the top waiter as it is blocked on hb2 lock if it
1907 * tries to do so. If something fiddled with it behind
1908 * our back the pi state lookup might unearth it. So
1909 * we rather use the known value than rereading and
1910 * handing potential crap to lookup_pi_state.
1911 *
1912 * If that call succeeds then we have pi_state and an
1913 * initial refcount on it.
Thomas Gleixner866293e2014-05-12 20:45:34 +00001914 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001915 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07001916 }
1917
1918 switch (ret) {
1919 case 0:
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001920 /* We hold a reference on the pi state. */
Darren Hart52400ba2009-04-03 13:40:49 -07001921 break;
Thomas Gleixner4959f2d2015-12-19 20:07:40 +00001922
1923 /* If the above failed, then pi_state is NULL */
Darren Hart52400ba2009-04-03 13:40:49 -07001924 case -EFAULT:
1925 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001926 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001927 put_futex_key(&key2);
1928 put_futex_key(&key1);
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001929 ret = fault_in_user_writeable(uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07001930 if (!ret)
1931 goto retry;
1932 goto out;
1933 case -EAGAIN:
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001934 /*
1935 * Two reasons for this:
1936 * - Owner is exiting and we just wait for the
1937 * exit to complete.
1938 * - The user space value changed.
1939 */
Darren Hart52400ba2009-04-03 13:40:49 -07001940 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001941 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001942 put_futex_key(&key2);
1943 put_futex_key(&key1);
Darren Hart52400ba2009-04-03 13:40:49 -07001944 cond_resched();
1945 goto retry;
1946 default:
1947 goto out_unlock;
1948 }
1949 }
1950
Jason Low0d00c7b2014-01-12 15:31:22 -08001951 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Darren Hart52400ba2009-04-03 13:40:49 -07001952 if (task_count - nr_wake >= nr_requeue)
1953 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954
Darren Hart52400ba2009-04-03 13:40:49 -07001955 if (!match_futex(&this->key, &key1))
1956 continue;
1957
Darren Hart392741e2009-08-07 15:20:48 -07001958 /*
1959 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1960 * be paired with each other and no other futex ops.
Darren Hartaa109902012-11-26 16:29:56 -08001961 *
1962 * We should never be requeueing a futex_q with a pi_state,
1963 * which is awaiting a futex_unlock_pi().
Darren Hart392741e2009-08-07 15:20:48 -07001964 */
1965 if ((requeue_pi && !this->rt_waiter) ||
Darren Hartaa109902012-11-26 16:29:56 -08001966 (!requeue_pi && this->rt_waiter) ||
1967 this->pi_state) {
Darren Hart392741e2009-08-07 15:20:48 -07001968 ret = -EINVAL;
1969 break;
1970 }
Darren Hart52400ba2009-04-03 13:40:49 -07001971
1972 /*
1973 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1974 * lock, we already woke the top_waiter. If not, it will be
1975 * woken by futex_unlock_pi().
1976 */
1977 if (++task_count <= nr_wake && !requeue_pi) {
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001978 mark_wake_futex(&wake_q, this);
Darren Hart52400ba2009-04-03 13:40:49 -07001979 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 }
Darren Hart52400ba2009-04-03 13:40:49 -07001981
Darren Hart84bc4af2009-08-13 17:36:53 -07001982 /* Ensure we requeue to the expected futex for requeue_pi. */
1983 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1984 ret = -EINVAL;
1985 break;
1986 }
1987
Darren Hart52400ba2009-04-03 13:40:49 -07001988 /*
1989 * Requeue nr_requeue waiters and possibly one more in the case
1990 * of requeue_pi if we couldn't acquire the lock atomically.
1991 */
1992 if (requeue_pi) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001993 /*
1994 * Prepare the waiter to take the rt_mutex. Take a
1995 * refcount on the pi_state and store the pointer in
1996 * the futex_q object of the waiter.
1997 */
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01001998 get_pi_state(pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07001999 this->pi_state = pi_state;
2000 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2001 this->rt_waiter,
Thomas Gleixnerc051b212014-05-22 03:25:50 +00002002 this->task);
Darren Hart52400ba2009-04-03 13:40:49 -07002003 if (ret == 1) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002004 /*
2005 * We got the lock. We do neither drop the
2006 * refcount on pi_state nor clear
2007 * this->pi_state because the waiter needs the
2008 * pi_state for cleaning up the user space
2009 * value. It will drop the refcount after
2010 * doing so.
2011 */
Darren Hartbeda2c72009-08-09 15:34:39 -07002012 requeue_pi_wake_futex(this, &key2, hb2);
Darren Hart89061d32009-10-15 15:30:48 -07002013 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07002014 continue;
2015 } else if (ret) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002016 /*
2017 * rt_mutex_start_proxy_lock() detected a
2018 * potential deadlock when we tried to queue
2019 * that waiter. Drop the pi_state reference
2020 * which we took above and remove the pointer
2021 * to the state from the waiters futex_q
2022 * object.
2023 */
Darren Hart52400ba2009-04-03 13:40:49 -07002024 this->pi_state = NULL;
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002025 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002026 /*
2027 * We stop queueing more waiters and let user
2028 * space deal with the mess.
2029 */
2030 break;
Darren Hart52400ba2009-04-03 13:40:49 -07002031 }
2032 }
2033 requeue_futex(this, hb1, hb2, &key2);
2034 drop_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 }
2036
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002037 /*
2038 * We took an extra initial reference to the pi_state either
2039 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2040 * need to drop it here again.
2041 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002042 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002043
2044out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07002045 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07002046 wake_up_q(&wake_q);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07002047 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048
Darren Hartcd84a422009-04-02 14:19:38 -07002049 /*
2050 * drop_futex_key_refs() must be called outside the spinlocks. During
2051 * the requeue we moved futex_q's from the hash bucket at key1 to the
2052 * one at key2 and updated their key pointer. We no longer need to
2053 * hold the references to key1.
2054 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 while (--drop_count >= 0)
Rusty Russell9adef582007-05-08 00:26:42 -07002056 drop_futex_key_refs(&key1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057
Darren Hart42d35d42008-12-29 15:49:53 -08002058out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002059 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08002060out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002061 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08002062out:
Darren Hart52400ba2009-04-03 13:40:49 -07002063 return ret ? ret : task_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064}
2065
2066/* The key must be already stored in q->key. */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002067static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002068 __acquires(&hb->lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069{
Ingo Molnare2970f22006-06-27 02:54:47 -07002070 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071
Ingo Molnare2970f22006-06-27 02:54:47 -07002072 hb = hash_futex(&q->key);
Linus Torvalds11d46162014-03-20 22:11:17 -07002073
2074 /*
2075 * Increment the counter before taking the lock so that
2076 * a potential waker won't miss a to-be-slept task that is
2077 * waiting for the spinlock. This is safe as all queue_lock()
2078 * users end up calling queue_me(). Similarly, for housekeeping,
2079 * decrement the counter at queue_unlock() when some error has
2080 * occurred and we don't end up adding the task to the list.
2081 */
2082 hb_waiters_inc(hb);
2083
Ingo Molnare2970f22006-06-27 02:54:47 -07002084 q->lock_ptr = &hb->lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -08002086 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
Ingo Molnare2970f22006-06-27 02:54:47 -07002087 return hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088}
2089
Darren Hartd40d65c2009-09-21 22:30:15 -07002090static inline void
Jason Low0d00c7b2014-01-12 15:31:22 -08002091queue_unlock(struct futex_hash_bucket *hb)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002092 __releases(&hb->lock)
Darren Hartd40d65c2009-09-21 22:30:15 -07002093{
2094 spin_unlock(&hb->lock);
Linus Torvalds11d46162014-03-20 22:11:17 -07002095 hb_waiters_dec(hb);
Darren Hartd40d65c2009-09-21 22:30:15 -07002096}
2097
2098/**
2099 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2100 * @q: The futex_q to enqueue
2101 * @hb: The destination hash bucket
2102 *
2103 * The hb->lock must be held by the caller, and is released here. A call to
2104 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2105 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2106 * or nothing if the unqueue is done as part of the wake process and the unqueue
2107 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2108 * an example).
2109 */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002110static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002111 __releases(&hb->lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112{
Pierre Peifferec92d082007-05-09 02:35:00 -07002113 int prio;
2114
2115 /*
2116 * The priority used to register this element is
2117 * - either the real thread-priority for the real-time threads
2118 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2119 * - or MAX_RT_PRIO for non-RT threads.
2120 * Thus, all RT-threads are woken first in priority order, and
2121 * the others are woken last, in FIFO order.
2122 */
2123 prio = min(current->normal_prio, MAX_RT_PRIO);
2124
2125 plist_node_init(&q->list, prio);
Pierre Peifferec92d082007-05-09 02:35:00 -07002126 plist_add(&q->list, &hb->chain);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002127 q->task = current;
Ingo Molnare2970f22006-06-27 02:54:47 -07002128 spin_unlock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129}
2130
Darren Hartd40d65c2009-09-21 22:30:15 -07002131/**
2132 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2133 * @q: The futex_q to unqueue
2134 *
2135 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2136 * be paired with exactly one earlier call to queue_me().
2137 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002138 * Return:
2139 * 1 - if the futex_q was still queued (and we removed unqueued it);
Darren Hartd40d65c2009-09-21 22:30:15 -07002140 * 0 - if the futex_q was already removed by the waking thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142static int unqueue_me(struct futex_q *q)
2143{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 spinlock_t *lock_ptr;
Ingo Molnare2970f22006-06-27 02:54:47 -07002145 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146
2147 /* In the common case we don't take the spinlock, which is nice. */
Darren Hart42d35d42008-12-29 15:49:53 -08002148retry:
Jianyu Zhan29b75eb2016-03-07 09:32:24 +08002149 /*
2150 * q->lock_ptr can change between this read and the following spin_lock.
2151 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2152 * optimizing lock_ptr out of the logic below.
2153 */
2154 lock_ptr = READ_ONCE(q->lock_ptr);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07002155 if (lock_ptr != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 spin_lock(lock_ptr);
2157 /*
2158 * q->lock_ptr can change between reading it and
2159 * spin_lock(), causing us to take the wrong lock. This
2160 * corrects the race condition.
2161 *
2162 * Reasoning goes like this: if we have the wrong lock,
2163 * q->lock_ptr must have changed (maybe several times)
2164 * between reading it and the spin_lock(). It can
2165 * change again after the spin_lock() but only if it was
2166 * already changed before the spin_lock(). It cannot,
2167 * however, change back to the original value. Therefore
2168 * we can detect whether we acquired the correct lock.
2169 */
2170 if (unlikely(lock_ptr != q->lock_ptr)) {
2171 spin_unlock(lock_ptr);
2172 goto retry;
2173 }
Lai Jiangshan2e129782010-12-22 14:18:50 +08002174 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002175
2176 BUG_ON(q->pi_state);
2177
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 spin_unlock(lock_ptr);
2179 ret = 1;
2180 }
2181
Rusty Russell9adef582007-05-08 00:26:42 -07002182 drop_futex_key_refs(&q->key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 return ret;
2184}
2185
Ingo Molnarc87e2832006-06-27 02:54:58 -07002186/*
2187 * PI futexes can not be requeued and must remove themself from the
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002188 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2189 * and dropped here.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002190 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002191static void unqueue_me_pi(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002192 __releases(q->lock_ptr)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002193{
Lai Jiangshan2e129782010-12-22 14:18:50 +08002194 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002195
2196 BUG_ON(!q->pi_state);
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002197 put_pi_state(q->pi_state);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002198 q->pi_state = NULL;
2199
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002200 spin_unlock(q->lock_ptr);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002201}
2202
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002203/*
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002204 * Fixup the pi_state owner with the new owner.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002205 *
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002206 * Must be called with hash bucket lock held and mm->sem held for non
2207 * private futexes.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002208 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002209static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002210 struct task_struct *newowner)
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002211{
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002212 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002213 struct futex_pi_state *pi_state = q->pi_state;
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03002214 u32 uval, uninitialized_var(curval), newval;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002215 struct task_struct *oldowner;
Darren Harte4dc5b72009-03-12 00:56:13 -07002216 int ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002217
Peter Zijlstra734009e2017-03-22 11:35:52 +01002218 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2219
2220 oldowner = pi_state->owner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002221 /* Owner died? */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002222 if (!pi_state->owner)
2223 newtid |= FUTEX_OWNER_DIED;
2224
2225 /*
2226 * We are here either because we stole the rtmutex from the
Lai Jiangshan81612392011-01-14 17:09:41 +08002227 * previous highest priority waiter or we are the highest priority
2228 * waiter but failed to get the rtmutex the first time.
2229 * We have to replace the newowner TID in the user space variable.
2230 * This must be atomic as we have to preserve the owner died bit here.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002231 *
Darren Hartb2d09942009-03-12 00:55:37 -07002232 * Note: We write the user space value _before_ changing the pi_state
2233 * because we can fault here. Imagine swapped out pages or a fork
2234 * that marked all the anonymous memory readonly for cow.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002235 *
Peter Zijlstra734009e2017-03-22 11:35:52 +01002236 * Modifying pi_state _before_ the user space value would leave the
2237 * pi_state in an inconsistent state when we fault here, because we
2238 * need to drop the locks to handle the fault. This might be observed
2239 * in the PID check in lookup_pi_state.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002240 */
2241retry:
2242 if (get_futex_value_locked(&uval, uaddr))
2243 goto handle_fault;
2244
2245 while (1) {
2246 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2247
Michel Lespinasse37a9d912011-03-10 18:48:51 -08002248 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002249 goto handle_fault;
2250 if (curval == uval)
2251 break;
2252 uval = curval;
2253 }
2254
2255 /*
2256 * We fixed up user space. Now we need to fix the pi_state
2257 * itself.
2258 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002259 if (pi_state->owner != NULL) {
Peter Zijlstra734009e2017-03-22 11:35:52 +01002260 raw_spin_lock(&pi_state->owner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002261 WARN_ON(list_empty(&pi_state->list));
2262 list_del_init(&pi_state->list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002263 raw_spin_unlock(&pi_state->owner->pi_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002264 }
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002265
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002266 pi_state->owner = newowner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002267
Peter Zijlstra734009e2017-03-22 11:35:52 +01002268 raw_spin_lock(&newowner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002269 WARN_ON(!list_empty(&pi_state->list));
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002270 list_add(&pi_state->list, &newowner->pi_state_list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002271 raw_spin_unlock(&newowner->pi_lock);
2272 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2273
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002274 return 0;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002275
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002276 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01002277 * To handle the page fault we need to drop the locks here. That gives
2278 * the other task (either the highest priority waiter itself or the
2279 * task which stole the rtmutex) the chance to try the fixup of the
2280 * pi_state. So once we are back from handling the fault we need to
2281 * check the pi_state after reacquiring the locks and before trying to
2282 * do another fixup. When the fixup has been done already we simply
2283 * return.
2284 *
2285 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2286 * drop hb->lock since the caller owns the hb -> futex_q relation.
2287 * Dropping the pi_mutex->wait_lock requires the state revalidate.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002288 */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002289handle_fault:
Peter Zijlstra734009e2017-03-22 11:35:52 +01002290 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002291 spin_unlock(q->lock_ptr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002292
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002293 ret = fault_in_user_writeable(uaddr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002294
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002295 spin_lock(q->lock_ptr);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002296 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002297
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002298 /*
2299 * Check if someone else fixed it for us:
2300 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01002301 if (pi_state->owner != oldowner) {
2302 ret = 0;
2303 goto out_unlock;
2304 }
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002305
2306 if (ret)
Peter Zijlstra734009e2017-03-22 11:35:52 +01002307 goto out_unlock;
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002308
2309 goto retry;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002310
2311out_unlock:
2312 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2313 return ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002314}
2315
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002316static long futex_wait_restart(struct restart_block *restart);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07002317
Darren Hartca5f9522009-04-03 13:39:33 -07002318/**
Darren Hartdd973992009-04-03 13:40:02 -07002319 * fixup_owner() - Post lock pi_state and corner case management
2320 * @uaddr: user address of the futex
Darren Hartdd973992009-04-03 13:40:02 -07002321 * @q: futex_q (contains pi_state and access to the rt_mutex)
2322 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2323 *
2324 * After attempting to lock an rt_mutex, this function is called to cleanup
2325 * the pi_state owner as well as handle race conditions that may allow us to
2326 * acquire the lock. Must be called with the hb lock held.
2327 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002328 * Return:
2329 * 1 - success, lock taken;
2330 * 0 - success, lock not taken;
Darren Hartdd973992009-04-03 13:40:02 -07002331 * <0 - on error (-EFAULT)
2332 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002333static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
Darren Hartdd973992009-04-03 13:40:02 -07002334{
2335 struct task_struct *owner;
2336 int ret = 0;
2337
2338 if (locked) {
2339 /*
2340 * Got the lock. We might not be the anticipated owner if we
2341 * did a lock-steal - fix up the PI-state in that case:
2342 */
2343 if (q->pi_state->owner != current)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002344 ret = fixup_pi_state_owner(uaddr, q, current);
Darren Hartdd973992009-04-03 13:40:02 -07002345 goto out;
2346 }
2347
2348 /*
2349 * Catch the rare case, where the lock was released when we were on the
2350 * way back before we locked the hash bucket.
2351 */
2352 if (q->pi_state->owner == current) {
2353 /*
2354 * Try to get the rt_mutex now. This might fail as some other
2355 * task acquired the rt_mutex after we removed ourself from the
2356 * rt_mutex waiters list.
2357 */
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01002358 if (rt_mutex_futex_trylock(&q->pi_state->pi_mutex)) {
Darren Hartdd973992009-04-03 13:40:02 -07002359 locked = 1;
2360 goto out;
2361 }
2362
2363 /*
2364 * pi_state is incorrect, some other task did a lock steal and
2365 * we returned due to timeout or signal without taking the
Lai Jiangshan81612392011-01-14 17:09:41 +08002366 * rt_mutex. Too late.
Darren Hartdd973992009-04-03 13:40:02 -07002367 */
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01002368 raw_spin_lock_irq(&q->pi_state->pi_mutex.wait_lock);
Darren Hartdd973992009-04-03 13:40:02 -07002369 owner = rt_mutex_owner(&q->pi_state->pi_mutex);
Lai Jiangshan81612392011-01-14 17:09:41 +08002370 if (!owner)
2371 owner = rt_mutex_next_owner(&q->pi_state->pi_mutex);
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01002372 raw_spin_unlock_irq(&q->pi_state->pi_mutex.wait_lock);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002373 ret = fixup_pi_state_owner(uaddr, q, owner);
Darren Hartdd973992009-04-03 13:40:02 -07002374 goto out;
2375 }
2376
2377 /*
2378 * Paranoia check. If we did not take the lock, then we should not be
Lai Jiangshan81612392011-01-14 17:09:41 +08002379 * the owner of the rt_mutex.
Darren Hartdd973992009-04-03 13:40:02 -07002380 */
2381 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current)
2382 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2383 "pi-state %p\n", ret,
2384 q->pi_state->pi_mutex.owner,
2385 q->pi_state->owner);
2386
2387out:
2388 return ret ? ret : locked;
2389}
2390
2391/**
Darren Hartca5f9522009-04-03 13:39:33 -07002392 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2393 * @hb: the futex hash bucket, must be locked by the caller
2394 * @q: the futex_q to queue up on
2395 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
Darren Hartca5f9522009-04-03 13:39:33 -07002396 */
2397static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002398 struct hrtimer_sleeper *timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002399{
Darren Hart9beba3c2009-09-24 11:54:47 -07002400 /*
2401 * The task state is guaranteed to be set before another task can
Peter Zijlstrab92b8b32015-05-12 10:51:55 +02002402 * wake it. set_current_state() is implemented using smp_store_mb() and
Darren Hart9beba3c2009-09-24 11:54:47 -07002403 * queue_me() calls spin_unlock() upon completion, both serializing
2404 * access to the hash list and forcing another memory barrier.
2405 */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002406 set_current_state(TASK_INTERRUPTIBLE);
Darren Hart0729e192009-09-21 22:30:38 -07002407 queue_me(q, hb);
Darren Hartca5f9522009-04-03 13:39:33 -07002408
2409 /* Arm the timer */
Thomas Gleixner2e4b0d32015-04-14 21:09:13 +00002410 if (timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002411 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002412
2413 /*
Darren Hart0729e192009-09-21 22:30:38 -07002414 * If we have been removed from the hash list, then another task
2415 * has tried to wake us, and we can skip the call to schedule().
Darren Hartca5f9522009-04-03 13:39:33 -07002416 */
2417 if (likely(!plist_node_empty(&q->list))) {
2418 /*
2419 * If the timer has already expired, current will already be
2420 * flagged for rescheduling. Only call schedule if there
2421 * is no timeout, or if it has yet to expire.
2422 */
2423 if (!timeout || timeout->task)
Colin Cross88c80042013-05-01 18:35:05 -07002424 freezable_schedule();
Darren Hartca5f9522009-04-03 13:39:33 -07002425 }
2426 __set_current_state(TASK_RUNNING);
2427}
2428
Darren Hartf8010732009-04-03 13:40:40 -07002429/**
2430 * futex_wait_setup() - Prepare to wait on a futex
2431 * @uaddr: the futex userspace address
2432 * @val: the expected value
Darren Hartb41277d2010-11-08 13:10:09 -08002433 * @flags: futex flags (FLAGS_SHARED, etc.)
Darren Hartf8010732009-04-03 13:40:40 -07002434 * @q: the associated futex_q
2435 * @hb: storage for hash_bucket pointer to be returned to caller
2436 *
2437 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2438 * compare it with the expected value. Handle atomic faults internally.
2439 * Return with the hb lock held and a q.key reference on success, and unlocked
2440 * with no q.key reference on failure.
2441 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002442 * Return:
2443 * 0 - uaddr contains val and hb has been locked;
Bart Van Asscheca4a04c2011-07-17 09:01:00 +02002444 * <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
Darren Hartf8010732009-04-03 13:40:40 -07002445 */
Darren Hartb41277d2010-11-08 13:10:09 -08002446static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
Darren Hartf8010732009-04-03 13:40:40 -07002447 struct futex_q *q, struct futex_hash_bucket **hb)
2448{
2449 u32 uval;
2450 int ret;
2451
2452 /*
2453 * Access the page AFTER the hash-bucket is locked.
2454 * Order is important:
2455 *
2456 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2457 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2458 *
2459 * The basic logical guarantee of a futex is that it blocks ONLY
2460 * if cond(var) is known to be true at the time of blocking, for
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002461 * any cond. If we locked the hash-bucket after testing *uaddr, that
2462 * would open a race condition where we could block indefinitely with
Darren Hartf8010732009-04-03 13:40:40 -07002463 * cond(var) false, which would violate the guarantee.
2464 *
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002465 * On the other hand, we insert q and release the hash-bucket only
2466 * after testing *uaddr. This guarantees that futex_wait() will NOT
2467 * absorb a wakeup if *uaddr does not match the desired values
2468 * while the syscall executes.
Darren Hartf8010732009-04-03 13:40:40 -07002469 */
2470retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002471 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
Darren Hartf8010732009-04-03 13:40:40 -07002472 if (unlikely(ret != 0))
Darren Harta5a2a0c2009-04-10 09:50:05 -07002473 return ret;
Darren Hartf8010732009-04-03 13:40:40 -07002474
2475retry_private:
2476 *hb = queue_lock(q);
2477
2478 ret = get_futex_value_locked(&uval, uaddr);
2479
2480 if (ret) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002481 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002482
2483 ret = get_user(uval, uaddr);
2484 if (ret)
2485 goto out;
2486
Darren Hartb41277d2010-11-08 13:10:09 -08002487 if (!(flags & FLAGS_SHARED))
Darren Hartf8010732009-04-03 13:40:40 -07002488 goto retry_private;
2489
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002490 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002491 goto retry;
2492 }
2493
2494 if (uval != val) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002495 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002496 ret = -EWOULDBLOCK;
2497 }
2498
2499out:
2500 if (ret)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002501 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002502 return ret;
2503}
2504
Darren Hartb41277d2010-11-08 13:10:09 -08002505static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2506 ktime_t *abs_time, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507{
Darren Hartca5f9522009-04-03 13:39:33 -07002508 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002509 struct restart_block *restart;
Ingo Molnare2970f22006-06-27 02:54:47 -07002510 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002511 struct futex_q q = futex_q_init;
Ingo Molnare2970f22006-06-27 02:54:47 -07002512 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513
Thomas Gleixnercd689982008-02-01 17:45:14 +01002514 if (!bitset)
2515 return -EINVAL;
Thomas Gleixnercd689982008-02-01 17:45:14 +01002516 q.bitset = bitset;
Darren Hartca5f9522009-04-03 13:39:33 -07002517
2518 if (abs_time) {
2519 to = &timeout;
2520
Darren Hartb41277d2010-11-08 13:10:09 -08002521 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2522 CLOCK_REALTIME : CLOCK_MONOTONIC,
2523 HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002524 hrtimer_init_sleeper(to, current);
2525 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2526 current->timer_slack_ns);
2527 }
2528
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002529retry:
Darren Hart7ada8762010-10-17 08:35:04 -07002530 /*
2531 * Prepare to wait on uaddr. On success, holds hb lock and increments
2532 * q.key refs.
2533 */
Darren Hartb41277d2010-11-08 13:10:09 -08002534 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Darren Hartf8010732009-04-03 13:40:40 -07002535 if (ret)
Darren Hart42d35d42008-12-29 15:49:53 -08002536 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537
Darren Hartca5f9522009-04-03 13:39:33 -07002538 /* queue_me and wait for wakeup, timeout, or a signal. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002539 futex_wait_queue_me(hb, &q, to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540
2541 /* If we were woken (and unqueued), we succeeded, whatever. */
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002542 ret = 0;
Darren Hart7ada8762010-10-17 08:35:04 -07002543 /* unqueue_me() drops q.key ref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 if (!unqueue_me(&q))
Darren Hart7ada8762010-10-17 08:35:04 -07002545 goto out;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002546 ret = -ETIMEDOUT;
Darren Hartca5f9522009-04-03 13:39:33 -07002547 if (to && !to->task)
Darren Hart7ada8762010-10-17 08:35:04 -07002548 goto out;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002549
Ingo Molnare2970f22006-06-27 02:54:47 -07002550 /*
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002551 * We expect signal_pending(current), but we might be the
2552 * victim of a spurious wakeup as well.
Ingo Molnare2970f22006-06-27 02:54:47 -07002553 */
Darren Hart7ada8762010-10-17 08:35:04 -07002554 if (!signal_pending(current))
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002555 goto retry;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002556
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002557 ret = -ERESTARTSYS;
Pierre Peifferc19384b2007-05-09 02:35:02 -07002558 if (!abs_time)
Darren Hart7ada8762010-10-17 08:35:04 -07002559 goto out;
Steven Rostedtce6bd422007-12-05 15:46:09 +01002560
Andy Lutomirskif56141e2015-02-12 15:01:14 -08002561 restart = &current->restart_block;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002562 restart->fn = futex_wait_restart;
Namhyung Kima3c74c52010-09-14 21:43:47 +09002563 restart->futex.uaddr = uaddr;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002564 restart->futex.val = val;
Thomas Gleixner2456e852016-12-25 11:38:40 +01002565 restart->futex.time = *abs_time;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002566 restart->futex.bitset = bitset;
Darren Hart0cd9c642011-04-14 15:41:57 -07002567 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002568
2569 ret = -ERESTART_RESTARTBLOCK;
2570
Darren Hart42d35d42008-12-29 15:49:53 -08002571out:
Darren Hartca5f9522009-04-03 13:39:33 -07002572 if (to) {
2573 hrtimer_cancel(&to->timer);
2574 destroy_hrtimer_on_stack(&to->timer);
2575 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002576 return ret;
2577}
2578
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002579
2580static long futex_wait_restart(struct restart_block *restart)
2581{
Namhyung Kima3c74c52010-09-14 21:43:47 +09002582 u32 __user *uaddr = restart->futex.uaddr;
Darren Harta72188d2009-04-03 13:40:22 -07002583 ktime_t t, *tp = NULL;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002584
Darren Harta72188d2009-04-03 13:40:22 -07002585 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
Thomas Gleixner2456e852016-12-25 11:38:40 +01002586 t = restart->futex.time;
Darren Harta72188d2009-04-03 13:40:22 -07002587 tp = &t;
2588 }
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002589 restart->fn = do_no_restart_syscall;
Darren Hartb41277d2010-11-08 13:10:09 -08002590
2591 return (long)futex_wait(uaddr, restart->futex.flags,
2592 restart->futex.val, tp, restart->futex.bitset);
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002593}
2594
2595
Ingo Molnarc87e2832006-06-27 02:54:58 -07002596/*
2597 * Userspace tried a 0 -> TID atomic transition of the futex value
2598 * and failed. The kernel side here does the whole locking operation:
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002599 * if there are waiters then it will block as a consequence of relying
2600 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2601 * a 0 value of the futex too.).
2602 *
2603 * Also serves as futex trylock_pi()'ing, and due semantics.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002604 */
Michael Kerrisk996636d2015-01-16 20:28:06 +01002605static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
Darren Hartb41277d2010-11-08 13:10:09 -08002606 ktime_t *time, int trylock)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002607{
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002608 struct hrtimer_sleeper timeout, *to = NULL;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002609 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002610 struct futex_q q = futex_q_init;
Darren Hartdd973992009-04-03 13:40:02 -07002611 int res, ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002612
2613 if (refill_pi_state_cache())
2614 return -ENOMEM;
2615
Pierre Peifferc19384b2007-05-09 02:35:02 -07002616 if (time) {
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002617 to = &timeout;
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002618 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2619 HRTIMER_MODE_ABS);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002620 hrtimer_init_sleeper(to, current);
Arjan van de Vencc584b22008-09-01 15:02:30 -07002621 hrtimer_set_expires(&to->timer, *time);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002622 }
2623
Darren Hart42d35d42008-12-29 15:49:53 -08002624retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002625 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002626 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08002627 goto out;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002628
Darren Harte4dc5b72009-03-12 00:56:13 -07002629retry_private:
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002630 hb = queue_lock(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002631
Darren Hartbab5bc92009-04-07 23:23:50 -07002632 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002633 if (unlikely(ret)) {
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002634 /*
2635 * Atomic work succeeded and we got the lock,
2636 * or failed. Either way, we do _not_ block.
2637 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002638 switch (ret) {
Darren Hart1a520842009-04-03 13:39:52 -07002639 case 1:
2640 /* We got the lock. */
2641 ret = 0;
2642 goto out_unlock_put_key;
2643 case -EFAULT:
2644 goto uaddr_faulted;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002645 case -EAGAIN:
2646 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00002647 * Two reasons for this:
2648 * - Task is exiting and we just wait for the
2649 * exit to complete.
2650 * - The user space value changed.
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002651 */
Jason Low0d00c7b2014-01-12 15:31:22 -08002652 queue_unlock(hb);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002653 put_futex_key(&q.key);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002654 cond_resched();
2655 goto retry;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002656 default:
Darren Hart42d35d42008-12-29 15:49:53 -08002657 goto out_unlock_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002658 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002659 }
2660
2661 /*
2662 * Only actually queue now that the atomic ops are done:
2663 */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002664 queue_me(&q, hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002665
Ingo Molnarc87e2832006-06-27 02:54:58 -07002666 WARN_ON(!q.pi_state);
2667 /*
2668 * Block on the PI mutex:
2669 */
Thomas Gleixnerc051b212014-05-22 03:25:50 +00002670 if (!trylock) {
2671 ret = rt_mutex_timed_futex_lock(&q.pi_state->pi_mutex, to);
2672 } else {
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01002673 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002674 /* Fixup the trylock return value: */
2675 ret = ret ? 0 : -EWOULDBLOCK;
2676 }
2677
Vernon Mauerya99e4e42006-07-01 04:35:42 -07002678 spin_lock(q.lock_ptr);
Darren Hartdd973992009-04-03 13:40:02 -07002679 /*
2680 * Fixup the pi_state owner and possibly acquire the lock if we
2681 * haven't already.
2682 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002683 res = fixup_owner(uaddr, &q, !ret);
Darren Hartdd973992009-04-03 13:40:02 -07002684 /*
2685 * If fixup_owner() returned an error, proprogate that. If it acquired
2686 * the lock, clear our -ETIMEDOUT or -EINTR.
2687 */
2688 if (res)
2689 ret = (res < 0) ? res : 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002690
Darren Harte8f63862009-03-12 00:56:06 -07002691 /*
Darren Hartdd973992009-04-03 13:40:02 -07002692 * If fixup_owner() faulted and was unable to handle the fault, unlock
2693 * it and return the fault to userspace.
Darren Harte8f63862009-03-12 00:56:06 -07002694 */
2695 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current))
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01002696 rt_mutex_futex_unlock(&q.pi_state->pi_mutex);
Darren Harte8f63862009-03-12 00:56:06 -07002697
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002698 /* Unqueue and drop the lock */
2699 unqueue_me_pi(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002700
Mikael Pettersson5ecb01c2010-01-23 22:36:29 +01002701 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002702
Darren Hart42d35d42008-12-29 15:49:53 -08002703out_unlock_put_key:
Jason Low0d00c7b2014-01-12 15:31:22 -08002704 queue_unlock(hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002705
Darren Hart42d35d42008-12-29 15:49:53 -08002706out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002707 put_futex_key(&q.key);
Darren Hart42d35d42008-12-29 15:49:53 -08002708out:
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002709 if (to)
2710 destroy_hrtimer_on_stack(&to->timer);
Darren Hartdd973992009-04-03 13:40:02 -07002711 return ret != -EINTR ? ret : -ERESTARTNOINTR;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002712
Darren Hart42d35d42008-12-29 15:49:53 -08002713uaddr_faulted:
Jason Low0d00c7b2014-01-12 15:31:22 -08002714 queue_unlock(hb);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002715
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002716 ret = fault_in_user_writeable(uaddr);
Darren Harte4dc5b72009-03-12 00:56:13 -07002717 if (ret)
2718 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002719
Darren Hartb41277d2010-11-08 13:10:09 -08002720 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07002721 goto retry_private;
2722
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002723 put_futex_key(&q.key);
Darren Harte4dc5b72009-03-12 00:56:13 -07002724 goto retry;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002725}
2726
2727/*
Ingo Molnarc87e2832006-06-27 02:54:58 -07002728 * Userspace attempted a TID -> 0 atomic transition, and failed.
2729 * This is the in-kernel slowpath: we look up the PI state (if any),
2730 * and do the rt-mutex unlock.
2731 */
Darren Hartb41277d2010-11-08 13:10:09 -08002732static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002733{
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002734 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
Peter Zijlstra38d47c12008-09-26 19:32:20 +02002735 union futex_key key = FUTEX_KEY_INIT;
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002736 struct futex_hash_bucket *hb;
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002737 struct futex_q *top_waiter;
Darren Harte4dc5b72009-03-12 00:56:13 -07002738 int ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002739
2740retry:
2741 if (get_user(uval, uaddr))
2742 return -EFAULT;
2743 /*
2744 * We release only a lock we actually own:
2745 */
Thomas Gleixnerc0c9ed12011-03-11 11:51:22 +01002746 if ((uval & FUTEX_TID_MASK) != vpid)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002747 return -EPERM;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002748
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002749 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002750 if (ret)
2751 return ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002752
2753 hb = hash_futex(&key);
2754 spin_lock(&hb->lock);
2755
Ingo Molnarc87e2832006-06-27 02:54:58 -07002756 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002757 * Check waiters first. We do not trust user space values at
2758 * all and we at least want to know if user space fiddled
2759 * with the futex value instead of blindly unlocking.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002760 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002761 top_waiter = futex_top_waiter(hb, &key);
2762 if (top_waiter) {
2763 ret = wake_futex_pi(uaddr, uval, top_waiter, hb);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002764 /*
2765 * In case of success wake_futex_pi dropped the hash
2766 * bucket lock.
2767 */
2768 if (!ret)
2769 goto out_putkey;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002770 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002771 * The atomic access to the futex value generated a
2772 * pagefault, so retry the user-access and the wakeup:
Ingo Molnarc87e2832006-06-27 02:54:58 -07002773 */
2774 if (ret == -EFAULT)
2775 goto pi_faulted;
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002776 /*
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02002777 * A unconditional UNLOCK_PI op raced against a waiter
2778 * setting the FUTEX_WAITERS bit. Try again.
2779 */
2780 if (ret == -EAGAIN) {
2781 spin_unlock(&hb->lock);
2782 put_futex_key(&key);
2783 goto retry;
2784 }
2785 /*
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002786 * wake_futex_pi has detected invalid state. Tell user
2787 * space.
2788 */
Ingo Molnarc87e2832006-06-27 02:54:58 -07002789 goto out_unlock;
2790 }
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002791
Ingo Molnarc87e2832006-06-27 02:54:58 -07002792 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002793 * We have no kernel internal state, i.e. no waiters in the
2794 * kernel. Waiters which are about to queue themselves are stuck
2795 * on hb->lock. So we can safely ignore them. We do neither
2796 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2797 * owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002798 */
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002799 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))
Thomas Gleixner13fbca42014-06-03 12:27:07 +00002800 goto pi_faulted;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002801
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002802 /*
2803 * If uval has changed, let user space handle it.
2804 */
2805 ret = (curval == uval) ? 0 : -EAGAIN;
2806
Ingo Molnarc87e2832006-06-27 02:54:58 -07002807out_unlock:
2808 spin_unlock(&hb->lock);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002809out_putkey:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002810 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002811 return ret;
2812
2813pi_faulted:
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002814 spin_unlock(&hb->lock);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002815 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002816
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002817 ret = fault_in_user_writeable(uaddr);
Darren Hartb5686362008-12-18 15:06:34 -08002818 if (!ret)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002819 goto retry;
2820
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 return ret;
2822}
2823
Darren Hart52400ba2009-04-03 13:40:49 -07002824/**
2825 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2826 * @hb: the hash_bucket futex_q was original enqueued on
2827 * @q: the futex_q woken while waiting to be requeued
2828 * @key2: the futex_key of the requeue target futex
2829 * @timeout: the timeout associated with the wait (NULL if none)
2830 *
2831 * Detect if the task was woken on the initial futex as opposed to the requeue
2832 * target futex. If so, determine if it was a timeout or a signal that caused
2833 * the wakeup and return the appropriate error code to the caller. Must be
2834 * called with the hb lock held.
2835 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002836 * Return:
2837 * 0 = no early wakeup detected;
2838 * <0 = -ETIMEDOUT or -ERESTARTNOINTR
Darren Hart52400ba2009-04-03 13:40:49 -07002839 */
2840static inline
2841int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2842 struct futex_q *q, union futex_key *key2,
2843 struct hrtimer_sleeper *timeout)
2844{
2845 int ret = 0;
2846
2847 /*
2848 * With the hb lock held, we avoid races while we process the wakeup.
2849 * We only need to hold hb (and not hb2) to ensure atomicity as the
2850 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2851 * It can't be requeued from uaddr2 to something else since we don't
2852 * support a PI aware source futex for requeue.
2853 */
2854 if (!match_futex(&q->key, key2)) {
2855 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2856 /*
2857 * We were woken prior to requeue by a timeout or a signal.
2858 * Unqueue the futex_q and determine which it was.
2859 */
Lai Jiangshan2e129782010-12-22 14:18:50 +08002860 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07002861 hb_waiters_dec(hb);
Darren Hart52400ba2009-04-03 13:40:49 -07002862
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002863 /* Handle spurious wakeups gracefully */
Thomas Gleixner11df6dd2009-10-28 20:26:48 +01002864 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07002865 if (timeout && !timeout->task)
2866 ret = -ETIMEDOUT;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002867 else if (signal_pending(current))
Thomas Gleixner1c840c12009-05-20 09:22:40 +02002868 ret = -ERESTARTNOINTR;
Darren Hart52400ba2009-04-03 13:40:49 -07002869 }
2870 return ret;
2871}
2872
2873/**
2874 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
Darren Hart56ec1602009-09-21 22:29:59 -07002875 * @uaddr: the futex we initially wait on (non-pi)
Darren Hartb41277d2010-11-08 13:10:09 -08002876 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07002877 * the same type, no requeueing from private to shared, etc.
Darren Hart52400ba2009-04-03 13:40:49 -07002878 * @val: the expected value of uaddr
2879 * @abs_time: absolute timeout
Darren Hart56ec1602009-09-21 22:29:59 -07002880 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
Darren Hart52400ba2009-04-03 13:40:49 -07002881 * @uaddr2: the pi futex we will take prior to returning to user-space
2882 *
2883 * The caller will wait on uaddr and will be requeued by futex_requeue() to
Darren Hart6f7b0a22012-07-20 11:53:31 -07002884 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2885 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2886 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2887 * without one, the pi logic would not know which task to boost/deboost, if
2888 * there was a need to.
Darren Hart52400ba2009-04-03 13:40:49 -07002889 *
2890 * We call schedule in futex_wait_queue_me() when we enqueue and return there
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002891 * via the following--
Darren Hart52400ba2009-04-03 13:40:49 -07002892 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
Darren Hartcc6db4e2009-07-31 16:20:10 -07002893 * 2) wakeup on uaddr2 after a requeue
2894 * 3) signal
2895 * 4) timeout
Darren Hart52400ba2009-04-03 13:40:49 -07002896 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002897 * If 3, cleanup and return -ERESTARTNOINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07002898 *
2899 * If 2, we may then block on trying to take the rt_mutex and return via:
2900 * 5) successful lock
2901 * 6) signal
2902 * 7) timeout
2903 * 8) other lock acquisition failure
2904 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002905 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
Darren Hart52400ba2009-04-03 13:40:49 -07002906 *
2907 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2908 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002909 * Return:
2910 * 0 - On success;
Darren Hart52400ba2009-04-03 13:40:49 -07002911 * <0 - On error
2912 */
Darren Hartb41277d2010-11-08 13:10:09 -08002913static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
Darren Hart52400ba2009-04-03 13:40:49 -07002914 u32 val, ktime_t *abs_time, u32 bitset,
Darren Hartb41277d2010-11-08 13:10:09 -08002915 u32 __user *uaddr2)
Darren Hart52400ba2009-04-03 13:40:49 -07002916{
2917 struct hrtimer_sleeper timeout, *to = NULL;
2918 struct rt_mutex_waiter rt_waiter;
Darren Hart52400ba2009-04-03 13:40:49 -07002919 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002920 union futex_key key2 = FUTEX_KEY_INIT;
2921 struct futex_q q = futex_q_init;
Darren Hart52400ba2009-04-03 13:40:49 -07002922 int res, ret;
Darren Hart52400ba2009-04-03 13:40:49 -07002923
Darren Hart6f7b0a22012-07-20 11:53:31 -07002924 if (uaddr == uaddr2)
2925 return -EINVAL;
2926
Darren Hart52400ba2009-04-03 13:40:49 -07002927 if (!bitset)
2928 return -EINVAL;
2929
2930 if (abs_time) {
2931 to = &timeout;
Darren Hartb41277d2010-11-08 13:10:09 -08002932 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2933 CLOCK_REALTIME : CLOCK_MONOTONIC,
2934 HRTIMER_MODE_ABS);
Darren Hart52400ba2009-04-03 13:40:49 -07002935 hrtimer_init_sleeper(to, current);
2936 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2937 current->timer_slack_ns);
2938 }
2939
2940 /*
2941 * The waiter is allocated on our stack, manipulated by the requeue
2942 * code while we sleep on uaddr.
2943 */
2944 debug_rt_mutex_init_waiter(&rt_waiter);
Peter Zijlstrafb00aca2013-11-07 14:43:43 +01002945 RB_CLEAR_NODE(&rt_waiter.pi_tree_entry);
2946 RB_CLEAR_NODE(&rt_waiter.tree_entry);
Darren Hart52400ba2009-04-03 13:40:49 -07002947 rt_waiter.task = NULL;
2948
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002949 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Darren Hart52400ba2009-04-03 13:40:49 -07002950 if (unlikely(ret != 0))
2951 goto out;
2952
Darren Hart84bc4af2009-08-13 17:36:53 -07002953 q.bitset = bitset;
2954 q.rt_waiter = &rt_waiter;
2955 q.requeue_pi_key = &key2;
2956
Darren Hart7ada8762010-10-17 08:35:04 -07002957 /*
2958 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
2959 * count.
2960 */
Darren Hartb41277d2010-11-08 13:10:09 -08002961 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02002962 if (ret)
2963 goto out_key2;
Darren Hart52400ba2009-04-03 13:40:49 -07002964
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00002965 /*
2966 * The check above which compares uaddrs is not sufficient for
2967 * shared futexes. We need to compare the keys:
2968 */
2969 if (match_futex(&q.key, &key2)) {
Thomas Gleixner13c42c22014-09-11 23:44:35 +02002970 queue_unlock(hb);
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00002971 ret = -EINVAL;
2972 goto out_put_keys;
2973 }
2974
Darren Hart52400ba2009-04-03 13:40:49 -07002975 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002976 futex_wait_queue_me(hb, &q, to);
Darren Hart52400ba2009-04-03 13:40:49 -07002977
2978 spin_lock(&hb->lock);
2979 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
2980 spin_unlock(&hb->lock);
2981 if (ret)
2982 goto out_put_keys;
2983
2984 /*
2985 * In order for us to be here, we know our q.key == key2, and since
2986 * we took the hb->lock above, we also know that futex_requeue() has
2987 * completed and we no longer have to concern ourselves with a wakeup
Darren Hart7ada8762010-10-17 08:35:04 -07002988 * race with the atomic proxy lock acquisition by the requeue code. The
2989 * futex_requeue dropped our key1 reference and incremented our key2
2990 * reference count.
Darren Hart52400ba2009-04-03 13:40:49 -07002991 */
2992
2993 /* Check if the requeue code acquired the second futex for us. */
2994 if (!q.rt_waiter) {
2995 /*
2996 * Got the lock. We might not be the anticipated owner if we
2997 * did a lock-steal - fix up the PI-state in that case.
2998 */
2999 if (q.pi_state && (q.pi_state->owner != current)) {
3000 spin_lock(q.lock_ptr);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003001 ret = fixup_pi_state_owner(uaddr2, &q, current);
Peter Zijlstra9bbb25a2017-03-04 10:27:19 +01003002 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current)
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01003003 rt_mutex_futex_unlock(&q.pi_state->pi_mutex);
Thomas Gleixnerfb75a422015-12-19 20:07:38 +00003004 /*
3005 * Drop the reference to the pi state which
3006 * the requeue_pi() code acquired for us.
3007 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00003008 put_pi_state(q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003009 spin_unlock(q.lock_ptr);
3010 }
3011 } else {
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003012 struct rt_mutex *pi_mutex;
3013
Darren Hart52400ba2009-04-03 13:40:49 -07003014 /*
3015 * We have been woken up by futex_unlock_pi(), a timeout, or a
3016 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3017 * the pi_state.
3018 */
Darren Hartf27071c2012-07-20 11:53:30 -07003019 WARN_ON(!q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003020 pi_mutex = &q.pi_state->pi_mutex;
Thomas Gleixnerc051b212014-05-22 03:25:50 +00003021 ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003022 debug_rt_mutex_free_waiter(&rt_waiter);
3023
3024 spin_lock(q.lock_ptr);
3025 /*
3026 * Fixup the pi_state owner and possibly acquire the lock if we
3027 * haven't already.
3028 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003029 res = fixup_owner(uaddr2, &q, !ret);
Darren Hart52400ba2009-04-03 13:40:49 -07003030 /*
3031 * If fixup_owner() returned an error, proprogate that. If it
Darren Hart56ec1602009-09-21 22:29:59 -07003032 * acquired the lock, clear -ETIMEDOUT or -EINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07003033 */
3034 if (res)
3035 ret = (res < 0) ? res : 0;
3036
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003037 /*
3038 * If fixup_pi_state_owner() faulted and was unable to handle
3039 * the fault, unlock the rt_mutex and return the fault to
3040 * userspace.
3041 */
3042 if (ret && rt_mutex_owner(pi_mutex) == current)
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01003043 rt_mutex_futex_unlock(pi_mutex);
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003044
Darren Hart52400ba2009-04-03 13:40:49 -07003045 /* Unqueue and drop the lock. */
3046 unqueue_me_pi(&q);
3047 }
3048
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003049 if (ret == -EINTR) {
Darren Hart52400ba2009-04-03 13:40:49 -07003050 /*
Darren Hartcc6db4e2009-07-31 16:20:10 -07003051 * We've already been requeued, but cannot restart by calling
3052 * futex_lock_pi() directly. We could restart this syscall, but
3053 * it would detect that the user space "val" changed and return
3054 * -EWOULDBLOCK. Save the overhead of the restart and return
3055 * -EWOULDBLOCK directly.
Darren Hart52400ba2009-04-03 13:40:49 -07003056 */
Thomas Gleixner20708872009-05-19 23:04:59 +02003057 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07003058 }
3059
3060out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003061 put_futex_key(&q.key);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02003062out_key2:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003063 put_futex_key(&key2);
Darren Hart52400ba2009-04-03 13:40:49 -07003064
3065out:
3066 if (to) {
3067 hrtimer_cancel(&to->timer);
3068 destroy_hrtimer_on_stack(&to->timer);
3069 }
3070 return ret;
3071}
3072
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003073/*
3074 * Support for robust futexes: the kernel cleans up held futexes at
3075 * thread exit time.
3076 *
3077 * Implementation: user-space maintains a per-thread list of locks it
3078 * is holding. Upon do_exit(), the kernel carefully walks this list,
3079 * and marks all locks that are owned by this thread with the
Ingo Molnarc87e2832006-06-27 02:54:58 -07003080 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003081 * always manipulated with the lock held, so the list is private and
3082 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3083 * field, to allow the kernel to clean up if the thread dies after
3084 * acquiring the lock, but just before it could have added itself to
3085 * the list. There can only be one such pending lock.
3086 */
3087
3088/**
Darren Hartd96ee562009-09-21 22:30:22 -07003089 * sys_set_robust_list() - Set the robust-futex list head of a task
3090 * @head: pointer to the list-head
3091 * @len: length of the list-head, as userspace expects
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003092 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003093SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3094 size_t, len)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003095{
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003096 if (!futex_cmpxchg_enabled)
3097 return -ENOSYS;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003098 /*
3099 * The kernel knows only one size for now:
3100 */
3101 if (unlikely(len != sizeof(*head)))
3102 return -EINVAL;
3103
3104 current->robust_list = head;
3105
3106 return 0;
3107}
3108
3109/**
Darren Hartd96ee562009-09-21 22:30:22 -07003110 * sys_get_robust_list() - Get the robust-futex list head of a task
3111 * @pid: pid of the process [zero for current task]
3112 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3113 * @len_ptr: pointer to a length field, the kernel fills in the header size
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003114 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003115SYSCALL_DEFINE3(get_robust_list, int, pid,
3116 struct robust_list_head __user * __user *, head_ptr,
3117 size_t __user *, len_ptr)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003118{
Al Viroba46df92006-10-10 22:46:07 +01003119 struct robust_list_head __user *head;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003120 unsigned long ret;
Kees Cookbdbb7762012-03-19 16:12:53 -07003121 struct task_struct *p;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003122
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003123 if (!futex_cmpxchg_enabled)
3124 return -ENOSYS;
3125
Kees Cookbdbb7762012-03-19 16:12:53 -07003126 rcu_read_lock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003127
Kees Cookbdbb7762012-03-19 16:12:53 -07003128 ret = -ESRCH;
3129 if (!pid)
3130 p = current;
3131 else {
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07003132 p = find_task_by_vpid(pid);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003133 if (!p)
3134 goto err_unlock;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003135 }
3136
Kees Cookbdbb7762012-03-19 16:12:53 -07003137 ret = -EPERM;
Jann Horncaaee622016-01-20 15:00:04 -08003138 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
Kees Cookbdbb7762012-03-19 16:12:53 -07003139 goto err_unlock;
3140
3141 head = p->robust_list;
3142 rcu_read_unlock();
3143
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003144 if (put_user(sizeof(*head), len_ptr))
3145 return -EFAULT;
3146 return put_user(head, head_ptr);
3147
3148err_unlock:
Oleg Nesterovaaa2a972006-09-29 02:00:55 -07003149 rcu_read_unlock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003150
3151 return ret;
3152}
3153
3154/*
3155 * Process a futex-list entry, check whether it's owned by the
3156 * dying task, and do notification if so:
3157 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003158int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003159{
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03003160 u32 uval, uninitialized_var(nval), mval;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003161
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003162retry:
3163 if (get_user(uval, uaddr))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003164 return -1;
3165
Pavel Emelyanovb4888932007-10-18 23:40:14 -07003166 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003167 /*
3168 * Ok, this dying thread is truly holding a futex
3169 * of interest. Set the OWNER_DIED bit atomically
3170 * via cmpxchg, and if the value had FUTEX_WAITERS
3171 * set, wake up a waiter (if any). (We have to do a
3172 * futex_wake() even if OWNER_DIED is already set -
3173 * to handle the rare but possible case of recursive
3174 * thread-death.) The rest of the cleanup is done in
3175 * userspace.
3176 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003177 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
Thomas Gleixner6e0aa9f2011-03-14 10:34:35 +01003178 /*
3179 * We are not holding a lock here, but we want to have
3180 * the pagefault_disable/enable() protection because
3181 * we want to handle the fault gracefully. If the
3182 * access fails we try to fault in the futex with R/W
3183 * verification via get_user_pages. get_user() above
3184 * does not guarantee R/W access. If that fails we
3185 * give up and leave the futex locked.
3186 */
3187 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3188 if (fault_in_user_writeable(uaddr))
3189 return -1;
3190 goto retry;
3191 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07003192 if (nval != uval)
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003193 goto retry;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003194
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003195 /*
3196 * Wake robust non-PI futexes here. The wakeup of
3197 * PI futexes happens in exit_pi_state():
3198 */
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07003199 if (!pi && (uval & FUTEX_WAITERS))
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02003200 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003201 }
3202 return 0;
3203}
3204
3205/*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003206 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3207 */
3208static inline int fetch_robust_entry(struct robust_list __user **entry,
Al Viroba46df92006-10-10 22:46:07 +01003209 struct robust_list __user * __user *head,
Namhyung Kim1dcc41b2010-09-14 21:43:46 +09003210 unsigned int *pi)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003211{
3212 unsigned long uentry;
3213
Al Viroba46df92006-10-10 22:46:07 +01003214 if (get_user(uentry, (unsigned long __user *)head))
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003215 return -EFAULT;
3216
Al Viroba46df92006-10-10 22:46:07 +01003217 *entry = (void __user *)(uentry & ~1UL);
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003218 *pi = uentry & 1;
3219
3220 return 0;
3221}
3222
3223/*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003224 * Walk curr->robust_list (very carefully, it's a userspace list!)
3225 * and mark any locks found there dead, and notify any waiters.
3226 *
3227 * We silently return on any sign of list-walking problem.
3228 */
3229void exit_robust_list(struct task_struct *curr)
3230{
3231 struct robust_list_head __user *head = curr->robust_list;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003232 struct robust_list __user *entry, *next_entry, *pending;
Darren Hart4c115e92010-11-04 15:00:00 -04003233 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3234 unsigned int uninitialized_var(next_pi);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003235 unsigned long futex_offset;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003236 int rc;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003237
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003238 if (!futex_cmpxchg_enabled)
3239 return;
3240
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003241 /*
3242 * Fetch the list head (which was registered earlier, via
3243 * sys_set_robust_list()):
3244 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003245 if (fetch_robust_entry(&entry, &head->list.next, &pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003246 return;
3247 /*
3248 * Fetch the relative futex offset:
3249 */
3250 if (get_user(futex_offset, &head->futex_offset))
3251 return;
3252 /*
3253 * Fetch any possibly pending lock-add first, and handle it
3254 * if it exists:
3255 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003256 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003257 return;
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003258
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003259 next_entry = NULL; /* avoid warning with gcc */
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003260 while (entry != &head->list) {
3261 /*
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003262 * Fetch the next entry in the list before calling
3263 * handle_futex_death:
3264 */
3265 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3266 /*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003267 * A pending lock might already be on the list, so
Ingo Molnarc87e2832006-06-27 02:54:58 -07003268 * don't process it twice:
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003269 */
3270 if (entry != pending)
Al Viroba46df92006-10-10 22:46:07 +01003271 if (handle_futex_death((void __user *)entry + futex_offset,
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003272 curr, pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003273 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003274 if (rc)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003275 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003276 entry = next_entry;
3277 pi = next_pi;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003278 /*
3279 * Avoid excessively long or circular lists:
3280 */
3281 if (!--limit)
3282 break;
3283
3284 cond_resched();
3285 }
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003286
3287 if (pending)
3288 handle_futex_death((void __user *)pending + futex_offset,
3289 curr, pip);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003290}
3291
Pierre Peifferc19384b2007-05-09 02:35:02 -07003292long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
Ingo Molnare2970f22006-06-27 02:54:47 -07003293 u32 __user *uaddr2, u32 val2, u32 val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294{
Thomas Gleixner81b40532012-02-15 12:17:09 +01003295 int cmd = op & FUTEX_CMD_MASK;
Darren Hartb41277d2010-11-08 13:10:09 -08003296 unsigned int flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003298 if (!(op & FUTEX_PRIVATE_FLAG))
Darren Hartb41277d2010-11-08 13:10:09 -08003299 flags |= FLAGS_SHARED;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003300
Darren Hartb41277d2010-11-08 13:10:09 -08003301 if (op & FUTEX_CLOCK_REALTIME) {
3302 flags |= FLAGS_CLOCKRT;
Darren Hart337f1302015-12-18 13:36:37 -08003303 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3304 cmd != FUTEX_WAIT_REQUEUE_PI)
Darren Hartb41277d2010-11-08 13:10:09 -08003305 return -ENOSYS;
3306 }
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003307
3308 switch (cmd) {
Thomas Gleixner59263b52012-02-15 12:08:34 +01003309 case FUTEX_LOCK_PI:
3310 case FUTEX_UNLOCK_PI:
3311 case FUTEX_TRYLOCK_PI:
3312 case FUTEX_WAIT_REQUEUE_PI:
3313 case FUTEX_CMP_REQUEUE_PI:
3314 if (!futex_cmpxchg_enabled)
3315 return -ENOSYS;
3316 }
3317
3318 switch (cmd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319 case FUTEX_WAIT:
Thomas Gleixnercd689982008-02-01 17:45:14 +01003320 val3 = FUTEX_BITSET_MATCH_ANY;
3321 case FUTEX_WAIT_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003322 return futex_wait(uaddr, flags, val, timeout, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 case FUTEX_WAKE:
Thomas Gleixnercd689982008-02-01 17:45:14 +01003324 val3 = FUTEX_BITSET_MATCH_ANY;
3325 case FUTEX_WAKE_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003326 return futex_wake(uaddr, flags, val, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 case FUTEX_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003328 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 case FUTEX_CMP_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003330 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07003331 case FUTEX_WAKE_OP:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003332 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003333 case FUTEX_LOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003334 return futex_lock_pi(uaddr, flags, timeout, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003335 case FUTEX_UNLOCK_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003336 return futex_unlock_pi(uaddr, flags);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003337 case FUTEX_TRYLOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003338 return futex_lock_pi(uaddr, flags, NULL, 1);
Darren Hart52400ba2009-04-03 13:40:49 -07003339 case FUTEX_WAIT_REQUEUE_PI:
3340 val3 = FUTEX_BITSET_MATCH_ANY;
Thomas Gleixner81b40532012-02-15 12:17:09 +01003341 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3342 uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07003343 case FUTEX_CMP_REQUEUE_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003344 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345 }
Thomas Gleixner81b40532012-02-15 12:17:09 +01003346 return -ENOSYS;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347}
3348
3349
Heiko Carstens17da2bd2009-01-14 14:14:10 +01003350SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3351 struct timespec __user *, utime, u32 __user *, uaddr2,
3352 u32, val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353{
Pierre Peifferc19384b2007-05-09 02:35:02 -07003354 struct timespec ts;
3355 ktime_t t, *tp = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07003356 u32 val2 = 0;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003357 int cmd = op & FUTEX_CMD_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358
Thomas Gleixnercd689982008-02-01 17:45:14 +01003359 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
Darren Hart52400ba2009-04-03 13:40:49 -07003360 cmd == FUTEX_WAIT_BITSET ||
3361 cmd == FUTEX_WAIT_REQUEUE_PI)) {
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07003362 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3363 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003364 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003366 if (!timespec_valid(&ts))
Thomas Gleixner9741ef962006-03-31 02:31:32 -08003367 return -EINVAL;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003368
3369 t = timespec_to_ktime(ts);
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003370 if (cmd == FUTEX_WAIT)
Thomas Gleixner5a7780e2008-02-13 09:20:43 +01003371 t = ktime_add_safe(ktime_get(), t);
Pierre Peifferc19384b2007-05-09 02:35:02 -07003372 tp = &t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 }
3374 /*
Darren Hart52400ba2009-04-03 13:40:49 -07003375 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
Andreas Schwabf54f0982007-07-31 00:38:51 -07003376 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377 */
Andreas Schwabf54f0982007-07-31 00:38:51 -07003378 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
Darren Hartba9c22f2009-04-20 22:22:22 -07003379 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
Ingo Molnare2970f22006-06-27 02:54:47 -07003380 val2 = (u32) (unsigned long) utime;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381
Pierre Peifferc19384b2007-05-09 02:35:02 -07003382 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383}
3384
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003385static void __init futex_detect_cmpxchg(void)
3386{
3387#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3388 u32 curval;
3389
3390 /*
3391 * This will fail and we want it. Some arch implementations do
3392 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3393 * functionality. We want to know that before we call in any
3394 * of the complex code paths. Also we want to prevent
3395 * registration of robust lists in that case. NULL is
3396 * guaranteed to fault and we get -EFAULT on functional
3397 * implementation, the non-functional ones will return
3398 * -ENOSYS.
3399 */
3400 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3401 futex_cmpxchg_enabled = 1;
3402#endif
3403}
3404
Benjamin Herrenschmidtf6d107f2008-03-27 14:52:15 +11003405static int __init futex_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406{
Heiko Carstens63b1a812014-01-16 14:54:50 +01003407 unsigned int futex_shift;
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003408 unsigned long i;
3409
3410#if CONFIG_BASE_SMALL
3411 futex_hashsize = 16;
3412#else
3413 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3414#endif
3415
3416 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3417 futex_hashsize, 0,
3418 futex_hashsize < 256 ? HASH_SMALL : 0,
Heiko Carstens63b1a812014-01-16 14:54:50 +01003419 &futex_shift, NULL,
3420 futex_hashsize, futex_hashsize);
3421 futex_hashsize = 1UL << futex_shift;
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003422
3423 futex_detect_cmpxchg();
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003424
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003425 for (i = 0; i < futex_hashsize; i++) {
Linus Torvalds11d46162014-03-20 22:11:17 -07003426 atomic_set(&futex_queues[i].waiters, 0);
Dima Zavin732375c2011-07-07 17:27:59 -07003427 plist_head_init(&futex_queues[i].chain);
Thomas Gleixner3e4ab742008-02-23 15:23:55 -08003428 spin_lock_init(&futex_queues[i].lock);
3429 }
3430
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 return 0;
3432}
Yang Yang25f71d12016-12-30 16:17:55 +08003433core_initcall(futex_init);