blob: 57d0b3657e16b90268fa3396668bb62e6e54d287 [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;
Kees Cook3859a272016-10-28 01:22:25 -0700215} __randomize_layout;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700216
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 *
Ingo Molnarac6424b2017-06-20 12:06:13 +0200228 * We use this hashed waitqueue, instead of a normal wait_queue_entry_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 Gleixnercd6899852008-02-01 17:45:14 +0100248 u32 bitset;
Kees Cook3859a272016-10-28 01:22:25 -0700249} __randomize_layout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250
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 *
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -0300491 * 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
Mel Gorman48fb6f42017-08-09 08:27:11 +0100673 * truncated in parallel which is almost certainly an
674 * application bug. In such a case, just retry.
Mel Gorman65d8fc72016-02-09 11:15:14 -0800675 *
676 * We are not calling into get_futex_key_refs() in file-backed
677 * cases, therefore a successful atomic_inc return below will
678 * guarantee that get_futex_key() will still imply smp_mb(); (B).
679 */
Mel Gorman48fb6f42017-08-09 08:27:11 +0100680 if (!atomic_inc_not_zero(&inode->i_count)) {
Mel Gorman65d8fc72016-02-09 11:15:14 -0800681 rcu_read_unlock();
682 put_page(page);
683
684 goto again;
685 }
686
687 /* Should be impossible but lets be paranoid for now */
688 if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
689 err = -EFAULT;
690 rcu_read_unlock();
691 iput(inode);
692
693 goto out;
694 }
695
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200696 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
Mel Gorman65d8fc72016-02-09 11:15:14 -0800697 key->shared.inode = inode;
Mel Gorman077fa7a2016-06-08 14:25:22 +0100698 key->shared.pgoff = basepage_index(tail);
Mel Gorman65d8fc72016-02-09 11:15:14 -0800699 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 }
701
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500702out:
Kirill A. Shutemov14d27ab2016-01-15 16:53:00 -0800703 put_page(page);
Shawn Bohrer9ea71502011-06-30 11:21:32 -0500704 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705}
706
Thomas Gleixnerae791a22010-11-10 13:30:36 +0100707static inline void put_futex_key(union futex_key *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708{
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200709 drop_futex_key_refs(key);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710}
711
Darren Hartd96ee562009-09-21 22:30:22 -0700712/**
713 * fault_in_user_writeable() - Fault in user address and verify RW access
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200714 * @uaddr: pointer to faulting user space address
715 *
716 * Slow path to fixup the fault we just took in the atomic write
717 * access to @uaddr.
718 *
Randy Dunlapfb62db22010-10-13 11:02:34 -0700719 * We have no generic implementation of a non-destructive write to the
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200720 * user address. We know that we faulted in the atomic pagefault
721 * disabled section so we can as well avoid the #PF overhead by
722 * calling get_user_pages() right away.
723 */
724static int fault_in_user_writeable(u32 __user *uaddr)
725{
Andi Kleen722d0172009-12-08 13:19:42 +0100726 struct mm_struct *mm = current->mm;
727 int ret;
728
729 down_read(&mm->mmap_sem);
Benjamin Herrenschmidt2efaca92011-07-25 17:12:32 -0700730 ret = fixup_user_fault(current, mm, (unsigned long)uaddr,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800731 FAULT_FLAG_WRITE, NULL);
Andi Kleen722d0172009-12-08 13:19:42 +0100732 up_read(&mm->mmap_sem);
733
Thomas Gleixnerd0725992009-06-11 23:15:43 +0200734 return ret < 0 ? ret : 0;
735}
736
Darren Hart4b1c4862009-04-03 13:39:42 -0700737/**
738 * futex_top_waiter() - Return the highest priority waiter on a futex
Darren Hartd96ee562009-09-21 22:30:22 -0700739 * @hb: the hash bucket the futex_q's reside in
740 * @key: the futex key (to distinguish it from other futex futex_q's)
Darren Hart4b1c4862009-04-03 13:39:42 -0700741 *
742 * Must be called with the hb lock held.
743 */
744static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
745 union futex_key *key)
746{
747 struct futex_q *this;
748
749 plist_for_each_entry(this, &hb->chain, list) {
750 if (match_futex(&this->key, key))
751 return this;
752 }
753 return NULL;
754}
755
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800756static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
757 u32 uval, u32 newval)
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700758{
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800759 int ret;
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700760
761 pagefault_disable();
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800762 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700763 pagefault_enable();
764
Michel Lespinasse37a9d912011-03-10 18:48:51 -0800765 return ret;
Thomas Gleixner36cf3b52007-07-15 23:41:20 -0700766}
767
768static int get_futex_value_locked(u32 *dest, u32 __user *from)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769{
770 int ret;
771
Peter Zijlstraa8663742006-12-06 20:32:20 -0800772 pagefault_disable();
Linus Torvaldsbd28b142016-05-22 17:21:27 -0700773 ret = __get_user(*dest, from);
Peter Zijlstraa8663742006-12-06 20:32:20 -0800774 pagefault_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
776 return ret ? -EFAULT : 0;
777}
778
Ingo Molnarc87e2832006-06-27 02:54:58 -0700779
780/*
781 * PI code:
782 */
783static int refill_pi_state_cache(void)
784{
785 struct futex_pi_state *pi_state;
786
787 if (likely(current->pi_state_cache))
788 return 0;
789
Burman Yan4668edc2006-12-06 20:38:51 -0800790 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700791
792 if (!pi_state)
793 return -ENOMEM;
794
Ingo Molnarc87e2832006-06-27 02:54:58 -0700795 INIT_LIST_HEAD(&pi_state->list);
796 /* pi_mutex gets initialized later */
797 pi_state->owner = NULL;
798 atomic_set(&pi_state->refcount, 1);
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200799 pi_state->key = FUTEX_KEY_INIT;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700800
801 current->pi_state_cache = pi_state;
802
803 return 0;
804}
805
Peter Zijlstrabf92cf32017-03-22 11:35:53 +0100806static struct futex_pi_state *alloc_pi_state(void)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700807{
808 struct futex_pi_state *pi_state = current->pi_state_cache;
809
810 WARN_ON(!pi_state);
811 current->pi_state_cache = NULL;
812
813 return pi_state;
814}
815
Peter Zijlstrabf92cf32017-03-22 11:35:53 +0100816static void get_pi_state(struct futex_pi_state *pi_state)
817{
818 WARN_ON_ONCE(!atomic_inc_not_zero(&pi_state->refcount));
819}
820
Brian Silverman30a6b802014-10-25 20:20:37 -0400821/*
Thomas Gleixner29e9ee52015-12-19 20:07:39 +0000822 * Drops a reference to the pi_state object and frees or caches it
823 * when the last reference is gone.
Brian Silverman30a6b802014-10-25 20:20:37 -0400824 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +0000825static void put_pi_state(struct futex_pi_state *pi_state)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700826{
Brian Silverman30a6b802014-10-25 20:20:37 -0400827 if (!pi_state)
828 return;
829
Ingo Molnarc87e2832006-06-27 02:54:58 -0700830 if (!atomic_dec_and_test(&pi_state->refcount))
831 return;
832
833 /*
834 * If pi_state->owner is NULL, the owner is most probably dying
835 * and has cleaned up the pi_state already
836 */
837 if (pi_state->owner) {
Peter Zijlstrac74aef22017-09-22 17:48:06 +0200838 struct task_struct *owner;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700839
Peter Zijlstrac74aef22017-09-22 17:48:06 +0200840 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
841 owner = pi_state->owner;
842 if (owner) {
843 raw_spin_lock(&owner->pi_lock);
844 list_del_init(&pi_state->list);
845 raw_spin_unlock(&owner->pi_lock);
846 }
847 rt_mutex_proxy_unlock(&pi_state->pi_mutex, owner);
848 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700849 }
850
Peter Zijlstrac74aef22017-09-22 17:48:06 +0200851 if (current->pi_state_cache) {
Ingo Molnarc87e2832006-06-27 02:54:58 -0700852 kfree(pi_state);
Peter Zijlstrac74aef22017-09-22 17:48:06 +0200853 } else {
Ingo Molnarc87e2832006-06-27 02:54:58 -0700854 /*
855 * pi_state->list is already empty.
856 * clear pi_state->owner.
857 * refcount is at 0 - put it back to 1.
858 */
859 pi_state->owner = NULL;
860 atomic_set(&pi_state->refcount, 1);
861 current->pi_state_cache = pi_state;
862 }
863}
864
865/*
866 * Look up the task based on what TID userspace gave us.
867 * We dont trust it.
868 */
Peter Zijlstrabf92cf32017-03-22 11:35:53 +0100869static struct task_struct *futex_find_get_task(pid_t pid)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700870{
871 struct task_struct *p;
872
Oleg Nesterovd359b542006-09-29 02:00:55 -0700873 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -0700874 p = find_task_by_vpid(pid);
Michal Hocko7a0ea092010-06-30 09:51:19 +0200875 if (p)
876 get_task_struct(p);
Thomas Gleixnera06381f2007-06-23 11:48:40 +0200877
Oleg Nesterovd359b542006-09-29 02:00:55 -0700878 rcu_read_unlock();
Ingo Molnarc87e2832006-06-27 02:54:58 -0700879
880 return p;
881}
882
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -0400883#ifdef CONFIG_FUTEX_PI
884
Ingo Molnarc87e2832006-06-27 02:54:58 -0700885/*
886 * This task is holding PI mutexes at exit time => bad.
887 * Kernel cleans up PI-state, but userspace is likely hosed.
888 * (Robust-futex cleanup is separate and might save the day for userspace.)
889 */
890void exit_pi_state_list(struct task_struct *curr)
891{
Ingo Molnarc87e2832006-06-27 02:54:58 -0700892 struct list_head *next, *head = &curr->pi_state_list;
893 struct futex_pi_state *pi_state;
Ingo Molnar627371d2006-07-29 05:16:20 +0200894 struct futex_hash_bucket *hb;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200895 union futex_key key = FUTEX_KEY_INIT;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700896
Thomas Gleixnera0c1e902008-02-23 15:23:57 -0800897 if (!futex_cmpxchg_enabled)
898 return;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700899 /*
900 * We are a ZOMBIE and nobody can enqueue itself on
901 * pi_state_list anymore, but we have to be careful
Ingo Molnar627371d2006-07-29 05:16:20 +0200902 * versus waiters unqueueing themselves:
Ingo Molnarc87e2832006-06-27 02:54:58 -0700903 */
Thomas Gleixner1d615482009-11-17 14:54:03 +0100904 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700905 while (!list_empty(head)) {
Ingo Molnarc87e2832006-06-27 02:54:58 -0700906 next = head->next;
907 pi_state = list_entry(next, struct futex_pi_state, list);
908 key = pi_state->key;
Ingo Molnar627371d2006-07-29 05:16:20 +0200909 hb = hash_futex(&key);
Peter Zijlstra153fbd12017-10-31 11:18:53 +0100910
911 /*
912 * We can race against put_pi_state() removing itself from the
913 * list (a waiter going away). put_pi_state() will first
914 * decrement the reference count and then modify the list, so
915 * its possible to see the list entry but fail this reference
916 * acquire.
917 *
918 * In that case; drop the locks to let put_pi_state() make
919 * progress and retry the loop.
920 */
921 if (!atomic_inc_not_zero(&pi_state->refcount)) {
922 raw_spin_unlock_irq(&curr->pi_lock);
923 cpu_relax();
924 raw_spin_lock_irq(&curr->pi_lock);
925 continue;
926 }
Thomas Gleixner1d615482009-11-17 14:54:03 +0100927 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700928
Ingo Molnarc87e2832006-06-27 02:54:58 -0700929 spin_lock(&hb->lock);
Peter Zijlstrac74aef22017-09-22 17:48:06 +0200930 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
931 raw_spin_lock(&curr->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +0200932 /*
933 * We dropped the pi-lock, so re-check whether this
934 * task still owns the PI-state:
935 */
Ingo Molnarc87e2832006-06-27 02:54:58 -0700936 if (head->next != next) {
Peter Zijlstra153fbd12017-10-31 11:18:53 +0100937 /* retain curr->pi_lock for the loop invariant */
Peter Zijlstrac74aef22017-09-22 17:48:06 +0200938 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700939 spin_unlock(&hb->lock);
Peter Zijlstra153fbd12017-10-31 11:18:53 +0100940 put_pi_state(pi_state);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700941 continue;
942 }
943
Ingo Molnarc87e2832006-06-27 02:54:58 -0700944 WARN_ON(pi_state->owner != curr);
Ingo Molnar627371d2006-07-29 05:16:20 +0200945 WARN_ON(list_empty(&pi_state->list));
946 list_del_init(&pi_state->list);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700947 pi_state->owner = NULL;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700948
Peter Zijlstra153fbd12017-10-31 11:18:53 +0100949 raw_spin_unlock(&curr->pi_lock);
Peter Zijlstrac74aef22017-09-22 17:48:06 +0200950 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700951 spin_unlock(&hb->lock);
952
Peter Zijlstra16ffa122017-03-22 11:35:55 +0100953 rt_mutex_futex_unlock(&pi_state->pi_mutex);
954 put_pi_state(pi_state);
955
Thomas Gleixner1d615482009-11-17 14:54:03 +0100956 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700957 }
Thomas Gleixner1d615482009-11-17 14:54:03 +0100958 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700959}
960
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -0400961#endif
962
Thomas Gleixner54a217882014-06-03 12:27:08 +0000963/*
964 * We need to check the following states:
965 *
966 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
967 *
968 * [1] NULL | --- | --- | 0 | 0/1 | Valid
969 * [2] NULL | --- | --- | >0 | 0/1 | Valid
970 *
971 * [3] Found | NULL | -- | Any | 0/1 | Invalid
972 *
973 * [4] Found | Found | NULL | 0 | 1 | Valid
974 * [5] Found | Found | NULL | >0 | 1 | Invalid
975 *
976 * [6] Found | Found | task | 0 | 1 | Valid
977 *
978 * [7] Found | Found | NULL | Any | 0 | Invalid
979 *
980 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
981 * [9] Found | Found | task | 0 | 0 | Invalid
982 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
983 *
984 * [1] Indicates that the kernel can acquire the futex atomically. We
985 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
986 *
987 * [2] Valid, if TID does not belong to a kernel thread. If no matching
988 * thread is found then it indicates that the owner TID has died.
989 *
990 * [3] Invalid. The waiter is queued on a non PI futex
991 *
992 * [4] Valid state after exit_robust_list(), which sets the user space
993 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
994 *
995 * [5] The user space value got manipulated between exit_robust_list()
996 * and exit_pi_state_list()
997 *
998 * [6] Valid state after exit_pi_state_list() which sets the new owner in
999 * the pi_state but cannot access the user space value.
1000 *
1001 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
1002 *
1003 * [8] Owner and user space value match
1004 *
1005 * [9] There is no transient state which sets the user space TID to 0
1006 * except exit_robust_list(), but this is indicated by the
1007 * FUTEX_OWNER_DIED bit. See [4]
1008 *
1009 * [10] There is no transient state which leaves owner and user space
1010 * TID out of sync.
Peter Zijlstra734009e2017-03-22 11:35:52 +01001011 *
1012 *
1013 * Serialization and lifetime rules:
1014 *
1015 * hb->lock:
1016 *
1017 * hb -> futex_q, relation
1018 * futex_q -> pi_state, relation
1019 *
1020 * (cannot be raw because hb can contain arbitrary amount
1021 * of futex_q's)
1022 *
1023 * pi_mutex->wait_lock:
1024 *
1025 * {uval, pi_state}
1026 *
1027 * (and pi_mutex 'obviously')
1028 *
1029 * p->pi_lock:
1030 *
1031 * p->pi_state_list -> pi_state->list, relation
1032 *
1033 * pi_state->refcount:
1034 *
1035 * pi_state lifetime
1036 *
1037 *
1038 * Lock order:
1039 *
1040 * hb->lock
1041 * pi_mutex->wait_lock
1042 * p->pi_lock
1043 *
Thomas Gleixner54a217882014-06-03 12:27:08 +00001044 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001045
1046/*
1047 * Validate that the existing waiter has a pi_state and sanity check
1048 * the pi_state against the user space value. If correct, attach to
1049 * it.
1050 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001051static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1052 struct futex_pi_state *pi_state,
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001053 struct futex_pi_state **ps)
1054{
1055 pid_t pid = uval & FUTEX_TID_MASK;
Peter Zijlstra94ffac52017-04-07 09:04:07 +02001056 u32 uval2;
1057 int ret;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001058
1059 /*
1060 * Userspace might have messed up non-PI and PI futexes [3]
1061 */
1062 if (unlikely(!pi_state))
1063 return -EINVAL;
1064
Peter Zijlstra734009e2017-03-22 11:35:52 +01001065 /*
1066 * We get here with hb->lock held, and having found a
1067 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1068 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1069 * which in turn means that futex_lock_pi() still has a reference on
1070 * our pi_state.
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001071 *
1072 * The waiter holding a reference on @pi_state also protects against
1073 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1074 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1075 * free pi_state before we can take a reference ourselves.
Peter Zijlstra734009e2017-03-22 11:35:52 +01001076 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001077 WARN_ON(!atomic_read(&pi_state->refcount));
1078
1079 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01001080 * Now that we have a pi_state, we can acquire wait_lock
1081 * and do the state validation.
1082 */
1083 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1084
1085 /*
1086 * Since {uval, pi_state} is serialized by wait_lock, and our current
1087 * uval was read without holding it, it can have changed. Verify it
1088 * still is what we expect it to be, otherwise retry the entire
1089 * operation.
1090 */
1091 if (get_futex_value_locked(&uval2, uaddr))
1092 goto out_efault;
1093
1094 if (uval != uval2)
1095 goto out_eagain;
1096
1097 /*
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001098 * Handle the owner died case:
1099 */
1100 if (uval & FUTEX_OWNER_DIED) {
1101 /*
1102 * exit_pi_state_list sets owner to NULL and wakes the
1103 * topmost waiter. The task which acquires the
1104 * pi_state->rt_mutex will fixup owner.
1105 */
1106 if (!pi_state->owner) {
1107 /*
1108 * No pi state owner, but the user space TID
1109 * is not 0. Inconsistent state. [5]
1110 */
1111 if (pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001112 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001113 /*
1114 * Take a ref on the state and return success. [4]
1115 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001116 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001117 }
1118
1119 /*
1120 * If TID is 0, then either the dying owner has not
1121 * yet executed exit_pi_state_list() or some waiter
1122 * acquired the rtmutex in the pi state, but did not
1123 * yet fixup the TID in user space.
1124 *
1125 * Take a ref on the state and return success. [6]
1126 */
1127 if (!pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001128 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001129 } else {
1130 /*
1131 * If the owner died bit is not set, then the pi_state
1132 * must have an owner. [7]
1133 */
1134 if (!pi_state->owner)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001135 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001136 }
1137
1138 /*
1139 * Bail out if user space manipulated the futex value. If pi
1140 * state exists then the owner TID must be the same as the
1141 * user space TID. [9/10]
1142 */
1143 if (pid != task_pid_vnr(pi_state->owner))
Peter Zijlstra734009e2017-03-22 11:35:52 +01001144 goto out_einval;
1145
1146out_attach:
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01001147 get_pi_state(pi_state);
Peter Zijlstra734009e2017-03-22 11:35:52 +01001148 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001149 *ps = pi_state;
1150 return 0;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001151
1152out_einval:
1153 ret = -EINVAL;
1154 goto out_error;
1155
1156out_eagain:
1157 ret = -EAGAIN;
1158 goto out_error;
1159
1160out_efault:
1161 ret = -EFAULT;
1162 goto out_error;
1163
1164out_error:
1165 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1166 return ret;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001167}
1168
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001169/*
1170 * Lookup the task for the TID provided from user space and attach to
1171 * it after doing proper sanity checks.
1172 */
1173static int attach_to_pi_owner(u32 uval, union futex_key *key,
1174 struct futex_pi_state **ps)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001175{
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001176 pid_t pid = uval & FUTEX_TID_MASK;
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001177 struct futex_pi_state *pi_state;
1178 struct task_struct *p;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001179
1180 /*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001181 * We are the first waiter - try to look up the real owner and attach
Thomas Gleixner54a217882014-06-03 12:27:08 +00001182 * the new pi_state to it, but bail out when TID = 0 [1]
Ingo Molnarc87e2832006-06-27 02:54:58 -07001183 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001184 if (!pid)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001185 return -ESRCH;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001186 p = futex_find_get_task(pid);
Michal Hocko7a0ea092010-06-30 09:51:19 +02001187 if (!p)
1188 return -ESRCH;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001189
Oleg Nesterova2129462015-02-02 15:05:36 +01001190 if (unlikely(p->flags & PF_KTHREAD)) {
Thomas Gleixnerf0d71b32014-05-12 20:45:35 +00001191 put_task_struct(p);
1192 return -EPERM;
1193 }
1194
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001195 /*
1196 * We need to look at the task state flags to figure out,
1197 * whether the task is exiting. To protect against the do_exit
1198 * change of the task flags, we do this protected by
1199 * p->pi_lock:
1200 */
Thomas Gleixner1d615482009-11-17 14:54:03 +01001201 raw_spin_lock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001202 if (unlikely(p->flags & PF_EXITING)) {
1203 /*
1204 * The task is on the way out. When PF_EXITPIDONE is
1205 * set, we know that the task has finished the
1206 * cleanup:
1207 */
1208 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1209
Thomas Gleixner1d615482009-11-17 14:54:03 +01001210 raw_spin_unlock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001211 put_task_struct(p);
1212 return ret;
1213 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001214
Thomas Gleixner54a217882014-06-03 12:27:08 +00001215 /*
1216 * No existing pi state. First waiter. [2]
Peter Zijlstra734009e2017-03-22 11:35:52 +01001217 *
1218 * This creates pi_state, we have hb->lock held, this means nothing can
1219 * observe this state, wait_lock is irrelevant.
Thomas Gleixner54a217882014-06-03 12:27:08 +00001220 */
Ingo Molnarc87e2832006-06-27 02:54:58 -07001221 pi_state = alloc_pi_state();
1222
1223 /*
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001224 * Initialize the pi_mutex in locked state and make @p
Ingo Molnarc87e2832006-06-27 02:54:58 -07001225 * the owner of it:
1226 */
1227 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1228
1229 /* Store the key for possible exit cleanups: */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001230 pi_state->key = *key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001231
Ingo Molnar627371d2006-07-29 05:16:20 +02001232 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001233 list_add(&pi_state->list, &p->pi_state_list);
Peter Zijlstrac74aef22017-09-22 17:48:06 +02001234 /*
1235 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1236 * because there is no concurrency as the object is not published yet.
1237 */
Ingo Molnarc87e2832006-06-27 02:54:58 -07001238 pi_state->owner = p;
Thomas Gleixner1d615482009-11-17 14:54:03 +01001239 raw_spin_unlock_irq(&p->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001240
1241 put_task_struct(p);
1242
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001243 *ps = pi_state;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001244
1245 return 0;
1246}
1247
Peter Zijlstra734009e2017-03-22 11:35:52 +01001248static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1249 struct futex_hash_bucket *hb,
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001250 union futex_key *key, struct futex_pi_state **ps)
1251{
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001252 struct futex_q *top_waiter = futex_top_waiter(hb, key);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001253
1254 /*
1255 * If there is a waiter on that futex, validate it and
1256 * attach to the pi_state when the validation succeeds.
1257 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001258 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001259 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001260
1261 /*
1262 * We are the first waiter - try to look up the owner based on
1263 * @uval and attach to it.
1264 */
1265 return attach_to_pi_owner(uval, key, ps);
1266}
1267
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001268static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1269{
1270 u32 uninitialized_var(curval);
1271
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001272 if (unlikely(should_fail_futex(true)))
1273 return -EFAULT;
1274
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001275 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1276 return -EFAULT;
1277
Peter Zijlstra734009e2017-03-22 11:35:52 +01001278 /* If user space value changed, let the caller retry */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001279 return curval != uval ? -EAGAIN : 0;
1280}
1281
Darren Hart1a520842009-04-03 13:39:52 -07001282/**
Darren Hartd96ee562009-09-21 22:30:22 -07001283 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
Darren Hartbab5bc92009-04-07 23:23:50 -07001284 * @uaddr: the pi futex user address
1285 * @hb: the pi futex hash bucket
1286 * @key: the futex key associated with uaddr and hb
1287 * @ps: the pi_state pointer where we store the result of the
1288 * lookup
1289 * @task: the task to perform the atomic lock work for. This will
1290 * be "current" except in the case of requeue pi.
1291 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart1a520842009-04-03 13:39:52 -07001292 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001293 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03001294 * - 0 - ready to wait;
1295 * - 1 - acquired the lock;
1296 * - <0 - error
Darren Hart1a520842009-04-03 13:39:52 -07001297 *
1298 * The hb->lock and futex_key refs shall be held by the caller.
1299 */
1300static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1301 union futex_key *key,
1302 struct futex_pi_state **ps,
Darren Hartbab5bc92009-04-07 23:23:50 -07001303 struct task_struct *task, int set_waiters)
Darren Hart1a520842009-04-03 13:39:52 -07001304{
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001305 u32 uval, newval, vpid = task_pid_vnr(task);
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001306 struct futex_q *top_waiter;
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001307 int ret;
Darren Hart1a520842009-04-03 13:39:52 -07001308
1309 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001310 * Read the user space value first so we can validate a few
1311 * things before proceeding further.
Darren Hart1a520842009-04-03 13:39:52 -07001312 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001313 if (get_futex_value_locked(&uval, uaddr))
Darren Hart1a520842009-04-03 13:39:52 -07001314 return -EFAULT;
1315
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001316 if (unlikely(should_fail_futex(true)))
1317 return -EFAULT;
1318
Darren Hart1a520842009-04-03 13:39:52 -07001319 /*
1320 * Detect deadlocks.
1321 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001322 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
Darren Hart1a520842009-04-03 13:39:52 -07001323 return -EDEADLK;
1324
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001325 if ((unlikely(should_fail_futex(true))))
1326 return -EDEADLK;
1327
Darren Hart1a520842009-04-03 13:39:52 -07001328 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001329 * Lookup existing state first. If it exists, try to attach to
1330 * its pi_state.
Darren Hart1a520842009-04-03 13:39:52 -07001331 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001332 top_waiter = futex_top_waiter(hb, key);
1333 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001334 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001335
1336 /*
1337 * No waiter and user TID is 0. We are here because the
1338 * waiters or the owner died bit is set or called from
1339 * requeue_cmp_pi or for whatever reason something took the
1340 * syscall.
1341 */
1342 if (!(uval & FUTEX_TID_MASK)) {
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001343 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001344 * We take over the futex. No other waiters and the user space
1345 * TID is 0. We preserve the owner died bit.
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001346 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001347 newval = uval & FUTEX_OWNER_DIED;
1348 newval |= vpid;
1349
1350 /* The futex requeue_pi code can enforce the waiters bit */
1351 if (set_waiters)
1352 newval |= FUTEX_WAITERS;
1353
1354 ret = lock_pi_update_atomic(uaddr, uval, newval);
1355 /* If the take over worked, return 1 */
1356 return ret < 0 ? ret : 1;
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001357 }
Darren Hart1a520842009-04-03 13:39:52 -07001358
Darren Hart1a520842009-04-03 13:39:52 -07001359 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001360 * First waiter. Set the waiters bit before attaching ourself to
1361 * the owner. If owner tries to unlock, it will be forced into
1362 * the kernel and blocked on hb->lock.
Darren Hart1a520842009-04-03 13:39:52 -07001363 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001364 newval = uval | FUTEX_WAITERS;
1365 ret = lock_pi_update_atomic(uaddr, uval, newval);
1366 if (ret)
1367 return ret;
Darren Hart1a520842009-04-03 13:39:52 -07001368 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001369 * If the update of the user space value succeeded, we try to
1370 * attach to the owner. If that fails, no harm done, we only
1371 * set the FUTEX_WAITERS bit in the user space variable.
Darren Hart1a520842009-04-03 13:39:52 -07001372 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001373 return attach_to_pi_owner(uval, key, ps);
Darren Hart1a520842009-04-03 13:39:52 -07001374}
1375
Lai Jiangshan2e129782010-12-22 14:18:50 +08001376/**
1377 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1378 * @q: The futex_q to unqueue
1379 *
1380 * The q->lock_ptr must not be NULL and must be held by the caller.
1381 */
1382static void __unqueue_futex(struct futex_q *q)
1383{
1384 struct futex_hash_bucket *hb;
1385
Steven Rostedt29096202011-03-17 15:21:07 -04001386 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1387 || WARN_ON(plist_node_empty(&q->list)))
Lai Jiangshan2e129782010-12-22 14:18:50 +08001388 return;
1389
1390 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1391 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001392 hb_waiters_dec(hb);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001393}
1394
Ingo Molnarc87e2832006-06-27 02:54:58 -07001395/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 * The hash bucket lock must be held when this is called.
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001397 * Afterwards, the futex_q must not be accessed. Callers
1398 * must ensure to later call wake_up_q() for the actual
1399 * wakeups to occur.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001401static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402{
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001403 struct task_struct *p = q->task;
1404
Darren Hartaa109902012-11-26 16:29:56 -08001405 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1406 return;
1407
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001408 /*
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001409 * Queue the task for later wakeup for after we've released
1410 * the hb->lock. wake_q_add() grabs reference to p.
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001411 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001412 wake_q_add(wake_q, p);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001413 __unqueue_futex(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 /*
Darren Hart (VMware)38fcd062017-04-14 15:31:38 -07001415 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1416 * is written, without taking any locks. This is possible in the event
1417 * of a spurious wakeup, for example. A memory barrier is required here
1418 * to prevent the following store to lock_ptr from getting ahead of the
1419 * plist_del in __unqueue_futex().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 */
Peter Zijlstra1b367ec2017-03-22 11:35:49 +01001421 smp_store_release(&q->lock_ptr, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422}
1423
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001424/*
1425 * Caller must hold a reference on @pi_state.
1426 */
1427static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001428{
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03001429 u32 uninitialized_var(curval), newval;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001430 struct task_struct *new_owner;
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001431 bool postunlock = false;
Waiman Long194a6b52016-11-17 11:46:38 -05001432 DEFINE_WAKE_Q(wake_q);
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001433 int ret = 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001434
Ingo Molnarc87e2832006-06-27 02:54:58 -07001435 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01001436 if (WARN_ON_ONCE(!new_owner)) {
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001437 /*
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01001438 * As per the comment in futex_unlock_pi() this should not happen.
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001439 *
1440 * When this happens, give up our locks and try again, giving
1441 * the futex_lock_pi() instance time to complete, either by
1442 * waiting on the rtmutex or removing itself from the futex
1443 * queue.
1444 */
1445 ret = -EAGAIN;
1446 goto out_unlock;
Peter Zijlstra73d786b2017-03-22 11:35:54 +01001447 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001448
1449 /*
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001450 * We pass it to the next owner. The WAITERS bit is always kept
1451 * enabled while there is PI state around. We cleanup the owner
1452 * died bit, because we are the owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07001453 */
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001454 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001455
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001456 if (unlikely(should_fail_futex(true)))
1457 ret = -EFAULT;
1458
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001459 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001460 ret = -EFAULT;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001461
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001462 } else if (curval != uval) {
1463 /*
1464 * If a unconditional UNLOCK_PI operation (user space did not
1465 * try the TID->0 transition) raced with a waiter setting the
1466 * FUTEX_WAITERS flag between get_user() and locking the hash
1467 * bucket lock, retry the operation.
1468 */
1469 if ((FUTEX_TID_MASK & curval) == uval)
1470 ret = -EAGAIN;
1471 else
1472 ret = -EINVAL;
1473 }
Peter Zijlstra734009e2017-03-22 11:35:52 +01001474
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001475 if (ret)
1476 goto out_unlock;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001477
Peter Zijlstra94ffac52017-04-07 09:04:07 +02001478 /*
1479 * This is a point of no return; once we modify the uval there is no
1480 * going back and subsequent operations must not fail.
1481 */
1482
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001483 raw_spin_lock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001484 WARN_ON(list_empty(&pi_state->list));
1485 list_del_init(&pi_state->list);
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001486 raw_spin_unlock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001487
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001488 raw_spin_lock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001489 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001490 list_add(&pi_state->list, &new_owner->pi_state_list);
1491 pi_state->owner = new_owner;
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001492 raw_spin_unlock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001493
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001494 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001495
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001496out_unlock:
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001497 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001498
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001499 if (postunlock)
1500 rt_mutex_postunlock(&wake_q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001501
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001502 return ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001503}
1504
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505/*
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001506 * Express the locking dependencies for lockdep:
1507 */
1508static inline void
1509double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1510{
1511 if (hb1 <= hb2) {
1512 spin_lock(&hb1->lock);
1513 if (hb1 < hb2)
1514 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1515 } else { /* hb1 > hb2 */
1516 spin_lock(&hb2->lock);
1517 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1518 }
1519}
1520
Darren Hart5eb3dc62009-03-12 00:55:52 -07001521static inline void
1522double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1523{
Darren Hartf061d352009-03-12 15:11:18 -07001524 spin_unlock(&hb1->lock);
Ingo Molnar88f502f2009-03-13 10:32:07 +01001525 if (hb1 != hb2)
1526 spin_unlock(&hb2->lock);
Darren Hart5eb3dc62009-03-12 00:55:52 -07001527}
1528
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001529/*
Darren Hartb2d09942009-03-12 00:55:37 -07001530 * Wake up waiters matching bitset queued on this futex (uaddr).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 */
Darren Hartb41277d2010-11-08 13:10:09 -08001532static int
1533futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534{
Ingo Molnare2970f22006-06-27 02:54:47 -07001535 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536 struct futex_q *this, *next;
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001537 union futex_key key = FUTEX_KEY_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 int ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001539 DEFINE_WAKE_Q(wake_q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540
Thomas Gleixnercd6899852008-02-01 17:45:14 +01001541 if (!bitset)
1542 return -EINVAL;
1543
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001544 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 if (unlikely(ret != 0))
1546 goto out;
1547
Ingo Molnare2970f22006-06-27 02:54:47 -07001548 hb = hash_futex(&key);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001549
1550 /* Make sure we really have tasks to wakeup */
1551 if (!hb_waiters_pending(hb))
1552 goto out_put_key;
1553
Ingo Molnare2970f22006-06-27 02:54:47 -07001554 spin_lock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555
Jason Low0d00c7b2014-01-12 15:31:22 -08001556 plist_for_each_entry_safe(this, next, &hb->chain, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557 if (match_futex (&this->key, &key)) {
Darren Hart52400ba2009-04-03 13:40:49 -07001558 if (this->pi_state || this->rt_waiter) {
Ingo Molnared6f7b12006-07-01 04:35:46 -07001559 ret = -EINVAL;
1560 break;
1561 }
Thomas Gleixnercd6899852008-02-01 17:45:14 +01001562
1563 /* Check if one of the bits is set in both bitsets */
1564 if (!(this->bitset & bitset))
1565 continue;
1566
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001567 mark_wake_futex(&wake_q, this);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568 if (++ret >= nr_wake)
1569 break;
1570 }
1571 }
1572
Ingo Molnare2970f22006-06-27 02:54:47 -07001573 spin_unlock(&hb->lock);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001574 wake_up_q(&wake_q);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001575out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001576 put_futex_key(&key);
Darren Hart42d35d42008-12-29 15:49:53 -08001577out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 return ret;
1579}
1580
Jiri Slaby30d6e0a2017-08-24 09:31:05 +02001581static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1582{
1583 unsigned int op = (encoded_op & 0x70000000) >> 28;
1584 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
Jiri Slabyd70ef222017-11-30 15:35:44 +01001585 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1586 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
Jiri Slaby30d6e0a2017-08-24 09:31:05 +02001587 int oldval, ret;
1588
1589 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
Jiri Slabye78c38f62017-10-23 13:41:51 +02001590 if (oparg < 0 || oparg > 31) {
1591 char comm[sizeof(current->comm)];
1592 /*
1593 * kill this print and return -EINVAL when userspace
1594 * is sane again
1595 */
1596 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1597 get_task_comm(comm, current), oparg);
1598 oparg &= 31;
1599 }
Jiri Slaby30d6e0a2017-08-24 09:31:05 +02001600 oparg = 1 << oparg;
1601 }
1602
1603 if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
1604 return -EFAULT;
1605
1606 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1607 if (ret)
1608 return ret;
1609
1610 switch (cmp) {
1611 case FUTEX_OP_CMP_EQ:
1612 return oldval == cmparg;
1613 case FUTEX_OP_CMP_NE:
1614 return oldval != cmparg;
1615 case FUTEX_OP_CMP_LT:
1616 return oldval < cmparg;
1617 case FUTEX_OP_CMP_GE:
1618 return oldval >= cmparg;
1619 case FUTEX_OP_CMP_LE:
1620 return oldval <= cmparg;
1621 case FUTEX_OP_CMP_GT:
1622 return oldval > cmparg;
1623 default:
1624 return -ENOSYS;
1625 }
1626}
1627
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628/*
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001629 * Wake up all waiters hashed on the physical page that is mapped
1630 * to this virtual address:
1631 */
Ingo Molnare2970f22006-06-27 02:54:47 -07001632static int
Darren Hartb41277d2010-11-08 13:10:09 -08001633futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
Ingo Molnare2970f22006-06-27 02:54:47 -07001634 int nr_wake, int nr_wake2, int op)
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001635{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001636 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Ingo Molnare2970f22006-06-27 02:54:47 -07001637 struct futex_hash_bucket *hb1, *hb2;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001638 struct futex_q *this, *next;
Darren Harte4dc5b72009-03-12 00:56:13 -07001639 int ret, op_ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001640 DEFINE_WAKE_Q(wake_q);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001641
Darren Harte4dc5b72009-03-12 00:56:13 -07001642retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001643 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001644 if (unlikely(ret != 0))
1645 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001646 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001647 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001648 goto out_put_key1;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001649
Ingo Molnare2970f22006-06-27 02:54:47 -07001650 hb1 = hash_futex(&key1);
1651 hb2 = hash_futex(&key2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001652
Darren Harte4dc5b72009-03-12 00:56:13 -07001653retry_private:
Thomas Gleixnereaaea802009-10-04 09:34:17 +02001654 double_lock_hb(hb1, hb2);
Ingo Molnare2970f22006-06-27 02:54:47 -07001655 op_ret = futex_atomic_op_inuser(op, uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001656 if (unlikely(op_ret < 0)) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001657
Darren Hart5eb3dc62009-03-12 00:55:52 -07001658 double_unlock_hb(hb1, hb2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001659
David Howells7ee1dd32006-01-06 00:11:44 -08001660#ifndef CONFIG_MMU
Ingo Molnare2970f22006-06-27 02:54:47 -07001661 /*
1662 * we don't get EFAULT from MMU faults if we don't have an MMU,
1663 * but we might get them from range checking
1664 */
David Howells7ee1dd32006-01-06 00:11:44 -08001665 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001666 goto out_put_keys;
David Howells7ee1dd32006-01-06 00:11:44 -08001667#endif
1668
David Gibson796f8d92005-11-07 00:59:33 -08001669 if (unlikely(op_ret != -EFAULT)) {
1670 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001671 goto out_put_keys;
David Gibson796f8d92005-11-07 00:59:33 -08001672 }
1673
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001674 ret = fault_in_user_writeable(uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001675 if (ret)
Darren Hartde87fcc2009-03-12 00:55:46 -07001676 goto out_put_keys;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001677
Darren Hartb41277d2010-11-08 13:10:09 -08001678 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001679 goto retry_private;
1680
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001681 put_futex_key(&key2);
1682 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001683 goto retry;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001684 }
1685
Jason Low0d00c7b2014-01-12 15:31:22 -08001686 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001687 if (match_futex (&this->key, &key1)) {
Darren Hartaa109902012-11-26 16:29:56 -08001688 if (this->pi_state || this->rt_waiter) {
1689 ret = -EINVAL;
1690 goto out_unlock;
1691 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001692 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001693 if (++ret >= nr_wake)
1694 break;
1695 }
1696 }
1697
1698 if (op_ret > 0) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001699 op_ret = 0;
Jason Low0d00c7b2014-01-12 15:31:22 -08001700 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001701 if (match_futex (&this->key, &key2)) {
Darren Hartaa109902012-11-26 16:29:56 -08001702 if (this->pi_state || this->rt_waiter) {
1703 ret = -EINVAL;
1704 goto out_unlock;
1705 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001706 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001707 if (++op_ret >= nr_wake2)
1708 break;
1709 }
1710 }
1711 ret += op_ret;
1712 }
1713
Darren Hartaa109902012-11-26 16:29:56 -08001714out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07001715 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001716 wake_up_q(&wake_q);
Darren Hart42d35d42008-12-29 15:49:53 -08001717out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001718 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08001719out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001720 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08001721out:
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001722 return ret;
1723}
1724
Darren Hart9121e472009-04-03 13:40:31 -07001725/**
1726 * requeue_futex() - Requeue a futex_q from one hb to another
1727 * @q: the futex_q to requeue
1728 * @hb1: the source hash_bucket
1729 * @hb2: the target hash_bucket
1730 * @key2: the new key for the requeued futex_q
1731 */
1732static inline
1733void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1734 struct futex_hash_bucket *hb2, union futex_key *key2)
1735{
1736
1737 /*
1738 * If key1 and key2 hash to the same bucket, no need to
1739 * requeue.
1740 */
1741 if (likely(&hb1->chain != &hb2->chain)) {
1742 plist_del(&q->list, &hb1->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001743 hb_waiters_dec(hb1);
Linus Torvalds11d46162014-03-20 22:11:17 -07001744 hb_waiters_inc(hb2);
Davidlohr Buesofe1bce92016-04-20 20:09:24 -07001745 plist_add(&q->list, &hb2->chain);
Darren Hart9121e472009-04-03 13:40:31 -07001746 q->lock_ptr = &hb2->lock;
Darren Hart9121e472009-04-03 13:40:31 -07001747 }
1748 get_futex_key_refs(key2);
1749 q->key = *key2;
1750}
1751
Darren Hart52400ba2009-04-03 13:40:49 -07001752/**
1753 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
Darren Hartd96ee562009-09-21 22:30:22 -07001754 * @q: the futex_q
1755 * @key: the key of the requeue target futex
1756 * @hb: the hash_bucket of the requeue target futex
Darren Hart52400ba2009-04-03 13:40:49 -07001757 *
1758 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1759 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1760 * to the requeue target futex so the waiter can detect the wakeup on the right
1761 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
Darren Hartbeda2c72009-08-09 15:34:39 -07001762 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1763 * to protect access to the pi_state to fixup the owner later. Must be called
1764 * with both q->lock_ptr and hb->lock held.
Darren Hart52400ba2009-04-03 13:40:49 -07001765 */
1766static inline
Darren Hartbeda2c72009-08-09 15:34:39 -07001767void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1768 struct futex_hash_bucket *hb)
Darren Hart52400ba2009-04-03 13:40:49 -07001769{
Darren Hart52400ba2009-04-03 13:40:49 -07001770 get_futex_key_refs(key);
1771 q->key = *key;
1772
Lai Jiangshan2e129782010-12-22 14:18:50 +08001773 __unqueue_futex(q);
Darren Hart52400ba2009-04-03 13:40:49 -07001774
1775 WARN_ON(!q->rt_waiter);
1776 q->rt_waiter = NULL;
1777
Darren Hartbeda2c72009-08-09 15:34:39 -07001778 q->lock_ptr = &hb->lock;
Darren Hartbeda2c72009-08-09 15:34:39 -07001779
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001780 wake_up_state(q->task, TASK_NORMAL);
Darren Hart52400ba2009-04-03 13:40:49 -07001781}
1782
1783/**
1784 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
Darren Hartbab5bc92009-04-07 23:23:50 -07001785 * @pifutex: the user address of the to futex
1786 * @hb1: the from futex hash bucket, must be locked by the caller
1787 * @hb2: the to futex hash bucket, must be locked by the caller
1788 * @key1: the from futex key
1789 * @key2: the to futex key
1790 * @ps: address to store the pi_state pointer
1791 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart52400ba2009-04-03 13:40:49 -07001792 *
1793 * Try and get the lock on behalf of the top waiter if we can do it atomically.
Darren Hartbab5bc92009-04-07 23:23:50 -07001794 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1795 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1796 * hb1 and hb2 must be held by the caller.
Darren Hart52400ba2009-04-03 13:40:49 -07001797 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001798 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03001799 * - 0 - failed to acquire the lock atomically;
1800 * - >0 - acquired the lock, return value is vpid of the top_waiter
1801 * - <0 - error
Darren Hart52400ba2009-04-03 13:40:49 -07001802 */
1803static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1804 struct futex_hash_bucket *hb1,
1805 struct futex_hash_bucket *hb2,
1806 union futex_key *key1, union futex_key *key2,
Darren Hartbab5bc92009-04-07 23:23:50 -07001807 struct futex_pi_state **ps, int set_waiters)
Darren Hart52400ba2009-04-03 13:40:49 -07001808{
Darren Hartbab5bc92009-04-07 23:23:50 -07001809 struct futex_q *top_waiter = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07001810 u32 curval;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001811 int ret, vpid;
Darren Hart52400ba2009-04-03 13:40:49 -07001812
1813 if (get_futex_value_locked(&curval, pifutex))
1814 return -EFAULT;
1815
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001816 if (unlikely(should_fail_futex(true)))
1817 return -EFAULT;
1818
Darren Hartbab5bc92009-04-07 23:23:50 -07001819 /*
1820 * Find the top_waiter and determine if there are additional waiters.
1821 * If the caller intends to requeue more than 1 waiter to pifutex,
1822 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1823 * as we have means to handle the possible fault. If not, don't set
1824 * the bit unecessarily as it will force the subsequent unlock to enter
1825 * the kernel.
1826 */
Darren Hart52400ba2009-04-03 13:40:49 -07001827 top_waiter = futex_top_waiter(hb1, key1);
1828
1829 /* There are no waiters, nothing for us to do. */
1830 if (!top_waiter)
1831 return 0;
1832
Darren Hart84bc4af2009-08-13 17:36:53 -07001833 /* Ensure we requeue to the expected futex. */
1834 if (!match_futex(top_waiter->requeue_pi_key, key2))
1835 return -EINVAL;
1836
Darren Hart52400ba2009-04-03 13:40:49 -07001837 /*
Darren Hartbab5bc92009-04-07 23:23:50 -07001838 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1839 * the contended case or if set_waiters is 1. The pi_state is returned
1840 * in ps in contended cases.
Darren Hart52400ba2009-04-03 13:40:49 -07001841 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001842 vpid = task_pid_vnr(top_waiter->task);
Darren Hartbab5bc92009-04-07 23:23:50 -07001843 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1844 set_waiters);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001845 if (ret == 1) {
Darren Hartbeda2c72009-08-09 15:34:39 -07001846 requeue_pi_wake_futex(top_waiter, key2, hb2);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001847 return vpid;
1848 }
Darren Hart52400ba2009-04-03 13:40:49 -07001849 return ret;
1850}
1851
1852/**
1853 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
Randy Dunlapfb62db22010-10-13 11:02:34 -07001854 * @uaddr1: source futex user address
Darren Hartb41277d2010-11-08 13:10:09 -08001855 * @flags: futex flags (FLAGS_SHARED, etc.)
Randy Dunlapfb62db22010-10-13 11:02:34 -07001856 * @uaddr2: target futex user address
1857 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1858 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1859 * @cmpval: @uaddr1 expected value (or %NULL)
1860 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
Darren Hartb41277d2010-11-08 13:10:09 -08001861 * pi futex (pi to pi requeue is not supported)
Darren Hart52400ba2009-04-03 13:40:49 -07001862 *
1863 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1864 * uaddr2 atomically on behalf of the top waiter.
1865 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001866 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03001867 * - >=0 - on success, the number of tasks requeued or woken;
1868 * - <0 - on error
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 */
Darren Hartb41277d2010-11-08 13:10:09 -08001870static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1871 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1872 u32 *cmpval, int requeue_pi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001874 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Darren Hart52400ba2009-04-03 13:40:49 -07001875 int drop_count = 0, task_count = 0, ret;
1876 struct futex_pi_state *pi_state = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07001877 struct futex_hash_bucket *hb1, *hb2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 struct futex_q *this, *next;
Waiman Long194a6b52016-11-17 11:46:38 -05001879 DEFINE_WAKE_Q(wake_q);
Darren Hart52400ba2009-04-03 13:40:49 -07001880
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04001881 /*
1882 * When PI not supported: return -ENOSYS if requeue_pi is true,
1883 * consequently the compiler knows requeue_pi is always false past
1884 * this point which will optimize away all the conditional code
1885 * further down.
1886 */
1887 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1888 return -ENOSYS;
1889
Darren Hart52400ba2009-04-03 13:40:49 -07001890 if (requeue_pi) {
1891 /*
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001892 * Requeue PI only works on two distinct uaddrs. This
1893 * check is only valid for private futexes. See below.
1894 */
1895 if (uaddr1 == uaddr2)
1896 return -EINVAL;
1897
1898 /*
Darren Hart52400ba2009-04-03 13:40:49 -07001899 * requeue_pi requires a pi_state, try to allocate it now
1900 * without any locks in case it fails.
1901 */
1902 if (refill_pi_state_cache())
1903 return -ENOMEM;
1904 /*
1905 * requeue_pi must wake as many tasks as it can, up to nr_wake
1906 * + nr_requeue, since it acquires the rt_mutex prior to
1907 * returning to userspace, so as to not leave the rt_mutex with
1908 * waiters and no owner. However, second and third wake-ups
1909 * cannot be predicted as they involve race conditions with the
1910 * first wake and a fault while looking up the pi_state. Both
1911 * pthread_cond_signal() and pthread_cond_broadcast() should
1912 * use nr_wake=1.
1913 */
1914 if (nr_wake != 1)
1915 return -EINVAL;
1916 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
Darren Hart42d35d42008-12-29 15:49:53 -08001918retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001919 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 if (unlikely(ret != 0))
1921 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001922 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1923 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001925 goto out_put_key1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001927 /*
1928 * The check above which compares uaddrs is not sufficient for
1929 * shared futexes. We need to compare the keys:
1930 */
1931 if (requeue_pi && match_futex(&key1, &key2)) {
1932 ret = -EINVAL;
1933 goto out_put_keys;
1934 }
1935
Ingo Molnare2970f22006-06-27 02:54:47 -07001936 hb1 = hash_futex(&key1);
1937 hb2 = hash_futex(&key2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
Darren Harte4dc5b72009-03-12 00:56:13 -07001939retry_private:
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001940 hb_waiters_inc(hb2);
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001941 double_lock_hb(hb1, hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942
Ingo Molnare2970f22006-06-27 02:54:47 -07001943 if (likely(cmpval != NULL)) {
1944 u32 curval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945
Ingo Molnare2970f22006-06-27 02:54:47 -07001946 ret = get_futex_value_locked(&curval, uaddr1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947
1948 if (unlikely(ret)) {
Darren Hart5eb3dc62009-03-12 00:55:52 -07001949 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001950 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951
Darren Harte4dc5b72009-03-12 00:56:13 -07001952 ret = get_user(curval, uaddr1);
1953 if (ret)
1954 goto out_put_keys;
1955
Darren Hartb41277d2010-11-08 13:10:09 -08001956 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001957 goto retry_private;
1958
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001959 put_futex_key(&key2);
1960 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001961 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 }
Ingo Molnare2970f22006-06-27 02:54:47 -07001963 if (curval != *cmpval) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 ret = -EAGAIN;
1965 goto out_unlock;
1966 }
1967 }
1968
Darren Hart52400ba2009-04-03 13:40:49 -07001969 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
Darren Hartbab5bc92009-04-07 23:23:50 -07001970 /*
1971 * Attempt to acquire uaddr2 and wake the top waiter. If we
1972 * intend to requeue waiters, force setting the FUTEX_WAITERS
1973 * bit. We force this here where we are able to easily handle
1974 * faults rather in the requeue loop below.
1975 */
Darren Hart52400ba2009-04-03 13:40:49 -07001976 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
Darren Hartbab5bc92009-04-07 23:23:50 -07001977 &key2, &pi_state, nr_requeue);
Darren Hart52400ba2009-04-03 13:40:49 -07001978
1979 /*
1980 * At this point the top_waiter has either taken uaddr2 or is
1981 * waiting on it. If the former, then the pi_state will not
1982 * exist yet, look it up one more time to ensure we have a
Thomas Gleixner866293e2014-05-12 20:45:34 +00001983 * reference to it. If the lock was taken, ret contains the
1984 * vpid of the top waiter task.
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001985 * If the lock was not taken, we have pi_state and an initial
1986 * refcount on it. In case of an error we have nothing.
Darren Hart52400ba2009-04-03 13:40:49 -07001987 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001988 if (ret > 0) {
Darren Hart52400ba2009-04-03 13:40:49 -07001989 WARN_ON(pi_state);
Darren Hart89061d32009-10-15 15:30:48 -07001990 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07001991 task_count++;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001992 /*
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001993 * If we acquired the lock, then the user space value
1994 * of uaddr2 should be vpid. It cannot be changed by
1995 * the top waiter as it is blocked on hb2 lock if it
1996 * tries to do so. If something fiddled with it behind
1997 * our back the pi state lookup might unearth it. So
1998 * we rather use the known value than rereading and
1999 * handing potential crap to lookup_pi_state.
2000 *
2001 * If that call succeeds then we have pi_state and an
2002 * initial refcount on it.
Thomas Gleixner866293e2014-05-12 20:45:34 +00002003 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01002004 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07002005 }
2006
2007 switch (ret) {
2008 case 0:
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002009 /* We hold a reference on the pi state. */
Darren Hart52400ba2009-04-03 13:40:49 -07002010 break;
Thomas Gleixner4959f2d2015-12-19 20:07:40 +00002011
2012 /* If the above failed, then pi_state is NULL */
Darren Hart52400ba2009-04-03 13:40:49 -07002013 case -EFAULT:
2014 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07002015 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002016 put_futex_key(&key2);
2017 put_futex_key(&key1);
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002018 ret = fault_in_user_writeable(uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07002019 if (!ret)
2020 goto retry;
2021 goto out;
2022 case -EAGAIN:
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00002023 /*
2024 * Two reasons for this:
2025 * - Owner is exiting and we just wait for the
2026 * exit to complete.
2027 * - The user space value changed.
2028 */
Darren Hart52400ba2009-04-03 13:40:49 -07002029 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07002030 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002031 put_futex_key(&key2);
2032 put_futex_key(&key1);
Darren Hart52400ba2009-04-03 13:40:49 -07002033 cond_resched();
2034 goto retry;
2035 default:
2036 goto out_unlock;
2037 }
2038 }
2039
Jason Low0d00c7b2014-01-12 15:31:22 -08002040 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Darren Hart52400ba2009-04-03 13:40:49 -07002041 if (task_count - nr_wake >= nr_requeue)
2042 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
Darren Hart52400ba2009-04-03 13:40:49 -07002044 if (!match_futex(&this->key, &key1))
2045 continue;
2046
Darren Hart392741e2009-08-07 15:20:48 -07002047 /*
2048 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2049 * be paired with each other and no other futex ops.
Darren Hartaa109902012-11-26 16:29:56 -08002050 *
2051 * We should never be requeueing a futex_q with a pi_state,
2052 * which is awaiting a futex_unlock_pi().
Darren Hart392741e2009-08-07 15:20:48 -07002053 */
2054 if ((requeue_pi && !this->rt_waiter) ||
Darren Hartaa109902012-11-26 16:29:56 -08002055 (!requeue_pi && this->rt_waiter) ||
2056 this->pi_state) {
Darren Hart392741e2009-08-07 15:20:48 -07002057 ret = -EINVAL;
2058 break;
2059 }
Darren Hart52400ba2009-04-03 13:40:49 -07002060
2061 /*
2062 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2063 * lock, we already woke the top_waiter. If not, it will be
2064 * woken by futex_unlock_pi().
2065 */
2066 if (++task_count <= nr_wake && !requeue_pi) {
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07002067 mark_wake_futex(&wake_q, this);
Darren Hart52400ba2009-04-03 13:40:49 -07002068 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 }
Darren Hart52400ba2009-04-03 13:40:49 -07002070
Darren Hart84bc4af2009-08-13 17:36:53 -07002071 /* Ensure we requeue to the expected futex for requeue_pi. */
2072 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2073 ret = -EINVAL;
2074 break;
2075 }
2076
Darren Hart52400ba2009-04-03 13:40:49 -07002077 /*
2078 * Requeue nr_requeue waiters and possibly one more in the case
2079 * of requeue_pi if we couldn't acquire the lock atomically.
2080 */
2081 if (requeue_pi) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002082 /*
2083 * Prepare the waiter to take the rt_mutex. Take a
2084 * refcount on the pi_state and store the pointer in
2085 * the futex_q object of the waiter.
2086 */
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01002087 get_pi_state(pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07002088 this->pi_state = pi_state;
2089 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2090 this->rt_waiter,
Thomas Gleixnerc051b212014-05-22 03:25:50 +00002091 this->task);
Darren Hart52400ba2009-04-03 13:40:49 -07002092 if (ret == 1) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002093 /*
2094 * We got the lock. We do neither drop the
2095 * refcount on pi_state nor clear
2096 * this->pi_state because the waiter needs the
2097 * pi_state for cleaning up the user space
2098 * value. It will drop the refcount after
2099 * doing so.
2100 */
Darren Hartbeda2c72009-08-09 15:34:39 -07002101 requeue_pi_wake_futex(this, &key2, hb2);
Darren Hart89061d32009-10-15 15:30:48 -07002102 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07002103 continue;
2104 } else if (ret) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002105 /*
2106 * rt_mutex_start_proxy_lock() detected a
2107 * potential deadlock when we tried to queue
2108 * that waiter. Drop the pi_state reference
2109 * which we took above and remove the pointer
2110 * to the state from the waiters futex_q
2111 * object.
2112 */
Darren Hart52400ba2009-04-03 13:40:49 -07002113 this->pi_state = NULL;
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002114 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002115 /*
2116 * We stop queueing more waiters and let user
2117 * space deal with the mess.
2118 */
2119 break;
Darren Hart52400ba2009-04-03 13:40:49 -07002120 }
2121 }
2122 requeue_futex(this, hb1, hb2, &key2);
2123 drop_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 }
2125
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002126 /*
2127 * We took an extra initial reference to the pi_state either
2128 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2129 * need to drop it here again.
2130 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002131 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002132
2133out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07002134 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07002135 wake_up_q(&wake_q);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07002136 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137
Darren Hartcd84a422009-04-02 14:19:38 -07002138 /*
2139 * drop_futex_key_refs() must be called outside the spinlocks. During
2140 * the requeue we moved futex_q's from the hash bucket at key1 to the
2141 * one at key2 and updated their key pointer. We no longer need to
2142 * hold the references to key1.
2143 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 while (--drop_count >= 0)
Rusty Russell9adef58b2007-05-08 00:26:42 -07002145 drop_futex_key_refs(&key1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146
Darren Hart42d35d42008-12-29 15:49:53 -08002147out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002148 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08002149out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002150 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08002151out:
Darren Hart52400ba2009-04-03 13:40:49 -07002152 return ret ? ret : task_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153}
2154
2155/* The key must be already stored in q->key. */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002156static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002157 __acquires(&hb->lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158{
Ingo Molnare2970f22006-06-27 02:54:47 -07002159 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160
Ingo Molnare2970f22006-06-27 02:54:47 -07002161 hb = hash_futex(&q->key);
Linus Torvalds11d46162014-03-20 22:11:17 -07002162
2163 /*
2164 * Increment the counter before taking the lock so that
2165 * a potential waker won't miss a to-be-slept task that is
2166 * waiting for the spinlock. This is safe as all queue_lock()
2167 * users end up calling queue_me(). Similarly, for housekeeping,
2168 * decrement the counter at queue_unlock() when some error has
2169 * occurred and we don't end up adding the task to the list.
2170 */
2171 hb_waiters_inc(hb);
2172
Ingo Molnare2970f22006-06-27 02:54:47 -07002173 q->lock_ptr = &hb->lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -08002175 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
Ingo Molnare2970f22006-06-27 02:54:47 -07002176 return hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177}
2178
Darren Hartd40d65c2009-09-21 22:30:15 -07002179static inline void
Jason Low0d00c7b2014-01-12 15:31:22 -08002180queue_unlock(struct futex_hash_bucket *hb)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002181 __releases(&hb->lock)
Darren Hartd40d65c2009-09-21 22:30:15 -07002182{
2183 spin_unlock(&hb->lock);
Linus Torvalds11d46162014-03-20 22:11:17 -07002184 hb_waiters_dec(hb);
Darren Hartd40d65c2009-09-21 22:30:15 -07002185}
2186
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002187static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188{
Pierre Peifferec92d082007-05-09 02:35:00 -07002189 int prio;
2190
2191 /*
2192 * The priority used to register this element is
2193 * - either the real thread-priority for the real-time threads
2194 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2195 * - or MAX_RT_PRIO for non-RT threads.
2196 * Thus, all RT-threads are woken first in priority order, and
2197 * the others are woken last, in FIFO order.
2198 */
2199 prio = min(current->normal_prio, MAX_RT_PRIO);
2200
2201 plist_node_init(&q->list, prio);
Pierre Peifferec92d082007-05-09 02:35:00 -07002202 plist_add(&q->list, &hb->chain);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002203 q->task = current;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002204}
2205
2206/**
2207 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2208 * @q: The futex_q to enqueue
2209 * @hb: The destination hash bucket
2210 *
2211 * The hb->lock must be held by the caller, and is released here. A call to
2212 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2213 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2214 * or nothing if the unqueue is done as part of the wake process and the unqueue
2215 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2216 * an example).
2217 */
2218static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2219 __releases(&hb->lock)
2220{
2221 __queue_me(q, hb);
Ingo Molnare2970f22006-06-27 02:54:47 -07002222 spin_unlock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223}
2224
Darren Hartd40d65c2009-09-21 22:30:15 -07002225/**
2226 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2227 * @q: The futex_q to unqueue
2228 *
2229 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2230 * be paired with exactly one earlier call to queue_me().
2231 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002232 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002233 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2234 * - 0 - if the futex_q was already removed by the waking thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236static int unqueue_me(struct futex_q *q)
2237{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 spinlock_t *lock_ptr;
Ingo Molnare2970f22006-06-27 02:54:47 -07002239 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240
2241 /* In the common case we don't take the spinlock, which is nice. */
Darren Hart42d35d42008-12-29 15:49:53 -08002242retry:
Jianyu Zhan29b75eb2016-03-07 09:32:24 +08002243 /*
2244 * q->lock_ptr can change between this read and the following spin_lock.
2245 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2246 * optimizing lock_ptr out of the logic below.
2247 */
2248 lock_ptr = READ_ONCE(q->lock_ptr);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07002249 if (lock_ptr != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 spin_lock(lock_ptr);
2251 /*
2252 * q->lock_ptr can change between reading it and
2253 * spin_lock(), causing us to take the wrong lock. This
2254 * corrects the race condition.
2255 *
2256 * Reasoning goes like this: if we have the wrong lock,
2257 * q->lock_ptr must have changed (maybe several times)
2258 * between reading it and the spin_lock(). It can
2259 * change again after the spin_lock() but only if it was
2260 * already changed before the spin_lock(). It cannot,
2261 * however, change back to the original value. Therefore
2262 * we can detect whether we acquired the correct lock.
2263 */
2264 if (unlikely(lock_ptr != q->lock_ptr)) {
2265 spin_unlock(lock_ptr);
2266 goto retry;
2267 }
Lai Jiangshan2e129782010-12-22 14:18:50 +08002268 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002269
2270 BUG_ON(q->pi_state);
2271
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 spin_unlock(lock_ptr);
2273 ret = 1;
2274 }
2275
Rusty Russell9adef58b2007-05-08 00:26:42 -07002276 drop_futex_key_refs(&q->key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 return ret;
2278}
2279
Ingo Molnarc87e2832006-06-27 02:54:58 -07002280/*
2281 * PI futexes can not be requeued and must remove themself from the
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002282 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2283 * and dropped here.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002284 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002285static void unqueue_me_pi(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002286 __releases(q->lock_ptr)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002287{
Lai Jiangshan2e129782010-12-22 14:18:50 +08002288 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002289
2290 BUG_ON(!q->pi_state);
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002291 put_pi_state(q->pi_state);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002292 q->pi_state = NULL;
2293
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002294 spin_unlock(q->lock_ptr);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002295}
2296
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002297/*
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002298 * Fixup the pi_state owner with the new owner.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002299 *
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002300 * Must be called with hash bucket lock held and mm->sem held for non
2301 * private futexes.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002302 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002303static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002304 struct task_struct *newowner)
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002305{
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002306 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002307 struct futex_pi_state *pi_state = q->pi_state;
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03002308 u32 uval, uninitialized_var(curval), newval;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002309 struct task_struct *oldowner;
Darren Harte4dc5b72009-03-12 00:56:13 -07002310 int ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002311
Peter Zijlstra734009e2017-03-22 11:35:52 +01002312 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2313
2314 oldowner = pi_state->owner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002315 /* Owner died? */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002316 if (!pi_state->owner)
2317 newtid |= FUTEX_OWNER_DIED;
2318
2319 /*
2320 * We are here either because we stole the rtmutex from the
Lai Jiangshan81612392011-01-14 17:09:41 +08002321 * previous highest priority waiter or we are the highest priority
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002322 * waiter but have failed to get the rtmutex the first time.
2323 *
Lai Jiangshan81612392011-01-14 17:09:41 +08002324 * We have to replace the newowner TID in the user space variable.
2325 * This must be atomic as we have to preserve the owner died bit here.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002326 *
Darren Hartb2d09942009-03-12 00:55:37 -07002327 * Note: We write the user space value _before_ changing the pi_state
2328 * because we can fault here. Imagine swapped out pages or a fork
2329 * that marked all the anonymous memory readonly for cow.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002330 *
Peter Zijlstra734009e2017-03-22 11:35:52 +01002331 * Modifying pi_state _before_ the user space value would leave the
2332 * pi_state in an inconsistent state when we fault here, because we
2333 * need to drop the locks to handle the fault. This might be observed
2334 * in the PID check in lookup_pi_state.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002335 */
2336retry:
2337 if (get_futex_value_locked(&uval, uaddr))
2338 goto handle_fault;
2339
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002340 for (;;) {
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002341 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2342
Michel Lespinasse37a9d912011-03-10 18:48:51 -08002343 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002344 goto handle_fault;
2345 if (curval == uval)
2346 break;
2347 uval = curval;
2348 }
2349
2350 /*
2351 * We fixed up user space. Now we need to fix the pi_state
2352 * itself.
2353 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002354 if (pi_state->owner != NULL) {
Peter Zijlstra734009e2017-03-22 11:35:52 +01002355 raw_spin_lock(&pi_state->owner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002356 WARN_ON(list_empty(&pi_state->list));
2357 list_del_init(&pi_state->list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002358 raw_spin_unlock(&pi_state->owner->pi_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002359 }
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002360
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002361 pi_state->owner = newowner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002362
Peter Zijlstra734009e2017-03-22 11:35:52 +01002363 raw_spin_lock(&newowner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002364 WARN_ON(!list_empty(&pi_state->list));
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002365 list_add(&pi_state->list, &newowner->pi_state_list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002366 raw_spin_unlock(&newowner->pi_lock);
2367 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2368
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002369 return 0;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002370
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002371 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01002372 * To handle the page fault we need to drop the locks here. That gives
2373 * the other task (either the highest priority waiter itself or the
2374 * task which stole the rtmutex) the chance to try the fixup of the
2375 * pi_state. So once we are back from handling the fault we need to
2376 * check the pi_state after reacquiring the locks and before trying to
2377 * do another fixup. When the fixup has been done already we simply
2378 * return.
2379 *
2380 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2381 * drop hb->lock since the caller owns the hb -> futex_q relation.
2382 * Dropping the pi_mutex->wait_lock requires the state revalidate.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002383 */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002384handle_fault:
Peter Zijlstra734009e2017-03-22 11:35:52 +01002385 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002386 spin_unlock(q->lock_ptr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002387
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002388 ret = fault_in_user_writeable(uaddr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002389
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002390 spin_lock(q->lock_ptr);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002391 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002392
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002393 /*
2394 * Check if someone else fixed it for us:
2395 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01002396 if (pi_state->owner != oldowner) {
2397 ret = 0;
2398 goto out_unlock;
2399 }
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002400
2401 if (ret)
Peter Zijlstra734009e2017-03-22 11:35:52 +01002402 goto out_unlock;
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002403
2404 goto retry;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002405
2406out_unlock:
2407 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2408 return ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002409}
2410
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002411static long futex_wait_restart(struct restart_block *restart);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07002412
Darren Hartca5f9522009-04-03 13:39:33 -07002413/**
Darren Hartdd973992009-04-03 13:40:02 -07002414 * fixup_owner() - Post lock pi_state and corner case management
2415 * @uaddr: user address of the futex
Darren Hartdd973992009-04-03 13:40:02 -07002416 * @q: futex_q (contains pi_state and access to the rt_mutex)
2417 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2418 *
2419 * After attempting to lock an rt_mutex, this function is called to cleanup
2420 * the pi_state owner as well as handle race conditions that may allow us to
2421 * acquire the lock. Must be called with the hb lock held.
2422 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002423 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002424 * - 1 - success, lock taken;
2425 * - 0 - success, lock not taken;
2426 * - <0 - on error (-EFAULT)
Darren Hartdd973992009-04-03 13:40:02 -07002427 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002428static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
Darren Hartdd973992009-04-03 13:40:02 -07002429{
Darren Hartdd973992009-04-03 13:40:02 -07002430 int ret = 0;
2431
2432 if (locked) {
2433 /*
2434 * Got the lock. We might not be the anticipated owner if we
2435 * did a lock-steal - fix up the PI-state in that case:
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002436 *
2437 * We can safely read pi_state->owner without holding wait_lock
2438 * because we now own the rt_mutex, only the owner will attempt
2439 * to change it.
Darren Hartdd973992009-04-03 13:40:02 -07002440 */
2441 if (q->pi_state->owner != current)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002442 ret = fixup_pi_state_owner(uaddr, q, current);
Darren Hartdd973992009-04-03 13:40:02 -07002443 goto out;
2444 }
2445
2446 /*
Darren Hartdd973992009-04-03 13:40:02 -07002447 * Paranoia check. If we did not take the lock, then we should not be
Lai Jiangshan81612392011-01-14 17:09:41 +08002448 * the owner of the rt_mutex.
Darren Hartdd973992009-04-03 13:40:02 -07002449 */
Peter Zijlstra73d786b2017-03-22 11:35:54 +01002450 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
Darren Hartdd973992009-04-03 13:40:02 -07002451 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2452 "pi-state %p\n", ret,
2453 q->pi_state->pi_mutex.owner,
2454 q->pi_state->owner);
Peter Zijlstra73d786b2017-03-22 11:35:54 +01002455 }
Darren Hartdd973992009-04-03 13:40:02 -07002456
2457out:
2458 return ret ? ret : locked;
2459}
2460
2461/**
Darren Hartca5f9522009-04-03 13:39:33 -07002462 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2463 * @hb: the futex hash bucket, must be locked by the caller
2464 * @q: the futex_q to queue up on
2465 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
Darren Hartca5f9522009-04-03 13:39:33 -07002466 */
2467static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002468 struct hrtimer_sleeper *timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002469{
Darren Hart9beba3c2009-09-24 11:54:47 -07002470 /*
2471 * The task state is guaranteed to be set before another task can
Peter Zijlstrab92b8b32015-05-12 10:51:55 +02002472 * wake it. set_current_state() is implemented using smp_store_mb() and
Darren Hart9beba3c2009-09-24 11:54:47 -07002473 * queue_me() calls spin_unlock() upon completion, both serializing
2474 * access to the hash list and forcing another memory barrier.
2475 */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002476 set_current_state(TASK_INTERRUPTIBLE);
Darren Hart0729e192009-09-21 22:30:38 -07002477 queue_me(q, hb);
Darren Hartca5f9522009-04-03 13:39:33 -07002478
2479 /* Arm the timer */
Thomas Gleixner2e4b0d32015-04-14 21:09:13 +00002480 if (timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002481 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002482
2483 /*
Darren Hart0729e192009-09-21 22:30:38 -07002484 * If we have been removed from the hash list, then another task
2485 * has tried to wake us, and we can skip the call to schedule().
Darren Hartca5f9522009-04-03 13:39:33 -07002486 */
2487 if (likely(!plist_node_empty(&q->list))) {
2488 /*
2489 * If the timer has already expired, current will already be
2490 * flagged for rescheduling. Only call schedule if there
2491 * is no timeout, or if it has yet to expire.
2492 */
2493 if (!timeout || timeout->task)
Colin Cross88c80042013-05-01 18:35:05 -07002494 freezable_schedule();
Darren Hartca5f9522009-04-03 13:39:33 -07002495 }
2496 __set_current_state(TASK_RUNNING);
2497}
2498
Darren Hartf8010732009-04-03 13:40:40 -07002499/**
2500 * futex_wait_setup() - Prepare to wait on a futex
2501 * @uaddr: the futex userspace address
2502 * @val: the expected value
Darren Hartb41277d2010-11-08 13:10:09 -08002503 * @flags: futex flags (FLAGS_SHARED, etc.)
Darren Hartf8010732009-04-03 13:40:40 -07002504 * @q: the associated futex_q
2505 * @hb: storage for hash_bucket pointer to be returned to caller
2506 *
2507 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2508 * compare it with the expected value. Handle atomic faults internally.
2509 * Return with the hb lock held and a q.key reference on success, and unlocked
2510 * with no q.key reference on failure.
2511 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002512 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002513 * - 0 - uaddr contains val and hb has been locked;
2514 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
Darren Hartf8010732009-04-03 13:40:40 -07002515 */
Darren Hartb41277d2010-11-08 13:10:09 -08002516static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
Darren Hartf8010732009-04-03 13:40:40 -07002517 struct futex_q *q, struct futex_hash_bucket **hb)
2518{
2519 u32 uval;
2520 int ret;
2521
2522 /*
2523 * Access the page AFTER the hash-bucket is locked.
2524 * Order is important:
2525 *
2526 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2527 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2528 *
2529 * The basic logical guarantee of a futex is that it blocks ONLY
2530 * if cond(var) is known to be true at the time of blocking, for
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002531 * any cond. If we locked the hash-bucket after testing *uaddr, that
2532 * would open a race condition where we could block indefinitely with
Darren Hartf8010732009-04-03 13:40:40 -07002533 * cond(var) false, which would violate the guarantee.
2534 *
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002535 * On the other hand, we insert q and release the hash-bucket only
2536 * after testing *uaddr. This guarantees that futex_wait() will NOT
2537 * absorb a wakeup if *uaddr does not match the desired values
2538 * while the syscall executes.
Darren Hartf8010732009-04-03 13:40:40 -07002539 */
2540retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002541 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
Darren Hartf8010732009-04-03 13:40:40 -07002542 if (unlikely(ret != 0))
Darren Harta5a2a0c2009-04-10 09:50:05 -07002543 return ret;
Darren Hartf8010732009-04-03 13:40:40 -07002544
2545retry_private:
2546 *hb = queue_lock(q);
2547
2548 ret = get_futex_value_locked(&uval, uaddr);
2549
2550 if (ret) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002551 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002552
2553 ret = get_user(uval, uaddr);
2554 if (ret)
2555 goto out;
2556
Darren Hartb41277d2010-11-08 13:10:09 -08002557 if (!(flags & FLAGS_SHARED))
Darren Hartf8010732009-04-03 13:40:40 -07002558 goto retry_private;
2559
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002560 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002561 goto retry;
2562 }
2563
2564 if (uval != val) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002565 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002566 ret = -EWOULDBLOCK;
2567 }
2568
2569out:
2570 if (ret)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002571 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002572 return ret;
2573}
2574
Darren Hartb41277d2010-11-08 13:10:09 -08002575static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2576 ktime_t *abs_time, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577{
Darren Hartca5f9522009-04-03 13:39:33 -07002578 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002579 struct restart_block *restart;
Ingo Molnare2970f22006-06-27 02:54:47 -07002580 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002581 struct futex_q q = futex_q_init;
Ingo Molnare2970f22006-06-27 02:54:47 -07002582 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583
Thomas Gleixnercd6899852008-02-01 17:45:14 +01002584 if (!bitset)
2585 return -EINVAL;
Thomas Gleixnercd6899852008-02-01 17:45:14 +01002586 q.bitset = bitset;
Darren Hartca5f9522009-04-03 13:39:33 -07002587
2588 if (abs_time) {
2589 to = &timeout;
2590
Darren Hartb41277d2010-11-08 13:10:09 -08002591 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2592 CLOCK_REALTIME : CLOCK_MONOTONIC,
2593 HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002594 hrtimer_init_sleeper(to, current);
2595 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2596 current->timer_slack_ns);
2597 }
2598
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002599retry:
Darren Hart7ada8762010-10-17 08:35:04 -07002600 /*
2601 * Prepare to wait on uaddr. On success, holds hb lock and increments
2602 * q.key refs.
2603 */
Darren Hartb41277d2010-11-08 13:10:09 -08002604 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Darren Hartf8010732009-04-03 13:40:40 -07002605 if (ret)
Darren Hart42d35d42008-12-29 15:49:53 -08002606 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607
Darren Hartca5f9522009-04-03 13:39:33 -07002608 /* queue_me and wait for wakeup, timeout, or a signal. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002609 futex_wait_queue_me(hb, &q, to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610
2611 /* If we were woken (and unqueued), we succeeded, whatever. */
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002612 ret = 0;
Darren Hart7ada8762010-10-17 08:35:04 -07002613 /* unqueue_me() drops q.key ref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 if (!unqueue_me(&q))
Darren Hart7ada8762010-10-17 08:35:04 -07002615 goto out;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002616 ret = -ETIMEDOUT;
Darren Hartca5f9522009-04-03 13:39:33 -07002617 if (to && !to->task)
Darren Hart7ada8762010-10-17 08:35:04 -07002618 goto out;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002619
Ingo Molnare2970f22006-06-27 02:54:47 -07002620 /*
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002621 * We expect signal_pending(current), but we might be the
2622 * victim of a spurious wakeup as well.
Ingo Molnare2970f22006-06-27 02:54:47 -07002623 */
Darren Hart7ada8762010-10-17 08:35:04 -07002624 if (!signal_pending(current))
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002625 goto retry;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002626
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002627 ret = -ERESTARTSYS;
Pierre Peifferc19384b2007-05-09 02:35:02 -07002628 if (!abs_time)
Darren Hart7ada8762010-10-17 08:35:04 -07002629 goto out;
Steven Rostedtce6bd422007-12-05 15:46:09 +01002630
Andy Lutomirskif56141e2015-02-12 15:01:14 -08002631 restart = &current->restart_block;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002632 restart->fn = futex_wait_restart;
Namhyung Kima3c74c52010-09-14 21:43:47 +09002633 restart->futex.uaddr = uaddr;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002634 restart->futex.val = val;
Thomas Gleixner2456e852016-12-25 11:38:40 +01002635 restart->futex.time = *abs_time;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002636 restart->futex.bitset = bitset;
Darren Hart0cd9c642011-04-14 15:41:57 -07002637 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002638
2639 ret = -ERESTART_RESTARTBLOCK;
2640
Darren Hart42d35d42008-12-29 15:49:53 -08002641out:
Darren Hartca5f9522009-04-03 13:39:33 -07002642 if (to) {
2643 hrtimer_cancel(&to->timer);
2644 destroy_hrtimer_on_stack(&to->timer);
2645 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002646 return ret;
2647}
2648
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002649
2650static long futex_wait_restart(struct restart_block *restart)
2651{
Namhyung Kima3c74c52010-09-14 21:43:47 +09002652 u32 __user *uaddr = restart->futex.uaddr;
Darren Harta72188d2009-04-03 13:40:22 -07002653 ktime_t t, *tp = NULL;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002654
Darren Harta72188d2009-04-03 13:40:22 -07002655 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
Thomas Gleixner2456e852016-12-25 11:38:40 +01002656 t = restart->futex.time;
Darren Harta72188d2009-04-03 13:40:22 -07002657 tp = &t;
2658 }
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002659 restart->fn = do_no_restart_syscall;
Darren Hartb41277d2010-11-08 13:10:09 -08002660
2661 return (long)futex_wait(uaddr, restart->futex.flags,
2662 restart->futex.val, tp, restart->futex.bitset);
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002663}
2664
2665
Ingo Molnarc87e2832006-06-27 02:54:58 -07002666/*
2667 * Userspace tried a 0 -> TID atomic transition of the futex value
2668 * and failed. The kernel side here does the whole locking operation:
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002669 * if there are waiters then it will block as a consequence of relying
2670 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2671 * a 0 value of the futex too.).
2672 *
2673 * Also serves as futex trylock_pi()'ing, and due semantics.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002674 */
Michael Kerrisk996636d2015-01-16 20:28:06 +01002675static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
Darren Hartb41277d2010-11-08 13:10:09 -08002676 ktime_t *time, int trylock)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002677{
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002678 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002679 struct futex_pi_state *pi_state = NULL;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002680 struct rt_mutex_waiter rt_waiter;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002681 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002682 struct futex_q q = futex_q_init;
Darren Hartdd973992009-04-03 13:40:02 -07002683 int res, ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002684
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04002685 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2686 return -ENOSYS;
2687
Ingo Molnarc87e2832006-06-27 02:54:58 -07002688 if (refill_pi_state_cache())
2689 return -ENOMEM;
2690
Pierre Peifferc19384b2007-05-09 02:35:02 -07002691 if (time) {
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002692 to = &timeout;
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002693 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2694 HRTIMER_MODE_ABS);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002695 hrtimer_init_sleeper(to, current);
Arjan van de Vencc584b22008-09-01 15:02:30 -07002696 hrtimer_set_expires(&to->timer, *time);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002697 }
2698
Darren Hart42d35d42008-12-29 15:49:53 -08002699retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002700 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002701 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08002702 goto out;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002703
Darren Harte4dc5b72009-03-12 00:56:13 -07002704retry_private:
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002705 hb = queue_lock(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002706
Darren Hartbab5bc92009-04-07 23:23:50 -07002707 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002708 if (unlikely(ret)) {
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002709 /*
2710 * Atomic work succeeded and we got the lock,
2711 * or failed. Either way, we do _not_ block.
2712 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002713 switch (ret) {
Darren Hart1a520842009-04-03 13:39:52 -07002714 case 1:
2715 /* We got the lock. */
2716 ret = 0;
2717 goto out_unlock_put_key;
2718 case -EFAULT:
2719 goto uaddr_faulted;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002720 case -EAGAIN:
2721 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00002722 * Two reasons for this:
2723 * - Task is exiting and we just wait for the
2724 * exit to complete.
2725 * - The user space value changed.
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002726 */
Jason Low0d00c7b2014-01-12 15:31:22 -08002727 queue_unlock(hb);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002728 put_futex_key(&q.key);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002729 cond_resched();
2730 goto retry;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002731 default:
Darren Hart42d35d42008-12-29 15:49:53 -08002732 goto out_unlock_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002733 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002734 }
2735
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002736 WARN_ON(!q.pi_state);
2737
Ingo Molnarc87e2832006-06-27 02:54:58 -07002738 /*
2739 * Only actually queue now that the atomic ops are done:
2740 */
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002741 __queue_me(&q, hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002742
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002743 if (trylock) {
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01002744 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002745 /* Fixup the trylock return value: */
2746 ret = ret ? 0 : -EWOULDBLOCK;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002747 goto no_block;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002748 }
2749
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002750 rt_mutex_init_waiter(&rt_waiter);
Peter Zijlstra56222b22017-03-22 11:36:00 +01002751
2752 /*
2753 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2754 * hold it while doing rt_mutex_start_proxy(), because then it will
2755 * include hb->lock in the blocking chain, even through we'll not in
2756 * fact hold it while blocking. This will lead it to report -EDEADLK
2757 * and BUG when futex_unlock_pi() interleaves with this.
2758 *
2759 * Therefore acquire wait_lock while holding hb->lock, but drop the
2760 * latter before calling rt_mutex_start_proxy_lock(). This still fully
2761 * serializes against futex_unlock_pi() as that does the exact same
2762 * lock handoff sequence.
2763 */
2764 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2765 spin_unlock(q.lock_ptr);
2766 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2767 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2768
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002769 if (ret) {
2770 if (ret == 1)
2771 ret = 0;
2772
Peter Zijlstra56222b22017-03-22 11:36:00 +01002773 spin_lock(q.lock_ptr);
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002774 goto no_block;
2775 }
2776
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002777
2778 if (unlikely(to))
2779 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2780
2781 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2782
Vernon Mauerya99e4e42006-07-01 04:35:42 -07002783 spin_lock(q.lock_ptr);
Darren Hartdd973992009-04-03 13:40:02 -07002784 /*
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002785 * If we failed to acquire the lock (signal/timeout), we must
2786 * first acquire the hb->lock before removing the lock from the
2787 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2788 * wait lists consistent.
Peter Zijlstra56222b22017-03-22 11:36:00 +01002789 *
2790 * In particular; it is important that futex_unlock_pi() can not
2791 * observe this inconsistency.
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002792 */
2793 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2794 ret = 0;
2795
2796no_block:
2797 /*
Darren Hartdd973992009-04-03 13:40:02 -07002798 * Fixup the pi_state owner and possibly acquire the lock if we
2799 * haven't already.
2800 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002801 res = fixup_owner(uaddr, &q, !ret);
Darren Hartdd973992009-04-03 13:40:02 -07002802 /*
2803 * If fixup_owner() returned an error, proprogate that. If it acquired
2804 * the lock, clear our -ETIMEDOUT or -EINTR.
2805 */
2806 if (res)
2807 ret = (res < 0) ? res : 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002808
Darren Harte8f63862009-03-12 00:56:06 -07002809 /*
Darren Hartdd973992009-04-03 13:40:02 -07002810 * If fixup_owner() faulted and was unable to handle the fault, unlock
2811 * it and return the fault to userspace.
Darren Harte8f63862009-03-12 00:56:06 -07002812 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002813 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2814 pi_state = q.pi_state;
2815 get_pi_state(pi_state);
2816 }
Darren Harte8f63862009-03-12 00:56:06 -07002817
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002818 /* Unqueue and drop the lock */
2819 unqueue_me_pi(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002820
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002821 if (pi_state) {
2822 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2823 put_pi_state(pi_state);
2824 }
2825
Mikael Pettersson5ecb01c2010-01-23 22:36:29 +01002826 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002827
Darren Hart42d35d42008-12-29 15:49:53 -08002828out_unlock_put_key:
Jason Low0d00c7b2014-01-12 15:31:22 -08002829 queue_unlock(hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002830
Darren Hart42d35d42008-12-29 15:49:53 -08002831out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002832 put_futex_key(&q.key);
Darren Hart42d35d42008-12-29 15:49:53 -08002833out:
Thomas Gleixner97181f92017-04-10 18:03:36 +02002834 if (to) {
2835 hrtimer_cancel(&to->timer);
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002836 destroy_hrtimer_on_stack(&to->timer);
Thomas Gleixner97181f92017-04-10 18:03:36 +02002837 }
Darren Hartdd973992009-04-03 13:40:02 -07002838 return ret != -EINTR ? ret : -ERESTARTNOINTR;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002839
Darren Hart42d35d42008-12-29 15:49:53 -08002840uaddr_faulted:
Jason Low0d00c7b2014-01-12 15:31:22 -08002841 queue_unlock(hb);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002842
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002843 ret = fault_in_user_writeable(uaddr);
Darren Harte4dc5b72009-03-12 00:56:13 -07002844 if (ret)
2845 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002846
Darren Hartb41277d2010-11-08 13:10:09 -08002847 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07002848 goto retry_private;
2849
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002850 put_futex_key(&q.key);
Darren Harte4dc5b72009-03-12 00:56:13 -07002851 goto retry;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002852}
2853
2854/*
Ingo Molnarc87e2832006-06-27 02:54:58 -07002855 * Userspace attempted a TID -> 0 atomic transition, and failed.
2856 * This is the in-kernel slowpath: we look up the PI state (if any),
2857 * and do the rt-mutex unlock.
2858 */
Darren Hartb41277d2010-11-08 13:10:09 -08002859static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002860{
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002861 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
Peter Zijlstra38d47c12008-09-26 19:32:20 +02002862 union futex_key key = FUTEX_KEY_INIT;
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002863 struct futex_hash_bucket *hb;
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002864 struct futex_q *top_waiter;
Darren Harte4dc5b72009-03-12 00:56:13 -07002865 int ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002866
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04002867 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2868 return -ENOSYS;
2869
Ingo Molnarc87e2832006-06-27 02:54:58 -07002870retry:
2871 if (get_user(uval, uaddr))
2872 return -EFAULT;
2873 /*
2874 * We release only a lock we actually own:
2875 */
Thomas Gleixnerc0c9ed12011-03-11 11:51:22 +01002876 if ((uval & FUTEX_TID_MASK) != vpid)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002877 return -EPERM;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002878
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002879 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002880 if (ret)
2881 return ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002882
2883 hb = hash_futex(&key);
2884 spin_lock(&hb->lock);
2885
Ingo Molnarc87e2832006-06-27 02:54:58 -07002886 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002887 * Check waiters first. We do not trust user space values at
2888 * all and we at least want to know if user space fiddled
2889 * with the futex value instead of blindly unlocking.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002890 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002891 top_waiter = futex_top_waiter(hb, &key);
2892 if (top_waiter) {
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002893 struct futex_pi_state *pi_state = top_waiter->pi_state;
2894
2895 ret = -EINVAL;
2896 if (!pi_state)
2897 goto out_unlock;
2898
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002899 /*
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002900 * If current does not own the pi_state then the futex is
2901 * inconsistent and user space fiddled with the futex value.
2902 */
2903 if (pi_state->owner != current)
2904 goto out_unlock;
2905
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002906 get_pi_state(pi_state);
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01002907 /*
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01002908 * By taking wait_lock while still holding hb->lock, we ensure
2909 * there is no point where we hold neither; and therefore
2910 * wake_futex_pi() must observe a state consistent with what we
2911 * observed.
2912 */
2913 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002914 spin_unlock(&hb->lock);
2915
Peter Zijlstrac74aef22017-09-22 17:48:06 +02002916 /* drops pi_state->pi_mutex.wait_lock */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002917 ret = wake_futex_pi(uaddr, uval, pi_state);
2918
2919 put_pi_state(pi_state);
2920
2921 /*
2922 * Success, we're done! No tricky corner cases.
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002923 */
2924 if (!ret)
2925 goto out_putkey;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002926 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002927 * The atomic access to the futex value generated a
2928 * pagefault, so retry the user-access and the wakeup:
Ingo Molnarc87e2832006-06-27 02:54:58 -07002929 */
2930 if (ret == -EFAULT)
2931 goto pi_faulted;
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002932 /*
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02002933 * A unconditional UNLOCK_PI op raced against a waiter
2934 * setting the FUTEX_WAITERS bit. Try again.
2935 */
2936 if (ret == -EAGAIN) {
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02002937 put_futex_key(&key);
2938 goto retry;
2939 }
2940 /*
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002941 * wake_futex_pi has detected invalid state. Tell user
2942 * space.
2943 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002944 goto out_putkey;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002945 }
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002946
Ingo Molnarc87e2832006-06-27 02:54:58 -07002947 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002948 * We have no kernel internal state, i.e. no waiters in the
2949 * kernel. Waiters which are about to queue themselves are stuck
2950 * on hb->lock. So we can safely ignore them. We do neither
2951 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2952 * owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002953 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002954 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
2955 spin_unlock(&hb->lock);
Thomas Gleixner13fbca42014-06-03 12:27:07 +00002956 goto pi_faulted;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002957 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002958
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002959 /*
2960 * If uval has changed, let user space handle it.
2961 */
2962 ret = (curval == uval) ? 0 : -EAGAIN;
2963
Ingo Molnarc87e2832006-06-27 02:54:58 -07002964out_unlock:
2965 spin_unlock(&hb->lock);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002966out_putkey:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002967 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002968 return ret;
2969
2970pi_faulted:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002971 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002972
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002973 ret = fault_in_user_writeable(uaddr);
Darren Hartb5686362008-12-18 15:06:34 -08002974 if (!ret)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002975 goto retry;
2976
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 return ret;
2978}
2979
Darren Hart52400ba2009-04-03 13:40:49 -07002980/**
2981 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2982 * @hb: the hash_bucket futex_q was original enqueued on
2983 * @q: the futex_q woken while waiting to be requeued
2984 * @key2: the futex_key of the requeue target futex
2985 * @timeout: the timeout associated with the wait (NULL if none)
2986 *
2987 * Detect if the task was woken on the initial futex as opposed to the requeue
2988 * target futex. If so, determine if it was a timeout or a signal that caused
2989 * the wakeup and return the appropriate error code to the caller. Must be
2990 * called with the hb lock held.
2991 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002992 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002993 * - 0 = no early wakeup detected;
2994 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
Darren Hart52400ba2009-04-03 13:40:49 -07002995 */
2996static inline
2997int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2998 struct futex_q *q, union futex_key *key2,
2999 struct hrtimer_sleeper *timeout)
3000{
3001 int ret = 0;
3002
3003 /*
3004 * With the hb lock held, we avoid races while we process the wakeup.
3005 * We only need to hold hb (and not hb2) to ensure atomicity as the
3006 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3007 * It can't be requeued from uaddr2 to something else since we don't
3008 * support a PI aware source futex for requeue.
3009 */
3010 if (!match_futex(&q->key, key2)) {
3011 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3012 /*
3013 * We were woken prior to requeue by a timeout or a signal.
3014 * Unqueue the futex_q and determine which it was.
3015 */
Lai Jiangshan2e129782010-12-22 14:18:50 +08003016 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07003017 hb_waiters_dec(hb);
Darren Hart52400ba2009-04-03 13:40:49 -07003018
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02003019 /* Handle spurious wakeups gracefully */
Thomas Gleixner11df6dd2009-10-28 20:26:48 +01003020 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07003021 if (timeout && !timeout->task)
3022 ret = -ETIMEDOUT;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02003023 else if (signal_pending(current))
Thomas Gleixner1c840c12009-05-20 09:22:40 +02003024 ret = -ERESTARTNOINTR;
Darren Hart52400ba2009-04-03 13:40:49 -07003025 }
3026 return ret;
3027}
3028
3029/**
3030 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
Darren Hart56ec1602009-09-21 22:29:59 -07003031 * @uaddr: the futex we initially wait on (non-pi)
Darren Hartb41277d2010-11-08 13:10:09 -08003032 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07003033 * the same type, no requeueing from private to shared, etc.
Darren Hart52400ba2009-04-03 13:40:49 -07003034 * @val: the expected value of uaddr
3035 * @abs_time: absolute timeout
Darren Hart56ec1602009-09-21 22:29:59 -07003036 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
Darren Hart52400ba2009-04-03 13:40:49 -07003037 * @uaddr2: the pi futex we will take prior to returning to user-space
3038 *
3039 * The caller will wait on uaddr and will be requeued by futex_requeue() to
Darren Hart6f7b0a22012-07-20 11:53:31 -07003040 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3041 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3042 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3043 * without one, the pi logic would not know which task to boost/deboost, if
3044 * there was a need to.
Darren Hart52400ba2009-04-03 13:40:49 -07003045 *
3046 * We call schedule in futex_wait_queue_me() when we enqueue and return there
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08003047 * via the following--
Darren Hart52400ba2009-04-03 13:40:49 -07003048 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
Darren Hartcc6db4e2009-07-31 16:20:10 -07003049 * 2) wakeup on uaddr2 after a requeue
3050 * 3) signal
3051 * 4) timeout
Darren Hart52400ba2009-04-03 13:40:49 -07003052 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07003053 * If 3, cleanup and return -ERESTARTNOINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07003054 *
3055 * If 2, we may then block on trying to take the rt_mutex and return via:
3056 * 5) successful lock
3057 * 6) signal
3058 * 7) timeout
3059 * 8) other lock acquisition failure
3060 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07003061 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
Darren Hart52400ba2009-04-03 13:40:49 -07003062 *
3063 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3064 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08003065 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03003066 * - 0 - On success;
3067 * - <0 - On error
Darren Hart52400ba2009-04-03 13:40:49 -07003068 */
Darren Hartb41277d2010-11-08 13:10:09 -08003069static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
Darren Hart52400ba2009-04-03 13:40:49 -07003070 u32 val, ktime_t *abs_time, u32 bitset,
Darren Hartb41277d2010-11-08 13:10:09 -08003071 u32 __user *uaddr2)
Darren Hart52400ba2009-04-03 13:40:49 -07003072{
3073 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003074 struct futex_pi_state *pi_state = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07003075 struct rt_mutex_waiter rt_waiter;
Darren Hart52400ba2009-04-03 13:40:49 -07003076 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08003077 union futex_key key2 = FUTEX_KEY_INIT;
3078 struct futex_q q = futex_q_init;
Darren Hart52400ba2009-04-03 13:40:49 -07003079 int res, ret;
Darren Hart52400ba2009-04-03 13:40:49 -07003080
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04003081 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3082 return -ENOSYS;
3083
Darren Hart6f7b0a22012-07-20 11:53:31 -07003084 if (uaddr == uaddr2)
3085 return -EINVAL;
3086
Darren Hart52400ba2009-04-03 13:40:49 -07003087 if (!bitset)
3088 return -EINVAL;
3089
3090 if (abs_time) {
3091 to = &timeout;
Darren Hartb41277d2010-11-08 13:10:09 -08003092 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3093 CLOCK_REALTIME : CLOCK_MONOTONIC,
3094 HRTIMER_MODE_ABS);
Darren Hart52400ba2009-04-03 13:40:49 -07003095 hrtimer_init_sleeper(to, current);
3096 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3097 current->timer_slack_ns);
3098 }
3099
3100 /*
3101 * The waiter is allocated on our stack, manipulated by the requeue
3102 * code while we sleep on uaddr.
3103 */
Peter Zijlstra50809352017-03-22 11:35:56 +01003104 rt_mutex_init_waiter(&rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003105
Shawn Bohrer9ea71502011-06-30 11:21:32 -05003106 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Darren Hart52400ba2009-04-03 13:40:49 -07003107 if (unlikely(ret != 0))
3108 goto out;
3109
Darren Hart84bc4af2009-08-13 17:36:53 -07003110 q.bitset = bitset;
3111 q.rt_waiter = &rt_waiter;
3112 q.requeue_pi_key = &key2;
3113
Darren Hart7ada8762010-10-17 08:35:04 -07003114 /*
3115 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3116 * count.
3117 */
Darren Hartb41277d2010-11-08 13:10:09 -08003118 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02003119 if (ret)
3120 goto out_key2;
Darren Hart52400ba2009-04-03 13:40:49 -07003121
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00003122 /*
3123 * The check above which compares uaddrs is not sufficient for
3124 * shared futexes. We need to compare the keys:
3125 */
3126 if (match_futex(&q.key, &key2)) {
Thomas Gleixner13c42c22014-09-11 23:44:35 +02003127 queue_unlock(hb);
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00003128 ret = -EINVAL;
3129 goto out_put_keys;
3130 }
3131
Darren Hart52400ba2009-04-03 13:40:49 -07003132 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02003133 futex_wait_queue_me(hb, &q, to);
Darren Hart52400ba2009-04-03 13:40:49 -07003134
3135 spin_lock(&hb->lock);
3136 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3137 spin_unlock(&hb->lock);
3138 if (ret)
3139 goto out_put_keys;
3140
3141 /*
3142 * In order for us to be here, we know our q.key == key2, and since
3143 * we took the hb->lock above, we also know that futex_requeue() has
3144 * completed and we no longer have to concern ourselves with a wakeup
Darren Hart7ada8762010-10-17 08:35:04 -07003145 * race with the atomic proxy lock acquisition by the requeue code. The
3146 * futex_requeue dropped our key1 reference and incremented our key2
3147 * reference count.
Darren Hart52400ba2009-04-03 13:40:49 -07003148 */
3149
3150 /* Check if the requeue code acquired the second futex for us. */
3151 if (!q.rt_waiter) {
3152 /*
3153 * Got the lock. We might not be the anticipated owner if we
3154 * did a lock-steal - fix up the PI-state in that case.
3155 */
3156 if (q.pi_state && (q.pi_state->owner != current)) {
3157 spin_lock(q.lock_ptr);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003158 ret = fixup_pi_state_owner(uaddr2, &q, current);
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003159 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3160 pi_state = q.pi_state;
3161 get_pi_state(pi_state);
3162 }
Thomas Gleixnerfb75a422015-12-19 20:07:38 +00003163 /*
3164 * Drop the reference to the pi state which
3165 * the requeue_pi() code acquired for us.
3166 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00003167 put_pi_state(q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003168 spin_unlock(q.lock_ptr);
3169 }
3170 } else {
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003171 struct rt_mutex *pi_mutex;
3172
Darren Hart52400ba2009-04-03 13:40:49 -07003173 /*
3174 * We have been woken up by futex_unlock_pi(), a timeout, or a
3175 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3176 * the pi_state.
3177 */
Darren Hartf27071c2012-07-20 11:53:30 -07003178 WARN_ON(!q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003179 pi_mutex = &q.pi_state->pi_mutex;
Peter Zijlstra38d589f2017-03-22 11:35:57 +01003180 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003181
3182 spin_lock(q.lock_ptr);
Peter Zijlstra38d589f2017-03-22 11:35:57 +01003183 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3184 ret = 0;
3185
3186 debug_rt_mutex_free_waiter(&rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003187 /*
3188 * Fixup the pi_state owner and possibly acquire the lock if we
3189 * haven't already.
3190 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003191 res = fixup_owner(uaddr2, &q, !ret);
Darren Hart52400ba2009-04-03 13:40:49 -07003192 /*
3193 * If fixup_owner() returned an error, proprogate that. If it
Darren Hart56ec1602009-09-21 22:29:59 -07003194 * acquired the lock, clear -ETIMEDOUT or -EINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07003195 */
3196 if (res)
3197 ret = (res < 0) ? res : 0;
3198
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003199 /*
3200 * If fixup_pi_state_owner() faulted and was unable to handle
3201 * the fault, unlock the rt_mutex and return the fault to
3202 * userspace.
3203 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003204 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3205 pi_state = q.pi_state;
3206 get_pi_state(pi_state);
3207 }
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003208
Darren Hart52400ba2009-04-03 13:40:49 -07003209 /* Unqueue and drop the lock. */
3210 unqueue_me_pi(&q);
3211 }
3212
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003213 if (pi_state) {
3214 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3215 put_pi_state(pi_state);
3216 }
3217
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003218 if (ret == -EINTR) {
Darren Hart52400ba2009-04-03 13:40:49 -07003219 /*
Darren Hartcc6db4e2009-07-31 16:20:10 -07003220 * We've already been requeued, but cannot restart by calling
3221 * futex_lock_pi() directly. We could restart this syscall, but
3222 * it would detect that the user space "val" changed and return
3223 * -EWOULDBLOCK. Save the overhead of the restart and return
3224 * -EWOULDBLOCK directly.
Darren Hart52400ba2009-04-03 13:40:49 -07003225 */
Thomas Gleixner20708872009-05-19 23:04:59 +02003226 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07003227 }
3228
3229out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003230 put_futex_key(&q.key);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02003231out_key2:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003232 put_futex_key(&key2);
Darren Hart52400ba2009-04-03 13:40:49 -07003233
3234out:
3235 if (to) {
3236 hrtimer_cancel(&to->timer);
3237 destroy_hrtimer_on_stack(&to->timer);
3238 }
3239 return ret;
3240}
3241
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003242/*
3243 * Support for robust futexes: the kernel cleans up held futexes at
3244 * thread exit time.
3245 *
3246 * Implementation: user-space maintains a per-thread list of locks it
3247 * is holding. Upon do_exit(), the kernel carefully walks this list,
3248 * and marks all locks that are owned by this thread with the
Ingo Molnarc87e2832006-06-27 02:54:58 -07003249 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003250 * always manipulated with the lock held, so the list is private and
3251 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3252 * field, to allow the kernel to clean up if the thread dies after
3253 * acquiring the lock, but just before it could have added itself to
3254 * the list. There can only be one such pending lock.
3255 */
3256
3257/**
Darren Hartd96ee562009-09-21 22:30:22 -07003258 * sys_set_robust_list() - Set the robust-futex list head of a task
3259 * @head: pointer to the list-head
3260 * @len: length of the list-head, as userspace expects
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003261 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003262SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3263 size_t, len)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003264{
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003265 if (!futex_cmpxchg_enabled)
3266 return -ENOSYS;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003267 /*
3268 * The kernel knows only one size for now:
3269 */
3270 if (unlikely(len != sizeof(*head)))
3271 return -EINVAL;
3272
3273 current->robust_list = head;
3274
3275 return 0;
3276}
3277
3278/**
Darren Hartd96ee562009-09-21 22:30:22 -07003279 * sys_get_robust_list() - Get the robust-futex list head of a task
3280 * @pid: pid of the process [zero for current task]
3281 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3282 * @len_ptr: pointer to a length field, the kernel fills in the header size
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003283 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003284SYSCALL_DEFINE3(get_robust_list, int, pid,
3285 struct robust_list_head __user * __user *, head_ptr,
3286 size_t __user *, len_ptr)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003287{
Al Viroba46df92006-10-10 22:46:07 +01003288 struct robust_list_head __user *head;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003289 unsigned long ret;
Kees Cookbdbb7762012-03-19 16:12:53 -07003290 struct task_struct *p;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003291
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003292 if (!futex_cmpxchg_enabled)
3293 return -ENOSYS;
3294
Kees Cookbdbb7762012-03-19 16:12:53 -07003295 rcu_read_lock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003296
Kees Cookbdbb7762012-03-19 16:12:53 -07003297 ret = -ESRCH;
3298 if (!pid)
3299 p = current;
3300 else {
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07003301 p = find_task_by_vpid(pid);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003302 if (!p)
3303 goto err_unlock;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003304 }
3305
Kees Cookbdbb7762012-03-19 16:12:53 -07003306 ret = -EPERM;
Jann Horncaaee622016-01-20 15:00:04 -08003307 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
Kees Cookbdbb7762012-03-19 16:12:53 -07003308 goto err_unlock;
3309
3310 head = p->robust_list;
3311 rcu_read_unlock();
3312
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003313 if (put_user(sizeof(*head), len_ptr))
3314 return -EFAULT;
3315 return put_user(head, head_ptr);
3316
3317err_unlock:
Oleg Nesterovaaa2a972006-09-29 02:00:55 -07003318 rcu_read_unlock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003319
3320 return ret;
3321}
3322
3323/*
3324 * Process a futex-list entry, check whether it's owned by the
3325 * dying task, and do notification if so:
3326 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003327int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003328{
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03003329 u32 uval, uninitialized_var(nval), mval;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003330
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003331retry:
3332 if (get_user(uval, uaddr))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003333 return -1;
3334
Pavel Emelyanovb4888932007-10-18 23:40:14 -07003335 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003336 /*
3337 * Ok, this dying thread is truly holding a futex
3338 * of interest. Set the OWNER_DIED bit atomically
3339 * via cmpxchg, and if the value had FUTEX_WAITERS
3340 * set, wake up a waiter (if any). (We have to do a
3341 * futex_wake() even if OWNER_DIED is already set -
3342 * to handle the rare but possible case of recursive
3343 * thread-death.) The rest of the cleanup is done in
3344 * userspace.
3345 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003346 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
Thomas Gleixner6e0aa9f2011-03-14 10:34:35 +01003347 /*
3348 * We are not holding a lock here, but we want to have
3349 * the pagefault_disable/enable() protection because
3350 * we want to handle the fault gracefully. If the
3351 * access fails we try to fault in the futex with R/W
3352 * verification via get_user_pages. get_user() above
3353 * does not guarantee R/W access. If that fails we
3354 * give up and leave the futex locked.
3355 */
3356 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3357 if (fault_in_user_writeable(uaddr))
3358 return -1;
3359 goto retry;
3360 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07003361 if (nval != uval)
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003362 goto retry;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003363
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003364 /*
3365 * Wake robust non-PI futexes here. The wakeup of
3366 * PI futexes happens in exit_pi_state():
3367 */
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07003368 if (!pi && (uval & FUTEX_WAITERS))
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02003369 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003370 }
3371 return 0;
3372}
3373
3374/*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003375 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3376 */
3377static inline int fetch_robust_entry(struct robust_list __user **entry,
Al Viroba46df92006-10-10 22:46:07 +01003378 struct robust_list __user * __user *head,
Namhyung Kim1dcc41b2010-09-14 21:43:46 +09003379 unsigned int *pi)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003380{
3381 unsigned long uentry;
3382
Al Viroba46df92006-10-10 22:46:07 +01003383 if (get_user(uentry, (unsigned long __user *)head))
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003384 return -EFAULT;
3385
Al Viroba46df92006-10-10 22:46:07 +01003386 *entry = (void __user *)(uentry & ~1UL);
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003387 *pi = uentry & 1;
3388
3389 return 0;
3390}
3391
3392/*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003393 * Walk curr->robust_list (very carefully, it's a userspace list!)
3394 * and mark any locks found there dead, and notify any waiters.
3395 *
3396 * We silently return on any sign of list-walking problem.
3397 */
3398void exit_robust_list(struct task_struct *curr)
3399{
3400 struct robust_list_head __user *head = curr->robust_list;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003401 struct robust_list __user *entry, *next_entry, *pending;
Darren Hart4c115e92010-11-04 15:00:00 -04003402 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3403 unsigned int uninitialized_var(next_pi);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003404 unsigned long futex_offset;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003405 int rc;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003406
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003407 if (!futex_cmpxchg_enabled)
3408 return;
3409
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003410 /*
3411 * Fetch the list head (which was registered earlier, via
3412 * sys_set_robust_list()):
3413 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003414 if (fetch_robust_entry(&entry, &head->list.next, &pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003415 return;
3416 /*
3417 * Fetch the relative futex offset:
3418 */
3419 if (get_user(futex_offset, &head->futex_offset))
3420 return;
3421 /*
3422 * Fetch any possibly pending lock-add first, and handle it
3423 * if it exists:
3424 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003425 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003426 return;
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003427
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003428 next_entry = NULL; /* avoid warning with gcc */
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003429 while (entry != &head->list) {
3430 /*
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003431 * Fetch the next entry in the list before calling
3432 * handle_futex_death:
3433 */
3434 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3435 /*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003436 * A pending lock might already be on the list, so
Ingo Molnarc87e2832006-06-27 02:54:58 -07003437 * don't process it twice:
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003438 */
3439 if (entry != pending)
Al Viroba46df92006-10-10 22:46:07 +01003440 if (handle_futex_death((void __user *)entry + futex_offset,
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003441 curr, pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003442 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003443 if (rc)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003444 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003445 entry = next_entry;
3446 pi = next_pi;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003447 /*
3448 * Avoid excessively long or circular lists:
3449 */
3450 if (!--limit)
3451 break;
3452
3453 cond_resched();
3454 }
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003455
3456 if (pending)
3457 handle_futex_death((void __user *)pending + futex_offset,
3458 curr, pip);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003459}
3460
Pierre Peifferc19384b2007-05-09 02:35:02 -07003461long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
Ingo Molnare2970f22006-06-27 02:54:47 -07003462 u32 __user *uaddr2, u32 val2, u32 val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463{
Thomas Gleixner81b40532012-02-15 12:17:09 +01003464 int cmd = op & FUTEX_CMD_MASK;
Darren Hartb41277d2010-11-08 13:10:09 -08003465 unsigned int flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003467 if (!(op & FUTEX_PRIVATE_FLAG))
Darren Hartb41277d2010-11-08 13:10:09 -08003468 flags |= FLAGS_SHARED;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003469
Darren Hartb41277d2010-11-08 13:10:09 -08003470 if (op & FUTEX_CLOCK_REALTIME) {
3471 flags |= FLAGS_CLOCKRT;
Darren Hart337f1302015-12-18 13:36:37 -08003472 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3473 cmd != FUTEX_WAIT_REQUEUE_PI)
Darren Hartb41277d2010-11-08 13:10:09 -08003474 return -ENOSYS;
3475 }
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003476
3477 switch (cmd) {
Thomas Gleixner59263b52012-02-15 12:08:34 +01003478 case FUTEX_LOCK_PI:
3479 case FUTEX_UNLOCK_PI:
3480 case FUTEX_TRYLOCK_PI:
3481 case FUTEX_WAIT_REQUEUE_PI:
3482 case FUTEX_CMP_REQUEUE_PI:
3483 if (!futex_cmpxchg_enabled)
3484 return -ENOSYS;
3485 }
3486
3487 switch (cmd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 case FUTEX_WAIT:
Thomas Gleixnercd6899852008-02-01 17:45:14 +01003489 val3 = FUTEX_BITSET_MATCH_ANY;
3490 case FUTEX_WAIT_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003491 return futex_wait(uaddr, flags, val, timeout, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492 case FUTEX_WAKE:
Thomas Gleixnercd6899852008-02-01 17:45:14 +01003493 val3 = FUTEX_BITSET_MATCH_ANY;
3494 case FUTEX_WAKE_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003495 return futex_wake(uaddr, flags, val, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 case FUTEX_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003497 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 case FUTEX_CMP_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003499 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07003500 case FUTEX_WAKE_OP:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003501 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003502 case FUTEX_LOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003503 return futex_lock_pi(uaddr, flags, timeout, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003504 case FUTEX_UNLOCK_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003505 return futex_unlock_pi(uaddr, flags);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003506 case FUTEX_TRYLOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003507 return futex_lock_pi(uaddr, flags, NULL, 1);
Darren Hart52400ba2009-04-03 13:40:49 -07003508 case FUTEX_WAIT_REQUEUE_PI:
3509 val3 = FUTEX_BITSET_MATCH_ANY;
Thomas Gleixner81b40532012-02-15 12:17:09 +01003510 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3511 uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07003512 case FUTEX_CMP_REQUEUE_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003513 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 }
Thomas Gleixner81b40532012-02-15 12:17:09 +01003515 return -ENOSYS;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516}
3517
3518
Heiko Carstens17da2bd2009-01-14 14:14:10 +01003519SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3520 struct timespec __user *, utime, u32 __user *, uaddr2,
3521 u32, val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522{
Pierre Peifferc19384b2007-05-09 02:35:02 -07003523 struct timespec ts;
3524 ktime_t t, *tp = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07003525 u32 val2 = 0;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003526 int cmd = op & FUTEX_CMD_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527
Thomas Gleixnercd6899852008-02-01 17:45:14 +01003528 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
Darren Hart52400ba2009-04-03 13:40:49 -07003529 cmd == FUTEX_WAIT_BITSET ||
3530 cmd == FUTEX_WAIT_REQUEUE_PI)) {
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07003531 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3532 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003533 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003535 if (!timespec_valid(&ts))
Thomas Gleixner9741ef962006-03-31 02:31:32 -08003536 return -EINVAL;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003537
3538 t = timespec_to_ktime(ts);
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003539 if (cmd == FUTEX_WAIT)
Thomas Gleixner5a7780e2008-02-13 09:20:43 +01003540 t = ktime_add_safe(ktime_get(), t);
Pierre Peifferc19384b2007-05-09 02:35:02 -07003541 tp = &t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 }
3543 /*
Darren Hart52400ba2009-04-03 13:40:49 -07003544 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
Andreas Schwabf54f0982007-07-31 00:38:51 -07003545 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546 */
Andreas Schwabf54f0982007-07-31 00:38:51 -07003547 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
Darren Hartba9c22f2009-04-20 22:22:22 -07003548 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
Ingo Molnare2970f22006-06-27 02:54:47 -07003549 val2 = (u32) (unsigned long) utime;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550
Pierre Peifferc19384b2007-05-09 02:35:02 -07003551 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552}
3553
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003554static void __init futex_detect_cmpxchg(void)
3555{
3556#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3557 u32 curval;
3558
3559 /*
3560 * This will fail and we want it. Some arch implementations do
3561 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3562 * functionality. We want to know that before we call in any
3563 * of the complex code paths. Also we want to prevent
3564 * registration of robust lists in that case. NULL is
3565 * guaranteed to fault and we get -EFAULT on functional
3566 * implementation, the non-functional ones will return
3567 * -ENOSYS.
3568 */
3569 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3570 futex_cmpxchg_enabled = 1;
3571#endif
3572}
3573
Benjamin Herrenschmidtf6d107f2008-03-27 14:52:15 +11003574static int __init futex_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575{
Heiko Carstens63b1a812014-01-16 14:54:50 +01003576 unsigned int futex_shift;
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003577 unsigned long i;
3578
3579#if CONFIG_BASE_SMALL
3580 futex_hashsize = 16;
3581#else
3582 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3583#endif
3584
3585 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3586 futex_hashsize, 0,
3587 futex_hashsize < 256 ? HASH_SMALL : 0,
Heiko Carstens63b1a812014-01-16 14:54:50 +01003588 &futex_shift, NULL,
3589 futex_hashsize, futex_hashsize);
3590 futex_hashsize = 1UL << futex_shift;
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003591
3592 futex_detect_cmpxchg();
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003593
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003594 for (i = 0; i < futex_hashsize; i++) {
Linus Torvalds11d46162014-03-20 22:11:17 -07003595 atomic_set(&futex_queues[i].waiters, 0);
Dima Zavin732375c2011-07-07 17:27:59 -07003596 plist_head_init(&futex_queues[i].chain);
Thomas Gleixner3e4ab742008-02-23 15:23:55 -08003597 spin_lock_init(&futex_queues[i].lock);
3598 }
3599
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 return 0;
3601}
Yang Yang25f71d12016-12-30 16:17:55 +08003602core_initcall(futex_init);