blob: 0939255fc75077f3f07e0ec8259b9c93f645b874 [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 Gleixnercd689982008-02-01 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.
824 *
Brian Silverman30a6b802014-10-25 20:20:37 -0400825 * Must be called with the hb lock held.
826 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +0000827static void put_pi_state(struct futex_pi_state *pi_state)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700828{
Brian Silverman30a6b802014-10-25 20:20:37 -0400829 if (!pi_state)
830 return;
831
Ingo Molnarc87e2832006-06-27 02:54:58 -0700832 if (!atomic_dec_and_test(&pi_state->refcount))
833 return;
834
835 /*
836 * If pi_state->owner is NULL, the owner is most probably dying
837 * and has cleaned up the pi_state already
838 */
839 if (pi_state->owner) {
Thomas Gleixner1d615482009-11-17 14:54:03 +0100840 raw_spin_lock_irq(&pi_state->owner->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700841 list_del_init(&pi_state->list);
Thomas Gleixner1d615482009-11-17 14:54:03 +0100842 raw_spin_unlock_irq(&pi_state->owner->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700843
844 rt_mutex_proxy_unlock(&pi_state->pi_mutex, pi_state->owner);
845 }
846
847 if (current->pi_state_cache)
848 kfree(pi_state);
849 else {
850 /*
851 * pi_state->list is already empty.
852 * clear pi_state->owner.
853 * refcount is at 0 - put it back to 1.
854 */
855 pi_state->owner = NULL;
856 atomic_set(&pi_state->refcount, 1);
857 current->pi_state_cache = pi_state;
858 }
859}
860
861/*
862 * Look up the task based on what TID userspace gave us.
863 * We dont trust it.
864 */
Peter Zijlstrabf92cf32017-03-22 11:35:53 +0100865static struct task_struct *futex_find_get_task(pid_t pid)
Ingo Molnarc87e2832006-06-27 02:54:58 -0700866{
867 struct task_struct *p;
868
Oleg Nesterovd359b542006-09-29 02:00:55 -0700869 rcu_read_lock();
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -0700870 p = find_task_by_vpid(pid);
Michal Hocko7a0ea092010-06-30 09:51:19 +0200871 if (p)
872 get_task_struct(p);
Thomas Gleixnera06381f2007-06-23 11:48:40 +0200873
Oleg Nesterovd359b542006-09-29 02:00:55 -0700874 rcu_read_unlock();
Ingo Molnarc87e2832006-06-27 02:54:58 -0700875
876 return p;
877}
878
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -0400879#ifdef CONFIG_FUTEX_PI
880
Ingo Molnarc87e2832006-06-27 02:54:58 -0700881/*
882 * This task is holding PI mutexes at exit time => bad.
883 * Kernel cleans up PI-state, but userspace is likely hosed.
884 * (Robust-futex cleanup is separate and might save the day for userspace.)
885 */
886void exit_pi_state_list(struct task_struct *curr)
887{
Ingo Molnarc87e2832006-06-27 02:54:58 -0700888 struct list_head *next, *head = &curr->pi_state_list;
889 struct futex_pi_state *pi_state;
Ingo Molnar627371d2006-07-29 05:16:20 +0200890 struct futex_hash_bucket *hb;
Peter Zijlstra38d47c12008-09-26 19:32:20 +0200891 union futex_key key = FUTEX_KEY_INIT;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700892
Thomas Gleixnera0c1e902008-02-23 15:23:57 -0800893 if (!futex_cmpxchg_enabled)
894 return;
Ingo Molnarc87e2832006-06-27 02:54:58 -0700895 /*
896 * We are a ZOMBIE and nobody can enqueue itself on
897 * pi_state_list anymore, but we have to be careful
Ingo Molnar627371d2006-07-29 05:16:20 +0200898 * versus waiters unqueueing themselves:
Ingo Molnarc87e2832006-06-27 02:54:58 -0700899 */
Thomas Gleixner1d615482009-11-17 14:54:03 +0100900 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700901 while (!list_empty(head)) {
902
903 next = head->next;
904 pi_state = list_entry(next, struct futex_pi_state, list);
905 key = pi_state->key;
Ingo Molnar627371d2006-07-29 05:16:20 +0200906 hb = hash_futex(&key);
Thomas Gleixner1d615482009-11-17 14:54:03 +0100907 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700908
Ingo Molnarc87e2832006-06-27 02:54:58 -0700909 spin_lock(&hb->lock);
910
Thomas Gleixner1d615482009-11-17 14:54:03 +0100911 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +0200912 /*
913 * We dropped the pi-lock, so re-check whether this
914 * task still owns the PI-state:
915 */
Ingo Molnarc87e2832006-06-27 02:54:58 -0700916 if (head->next != next) {
917 spin_unlock(&hb->lock);
918 continue;
919 }
920
Ingo Molnarc87e2832006-06-27 02:54:58 -0700921 WARN_ON(pi_state->owner != curr);
Ingo Molnar627371d2006-07-29 05:16:20 +0200922 WARN_ON(list_empty(&pi_state->list));
923 list_del_init(&pi_state->list);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700924 pi_state->owner = NULL;
Thomas Gleixner1d615482009-11-17 14:54:03 +0100925 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700926
Peter Zijlstra16ffa122017-03-22 11:35:55 +0100927 get_pi_state(pi_state);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700928 spin_unlock(&hb->lock);
929
Peter Zijlstra16ffa122017-03-22 11:35:55 +0100930 rt_mutex_futex_unlock(&pi_state->pi_mutex);
931 put_pi_state(pi_state);
932
Thomas Gleixner1d615482009-11-17 14:54:03 +0100933 raw_spin_lock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700934 }
Thomas Gleixner1d615482009-11-17 14:54:03 +0100935 raw_spin_unlock_irq(&curr->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -0700936}
937
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -0400938#endif
939
Thomas Gleixner54a21782014-06-03 12:27:08 +0000940/*
941 * We need to check the following states:
942 *
943 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
944 *
945 * [1] NULL | --- | --- | 0 | 0/1 | Valid
946 * [2] NULL | --- | --- | >0 | 0/1 | Valid
947 *
948 * [3] Found | NULL | -- | Any | 0/1 | Invalid
949 *
950 * [4] Found | Found | NULL | 0 | 1 | Valid
951 * [5] Found | Found | NULL | >0 | 1 | Invalid
952 *
953 * [6] Found | Found | task | 0 | 1 | Valid
954 *
955 * [7] Found | Found | NULL | Any | 0 | Invalid
956 *
957 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
958 * [9] Found | Found | task | 0 | 0 | Invalid
959 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
960 *
961 * [1] Indicates that the kernel can acquire the futex atomically. We
962 * came came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
963 *
964 * [2] Valid, if TID does not belong to a kernel thread. If no matching
965 * thread is found then it indicates that the owner TID has died.
966 *
967 * [3] Invalid. The waiter is queued on a non PI futex
968 *
969 * [4] Valid state after exit_robust_list(), which sets the user space
970 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
971 *
972 * [5] The user space value got manipulated between exit_robust_list()
973 * and exit_pi_state_list()
974 *
975 * [6] Valid state after exit_pi_state_list() which sets the new owner in
976 * the pi_state but cannot access the user space value.
977 *
978 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
979 *
980 * [8] Owner and user space value match
981 *
982 * [9] There is no transient state which sets the user space TID to 0
983 * except exit_robust_list(), but this is indicated by the
984 * FUTEX_OWNER_DIED bit. See [4]
985 *
986 * [10] There is no transient state which leaves owner and user space
987 * TID out of sync.
Peter Zijlstra734009e2017-03-22 11:35:52 +0100988 *
989 *
990 * Serialization and lifetime rules:
991 *
992 * hb->lock:
993 *
994 * hb -> futex_q, relation
995 * futex_q -> pi_state, relation
996 *
997 * (cannot be raw because hb can contain arbitrary amount
998 * of futex_q's)
999 *
1000 * pi_mutex->wait_lock:
1001 *
1002 * {uval, pi_state}
1003 *
1004 * (and pi_mutex 'obviously')
1005 *
1006 * p->pi_lock:
1007 *
1008 * p->pi_state_list -> pi_state->list, relation
1009 *
1010 * pi_state->refcount:
1011 *
1012 * pi_state lifetime
1013 *
1014 *
1015 * Lock order:
1016 *
1017 * hb->lock
1018 * pi_mutex->wait_lock
1019 * p->pi_lock
1020 *
Thomas Gleixner54a21782014-06-03 12:27:08 +00001021 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001022
1023/*
1024 * Validate that the existing waiter has a pi_state and sanity check
1025 * the pi_state against the user space value. If correct, attach to
1026 * it.
1027 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001028static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1029 struct futex_pi_state *pi_state,
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001030 struct futex_pi_state **ps)
1031{
1032 pid_t pid = uval & FUTEX_TID_MASK;
Peter Zijlstra94ffac52017-04-07 09:04:07 +02001033 u32 uval2;
1034 int ret;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001035
1036 /*
1037 * Userspace might have messed up non-PI and PI futexes [3]
1038 */
1039 if (unlikely(!pi_state))
1040 return -EINVAL;
1041
Peter Zijlstra734009e2017-03-22 11:35:52 +01001042 /*
1043 * We get here with hb->lock held, and having found a
1044 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1045 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1046 * which in turn means that futex_lock_pi() still has a reference on
1047 * our pi_state.
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001048 *
1049 * The waiter holding a reference on @pi_state also protects against
1050 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1051 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1052 * free pi_state before we can take a reference ourselves.
Peter Zijlstra734009e2017-03-22 11:35:52 +01001053 */
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001054 WARN_ON(!atomic_read(&pi_state->refcount));
1055
1056 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01001057 * Now that we have a pi_state, we can acquire wait_lock
1058 * and do the state validation.
1059 */
1060 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1061
1062 /*
1063 * Since {uval, pi_state} is serialized by wait_lock, and our current
1064 * uval was read without holding it, it can have changed. Verify it
1065 * still is what we expect it to be, otherwise retry the entire
1066 * operation.
1067 */
1068 if (get_futex_value_locked(&uval2, uaddr))
1069 goto out_efault;
1070
1071 if (uval != uval2)
1072 goto out_eagain;
1073
1074 /*
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001075 * Handle the owner died case:
1076 */
1077 if (uval & FUTEX_OWNER_DIED) {
1078 /*
1079 * exit_pi_state_list sets owner to NULL and wakes the
1080 * topmost waiter. The task which acquires the
1081 * pi_state->rt_mutex will fixup owner.
1082 */
1083 if (!pi_state->owner) {
1084 /*
1085 * No pi state owner, but the user space TID
1086 * is not 0. Inconsistent state. [5]
1087 */
1088 if (pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001089 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001090 /*
1091 * Take a ref on the state and return success. [4]
1092 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001093 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001094 }
1095
1096 /*
1097 * If TID is 0, then either the dying owner has not
1098 * yet executed exit_pi_state_list() or some waiter
1099 * acquired the rtmutex in the pi state, but did not
1100 * yet fixup the TID in user space.
1101 *
1102 * Take a ref on the state and return success. [6]
1103 */
1104 if (!pid)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001105 goto out_attach;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001106 } else {
1107 /*
1108 * If the owner died bit is not set, then the pi_state
1109 * must have an owner. [7]
1110 */
1111 if (!pi_state->owner)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001112 goto out_einval;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001113 }
1114
1115 /*
1116 * Bail out if user space manipulated the futex value. If pi
1117 * state exists then the owner TID must be the same as the
1118 * user space TID. [9/10]
1119 */
1120 if (pid != task_pid_vnr(pi_state->owner))
Peter Zijlstra734009e2017-03-22 11:35:52 +01001121 goto out_einval;
1122
1123out_attach:
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01001124 get_pi_state(pi_state);
Peter Zijlstra734009e2017-03-22 11:35:52 +01001125 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001126 *ps = pi_state;
1127 return 0;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001128
1129out_einval:
1130 ret = -EINVAL;
1131 goto out_error;
1132
1133out_eagain:
1134 ret = -EAGAIN;
1135 goto out_error;
1136
1137out_efault:
1138 ret = -EFAULT;
1139 goto out_error;
1140
1141out_error:
1142 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1143 return ret;
Thomas Gleixnere60cbc52014-06-11 20:45:39 +00001144}
1145
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001146/*
1147 * Lookup the task for the TID provided from user space and attach to
1148 * it after doing proper sanity checks.
1149 */
1150static int attach_to_pi_owner(u32 uval, union futex_key *key,
1151 struct futex_pi_state **ps)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001152{
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001153 pid_t pid = uval & FUTEX_TID_MASK;
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001154 struct futex_pi_state *pi_state;
1155 struct task_struct *p;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001156
1157 /*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001158 * We are the first waiter - try to look up the real owner and attach
Thomas Gleixner54a21782014-06-03 12:27:08 +00001159 * the new pi_state to it, but bail out when TID = 0 [1]
Ingo Molnarc87e2832006-06-27 02:54:58 -07001160 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001161 if (!pid)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02001162 return -ESRCH;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001163 p = futex_find_get_task(pid);
Michal Hocko7a0ea092010-06-30 09:51:19 +02001164 if (!p)
1165 return -ESRCH;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001166
Oleg Nesterova2129462015-02-02 15:05:36 +01001167 if (unlikely(p->flags & PF_KTHREAD)) {
Thomas Gleixnerf0d71b32014-05-12 20:45:35 +00001168 put_task_struct(p);
1169 return -EPERM;
1170 }
1171
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001172 /*
1173 * We need to look at the task state flags to figure out,
1174 * whether the task is exiting. To protect against the do_exit
1175 * change of the task flags, we do this protected by
1176 * p->pi_lock:
1177 */
Thomas Gleixner1d615482009-11-17 14:54:03 +01001178 raw_spin_lock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001179 if (unlikely(p->flags & PF_EXITING)) {
1180 /*
1181 * The task is on the way out. When PF_EXITPIDONE is
1182 * set, we know that the task has finished the
1183 * cleanup:
1184 */
1185 int ret = (p->flags & PF_EXITPIDONE) ? -ESRCH : -EAGAIN;
1186
Thomas Gleixner1d615482009-11-17 14:54:03 +01001187 raw_spin_unlock_irq(&p->pi_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001188 put_task_struct(p);
1189 return ret;
1190 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001191
Thomas Gleixner54a21782014-06-03 12:27:08 +00001192 /*
1193 * No existing pi state. First waiter. [2]
Peter Zijlstra734009e2017-03-22 11:35:52 +01001194 *
1195 * This creates pi_state, we have hb->lock held, this means nothing can
1196 * observe this state, wait_lock is irrelevant.
Thomas Gleixner54a21782014-06-03 12:27:08 +00001197 */
Ingo Molnarc87e2832006-06-27 02:54:58 -07001198 pi_state = alloc_pi_state();
1199
1200 /*
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001201 * Initialize the pi_mutex in locked state and make @p
Ingo Molnarc87e2832006-06-27 02:54:58 -07001202 * the owner of it:
1203 */
1204 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1205
1206 /* Store the key for possible exit cleanups: */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001207 pi_state->key = *key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001208
Ingo Molnar627371d2006-07-29 05:16:20 +02001209 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001210 list_add(&pi_state->list, &p->pi_state_list);
1211 pi_state->owner = p;
Thomas Gleixner1d615482009-11-17 14:54:03 +01001212 raw_spin_unlock_irq(&p->pi_lock);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001213
1214 put_task_struct(p);
1215
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07001216 *ps = pi_state;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001217
1218 return 0;
1219}
1220
Peter Zijlstra734009e2017-03-22 11:35:52 +01001221static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1222 struct futex_hash_bucket *hb,
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001223 union futex_key *key, struct futex_pi_state **ps)
1224{
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001225 struct futex_q *top_waiter = futex_top_waiter(hb, key);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001226
1227 /*
1228 * If there is a waiter on that futex, validate it and
1229 * attach to the pi_state when the validation succeeds.
1230 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001231 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001232 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixner04e1b2e2014-06-11 20:45:40 +00001233
1234 /*
1235 * We are the first waiter - try to look up the owner based on
1236 * @uval and attach to it.
1237 */
1238 return attach_to_pi_owner(uval, key, ps);
1239}
1240
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001241static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1242{
1243 u32 uninitialized_var(curval);
1244
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001245 if (unlikely(should_fail_futex(true)))
1246 return -EFAULT;
1247
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001248 if (unlikely(cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)))
1249 return -EFAULT;
1250
Peter Zijlstra734009e2017-03-22 11:35:52 +01001251 /* If user space value changed, let the caller retry */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001252 return curval != uval ? -EAGAIN : 0;
1253}
1254
Darren Hart1a520842009-04-03 13:39:52 -07001255/**
Darren Hartd96ee562009-09-21 22:30:22 -07001256 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
Darren Hartbab5bc92009-04-07 23:23:50 -07001257 * @uaddr: the pi futex user address
1258 * @hb: the pi futex hash bucket
1259 * @key: the futex key associated with uaddr and hb
1260 * @ps: the pi_state pointer where we store the result of the
1261 * lookup
1262 * @task: the task to perform the atomic lock work for. This will
1263 * be "current" except in the case of requeue pi.
1264 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart1a520842009-04-03 13:39:52 -07001265 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001266 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03001267 * - 0 - ready to wait;
1268 * - 1 - acquired the lock;
1269 * - <0 - error
Darren Hart1a520842009-04-03 13:39:52 -07001270 *
1271 * The hb->lock and futex_key refs shall be held by the caller.
1272 */
1273static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1274 union futex_key *key,
1275 struct futex_pi_state **ps,
Darren Hartbab5bc92009-04-07 23:23:50 -07001276 struct task_struct *task, int set_waiters)
Darren Hart1a520842009-04-03 13:39:52 -07001277{
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001278 u32 uval, newval, vpid = task_pid_vnr(task);
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001279 struct futex_q *top_waiter;
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001280 int ret;
Darren Hart1a520842009-04-03 13:39:52 -07001281
1282 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001283 * Read the user space value first so we can validate a few
1284 * things before proceeding further.
Darren Hart1a520842009-04-03 13:39:52 -07001285 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001286 if (get_futex_value_locked(&uval, uaddr))
Darren Hart1a520842009-04-03 13:39:52 -07001287 return -EFAULT;
1288
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001289 if (unlikely(should_fail_futex(true)))
1290 return -EFAULT;
1291
Darren Hart1a520842009-04-03 13:39:52 -07001292 /*
1293 * Detect deadlocks.
1294 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001295 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
Darren Hart1a520842009-04-03 13:39:52 -07001296 return -EDEADLK;
1297
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001298 if ((unlikely(should_fail_futex(true))))
1299 return -EDEADLK;
1300
Darren Hart1a520842009-04-03 13:39:52 -07001301 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001302 * Lookup existing state first. If it exists, try to attach to
1303 * its pi_state.
Darren Hart1a520842009-04-03 13:39:52 -07001304 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01001305 top_waiter = futex_top_waiter(hb, key);
1306 if (top_waiter)
Peter Zijlstra734009e2017-03-22 11:35:52 +01001307 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001308
1309 /*
1310 * No waiter and user TID is 0. We are here because the
1311 * waiters or the owner died bit is set or called from
1312 * requeue_cmp_pi or for whatever reason something took the
1313 * syscall.
1314 */
1315 if (!(uval & FUTEX_TID_MASK)) {
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001316 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001317 * We take over the futex. No other waiters and the user space
1318 * TID is 0. We preserve the owner died bit.
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001319 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001320 newval = uval & FUTEX_OWNER_DIED;
1321 newval |= vpid;
1322
1323 /* The futex requeue_pi code can enforce the waiters bit */
1324 if (set_waiters)
1325 newval |= FUTEX_WAITERS;
1326
1327 ret = lock_pi_update_atomic(uaddr, uval, newval);
1328 /* If the take over worked, return 1 */
1329 return ret < 0 ? ret : 1;
Thomas Gleixnerb3eaa9f2014-06-03 12:27:06 +00001330 }
Darren Hart1a520842009-04-03 13:39:52 -07001331
Darren Hart1a520842009-04-03 13:39:52 -07001332 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001333 * First waiter. Set the waiters bit before attaching ourself to
1334 * the owner. If owner tries to unlock, it will be forced into
1335 * the kernel and blocked on hb->lock.
Darren Hart1a520842009-04-03 13:39:52 -07001336 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001337 newval = uval | FUTEX_WAITERS;
1338 ret = lock_pi_update_atomic(uaddr, uval, newval);
1339 if (ret)
1340 return ret;
Darren Hart1a520842009-04-03 13:39:52 -07001341 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001342 * If the update of the user space value succeeded, we try to
1343 * attach to the owner. If that fails, no harm done, we only
1344 * set the FUTEX_WAITERS bit in the user space variable.
Darren Hart1a520842009-04-03 13:39:52 -07001345 */
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001346 return attach_to_pi_owner(uval, key, ps);
Darren Hart1a520842009-04-03 13:39:52 -07001347}
1348
Lai Jiangshan2e129782010-12-22 14:18:50 +08001349/**
1350 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1351 * @q: The futex_q to unqueue
1352 *
1353 * The q->lock_ptr must not be NULL and must be held by the caller.
1354 */
1355static void __unqueue_futex(struct futex_q *q)
1356{
1357 struct futex_hash_bucket *hb;
1358
Steven Rostedt29096202011-03-17 15:21:07 -04001359 if (WARN_ON_SMP(!q->lock_ptr || !spin_is_locked(q->lock_ptr))
1360 || WARN_ON(plist_node_empty(&q->list)))
Lai Jiangshan2e129782010-12-22 14:18:50 +08001361 return;
1362
1363 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1364 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001365 hb_waiters_dec(hb);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001366}
1367
Ingo Molnarc87e2832006-06-27 02:54:58 -07001368/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 * The hash bucket lock must be held when this is called.
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001370 * Afterwards, the futex_q must not be accessed. Callers
1371 * must ensure to later call wake_up_q() for the actual
1372 * wakeups to occur.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001374static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375{
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001376 struct task_struct *p = q->task;
1377
Darren Hartaa109902012-11-26 16:29:56 -08001378 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1379 return;
1380
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001381 /*
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001382 * Queue the task for later wakeup for after we've released
1383 * the hb->lock. wake_q_add() grabs reference to p.
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001384 */
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001385 wake_q_add(wake_q, p);
Lai Jiangshan2e129782010-12-22 14:18:50 +08001386 __unqueue_futex(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 /*
Darren Hart (VMware)38fcd062017-04-14 15:31:38 -07001388 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1389 * is written, without taking any locks. This is possible in the event
1390 * of a spurious wakeup, for example. A memory barrier is required here
1391 * to prevent the following store to lock_ptr from getting ahead of the
1392 * plist_del in __unqueue_futex().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 */
Peter Zijlstra1b367ec2017-03-22 11:35:49 +01001394 smp_store_release(&q->lock_ptr, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395}
1396
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001397/*
1398 * Caller must hold a reference on @pi_state.
1399 */
1400static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
Ingo Molnarc87e2832006-06-27 02:54:58 -07001401{
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03001402 u32 uninitialized_var(curval), newval;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001403 struct task_struct *new_owner;
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001404 bool postunlock = false;
Waiman Long194a6b52016-11-17 11:46:38 -05001405 DEFINE_WAKE_Q(wake_q);
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001406 int ret = 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001407
Ingo Molnarc87e2832006-06-27 02:54:58 -07001408 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01001409 if (WARN_ON_ONCE(!new_owner)) {
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001410 /*
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01001411 * As per the comment in futex_unlock_pi() this should not happen.
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001412 *
1413 * When this happens, give up our locks and try again, giving
1414 * the futex_lock_pi() instance time to complete, either by
1415 * waiting on the rtmutex or removing itself from the futex
1416 * queue.
1417 */
1418 ret = -EAGAIN;
1419 goto out_unlock;
Peter Zijlstra73d786b2017-03-22 11:35:54 +01001420 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07001421
1422 /*
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001423 * We pass it to the next owner. The WAITERS bit is always kept
1424 * enabled while there is PI state around. We cleanup the owner
1425 * died bit, because we are the owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07001426 */
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001427 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07001428
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001429 if (unlikely(should_fail_futex(true)))
1430 ret = -EFAULT;
1431
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001432 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
Thomas Gleixner13fbca42014-06-03 12:27:07 +00001433 ret = -EFAULT;
Peter Zijlstra734009e2017-03-22 11:35:52 +01001434
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02001435 } else if (curval != uval) {
1436 /*
1437 * If a unconditional UNLOCK_PI operation (user space did not
1438 * try the TID->0 transition) raced with a waiter setting the
1439 * FUTEX_WAITERS flag between get_user() and locking the hash
1440 * bucket lock, retry the operation.
1441 */
1442 if ((FUTEX_TID_MASK & curval) == uval)
1443 ret = -EAGAIN;
1444 else
1445 ret = -EINVAL;
1446 }
Peter Zijlstra734009e2017-03-22 11:35:52 +01001447
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001448 if (ret)
1449 goto out_unlock;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001450
Peter Zijlstra94ffac52017-04-07 09:04:07 +02001451 /*
1452 * This is a point of no return; once we modify the uval there is no
1453 * going back and subsequent operations must not fail.
1454 */
1455
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001456 raw_spin_lock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001457 WARN_ON(list_empty(&pi_state->list));
1458 list_del_init(&pi_state->list);
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001459 raw_spin_unlock(&pi_state->owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001460
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001461 raw_spin_lock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001462 WARN_ON(!list_empty(&pi_state->list));
Ingo Molnarc87e2832006-06-27 02:54:58 -07001463 list_add(&pi_state->list, &new_owner->pi_state_list);
1464 pi_state->owner = new_owner;
Thomas Gleixnerb4abf912016-01-13 11:25:38 +01001465 raw_spin_unlock(&new_owner->pi_lock);
Ingo Molnar627371d2006-07-29 05:16:20 +02001466
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001467 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001468
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001469out_unlock:
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001470 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01001471
Peter Zijlstraaa2bfe52017-03-23 15:56:10 +01001472 if (postunlock)
1473 rt_mutex_postunlock(&wake_q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07001474
Peter Zijlstra16ffa122017-03-22 11:35:55 +01001475 return ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07001476}
1477
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478/*
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001479 * Express the locking dependencies for lockdep:
1480 */
1481static inline void
1482double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1483{
1484 if (hb1 <= hb2) {
1485 spin_lock(&hb1->lock);
1486 if (hb1 < hb2)
1487 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1488 } else { /* hb1 > hb2 */
1489 spin_lock(&hb2->lock);
1490 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1491 }
1492}
1493
Darren Hart5eb3dc62009-03-12 00:55:52 -07001494static inline void
1495double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1496{
Darren Hartf061d352009-03-12 15:11:18 -07001497 spin_unlock(&hb1->lock);
Ingo Molnar88f502f2009-03-13 10:32:07 +01001498 if (hb1 != hb2)
1499 spin_unlock(&hb2->lock);
Darren Hart5eb3dc62009-03-12 00:55:52 -07001500}
1501
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001502/*
Darren Hartb2d09942009-03-12 00:55:37 -07001503 * Wake up waiters matching bitset queued on this futex (uaddr).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504 */
Darren Hartb41277d2010-11-08 13:10:09 -08001505static int
1506futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507{
Ingo Molnare2970f22006-06-27 02:54:47 -07001508 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 struct futex_q *this, *next;
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001510 union futex_key key = FUTEX_KEY_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 int ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001512 DEFINE_WAKE_Q(wake_q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513
Thomas Gleixnercd689982008-02-01 17:45:14 +01001514 if (!bitset)
1515 return -EINVAL;
1516
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001517 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 if (unlikely(ret != 0))
1519 goto out;
1520
Ingo Molnare2970f22006-06-27 02:54:47 -07001521 hb = hash_futex(&key);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001522
1523 /* Make sure we really have tasks to wakeup */
1524 if (!hb_waiters_pending(hb))
1525 goto out_put_key;
1526
Ingo Molnare2970f22006-06-27 02:54:47 -07001527 spin_lock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528
Jason Low0d00c7b2014-01-12 15:31:22 -08001529 plist_for_each_entry_safe(this, next, &hb->chain, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 if (match_futex (&this->key, &key)) {
Darren Hart52400ba2009-04-03 13:40:49 -07001531 if (this->pi_state || this->rt_waiter) {
Ingo Molnared6f7b12006-07-01 04:35:46 -07001532 ret = -EINVAL;
1533 break;
1534 }
Thomas Gleixnercd689982008-02-01 17:45:14 +01001535
1536 /* Check if one of the bits is set in both bitsets */
1537 if (!(this->bitset & bitset))
1538 continue;
1539
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001540 mark_wake_futex(&wake_q, this);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 if (++ret >= nr_wake)
1542 break;
1543 }
1544 }
1545
Ingo Molnare2970f22006-06-27 02:54:47 -07001546 spin_unlock(&hb->lock);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001547 wake_up_q(&wake_q);
Davidlohr Buesob0c29f72014-01-12 15:31:25 -08001548out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001549 put_futex_key(&key);
Darren Hart42d35d42008-12-29 15:49:53 -08001550out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551 return ret;
1552}
1553
1554/*
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001555 * Wake up all waiters hashed on the physical page that is mapped
1556 * to this virtual address:
1557 */
Ingo Molnare2970f22006-06-27 02:54:47 -07001558static int
Darren Hartb41277d2010-11-08 13:10:09 -08001559futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
Ingo Molnare2970f22006-06-27 02:54:47 -07001560 int nr_wake, int nr_wake2, int op)
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001561{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001562 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Ingo Molnare2970f22006-06-27 02:54:47 -07001563 struct futex_hash_bucket *hb1, *hb2;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001564 struct futex_q *this, *next;
Darren Harte4dc5b72009-03-12 00:56:13 -07001565 int ret, op_ret;
Waiman Long194a6b52016-11-17 11:46:38 -05001566 DEFINE_WAKE_Q(wake_q);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001567
Darren Harte4dc5b72009-03-12 00:56:13 -07001568retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001569 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001570 if (unlikely(ret != 0))
1571 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001572 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001573 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001574 goto out_put_key1;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001575
Ingo Molnare2970f22006-06-27 02:54:47 -07001576 hb1 = hash_futex(&key1);
1577 hb2 = hash_futex(&key2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001578
Darren Harte4dc5b72009-03-12 00:56:13 -07001579retry_private:
Thomas Gleixnereaaea802009-10-04 09:34:17 +02001580 double_lock_hb(hb1, hb2);
Ingo Molnare2970f22006-06-27 02:54:47 -07001581 op_ret = futex_atomic_op_inuser(op, uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001582 if (unlikely(op_ret < 0)) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001583
Darren Hart5eb3dc62009-03-12 00:55:52 -07001584 double_unlock_hb(hb1, hb2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001585
David Howells7ee1dd32006-01-06 00:11:44 -08001586#ifndef CONFIG_MMU
Ingo Molnare2970f22006-06-27 02:54:47 -07001587 /*
1588 * we don't get EFAULT from MMU faults if we don't have an MMU,
1589 * but we might get them from range checking
1590 */
David Howells7ee1dd32006-01-06 00:11:44 -08001591 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001592 goto out_put_keys;
David Howells7ee1dd32006-01-06 00:11:44 -08001593#endif
1594
David Gibson796f8d92005-11-07 00:59:33 -08001595 if (unlikely(op_ret != -EFAULT)) {
1596 ret = op_ret;
Darren Hart42d35d42008-12-29 15:49:53 -08001597 goto out_put_keys;
David Gibson796f8d92005-11-07 00:59:33 -08001598 }
1599
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001600 ret = fault_in_user_writeable(uaddr2);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001601 if (ret)
Darren Hartde87fcc2009-03-12 00:55:46 -07001602 goto out_put_keys;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001603
Darren Hartb41277d2010-11-08 13:10:09 -08001604 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001605 goto retry_private;
1606
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001607 put_futex_key(&key2);
1608 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001609 goto retry;
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001610 }
1611
Jason Low0d00c7b2014-01-12 15:31:22 -08001612 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001613 if (match_futex (&this->key, &key1)) {
Darren Hartaa109902012-11-26 16:29:56 -08001614 if (this->pi_state || this->rt_waiter) {
1615 ret = -EINVAL;
1616 goto out_unlock;
1617 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001618 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001619 if (++ret >= nr_wake)
1620 break;
1621 }
1622 }
1623
1624 if (op_ret > 0) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001625 op_ret = 0;
Jason Low0d00c7b2014-01-12 15:31:22 -08001626 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001627 if (match_futex (&this->key, &key2)) {
Darren Hartaa109902012-11-26 16:29:56 -08001628 if (this->pi_state || this->rt_waiter) {
1629 ret = -EINVAL;
1630 goto out_unlock;
1631 }
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001632 mark_wake_futex(&wake_q, this);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001633 if (++op_ret >= nr_wake2)
1634 break;
1635 }
1636 }
1637 ret += op_ret;
1638 }
1639
Darren Hartaa109902012-11-26 16:29:56 -08001640out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07001641 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001642 wake_up_q(&wake_q);
Darren Hart42d35d42008-12-29 15:49:53 -08001643out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001644 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08001645out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001646 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08001647out:
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07001648 return ret;
1649}
1650
Darren Hart9121e472009-04-03 13:40:31 -07001651/**
1652 * requeue_futex() - Requeue a futex_q from one hb to another
1653 * @q: the futex_q to requeue
1654 * @hb1: the source hash_bucket
1655 * @hb2: the target hash_bucket
1656 * @key2: the new key for the requeued futex_q
1657 */
1658static inline
1659void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1660 struct futex_hash_bucket *hb2, union futex_key *key2)
1661{
1662
1663 /*
1664 * If key1 and key2 hash to the same bucket, no need to
1665 * requeue.
1666 */
1667 if (likely(&hb1->chain != &hb2->chain)) {
1668 plist_del(&q->list, &hb1->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07001669 hb_waiters_dec(hb1);
Linus Torvalds11d46162014-03-20 22:11:17 -07001670 hb_waiters_inc(hb2);
Davidlohr Buesofe1bce92016-04-20 20:09:24 -07001671 plist_add(&q->list, &hb2->chain);
Darren Hart9121e472009-04-03 13:40:31 -07001672 q->lock_ptr = &hb2->lock;
Darren Hart9121e472009-04-03 13:40:31 -07001673 }
1674 get_futex_key_refs(key2);
1675 q->key = *key2;
1676}
1677
Darren Hart52400ba2009-04-03 13:40:49 -07001678/**
1679 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
Darren Hartd96ee562009-09-21 22:30:22 -07001680 * @q: the futex_q
1681 * @key: the key of the requeue target futex
1682 * @hb: the hash_bucket of the requeue target futex
Darren Hart52400ba2009-04-03 13:40:49 -07001683 *
1684 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1685 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1686 * to the requeue target futex so the waiter can detect the wakeup on the right
1687 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
Darren Hartbeda2c72009-08-09 15:34:39 -07001688 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1689 * to protect access to the pi_state to fixup the owner later. Must be called
1690 * with both q->lock_ptr and hb->lock held.
Darren Hart52400ba2009-04-03 13:40:49 -07001691 */
1692static inline
Darren Hartbeda2c72009-08-09 15:34:39 -07001693void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1694 struct futex_hash_bucket *hb)
Darren Hart52400ba2009-04-03 13:40:49 -07001695{
Darren Hart52400ba2009-04-03 13:40:49 -07001696 get_futex_key_refs(key);
1697 q->key = *key;
1698
Lai Jiangshan2e129782010-12-22 14:18:50 +08001699 __unqueue_futex(q);
Darren Hart52400ba2009-04-03 13:40:49 -07001700
1701 WARN_ON(!q->rt_waiter);
1702 q->rt_waiter = NULL;
1703
Darren Hartbeda2c72009-08-09 15:34:39 -07001704 q->lock_ptr = &hb->lock;
Darren Hartbeda2c72009-08-09 15:34:39 -07001705
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02001706 wake_up_state(q->task, TASK_NORMAL);
Darren Hart52400ba2009-04-03 13:40:49 -07001707}
1708
1709/**
1710 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
Darren Hartbab5bc92009-04-07 23:23:50 -07001711 * @pifutex: the user address of the to futex
1712 * @hb1: the from futex hash bucket, must be locked by the caller
1713 * @hb2: the to futex hash bucket, must be locked by the caller
1714 * @key1: the from futex key
1715 * @key2: the to futex key
1716 * @ps: address to store the pi_state pointer
1717 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
Darren Hart52400ba2009-04-03 13:40:49 -07001718 *
1719 * Try and get the lock on behalf of the top waiter if we can do it atomically.
Darren Hartbab5bc92009-04-07 23:23:50 -07001720 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1721 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1722 * hb1 and hb2 must be held by the caller.
Darren Hart52400ba2009-04-03 13:40:49 -07001723 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001724 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03001725 * - 0 - failed to acquire the lock atomically;
1726 * - >0 - acquired the lock, return value is vpid of the top_waiter
1727 * - <0 - error
Darren Hart52400ba2009-04-03 13:40:49 -07001728 */
1729static int futex_proxy_trylock_atomic(u32 __user *pifutex,
1730 struct futex_hash_bucket *hb1,
1731 struct futex_hash_bucket *hb2,
1732 union futex_key *key1, union futex_key *key2,
Darren Hartbab5bc92009-04-07 23:23:50 -07001733 struct futex_pi_state **ps, int set_waiters)
Darren Hart52400ba2009-04-03 13:40:49 -07001734{
Darren Hartbab5bc92009-04-07 23:23:50 -07001735 struct futex_q *top_waiter = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07001736 u32 curval;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001737 int ret, vpid;
Darren Hart52400ba2009-04-03 13:40:49 -07001738
1739 if (get_futex_value_locked(&curval, pifutex))
1740 return -EFAULT;
1741
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07001742 if (unlikely(should_fail_futex(true)))
1743 return -EFAULT;
1744
Darren Hartbab5bc92009-04-07 23:23:50 -07001745 /*
1746 * Find the top_waiter and determine if there are additional waiters.
1747 * If the caller intends to requeue more than 1 waiter to pifutex,
1748 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1749 * as we have means to handle the possible fault. If not, don't set
1750 * the bit unecessarily as it will force the subsequent unlock to enter
1751 * the kernel.
1752 */
Darren Hart52400ba2009-04-03 13:40:49 -07001753 top_waiter = futex_top_waiter(hb1, key1);
1754
1755 /* There are no waiters, nothing for us to do. */
1756 if (!top_waiter)
1757 return 0;
1758
Darren Hart84bc4af2009-08-13 17:36:53 -07001759 /* Ensure we requeue to the expected futex. */
1760 if (!match_futex(top_waiter->requeue_pi_key, key2))
1761 return -EINVAL;
1762
Darren Hart52400ba2009-04-03 13:40:49 -07001763 /*
Darren Hartbab5bc92009-04-07 23:23:50 -07001764 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1765 * the contended case or if set_waiters is 1. The pi_state is returned
1766 * in ps in contended cases.
Darren Hart52400ba2009-04-03 13:40:49 -07001767 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001768 vpid = task_pid_vnr(top_waiter->task);
Darren Hartbab5bc92009-04-07 23:23:50 -07001769 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1770 set_waiters);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001771 if (ret == 1) {
Darren Hartbeda2c72009-08-09 15:34:39 -07001772 requeue_pi_wake_futex(top_waiter, key2, hb2);
Thomas Gleixner866293e2014-05-12 20:45:34 +00001773 return vpid;
1774 }
Darren Hart52400ba2009-04-03 13:40:49 -07001775 return ret;
1776}
1777
1778/**
1779 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
Randy Dunlapfb62db22010-10-13 11:02:34 -07001780 * @uaddr1: source futex user address
Darren Hartb41277d2010-11-08 13:10:09 -08001781 * @flags: futex flags (FLAGS_SHARED, etc.)
Randy Dunlapfb62db22010-10-13 11:02:34 -07001782 * @uaddr2: target futex user address
1783 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1784 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1785 * @cmpval: @uaddr1 expected value (or %NULL)
1786 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
Darren Hartb41277d2010-11-08 13:10:09 -08001787 * pi futex (pi to pi requeue is not supported)
Darren Hart52400ba2009-04-03 13:40:49 -07001788 *
1789 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1790 * uaddr2 atomically on behalf of the top waiter.
1791 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08001792 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03001793 * - >=0 - on success, the number of tasks requeued or woken;
1794 * - <0 - on error
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 */
Darren Hartb41277d2010-11-08 13:10:09 -08001796static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1797 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1798 u32 *cmpval, int requeue_pi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799{
Peter Zijlstra38d47c12008-09-26 19:32:20 +02001800 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
Darren Hart52400ba2009-04-03 13:40:49 -07001801 int drop_count = 0, task_count = 0, ret;
1802 struct futex_pi_state *pi_state = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07001803 struct futex_hash_bucket *hb1, *hb2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804 struct futex_q *this, *next;
Waiman Long194a6b52016-11-17 11:46:38 -05001805 DEFINE_WAKE_Q(wake_q);
Darren Hart52400ba2009-04-03 13:40:49 -07001806
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04001807 /*
1808 * When PI not supported: return -ENOSYS if requeue_pi is true,
1809 * consequently the compiler knows requeue_pi is always false past
1810 * this point which will optimize away all the conditional code
1811 * further down.
1812 */
1813 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1814 return -ENOSYS;
1815
Darren Hart52400ba2009-04-03 13:40:49 -07001816 if (requeue_pi) {
1817 /*
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001818 * Requeue PI only works on two distinct uaddrs. This
1819 * check is only valid for private futexes. See below.
1820 */
1821 if (uaddr1 == uaddr2)
1822 return -EINVAL;
1823
1824 /*
Darren Hart52400ba2009-04-03 13:40:49 -07001825 * requeue_pi requires a pi_state, try to allocate it now
1826 * without any locks in case it fails.
1827 */
1828 if (refill_pi_state_cache())
1829 return -ENOMEM;
1830 /*
1831 * requeue_pi must wake as many tasks as it can, up to nr_wake
1832 * + nr_requeue, since it acquires the rt_mutex prior to
1833 * returning to userspace, so as to not leave the rt_mutex with
1834 * waiters and no owner. However, second and third wake-ups
1835 * cannot be predicted as they involve race conditions with the
1836 * first wake and a fault while looking up the pi_state. Both
1837 * pthread_cond_signal() and pthread_cond_broadcast() should
1838 * use nr_wake=1.
1839 */
1840 if (nr_wake != 1)
1841 return -EINVAL;
1842 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843
Darren Hart42d35d42008-12-29 15:49:53 -08001844retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001845 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 if (unlikely(ret != 0))
1847 goto out;
Shawn Bohrer9ea71502011-06-30 11:21:32 -05001848 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1849 requeue_pi ? VERIFY_WRITE : VERIFY_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08001851 goto out_put_key1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00001853 /*
1854 * The check above which compares uaddrs is not sufficient for
1855 * shared futexes. We need to compare the keys:
1856 */
1857 if (requeue_pi && match_futex(&key1, &key2)) {
1858 ret = -EINVAL;
1859 goto out_put_keys;
1860 }
1861
Ingo Molnare2970f22006-06-27 02:54:47 -07001862 hb1 = hash_futex(&key1);
1863 hb2 = hash_futex(&key2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864
Darren Harte4dc5b72009-03-12 00:56:13 -07001865retry_private:
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001866 hb_waiters_inc(hb2);
Ingo Molnar8b8f3192006-07-03 00:25:05 -07001867 double_lock_hb(hb1, hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868
Ingo Molnare2970f22006-06-27 02:54:47 -07001869 if (likely(cmpval != NULL)) {
1870 u32 curval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871
Ingo Molnare2970f22006-06-27 02:54:47 -07001872 ret = get_futex_value_locked(&curval, uaddr1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873
1874 if (unlikely(ret)) {
Darren Hart5eb3dc62009-03-12 00:55:52 -07001875 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001876 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877
Darren Harte4dc5b72009-03-12 00:56:13 -07001878 ret = get_user(curval, uaddr1);
1879 if (ret)
1880 goto out_put_keys;
1881
Darren Hartb41277d2010-11-08 13:10:09 -08001882 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07001883 goto retry_private;
1884
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001885 put_futex_key(&key2);
1886 put_futex_key(&key1);
Darren Harte4dc5b72009-03-12 00:56:13 -07001887 goto retry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 }
Ingo Molnare2970f22006-06-27 02:54:47 -07001889 if (curval != *cmpval) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 ret = -EAGAIN;
1891 goto out_unlock;
1892 }
1893 }
1894
Darren Hart52400ba2009-04-03 13:40:49 -07001895 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
Darren Hartbab5bc92009-04-07 23:23:50 -07001896 /*
1897 * Attempt to acquire uaddr2 and wake the top waiter. If we
1898 * intend to requeue waiters, force setting the FUTEX_WAITERS
1899 * bit. We force this here where we are able to easily handle
1900 * faults rather in the requeue loop below.
1901 */
Darren Hart52400ba2009-04-03 13:40:49 -07001902 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
Darren Hartbab5bc92009-04-07 23:23:50 -07001903 &key2, &pi_state, nr_requeue);
Darren Hart52400ba2009-04-03 13:40:49 -07001904
1905 /*
1906 * At this point the top_waiter has either taken uaddr2 or is
1907 * waiting on it. If the former, then the pi_state will not
1908 * exist yet, look it up one more time to ensure we have a
Thomas Gleixner866293e2014-05-12 20:45:34 +00001909 * reference to it. If the lock was taken, ret contains the
1910 * vpid of the top waiter task.
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001911 * If the lock was not taken, we have pi_state and an initial
1912 * refcount on it. In case of an error we have nothing.
Darren Hart52400ba2009-04-03 13:40:49 -07001913 */
Thomas Gleixner866293e2014-05-12 20:45:34 +00001914 if (ret > 0) {
Darren Hart52400ba2009-04-03 13:40:49 -07001915 WARN_ON(pi_state);
Darren Hart89061d32009-10-15 15:30:48 -07001916 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07001917 task_count++;
Thomas Gleixner866293e2014-05-12 20:45:34 +00001918 /*
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001919 * If we acquired the lock, then the user space value
1920 * of uaddr2 should be vpid. It cannot be changed by
1921 * the top waiter as it is blocked on hb2 lock if it
1922 * tries to do so. If something fiddled with it behind
1923 * our back the pi state lookup might unearth it. So
1924 * we rather use the known value than rereading and
1925 * handing potential crap to lookup_pi_state.
1926 *
1927 * If that call succeeds then we have pi_state and an
1928 * initial refcount on it.
Thomas Gleixner866293e2014-05-12 20:45:34 +00001929 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01001930 ret = lookup_pi_state(uaddr2, ret, hb2, &key2, &pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07001931 }
1932
1933 switch (ret) {
1934 case 0:
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00001935 /* We hold a reference on the pi state. */
Darren Hart52400ba2009-04-03 13:40:49 -07001936 break;
Thomas Gleixner4959f2d2015-12-19 20:07:40 +00001937
1938 /* If the above failed, then pi_state is NULL */
Darren Hart52400ba2009-04-03 13:40:49 -07001939 case -EFAULT:
1940 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001941 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001942 put_futex_key(&key2);
1943 put_futex_key(&key1);
Thomas Gleixnerd0725992009-06-11 23:15:43 +02001944 ret = fault_in_user_writeable(uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07001945 if (!ret)
1946 goto retry;
1947 goto out;
1948 case -EAGAIN:
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00001949 /*
1950 * Two reasons for this:
1951 * - Owner is exiting and we just wait for the
1952 * exit to complete.
1953 * - The user space value changed.
1954 */
Darren Hart52400ba2009-04-03 13:40:49 -07001955 double_unlock_hb(hb1, hb2);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07001956 hb_waiters_dec(hb2);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01001957 put_futex_key(&key2);
1958 put_futex_key(&key1);
Darren Hart52400ba2009-04-03 13:40:49 -07001959 cond_resched();
1960 goto retry;
1961 default:
1962 goto out_unlock;
1963 }
1964 }
1965
Jason Low0d00c7b2014-01-12 15:31:22 -08001966 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
Darren Hart52400ba2009-04-03 13:40:49 -07001967 if (task_count - nr_wake >= nr_requeue)
1968 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969
Darren Hart52400ba2009-04-03 13:40:49 -07001970 if (!match_futex(&this->key, &key1))
1971 continue;
1972
Darren Hart392741e2009-08-07 15:20:48 -07001973 /*
1974 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
1975 * be paired with each other and no other futex ops.
Darren Hartaa109902012-11-26 16:29:56 -08001976 *
1977 * We should never be requeueing a futex_q with a pi_state,
1978 * which is awaiting a futex_unlock_pi().
Darren Hart392741e2009-08-07 15:20:48 -07001979 */
1980 if ((requeue_pi && !this->rt_waiter) ||
Darren Hartaa109902012-11-26 16:29:56 -08001981 (!requeue_pi && this->rt_waiter) ||
1982 this->pi_state) {
Darren Hart392741e2009-08-07 15:20:48 -07001983 ret = -EINVAL;
1984 break;
1985 }
Darren Hart52400ba2009-04-03 13:40:49 -07001986
1987 /*
1988 * Wake nr_wake waiters. For requeue_pi, if we acquired the
1989 * lock, we already woke the top_waiter. If not, it will be
1990 * woken by futex_unlock_pi().
1991 */
1992 if (++task_count <= nr_wake && !requeue_pi) {
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07001993 mark_wake_futex(&wake_q, this);
Darren Hart52400ba2009-04-03 13:40:49 -07001994 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 }
Darren Hart52400ba2009-04-03 13:40:49 -07001996
Darren Hart84bc4af2009-08-13 17:36:53 -07001997 /* Ensure we requeue to the expected futex for requeue_pi. */
1998 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
1999 ret = -EINVAL;
2000 break;
2001 }
2002
Darren Hart52400ba2009-04-03 13:40:49 -07002003 /*
2004 * Requeue nr_requeue waiters and possibly one more in the case
2005 * of requeue_pi if we couldn't acquire the lock atomically.
2006 */
2007 if (requeue_pi) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002008 /*
2009 * Prepare the waiter to take the rt_mutex. Take a
2010 * refcount on the pi_state and store the pointer in
2011 * the futex_q object of the waiter.
2012 */
Peter Zijlstrabf92cf32017-03-22 11:35:53 +01002013 get_pi_state(pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07002014 this->pi_state = pi_state;
2015 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2016 this->rt_waiter,
Thomas Gleixnerc051b212014-05-22 03:25:50 +00002017 this->task);
Darren Hart52400ba2009-04-03 13:40:49 -07002018 if (ret == 1) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002019 /*
2020 * We got the lock. We do neither drop the
2021 * refcount on pi_state nor clear
2022 * this->pi_state because the waiter needs the
2023 * pi_state for cleaning up the user space
2024 * value. It will drop the refcount after
2025 * doing so.
2026 */
Darren Hartbeda2c72009-08-09 15:34:39 -07002027 requeue_pi_wake_futex(this, &key2, hb2);
Darren Hart89061d32009-10-15 15:30:48 -07002028 drop_count++;
Darren Hart52400ba2009-04-03 13:40:49 -07002029 continue;
2030 } else if (ret) {
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002031 /*
2032 * rt_mutex_start_proxy_lock() detected a
2033 * potential deadlock when we tried to queue
2034 * that waiter. Drop the pi_state reference
2035 * which we took above and remove the pointer
2036 * to the state from the waiters futex_q
2037 * object.
2038 */
Darren Hart52400ba2009-04-03 13:40:49 -07002039 this->pi_state = NULL;
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002040 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002041 /*
2042 * We stop queueing more waiters and let user
2043 * space deal with the mess.
2044 */
2045 break;
Darren Hart52400ba2009-04-03 13:40:49 -07002046 }
2047 }
2048 requeue_futex(this, hb1, hb2, &key2);
2049 drop_count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 }
2051
Thomas Gleixnerecb38b72015-12-19 20:07:39 +00002052 /*
2053 * We took an extra initial reference to the pi_state either
2054 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2055 * need to drop it here again.
2056 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002057 put_pi_state(pi_state);
Thomas Gleixner885c2cb2015-12-19 20:07:41 +00002058
2059out_unlock:
Darren Hart5eb3dc62009-03-12 00:55:52 -07002060 double_unlock_hb(hb1, hb2);
Davidlohr Bueso1d0dcb32015-05-01 08:27:51 -07002061 wake_up_q(&wake_q);
Linus Torvalds69cd9eb2014-04-08 15:30:07 -07002062 hb_waiters_dec(hb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063
Darren Hartcd84a422009-04-02 14:19:38 -07002064 /*
2065 * drop_futex_key_refs() must be called outside the spinlocks. During
2066 * the requeue we moved futex_q's from the hash bucket at key1 to the
2067 * one at key2 and updated their key pointer. We no longer need to
2068 * hold the references to key1.
2069 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 while (--drop_count >= 0)
Rusty Russell9adef582007-05-08 00:26:42 -07002071 drop_futex_key_refs(&key1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072
Darren Hart42d35d42008-12-29 15:49:53 -08002073out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002074 put_futex_key(&key2);
Darren Hart42d35d42008-12-29 15:49:53 -08002075out_put_key1:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002076 put_futex_key(&key1);
Darren Hart42d35d42008-12-29 15:49:53 -08002077out:
Darren Hart52400ba2009-04-03 13:40:49 -07002078 return ret ? ret : task_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079}
2080
2081/* The key must be already stored in q->key. */
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002082static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002083 __acquires(&hb->lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084{
Ingo Molnare2970f22006-06-27 02:54:47 -07002085 struct futex_hash_bucket *hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086
Ingo Molnare2970f22006-06-27 02:54:47 -07002087 hb = hash_futex(&q->key);
Linus Torvalds11d46162014-03-20 22:11:17 -07002088
2089 /*
2090 * Increment the counter before taking the lock so that
2091 * a potential waker won't miss a to-be-slept task that is
2092 * waiting for the spinlock. This is safe as all queue_lock()
2093 * users end up calling queue_me(). Similarly, for housekeeping,
2094 * decrement the counter at queue_unlock() when some error has
2095 * occurred and we don't end up adding the task to the list.
2096 */
2097 hb_waiters_inc(hb);
2098
Ingo Molnare2970f22006-06-27 02:54:47 -07002099 q->lock_ptr = &hb->lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100
Davidlohr Bueso8ad7b372016-02-09 11:15:13 -08002101 spin_lock(&hb->lock); /* implies smp_mb(); (A) */
Ingo Molnare2970f22006-06-27 02:54:47 -07002102 return hb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103}
2104
Darren Hartd40d65c2009-09-21 22:30:15 -07002105static inline void
Jason Low0d00c7b2014-01-12 15:31:22 -08002106queue_unlock(struct futex_hash_bucket *hb)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002107 __releases(&hb->lock)
Darren Hartd40d65c2009-09-21 22:30:15 -07002108{
2109 spin_unlock(&hb->lock);
Linus Torvalds11d46162014-03-20 22:11:17 -07002110 hb_waiters_dec(hb);
Darren Hartd40d65c2009-09-21 22:30:15 -07002111}
2112
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002113static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114{
Pierre Peifferec92d082007-05-09 02:35:00 -07002115 int prio;
2116
2117 /*
2118 * The priority used to register this element is
2119 * - either the real thread-priority for the real-time threads
2120 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2121 * - or MAX_RT_PRIO for non-RT threads.
2122 * Thus, all RT-threads are woken first in priority order, and
2123 * the others are woken last, in FIFO order.
2124 */
2125 prio = min(current->normal_prio, MAX_RT_PRIO);
2126
2127 plist_node_init(&q->list, prio);
Pierre Peifferec92d082007-05-09 02:35:00 -07002128 plist_add(&q->list, &hb->chain);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002129 q->task = current;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002130}
2131
2132/**
2133 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2134 * @q: The futex_q to enqueue
2135 * @hb: The destination hash bucket
2136 *
2137 * The hb->lock must be held by the caller, and is released here. A call to
2138 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2139 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2140 * or nothing if the unqueue is done as part of the wake process and the unqueue
2141 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2142 * an example).
2143 */
2144static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2145 __releases(&hb->lock)
2146{
2147 __queue_me(q, hb);
Ingo Molnare2970f22006-06-27 02:54:47 -07002148 spin_unlock(&hb->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149}
2150
Darren Hartd40d65c2009-09-21 22:30:15 -07002151/**
2152 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2153 * @q: The futex_q to unqueue
2154 *
2155 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2156 * be paired with exactly one earlier call to queue_me().
2157 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002158 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002159 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2160 * - 0 - if the futex_q was already removed by the waking thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162static int unqueue_me(struct futex_q *q)
2163{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 spinlock_t *lock_ptr;
Ingo Molnare2970f22006-06-27 02:54:47 -07002165 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166
2167 /* In the common case we don't take the spinlock, which is nice. */
Darren Hart42d35d42008-12-29 15:49:53 -08002168retry:
Jianyu Zhan29b75eb2016-03-07 09:32:24 +08002169 /*
2170 * q->lock_ptr can change between this read and the following spin_lock.
2171 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2172 * optimizing lock_ptr out of the logic below.
2173 */
2174 lock_ptr = READ_ONCE(q->lock_ptr);
Stephen Hemmingerc80544d2007-10-18 03:07:05 -07002175 if (lock_ptr != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 spin_lock(lock_ptr);
2177 /*
2178 * q->lock_ptr can change between reading it and
2179 * spin_lock(), causing us to take the wrong lock. This
2180 * corrects the race condition.
2181 *
2182 * Reasoning goes like this: if we have the wrong lock,
2183 * q->lock_ptr must have changed (maybe several times)
2184 * between reading it and the spin_lock(). It can
2185 * change again after the spin_lock() but only if it was
2186 * already changed before the spin_lock(). It cannot,
2187 * however, change back to the original value. Therefore
2188 * we can detect whether we acquired the correct lock.
2189 */
2190 if (unlikely(lock_ptr != q->lock_ptr)) {
2191 spin_unlock(lock_ptr);
2192 goto retry;
2193 }
Lai Jiangshan2e129782010-12-22 14:18:50 +08002194 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002195
2196 BUG_ON(q->pi_state);
2197
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 spin_unlock(lock_ptr);
2199 ret = 1;
2200 }
2201
Rusty Russell9adef582007-05-08 00:26:42 -07002202 drop_futex_key_refs(&q->key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 return ret;
2204}
2205
Ingo Molnarc87e2832006-06-27 02:54:58 -07002206/*
2207 * PI futexes can not be requeued and must remove themself from the
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002208 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2209 * and dropped here.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002210 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002211static void unqueue_me_pi(struct futex_q *q)
Namhyung Kim15e408c2010-09-14 21:43:48 +09002212 __releases(q->lock_ptr)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002213{
Lai Jiangshan2e129782010-12-22 14:18:50 +08002214 __unqueue_futex(q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002215
2216 BUG_ON(!q->pi_state);
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00002217 put_pi_state(q->pi_state);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002218 q->pi_state = NULL;
2219
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002220 spin_unlock(q->lock_ptr);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002221}
2222
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002223/*
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002224 * Fixup the pi_state owner with the new owner.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002225 *
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002226 * Must be called with hash bucket lock held and mm->sem held for non
2227 * private futexes.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002228 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002229static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002230 struct task_struct *newowner)
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002231{
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002232 u32 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002233 struct futex_pi_state *pi_state = q->pi_state;
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03002234 u32 uval, uninitialized_var(curval), newval;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002235 struct task_struct *oldowner;
Darren Harte4dc5b72009-03-12 00:56:13 -07002236 int ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002237
Peter Zijlstra734009e2017-03-22 11:35:52 +01002238 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2239
2240 oldowner = pi_state->owner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002241 /* Owner died? */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002242 if (!pi_state->owner)
2243 newtid |= FUTEX_OWNER_DIED;
2244
2245 /*
2246 * We are here either because we stole the rtmutex from the
Lai Jiangshan81612392011-01-14 17:09:41 +08002247 * previous highest priority waiter or we are the highest priority
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002248 * waiter but have failed to get the rtmutex the first time.
2249 *
Lai Jiangshan81612392011-01-14 17:09:41 +08002250 * We have to replace the newowner TID in the user space variable.
2251 * This must be atomic as we have to preserve the owner died bit here.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002252 *
Darren Hartb2d09942009-03-12 00:55:37 -07002253 * Note: We write the user space value _before_ changing the pi_state
2254 * because we can fault here. Imagine swapped out pages or a fork
2255 * that marked all the anonymous memory readonly for cow.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002256 *
Peter Zijlstra734009e2017-03-22 11:35:52 +01002257 * Modifying pi_state _before_ the user space value would leave the
2258 * pi_state in an inconsistent state when we fault here, because we
2259 * need to drop the locks to handle the fault. This might be observed
2260 * in the PID check in lookup_pi_state.
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002261 */
2262retry:
2263 if (get_futex_value_locked(&uval, uaddr))
2264 goto handle_fault;
2265
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002266 for (;;) {
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002267 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2268
Michel Lespinasse37a9d912011-03-10 18:48:51 -08002269 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002270 goto handle_fault;
2271 if (curval == uval)
2272 break;
2273 uval = curval;
2274 }
2275
2276 /*
2277 * We fixed up user space. Now we need to fix the pi_state
2278 * itself.
2279 */
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002280 if (pi_state->owner != NULL) {
Peter Zijlstra734009e2017-03-22 11:35:52 +01002281 raw_spin_lock(&pi_state->owner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002282 WARN_ON(list_empty(&pi_state->list));
2283 list_del_init(&pi_state->list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002284 raw_spin_unlock(&pi_state->owner->pi_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002285 }
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002286
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002287 pi_state->owner = newowner;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002288
Peter Zijlstra734009e2017-03-22 11:35:52 +01002289 raw_spin_lock(&newowner->pi_lock);
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002290 WARN_ON(!list_empty(&pi_state->list));
Thomas Gleixnercdf71a12008-01-08 19:47:38 +01002291 list_add(&pi_state->list, &newowner->pi_state_list);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002292 raw_spin_unlock(&newowner->pi_lock);
2293 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2294
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002295 return 0;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002296
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002297 /*
Peter Zijlstra734009e2017-03-22 11:35:52 +01002298 * To handle the page fault we need to drop the locks here. That gives
2299 * the other task (either the highest priority waiter itself or the
2300 * task which stole the rtmutex) the chance to try the fixup of the
2301 * pi_state. So once we are back from handling the fault we need to
2302 * check the pi_state after reacquiring the locks and before trying to
2303 * do another fixup. When the fixup has been done already we simply
2304 * return.
2305 *
2306 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2307 * drop hb->lock since the caller owns the hb -> futex_q relation.
2308 * Dropping the pi_mutex->wait_lock requires the state revalidate.
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002309 */
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002310handle_fault:
Peter Zijlstra734009e2017-03-22 11:35:52 +01002311 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002312 spin_unlock(q->lock_ptr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002313
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002314 ret = fault_in_user_writeable(uaddr);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002315
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002316 spin_lock(q->lock_ptr);
Peter Zijlstra734009e2017-03-22 11:35:52 +01002317 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002318
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002319 /*
2320 * Check if someone else fixed it for us:
2321 */
Peter Zijlstra734009e2017-03-22 11:35:52 +01002322 if (pi_state->owner != oldowner) {
2323 ret = 0;
2324 goto out_unlock;
2325 }
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002326
2327 if (ret)
Peter Zijlstra734009e2017-03-22 11:35:52 +01002328 goto out_unlock;
Thomas Gleixner1b7558e2008-06-23 11:21:58 +02002329
2330 goto retry;
Peter Zijlstra734009e2017-03-22 11:35:52 +01002331
2332out_unlock:
2333 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2334 return ret;
Pierre Peifferd0aa7a72007-05-09 02:35:02 -07002335}
2336
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002337static long futex_wait_restart(struct restart_block *restart);
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07002338
Darren Hartca5f9522009-04-03 13:39:33 -07002339/**
Darren Hartdd973992009-04-03 13:40:02 -07002340 * fixup_owner() - Post lock pi_state and corner case management
2341 * @uaddr: user address of the futex
Darren Hartdd973992009-04-03 13:40:02 -07002342 * @q: futex_q (contains pi_state and access to the rt_mutex)
2343 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2344 *
2345 * After attempting to lock an rt_mutex, this function is called to cleanup
2346 * the pi_state owner as well as handle race conditions that may allow us to
2347 * acquire the lock. Must be called with the hb lock held.
2348 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002349 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002350 * - 1 - success, lock taken;
2351 * - 0 - success, lock not taken;
2352 * - <0 - on error (-EFAULT)
Darren Hartdd973992009-04-03 13:40:02 -07002353 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002354static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
Darren Hartdd973992009-04-03 13:40:02 -07002355{
Darren Hartdd973992009-04-03 13:40:02 -07002356 int ret = 0;
2357
2358 if (locked) {
2359 /*
2360 * Got the lock. We might not be the anticipated owner if we
2361 * did a lock-steal - fix up the PI-state in that case:
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002362 *
2363 * We can safely read pi_state->owner without holding wait_lock
2364 * because we now own the rt_mutex, only the owner will attempt
2365 * to change it.
Darren Hartdd973992009-04-03 13:40:02 -07002366 */
2367 if (q->pi_state->owner != current)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002368 ret = fixup_pi_state_owner(uaddr, q, current);
Darren Hartdd973992009-04-03 13:40:02 -07002369 goto out;
2370 }
2371
2372 /*
Darren Hartdd973992009-04-03 13:40:02 -07002373 * Paranoia check. If we did not take the lock, then we should not be
Lai Jiangshan81612392011-01-14 17:09:41 +08002374 * the owner of the rt_mutex.
Darren Hartdd973992009-04-03 13:40:02 -07002375 */
Peter Zijlstra73d786b2017-03-22 11:35:54 +01002376 if (rt_mutex_owner(&q->pi_state->pi_mutex) == current) {
Darren Hartdd973992009-04-03 13:40:02 -07002377 printk(KERN_ERR "fixup_owner: ret = %d pi-mutex: %p "
2378 "pi-state %p\n", ret,
2379 q->pi_state->pi_mutex.owner,
2380 q->pi_state->owner);
Peter Zijlstra73d786b2017-03-22 11:35:54 +01002381 }
Darren Hartdd973992009-04-03 13:40:02 -07002382
2383out:
2384 return ret ? ret : locked;
2385}
2386
2387/**
Darren Hartca5f9522009-04-03 13:39:33 -07002388 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2389 * @hb: the futex hash bucket, must be locked by the caller
2390 * @q: the futex_q to queue up on
2391 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
Darren Hartca5f9522009-04-03 13:39:33 -07002392 */
2393static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002394 struct hrtimer_sleeper *timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002395{
Darren Hart9beba3c2009-09-24 11:54:47 -07002396 /*
2397 * The task state is guaranteed to be set before another task can
Peter Zijlstrab92b8b32015-05-12 10:51:55 +02002398 * wake it. set_current_state() is implemented using smp_store_mb() and
Darren Hart9beba3c2009-09-24 11:54:47 -07002399 * queue_me() calls spin_unlock() upon completion, both serializing
2400 * access to the hash list and forcing another memory barrier.
2401 */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002402 set_current_state(TASK_INTERRUPTIBLE);
Darren Hart0729e192009-09-21 22:30:38 -07002403 queue_me(q, hb);
Darren Hartca5f9522009-04-03 13:39:33 -07002404
2405 /* Arm the timer */
Thomas Gleixner2e4b0d32015-04-14 21:09:13 +00002406 if (timeout)
Darren Hartca5f9522009-04-03 13:39:33 -07002407 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002408
2409 /*
Darren Hart0729e192009-09-21 22:30:38 -07002410 * If we have been removed from the hash list, then another task
2411 * has tried to wake us, and we can skip the call to schedule().
Darren Hartca5f9522009-04-03 13:39:33 -07002412 */
2413 if (likely(!plist_node_empty(&q->list))) {
2414 /*
2415 * If the timer has already expired, current will already be
2416 * flagged for rescheduling. Only call schedule if there
2417 * is no timeout, or if it has yet to expire.
2418 */
2419 if (!timeout || timeout->task)
Colin Cross88c80042013-05-01 18:35:05 -07002420 freezable_schedule();
Darren Hartca5f9522009-04-03 13:39:33 -07002421 }
2422 __set_current_state(TASK_RUNNING);
2423}
2424
Darren Hartf8010732009-04-03 13:40:40 -07002425/**
2426 * futex_wait_setup() - Prepare to wait on a futex
2427 * @uaddr: the futex userspace address
2428 * @val: the expected value
Darren Hartb41277d2010-11-08 13:10:09 -08002429 * @flags: futex flags (FLAGS_SHARED, etc.)
Darren Hartf8010732009-04-03 13:40:40 -07002430 * @q: the associated futex_q
2431 * @hb: storage for hash_bucket pointer to be returned to caller
2432 *
2433 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2434 * compare it with the expected value. Handle atomic faults internally.
2435 * Return with the hb lock held and a q.key reference on success, and unlocked
2436 * with no q.key reference on failure.
2437 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002438 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002439 * - 0 - uaddr contains val and hb has been locked;
2440 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
Darren Hartf8010732009-04-03 13:40:40 -07002441 */
Darren Hartb41277d2010-11-08 13:10:09 -08002442static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
Darren Hartf8010732009-04-03 13:40:40 -07002443 struct futex_q *q, struct futex_hash_bucket **hb)
2444{
2445 u32 uval;
2446 int ret;
2447
2448 /*
2449 * Access the page AFTER the hash-bucket is locked.
2450 * Order is important:
2451 *
2452 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2453 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2454 *
2455 * The basic logical guarantee of a futex is that it blocks ONLY
2456 * if cond(var) is known to be true at the time of blocking, for
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002457 * any cond. If we locked the hash-bucket after testing *uaddr, that
2458 * would open a race condition where we could block indefinitely with
Darren Hartf8010732009-04-03 13:40:40 -07002459 * cond(var) false, which would violate the guarantee.
2460 *
Michel Lespinasse8fe8f542011-03-06 18:07:50 -08002461 * On the other hand, we insert q and release the hash-bucket only
2462 * after testing *uaddr. This guarantees that futex_wait() will NOT
2463 * absorb a wakeup if *uaddr does not match the desired values
2464 * while the syscall executes.
Darren Hartf8010732009-04-03 13:40:40 -07002465 */
2466retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002467 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, VERIFY_READ);
Darren Hartf8010732009-04-03 13:40:40 -07002468 if (unlikely(ret != 0))
Darren Harta5a2a0c2009-04-10 09:50:05 -07002469 return ret;
Darren Hartf8010732009-04-03 13:40:40 -07002470
2471retry_private:
2472 *hb = queue_lock(q);
2473
2474 ret = get_futex_value_locked(&uval, uaddr);
2475
2476 if (ret) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002477 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002478
2479 ret = get_user(uval, uaddr);
2480 if (ret)
2481 goto out;
2482
Darren Hartb41277d2010-11-08 13:10:09 -08002483 if (!(flags & FLAGS_SHARED))
Darren Hartf8010732009-04-03 13:40:40 -07002484 goto retry_private;
2485
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002486 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002487 goto retry;
2488 }
2489
2490 if (uval != val) {
Jason Low0d00c7b2014-01-12 15:31:22 -08002491 queue_unlock(*hb);
Darren Hartf8010732009-04-03 13:40:40 -07002492 ret = -EWOULDBLOCK;
2493 }
2494
2495out:
2496 if (ret)
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002497 put_futex_key(&q->key);
Darren Hartf8010732009-04-03 13:40:40 -07002498 return ret;
2499}
2500
Darren Hartb41277d2010-11-08 13:10:09 -08002501static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2502 ktime_t *abs_time, u32 bitset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503{
Darren Hartca5f9522009-04-03 13:39:33 -07002504 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002505 struct restart_block *restart;
Ingo Molnare2970f22006-06-27 02:54:47 -07002506 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002507 struct futex_q q = futex_q_init;
Ingo Molnare2970f22006-06-27 02:54:47 -07002508 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
Thomas Gleixnercd689982008-02-01 17:45:14 +01002510 if (!bitset)
2511 return -EINVAL;
Thomas Gleixnercd689982008-02-01 17:45:14 +01002512 q.bitset = bitset;
Darren Hartca5f9522009-04-03 13:39:33 -07002513
2514 if (abs_time) {
2515 to = &timeout;
2516
Darren Hartb41277d2010-11-08 13:10:09 -08002517 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
2518 CLOCK_REALTIME : CLOCK_MONOTONIC,
2519 HRTIMER_MODE_ABS);
Darren Hartca5f9522009-04-03 13:39:33 -07002520 hrtimer_init_sleeper(to, current);
2521 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
2522 current->timer_slack_ns);
2523 }
2524
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002525retry:
Darren Hart7ada8762010-10-17 08:35:04 -07002526 /*
2527 * Prepare to wait on uaddr. On success, holds hb lock and increments
2528 * q.key refs.
2529 */
Darren Hartb41277d2010-11-08 13:10:09 -08002530 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Darren Hartf8010732009-04-03 13:40:40 -07002531 if (ret)
Darren Hart42d35d42008-12-29 15:49:53 -08002532 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
Darren Hartca5f9522009-04-03 13:39:33 -07002534 /* queue_me and wait for wakeup, timeout, or a signal. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02002535 futex_wait_queue_me(hb, &q, to);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536
2537 /* If we were woken (and unqueued), we succeeded, whatever. */
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002538 ret = 0;
Darren Hart7ada8762010-10-17 08:35:04 -07002539 /* unqueue_me() drops q.key ref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 if (!unqueue_me(&q))
Darren Hart7ada8762010-10-17 08:35:04 -07002541 goto out;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002542 ret = -ETIMEDOUT;
Darren Hartca5f9522009-04-03 13:39:33 -07002543 if (to && !to->task)
Darren Hart7ada8762010-10-17 08:35:04 -07002544 goto out;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002545
Ingo Molnare2970f22006-06-27 02:54:47 -07002546 /*
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002547 * We expect signal_pending(current), but we might be the
2548 * victim of a spurious wakeup as well.
Ingo Molnare2970f22006-06-27 02:54:47 -07002549 */
Darren Hart7ada8762010-10-17 08:35:04 -07002550 if (!signal_pending(current))
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002551 goto retry;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002552
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002553 ret = -ERESTARTSYS;
Pierre Peifferc19384b2007-05-09 02:35:02 -07002554 if (!abs_time)
Darren Hart7ada8762010-10-17 08:35:04 -07002555 goto out;
Steven Rostedtce6bd422007-12-05 15:46:09 +01002556
Andy Lutomirskif56141e2015-02-12 15:01:14 -08002557 restart = &current->restart_block;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002558 restart->fn = futex_wait_restart;
Namhyung Kima3c74c52010-09-14 21:43:47 +09002559 restart->futex.uaddr = uaddr;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002560 restart->futex.val = val;
Thomas Gleixner2456e852016-12-25 11:38:40 +01002561 restart->futex.time = *abs_time;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002562 restart->futex.bitset = bitset;
Darren Hart0cd9c642011-04-14 15:41:57 -07002563 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
Peter Zijlstra2fff78c2009-02-11 18:10:10 +01002564
2565 ret = -ERESTART_RESTARTBLOCK;
2566
Darren Hart42d35d42008-12-29 15:49:53 -08002567out:
Darren Hartca5f9522009-04-03 13:39:33 -07002568 if (to) {
2569 hrtimer_cancel(&to->timer);
2570 destroy_hrtimer_on_stack(&to->timer);
2571 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002572 return ret;
2573}
2574
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002575
2576static long futex_wait_restart(struct restart_block *restart)
2577{
Namhyung Kima3c74c52010-09-14 21:43:47 +09002578 u32 __user *uaddr = restart->futex.uaddr;
Darren Harta72188d2009-04-03 13:40:22 -07002579 ktime_t t, *tp = NULL;
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002580
Darren Harta72188d2009-04-03 13:40:22 -07002581 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
Thomas Gleixner2456e852016-12-25 11:38:40 +01002582 t = restart->futex.time;
Darren Harta72188d2009-04-03 13:40:22 -07002583 tp = &t;
2584 }
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002585 restart->fn = do_no_restart_syscall;
Darren Hartb41277d2010-11-08 13:10:09 -08002586
2587 return (long)futex_wait(uaddr, restart->futex.flags,
2588 restart->futex.val, tp, restart->futex.bitset);
Nick Piggin72c1bbf2007-05-08 00:26:43 -07002589}
2590
2591
Ingo Molnarc87e2832006-06-27 02:54:58 -07002592/*
2593 * Userspace tried a 0 -> TID atomic transition of the futex value
2594 * and failed. The kernel side here does the whole locking operation:
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002595 * if there are waiters then it will block as a consequence of relying
2596 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2597 * a 0 value of the futex too.).
2598 *
2599 * Also serves as futex trylock_pi()'ing, and due semantics.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002600 */
Michael Kerrisk996636d2015-01-16 20:28:06 +01002601static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
Darren Hartb41277d2010-11-08 13:10:09 -08002602 ktime_t *time, int trylock)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002603{
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002604 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002605 struct futex_pi_state *pi_state = NULL;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002606 struct rt_mutex_waiter rt_waiter;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002607 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08002608 struct futex_q q = futex_q_init;
Darren Hartdd973992009-04-03 13:40:02 -07002609 int res, ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002610
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04002611 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2612 return -ENOSYS;
2613
Ingo Molnarc87e2832006-06-27 02:54:58 -07002614 if (refill_pi_state_cache())
2615 return -ENOMEM;
2616
Pierre Peifferc19384b2007-05-09 02:35:02 -07002617 if (time) {
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002618 to = &timeout;
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002619 hrtimer_init_on_stack(&to->timer, CLOCK_REALTIME,
2620 HRTIMER_MODE_ABS);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002621 hrtimer_init_sleeper(to, current);
Arjan van de Vencc584b22008-09-01 15:02:30 -07002622 hrtimer_set_expires(&to->timer, *time);
Thomas Gleixnerc5780e92006-09-08 09:47:15 -07002623 }
2624
Darren Hart42d35d42008-12-29 15:49:53 -08002625retry:
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002626 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, VERIFY_WRITE);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002627 if (unlikely(ret != 0))
Darren Hart42d35d42008-12-29 15:49:53 -08002628 goto out;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002629
Darren Harte4dc5b72009-03-12 00:56:13 -07002630retry_private:
Eric Sesterhenn82af7ac2008-01-25 10:40:46 +01002631 hb = queue_lock(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002632
Darren Hartbab5bc92009-04-07 23:23:50 -07002633 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002634 if (unlikely(ret)) {
Davidlohr Bueso767f5092015-06-29 23:26:01 -07002635 /*
2636 * Atomic work succeeded and we got the lock,
2637 * or failed. Either way, we do _not_ block.
2638 */
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002639 switch (ret) {
Darren Hart1a520842009-04-03 13:39:52 -07002640 case 1:
2641 /* We got the lock. */
2642 ret = 0;
2643 goto out_unlock_put_key;
2644 case -EFAULT:
2645 goto uaddr_faulted;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002646 case -EAGAIN:
2647 /*
Thomas Gleixneraf54d6a2014-06-11 20:45:41 +00002648 * Two reasons for this:
2649 * - Task is exiting and we just wait for the
2650 * exit to complete.
2651 * - The user space value changed.
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002652 */
Jason Low0d00c7b2014-01-12 15:31:22 -08002653 queue_unlock(hb);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002654 put_futex_key(&q.key);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002655 cond_resched();
2656 goto retry;
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002657 default:
Darren Hart42d35d42008-12-29 15:49:53 -08002658 goto out_unlock_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002659 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002660 }
2661
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002662 WARN_ON(!q.pi_state);
2663
Ingo Molnarc87e2832006-06-27 02:54:58 -07002664 /*
2665 * Only actually queue now that the atomic ops are done:
2666 */
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002667 __queue_me(&q, hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002668
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002669 if (trylock) {
Peter Zijlstra5293c2e2017-03-22 11:35:51 +01002670 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002671 /* Fixup the trylock return value: */
2672 ret = ret ? 0 : -EWOULDBLOCK;
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002673 goto no_block;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002674 }
2675
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002676 rt_mutex_init_waiter(&rt_waiter);
Peter Zijlstra56222b22017-03-22 11:36:00 +01002677
2678 /*
2679 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2680 * hold it while doing rt_mutex_start_proxy(), because then it will
2681 * include hb->lock in the blocking chain, even through we'll not in
2682 * fact hold it while blocking. This will lead it to report -EDEADLK
2683 * and BUG when futex_unlock_pi() interleaves with this.
2684 *
2685 * Therefore acquire wait_lock while holding hb->lock, but drop the
2686 * latter before calling rt_mutex_start_proxy_lock(). This still fully
2687 * serializes against futex_unlock_pi() as that does the exact same
2688 * lock handoff sequence.
2689 */
2690 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2691 spin_unlock(q.lock_ptr);
2692 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2693 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2694
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002695 if (ret) {
2696 if (ret == 1)
2697 ret = 0;
2698
Peter Zijlstra56222b22017-03-22 11:36:00 +01002699 spin_lock(q.lock_ptr);
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002700 goto no_block;
2701 }
2702
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002703
2704 if (unlikely(to))
2705 hrtimer_start_expires(&to->timer, HRTIMER_MODE_ABS);
2706
2707 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2708
Vernon Mauerya99e4e42006-07-01 04:35:42 -07002709 spin_lock(q.lock_ptr);
Darren Hartdd973992009-04-03 13:40:02 -07002710 /*
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002711 * If we failed to acquire the lock (signal/timeout), we must
2712 * first acquire the hb->lock before removing the lock from the
2713 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex
2714 * wait lists consistent.
Peter Zijlstra56222b22017-03-22 11:36:00 +01002715 *
2716 * In particular; it is important that futex_unlock_pi() can not
2717 * observe this inconsistency.
Peter Zijlstracfafcd12017-03-22 11:35:58 +01002718 */
2719 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2720 ret = 0;
2721
2722no_block:
2723 /*
Darren Hartdd973992009-04-03 13:40:02 -07002724 * Fixup the pi_state owner and possibly acquire the lock if we
2725 * haven't already.
2726 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002727 res = fixup_owner(uaddr, &q, !ret);
Darren Hartdd973992009-04-03 13:40:02 -07002728 /*
2729 * If fixup_owner() returned an error, proprogate that. If it acquired
2730 * the lock, clear our -ETIMEDOUT or -EINTR.
2731 */
2732 if (res)
2733 ret = (res < 0) ? res : 0;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002734
Darren Harte8f63862009-03-12 00:56:06 -07002735 /*
Darren Hartdd973992009-04-03 13:40:02 -07002736 * If fixup_owner() faulted and was unable to handle the fault, unlock
2737 * it and return the fault to userspace.
Darren Harte8f63862009-03-12 00:56:06 -07002738 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002739 if (ret && (rt_mutex_owner(&q.pi_state->pi_mutex) == current)) {
2740 pi_state = q.pi_state;
2741 get_pi_state(pi_state);
2742 }
Darren Harte8f63862009-03-12 00:56:06 -07002743
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002744 /* Unqueue and drop the lock */
2745 unqueue_me_pi(&q);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002746
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002747 if (pi_state) {
2748 rt_mutex_futex_unlock(&pi_state->pi_mutex);
2749 put_pi_state(pi_state);
2750 }
2751
Mikael Pettersson5ecb01c2010-01-23 22:36:29 +01002752 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002753
Darren Hart42d35d42008-12-29 15:49:53 -08002754out_unlock_put_key:
Jason Low0d00c7b2014-01-12 15:31:22 -08002755 queue_unlock(hb);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002756
Darren Hart42d35d42008-12-29 15:49:53 -08002757out_put_key:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002758 put_futex_key(&q.key);
Darren Hart42d35d42008-12-29 15:49:53 -08002759out:
Thomas Gleixner97181f92017-04-10 18:03:36 +02002760 if (to) {
2761 hrtimer_cancel(&to->timer);
Thomas Gleixner237fc6e2008-04-30 00:55:04 -07002762 destroy_hrtimer_on_stack(&to->timer);
Thomas Gleixner97181f92017-04-10 18:03:36 +02002763 }
Darren Hartdd973992009-04-03 13:40:02 -07002764 return ret != -EINTR ? ret : -ERESTARTNOINTR;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002765
Darren Hart42d35d42008-12-29 15:49:53 -08002766uaddr_faulted:
Jason Low0d00c7b2014-01-12 15:31:22 -08002767 queue_unlock(hb);
Alexey Kuznetsov778e9a92007-06-08 13:47:00 -07002768
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002769 ret = fault_in_user_writeable(uaddr);
Darren Harte4dc5b72009-03-12 00:56:13 -07002770 if (ret)
2771 goto out_put_key;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002772
Darren Hartb41277d2010-11-08 13:10:09 -08002773 if (!(flags & FLAGS_SHARED))
Darren Harte4dc5b72009-03-12 00:56:13 -07002774 goto retry_private;
2775
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002776 put_futex_key(&q.key);
Darren Harte4dc5b72009-03-12 00:56:13 -07002777 goto retry;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002778}
2779
2780/*
Ingo Molnarc87e2832006-06-27 02:54:58 -07002781 * Userspace attempted a TID -> 0 atomic transition, and failed.
2782 * This is the in-kernel slowpath: we look up the PI state (if any),
2783 * and do the rt-mutex unlock.
2784 */
Darren Hartb41277d2010-11-08 13:10:09 -08002785static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002786{
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002787 u32 uninitialized_var(curval), uval, vpid = task_pid_vnr(current);
Peter Zijlstra38d47c12008-09-26 19:32:20 +02002788 union futex_key key = FUTEX_KEY_INIT;
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002789 struct futex_hash_bucket *hb;
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002790 struct futex_q *top_waiter;
Darren Harte4dc5b72009-03-12 00:56:13 -07002791 int ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002792
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04002793 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2794 return -ENOSYS;
2795
Ingo Molnarc87e2832006-06-27 02:54:58 -07002796retry:
2797 if (get_user(uval, uaddr))
2798 return -EFAULT;
2799 /*
2800 * We release only a lock we actually own:
2801 */
Thomas Gleixnerc0c9ed12011-03-11 11:51:22 +01002802 if ((uval & FUTEX_TID_MASK) != vpid)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002803 return -EPERM;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002804
Shawn Bohrer9ea71502011-06-30 11:21:32 -05002805 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, VERIFY_WRITE);
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002806 if (ret)
2807 return ret;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002808
2809 hb = hash_futex(&key);
2810 spin_lock(&hb->lock);
2811
Ingo Molnarc87e2832006-06-27 02:54:58 -07002812 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002813 * Check waiters first. We do not trust user space values at
2814 * all and we at least want to know if user space fiddled
2815 * with the futex value instead of blindly unlocking.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002816 */
Peter Zijlstra499f5ac2017-03-22 11:35:48 +01002817 top_waiter = futex_top_waiter(hb, &key);
2818 if (top_waiter) {
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002819 struct futex_pi_state *pi_state = top_waiter->pi_state;
2820
2821 ret = -EINVAL;
2822 if (!pi_state)
2823 goto out_unlock;
2824
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002825 /*
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002826 * If current does not own the pi_state then the futex is
2827 * inconsistent and user space fiddled with the futex value.
2828 */
2829 if (pi_state->owner != current)
2830 goto out_unlock;
2831
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002832 get_pi_state(pi_state);
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01002833 /*
Peter Zijlstrabebe5b52017-03-22 11:35:59 +01002834 * By taking wait_lock while still holding hb->lock, we ensure
2835 * there is no point where we hold neither; and therefore
2836 * wake_futex_pi() must observe a state consistent with what we
2837 * observed.
2838 */
2839 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002840 spin_unlock(&hb->lock);
2841
2842 ret = wake_futex_pi(uaddr, uval, pi_state);
2843
2844 put_pi_state(pi_state);
2845
2846 /*
2847 * Success, we're done! No tricky corner cases.
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002848 */
2849 if (!ret)
2850 goto out_putkey;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002851 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002852 * The atomic access to the futex value generated a
2853 * pagefault, so retry the user-access and the wakeup:
Ingo Molnarc87e2832006-06-27 02:54:58 -07002854 */
2855 if (ret == -EFAULT)
2856 goto pi_faulted;
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002857 /*
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02002858 * A unconditional UNLOCK_PI op raced against a waiter
2859 * setting the FUTEX_WAITERS bit. Try again.
2860 */
2861 if (ret == -EAGAIN) {
Sebastian Andrzej Siewior89e9e662016-04-15 14:35:39 +02002862 put_futex_key(&key);
2863 goto retry;
2864 }
2865 /*
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002866 * wake_futex_pi has detected invalid state. Tell user
2867 * space.
2868 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002869 goto out_putkey;
Ingo Molnarc87e2832006-06-27 02:54:58 -07002870 }
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002871
Ingo Molnarc87e2832006-06-27 02:54:58 -07002872 /*
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002873 * We have no kernel internal state, i.e. no waiters in the
2874 * kernel. Waiters which are about to queue themselves are stuck
2875 * on hb->lock. So we can safely ignore them. We do neither
2876 * preserve the WAITERS bit not the OWNER_DIED one. We are the
2877 * owner.
Ingo Molnarc87e2832006-06-27 02:54:58 -07002878 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002879 if (cmpxchg_futex_value_locked(&curval, uaddr, uval, 0)) {
2880 spin_unlock(&hb->lock);
Thomas Gleixner13fbca42014-06-03 12:27:07 +00002881 goto pi_faulted;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002882 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07002883
Thomas Gleixnerccf9e6a2014-06-11 20:45:38 +00002884 /*
2885 * If uval has changed, let user space handle it.
2886 */
2887 ret = (curval == uval) ? 0 : -EAGAIN;
2888
Ingo Molnarc87e2832006-06-27 02:54:58 -07002889out_unlock:
2890 spin_unlock(&hb->lock);
Sebastian Andrzej Siewior802ab582015-06-17 10:33:50 +02002891out_putkey:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002892 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002893 return ret;
2894
2895pi_faulted:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01002896 put_futex_key(&key);
Ingo Molnarc87e2832006-06-27 02:54:58 -07002897
Thomas Gleixnerd0725992009-06-11 23:15:43 +02002898 ret = fault_in_user_writeable(uaddr);
Darren Hartb5686362008-12-18 15:06:34 -08002899 if (!ret)
Ingo Molnarc87e2832006-06-27 02:54:58 -07002900 goto retry;
2901
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 return ret;
2903}
2904
Darren Hart52400ba2009-04-03 13:40:49 -07002905/**
2906 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
2907 * @hb: the hash_bucket futex_q was original enqueued on
2908 * @q: the futex_q woken while waiting to be requeued
2909 * @key2: the futex_key of the requeue target futex
2910 * @timeout: the timeout associated with the wait (NULL if none)
2911 *
2912 * Detect if the task was woken on the initial futex as opposed to the requeue
2913 * target futex. If so, determine if it was a timeout or a signal that caused
2914 * the wakeup and return the appropriate error code to the caller. Must be
2915 * called with the hb lock held.
2916 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002917 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002918 * - 0 = no early wakeup detected;
2919 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
Darren Hart52400ba2009-04-03 13:40:49 -07002920 */
2921static inline
2922int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
2923 struct futex_q *q, union futex_key *key2,
2924 struct hrtimer_sleeper *timeout)
2925{
2926 int ret = 0;
2927
2928 /*
2929 * With the hb lock held, we avoid races while we process the wakeup.
2930 * We only need to hold hb (and not hb2) to ensure atomicity as the
2931 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
2932 * It can't be requeued from uaddr2 to something else since we don't
2933 * support a PI aware source futex for requeue.
2934 */
2935 if (!match_futex(&q->key, key2)) {
2936 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
2937 /*
2938 * We were woken prior to requeue by a timeout or a signal.
2939 * Unqueue the futex_q and determine which it was.
2940 */
Lai Jiangshan2e129782010-12-22 14:18:50 +08002941 plist_del(&q->list, &hb->chain);
Linus Torvalds11d46162014-03-20 22:11:17 -07002942 hb_waiters_dec(hb);
Darren Hart52400ba2009-04-03 13:40:49 -07002943
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002944 /* Handle spurious wakeups gracefully */
Thomas Gleixner11df6dd2009-10-28 20:26:48 +01002945 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07002946 if (timeout && !timeout->task)
2947 ret = -ETIMEDOUT;
Thomas Gleixnerd58e6572009-10-13 20:40:43 +02002948 else if (signal_pending(current))
Thomas Gleixner1c840c12009-05-20 09:22:40 +02002949 ret = -ERESTARTNOINTR;
Darren Hart52400ba2009-04-03 13:40:49 -07002950 }
2951 return ret;
2952}
2953
2954/**
2955 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
Darren Hart56ec1602009-09-21 22:29:59 -07002956 * @uaddr: the futex we initially wait on (non-pi)
Darren Hartb41277d2010-11-08 13:10:09 -08002957 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07002958 * the same type, no requeueing from private to shared, etc.
Darren Hart52400ba2009-04-03 13:40:49 -07002959 * @val: the expected value of uaddr
2960 * @abs_time: absolute timeout
Darren Hart56ec1602009-09-21 22:29:59 -07002961 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
Darren Hart52400ba2009-04-03 13:40:49 -07002962 * @uaddr2: the pi futex we will take prior to returning to user-space
2963 *
2964 * The caller will wait on uaddr and will be requeued by futex_requeue() to
Darren Hart6f7b0a22012-07-20 11:53:31 -07002965 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
2966 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
2967 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
2968 * without one, the pi logic would not know which task to boost/deboost, if
2969 * there was a need to.
Darren Hart52400ba2009-04-03 13:40:49 -07002970 *
2971 * We call schedule in futex_wait_queue_me() when we enqueue and return there
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002972 * via the following--
Darren Hart52400ba2009-04-03 13:40:49 -07002973 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
Darren Hartcc6db4e2009-07-31 16:20:10 -07002974 * 2) wakeup on uaddr2 after a requeue
2975 * 3) signal
2976 * 4) timeout
Darren Hart52400ba2009-04-03 13:40:49 -07002977 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002978 * If 3, cleanup and return -ERESTARTNOINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07002979 *
2980 * If 2, we may then block on trying to take the rt_mutex and return via:
2981 * 5) successful lock
2982 * 6) signal
2983 * 7) timeout
2984 * 8) other lock acquisition failure
2985 *
Darren Hartcc6db4e2009-07-31 16:20:10 -07002986 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
Darren Hart52400ba2009-04-03 13:40:49 -07002987 *
2988 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
2989 *
Randy Dunlap6c23cbb2013-03-05 10:00:24 -08002990 * Return:
Mauro Carvalho Chehab7b4ff1a2017-05-11 10:17:45 -03002991 * - 0 - On success;
2992 * - <0 - On error
Darren Hart52400ba2009-04-03 13:40:49 -07002993 */
Darren Hartb41277d2010-11-08 13:10:09 -08002994static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
Darren Hart52400ba2009-04-03 13:40:49 -07002995 u32 val, ktime_t *abs_time, u32 bitset,
Darren Hartb41277d2010-11-08 13:10:09 -08002996 u32 __user *uaddr2)
Darren Hart52400ba2009-04-03 13:40:49 -07002997{
2998 struct hrtimer_sleeper timeout, *to = NULL;
Peter Zijlstra16ffa122017-03-22 11:35:55 +01002999 struct futex_pi_state *pi_state = NULL;
Darren Hart52400ba2009-04-03 13:40:49 -07003000 struct rt_mutex_waiter rt_waiter;
Darren Hart52400ba2009-04-03 13:40:49 -07003001 struct futex_hash_bucket *hb;
Darren Hart5bdb05f2010-11-08 13:40:28 -08003002 union futex_key key2 = FUTEX_KEY_INIT;
3003 struct futex_q q = futex_q_init;
Darren Hart52400ba2009-04-03 13:40:49 -07003004 int res, ret;
Darren Hart52400ba2009-04-03 13:40:49 -07003005
Nicolas Pitrebc2eecd2017-08-01 00:31:32 -04003006 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3007 return -ENOSYS;
3008
Darren Hart6f7b0a22012-07-20 11:53:31 -07003009 if (uaddr == uaddr2)
3010 return -EINVAL;
3011
Darren Hart52400ba2009-04-03 13:40:49 -07003012 if (!bitset)
3013 return -EINVAL;
3014
3015 if (abs_time) {
3016 to = &timeout;
Darren Hartb41277d2010-11-08 13:10:09 -08003017 hrtimer_init_on_stack(&to->timer, (flags & FLAGS_CLOCKRT) ?
3018 CLOCK_REALTIME : CLOCK_MONOTONIC,
3019 HRTIMER_MODE_ABS);
Darren Hart52400ba2009-04-03 13:40:49 -07003020 hrtimer_init_sleeper(to, current);
3021 hrtimer_set_expires_range_ns(&to->timer, *abs_time,
3022 current->timer_slack_ns);
3023 }
3024
3025 /*
3026 * The waiter is allocated on our stack, manipulated by the requeue
3027 * code while we sleep on uaddr.
3028 */
Peter Zijlstra50809352017-03-22 11:35:56 +01003029 rt_mutex_init_waiter(&rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003030
Shawn Bohrer9ea71502011-06-30 11:21:32 -05003031 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, VERIFY_WRITE);
Darren Hart52400ba2009-04-03 13:40:49 -07003032 if (unlikely(ret != 0))
3033 goto out;
3034
Darren Hart84bc4af2009-08-13 17:36:53 -07003035 q.bitset = bitset;
3036 q.rt_waiter = &rt_waiter;
3037 q.requeue_pi_key = &key2;
3038
Darren Hart7ada8762010-10-17 08:35:04 -07003039 /*
3040 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3041 * count.
3042 */
Darren Hartb41277d2010-11-08 13:10:09 -08003043 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02003044 if (ret)
3045 goto out_key2;
Darren Hart52400ba2009-04-03 13:40:49 -07003046
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00003047 /*
3048 * The check above which compares uaddrs is not sufficient for
3049 * shared futexes. We need to compare the keys:
3050 */
3051 if (match_futex(&q.key, &key2)) {
Thomas Gleixner13c42c22014-09-11 23:44:35 +02003052 queue_unlock(hb);
Thomas Gleixnere9c243a2014-06-03 12:27:06 +00003053 ret = -EINVAL;
3054 goto out_put_keys;
3055 }
3056
Darren Hart52400ba2009-04-03 13:40:49 -07003057 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
Thomas Gleixnerf1a11e02009-05-05 19:21:40 +02003058 futex_wait_queue_me(hb, &q, to);
Darren Hart52400ba2009-04-03 13:40:49 -07003059
3060 spin_lock(&hb->lock);
3061 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3062 spin_unlock(&hb->lock);
3063 if (ret)
3064 goto out_put_keys;
3065
3066 /*
3067 * In order for us to be here, we know our q.key == key2, and since
3068 * we took the hb->lock above, we also know that futex_requeue() has
3069 * completed and we no longer have to concern ourselves with a wakeup
Darren Hart7ada8762010-10-17 08:35:04 -07003070 * race with the atomic proxy lock acquisition by the requeue code. The
3071 * futex_requeue dropped our key1 reference and incremented our key2
3072 * reference count.
Darren Hart52400ba2009-04-03 13:40:49 -07003073 */
3074
3075 /* Check if the requeue code acquired the second futex for us. */
3076 if (!q.rt_waiter) {
3077 /*
3078 * Got the lock. We might not be the anticipated owner if we
3079 * did a lock-steal - fix up the PI-state in that case.
3080 */
3081 if (q.pi_state && (q.pi_state->owner != current)) {
3082 spin_lock(q.lock_ptr);
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003083 ret = fixup_pi_state_owner(uaddr2, &q, current);
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003084 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3085 pi_state = q.pi_state;
3086 get_pi_state(pi_state);
3087 }
Thomas Gleixnerfb75a422015-12-19 20:07:38 +00003088 /*
3089 * Drop the reference to the pi state which
3090 * the requeue_pi() code acquired for us.
3091 */
Thomas Gleixner29e9ee52015-12-19 20:07:39 +00003092 put_pi_state(q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003093 spin_unlock(q.lock_ptr);
3094 }
3095 } else {
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003096 struct rt_mutex *pi_mutex;
3097
Darren Hart52400ba2009-04-03 13:40:49 -07003098 /*
3099 * We have been woken up by futex_unlock_pi(), a timeout, or a
3100 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3101 * the pi_state.
3102 */
Darren Hartf27071c2012-07-20 11:53:30 -07003103 WARN_ON(!q.pi_state);
Darren Hart52400ba2009-04-03 13:40:49 -07003104 pi_mutex = &q.pi_state->pi_mutex;
Peter Zijlstra38d589f2017-03-22 11:35:57 +01003105 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003106
3107 spin_lock(q.lock_ptr);
Peter Zijlstra38d589f2017-03-22 11:35:57 +01003108 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3109 ret = 0;
3110
3111 debug_rt_mutex_free_waiter(&rt_waiter);
Darren Hart52400ba2009-04-03 13:40:49 -07003112 /*
3113 * Fixup the pi_state owner and possibly acquire the lock if we
3114 * haven't already.
3115 */
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003116 res = fixup_owner(uaddr2, &q, !ret);
Darren Hart52400ba2009-04-03 13:40:49 -07003117 /*
3118 * If fixup_owner() returned an error, proprogate that. If it
Darren Hart56ec1602009-09-21 22:29:59 -07003119 * acquired the lock, clear -ETIMEDOUT or -EINTR.
Darren Hart52400ba2009-04-03 13:40:49 -07003120 */
3121 if (res)
3122 ret = (res < 0) ? res : 0;
3123
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003124 /*
3125 * If fixup_pi_state_owner() faulted and was unable to handle
3126 * the fault, unlock the rt_mutex and return the fault to
3127 * userspace.
3128 */
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003129 if (ret && rt_mutex_owner(&q.pi_state->pi_mutex) == current) {
3130 pi_state = q.pi_state;
3131 get_pi_state(pi_state);
3132 }
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003133
Darren Hart52400ba2009-04-03 13:40:49 -07003134 /* Unqueue and drop the lock. */
3135 unqueue_me_pi(&q);
3136 }
3137
Peter Zijlstra16ffa122017-03-22 11:35:55 +01003138 if (pi_state) {
3139 rt_mutex_futex_unlock(&pi_state->pi_mutex);
3140 put_pi_state(pi_state);
3141 }
3142
Peter Zijlstrac236c8e2017-03-04 10:27:18 +01003143 if (ret == -EINTR) {
Darren Hart52400ba2009-04-03 13:40:49 -07003144 /*
Darren Hartcc6db4e2009-07-31 16:20:10 -07003145 * We've already been requeued, but cannot restart by calling
3146 * futex_lock_pi() directly. We could restart this syscall, but
3147 * it would detect that the user space "val" changed and return
3148 * -EWOULDBLOCK. Save the overhead of the restart and return
3149 * -EWOULDBLOCK directly.
Darren Hart52400ba2009-04-03 13:40:49 -07003150 */
Thomas Gleixner20708872009-05-19 23:04:59 +02003151 ret = -EWOULDBLOCK;
Darren Hart52400ba2009-04-03 13:40:49 -07003152 }
3153
3154out_put_keys:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003155 put_futex_key(&q.key);
Thomas Gleixnerc8b15a72009-05-20 09:18:50 +02003156out_key2:
Thomas Gleixnerae791a22010-11-10 13:30:36 +01003157 put_futex_key(&key2);
Darren Hart52400ba2009-04-03 13:40:49 -07003158
3159out:
3160 if (to) {
3161 hrtimer_cancel(&to->timer);
3162 destroy_hrtimer_on_stack(&to->timer);
3163 }
3164 return ret;
3165}
3166
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003167/*
3168 * Support for robust futexes: the kernel cleans up held futexes at
3169 * thread exit time.
3170 *
3171 * Implementation: user-space maintains a per-thread list of locks it
3172 * is holding. Upon do_exit(), the kernel carefully walks this list,
3173 * and marks all locks that are owned by this thread with the
Ingo Molnarc87e2832006-06-27 02:54:58 -07003174 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003175 * always manipulated with the lock held, so the list is private and
3176 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3177 * field, to allow the kernel to clean up if the thread dies after
3178 * acquiring the lock, but just before it could have added itself to
3179 * the list. There can only be one such pending lock.
3180 */
3181
3182/**
Darren Hartd96ee562009-09-21 22:30:22 -07003183 * sys_set_robust_list() - Set the robust-futex list head of a task
3184 * @head: pointer to the list-head
3185 * @len: length of the list-head, as userspace expects
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003186 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003187SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3188 size_t, len)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003189{
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003190 if (!futex_cmpxchg_enabled)
3191 return -ENOSYS;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003192 /*
3193 * The kernel knows only one size for now:
3194 */
3195 if (unlikely(len != sizeof(*head)))
3196 return -EINVAL;
3197
3198 current->robust_list = head;
3199
3200 return 0;
3201}
3202
3203/**
Darren Hartd96ee562009-09-21 22:30:22 -07003204 * sys_get_robust_list() - Get the robust-futex list head of a task
3205 * @pid: pid of the process [zero for current task]
3206 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3207 * @len_ptr: pointer to a length field, the kernel fills in the header size
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003208 */
Heiko Carstens836f92a2009-01-14 14:14:33 +01003209SYSCALL_DEFINE3(get_robust_list, int, pid,
3210 struct robust_list_head __user * __user *, head_ptr,
3211 size_t __user *, len_ptr)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003212{
Al Viroba46df92006-10-10 22:46:07 +01003213 struct robust_list_head __user *head;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003214 unsigned long ret;
Kees Cookbdbb7762012-03-19 16:12:53 -07003215 struct task_struct *p;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003216
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003217 if (!futex_cmpxchg_enabled)
3218 return -ENOSYS;
3219
Kees Cookbdbb7762012-03-19 16:12:53 -07003220 rcu_read_lock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003221
Kees Cookbdbb7762012-03-19 16:12:53 -07003222 ret = -ESRCH;
3223 if (!pid)
3224 p = current;
3225 else {
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07003226 p = find_task_by_vpid(pid);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003227 if (!p)
3228 goto err_unlock;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003229 }
3230
Kees Cookbdbb7762012-03-19 16:12:53 -07003231 ret = -EPERM;
Jann Horncaaee622016-01-20 15:00:04 -08003232 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
Kees Cookbdbb7762012-03-19 16:12:53 -07003233 goto err_unlock;
3234
3235 head = p->robust_list;
3236 rcu_read_unlock();
3237
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003238 if (put_user(sizeof(*head), len_ptr))
3239 return -EFAULT;
3240 return put_user(head, head_ptr);
3241
3242err_unlock:
Oleg Nesterovaaa2a972006-09-29 02:00:55 -07003243 rcu_read_unlock();
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003244
3245 return ret;
3246}
3247
3248/*
3249 * Process a futex-list entry, check whether it's owned by the
3250 * dying task, and do notification if so:
3251 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003252int handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003253{
Vitaliy Ivanov7cfdaf32011-07-07 15:10:31 +03003254 u32 uval, uninitialized_var(nval), mval;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003255
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003256retry:
3257 if (get_user(uval, uaddr))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003258 return -1;
3259
Pavel Emelyanovb4888932007-10-18 23:40:14 -07003260 if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003261 /*
3262 * Ok, this dying thread is truly holding a futex
3263 * of interest. Set the OWNER_DIED bit atomically
3264 * via cmpxchg, and if the value had FUTEX_WAITERS
3265 * set, wake up a waiter (if any). (We have to do a
3266 * futex_wake() even if OWNER_DIED is already set -
3267 * to handle the rare but possible case of recursive
3268 * thread-death.) The rest of the cleanup is done in
3269 * userspace.
3270 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003271 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
Thomas Gleixner6e0aa9f2011-03-14 10:34:35 +01003272 /*
3273 * We are not holding a lock here, but we want to have
3274 * the pagefault_disable/enable() protection because
3275 * we want to handle the fault gracefully. If the
3276 * access fails we try to fault in the futex with R/W
3277 * verification via get_user_pages. get_user() above
3278 * does not guarantee R/W access. If that fails we
3279 * give up and leave the futex locked.
3280 */
3281 if (cmpxchg_futex_value_locked(&nval, uaddr, uval, mval)) {
3282 if (fault_in_user_writeable(uaddr))
3283 return -1;
3284 goto retry;
3285 }
Ingo Molnarc87e2832006-06-27 02:54:58 -07003286 if (nval != uval)
Ingo Molnar8f17d3a2006-03-27 01:16:27 -08003287 goto retry;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003288
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003289 /*
3290 * Wake robust non-PI futexes here. The wakeup of
3291 * PI futexes happens in exit_pi_state():
3292 */
Thomas Gleixner36cf3b52007-07-15 23:41:20 -07003293 if (!pi && (uval & FUTEX_WAITERS))
Peter Zijlstrac2f9f202008-09-26 19:32:23 +02003294 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003295 }
3296 return 0;
3297}
3298
3299/*
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003300 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3301 */
3302static inline int fetch_robust_entry(struct robust_list __user **entry,
Al Viroba46df92006-10-10 22:46:07 +01003303 struct robust_list __user * __user *head,
Namhyung Kim1dcc41b2010-09-14 21:43:46 +09003304 unsigned int *pi)
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003305{
3306 unsigned long uentry;
3307
Al Viroba46df92006-10-10 22:46:07 +01003308 if (get_user(uentry, (unsigned long __user *)head))
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003309 return -EFAULT;
3310
Al Viroba46df92006-10-10 22:46:07 +01003311 *entry = (void __user *)(uentry & ~1UL);
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003312 *pi = uentry & 1;
3313
3314 return 0;
3315}
3316
3317/*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003318 * Walk curr->robust_list (very carefully, it's a userspace list!)
3319 * and mark any locks found there dead, and notify any waiters.
3320 *
3321 * We silently return on any sign of list-walking problem.
3322 */
3323void exit_robust_list(struct task_struct *curr)
3324{
3325 struct robust_list_head __user *head = curr->robust_list;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003326 struct robust_list __user *entry, *next_entry, *pending;
Darren Hart4c115e92010-11-04 15:00:00 -04003327 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3328 unsigned int uninitialized_var(next_pi);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003329 unsigned long futex_offset;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003330 int rc;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003331
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003332 if (!futex_cmpxchg_enabled)
3333 return;
3334
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003335 /*
3336 * Fetch the list head (which was registered earlier, via
3337 * sys_set_robust_list()):
3338 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003339 if (fetch_robust_entry(&entry, &head->list.next, &pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003340 return;
3341 /*
3342 * Fetch the relative futex offset:
3343 */
3344 if (get_user(futex_offset, &head->futex_offset))
3345 return;
3346 /*
3347 * Fetch any possibly pending lock-add first, and handle it
3348 * if it exists:
3349 */
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003350 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003351 return;
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003352
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003353 next_entry = NULL; /* avoid warning with gcc */
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003354 while (entry != &head->list) {
3355 /*
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003356 * Fetch the next entry in the list before calling
3357 * handle_futex_death:
3358 */
3359 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3360 /*
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003361 * A pending lock might already be on the list, so
Ingo Molnarc87e2832006-06-27 02:54:58 -07003362 * don't process it twice:
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003363 */
3364 if (entry != pending)
Al Viroba46df92006-10-10 22:46:07 +01003365 if (handle_futex_death((void __user *)entry + futex_offset,
Ingo Molnare3f2dde2006-07-29 05:17:57 +02003366 curr, pi))
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003367 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003368 if (rc)
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003369 return;
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003370 entry = next_entry;
3371 pi = next_pi;
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003372 /*
3373 * Avoid excessively long or circular lists:
3374 */
3375 if (!--limit)
3376 break;
3377
3378 cond_resched();
3379 }
Martin Schwidefsky9f96cb12007-10-01 01:20:13 -07003380
3381 if (pending)
3382 handle_futex_death((void __user *)pending + futex_offset,
3383 curr, pip);
Ingo Molnar0771dfe2006-03-27 01:16:22 -08003384}
3385
Pierre Peifferc19384b2007-05-09 02:35:02 -07003386long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
Ingo Molnare2970f22006-06-27 02:54:47 -07003387 u32 __user *uaddr2, u32 val2, u32 val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388{
Thomas Gleixner81b40532012-02-15 12:17:09 +01003389 int cmd = op & FUTEX_CMD_MASK;
Darren Hartb41277d2010-11-08 13:10:09 -08003390 unsigned int flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003392 if (!(op & FUTEX_PRIVATE_FLAG))
Darren Hartb41277d2010-11-08 13:10:09 -08003393 flags |= FLAGS_SHARED;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003394
Darren Hartb41277d2010-11-08 13:10:09 -08003395 if (op & FUTEX_CLOCK_REALTIME) {
3396 flags |= FLAGS_CLOCKRT;
Darren Hart337f1302015-12-18 13:36:37 -08003397 if (cmd != FUTEX_WAIT && cmd != FUTEX_WAIT_BITSET && \
3398 cmd != FUTEX_WAIT_REQUEUE_PI)
Darren Hartb41277d2010-11-08 13:10:09 -08003399 return -ENOSYS;
3400 }
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003401
3402 switch (cmd) {
Thomas Gleixner59263b52012-02-15 12:08:34 +01003403 case FUTEX_LOCK_PI:
3404 case FUTEX_UNLOCK_PI:
3405 case FUTEX_TRYLOCK_PI:
3406 case FUTEX_WAIT_REQUEUE_PI:
3407 case FUTEX_CMP_REQUEUE_PI:
3408 if (!futex_cmpxchg_enabled)
3409 return -ENOSYS;
3410 }
3411
3412 switch (cmd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 case FUTEX_WAIT:
Thomas Gleixnercd689982008-02-01 17:45:14 +01003414 val3 = FUTEX_BITSET_MATCH_ANY;
3415 case FUTEX_WAIT_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003416 return futex_wait(uaddr, flags, val, timeout, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417 case FUTEX_WAKE:
Thomas Gleixnercd689982008-02-01 17:45:14 +01003418 val3 = FUTEX_BITSET_MATCH_ANY;
3419 case FUTEX_WAKE_BITSET:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003420 return futex_wake(uaddr, flags, val, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 case FUTEX_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003422 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423 case FUTEX_CMP_REQUEUE:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003424 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
Jakub Jelinek4732efbe2005-09-06 15:16:25 -07003425 case FUTEX_WAKE_OP:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003426 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003427 case FUTEX_LOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003428 return futex_lock_pi(uaddr, flags, timeout, 0);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003429 case FUTEX_UNLOCK_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003430 return futex_unlock_pi(uaddr, flags);
Ingo Molnarc87e2832006-06-27 02:54:58 -07003431 case FUTEX_TRYLOCK_PI:
Michael Kerrisk996636d2015-01-16 20:28:06 +01003432 return futex_lock_pi(uaddr, flags, NULL, 1);
Darren Hart52400ba2009-04-03 13:40:49 -07003433 case FUTEX_WAIT_REQUEUE_PI:
3434 val3 = FUTEX_BITSET_MATCH_ANY;
Thomas Gleixner81b40532012-02-15 12:17:09 +01003435 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3436 uaddr2);
Darren Hart52400ba2009-04-03 13:40:49 -07003437 case FUTEX_CMP_REQUEUE_PI:
Thomas Gleixner81b40532012-02-15 12:17:09 +01003438 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 }
Thomas Gleixner81b40532012-02-15 12:17:09 +01003440 return -ENOSYS;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441}
3442
3443
Heiko Carstens17da2bd2009-01-14 14:14:10 +01003444SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3445 struct timespec __user *, utime, u32 __user *, uaddr2,
3446 u32, val3)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447{
Pierre Peifferc19384b2007-05-09 02:35:02 -07003448 struct timespec ts;
3449 ktime_t t, *tp = NULL;
Ingo Molnare2970f22006-06-27 02:54:47 -07003450 u32 val2 = 0;
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003451 int cmd = op & FUTEX_CMD_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452
Thomas Gleixnercd689982008-02-01 17:45:14 +01003453 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
Darren Hart52400ba2009-04-03 13:40:49 -07003454 cmd == FUTEX_WAIT_BITSET ||
3455 cmd == FUTEX_WAIT_REQUEUE_PI)) {
Davidlohr Buesoab51fba2015-06-29 23:26:02 -07003456 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3457 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003458 if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459 return -EFAULT;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003460 if (!timespec_valid(&ts))
Thomas Gleixner9741ef962006-03-31 02:31:32 -08003461 return -EINVAL;
Pierre Peifferc19384b2007-05-09 02:35:02 -07003462
3463 t = timespec_to_ktime(ts);
Eric Dumazet34f01cc2007-05-09 02:35:04 -07003464 if (cmd == FUTEX_WAIT)
Thomas Gleixner5a7780e2008-02-13 09:20:43 +01003465 t = ktime_add_safe(ktime_get(), t);
Pierre Peifferc19384b2007-05-09 02:35:02 -07003466 tp = &t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 }
3468 /*
Darren Hart52400ba2009-04-03 13:40:49 -07003469 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
Andreas Schwabf54f0982007-07-31 00:38:51 -07003470 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 */
Andreas Schwabf54f0982007-07-31 00:38:51 -07003472 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
Darren Hartba9c22f2009-04-20 22:22:22 -07003473 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
Ingo Molnare2970f22006-06-27 02:54:47 -07003474 val2 = (u32) (unsigned long) utime;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475
Pierre Peifferc19384b2007-05-09 02:35:02 -07003476 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477}
3478
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003479static void __init futex_detect_cmpxchg(void)
3480{
3481#ifndef CONFIG_HAVE_FUTEX_CMPXCHG
3482 u32 curval;
3483
3484 /*
3485 * This will fail and we want it. Some arch implementations do
3486 * runtime detection of the futex_atomic_cmpxchg_inatomic()
3487 * functionality. We want to know that before we call in any
3488 * of the complex code paths. Also we want to prevent
3489 * registration of robust lists in that case. NULL is
3490 * guaranteed to fault and we get -EFAULT on functional
3491 * implementation, the non-functional ones will return
3492 * -ENOSYS.
3493 */
3494 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
3495 futex_cmpxchg_enabled = 1;
3496#endif
3497}
3498
Benjamin Herrenschmidtf6d107f2008-03-27 14:52:15 +11003499static int __init futex_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500{
Heiko Carstens63b1a812014-01-16 14:54:50 +01003501 unsigned int futex_shift;
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003502 unsigned long i;
3503
3504#if CONFIG_BASE_SMALL
3505 futex_hashsize = 16;
3506#else
3507 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
3508#endif
3509
3510 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
3511 futex_hashsize, 0,
3512 futex_hashsize < 256 ? HASH_SMALL : 0,
Heiko Carstens63b1a812014-01-16 14:54:50 +01003513 &futex_shift, NULL,
3514 futex_hashsize, futex_hashsize);
3515 futex_hashsize = 1UL << futex_shift;
Heiko Carstens03b8c7b2014-03-02 13:09:47 +01003516
3517 futex_detect_cmpxchg();
Thomas Gleixnera0c1e902008-02-23 15:23:57 -08003518
Davidlohr Buesoa52b89e2014-01-12 15:31:23 -08003519 for (i = 0; i < futex_hashsize; i++) {
Linus Torvalds11d46162014-03-20 22:11:17 -07003520 atomic_set(&futex_queues[i].waiters, 0);
Dima Zavin732375c2011-07-07 17:27:59 -07003521 plist_head_init(&futex_queues[i].chain);
Thomas Gleixner3e4ab742008-02-23 15:23:55 -08003522 spin_lock_init(&futex_queues[i].lock);
3523 }
3524
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 return 0;
3526}
Yang Yang25f71d12016-12-30 16:17:55 +08003527core_initcall(futex_init);